diff --git a/sys/dev/cxgbe/crypto/t6_kern_tls.c b/sys/dev/cxgbe/crypto/t6_kern_tls.c index 1374530f941f..4340c5484763 100644 --- a/sys/dev/cxgbe/crypto/t6_kern_tls.c +++ b/sys/dev/cxgbe/crypto/t6_kern_tls.c @@ -1,2151 +1,2151 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2018-2019 Chelsio Communications, Inc. * All rights reserved. * Written by: John Baldwin * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include "opt_inet.h" #include "opt_inet6.h" #include "opt_kern_tls.h" #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include "common/common.h" #include "common/t4_regs.h" #include "common/t4_regs_values.h" #include "common/t4_tcb.h" #include "t4_l2t.h" #include "t4_clip.h" #include "t4_mp_ring.h" #include "crypto/t4_crypto.h" #if defined(INET) || defined(INET6) #define TLS_HEADER_LENGTH 5 struct tls_scmd { __be32 seqno_numivs; __be32 ivgen_hdrlen; }; struct tlspcb { struct m_snd_tag com; struct vi_info *vi; /* virtual interface */ struct adapter *sc; struct l2t_entry *l2te; /* L2 table entry used by this connection */ int tid; /* Connection identifier */ int tx_key_addr; bool inline_key; bool using_timestamps; unsigned char enc_mode; struct tls_scmd scmd0; struct tls_scmd scmd0_short; unsigned int tx_key_info_size; uint32_t prev_seq; uint32_t prev_ack; uint32_t prev_tsecr; uint16_t prev_win; uint16_t prev_mss; /* Only used outside of setup and teardown when using inline keys. */ struct tls_keyctx keyctx; /* Fields only used during setup and teardown. */ struct inpcb *inp; /* backpointer to host stack's PCB */ struct sge_txq *txq; struct sge_wrq *ctrlq; struct clip_entry *ce; /* CLIP table entry used by this tid */ bool open_pending; }; static void t6_tls_tag_free(struct m_snd_tag *mst); static int ktls_setup_keys(struct tlspcb *tlsp, const struct ktls_session *tls, struct sge_txq *txq); static const struct if_snd_tag_sw t6_tls_tag_sw = { .snd_tag_free = t6_tls_tag_free, .type = IF_SND_TAG_TYPE_TLS }; static inline struct tlspcb * mst_to_tls(struct m_snd_tag *t) { return (__containerof(t, struct tlspcb, com)); } static struct tlspcb * alloc_tlspcb(struct ifnet *ifp, struct vi_info *vi, int flags) { struct port_info *pi = vi->pi; struct adapter *sc = pi->adapter; struct tlspcb *tlsp; tlsp = malloc(sizeof(*tlsp), M_CXGBE, M_ZERO | flags); if (tlsp == NULL) return (NULL); m_snd_tag_init(&tlsp->com, ifp, &t6_tls_tag_sw); tlsp->vi = vi; tlsp->sc = sc; tlsp->ctrlq = &sc->sge.ctrlq[pi->port_id]; tlsp->tid = -1; tlsp->tx_key_addr = -1; return (tlsp); } static int ktls_act_open_cpl_size(bool isipv6) { if (isipv6) return (sizeof(struct cpl_t6_act_open_req6)); else return (sizeof(struct cpl_t6_act_open_req)); } static void mk_ktls_act_open_req(struct adapter *sc, struct vi_info *vi, struct inpcb *inp, struct tlspcb *tlsp, int atid, void *dst) { struct tcpcb *tp = intotcpcb(inp); struct cpl_t6_act_open_req *cpl6; struct cpl_act_open_req *cpl; uint64_t options; int qid_atid; cpl6 = dst; cpl = (struct cpl_act_open_req *)cpl6; INIT_TP_WR(cpl6, 0); qid_atid = V_TID_QID(sc->sge.fwq.abs_id) | V_TID_TID(atid) | V_TID_COOKIE(CPL_COOKIE_KERN_TLS); OPCODE_TID(cpl) = htobe32(MK_OPCODE_TID(CPL_ACT_OPEN_REQ, qid_atid)); inp_4tuple_get(inp, &cpl->local_ip, &cpl->local_port, &cpl->peer_ip, &cpl->peer_port); options = F_TCAM_BYPASS | V_ULP_MODE(ULP_MODE_NONE); options |= V_SMAC_SEL(vi->smt_idx) | V_TX_CHAN(vi->pi->tx_chan); options |= F_NON_OFFLOAD; cpl->opt0 = htobe64(options); options = V_TX_QUEUE(sc->params.tp.tx_modq[vi->pi->tx_chan]); if (tp->t_flags & TF_REQ_TSTMP) options |= F_TSTAMPS_EN; cpl->opt2 = htobe32(options); } static void mk_ktls_act_open_req6(struct adapter *sc, struct vi_info *vi, struct inpcb *inp, struct tlspcb *tlsp, int atid, void *dst) { struct tcpcb *tp = intotcpcb(inp); struct cpl_t6_act_open_req6 *cpl6; struct cpl_act_open_req6 *cpl; uint64_t options; int qid_atid; cpl6 = dst; cpl = (struct cpl_act_open_req6 *)cpl6; INIT_TP_WR(cpl6, 0); qid_atid = V_TID_QID(sc->sge.fwq.abs_id) | V_TID_TID(atid) | V_TID_COOKIE(CPL_COOKIE_KERN_TLS); OPCODE_TID(cpl) = htobe32(MK_OPCODE_TID(CPL_ACT_OPEN_REQ6, qid_atid)); cpl->local_port = inp->inp_lport; cpl->local_ip_hi = *(uint64_t *)&inp->in6p_laddr.s6_addr[0]; cpl->local_ip_lo = *(uint64_t *)&inp->in6p_laddr.s6_addr[8]; cpl->peer_port = inp->inp_fport; cpl->peer_ip_hi = *(uint64_t *)&inp->in6p_faddr.s6_addr[0]; cpl->peer_ip_lo = *(uint64_t *)&inp->in6p_faddr.s6_addr[8]; options = F_TCAM_BYPASS | V_ULP_MODE(ULP_MODE_NONE); options |= V_SMAC_SEL(vi->smt_idx) | V_TX_CHAN(vi->pi->tx_chan); options |= F_NON_OFFLOAD; cpl->opt0 = htobe64(options); options = V_TX_QUEUE(sc->params.tp.tx_modq[vi->pi->tx_chan]); if (tp->t_flags & TF_REQ_TSTMP) options |= F_TSTAMPS_EN; cpl->opt2 = htobe32(options); } static int send_ktls_act_open_req(struct adapter *sc, struct vi_info *vi, struct inpcb *inp, struct tlspcb *tlsp, int atid) { struct wrqe *wr; bool isipv6; isipv6 = (inp->inp_vflag & INP_IPV6) != 0; if (isipv6) { tlsp->ce = t4_get_clip_entry(sc, &inp->in6p_laddr, true); if (tlsp->ce == NULL) return (ENOENT); } wr = alloc_wrqe(ktls_act_open_cpl_size(isipv6), tlsp->ctrlq); if (wr == NULL) { CTR2(KTR_CXGBE, "%s: atid %d failed to alloc WR", __func__, atid); return (ENOMEM); } if (isipv6) mk_ktls_act_open_req6(sc, vi, inp, tlsp, atid, wrtod(wr)); else mk_ktls_act_open_req(sc, vi, inp, tlsp, atid, wrtod(wr)); tlsp->open_pending = true; t4_wrq_tx(sc, wr); return (0); } static int ktls_act_open_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { struct adapter *sc = iq->adapter; const struct cpl_act_open_rpl *cpl = (const void *)(rss + 1); u_int atid = G_TID_TID(G_AOPEN_ATID(be32toh(cpl->atid_status))); u_int status = G_AOPEN_STATUS(be32toh(cpl->atid_status)); struct tlspcb *tlsp = lookup_atid(sc, atid); struct inpcb *inp = tlsp->inp; CTR3(KTR_CXGBE, "%s: atid %d status %d", __func__, atid, status); free_atid(sc, atid); if (status == 0) tlsp->tid = GET_TID(cpl); INP_WLOCK(inp); tlsp->open_pending = false; wakeup(tlsp); INP_WUNLOCK(inp); return (0); } /* SET_TCB_FIELD sent as a ULP command looks like this */ #define LEN__SET_TCB_FIELD_ULP (sizeof(struct ulp_txpkt) + \ sizeof(struct ulptx_idata) + sizeof(struct cpl_set_tcb_field_core)) _Static_assert((LEN__SET_TCB_FIELD_ULP + sizeof(struct ulptx_idata)) % 16 == 0, "CPL_SET_TCB_FIELD ULP command not 16-byte aligned"); static void write_set_tcb_field_ulp(struct tlspcb *tlsp, void *dst, struct sge_txq *txq, uint16_t word, uint64_t mask, uint64_t val) { struct ulp_txpkt *txpkt; struct ulptx_idata *idata; struct cpl_set_tcb_field_core *cpl; /* ULP_TXPKT */ txpkt = dst; txpkt->cmd_dest = htobe32(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DATAMODIFY(0) | V_ULP_TXPKT_CHANNELID(tlsp->vi->pi->port_id) | V_ULP_TXPKT_DEST(0) | V_ULP_TXPKT_FID(txq->eq.cntxt_id) | V_ULP_TXPKT_RO(1)); txpkt->len = htobe32(howmany(LEN__SET_TCB_FIELD_ULP, 16)); /* ULPTX_IDATA sub-command */ idata = (struct ulptx_idata *)(txpkt + 1); idata->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM)); idata->len = htobe32(sizeof(*cpl)); /* CPL_SET_TCB_FIELD */ cpl = (struct cpl_set_tcb_field_core *)(idata + 1); OPCODE_TID(cpl) = htobe32(MK_OPCODE_TID(CPL_SET_TCB_FIELD, tlsp->tid)); cpl->reply_ctrl = htobe16(F_NO_REPLY); cpl->word_cookie = htobe16(V_WORD(word)); cpl->mask = htobe64(mask); cpl->val = htobe64(val); /* ULPTX_NOOP */ idata = (struct ulptx_idata *)(cpl + 1); idata->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_NOOP)); idata->len = htobe32(0); } static int ktls_set_tcb_fields(struct tlspcb *tlsp, struct tcpcb *tp, struct sge_txq *txq) { struct fw_ulptx_wr *wr; struct mbuf *m; char *dst; void *items[1]; int error, len; len = sizeof(*wr) + 3 * roundup2(LEN__SET_TCB_FIELD_ULP, 16); if (tp->t_flags & TF_REQ_TSTMP) len += roundup2(LEN__SET_TCB_FIELD_ULP, 16); m = alloc_wr_mbuf(len, M_NOWAIT); if (m == NULL) { CTR2(KTR_CXGBE, "%s: tid %d failed to alloc WR mbuf", __func__, tlsp->tid); return (ENOMEM); } m->m_pkthdr.snd_tag = m_snd_tag_ref(&tlsp->com); m->m_pkthdr.csum_flags |= CSUM_SND_TAG; /* FW_ULPTX_WR */ wr = mtod(m, void *); wr->op_to_compl = htobe32(V_FW_WR_OP(FW_ULPTX_WR)); wr->flowid_len16 = htobe32(F_FW_ULPTX_WR_DATA | V_FW_WR_LEN16(len / 16)); wr->cookie = 0; dst = (char *)(wr + 1); /* Clear TF_NON_OFFLOAD and set TF_CORE_BYPASS */ write_set_tcb_field_ulp(tlsp, dst, txq, W_TCB_T_FLAGS, V_TCB_T_FLAGS(V_TF_CORE_BYPASS(1) | V_TF_NON_OFFLOAD(1)), V_TCB_T_FLAGS(V_TF_CORE_BYPASS(1))); dst += roundup2(LEN__SET_TCB_FIELD_ULP, 16); /* Clear the SND_UNA_RAW, SND_NXT_RAW, and SND_MAX_RAW offsets. */ write_set_tcb_field_ulp(tlsp, dst, txq, W_TCB_SND_UNA_RAW, V_TCB_SND_NXT_RAW(M_TCB_SND_NXT_RAW) | V_TCB_SND_UNA_RAW(M_TCB_SND_UNA_RAW), V_TCB_SND_NXT_RAW(0) | V_TCB_SND_UNA_RAW(0)); dst += roundup2(LEN__SET_TCB_FIELD_ULP, 16); write_set_tcb_field_ulp(tlsp, dst, txq, W_TCB_SND_MAX_RAW, V_TCB_SND_MAX_RAW(M_TCB_SND_MAX_RAW), V_TCB_SND_MAX_RAW(0)); dst += roundup2(LEN__SET_TCB_FIELD_ULP, 16); if (tp->t_flags & TF_REQ_TSTMP) { write_set_tcb_field_ulp(tlsp, dst, txq, W_TCB_TIMESTAMP_OFFSET, V_TCB_TIMESTAMP_OFFSET(M_TCB_TIMESTAMP_OFFSET), V_TCB_TIMESTAMP_OFFSET(tp->ts_offset >> 28)); dst += roundup2(LEN__SET_TCB_FIELD_ULP, 16); } KASSERT(dst - (char *)wr == len, ("%s: length mismatch", __func__)); items[0] = m; error = mp_ring_enqueue(txq->r, items, 1, 1); if (error) m_free(m); return (error); } int t6_tls_tag_alloc(struct ifnet *ifp, union if_snd_tag_alloc_params *params, struct m_snd_tag **pt) { const struct ktls_session *tls; struct tlspcb *tlsp; struct adapter *sc; struct vi_info *vi; struct inpcb *inp; struct tcpcb *tp; struct sge_txq *txq; int atid, error, explicit_iv_size, keyid, mac_first; tls = params->tls.tls; /* Only TLS 1.1 and TLS 1.2 are currently supported. */ if (tls->params.tls_vmajor != TLS_MAJOR_VER_ONE || tls->params.tls_vminor < TLS_MINOR_VER_ONE || tls->params.tls_vminor > TLS_MINOR_VER_TWO) return (EPROTONOSUPPORT); /* Sanity check values in *tls. */ switch (tls->params.cipher_algorithm) { case CRYPTO_AES_CBC: /* XXX: Explicitly ignore any provided IV. */ switch (tls->params.cipher_key_len) { case 128 / 8: case 192 / 8: case 256 / 8: break; default: return (EINVAL); } switch (tls->params.auth_algorithm) { case CRYPTO_SHA1_HMAC: case CRYPTO_SHA2_256_HMAC: case CRYPTO_SHA2_384_HMAC: break; default: return (EPROTONOSUPPORT); } explicit_iv_size = AES_BLOCK_LEN; mac_first = 1; break; case CRYPTO_AES_NIST_GCM_16: if (tls->params.iv_len != SALT_SIZE) return (EINVAL); switch (tls->params.cipher_key_len) { case 128 / 8: case 192 / 8: case 256 / 8: break; default: return (EINVAL); } explicit_iv_size = 8; mac_first = 0; break; default: return (EPROTONOSUPPORT); } vi = ifp->if_softc; sc = vi->adapter; tlsp = alloc_tlspcb(ifp, vi, M_WAITOK); atid = alloc_atid(sc, tlsp); if (atid < 0) { error = ENOMEM; goto failed; } if (sc->tlst.inline_keys) keyid = -1; else keyid = t4_alloc_tls_keyid(sc); if (keyid < 0) { CTR2(KTR_CXGBE, "%s: atid %d using immediate key ctx", __func__, atid); tlsp->inline_key = true; } else { tlsp->tx_key_addr = keyid; CTR3(KTR_CXGBE, "%s: atid %d allocated TX key addr %#x", __func__, atid, tlsp->tx_key_addr); } inp = params->tls.inp; INP_RLOCK(inp); - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { INP_RUNLOCK(inp); error = ECONNRESET; goto failed; } tlsp->inp = inp; tp = inp->inp_ppcb; if (tp->t_flags & TF_REQ_TSTMP) { tlsp->using_timestamps = true; if ((tp->ts_offset & 0xfffffff) != 0) { INP_RUNLOCK(inp); error = EINVAL; goto failed; } } else tlsp->using_timestamps = false; error = send_ktls_act_open_req(sc, vi, inp, tlsp, atid); if (error) { INP_RUNLOCK(inp); goto failed; } /* Wait for reply to active open. */ CTR2(KTR_CXGBE, "%s: atid %d sent CPL_ACT_OPEN_REQ", __func__, atid); while (tlsp->open_pending) { /* * XXX: PCATCH? We would then have to discard the PCB * when the completion CPL arrived. */ error = rw_sleep(tlsp, &inp->inp_lock, 0, "t6tlsop", 0); } atid = -1; if (tlsp->tid < 0) { INP_RUNLOCK(inp); error = ENOMEM; goto failed; } - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { INP_RUNLOCK(inp); error = ECONNRESET; goto failed; } txq = &sc->sge.txq[vi->first_txq]; if (inp->inp_flowtype != M_HASHTYPE_NONE) txq += ((inp->inp_flowid % (vi->ntxq - vi->rsrv_noflowq)) + vi->rsrv_noflowq); tlsp->txq = txq; error = ktls_set_tcb_fields(tlsp, tp, txq); INP_RUNLOCK(inp); if (error) goto failed; error = ktls_setup_keys(tlsp, tls, txq); if (error) goto failed; tlsp->enc_mode = t4_tls_cipher_mode(tls); tlsp->tx_key_info_size = t4_tls_key_info_size(tls); /* The SCMD fields used when encrypting a full TLS record. */ tlsp->scmd0.seqno_numivs = htobe32(V_SCMD_SEQ_NO_CTRL(3) | V_SCMD_PROTO_VERSION(t4_tls_proto_ver(tls)) | V_SCMD_ENC_DEC_CTRL(SCMD_ENCDECCTRL_ENCRYPT) | V_SCMD_CIPH_AUTH_SEQ_CTRL((mac_first == 0)) | V_SCMD_CIPH_MODE(tlsp->enc_mode) | V_SCMD_AUTH_MODE(t4_tls_auth_mode(tls)) | V_SCMD_HMAC_CTRL(t4_tls_hmac_ctrl(tls)) | V_SCMD_IV_SIZE(explicit_iv_size / 2) | V_SCMD_NUM_IVS(1)); tlsp->scmd0.ivgen_hdrlen = V_SCMD_IV_GEN_CTRL(0) | V_SCMD_TLS_FRAG_ENABLE(0); if (tlsp->inline_key) tlsp->scmd0.ivgen_hdrlen |= V_SCMD_KEY_CTX_INLINE(1); tlsp->scmd0.ivgen_hdrlen = htobe32(tlsp->scmd0.ivgen_hdrlen); /* * The SCMD fields used when encrypting a partial TLS record * (no trailer and possibly a truncated payload). */ tlsp->scmd0_short.seqno_numivs = V_SCMD_SEQ_NO_CTRL(0) | V_SCMD_PROTO_VERSION(SCMD_PROTO_VERSION_GENERIC) | V_SCMD_ENC_DEC_CTRL(SCMD_ENCDECCTRL_ENCRYPT) | V_SCMD_CIPH_AUTH_SEQ_CTRL((mac_first == 0)) | V_SCMD_AUTH_MODE(SCMD_AUTH_MODE_NOP) | V_SCMD_HMAC_CTRL(SCMD_HMAC_CTRL_NOP) | V_SCMD_IV_SIZE(AES_BLOCK_LEN / 2) | V_SCMD_NUM_IVS(0); if (tlsp->enc_mode == SCMD_CIPH_MODE_AES_GCM) tlsp->scmd0_short.seqno_numivs |= V_SCMD_CIPH_MODE(SCMD_CIPH_MODE_AES_CTR); else tlsp->scmd0_short.seqno_numivs |= V_SCMD_CIPH_MODE(tlsp->enc_mode); tlsp->scmd0_short.seqno_numivs = htobe32(tlsp->scmd0_short.seqno_numivs); tlsp->scmd0_short.ivgen_hdrlen = V_SCMD_IV_GEN_CTRL(0) | V_SCMD_TLS_FRAG_ENABLE(0) | V_SCMD_AADIVDROP(1); if (tlsp->inline_key) tlsp->scmd0_short.ivgen_hdrlen |= V_SCMD_KEY_CTX_INLINE(1); TXQ_LOCK(txq); if (tlsp->enc_mode == SCMD_CIPH_MODE_AES_GCM) txq->kern_tls_gcm++; else txq->kern_tls_cbc++; TXQ_UNLOCK(txq); *pt = &tlsp->com; return (0); failed: if (atid >= 0) free_atid(sc, atid); m_snd_tag_rele(&tlsp->com); return (error); } static int ktls_setup_keys(struct tlspcb *tlsp, const struct ktls_session *tls, struct sge_txq *txq) { struct tls_key_req *kwr; struct tls_keyctx *kctx; void *items[1]; struct mbuf *m; int error; /* * Store the salt and keys in the key context. For * connections with an inline key, this key context is passed * as immediate data in each work request. For connections * storing the key in DDR, a work request is used to store a * copy of the key context in DDR. */ t4_tls_key_ctx(tls, KTLS_TX, &tlsp->keyctx); if (tlsp->inline_key) return (0); /* Populate key work request. */ m = alloc_wr_mbuf(TLS_KEY_WR_SZ, M_NOWAIT); if (m == NULL) { CTR2(KTR_CXGBE, "%s: tid %d failed to alloc WR mbuf", __func__, tlsp->tid); return (ENOMEM); } m->m_pkthdr.snd_tag = m_snd_tag_ref(&tlsp->com); m->m_pkthdr.csum_flags |= CSUM_SND_TAG; kwr = mtod(m, void *); memset(kwr, 0, TLS_KEY_WR_SZ); t4_write_tlskey_wr(tls, KTLS_TX, tlsp->tid, 0, tlsp->tx_key_addr, kwr); kctx = (struct tls_keyctx *)(kwr + 1); memcpy(kctx, &tlsp->keyctx, sizeof(*kctx)); /* * Place the key work request in the transmit queue. It * should be sent to the NIC before any TLS packets using this * session. */ items[0] = m; error = mp_ring_enqueue(txq->r, items, 1, 1); if (error) m_free(m); else CTR2(KTR_CXGBE, "%s: tid %d sent key WR", __func__, tlsp->tid); return (error); } static u_int ktls_base_wr_size(struct tlspcb *tlsp) { u_int wr_len; wr_len = sizeof(struct fw_ulptx_wr); // 16 wr_len += sizeof(struct ulp_txpkt); // 8 wr_len += sizeof(struct ulptx_idata); // 8 wr_len += sizeof(struct cpl_tx_sec_pdu);// 32 if (tlsp->inline_key) wr_len += tlsp->tx_key_info_size; else { wr_len += sizeof(struct ulptx_sc_memrd);// 8 wr_len += sizeof(struct ulptx_idata); // 8 } wr_len += sizeof(struct cpl_tx_data); // 16 return (wr_len); } /* How many bytes of TCP payload to send for a given TLS record. */ static u_int ktls_tcp_payload_length(struct tlspcb *tlsp, struct mbuf *m_tls) { struct tls_record_layer *hdr; u_int plen, mlen; M_ASSERTEXTPG(m_tls); hdr = (void *)m_tls->m_epg_hdr; plen = ntohs(hdr->tls_length); /* * What range of the TLS record is the mbuf requesting to be * sent. */ mlen = mtod(m_tls, vm_offset_t) + m_tls->m_len; /* Always send complete records. */ if (mlen == TLS_HEADER_LENGTH + plen) return (mlen); /* * If the host stack has asked to send part of the trailer, * trim the length to avoid sending any of the trailer. There * is no way to send a partial trailer currently. */ if (mlen > TLS_HEADER_LENGTH + plen - m_tls->m_epg_trllen) mlen = TLS_HEADER_LENGTH + plen - m_tls->m_epg_trllen; /* * For AES-CBC adjust the ciphertext length for the block * size. */ if (tlsp->enc_mode == SCMD_CIPH_MODE_AES_CBC && mlen > TLS_HEADER_LENGTH) { mlen = TLS_HEADER_LENGTH + rounddown(mlen - TLS_HEADER_LENGTH, AES_BLOCK_LEN); } #ifdef VERBOSE_TRACES CTR4(KTR_CXGBE, "%s: tid %d short TLS record (%u vs %u)", __func__, tlsp->tid, mlen, TLS_HEADER_LENGTH + plen); #endif return (mlen); } /* * For a "short" TLS record, determine the offset into the TLS record * payload to send. This offset does not include the TLS header, but * a non-zero offset implies that a header will not be sent. */ static u_int ktls_payload_offset(struct tlspcb *tlsp, struct mbuf *m_tls) { struct tls_record_layer *hdr; u_int offset, plen; #ifdef INVARIANTS u_int mlen; #endif M_ASSERTEXTPG(m_tls); hdr = (void *)m_tls->m_epg_hdr; plen = ntohs(hdr->tls_length); #ifdef INVARIANTS mlen = mtod(m_tls, vm_offset_t) + m_tls->m_len; MPASS(mlen < TLS_HEADER_LENGTH + plen); #endif if (mtod(m_tls, vm_offset_t) <= m_tls->m_epg_hdrlen) return (0); if (tlsp->enc_mode == SCMD_CIPH_MODE_AES_GCM) { /* * Always send something. This function is only called * if we aren't sending the tag at all, but if the * request starts in the tag then we are in an odd * state where would effectively send nothing. Cap * the offset at the last byte of the record payload * to send the last cipher block. */ offset = min(mtod(m_tls, vm_offset_t) - m_tls->m_epg_hdrlen, (plen - TLS_HEADER_LENGTH - m_tls->m_epg_trllen) - 1); return (rounddown(offset, AES_BLOCK_LEN)); } return (0); } static u_int ktls_sgl_size(u_int nsegs) { u_int wr_len; /* First segment is part of ulptx_sgl. */ nsegs--; wr_len = sizeof(struct ulptx_sgl); wr_len += 8 * ((3 * nsegs) / 2 + (nsegs & 1)); return (wr_len); } static int ktls_wr_len(struct tlspcb *tlsp, struct mbuf *m, struct mbuf *m_tls, int *nsegsp) { struct tls_record_layer *hdr; u_int imm_len, offset, plen, wr_len, tlen; M_ASSERTEXTPG(m_tls); /* * Determine the size of the TLS record payload to send * excluding header and trailer. */ tlen = ktls_tcp_payload_length(tlsp, m_tls); if (tlen <= m_tls->m_epg_hdrlen) { /* * For requests that only want to send the TLS header, * send a tunnelled packet as immediate data. */ wr_len = sizeof(struct fw_eth_tx_pkt_wr) + sizeof(struct cpl_tx_pkt_core) + roundup2(m->m_len + m_tls->m_len, 16); if (wr_len > SGE_MAX_WR_LEN) { CTR3(KTR_CXGBE, "%s: tid %d TLS header-only packet too long (len %d)", __func__, tlsp->tid, m->m_len + m_tls->m_len); } /* This should always be the last TLS record in a chain. */ MPASS(m_tls->m_next == NULL); /* * XXX: Set a bogus 'nsegs' value to avoid tripping an * assertion in mbuf_nsegs() in t4_sge.c. */ *nsegsp = 1; return (wr_len); } hdr = (void *)m_tls->m_epg_hdr; plen = TLS_HEADER_LENGTH + ntohs(hdr->tls_length) - m_tls->m_epg_trllen; if (tlen < plen) { plen = tlen; offset = ktls_payload_offset(tlsp, m_tls); } else offset = 0; /* Calculate the size of the work request. */ wr_len = ktls_base_wr_size(tlsp); /* * Full records and short records with an offset of 0 include * the TLS header as immediate data. Short records include a * raw AES IV as immediate data. */ imm_len = 0; if (offset == 0) imm_len += m_tls->m_epg_hdrlen; if (plen == tlen) imm_len += AES_BLOCK_LEN; wr_len += roundup2(imm_len, 16); /* TLS record payload via DSGL. */ *nsegsp = sglist_count_mbuf_epg(m_tls, m_tls->m_epg_hdrlen + offset, plen - (m_tls->m_epg_hdrlen + offset)); wr_len += ktls_sgl_size(*nsegsp); wr_len = roundup2(wr_len, 16); return (wr_len); } /* * See if we have any TCP options requiring a dedicated options-only * packet. */ static int ktls_has_tcp_options(struct tcphdr *tcp) { u_char *cp; int cnt, opt, optlen; cp = (u_char *)(tcp + 1); cnt = tcp->th_off * 4 - sizeof(struct tcphdr); for (; cnt > 0; cnt -= optlen, cp += optlen) { opt = cp[0]; if (opt == TCPOPT_EOL) break; if (opt == TCPOPT_NOP) optlen = 1; else { if (cnt < 2) break; optlen = cp[1]; if (optlen < 2 || optlen > cnt) break; } switch (opt) { case TCPOPT_NOP: case TCPOPT_TIMESTAMP: break; default: return (1); } } return (0); } /* * Find the TCP timestamp option. */ static void * ktls_find_tcp_timestamps(struct tcphdr *tcp) { u_char *cp; int cnt, opt, optlen; cp = (u_char *)(tcp + 1); cnt = tcp->th_off * 4 - sizeof(struct tcphdr); for (; cnt > 0; cnt -= optlen, cp += optlen) { opt = cp[0]; if (opt == TCPOPT_EOL) break; if (opt == TCPOPT_NOP) optlen = 1; else { if (cnt < 2) break; optlen = cp[1]; if (optlen < 2 || optlen > cnt) break; } if (opt == TCPOPT_TIMESTAMP && optlen == TCPOLEN_TIMESTAMP) return (cp + 2); } return (NULL); } int t6_ktls_parse_pkt(struct mbuf *m, int *nsegsp, int *len16p) { struct tlspcb *tlsp; struct ether_header *eh; struct ip *ip; struct ip6_hdr *ip6; struct tcphdr *tcp; struct mbuf *m_tls; int nsegs; u_int wr_len, tot_len; /* * Locate headers in initial mbuf. * * XXX: This assumes all of the headers are in the initial mbuf. * Could perhaps use m_advance() like parse_pkt() if that turns * out to not be true. */ M_ASSERTPKTHDR(m); MPASS(m->m_pkthdr.snd_tag != NULL); tlsp = mst_to_tls(m->m_pkthdr.snd_tag); if (m->m_len <= sizeof(*eh) + sizeof(*ip)) { CTR2(KTR_CXGBE, "%s: tid %d header mbuf too short", __func__, tlsp->tid); return (EINVAL); } eh = mtod(m, struct ether_header *); if (ntohs(eh->ether_type) != ETHERTYPE_IP && ntohs(eh->ether_type) != ETHERTYPE_IPV6) { CTR2(KTR_CXGBE, "%s: tid %d mbuf not ETHERTYPE_IP{,V6}", __func__, tlsp->tid); return (EINVAL); } m->m_pkthdr.l2hlen = sizeof(*eh); /* XXX: Reject unsupported IP options? */ if (ntohs(eh->ether_type) == ETHERTYPE_IP) { ip = (struct ip *)(eh + 1); if (ip->ip_p != IPPROTO_TCP) { CTR2(KTR_CXGBE, "%s: tid %d mbuf not IPPROTO_TCP", __func__, tlsp->tid); return (EINVAL); } m->m_pkthdr.l3hlen = ip->ip_hl * 4; } else { ip6 = (struct ip6_hdr *)(eh + 1); if (ip6->ip6_nxt != IPPROTO_TCP) { CTR3(KTR_CXGBE, "%s: tid %d mbuf not IPPROTO_TCP (%u)", __func__, tlsp->tid, ip6->ip6_nxt); return (EINVAL); } m->m_pkthdr.l3hlen = sizeof(struct ip6_hdr); } if (m->m_len < m->m_pkthdr.l2hlen + m->m_pkthdr.l3hlen + sizeof(*tcp)) { CTR2(KTR_CXGBE, "%s: tid %d header mbuf too short (2)", __func__, tlsp->tid); return (EINVAL); } tcp = (struct tcphdr *)((char *)(eh + 1) + m->m_pkthdr.l3hlen); m->m_pkthdr.l4hlen = tcp->th_off * 4; /* Bail if there is TCP payload before the TLS record. */ if (m->m_len != m->m_pkthdr.l2hlen + m->m_pkthdr.l3hlen + m->m_pkthdr.l4hlen) { CTR6(KTR_CXGBE, "%s: tid %d header mbuf bad length (%d + %d + %d != %d)", __func__, tlsp->tid, m->m_pkthdr.l2hlen, m->m_pkthdr.l3hlen, m->m_pkthdr.l4hlen, m->m_len); return (EINVAL); } /* Assume all headers are in 'm' for now. */ MPASS(m->m_next != NULL); MPASS(m->m_next->m_flags & M_EXTPG); tot_len = 0; /* * Each of the remaining mbufs in the chain should reference a * TLS record. */ *nsegsp = 0; for (m_tls = m->m_next; m_tls != NULL; m_tls = m_tls->m_next) { MPASS(m_tls->m_flags & M_EXTPG); wr_len = ktls_wr_len(tlsp, m, m_tls, &nsegs); #ifdef VERBOSE_TRACES CTR4(KTR_CXGBE, "%s: tid %d wr_len %d nsegs %d", __func__, tlsp->tid, wr_len, nsegs); #endif if (wr_len > SGE_MAX_WR_LEN || nsegs > TX_SGL_SEGS) return (EFBIG); tot_len += roundup2(wr_len, EQ_ESIZE); /* * Store 'nsegs' for the first TLS record in the * header mbuf's metadata. */ if (*nsegsp == 0) *nsegsp = nsegs; } MPASS(tot_len != 0); /* * See if we have any TCP options or a FIN requiring a * dedicated packet. */ if ((tcp->th_flags & TH_FIN) != 0 || ktls_has_tcp_options(tcp)) { wr_len = sizeof(struct fw_eth_tx_pkt_wr) + sizeof(struct cpl_tx_pkt_core) + roundup2(m->m_len, 16); if (wr_len > SGE_MAX_WR_LEN) { CTR3(KTR_CXGBE, "%s: tid %d options-only packet too long (len %d)", __func__, tlsp->tid, m->m_len); return (EINVAL); } tot_len += roundup2(wr_len, EQ_ESIZE); } /* Include room for a TP work request to program an L2T entry. */ tot_len += EQ_ESIZE; /* * Include room for a ULPTX work request including up to 5 * CPL_SET_TCB_FIELD commands before the first TLS work * request. */ wr_len = sizeof(struct fw_ulptx_wr) + 5 * roundup2(LEN__SET_TCB_FIELD_ULP, 16); /* * If timestamps are present, reserve 1 more command for * setting the echoed timestamp. */ if (tlsp->using_timestamps) wr_len += roundup2(LEN__SET_TCB_FIELD_ULP, 16); tot_len += roundup2(wr_len, EQ_ESIZE); *len16p = tot_len / 16; #ifdef VERBOSE_TRACES CTR4(KTR_CXGBE, "%s: tid %d len16 %d nsegs %d", __func__, tlsp->tid, *len16p, *nsegsp); #endif return (0); } /* * If the SGL ends on an address that is not 16 byte aligned, this function will * add a 0 filled flit at the end. */ static void write_gl_to_buf(struct sglist *gl, caddr_t to) { struct sglist_seg *seg; __be64 *flitp; struct ulptx_sgl *usgl; int i, nflits, nsegs; KASSERT(((uintptr_t)to & 0xf) == 0, ("%s: SGL must start at a 16 byte boundary: %p", __func__, to)); nsegs = gl->sg_nseg; MPASS(nsegs > 0); nflits = (3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1) + 2; flitp = (__be64 *)to; seg = &gl->sg_segs[0]; usgl = (void *)flitp; usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) | V_ULPTX_NSGE(nsegs)); usgl->len0 = htobe32(seg->ss_len); usgl->addr0 = htobe64(seg->ss_paddr); seg++; for (i = 0; i < nsegs - 1; i++, seg++) { usgl->sge[i / 2].len[i & 1] = htobe32(seg->ss_len); usgl->sge[i / 2].addr[i & 1] = htobe64(seg->ss_paddr); } if (i & 1) usgl->sge[i / 2].len[1] = htobe32(0); flitp += nflits; if (nflits & 1) { MPASS(((uintptr_t)flitp) & 0xf); *flitp++ = 0; } MPASS((((uintptr_t)flitp) & 0xf) == 0); } static inline void copy_to_txd(struct sge_eq *eq, caddr_t from, caddr_t *to, int len) { MPASS((uintptr_t)(*to) >= (uintptr_t)&eq->desc[0]); MPASS((uintptr_t)(*to) < (uintptr_t)&eq->desc[eq->sidx]); if (__predict_true((uintptr_t)(*to) + len <= (uintptr_t)&eq->desc[eq->sidx])) { bcopy(from, *to, len); (*to) += len; if ((uintptr_t)(*to) == (uintptr_t)&eq->desc[eq->sidx]) (*to) = (caddr_t)eq->desc; } else { int portion = (uintptr_t)&eq->desc[eq->sidx] - (uintptr_t)(*to); bcopy(from, *to, portion); from += portion; portion = len - portion; /* remaining */ bcopy(from, (void *)eq->desc, portion); (*to) = (caddr_t)eq->desc + portion; } } static int ktls_write_tcp_options(struct sge_txq *txq, void *dst, struct mbuf *m, u_int available, u_int pidx) { struct tx_sdesc *txsd; struct fw_eth_tx_pkt_wr *wr; struct cpl_tx_pkt_core *cpl; uint32_t ctrl; uint64_t ctrl1; int len16, ndesc, pktlen; struct ether_header *eh; struct ip *ip, newip; struct ip6_hdr *ip6, newip6; struct tcphdr *tcp, newtcp; caddr_t out; TXQ_LOCK_ASSERT_OWNED(txq); M_ASSERTPKTHDR(m); wr = dst; pktlen = m->m_len; ctrl = sizeof(struct cpl_tx_pkt_core) + pktlen; len16 = howmany(sizeof(struct fw_eth_tx_pkt_wr) + ctrl, 16); ndesc = tx_len16_to_desc(len16); MPASS(ndesc <= available); /* Firmware work request header */ wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) | V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl)); ctrl = V_FW_WR_LEN16(len16); wr->equiq_to_len16 = htobe32(ctrl); wr->r3 = 0; cpl = (void *)(wr + 1); /* CPL header */ cpl->ctrl0 = txq->cpl_ctrl0; cpl->pack = 0; cpl->len = htobe16(pktlen); out = (void *)(cpl + 1); /* Copy over Ethernet header. */ eh = mtod(m, struct ether_header *); copy_to_txd(&txq->eq, (caddr_t)eh, &out, m->m_pkthdr.l2hlen); /* Fixup length in IP header and copy out. */ if (ntohs(eh->ether_type) == ETHERTYPE_IP) { ip = (void *)((char *)eh + m->m_pkthdr.l2hlen); newip = *ip; newip.ip_len = htons(pktlen - m->m_pkthdr.l2hlen); copy_to_txd(&txq->eq, (caddr_t)&newip, &out, sizeof(newip)); if (m->m_pkthdr.l3hlen > sizeof(*ip)) copy_to_txd(&txq->eq, (caddr_t)(ip + 1), &out, m->m_pkthdr.l3hlen - sizeof(*ip)); ctrl1 = V_TXPKT_CSUM_TYPE(TX_CSUM_TCPIP) | V_T6_TXPKT_ETHHDR_LEN(m->m_pkthdr.l2hlen - ETHER_HDR_LEN) | V_TXPKT_IPHDR_LEN(m->m_pkthdr.l3hlen); } else { ip6 = (void *)((char *)eh + m->m_pkthdr.l2hlen); newip6 = *ip6; newip6.ip6_plen = htons(pktlen - m->m_pkthdr.l2hlen); copy_to_txd(&txq->eq, (caddr_t)&newip6, &out, sizeof(newip6)); MPASS(m->m_pkthdr.l3hlen == sizeof(*ip6)); ctrl1 = V_TXPKT_CSUM_TYPE(TX_CSUM_TCPIP6) | V_T6_TXPKT_ETHHDR_LEN(m->m_pkthdr.l2hlen - ETHER_HDR_LEN) | V_TXPKT_IPHDR_LEN(m->m_pkthdr.l3hlen); } cpl->ctrl1 = htobe64(ctrl1); txq->txcsum++; /* Clear PUSH and FIN in the TCP header if present. */ tcp = (void *)((char *)eh + m->m_pkthdr.l2hlen + m->m_pkthdr.l3hlen); newtcp = *tcp; newtcp.th_flags &= ~(TH_PUSH | TH_FIN); copy_to_txd(&txq->eq, (caddr_t)&newtcp, &out, sizeof(newtcp)); /* Copy rest of packet. */ copy_to_txd(&txq->eq, (caddr_t)(tcp + 1), &out, pktlen - (m->m_pkthdr.l2hlen + m->m_pkthdr.l3hlen + sizeof(*tcp))); txq->imm_wrs++; txq->txpkt_wrs++; txq->kern_tls_options++; txsd = &txq->sdesc[pidx]; txsd->m = NULL; txsd->desc_used = ndesc; return (ndesc); } static int ktls_write_tunnel_packet(struct sge_txq *txq, void *dst, struct mbuf *m, struct mbuf *m_tls, u_int available, tcp_seq tcp_seqno, u_int pidx) { struct tx_sdesc *txsd; struct fw_eth_tx_pkt_wr *wr; struct cpl_tx_pkt_core *cpl; uint32_t ctrl; uint64_t ctrl1; int len16, ndesc, pktlen; struct ether_header *eh; struct ip *ip, newip; struct ip6_hdr *ip6, newip6; struct tcphdr *tcp, newtcp; caddr_t out; TXQ_LOCK_ASSERT_OWNED(txq); M_ASSERTPKTHDR(m); /* Locate the template TLS header. */ M_ASSERTEXTPG(m_tls); /* This should always be the last TLS record in a chain. */ MPASS(m_tls->m_next == NULL); wr = dst; pktlen = m->m_len + m_tls->m_len; ctrl = sizeof(struct cpl_tx_pkt_core) + pktlen; len16 = howmany(sizeof(struct fw_eth_tx_pkt_wr) + ctrl, 16); ndesc = tx_len16_to_desc(len16); MPASS(ndesc <= available); /* Firmware work request header */ wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) | V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl)); ctrl = V_FW_WR_LEN16(len16); wr->equiq_to_len16 = htobe32(ctrl); wr->r3 = 0; cpl = (void *)(wr + 1); /* CPL header */ cpl->ctrl0 = txq->cpl_ctrl0; cpl->pack = 0; cpl->len = htobe16(pktlen); out = (void *)(cpl + 1); /* Copy over Ethernet header. */ eh = mtod(m, struct ether_header *); copy_to_txd(&txq->eq, (caddr_t)eh, &out, m->m_pkthdr.l2hlen); /* Fixup length in IP header and copy out. */ if (ntohs(eh->ether_type) == ETHERTYPE_IP) { ip = (void *)((char *)eh + m->m_pkthdr.l2hlen); newip = *ip; newip.ip_len = htons(pktlen - m->m_pkthdr.l2hlen); copy_to_txd(&txq->eq, (caddr_t)&newip, &out, sizeof(newip)); if (m->m_pkthdr.l3hlen > sizeof(*ip)) copy_to_txd(&txq->eq, (caddr_t)(ip + 1), &out, m->m_pkthdr.l3hlen - sizeof(*ip)); ctrl1 = V_TXPKT_CSUM_TYPE(TX_CSUM_TCPIP) | V_T6_TXPKT_ETHHDR_LEN(m->m_pkthdr.l2hlen - ETHER_HDR_LEN) | V_TXPKT_IPHDR_LEN(m->m_pkthdr.l3hlen); } else { ip6 = (void *)((char *)eh + m->m_pkthdr.l2hlen); newip6 = *ip6; newip6.ip6_plen = htons(pktlen - m->m_pkthdr.l2hlen); copy_to_txd(&txq->eq, (caddr_t)&newip6, &out, sizeof(newip6)); MPASS(m->m_pkthdr.l3hlen == sizeof(*ip6)); ctrl1 = V_TXPKT_CSUM_TYPE(TX_CSUM_TCPIP6) | V_T6_TXPKT_ETHHDR_LEN(m->m_pkthdr.l2hlen - ETHER_HDR_LEN) | V_TXPKT_IPHDR_LEN(m->m_pkthdr.l3hlen); } cpl->ctrl1 = htobe64(ctrl1); txq->txcsum++; /* Set sequence number in TCP header. */ tcp = (void *)((char *)eh + m->m_pkthdr.l2hlen + m->m_pkthdr.l3hlen); newtcp = *tcp; newtcp.th_seq = htonl(tcp_seqno + mtod(m_tls, vm_offset_t)); copy_to_txd(&txq->eq, (caddr_t)&newtcp, &out, sizeof(newtcp)); /* Copy rest of TCP header. */ copy_to_txd(&txq->eq, (caddr_t)(tcp + 1), &out, m->m_len - (m->m_pkthdr.l2hlen + m->m_pkthdr.l3hlen + sizeof(*tcp))); /* Copy the subset of the TLS header requested. */ copy_to_txd(&txq->eq, (char *)m_tls->m_epg_hdr + mtod(m_tls, vm_offset_t), &out, m_tls->m_len); txq->imm_wrs++; txq->txpkt_wrs++; txq->kern_tls_header++; txsd = &txq->sdesc[pidx]; txsd->m = m; txsd->desc_used = ndesc; return (ndesc); } _Static_assert(sizeof(struct cpl_set_tcb_field) <= EQ_ESIZE, "CPL_SET_TCB_FIELD must be smaller than a single TX descriptor"); _Static_assert(W_TCB_SND_UNA_RAW == W_TCB_SND_NXT_RAW, "SND_NXT_RAW and SND_UNA_RAW are in different words"); static int ktls_write_tls_wr(struct tlspcb *tlsp, struct sge_txq *txq, void *dst, struct mbuf *m, struct tcphdr *tcp, struct mbuf *m_tls, u_int nsegs, u_int available, tcp_seq tcp_seqno, uint32_t *tsopt, u_int pidx, bool set_l2t_idx) { struct sge_eq *eq = &txq->eq; struct tx_sdesc *txsd; struct fw_ulptx_wr *wr; struct ulp_txpkt *txpkt; struct ulptx_sc_memrd *memrd; struct ulptx_idata *idata; struct cpl_tx_sec_pdu *sec_pdu; struct cpl_tx_data *tx_data; struct tls_record_layer *hdr; char *iv, *out; u_int aad_start, aad_stop; u_int auth_start, auth_stop, auth_insert; u_int cipher_start, cipher_stop, iv_offset; u_int imm_len, mss, ndesc, offset, plen, tlen, twr_len, wr_len; u_int fields, tx_max_offset, tx_max; bool first_wr, last_wr, using_scratch; ndesc = 0; MPASS(tlsp->txq == txq); first_wr = (tlsp->prev_seq == 0 && tlsp->prev_ack == 0 && tlsp->prev_win == 0); /* * Use the per-txq scratch pad if near the end of the ring to * simplify handling of wrap-around. This uses a simple but * not quite perfect test of using the scratch buffer if we * can't fit a maximal work request in without wrapping. */ using_scratch = (eq->sidx - pidx < SGE_MAX_WR_LEN / EQ_ESIZE); /* Locate the TLS header. */ M_ASSERTEXTPG(m_tls); hdr = (void *)m_tls->m_epg_hdr; plen = TLS_HEADER_LENGTH + ntohs(hdr->tls_length) - m_tls->m_epg_trllen; /* Determine how much of the TLS record to send. */ tlen = ktls_tcp_payload_length(tlsp, m_tls); if (tlen <= m_tls->m_epg_hdrlen) { /* * For requests that only want to send the TLS header, * send a tunnelled packet as immediate data. */ #ifdef VERBOSE_TRACES CTR3(KTR_CXGBE, "%s: tid %d header-only TLS record %u", __func__, tlsp->tid, (u_int)m_tls->m_epg_seqno); #endif return (ktls_write_tunnel_packet(txq, dst, m, m_tls, available, tcp_seqno, pidx)); } if (tlen < plen) { plen = tlen; offset = ktls_payload_offset(tlsp, m_tls); #ifdef VERBOSE_TRACES CTR4(KTR_CXGBE, "%s: tid %d short TLS record %u with offset %u", __func__, tlsp->tid, (u_int)m_tls->m_epg_seqno, offset); #endif if (m_tls->m_next == NULL && (tcp->th_flags & TH_FIN) != 0) { txq->kern_tls_fin_short++; #ifdef INVARIANTS panic("%s: FIN on short TLS record", __func__); #endif } } else offset = 0; /* * This is the last work request for a given TLS mbuf chain if * it is the last mbuf in the chain and FIN is not set. If * FIN is set, then ktls_write_tcp_fin() will write out the * last work request. */ last_wr = m_tls->m_next == NULL && (tcp->th_flags & TH_FIN) == 0; /* * The host stack may ask us to not send part of the start of * a TLS record. (For example, the stack might have * previously sent a "short" TLS record and might later send * down an mbuf that requests to send the remainder of the TLS * record.) The crypto engine must process a TLS record from * the beginning if computing a GCM tag or HMAC, so we always * send the TLS record from the beginning as input to the * crypto engine and via CPL_TX_DATA to TP. However, TP will * drop individual packets after they have been chopped up * into MSS-sized chunks if the entire sequence range of those * packets is less than SND_UNA. SND_UNA is computed as * TX_MAX - SND_UNA_RAW. Thus, use the offset stored in * m_data to set TX_MAX to the first byte in the TCP sequence * space the host actually wants us to send and set * SND_UNA_RAW to 0. * * If the host sends us back to back requests that span the * trailer of a single TLS record (first request ends "in" the * trailer and second request starts at the next byte but * still "in" the trailer), the initial bytes of the trailer * that the first request drops will not be retransmitted. If * the host uses the same requests when retransmitting the * connection will hang. To handle this, always transmit the * full trailer for a request that begins "in" the trailer * (the second request in the example above). This should * also help to avoid retransmits for the common case. * * A similar condition exists when using CBC for back to back * requests that span a single AES block. The first request * will be truncated to end at the end of the previous AES * block. To handle this, always begin transmission at the * start of the current AES block. */ tx_max_offset = mtod(m_tls, vm_offset_t); if (tx_max_offset > TLS_HEADER_LENGTH + ntohs(hdr->tls_length) - m_tls->m_epg_trllen) { /* Always send the full trailer. */ tx_max_offset = TLS_HEADER_LENGTH + ntohs(hdr->tls_length) - m_tls->m_epg_trllen; } if (tlsp->enc_mode == SCMD_CIPH_MODE_AES_CBC && tx_max_offset > TLS_HEADER_LENGTH) { /* Always send all of the first AES block. */ tx_max_offset = TLS_HEADER_LENGTH + rounddown(tx_max_offset - TLS_HEADER_LENGTH, AES_BLOCK_LEN); } tx_max = tcp_seqno + tx_max_offset; /* * Update TCB fields. Reserve space for the FW_ULPTX_WR header * but don't populate it until we know how many field updates * are required. */ if (using_scratch) wr = (void *)txq->ss; else wr = dst; out = (void *)(wr + 1); fields = 0; if (set_l2t_idx) { KASSERT(nsegs != 0, ("trying to set L2T_IX for subsequent TLS WR")); #ifdef VERBOSE_TRACES CTR3(KTR_CXGBE, "%s: tid %d set L2T_IX to %d", __func__, tlsp->tid, tlsp->l2te->idx); #endif write_set_tcb_field_ulp(tlsp, out, txq, W_TCB_L2T_IX, V_TCB_L2T_IX(M_TCB_L2T_IX), V_TCB_L2T_IX(tlsp->l2te->idx)); out += roundup2(LEN__SET_TCB_FIELD_ULP, 16); fields++; } if (tsopt != NULL && tlsp->prev_tsecr != ntohl(tsopt[1])) { KASSERT(nsegs != 0, ("trying to set T_RTSEQ_RECENT for subsequent TLS WR")); #ifdef VERBOSE_TRACES CTR2(KTR_CXGBE, "%s: tid %d wrote updated T_RTSEQ_RECENT", __func__, tlsp->tid); #endif write_set_tcb_field_ulp(tlsp, out, txq, W_TCB_T_RTSEQ_RECENT, V_TCB_T_RTSEQ_RECENT(M_TCB_T_RTSEQ_RECENT), V_TCB_T_RTSEQ_RECENT(ntohl(tsopt[1]))); out += roundup2(LEN__SET_TCB_FIELD_ULP, 16); fields++; tlsp->prev_tsecr = ntohl(tsopt[1]); } if (first_wr || tlsp->prev_seq != tx_max) { KASSERT(nsegs != 0, ("trying to set TX_MAX for subsequent TLS WR")); #ifdef VERBOSE_TRACES CTR4(KTR_CXGBE, "%s: tid %d setting TX_MAX to %u (tcp_seqno %u)", __func__, tlsp->tid, tx_max, tcp_seqno); #endif write_set_tcb_field_ulp(tlsp, out, txq, W_TCB_TX_MAX, V_TCB_TX_MAX(M_TCB_TX_MAX), V_TCB_TX_MAX(tx_max)); out += roundup2(LEN__SET_TCB_FIELD_ULP, 16); fields++; } /* * If there is data to drop at the beginning of this TLS * record or if this is a retransmit, * reset SND_UNA_RAW to 0 so that SND_UNA == TX_MAX. */ if (tlsp->prev_seq != tx_max || mtod(m_tls, vm_offset_t) != 0) { KASSERT(nsegs != 0, ("trying to clear SND_UNA_RAW for subsequent TLS WR")); #ifdef VERBOSE_TRACES CTR2(KTR_CXGBE, "%s: tid %d clearing SND_UNA_RAW", __func__, tlsp->tid); #endif write_set_tcb_field_ulp(tlsp, out, txq, W_TCB_SND_UNA_RAW, V_TCB_SND_UNA_RAW(M_TCB_SND_UNA_RAW), V_TCB_SND_UNA_RAW(0)); out += roundup2(LEN__SET_TCB_FIELD_ULP, 16); fields++; } /* * Store the expected sequence number of the next byte after * this record. */ tlsp->prev_seq = tcp_seqno + tlen; if (first_wr || tlsp->prev_ack != ntohl(tcp->th_ack)) { KASSERT(nsegs != 0, ("trying to set RCV_NXT for subsequent TLS WR")); write_set_tcb_field_ulp(tlsp, out, txq, W_TCB_RCV_NXT, V_TCB_RCV_NXT(M_TCB_RCV_NXT), V_TCB_RCV_NXT(ntohl(tcp->th_ack))); out += roundup2(LEN__SET_TCB_FIELD_ULP, 16); fields++; tlsp->prev_ack = ntohl(tcp->th_ack); } if (first_wr || tlsp->prev_win != ntohs(tcp->th_win)) { KASSERT(nsegs != 0, ("trying to set RCV_WND for subsequent TLS WR")); write_set_tcb_field_ulp(tlsp, out, txq, W_TCB_RCV_WND, V_TCB_RCV_WND(M_TCB_RCV_WND), V_TCB_RCV_WND(ntohs(tcp->th_win))); out += roundup2(LEN__SET_TCB_FIELD_ULP, 16); fields++; tlsp->prev_win = ntohs(tcp->th_win); } /* Recalculate 'nsegs' if cached value is not available. */ if (nsegs == 0) nsegs = sglist_count_mbuf_epg(m_tls, m_tls->m_epg_hdrlen + offset, plen - (m_tls->m_epg_hdrlen + offset)); /* Calculate the size of the TLS work request. */ twr_len = ktls_base_wr_size(tlsp); imm_len = 0; if (offset == 0) imm_len += m_tls->m_epg_hdrlen; if (plen == tlen) imm_len += AES_BLOCK_LEN; twr_len += roundup2(imm_len, 16); twr_len += ktls_sgl_size(nsegs); /* * If any field updates were required, determine if they can * be included in the TLS work request. If not, use the * FW_ULPTX_WR work request header at 'wr' as a dedicated work * request for the field updates and start a new work request * for the TLS work request afterward. */ if (fields != 0) { wr_len = fields * roundup2(LEN__SET_TCB_FIELD_ULP, 16); if (twr_len + wr_len <= SGE_MAX_WR_LEN && tlsp->sc->tlst.combo_wrs) { wr_len += twr_len; txpkt = (void *)out; } else { wr_len += sizeof(*wr); wr->op_to_compl = htobe32(V_FW_WR_OP(FW_ULPTX_WR)); wr->flowid_len16 = htobe32(F_FW_ULPTX_WR_DATA | V_FW_WR_LEN16(wr_len / 16)); wr->cookie = 0; /* * If we were using scratch space, copy the * field updates work request to the ring. */ if (using_scratch) { out = dst; copy_to_txd(eq, txq->ss, &out, wr_len); } ndesc = howmany(wr_len, EQ_ESIZE); MPASS(ndesc <= available); txq->raw_wrs++; txsd = &txq->sdesc[pidx]; txsd->m = NULL; txsd->desc_used = ndesc; IDXINCR(pidx, ndesc, eq->sidx); dst = &eq->desc[pidx]; /* * Determine if we should use scratch space * for the TLS work request based on the * available space after advancing pidx for * the field updates work request. */ wr_len = twr_len; using_scratch = (eq->sidx - pidx < howmany(wr_len, EQ_ESIZE)); if (using_scratch) wr = (void *)txq->ss; else wr = dst; txpkt = (void *)(wr + 1); } } else { wr_len = twr_len; txpkt = (void *)out; } wr_len = roundup2(wr_len, 16); MPASS(ndesc + howmany(wr_len, EQ_ESIZE) <= available); /* FW_ULPTX_WR */ wr->op_to_compl = htobe32(V_FW_WR_OP(FW_ULPTX_WR)); wr->flowid_len16 = htobe32(F_FW_ULPTX_WR_DATA | V_FW_WR_LEN16(wr_len / 16)); wr->cookie = 0; /* ULP_TXPKT */ txpkt->cmd_dest = htobe32(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DATAMODIFY(0) | V_ULP_TXPKT_CHANNELID(tlsp->vi->pi->port_id) | V_ULP_TXPKT_DEST(0) | V_ULP_TXPKT_FID(txq->eq.cntxt_id) | V_ULP_TXPKT_RO(1)); txpkt->len = htobe32(howmany(twr_len - sizeof(*wr), 16)); /* ULPTX_IDATA sub-command */ idata = (void *)(txpkt + 1); idata->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM) | V_ULP_TX_SC_MORE(1)); idata->len = sizeof(struct cpl_tx_sec_pdu); /* * The key context, CPL_TX_DATA, and immediate data are part * of this ULPTX_IDATA when using an inline key. When reading * the key from memory, the CPL_TX_DATA and immediate data are * part of a separate ULPTX_IDATA. */ if (tlsp->inline_key) idata->len += tlsp->tx_key_info_size + sizeof(struct cpl_tx_data) + imm_len; idata->len = htobe32(idata->len); /* CPL_TX_SEC_PDU */ sec_pdu = (void *)(idata + 1); /* * For short records, AAD is counted as header data in SCMD0, * the IV is next followed by a cipher region for the payload. */ if (plen == tlen) { aad_start = 0; aad_stop = 0; iv_offset = 1; auth_start = 0; auth_stop = 0; auth_insert = 0; cipher_start = AES_BLOCK_LEN + 1; cipher_stop = 0; sec_pdu->pldlen = htobe32(16 + plen - (m_tls->m_epg_hdrlen + offset)); /* These two flits are actually a CPL_TLS_TX_SCMD_FMT. */ sec_pdu->seqno_numivs = tlsp->scmd0_short.seqno_numivs; sec_pdu->ivgen_hdrlen = htobe32( tlsp->scmd0_short.ivgen_hdrlen | V_SCMD_HDR_LEN(offset == 0 ? m_tls->m_epg_hdrlen : 0)); txq->kern_tls_short++; } else { /* * AAD is TLS header. IV is after AAD. The cipher region * starts after the IV. See comments in ccr_authenc() and * ccr_gmac() in t4_crypto.c regarding cipher and auth * start/stop values. */ aad_start = 1; aad_stop = TLS_HEADER_LENGTH; iv_offset = TLS_HEADER_LENGTH + 1; cipher_start = m_tls->m_epg_hdrlen + 1; if (tlsp->enc_mode == SCMD_CIPH_MODE_AES_GCM) { cipher_stop = 0; auth_start = cipher_start; auth_stop = 0; auth_insert = 0; } else { cipher_stop = 0; auth_start = cipher_start; auth_stop = 0; auth_insert = 0; } sec_pdu->pldlen = htobe32(plen); /* These two flits are actually a CPL_TLS_TX_SCMD_FMT. */ sec_pdu->seqno_numivs = tlsp->scmd0.seqno_numivs; sec_pdu->ivgen_hdrlen = tlsp->scmd0.ivgen_hdrlen; if (mtod(m_tls, vm_offset_t) == 0) txq->kern_tls_full++; else txq->kern_tls_partial++; } sec_pdu->op_ivinsrtofst = htobe32( V_CPL_TX_SEC_PDU_OPCODE(CPL_TX_SEC_PDU) | V_CPL_TX_SEC_PDU_CPLLEN(2) | V_CPL_TX_SEC_PDU_PLACEHOLDER(0) | V_CPL_TX_SEC_PDU_IVINSRTOFST(iv_offset)); sec_pdu->aadstart_cipherstop_hi = htobe32( V_CPL_TX_SEC_PDU_AADSTART(aad_start) | V_CPL_TX_SEC_PDU_AADSTOP(aad_stop) | V_CPL_TX_SEC_PDU_CIPHERSTART(cipher_start) | V_CPL_TX_SEC_PDU_CIPHERSTOP_HI(cipher_stop >> 4)); sec_pdu->cipherstop_lo_authinsert = htobe32( V_CPL_TX_SEC_PDU_CIPHERSTOP_LO(cipher_stop & 0xf) | V_CPL_TX_SEC_PDU_AUTHSTART(auth_start) | V_CPL_TX_SEC_PDU_AUTHSTOP(auth_stop) | V_CPL_TX_SEC_PDU_AUTHINSERT(auth_insert)); sec_pdu->scmd1 = htobe64(m_tls->m_epg_seqno); /* Key context */ out = (void *)(sec_pdu + 1); if (tlsp->inline_key) { memcpy(out, &tlsp->keyctx, tlsp->tx_key_info_size); out += tlsp->tx_key_info_size; } else { /* ULPTX_SC_MEMRD to read key context. */ memrd = (void *)out; memrd->cmd_to_len = htobe32(V_ULPTX_CMD(ULP_TX_SC_MEMRD) | V_ULP_TX_SC_MORE(1) | V_ULPTX_LEN16(tlsp->tx_key_info_size >> 4)); memrd->addr = htobe32(tlsp->tx_key_addr >> 5); /* ULPTX_IDATA for CPL_TX_DATA and TLS header. */ idata = (void *)(memrd + 1); idata->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM) | V_ULP_TX_SC_MORE(1)); idata->len = htobe32(sizeof(struct cpl_tx_data) + imm_len); out = (void *)(idata + 1); } /* CPL_TX_DATA */ tx_data = (void *)out; OPCODE_TID(tx_data) = htonl(MK_OPCODE_TID(CPL_TX_DATA, tlsp->tid)); if (m->m_pkthdr.csum_flags & CSUM_TSO) { mss = m->m_pkthdr.tso_segsz; tlsp->prev_mss = mss; } else if (tlsp->prev_mss != 0) mss = tlsp->prev_mss; else mss = tlsp->vi->ifp->if_mtu - (m->m_pkthdr.l3hlen + m->m_pkthdr.l4hlen); if (offset == 0) { tx_data->len = htobe32(V_TX_DATA_MSS(mss) | V_TX_LENGTH(tlen)); tx_data->rsvd = htobe32(tcp_seqno); } else { tx_data->len = htobe32(V_TX_DATA_MSS(mss) | V_TX_LENGTH(tlen - (m_tls->m_epg_hdrlen + offset))); tx_data->rsvd = htobe32(tcp_seqno + m_tls->m_epg_hdrlen + offset); } tx_data->flags = htobe32(F_TX_BYPASS); if (last_wr && tcp->th_flags & TH_PUSH) tx_data->flags |= htobe32(F_TX_PUSH | F_TX_SHOVE); /* Populate the TLS header */ out = (void *)(tx_data + 1); if (offset == 0) { memcpy(out, m_tls->m_epg_hdr, m_tls->m_epg_hdrlen); out += m_tls->m_epg_hdrlen; } /* AES IV for a short record. */ if (plen == tlen) { iv = out; if (tlsp->enc_mode == SCMD_CIPH_MODE_AES_GCM) { memcpy(iv, tlsp->keyctx.u.txhdr.txsalt, SALT_SIZE); memcpy(iv + 4, hdr + 1, 8); *(uint32_t *)(iv + 12) = htobe32(2 + offset / AES_BLOCK_LEN); } else memcpy(iv, hdr + 1, AES_BLOCK_LEN); out += AES_BLOCK_LEN; } if (imm_len % 16 != 0) { /* Zero pad to an 8-byte boundary. */ memset(out, 0, 8 - (imm_len % 8)); out += 8 - (imm_len % 8); /* * Insert a ULP_TX_SC_NOOP if needed so the SGL is * 16-byte aligned. */ if (imm_len % 16 <= 8) { idata = (void *)out; idata->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_NOOP)); idata->len = htobe32(0); out = (void *)(idata + 1); } } /* SGL for record payload */ sglist_reset(txq->gl); if (sglist_append_mbuf_epg(txq->gl, m_tls, m_tls->m_epg_hdrlen + offset, plen - (m_tls->m_epg_hdrlen + offset)) != 0) { #ifdef INVARIANTS panic("%s: failed to append sglist", __func__); #endif } write_gl_to_buf(txq->gl, out); if (using_scratch) { out = dst; copy_to_txd(eq, txq->ss, &out, wr_len); } ndesc += howmany(wr_len, EQ_ESIZE); MPASS(ndesc <= available); txq->kern_tls_records++; txq->kern_tls_octets += tlen - mtod(m_tls, vm_offset_t); if (mtod(m_tls, vm_offset_t) != 0) { if (offset == 0) txq->kern_tls_waste += mtod(m_tls, vm_offset_t); else txq->kern_tls_waste += mtod(m_tls, vm_offset_t) - (m_tls->m_epg_hdrlen + offset); } txsd = &txq->sdesc[pidx]; if (last_wr) txsd->m = m; else txsd->m = NULL; txsd->desc_used = howmany(wr_len, EQ_ESIZE); return (ndesc); } static int ktls_write_tcp_fin(struct sge_txq *txq, void *dst, struct mbuf *m, u_int available, tcp_seq tcp_seqno, u_int pidx) { struct tx_sdesc *txsd; struct fw_eth_tx_pkt_wr *wr; struct cpl_tx_pkt_core *cpl; uint32_t ctrl; uint64_t ctrl1; int len16, ndesc, pktlen; struct ether_header *eh; struct ip *ip, newip; struct ip6_hdr *ip6, newip6; struct tcphdr *tcp, newtcp; caddr_t out; TXQ_LOCK_ASSERT_OWNED(txq); M_ASSERTPKTHDR(m); wr = dst; pktlen = m->m_len; ctrl = sizeof(struct cpl_tx_pkt_core) + pktlen; len16 = howmany(sizeof(struct fw_eth_tx_pkt_wr) + ctrl, 16); ndesc = tx_len16_to_desc(len16); MPASS(ndesc <= available); /* Firmware work request header */ wr->op_immdlen = htobe32(V_FW_WR_OP(FW_ETH_TX_PKT_WR) | V_FW_ETH_TX_PKT_WR_IMMDLEN(ctrl)); ctrl = V_FW_WR_LEN16(len16); wr->equiq_to_len16 = htobe32(ctrl); wr->r3 = 0; cpl = (void *)(wr + 1); /* CPL header */ cpl->ctrl0 = txq->cpl_ctrl0; cpl->pack = 0; cpl->len = htobe16(pktlen); out = (void *)(cpl + 1); /* Copy over Ethernet header. */ eh = mtod(m, struct ether_header *); copy_to_txd(&txq->eq, (caddr_t)eh, &out, m->m_pkthdr.l2hlen); /* Fixup length in IP header and copy out. */ if (ntohs(eh->ether_type) == ETHERTYPE_IP) { ip = (void *)((char *)eh + m->m_pkthdr.l2hlen); newip = *ip; newip.ip_len = htons(pktlen - m->m_pkthdr.l2hlen); copy_to_txd(&txq->eq, (caddr_t)&newip, &out, sizeof(newip)); if (m->m_pkthdr.l3hlen > sizeof(*ip)) copy_to_txd(&txq->eq, (caddr_t)(ip + 1), &out, m->m_pkthdr.l3hlen - sizeof(*ip)); ctrl1 = V_TXPKT_CSUM_TYPE(TX_CSUM_TCPIP) | V_T6_TXPKT_ETHHDR_LEN(m->m_pkthdr.l2hlen - ETHER_HDR_LEN) | V_TXPKT_IPHDR_LEN(m->m_pkthdr.l3hlen); } else { ip6 = (void *)((char *)eh + m->m_pkthdr.l2hlen); newip6 = *ip6; newip6.ip6_plen = htons(pktlen - m->m_pkthdr.l2hlen); copy_to_txd(&txq->eq, (caddr_t)&newip6, &out, sizeof(newip6)); MPASS(m->m_pkthdr.l3hlen == sizeof(*ip6)); ctrl1 = V_TXPKT_CSUM_TYPE(TX_CSUM_TCPIP6) | V_T6_TXPKT_ETHHDR_LEN(m->m_pkthdr.l2hlen - ETHER_HDR_LEN) | V_TXPKT_IPHDR_LEN(m->m_pkthdr.l3hlen); } cpl->ctrl1 = htobe64(ctrl1); txq->txcsum++; /* Set sequence number in TCP header. */ tcp = (void *)((char *)eh + m->m_pkthdr.l2hlen + m->m_pkthdr.l3hlen); newtcp = *tcp; newtcp.th_seq = htonl(tcp_seqno); copy_to_txd(&txq->eq, (caddr_t)&newtcp, &out, sizeof(newtcp)); /* Copy rest of packet. */ copy_to_txd(&txq->eq, (caddr_t)(tcp + 1), &out, m->m_len - (m->m_pkthdr.l2hlen + m->m_pkthdr.l3hlen + sizeof(*tcp))); txq->imm_wrs++; txq->txpkt_wrs++; txq->kern_tls_fin++; txsd = &txq->sdesc[pidx]; txsd->m = m; txsd->desc_used = ndesc; return (ndesc); } int t6_ktls_write_wr(struct sge_txq *txq, void *dst, struct mbuf *m, u_int nsegs, u_int available) { struct sge_eq *eq = &txq->eq; struct tx_sdesc *txsd; struct tlspcb *tlsp; struct tcphdr *tcp; struct mbuf *m_tls; struct ether_header *eh; tcp_seq tcp_seqno; u_int ndesc, pidx, totdesc; uint16_t vlan_tag; bool has_fin, set_l2t_idx; void *tsopt; M_ASSERTPKTHDR(m); MPASS(m->m_pkthdr.snd_tag != NULL); tlsp = mst_to_tls(m->m_pkthdr.snd_tag); totdesc = 0; eh = mtod(m, struct ether_header *); tcp = (struct tcphdr *)((char *)eh + m->m_pkthdr.l2hlen + m->m_pkthdr.l3hlen); pidx = eq->pidx; has_fin = (tcp->th_flags & TH_FIN) != 0; /* * If this TLS record has a FIN, then we will send any * requested options as part of the FIN packet. */ if (!has_fin && ktls_has_tcp_options(tcp)) { ndesc = ktls_write_tcp_options(txq, dst, m, available, pidx); totdesc += ndesc; IDXINCR(pidx, ndesc, eq->sidx); dst = &eq->desc[pidx]; #ifdef VERBOSE_TRACES CTR2(KTR_CXGBE, "%s: tid %d wrote TCP options packet", __func__, tlsp->tid); #endif } /* * Allocate a new L2T entry if necessary. This may write out * a work request to the txq. */ if (m->m_flags & M_VLANTAG) vlan_tag = m->m_pkthdr.ether_vtag; else vlan_tag = 0xfff; set_l2t_idx = false; if (tlsp->l2te == NULL || tlsp->l2te->vlan != vlan_tag || memcmp(tlsp->l2te->dmac, eh->ether_dhost, ETHER_ADDR_LEN) != 0) { set_l2t_idx = true; if (tlsp->l2te) t4_l2t_release(tlsp->l2te); tlsp->l2te = t4_l2t_alloc_tls(tlsp->sc, txq, dst, &ndesc, vlan_tag, tlsp->vi->pi->lport, eh->ether_dhost); if (tlsp->l2te == NULL) CXGBE_UNIMPLEMENTED("failed to allocate TLS L2TE"); if (ndesc != 0) { MPASS(ndesc <= available - totdesc); txq->raw_wrs++; txsd = &txq->sdesc[pidx]; txsd->m = NULL; txsd->desc_used = ndesc; totdesc += ndesc; IDXINCR(pidx, ndesc, eq->sidx); dst = &eq->desc[pidx]; } } /* * Iterate over each TLS record constructing a work request * for that record. */ for (m_tls = m->m_next; m_tls != NULL; m_tls = m_tls->m_next) { MPASS(m_tls->m_flags & M_EXTPG); /* * Determine the initial TCP sequence number for this * record. */ tsopt = NULL; if (m_tls == m->m_next) { tcp_seqno = ntohl(tcp->th_seq) - mtod(m_tls, vm_offset_t); if (tlsp->using_timestamps) tsopt = ktls_find_tcp_timestamps(tcp); } else { MPASS(mtod(m_tls, vm_offset_t) == 0); tcp_seqno = tlsp->prev_seq; } ndesc = ktls_write_tls_wr(tlsp, txq, dst, m, tcp, m_tls, nsegs, available - totdesc, tcp_seqno, tsopt, pidx, set_l2t_idx); totdesc += ndesc; IDXINCR(pidx, ndesc, eq->sidx); dst = &eq->desc[pidx]; /* * The value of nsegs from the header mbuf's metadata * is only valid for the first TLS record. */ nsegs = 0; /* Only need to set the L2T index once. */ set_l2t_idx = false; } if (has_fin) { /* * If the TCP header for this chain has FIN sent, then * explicitly send a packet that has FIN set. This * will also have PUSH set if requested. This assumes * we sent at least one TLS record work request and * uses the TCP sequence number after that reqeust as * the sequence number for the FIN packet. */ ndesc = ktls_write_tcp_fin(txq, dst, m, available, tlsp->prev_seq, pidx); totdesc += ndesc; } MPASS(totdesc <= available); return (totdesc); } static void t6_tls_tag_free(struct m_snd_tag *mst) { struct adapter *sc; struct tlspcb *tlsp; tlsp = mst_to_tls(mst); sc = tlsp->sc; CTR2(KTR_CXGBE, "%s: tid %d", __func__, tlsp->tid); if (tlsp->l2te) t4_l2t_release(tlsp->l2te); if (tlsp->tid >= 0) release_tid(sc, tlsp->tid, tlsp->ctrlq); if (tlsp->ce) t4_release_clip_entry(sc, tlsp->ce); if (tlsp->tx_key_addr >= 0) t4_free_tls_keyid(sc, tlsp->tx_key_addr); zfree(tlsp, M_CXGBE); } void t6_ktls_modload(void) { t4_register_shared_cpl_handler(CPL_ACT_OPEN_RPL, ktls_act_open_rpl, CPL_COOKIE_KERN_TLS); } void t6_ktls_modunload(void) { t4_register_shared_cpl_handler(CPL_ACT_OPEN_RPL, NULL, CPL_COOKIE_KERN_TLS); } #else int t6_tls_tag_alloc(struct ifnet *ifp, union if_snd_tag_alloc_params *params, struct m_snd_tag **pt) { return (ENXIO); } int t6_ktls_parse_pkt(struct mbuf *m, int *nsegsp, int *len16p) { return (EINVAL); } int t6_ktls_write_wr(struct sge_txq *txq, void *dst, struct mbuf *m, u_int nsegs, u_int available) { panic("can't happen"); } void t6_ktls_modload(void) { } void t6_ktls_modunload(void) { } #endif diff --git a/sys/dev/cxgbe/cxgbei/cxgbei.c b/sys/dev/cxgbe/cxgbei/cxgbei.c index b6dc5d5c3577..d11465e4567d 100644 --- a/sys/dev/cxgbe/cxgbei/cxgbei.c +++ b/sys/dev/cxgbe/cxgbei/cxgbei.c @@ -1,988 +1,988 @@ /*- * Copyright (c) 2012 Chelsio Communications, Inc. * All rights reserved. * * Chelsio T5xx iSCSI driver * * Written by: Sreenivasa Honnur * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include #include #include #include #include #include #ifdef TCP_OFFLOAD #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include "common/common.h" #include "common/t4_msg.h" #include "common/t4_regs.h" /* for PCIE_MEM_ACCESS */ #include "tom/t4_tom.h" #include "cxgbei.h" static void read_pdu_limits(struct adapter *sc, uint32_t *max_tx_data_len, uint32_t *max_rx_data_len, struct ppod_region *pr) { uint32_t tx_len, rx_len, r, v; rx_len = t4_read_reg(sc, A_TP_PMM_RX_PAGE_SIZE); tx_len = t4_read_reg(sc, A_TP_PMM_TX_PAGE_SIZE); r = t4_read_reg(sc, A_TP_PARA_REG2); rx_len = min(rx_len, G_MAXRXDATA(r)); tx_len = min(tx_len, G_MAXRXDATA(r)); r = t4_read_reg(sc, A_TP_PARA_REG7); v = min(G_PMMAXXFERLEN0(r), G_PMMAXXFERLEN1(r)); rx_len = min(rx_len, v); tx_len = min(tx_len, v); /* * AHS is not supported by the kernel so we'll not account for * it either in our PDU len -> data segment len conversions. */ rx_len -= ISCSI_BHS_SIZE + ISCSI_HEADER_DIGEST_SIZE + ISCSI_DATA_DIGEST_SIZE; tx_len -= ISCSI_BHS_SIZE + ISCSI_HEADER_DIGEST_SIZE + ISCSI_DATA_DIGEST_SIZE; /* * DDP can place only 4 pages for a single PDU. A single * request might use larger pages than the smallest page size, * but that cannot be guaranteed. Assume the smallest DDP * page size for this limit. */ rx_len = min(rx_len, 4 * (1U << pr->pr_page_shift[0])); if (chip_id(sc) == CHELSIO_T5) { tx_len = min(tx_len, 15360); rx_len = rounddown2(rx_len, 512); tx_len = rounddown2(tx_len, 512); } *max_tx_data_len = tx_len; *max_rx_data_len = rx_len; } /* * Initialize the software state of the iSCSI ULP driver. * * ENXIO means firmware didn't set up something that it was supposed to. */ static int cxgbei_init(struct adapter *sc, struct cxgbei_data *ci) { struct sysctl_oid *oid; struct sysctl_oid_list *children; struct ppod_region *pr; uint32_t r; int rc; MPASS(sc->vres.iscsi.size > 0); MPASS(ci != NULL); pr = &ci->pr; r = t4_read_reg(sc, A_ULP_RX_ISCSI_PSZ); rc = t4_init_ppod_region(pr, &sc->vres.iscsi, r, "iSCSI page pods"); if (rc != 0) { device_printf(sc->dev, "%s: failed to initialize the iSCSI page pod region: %u.\n", __func__, rc); return (rc); } read_pdu_limits(sc, &ci->max_tx_data_len, &ci->max_rx_data_len, pr); sysctl_ctx_init(&ci->ctx); oid = device_get_sysctl_tree(sc->dev); /* dev.t5nex.X */ children = SYSCTL_CHILDREN(oid); oid = SYSCTL_ADD_NODE(&ci->ctx, children, OID_AUTO, "iscsi", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "iSCSI ULP settings"); children = SYSCTL_CHILDREN(oid); ci->ddp_threshold = 2048; SYSCTL_ADD_UINT(&ci->ctx, children, OID_AUTO, "ddp_threshold", CTLFLAG_RW, &ci->ddp_threshold, 0, "Rx zero copy threshold"); SYSCTL_ADD_UINT(&ci->ctx, children, OID_AUTO, "max_rx_data_len", CTLFLAG_RW, &ci->max_rx_data_len, 0, "Maximum receive data segment length"); SYSCTL_ADD_UINT(&ci->ctx, children, OID_AUTO, "max_tx_data_len", CTLFLAG_RW, &ci->max_tx_data_len, 0, "Maximum transmit data segment length"); return (0); } static int do_rx_iscsi_hdr(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { struct adapter *sc = iq->adapter; struct cpl_iscsi_hdr *cpl = mtod(m, struct cpl_iscsi_hdr *); u_int tid = GET_TID(cpl); struct toepcb *toep = lookup_tid(sc, tid); struct icl_pdu *ip; struct icl_cxgbei_pdu *icp; uint16_t len_ddp = be16toh(cpl->pdu_len_ddp); uint16_t len = be16toh(cpl->len); M_ASSERTPKTHDR(m); MPASS(m->m_pkthdr.len == len + sizeof(*cpl)); ip = icl_cxgbei_new_pdu(M_NOWAIT); if (ip == NULL) CXGBE_UNIMPLEMENTED("PDU allocation failure"); m_copydata(m, sizeof(*cpl), ISCSI_BHS_SIZE, (caddr_t)ip->ip_bhs); ip->ip_data_len = G_ISCSI_PDU_LEN(len_ddp) - len; icp = ip_to_icp(ip); icp->icp_seq = ntohl(cpl->seq); icp->icp_flags = ICPF_RX_HDR; /* This is the start of a new PDU. There should be no old state. */ MPASS(toep->ulpcb2 == NULL); toep->ulpcb2 = icp; #if 0 CTR5(KTR_CXGBE, "%s: tid %u, cpl->len %u, pdu_len_ddp 0x%04x, icp %p", __func__, tid, len, len_ddp, icp); #endif m_freem(m); return (0); } static int do_rx_iscsi_data(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { struct adapter *sc = iq->adapter; struct cpl_iscsi_data *cpl = mtod(m, struct cpl_iscsi_data *); u_int tid = GET_TID(cpl); struct toepcb *toep = lookup_tid(sc, tid); struct icl_cxgbei_pdu *icp = toep->ulpcb2; struct icl_pdu *ip; M_ASSERTPKTHDR(m); MPASS(m->m_pkthdr.len == be16toh(cpl->len) + sizeof(*cpl)); if (icp == NULL) { /* * T6 completion enabled, start of a new pdu. Header * will come in completion CPL. */ ip = icl_cxgbei_new_pdu(M_NOWAIT); if (ip == NULL) CXGBE_UNIMPLEMENTED("PDU allocation failure"); icp = ip_to_icp(ip); } else { /* T5 mode, header is already received. */ MPASS(icp->icp_flags == ICPF_RX_HDR); MPASS(icp->ip.ip_data_mbuf == NULL); MPASS(icp->ip.ip_data_len == m->m_pkthdr.len - sizeof(*cpl)); } /* Trim the cpl header from mbuf. */ m_adj(m, sizeof(*cpl)); icp->icp_flags |= ICPF_RX_FLBUF; icp->ip.ip_data_mbuf = m; toep->ofld_rxq->rx_iscsi_fl_pdus++; toep->ofld_rxq->rx_iscsi_fl_octets += m->m_pkthdr.len; /* * For T6, save the icp for further processing in the * completion handler. */ if (icp->icp_flags == ICPF_RX_FLBUF) { MPASS(toep->ulpcb2 == NULL); toep->ulpcb2 = icp; } #if 0 CTR4(KTR_CXGBE, "%s: tid %u, cpl->len %u, icp %p", __func__, tid, be16toh(cpl->len), icp); #endif return (0); } static int mbuf_crc32c_helper(void *arg, void *data, u_int len) { uint32_t *digestp = arg; *digestp = calculate_crc32c(*digestp, data, len); return (0); } static struct icl_pdu * parse_pdu(struct socket *so, struct toepcb *toep, struct icl_cxgbei_conn *icc, struct sockbuf *sb, u_int total_len) { struct uio uio; struct iovec iov[2]; struct iscsi_bhs bhs; struct mbuf *m; struct icl_pdu *ip; u_int ahs_len, data_len, header_len, pdu_len; uint32_t calc_digest, wire_digest; int error; uio.uio_segflg = UIO_SYSSPACE; uio.uio_rw = UIO_READ; uio.uio_td = curthread; header_len = sizeof(struct iscsi_bhs); if (icc->ic.ic_header_crc32c) header_len += ISCSI_HEADER_DIGEST_SIZE; if (total_len < header_len) { ICL_WARN("truncated pre-offload PDU with len %u", total_len); return (NULL); } iov[0].iov_base = &bhs; iov[0].iov_len = sizeof(bhs); iov[1].iov_base = &wire_digest; iov[1].iov_len = sizeof(wire_digest); uio.uio_iov = iov; uio.uio_iovcnt = 1; uio.uio_offset = 0; uio.uio_resid = header_len; error = soreceive(so, NULL, &uio, NULL, NULL, NULL); if (error != 0) { ICL_WARN("failed to read BHS from pre-offload PDU: %d", error); return (NULL); } ahs_len = bhs.bhs_total_ahs_len * 4; data_len = bhs.bhs_data_segment_len[0] << 16 | bhs.bhs_data_segment_len[1] << 8 | bhs.bhs_data_segment_len[2]; pdu_len = header_len + ahs_len + roundup2(data_len, 4); if (icc->ic.ic_data_crc32c && data_len != 0) pdu_len += ISCSI_DATA_DIGEST_SIZE; if (total_len < pdu_len) { ICL_WARN("truncated pre-offload PDU len %u vs %u", total_len, pdu_len); return (NULL); } if (ahs_len != 0) { ICL_WARN("received pre-offload PDU with AHS"); return (NULL); } if (icc->ic.ic_header_crc32c) { calc_digest = calculate_crc32c(0xffffffff, (caddr_t)&bhs, sizeof(bhs)); calc_digest ^= 0xffffffff; if (calc_digest != wire_digest) { ICL_WARN("received pre-offload PDU 0x%02x with " "invalid header digest (0x%x vs 0x%x)", bhs.bhs_opcode, wire_digest, calc_digest); toep->ofld_rxq->rx_iscsi_header_digest_errors++; return (NULL); } } m = NULL; if (data_len != 0) { uio.uio_iov = NULL; uio.uio_resid = roundup2(data_len, 4); if (icc->ic.ic_data_crc32c) uio.uio_resid += ISCSI_DATA_DIGEST_SIZE; error = soreceive(so, NULL, &uio, &m, NULL, NULL); if (error != 0) { ICL_WARN("failed to read data payload from " "pre-offload PDU: %d", error); return (NULL); } if (icc->ic.ic_data_crc32c) { m_copydata(m, roundup2(data_len, 4), sizeof(wire_digest), (caddr_t)&wire_digest); calc_digest = 0xffffffff; m_apply(m, 0, roundup2(data_len, 4), mbuf_crc32c_helper, &calc_digest); calc_digest ^= 0xffffffff; if (calc_digest != wire_digest) { ICL_WARN("received pre-offload PDU 0x%02x " "with invalid data digest (0x%x vs 0x%x)", bhs.bhs_opcode, wire_digest, calc_digest); toep->ofld_rxq->rx_iscsi_data_digest_errors++; m_freem(m); return (NULL); } } } ip = icl_cxgbei_new_pdu(M_WAITOK); icl_cxgbei_new_pdu_set_conn(ip, &icc->ic); *ip->ip_bhs = bhs; ip->ip_data_len = data_len; ip->ip_data_mbuf = m; return (ip); } void parse_pdus(struct icl_cxgbei_conn *icc, struct sockbuf *sb) { struct icl_conn *ic = &icc->ic; struct socket *so = ic->ic_socket; struct toepcb *toep = icc->toep; struct icl_pdu *ip, *lastip; u_int total_len; SOCKBUF_LOCK_ASSERT(sb); CTR3(KTR_CXGBE, "%s: tid %u, %u bytes in so_rcv", __func__, toep->tid, sbused(sb)); lastip = NULL; while (sbused(sb) != 0 && (sb->sb_state & SBS_CANTRCVMORE) == 0) { total_len = sbused(sb); SOCKBUF_UNLOCK(sb); ip = parse_pdu(so, toep, icc, sb, total_len); if (ip == NULL) { ic->ic_error(ic); SOCKBUF_LOCK(sb); return; } if (lastip == NULL) STAILQ_INSERT_HEAD(&icc->rcvd_pdus, ip, ip_next); else STAILQ_INSERT_AFTER(&icc->rcvd_pdus, lastip, ip, ip_next); lastip = ip; SOCKBUF_LOCK(sb); } } static int do_rx_iscsi_ddp(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { struct adapter *sc = iq->adapter; const struct cpl_rx_data_ddp *cpl = (const void *)(rss + 1); u_int tid = GET_TID(cpl); struct toepcb *toep = lookup_tid(sc, tid); struct inpcb *inp = toep->inp; struct socket *so; struct sockbuf *sb; struct tcpcb *tp; struct icl_cxgbei_conn *icc; struct icl_conn *ic; struct icl_cxgbei_pdu *icp = toep->ulpcb2; struct icl_pdu *ip; u_int pdu_len, val; struct epoch_tracker et; MPASS(m == NULL); /* Must already be assembling a PDU. */ MPASS(icp != NULL); MPASS(icp->icp_flags & ICPF_RX_HDR); /* Data is optional. */ MPASS((icp->icp_flags & ICPF_RX_STATUS) == 0); pdu_len = be16toh(cpl->len); /* includes everything. */ val = be32toh(cpl->ddpvld); #if 0 CTR5(KTR_CXGBE, "%s: tid %u, cpl->len %u, ddpvld 0x%08x, icp_flags 0x%08x", __func__, tid, pdu_len, val, icp->icp_flags); #endif icp->icp_flags |= ICPF_RX_STATUS; ip = &icp->ip; if (val & F_DDP_PADDING_ERR) { ICL_WARN("received PDU 0x%02x with invalid padding", ip->ip_bhs->bhs_opcode); toep->ofld_rxq->rx_iscsi_padding_errors++; } if (val & F_DDP_HDRCRC_ERR) { ICL_WARN("received PDU 0x%02x with invalid header digest", ip->ip_bhs->bhs_opcode); toep->ofld_rxq->rx_iscsi_header_digest_errors++; } if (val & F_DDP_DATACRC_ERR) { ICL_WARN("received PDU 0x%02x with invalid data digest", ip->ip_bhs->bhs_opcode); toep->ofld_rxq->rx_iscsi_data_digest_errors++; } if (val & F_DDP_PDU && ip->ip_data_mbuf == NULL) { MPASS((icp->icp_flags & ICPF_RX_FLBUF) == 0); MPASS(ip->ip_data_len > 0); icp->icp_flags |= ICPF_RX_DDP; toep->ofld_rxq->rx_iscsi_ddp_pdus++; toep->ofld_rxq->rx_iscsi_ddp_octets += ip->ip_data_len; } INP_WLOCK(inp); - if (__predict_false(inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT))) { + if (__predict_false(inp->inp_flags & INP_DROPPED)) { CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x", __func__, tid, pdu_len, inp->inp_flags); INP_WUNLOCK(inp); icl_cxgbei_conn_pdu_free(NULL, ip); toep->ulpcb2 = NULL; return (0); } /* * T6+ does not report data PDUs received via DDP without F * set. This can result in gaps in the TCP sequence space. */ tp = intotcpcb(inp); MPASS(chip_id(sc) >= CHELSIO_T6 || icp->icp_seq == tp->rcv_nxt); tp->rcv_nxt = icp->icp_seq + pdu_len; tp->t_rcvtime = ticks; /* * Don't update the window size or return credits since RX * flow control is disabled. */ so = inp->inp_socket; sb = &so->so_rcv; SOCKBUF_LOCK(sb); icc = toep->ulpcb; if (__predict_false(icc == NULL || sb->sb_state & SBS_CANTRCVMORE)) { CTR5(KTR_CXGBE, "%s: tid %u, excess rx (%d bytes), icc %p, sb_state 0x%x", __func__, tid, pdu_len, icc, sb->sb_state); SOCKBUF_UNLOCK(sb); INP_WUNLOCK(inp); CURVNET_SET(so->so_vnet); NET_EPOCH_ENTER(et); INP_WLOCK(inp); tp = tcp_drop(tp, ECONNRESET); if (tp) INP_WUNLOCK(inp); NET_EPOCH_EXIT(et); CURVNET_RESTORE(); icl_cxgbei_conn_pdu_free(NULL, ip); toep->ulpcb2 = NULL; return (0); } MPASS(icc->icc_signature == CXGBEI_CONN_SIGNATURE); ic = &icc->ic; if ((val & (F_DDP_PADDING_ERR | F_DDP_HDRCRC_ERR | F_DDP_DATACRC_ERR)) != 0) { SOCKBUF_UNLOCK(sb); INP_WUNLOCK(inp); icl_cxgbei_conn_pdu_free(NULL, ip); toep->ulpcb2 = NULL; ic->ic_error(ic); return (0); } icl_cxgbei_new_pdu_set_conn(ip, ic); STAILQ_INSERT_TAIL(&icc->rcvd_pdus, ip, ip_next); if (!icc->rx_active) { icc->rx_active = true; wakeup(&icc->rx_active); } SOCKBUF_UNLOCK(sb); INP_WUNLOCK(inp); toep->ulpcb2 = NULL; return (0); } static int do_rx_iscsi_cmp(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { struct epoch_tracker et; struct adapter *sc = iq->adapter; struct cpl_rx_iscsi_cmp *cpl = mtod(m, struct cpl_rx_iscsi_cmp *); u_int tid = GET_TID(cpl); struct toepcb *toep = lookup_tid(sc, tid); struct icl_cxgbei_pdu *icp = toep->ulpcb2; struct icl_pdu *ip; struct cxgbei_cmp *cmp; struct inpcb *inp = toep->inp; #ifdef INVARIANTS uint16_t len = be16toh(cpl->len); u_int data_digest_len; #endif struct socket *so; struct sockbuf *sb; struct tcpcb *tp; struct icl_cxgbei_conn *icc; struct icl_conn *ic; struct iscsi_bhs_data_out *bhsdo; u_int val = be32toh(cpl->ddpvld); u_int npdus, pdu_len; uint32_t prev_seg_len; M_ASSERTPKTHDR(m); MPASS(m->m_pkthdr.len == len + sizeof(*cpl)); if ((val & F_DDP_PDU) == 0) { MPASS(icp != NULL); MPASS((icp->icp_flags & ICPF_RX_STATUS) == 0); ip = &icp->ip; } if (icp == NULL) { /* T6 completion enabled, start of a new PDU. */ ip = icl_cxgbei_new_pdu(M_NOWAIT); if (ip == NULL) CXGBE_UNIMPLEMENTED("PDU allocation failure"); icp = ip_to_icp(ip); } pdu_len = G_ISCSI_PDU_LEN(be16toh(cpl->pdu_len_ddp)); #if 0 CTR5(KTR_CXGBE, "%s: tid %u, cpl->len %u, ddpvld 0x%08x, icp %p", __func__, tid, pdu_len, val, icp); #endif /* Copy header */ m_copydata(m, sizeof(*cpl), ISCSI_BHS_SIZE, (caddr_t)ip->ip_bhs); bhsdo = (struct iscsi_bhs_data_out *)ip->ip_bhs; ip->ip_data_len = bhsdo->bhsdo_data_segment_len[0] << 16 | bhsdo->bhsdo_data_segment_len[1] << 8 | bhsdo->bhsdo_data_segment_len[2]; icp->icp_seq = ntohl(cpl->seq); icp->icp_flags |= ICPF_RX_HDR; icp->icp_flags |= ICPF_RX_STATUS; if (val & F_DDP_PADDING_ERR) { ICL_WARN("received PDU 0x%02x with invalid padding", ip->ip_bhs->bhs_opcode); toep->ofld_rxq->rx_iscsi_padding_errors++; } if (val & F_DDP_HDRCRC_ERR) { ICL_WARN("received PDU 0x%02x with invalid header digest", ip->ip_bhs->bhs_opcode); toep->ofld_rxq->rx_iscsi_header_digest_errors++; } if (val & F_DDP_DATACRC_ERR) { ICL_WARN("received PDU 0x%02x with invalid data digest", ip->ip_bhs->bhs_opcode); toep->ofld_rxq->rx_iscsi_data_digest_errors++; } INP_WLOCK(inp); - if (__predict_false(inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT))) { + if (__predict_false(inp->inp_flags & INP_DROPPED)) { CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x", __func__, tid, pdu_len, inp->inp_flags); INP_WUNLOCK(inp); icl_cxgbei_conn_pdu_free(NULL, ip); toep->ulpcb2 = NULL; m_freem(m); return (0); } tp = intotcpcb(inp); /* * If icc is NULL, the connection is being closed in * icl_cxgbei_conn_close(), just drop this data. */ icc = toep->ulpcb; if (__predict_false(icc == NULL)) { CTR4(KTR_CXGBE, "%s: tid %u, excess rx (%d bytes), icc %p", __func__, tid, pdu_len, icc); /* * Update rcv_nxt so the sequence number of the FIN * doesn't appear wrong. */ tp->rcv_nxt = icp->icp_seq + pdu_len; tp->t_rcvtime = ticks; INP_WUNLOCK(inp); icl_cxgbei_conn_pdu_free(NULL, ip); toep->ulpcb2 = NULL; m_freem(m); return (0); } MPASS(icc->icc_signature == CXGBEI_CONN_SIGNATURE); ic = &icc->ic; if ((val & (F_DDP_PADDING_ERR | F_DDP_HDRCRC_ERR | F_DDP_DATACRC_ERR)) != 0) { INP_WUNLOCK(inp); icl_cxgbei_conn_pdu_free(NULL, ip); toep->ulpcb2 = NULL; m_freem(m); ic->ic_error(ic); return (0); } #ifdef INVARIANTS data_digest_len = (icc->ulp_submode & ULP_CRC_DATA) ? ISCSI_DATA_DIGEST_SIZE : 0; MPASS(roundup2(ip->ip_data_len, 4) == pdu_len - len - data_digest_len); #endif if (val & F_DDP_PDU && ip->ip_data_mbuf == NULL) { MPASS((icp->icp_flags & ICPF_RX_FLBUF) == 0); MPASS(ip->ip_data_len > 0); icp->icp_flags |= ICPF_RX_DDP; bhsdo = (struct iscsi_bhs_data_out *)ip->ip_bhs; switch (ip->ip_bhs->bhs_opcode & ~ISCSI_BHS_OPCODE_IMMEDIATE) { case ISCSI_BHS_OPCODE_SCSI_DATA_IN: cmp = cxgbei_find_cmp(icc, be32toh(bhsdo->bhsdo_initiator_task_tag)); break; case ISCSI_BHS_OPCODE_SCSI_DATA_OUT: cmp = cxgbei_find_cmp(icc, be32toh(bhsdo->bhsdo_target_transfer_tag)); break; default: __assert_unreachable(); } MPASS(cmp != NULL); /* * The difference between the end of the last burst * and the offset of the last PDU in this burst is * the additional data received via DDP. */ prev_seg_len = be32toh(bhsdo->bhsdo_buffer_offset) - cmp->next_buffer_offset; if (prev_seg_len != 0) { uint32_t orig_datasn; /* * Return a "large" PDU representing the burst * of PDUs. Adjust the offset and length of * this PDU to represent the entire burst. */ ip->ip_data_len += prev_seg_len; bhsdo->bhsdo_data_segment_len[2] = ip->ip_data_len; bhsdo->bhsdo_data_segment_len[1] = ip->ip_data_len >> 8; bhsdo->bhsdo_data_segment_len[0] = ip->ip_data_len >> 16; bhsdo->bhsdo_buffer_offset = htobe32(cmp->next_buffer_offset); orig_datasn = htobe32(bhsdo->bhsdo_datasn); npdus = orig_datasn - cmp->last_datasn; bhsdo->bhsdo_datasn = htobe32(cmp->last_datasn + 1); cmp->last_datasn = orig_datasn; ip->ip_additional_pdus = npdus - 1; } else { MPASS(htobe32(bhsdo->bhsdo_datasn) == cmp->last_datasn + 1); npdus = 1; cmp->last_datasn = htobe32(bhsdo->bhsdo_datasn); } cmp->next_buffer_offset += ip->ip_data_len; toep->ofld_rxq->rx_iscsi_ddp_pdus += npdus; toep->ofld_rxq->rx_iscsi_ddp_octets += ip->ip_data_len; } else { MPASS(icp->icp_flags & (ICPF_RX_FLBUF)); MPASS(ip->ip_data_len == ip->ip_data_mbuf->m_pkthdr.len); } tp->rcv_nxt = icp->icp_seq + pdu_len; tp->t_rcvtime = ticks; /* * Don't update the window size or return credits since RX * flow control is disabled. */ so = inp->inp_socket; sb = &so->so_rcv; SOCKBUF_LOCK(sb); if (__predict_false(sb->sb_state & SBS_CANTRCVMORE)) { CTR5(KTR_CXGBE, "%s: tid %u, excess rx (%d bytes), icc %p, sb_state 0x%x", __func__, tid, pdu_len, icc, sb->sb_state); SOCKBUF_UNLOCK(sb); INP_WUNLOCK(inp); CURVNET_SET(so->so_vnet); NET_EPOCH_ENTER(et); INP_WLOCK(inp); tp = tcp_drop(tp, ECONNRESET); if (tp != NULL) INP_WUNLOCK(inp); NET_EPOCH_EXIT(et); CURVNET_RESTORE(); icl_cxgbei_conn_pdu_free(NULL, ip); toep->ulpcb2 = NULL; m_freem(m); return (0); } icl_cxgbei_new_pdu_set_conn(ip, ic); /* Enqueue the PDU to the received pdus queue. */ STAILQ_INSERT_TAIL(&icc->rcvd_pdus, ip, ip_next); if (!icc->rx_active) { icc->rx_active = true; wakeup(&icc->rx_active); } SOCKBUF_UNLOCK(sb); INP_WUNLOCK(inp); toep->ulpcb2 = NULL; m_freem(m); return (0); } static int cxgbei_activate(struct adapter *sc) { struct cxgbei_data *ci; int rc; ASSERT_SYNCHRONIZED_OP(sc); if (uld_active(sc, ULD_ISCSI)) { KASSERT(0, ("%s: iSCSI offload already enabled on adapter %p", __func__, sc)); return (0); } if (sc->iscsicaps == 0 || sc->vres.iscsi.size == 0) { device_printf(sc->dev, "not iSCSI offload capable, or capability disabled.\n"); return (ENOSYS); } /* per-adapter softc for iSCSI */ ci = malloc(sizeof(*ci), M_CXGBE, M_ZERO | M_WAITOK); if (ci == NULL) return (ENOMEM); rc = cxgbei_init(sc, ci); if (rc != 0) { free(ci, M_CXGBE); return (rc); } sc->iscsi_ulp_softc = ci; return (0); } static int cxgbei_deactivate(struct adapter *sc) { struct cxgbei_data *ci = sc->iscsi_ulp_softc; ASSERT_SYNCHRONIZED_OP(sc); if (ci != NULL) { sysctl_ctx_free(&ci->ctx); t4_free_ppod_region(&ci->pr); free(ci, M_CXGBE); sc->iscsi_ulp_softc = NULL; } return (0); } static void cxgbei_activate_all(struct adapter *sc, void *arg __unused) { if (begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4isact") != 0) return; /* Activate iSCSI if any port on this adapter has IFCAP_TOE enabled. */ if (sc->offload_map && !uld_active(sc, ULD_ISCSI)) (void) t4_activate_uld(sc, ULD_ISCSI); end_synchronized_op(sc, 0); } static void cxgbei_deactivate_all(struct adapter *sc, void *arg __unused) { if (begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4isdea") != 0) return; if (uld_active(sc, ULD_ISCSI)) (void) t4_deactivate_uld(sc, ULD_ISCSI); end_synchronized_op(sc, 0); } static struct uld_info cxgbei_uld_info = { .uld_id = ULD_ISCSI, .activate = cxgbei_activate, .deactivate = cxgbei_deactivate, }; static int cxgbei_mod_load(void) { int rc; t4_register_cpl_handler(CPL_ISCSI_HDR, do_rx_iscsi_hdr); t4_register_cpl_handler(CPL_ISCSI_DATA, do_rx_iscsi_data); t4_register_cpl_handler(CPL_RX_ISCSI_DDP, do_rx_iscsi_ddp); t4_register_cpl_handler(CPL_RX_ISCSI_CMP, do_rx_iscsi_cmp); rc = t4_register_uld(&cxgbei_uld_info); if (rc != 0) return (rc); t4_iterate(cxgbei_activate_all, NULL); return (rc); } static int cxgbei_mod_unload(void) { t4_iterate(cxgbei_deactivate_all, NULL); if (t4_unregister_uld(&cxgbei_uld_info) == EBUSY) return (EBUSY); t4_register_cpl_handler(CPL_ISCSI_HDR, NULL); t4_register_cpl_handler(CPL_ISCSI_DATA, NULL); t4_register_cpl_handler(CPL_RX_ISCSI_DDP, NULL); t4_register_cpl_handler(CPL_RX_ISCSI_CMP, NULL); return (0); } #endif static int cxgbei_modevent(module_t mod, int cmd, void *arg) { int rc = 0; #ifdef TCP_OFFLOAD switch (cmd) { case MOD_LOAD: rc = cxgbei_mod_load(); if (rc == 0) rc = icl_cxgbei_mod_load(); break; case MOD_UNLOAD: rc = icl_cxgbei_mod_unload(); if (rc == 0) rc = cxgbei_mod_unload(); break; default: rc = EINVAL; } #else printf("cxgbei: compiled without TCP_OFFLOAD support.\n"); rc = EOPNOTSUPP; #endif return (rc); } static moduledata_t cxgbei_mod = { "cxgbei", cxgbei_modevent, NULL, }; MODULE_VERSION(cxgbei, 1); DECLARE_MODULE(cxgbei, cxgbei_mod, SI_SUB_EXEC, SI_ORDER_ANY); MODULE_DEPEND(cxgbei, t4_tom, 1, 1, 1); MODULE_DEPEND(cxgbei, cxgbe, 1, 1, 1); MODULE_DEPEND(cxgbei, icl, 1, 1, 1); diff --git a/sys/dev/cxgbe/cxgbei/icl_cxgbei.c b/sys/dev/cxgbe/cxgbei/icl_cxgbei.c index 1b896516d546..82201b358e91 100644 --- a/sys/dev/cxgbe/cxgbei/icl_cxgbei.c +++ b/sys/dev/cxgbe/cxgbei/icl_cxgbei.c @@ -1,1880 +1,1879 @@ /*- * Copyright (c) 2012 The FreeBSD Foundation * Copyright (c) 2015 Chelsio Communications, Inc. * All rights reserved. * * This software was developed by Edward Tomasz Napierala under sponsorship * from the FreeBSD Foundation. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * */ /* * cxgbei implementation of iSCSI Common Layer kobj(9) interface. */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #ifdef TCP_OFFLOAD #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include "common/common.h" #include "common/t4_regs.h" #include "common/t4_tcb.h" #include "tom/t4_tom.h" #include "cxgbei.h" /* * Use the page pod tag for the TT hash. */ #define TT_HASH(icc, tt) (G_PPOD_TAG(tt) & (icc)->cmp_hash_mask) struct cxgbei_ddp_state { struct ppod_reservation prsv; struct cxgbei_cmp cmp; }; static MALLOC_DEFINE(M_CXGBEI, "cxgbei", "cxgbei(4)"); SYSCTL_NODE(_kern_icl, OID_AUTO, cxgbei, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "Chelsio iSCSI offload"); static int first_burst_length = 8192; SYSCTL_INT(_kern_icl_cxgbei, OID_AUTO, first_burst_length, CTLFLAG_RWTUN, &first_burst_length, 0, "First burst length"); static int max_burst_length = 2 * 1024 * 1024; SYSCTL_INT(_kern_icl_cxgbei, OID_AUTO, max_burst_length, CTLFLAG_RWTUN, &max_burst_length, 0, "Maximum burst length"); static int sendspace = 1048576; SYSCTL_INT(_kern_icl_cxgbei, OID_AUTO, sendspace, CTLFLAG_RWTUN, &sendspace, 0, "Default send socket buffer size"); static int recvspace = 1048576; SYSCTL_INT(_kern_icl_cxgbei, OID_AUTO, recvspace, CTLFLAG_RWTUN, &recvspace, 0, "Default receive socket buffer size"); static volatile u_int icl_cxgbei_ncons; static icl_conn_new_pdu_t icl_cxgbei_conn_new_pdu; static icl_conn_pdu_data_segment_length_t icl_cxgbei_conn_pdu_data_segment_length; static icl_conn_pdu_append_bio_t icl_cxgbei_conn_pdu_append_bio; static icl_conn_pdu_append_data_t icl_cxgbei_conn_pdu_append_data; static icl_conn_pdu_get_bio_t icl_cxgbei_conn_pdu_get_bio; static icl_conn_pdu_get_data_t icl_cxgbei_conn_pdu_get_data; static icl_conn_pdu_queue_t icl_cxgbei_conn_pdu_queue; static icl_conn_pdu_queue_cb_t icl_cxgbei_conn_pdu_queue_cb; static icl_conn_handoff_t icl_cxgbei_conn_handoff; static icl_conn_free_t icl_cxgbei_conn_free; static icl_conn_close_t icl_cxgbei_conn_close; static icl_conn_task_setup_t icl_cxgbei_conn_task_setup; static icl_conn_task_done_t icl_cxgbei_conn_task_done; static icl_conn_transfer_setup_t icl_cxgbei_conn_transfer_setup; static icl_conn_transfer_done_t icl_cxgbei_conn_transfer_done; static kobj_method_t icl_cxgbei_methods[] = { KOBJMETHOD(icl_conn_new_pdu, icl_cxgbei_conn_new_pdu), KOBJMETHOD(icl_conn_pdu_free, icl_cxgbei_conn_pdu_free), KOBJMETHOD(icl_conn_pdu_data_segment_length, icl_cxgbei_conn_pdu_data_segment_length), KOBJMETHOD(icl_conn_pdu_append_bio, icl_cxgbei_conn_pdu_append_bio), KOBJMETHOD(icl_conn_pdu_append_data, icl_cxgbei_conn_pdu_append_data), KOBJMETHOD(icl_conn_pdu_get_bio, icl_cxgbei_conn_pdu_get_bio), KOBJMETHOD(icl_conn_pdu_get_data, icl_cxgbei_conn_pdu_get_data), KOBJMETHOD(icl_conn_pdu_queue, icl_cxgbei_conn_pdu_queue), KOBJMETHOD(icl_conn_pdu_queue_cb, icl_cxgbei_conn_pdu_queue_cb), KOBJMETHOD(icl_conn_handoff, icl_cxgbei_conn_handoff), KOBJMETHOD(icl_conn_free, icl_cxgbei_conn_free), KOBJMETHOD(icl_conn_close, icl_cxgbei_conn_close), KOBJMETHOD(icl_conn_task_setup, icl_cxgbei_conn_task_setup), KOBJMETHOD(icl_conn_task_done, icl_cxgbei_conn_task_done), KOBJMETHOD(icl_conn_transfer_setup, icl_cxgbei_conn_transfer_setup), KOBJMETHOD(icl_conn_transfer_done, icl_cxgbei_conn_transfer_done), { 0, 0 } }; DEFINE_CLASS(icl_cxgbei, icl_cxgbei_methods, sizeof(struct icl_cxgbei_conn)); void icl_cxgbei_conn_pdu_free(struct icl_conn *ic, struct icl_pdu *ip) { struct icl_cxgbei_pdu *icp = ip_to_icp(ip); KASSERT(icp->ref_cnt != 0, ("freeing deleted PDU")); MPASS(icp->icp_signature == CXGBEI_PDU_SIGNATURE); MPASS(ic == ip->ip_conn); m_freem(ip->ip_ahs_mbuf); m_freem(ip->ip_data_mbuf); m_freem(ip->ip_bhs_mbuf); KASSERT(ic != NULL || icp->ref_cnt == 1, ("orphaned PDU has oustanding references")); if (atomic_fetchadd_int(&icp->ref_cnt, -1) != 1) return; free(icp, M_CXGBEI); #ifdef DIAGNOSTIC if (__predict_true(ic != NULL)) refcount_release(&ic->ic_outstanding_pdus); #endif } static void icl_cxgbei_pdu_call_cb(struct icl_pdu *ip) { struct icl_cxgbei_pdu *icp = ip_to_icp(ip); MPASS(icp->icp_signature == CXGBEI_PDU_SIGNATURE); if (icp->cb != NULL) icp->cb(ip, icp->error); #ifdef DIAGNOSTIC if (__predict_true(ip->ip_conn != NULL)) refcount_release(&ip->ip_conn->ic_outstanding_pdus); #endif free(icp, M_CXGBEI); } static void icl_cxgbei_pdu_done(struct icl_pdu *ip, int error) { struct icl_cxgbei_pdu *icp = ip_to_icp(ip); if (error != 0) icp->error = error; m_freem(ip->ip_ahs_mbuf); ip->ip_ahs_mbuf = NULL; m_freem(ip->ip_data_mbuf); ip->ip_data_mbuf = NULL; m_freem(ip->ip_bhs_mbuf); ip->ip_bhs_mbuf = NULL; /* * All other references to this PDU should have been dropped * by the m_freem() of ip_data_mbuf. */ if (atomic_fetchadd_int(&icp->ref_cnt, -1) == 1) icl_cxgbei_pdu_call_cb(ip); else __assert_unreachable(); } static void icl_cxgbei_mbuf_done(struct mbuf *mb) { struct icl_cxgbei_pdu *icp = (struct icl_cxgbei_pdu *)mb->m_ext.ext_arg1; /* * NB: mb_free_mext() might leave ref_cnt as 1 without * decrementing it if it hits the fast path in the ref_cnt * check. */ icl_cxgbei_pdu_call_cb(&icp->ip); } struct icl_pdu * icl_cxgbei_new_pdu(int flags) { struct icl_cxgbei_pdu *icp; struct icl_pdu *ip; struct mbuf *m; icp = malloc(sizeof(*icp), M_CXGBEI, flags | M_ZERO); if (__predict_false(icp == NULL)) return (NULL); icp->icp_signature = CXGBEI_PDU_SIGNATURE; icp->ref_cnt = 1; ip = &icp->ip; m = m_gethdr(flags, MT_DATA); if (__predict_false(m == NULL)) { free(icp, M_CXGBEI); return (NULL); } ip->ip_bhs_mbuf = m; ip->ip_bhs = mtod(m, struct iscsi_bhs *); memset(ip->ip_bhs, 0, sizeof(*ip->ip_bhs)); m->m_len = sizeof(struct iscsi_bhs); m->m_pkthdr.len = m->m_len; return (ip); } void icl_cxgbei_new_pdu_set_conn(struct icl_pdu *ip, struct icl_conn *ic) { ip->ip_conn = ic; #ifdef DIAGNOSTIC refcount_acquire(&ic->ic_outstanding_pdus); #endif } /* * Allocate icl_pdu with empty BHS to fill up by the caller. */ static struct icl_pdu * icl_cxgbei_conn_new_pdu(struct icl_conn *ic, int flags) { struct icl_pdu *ip; ip = icl_cxgbei_new_pdu(flags); if (__predict_false(ip == NULL)) return (NULL); icl_cxgbei_new_pdu_set_conn(ip, ic); return (ip); } static size_t icl_pdu_data_segment_length(const struct icl_pdu *request) { uint32_t len = 0; len += request->ip_bhs->bhs_data_segment_len[0]; len <<= 8; len += request->ip_bhs->bhs_data_segment_len[1]; len <<= 8; len += request->ip_bhs->bhs_data_segment_len[2]; return (len); } size_t icl_cxgbei_conn_pdu_data_segment_length(struct icl_conn *ic, const struct icl_pdu *request) { return (icl_pdu_data_segment_length(request)); } static struct mbuf * finalize_pdu(struct icl_cxgbei_conn *icc, struct icl_cxgbei_pdu *icp) { struct icl_pdu *ip = &icp->ip; uint8_t ulp_submode, padding; struct mbuf *m, *last; struct iscsi_bhs *bhs; int data_len; /* * Fix up the data segment mbuf first. */ m = ip->ip_data_mbuf; ulp_submode = icc->ulp_submode; if (m != NULL) { last = m_last(m); /* * Round up the data segment to a 4B boundary. Pad with 0 if * necessary. There will definitely be room in the mbuf. */ padding = roundup2(ip->ip_data_len, 4) - ip->ip_data_len; if (padding != 0) { MPASS(padding <= M_TRAILINGSPACE(last)); bzero(mtod(last, uint8_t *) + last->m_len, padding); last->m_len += padding; } } else { MPASS(ip->ip_data_len == 0); ulp_submode &= ~ULP_CRC_DATA; padding = 0; } /* * Now the header mbuf that has the BHS. */ m = ip->ip_bhs_mbuf; MPASS(m->m_pkthdr.len == sizeof(struct iscsi_bhs)); MPASS(m->m_len == sizeof(struct iscsi_bhs)); bhs = ip->ip_bhs; data_len = ip->ip_data_len; if (data_len > icc->ic.ic_max_send_data_segment_length) { struct iscsi_bhs_data_in *bhsdi; int flags; KASSERT(padding == 0, ("%s: ISO with padding %d for icp %p", __func__, padding, icp)); switch (bhs->bhs_opcode) { case ISCSI_BHS_OPCODE_SCSI_DATA_OUT: flags = 1; break; case ISCSI_BHS_OPCODE_SCSI_DATA_IN: flags = 2; break; default: panic("invalid opcode %#x for ISO", bhs->bhs_opcode); } data_len = icc->ic.ic_max_send_data_segment_length; bhsdi = (struct iscsi_bhs_data_in *)bhs; if (bhsdi->bhsdi_flags & BHSDI_FLAGS_F) { /* * Firmware will set F on the final PDU in the * burst. */ flags |= CXGBE_ISO_F; bhsdi->bhsdi_flags &= ~BHSDI_FLAGS_F; } set_mbuf_iscsi_iso(m, true); set_mbuf_iscsi_iso_flags(m, flags); set_mbuf_iscsi_iso_mss(m, data_len); } bhs->bhs_data_segment_len[2] = data_len; bhs->bhs_data_segment_len[1] = data_len >> 8; bhs->bhs_data_segment_len[0] = data_len >> 16; /* * Extract mbuf chain from PDU. */ m->m_pkthdr.len += ip->ip_data_len + padding; m->m_next = ip->ip_data_mbuf; set_mbuf_ulp_submode(m, ulp_submode); ip->ip_bhs_mbuf = NULL; ip->ip_data_mbuf = NULL; ip->ip_bhs = NULL; /* * Drop PDU reference on icp. Additional references might * still be held by zero-copy PDU buffers (ICL_NOCOPY). */ if (atomic_fetchadd_int(&icp->ref_cnt, -1) == 1) icl_cxgbei_pdu_call_cb(ip); return (m); } static void icl_cxgbei_tx_main(void *arg) { struct epoch_tracker et; struct icl_cxgbei_conn *icc = arg; struct icl_conn *ic = &icc->ic; struct toepcb *toep = icc->toep; struct socket *so = ic->ic_socket; struct inpcb *inp = sotoinpcb(so); struct icl_pdu *ip; struct mbuf *m; struct mbufq mq; STAILQ_HEAD(, icl_pdu) tx_pdus = STAILQ_HEAD_INITIALIZER(tx_pdus); mbufq_init(&mq, INT_MAX); ICL_CONN_LOCK(ic); while (__predict_true(!ic->ic_disconnecting)) { while (STAILQ_EMPTY(&icc->sent_pdus)) { icc->tx_active = false; mtx_sleep(&icc->tx_active, ic->ic_lock, 0, "-", 0); if (__predict_false(ic->ic_disconnecting)) goto out; MPASS(icc->tx_active); } STAILQ_SWAP(&icc->sent_pdus, &tx_pdus, icl_pdu); ICL_CONN_UNLOCK(ic); while ((ip = STAILQ_FIRST(&tx_pdus)) != NULL) { STAILQ_REMOVE_HEAD(&tx_pdus, ip_next); m = finalize_pdu(icc, ip_to_icp(ip)); M_ASSERTPKTHDR(m); MPASS((m->m_pkthdr.len & 3) == 0); mbufq_enqueue(&mq, m); } ICL_CONN_LOCK(ic); if (__predict_false(ic->ic_disconnecting) || __predict_false(ic->ic_socket == NULL)) { mbufq_drain(&mq); break; } CURVNET_SET(toep->vnet); NET_EPOCH_ENTER(et); INP_WLOCK(inp); ICL_CONN_UNLOCK(ic); - if (__predict_false(inp->inp_flags & (INP_DROPPED | - INP_TIMEWAIT)) || + if (__predict_false(inp->inp_flags & INP_DROPPED) || __predict_false((toep->flags & TPF_ATTACHED) == 0)) { mbufq_drain(&mq); } else { mbufq_concat(&toep->ulp_pduq, &mq); t4_push_pdus(icc->sc, toep, 0); } INP_WUNLOCK(inp); NET_EPOCH_EXIT(et); CURVNET_RESTORE(); ICL_CONN_LOCK(ic); } out: ICL_CONN_UNLOCK(ic); kthread_exit(); } static void icl_cxgbei_rx_main(void *arg) { struct icl_cxgbei_conn *icc = arg; struct icl_conn *ic = &icc->ic; struct icl_pdu *ip; struct sockbuf *sb; STAILQ_HEAD(, icl_pdu) rx_pdus = STAILQ_HEAD_INITIALIZER(rx_pdus); bool cantrcvmore; sb = &ic->ic_socket->so_rcv; SOCKBUF_LOCK(sb); while (__predict_true(!ic->ic_disconnecting)) { while (STAILQ_EMPTY(&icc->rcvd_pdus)) { icc->rx_active = false; mtx_sleep(&icc->rx_active, SOCKBUF_MTX(sb), 0, "-", 0); if (__predict_false(ic->ic_disconnecting)) goto out; MPASS(icc->rx_active); } if (__predict_false(sbused(sb)) != 0) { /* * PDUs were received before the tid * transitioned to ULP mode. Convert * them to icl_cxgbei_pdus and insert * them into the head of rcvd_pdus. */ parse_pdus(icc, sb); } cantrcvmore = (sb->sb_state & SBS_CANTRCVMORE) != 0; MPASS(STAILQ_EMPTY(&rx_pdus)); STAILQ_SWAP(&icc->rcvd_pdus, &rx_pdus, icl_pdu); SOCKBUF_UNLOCK(sb); /* Hand over PDUs to ICL. */ while ((ip = STAILQ_FIRST(&rx_pdus)) != NULL) { STAILQ_REMOVE_HEAD(&rx_pdus, ip_next); if (cantrcvmore) icl_cxgbei_pdu_done(ip, ENOTCONN); else ic->ic_receive(ip); } SOCKBUF_LOCK(sb); } out: /* * Since ic_disconnecting is set before the SOCKBUF_MTX is * locked in icl_cxgbei_conn_close, the loop above can exit * before icl_cxgbei_conn_close can lock SOCKBUF_MTX and block * waiting for the thread exit. */ while (!icc->rx_exiting) mtx_sleep(&icc->rx_active, SOCKBUF_MTX(sb), 0, "-", 0); SOCKBUF_UNLOCK(sb); kthread_exit(); } static void cxgbei_free_mext_pg(struct mbuf *m) { struct icl_cxgbei_pdu *icp; M_ASSERTEXTPG(m); /* * Nothing to do for the pages; they are owned by the PDU / * I/O request. */ /* Drop reference on the PDU. */ icp = m->m_ext.ext_arg1; if (atomic_fetchadd_int(&icp->ref_cnt, -1) == 1) icl_cxgbei_pdu_call_cb(&icp->ip); } static struct mbuf * cxgbei_getm(size_t len, int flags) { struct mbuf *m, *m0, *m_tail; m_tail = m0 = NULL; /* Allocate as jumbo mbufs of size MJUM16BYTES. */ while (len >= MJUM16BYTES) { m = m_getjcl(M_NOWAIT, MT_DATA, 0, MJUM16BYTES); if (__predict_false(m == NULL)) { if ((flags & M_WAITOK) != 0) { /* Fall back to non-jumbo mbufs. */ break; } return (NULL); } if (m0 == NULL) { m0 = m_tail = m; } else { m_tail->m_next = m; m_tail = m; } len -= MJUM16BYTES; } /* Allocate mbuf chain for the remaining data. */ if (len != 0) { m = m_getm2(NULL, len, flags, MT_DATA, 0); if (__predict_false(m == NULL)) { m_freem(m0); return (NULL); } if (m0 == NULL) m0 = m; else m_tail->m_next = m; } return (m0); } int icl_cxgbei_conn_pdu_append_bio(struct icl_conn *ic, struct icl_pdu *ip, struct bio *bp, size_t offset, size_t len, int flags) { struct icl_cxgbei_pdu *icp = ip_to_icp(ip); struct mbuf *m, *m_tail; vm_offset_t vaddr; size_t page_offset, todo, mtodo; boolean_t mapped; int i; MPASS(icp->icp_signature == CXGBEI_PDU_SIGNATURE); MPASS(ic == ip->ip_conn); KASSERT(len > 0, ("%s: len is %jd", __func__, (intmax_t)len)); m_tail = ip->ip_data_mbuf; if (m_tail != NULL) for (; m_tail->m_next != NULL; m_tail = m_tail->m_next) ; MPASS(bp->bio_flags & BIO_UNMAPPED); if (offset < PAGE_SIZE - bp->bio_ma_offset) { page_offset = bp->bio_ma_offset + offset; i = 0; } else { offset -= PAGE_SIZE - bp->bio_ma_offset; for (i = 1; offset >= PAGE_SIZE; i++) offset -= PAGE_SIZE; page_offset = offset; } if (flags & ICL_NOCOPY) { m = NULL; while (len > 0) { if (m == NULL) { m = mb_alloc_ext_pgs(flags & ~ICL_NOCOPY, cxgbei_free_mext_pg); if (__predict_false(m == NULL)) return (ENOMEM); atomic_add_int(&icp->ref_cnt, 1); m->m_ext.ext_arg1 = icp; m->m_epg_1st_off = page_offset; } todo = MIN(len, PAGE_SIZE - page_offset); m->m_epg_pa[m->m_epg_npgs] = VM_PAGE_TO_PHYS(bp->bio_ma[i]); m->m_epg_npgs++; m->m_epg_last_len = todo; m->m_len += todo; m->m_ext.ext_size += PAGE_SIZE; MBUF_EXT_PGS_ASSERT_SANITY(m); if (m->m_epg_npgs == MBUF_PEXT_MAX_PGS) { if (m_tail != NULL) m_tail->m_next = m; else ip->ip_data_mbuf = m; m_tail = m; ip->ip_data_len += m->m_len; m = NULL; } page_offset = 0; len -= todo; i++; } if (m != NULL) { if (m_tail != NULL) m_tail->m_next = m; else ip->ip_data_mbuf = m; ip->ip_data_len += m->m_len; } return (0); } m = cxgbei_getm(len, flags); if (__predict_false(m == NULL)) return (ENOMEM); if (ip->ip_data_mbuf == NULL) { ip->ip_data_mbuf = m; ip->ip_data_len = len; } else { m_tail->m_next = m; ip->ip_data_len += len; } while (len > 0) { todo = MIN(len, PAGE_SIZE - page_offset); mapped = pmap_map_io_transient(bp->bio_ma + i, &vaddr, 1, FALSE); do { mtodo = min(todo, M_SIZE(m) - m->m_len); memcpy(mtod(m, char *) + m->m_len, (char *)vaddr + page_offset, mtodo); m->m_len += mtodo; if (m->m_len == M_SIZE(m)) m = m->m_next; page_offset += mtodo; todo -= mtodo; } while (todo > 0); if (__predict_false(mapped)) pmap_unmap_io_transient(bp->bio_ma + 1, &vaddr, 1, FALSE); page_offset = 0; len -= todo; i++; } MPASS(ip->ip_data_len <= max(ic->ic_max_send_data_segment_length, ic->ic_hw_isomax)); return (0); } int icl_cxgbei_conn_pdu_append_data(struct icl_conn *ic, struct icl_pdu *ip, const void *addr, size_t len, int flags) { struct icl_cxgbei_pdu *icp = ip_to_icp(ip); struct mbuf *m, *m_tail; const char *src; MPASS(icp->icp_signature == CXGBEI_PDU_SIGNATURE); MPASS(ic == ip->ip_conn); KASSERT(len > 0, ("%s: len is %jd", __func__, (intmax_t)len)); m_tail = ip->ip_data_mbuf; if (m_tail != NULL) for (; m_tail->m_next != NULL; m_tail = m_tail->m_next) ; if (flags & ICL_NOCOPY) { m = m_get(flags & ~ICL_NOCOPY, MT_DATA); if (m == NULL) { ICL_WARN("failed to allocate mbuf"); return (ENOMEM); } m->m_flags |= M_RDONLY; m_extaddref(m, __DECONST(char *, addr), len, &icp->ref_cnt, icl_cxgbei_mbuf_done, icp, NULL); m->m_len = len; if (ip->ip_data_mbuf == NULL) { ip->ip_data_mbuf = m; ip->ip_data_len = len; } else { m_tail->m_next = m; m_tail = m_tail->m_next; ip->ip_data_len += len; } return (0); } m = cxgbei_getm(len, flags); if (__predict_false(m == NULL)) return (ENOMEM); if (ip->ip_data_mbuf == NULL) { ip->ip_data_mbuf = m; ip->ip_data_len = len; } else { m_tail->m_next = m; ip->ip_data_len += len; } src = (const char *)addr; for (; m != NULL; m = m->m_next) { m->m_len = min(len, M_SIZE(m)); memcpy(mtod(m, void *), src, m->m_len); src += m->m_len; len -= m->m_len; } MPASS(len == 0); MPASS(ip->ip_data_len <= max(ic->ic_max_send_data_segment_length, ic->ic_hw_isomax)); return (0); } void icl_cxgbei_conn_pdu_get_bio(struct icl_conn *ic, struct icl_pdu *ip, size_t pdu_off, struct bio *bp, size_t bio_off, size_t len) { struct icl_cxgbei_pdu *icp = ip_to_icp(ip); vm_offset_t vaddr; size_t page_offset, todo; boolean_t mapped; int i; if (icp->icp_flags & ICPF_RX_DDP) return; /* data is DDP'ed, no need to copy */ MPASS(bp->bio_flags & BIO_UNMAPPED); if (bio_off < PAGE_SIZE - bp->bio_ma_offset) { page_offset = bp->bio_ma_offset + bio_off; i = 0; } else { bio_off -= PAGE_SIZE - bp->bio_ma_offset; for (i = 1; bio_off >= PAGE_SIZE; i++) bio_off -= PAGE_SIZE; page_offset = bio_off; } while (len > 0) { todo = MIN(len, PAGE_SIZE - page_offset); mapped = pmap_map_io_transient(bp->bio_ma + i, &vaddr, 1, FALSE); m_copydata(ip->ip_data_mbuf, pdu_off, todo, (char *)vaddr + page_offset); if (__predict_false(mapped)) pmap_unmap_io_transient(bp->bio_ma + 1, &vaddr, 1, FALSE); page_offset = 0; pdu_off += todo; len -= todo; i++; } } void icl_cxgbei_conn_pdu_get_data(struct icl_conn *ic, struct icl_pdu *ip, size_t off, void *addr, size_t len) { struct icl_cxgbei_pdu *icp = ip_to_icp(ip); if (icp->icp_flags & ICPF_RX_DDP) return; /* data is DDP'ed, no need to copy */ m_copydata(ip->ip_data_mbuf, off, len, addr); } void icl_cxgbei_conn_pdu_queue(struct icl_conn *ic, struct icl_pdu *ip) { icl_cxgbei_conn_pdu_queue_cb(ic, ip, NULL); } void icl_cxgbei_conn_pdu_queue_cb(struct icl_conn *ic, struct icl_pdu *ip, icl_pdu_cb cb) { struct icl_cxgbei_conn *icc = ic_to_icc(ic); struct icl_cxgbei_pdu *icp = ip_to_icp(ip); struct socket *so = ic->ic_socket; MPASS(ic == ip->ip_conn); MPASS(ip->ip_bhs_mbuf != NULL); /* The kernel doesn't generate PDUs with AHS. */ MPASS(ip->ip_ahs_mbuf == NULL && ip->ip_ahs_len == 0); ICL_CONN_LOCK_ASSERT(ic); icp->cb = cb; /* NOTE: sowriteable without so_snd lock is a mostly harmless race. */ if (ic->ic_disconnecting || so == NULL || !sowriteable(so)) { icl_cxgbei_pdu_done(ip, ENOTCONN); return; } STAILQ_INSERT_TAIL(&icc->sent_pdus, ip, ip_next); if (!icc->tx_active) { icc->tx_active = true; wakeup(&icc->tx_active); } } static struct icl_conn * icl_cxgbei_new_conn(const char *name, struct mtx *lock) { struct icl_cxgbei_conn *icc; struct icl_conn *ic; refcount_acquire(&icl_cxgbei_ncons); icc = (struct icl_cxgbei_conn *)kobj_create(&icl_cxgbei_class, M_CXGBE, M_WAITOK | M_ZERO); icc->icc_signature = CXGBEI_CONN_SIGNATURE; STAILQ_INIT(&icc->rcvd_pdus); STAILQ_INIT(&icc->sent_pdus); icc->cmp_table = hashinit(64, M_CXGBEI, &icc->cmp_hash_mask); mtx_init(&icc->cmp_lock, "cxgbei_cmp", NULL, MTX_DEF); ic = &icc->ic; ic->ic_lock = lock; #ifdef DIAGNOSTIC refcount_init(&ic->ic_outstanding_pdus, 0); #endif ic->ic_name = name; ic->ic_offload = "cxgbei"; ic->ic_unmapped = true; CTR2(KTR_CXGBE, "%s: icc %p", __func__, icc); return (ic); } void icl_cxgbei_conn_free(struct icl_conn *ic) { struct icl_cxgbei_conn *icc = ic_to_icc(ic); MPASS(icc->icc_signature == CXGBEI_CONN_SIGNATURE); CTR2(KTR_CXGBE, "%s: icc %p", __func__, icc); mtx_destroy(&icc->cmp_lock); hashdestroy(icc->cmp_table, M_CXGBEI, icc->cmp_hash_mask); kobj_delete((struct kobj *)icc, M_CXGBE); refcount_release(&icl_cxgbei_ncons); } static int icl_cxgbei_setsockopt(struct icl_conn *ic, struct socket *so, int sspace, int rspace) { struct sockopt opt; int error, one = 1, ss, rs; ss = max(sendspace, sspace); rs = max(recvspace, rspace); error = soreserve(so, ss, rs); if (error != 0) return (error); SOCKBUF_LOCK(&so->so_snd); so->so_snd.sb_flags |= SB_AUTOSIZE; SOCKBUF_UNLOCK(&so->so_snd); SOCKBUF_LOCK(&so->so_rcv); so->so_rcv.sb_flags |= SB_AUTOSIZE; SOCKBUF_UNLOCK(&so->so_rcv); /* * Disable Nagle. */ bzero(&opt, sizeof(opt)); opt.sopt_dir = SOPT_SET; opt.sopt_level = IPPROTO_TCP; opt.sopt_name = TCP_NODELAY; opt.sopt_val = &one; opt.sopt_valsize = sizeof(one); error = sosetopt(so, &opt); if (error != 0) return (error); return (0); } /* * Request/response structure used to find out the adapter offloading a socket. */ struct find_ofld_adapter_rr { struct socket *so; struct adapter *sc; /* result */ }; static void find_offload_adapter(struct adapter *sc, void *arg) { struct find_ofld_adapter_rr *fa = arg; struct socket *so = fa->so; struct tom_data *td = sc->tom_softc; struct tcpcb *tp; struct inpcb *inp; /* Non-TCP were filtered out earlier. */ MPASS(so->so_proto->pr_protocol == IPPROTO_TCP); if (fa->sc != NULL) return; /* Found already. */ if (td == NULL) return; /* TOE not enabled on this adapter. */ inp = sotoinpcb(so); INP_WLOCK(inp); - if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) { + if ((inp->inp_flags & INP_DROPPED) == 0) { tp = intotcpcb(inp); if (tp->t_flags & TF_TOE && tp->tod == &td->tod) fa->sc = sc; /* Found. */ } INP_WUNLOCK(inp); } static bool is_memfree(struct adapter *sc) { uint32_t em; em = t4_read_reg(sc, A_MA_TARGET_MEM_ENABLE); if ((em & F_EXT_MEM_ENABLE) != 0) return (false); if (is_t5(sc) && (em & F_EXT_MEM1_ENABLE) != 0) return (false); return (true); } /* XXXNP: move this to t4_tom. */ static void send_iscsi_flowc_wr(struct adapter *sc, struct toepcb *toep, int maxlen) { struct wrqe *wr; struct fw_flowc_wr *flowc; const u_int nparams = 1; u_int flowclen; struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx]; flowclen = sizeof(*flowc) + nparams * sizeof(struct fw_flowc_mnemval); wr = alloc_wrqe(roundup2(flowclen, 16), &toep->ofld_txq->wrq); if (wr == NULL) { /* XXX */ panic("%s: allocation failure.", __func__); } flowc = wrtod(wr); memset(flowc, 0, wr->wr_len); flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) | V_FW_FLOWC_WR_NPARAMS(nparams)); flowc->flowid_len16 = htonl(V_FW_WR_LEN16(howmany(flowclen, 16)) | V_FW_WR_FLOWID(toep->tid)); flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_TXDATAPLEN_MAX; flowc->mnemval[0].val = htobe32(maxlen); txsd->tx_credits = howmany(flowclen, 16); txsd->plen = 0; KASSERT(toep->tx_credits >= txsd->tx_credits && toep->txsd_avail > 0, ("%s: not enough credits (%d)", __func__, toep->tx_credits)); toep->tx_credits -= txsd->tx_credits; if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) toep->txsd_pidx = 0; toep->txsd_avail--; t4_wrq_tx(sc, wr); } static void set_ulp_mode_iscsi(struct adapter *sc, struct toepcb *toep, u_int ulp_submode) { uint64_t val; CTR3(KTR_CXGBE, "%s: tid %u, ULP_MODE_ISCSI, submode=%#x", __func__, toep->tid, ulp_submode); val = V_TCB_ULP_TYPE(ULP_MODE_ISCSI) | V_TCB_ULP_RAW(ulp_submode); t4_set_tcb_field(sc, toep->ctrlq, toep, W_TCB_ULP_TYPE, V_TCB_ULP_TYPE(M_TCB_ULP_TYPE) | V_TCB_ULP_RAW(M_TCB_ULP_RAW), val, 0, 0); val = V_TF_RX_FLOW_CONTROL_DISABLE(1ULL); t4_set_tcb_field(sc, toep->ctrlq, toep, W_TCB_T_FLAGS, val, val, 0, 0); } /* * XXXNP: Who is responsible for cleaning up the socket if this returns with an * error? Review all error paths. * * XXXNP: What happens to the socket's fd reference if the operation is * successful, and how does that affect the socket's life cycle? */ int icl_cxgbei_conn_handoff(struct icl_conn *ic, int fd) { struct icl_cxgbei_conn *icc = ic_to_icc(ic); struct find_ofld_adapter_rr fa; struct file *fp; struct socket *so; struct inpcb *inp; struct tcpcb *tp; struct toepcb *toep; cap_rights_t rights; u_int max_iso_payload, max_rx_pdu_len, max_tx_pdu_len; int error, max_iso_pdus; MPASS(icc->icc_signature == CXGBEI_CONN_SIGNATURE); ICL_CONN_LOCK_ASSERT_NOT(ic); /* * Steal the socket from userland. */ error = fget(curthread, fd, cap_rights_init_one(&rights, CAP_SOCK_CLIENT), &fp); if (error != 0) return (error); if (fp->f_type != DTYPE_SOCKET) { fdrop(fp, curthread); return (EINVAL); } so = fp->f_data; if (so->so_type != SOCK_STREAM || so->so_proto->pr_protocol != IPPROTO_TCP) { fdrop(fp, curthread); return (EINVAL); } ICL_CONN_LOCK(ic); if (ic->ic_socket != NULL) { ICL_CONN_UNLOCK(ic); fdrop(fp, curthread); return (EBUSY); } ic->ic_disconnecting = false; ic->ic_socket = so; fp->f_ops = &badfileops; fp->f_data = NULL; fdrop(fp, curthread); ICL_CONN_UNLOCK(ic); /* Find the adapter offloading this socket. */ fa.sc = NULL; fa.so = so; t4_iterate(find_offload_adapter, &fa); if (fa.sc == NULL) { error = EINVAL; goto out; } icc->sc = fa.sc; max_rx_pdu_len = ISCSI_BHS_SIZE + ic->ic_max_recv_data_segment_length; max_tx_pdu_len = ISCSI_BHS_SIZE + ic->ic_max_send_data_segment_length; if (ic->ic_header_crc32c) { max_rx_pdu_len += ISCSI_HEADER_DIGEST_SIZE; max_tx_pdu_len += ISCSI_HEADER_DIGEST_SIZE; } if (ic->ic_data_crc32c) { max_rx_pdu_len += ISCSI_DATA_DIGEST_SIZE; max_tx_pdu_len += ISCSI_DATA_DIGEST_SIZE; } inp = sotoinpcb(so); INP_WLOCK(inp); tp = intotcpcb(inp); - if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) { + if (inp->inp_flags & INP_DROPPED) { INP_WUNLOCK(inp); error = ENOTCONN; goto out; } /* * socket could not have been "unoffloaded" if here. */ MPASS(tp->t_flags & TF_TOE); MPASS(tp->tod != NULL); MPASS(tp->t_toe != NULL); toep = tp->t_toe; MPASS(toep->vi->adapter == icc->sc); if (ulp_mode(toep) != ULP_MODE_NONE) { INP_WUNLOCK(inp); error = EINVAL; goto out; } icc->toep = toep; icc->ulp_submode = 0; if (ic->ic_header_crc32c) icc->ulp_submode |= ULP_CRC_HEADER; if (ic->ic_data_crc32c) icc->ulp_submode |= ULP_CRC_DATA; if (icc->sc->tt.iso && chip_id(icc->sc) >= CHELSIO_T5 && !is_memfree(icc->sc)) { max_iso_payload = rounddown(CXGBEI_MAX_ISO_PAYLOAD, tp->t_maxseg); max_iso_pdus = max_iso_payload / max_tx_pdu_len; ic->ic_hw_isomax = max_iso_pdus * ic->ic_max_send_data_segment_length; } else max_iso_pdus = 1; toep->params.ulp_mode = ULP_MODE_ISCSI; toep->ulpcb = icc; send_iscsi_flowc_wr(icc->sc, toep, roundup(max_iso_pdus * max_tx_pdu_len, tp->t_maxseg)); set_ulp_mode_iscsi(icc->sc, toep, icc->ulp_submode); INP_WUNLOCK(inp); error = kthread_add(icl_cxgbei_tx_main, icc, NULL, &icc->tx_thread, 0, 0, "%stx (cxgbei)", ic->ic_name); if (error != 0) goto out; error = kthread_add(icl_cxgbei_rx_main, icc, NULL, &icc->rx_thread, 0, 0, "%srx (cxgbei)", ic->ic_name); if (error != 0) goto out; error = icl_cxgbei_setsockopt(ic, so, max_tx_pdu_len, max_rx_pdu_len); out: if (error != 0) icl_cxgbei_conn_close(ic); return (error); } void icl_cxgbei_conn_close(struct icl_conn *ic) { struct icl_cxgbei_conn *icc = ic_to_icc(ic); struct icl_pdu *ip; struct socket *so; struct sockbuf *sb; struct inpcb *inp; struct toepcb *toep = icc->toep; MPASS(icc->icc_signature == CXGBEI_CONN_SIGNATURE); ICL_CONN_LOCK_ASSERT_NOT(ic); ICL_CONN_LOCK(ic); so = ic->ic_socket; if (ic->ic_disconnecting || so == NULL) { CTR4(KTR_CXGBE, "%s: icc %p (disconnecting = %d), so %p", __func__, icc, ic->ic_disconnecting, so); ICL_CONN_UNLOCK(ic); return; } ic->ic_disconnecting = true; #ifdef DIAGNOSTIC KASSERT(ic->ic_outstanding_pdus == 0, ("destroying session with %d outstanding PDUs", ic->ic_outstanding_pdus)); #endif CTR3(KTR_CXGBE, "%s: tid %d, icc %p", __func__, toep ? toep->tid : -1, icc); /* * Wait for the transmit thread to stop processing * this connection. */ if (icc->tx_thread != NULL) { wakeup(&icc->tx_active); mtx_sleep(icc->tx_thread, ic->ic_lock, 0, "conclo", 0); } /* Discard PDUs queued for TX. */ while (!STAILQ_EMPTY(&icc->sent_pdus)) { ip = STAILQ_FIRST(&icc->sent_pdus); STAILQ_REMOVE_HEAD(&icc->sent_pdus, ip_next); icl_cxgbei_pdu_done(ip, ENOTCONN); } ICL_CONN_UNLOCK(ic); inp = sotoinpcb(so); sb = &so->so_rcv; /* * Wait for the receive thread to stop processing this * connection. */ SOCKBUF_LOCK(sb); if (icc->rx_thread != NULL) { icc->rx_exiting = true; wakeup(&icc->rx_active); mtx_sleep(icc->rx_thread, SOCKBUF_MTX(sb), 0, "conclo", 0); } /* * Discard received PDUs not passed to the iSCSI layer. */ while (!STAILQ_EMPTY(&icc->rcvd_pdus)) { ip = STAILQ_FIRST(&icc->rcvd_pdus); STAILQ_REMOVE_HEAD(&icc->rcvd_pdus, ip_next); icl_cxgbei_pdu_done(ip, ENOTCONN); } SOCKBUF_UNLOCK(sb); INP_WLOCK(inp); if (toep != NULL) { /* NULL if connection was never offloaded. */ toep->ulpcb = NULL; /* Discard mbufs queued for TX. */ mbufq_drain(&toep->ulp_pduq); /* * Grab a reference to use when waiting for the final * CPL to be received. If toep->inp is NULL, then * final_cpl_received() has already been called (e.g. * due to the peer sending a RST). */ if (toep->inp != NULL) { toep = hold_toepcb(toep); toep->flags |= TPF_WAITING_FOR_FINAL; } else toep = NULL; } INP_WUNLOCK(inp); ICL_CONN_LOCK(ic); ic->ic_socket = NULL; ICL_CONN_UNLOCK(ic); /* * XXXNP: we should send RST instead of FIN when PDUs held in various * queues were purged instead of delivered reliably but soabort isn't * really general purpose and wouldn't do the right thing here. */ soclose(so); /* * Wait for the socket to fully close. This ensures any * pending received data has been received (and in particular, * any data that would be received by DDP has been handled). * Callers assume that it is safe to free buffers for tasks * and transfers after this function returns. */ if (toep != NULL) { struct mtx *lock = mtx_pool_find(mtxpool_sleep, toep); mtx_lock(lock); while ((toep->flags & TPF_WAITING_FOR_FINAL) != 0) mtx_sleep(toep, lock, PSOCK, "conclo2", 0); mtx_unlock(lock); free_toepcb(toep); } } static void cxgbei_insert_cmp(struct icl_cxgbei_conn *icc, struct cxgbei_cmp *cmp, uint32_t tt) { #ifdef INVARIANTS struct cxgbei_cmp *cmp2; #endif cmp->tt = tt; mtx_lock(&icc->cmp_lock); #ifdef INVARIANTS LIST_FOREACH(cmp2, &icc->cmp_table[TT_HASH(icc, tt)], link) { KASSERT(cmp2->tt != tt, ("%s: duplicate cmp", __func__)); } #endif LIST_INSERT_HEAD(&icc->cmp_table[TT_HASH(icc, tt)], cmp, link); mtx_unlock(&icc->cmp_lock); } struct cxgbei_cmp * cxgbei_find_cmp(struct icl_cxgbei_conn *icc, uint32_t tt) { struct cxgbei_cmp *cmp; mtx_lock(&icc->cmp_lock); LIST_FOREACH(cmp, &icc->cmp_table[TT_HASH(icc, tt)], link) { if (cmp->tt == tt) break; } mtx_unlock(&icc->cmp_lock); return (cmp); } static void cxgbei_rm_cmp(struct icl_cxgbei_conn *icc, struct cxgbei_cmp *cmp) { #ifdef INVARIANTS struct cxgbei_cmp *cmp2; #endif mtx_lock(&icc->cmp_lock); #ifdef INVARIANTS LIST_FOREACH(cmp2, &icc->cmp_table[TT_HASH(icc, cmp->tt)], link) { if (cmp2 == cmp) goto found; } panic("%s: could not find cmp", __func__); found: #endif LIST_REMOVE(cmp, link); mtx_unlock(&icc->cmp_lock); } int icl_cxgbei_conn_task_setup(struct icl_conn *ic, struct icl_pdu *ip, struct ccb_scsiio *csio, uint32_t *ittp, void **arg) { struct icl_cxgbei_conn *icc = ic_to_icc(ic); struct toepcb *toep = icc->toep; struct adapter *sc = icc->sc; struct cxgbei_data *ci = sc->iscsi_ulp_softc; struct ppod_region *pr = &ci->pr; struct cxgbei_ddp_state *ddp; struct ppod_reservation *prsv; struct inpcb *inp; struct mbufq mq; uint32_t itt; int rc = 0; ICL_CONN_LOCK_ASSERT(ic); /* This is for the offload driver's state. Must not be set already. */ MPASS(arg != NULL); MPASS(*arg == NULL); if ((csio->ccb_h.flags & CAM_DIR_MASK) != CAM_DIR_IN || csio->dxfer_len < ci->ddp_threshold || ic->ic_disconnecting || ic->ic_socket == NULL) { no_ddp: /* * No DDP for this I/O. Allocate an ITT (based on the one * passed in) that cannot be a valid hardware DDP tag in the * iSCSI region. */ itt = *ittp & M_PPOD_TAG; itt = V_PPOD_TAG(itt) | pr->pr_invalid_bit; *ittp = htobe32(itt); MPASS(*arg == NULL); /* State is maintained for DDP only. */ if (rc != 0) counter_u64_add( toep->ofld_rxq->rx_iscsi_ddp_setup_error, 1); return (0); } /* * Reserve resources for DDP, update the itt that should be used in the * PDU, and save DDP specific state for this I/O in *arg. */ ddp = malloc(sizeof(*ddp), M_CXGBEI, M_NOWAIT | M_ZERO); if (ddp == NULL) { rc = ENOMEM; goto no_ddp; } prsv = &ddp->prsv; mbufq_init(&mq, INT_MAX); switch (csio->ccb_h.flags & CAM_DATA_MASK) { case CAM_DATA_BIO: rc = t4_alloc_page_pods_for_bio(pr, (struct bio *)csio->data_ptr, prsv); if (rc != 0) { free(ddp, M_CXGBEI); goto no_ddp; } rc = t4_write_page_pods_for_bio(sc, toep, prsv, (struct bio *)csio->data_ptr, &mq); if (__predict_false(rc != 0)) { mbufq_drain(&mq); t4_free_page_pods(prsv); free(ddp, M_CXGBEI); goto no_ddp; } break; case CAM_DATA_VADDR: rc = t4_alloc_page_pods_for_buf(pr, (vm_offset_t)csio->data_ptr, csio->dxfer_len, prsv); if (rc != 0) { free(ddp, M_CXGBEI); goto no_ddp; } rc = t4_write_page_pods_for_buf(sc, toep, prsv, (vm_offset_t)csio->data_ptr, csio->dxfer_len, &mq); if (__predict_false(rc != 0)) { mbufq_drain(&mq); t4_free_page_pods(prsv); free(ddp, M_CXGBEI); goto no_ddp; } break; default: free(ddp, M_CXGBEI); rc = EINVAL; goto no_ddp; } /* * Do not get inp from toep->inp as the toepcb might have * detached already. */ inp = sotoinpcb(ic->ic_socket); INP_WLOCK(inp); - if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) != 0) { + if ((inp->inp_flags & INP_DROPPED) != 0) { INP_WUNLOCK(inp); mbufq_drain(&mq); t4_free_page_pods(prsv); free(ddp, M_CXGBEI); goto no_ddp; } mbufq_concat(&toep->ulp_pduq, &mq); INP_WUNLOCK(inp); ddp->cmp.last_datasn = -1; cxgbei_insert_cmp(icc, &ddp->cmp, prsv->prsv_tag); *ittp = htobe32(prsv->prsv_tag); *arg = prsv; counter_u64_add(toep->ofld_rxq->rx_iscsi_ddp_setup_ok, 1); return (0); } void icl_cxgbei_conn_task_done(struct icl_conn *ic, void *arg) { if (arg != NULL) { struct cxgbei_ddp_state *ddp = arg; cxgbei_rm_cmp(ic_to_icc(ic), &ddp->cmp); t4_free_page_pods(&ddp->prsv); free(ddp, M_CXGBEI); } } static inline bool ddp_sgl_check(struct ctl_sg_entry *sg, int entries, int xferlen) { #ifdef INVARIANTS int total_len = 0; #endif MPASS(entries > 0); if (((vm_offset_t)sg[--entries].addr & 3U) != 0) return (false); #ifdef INVARIANTS total_len += sg[entries].len; #endif while (--entries >= 0) { if (((vm_offset_t)sg[entries].addr & PAGE_MASK) != 0 || (sg[entries].len % PAGE_SIZE) != 0) return (false); #ifdef INVARIANTS total_len += sg[entries].len; #endif } MPASS(total_len == xferlen); return (true); } #define io_to_ddp_state(io) ((io)->io_hdr.ctl_private[CTL_PRIV_FRONTEND2].ptr) int icl_cxgbei_conn_transfer_setup(struct icl_conn *ic, struct icl_pdu *ip, union ctl_io *io, uint32_t *tttp, void **arg) { struct icl_cxgbei_conn *icc = ic_to_icc(ic); struct toepcb *toep = icc->toep; struct ctl_scsiio *ctsio = &io->scsiio; struct adapter *sc = icc->sc; struct cxgbei_data *ci = sc->iscsi_ulp_softc; struct ppod_region *pr = &ci->pr; struct cxgbei_ddp_state *ddp; struct ppod_reservation *prsv; struct ctl_sg_entry *sgl, sg_entry; struct inpcb *inp; struct mbufq mq; int sg_entries = ctsio->kern_sg_entries; uint32_t ttt; int xferlen, rc = 0, alias; /* This is for the offload driver's state. Must not be set already. */ MPASS(arg != NULL); MPASS(*arg == NULL); if (ctsio->ext_data_filled == 0) { int first_burst; #ifdef INVARIANTS struct icl_cxgbei_pdu *icp = ip_to_icp(ip); MPASS(icp->icp_signature == CXGBEI_PDU_SIGNATURE); MPASS(ic == ip->ip_conn); MPASS(ip->ip_bhs_mbuf != NULL); #endif first_burst = icl_pdu_data_segment_length(ip); /* * Note that ICL calls conn_transfer_setup even if the first * burst had everything and there's nothing left to transfer. * * NB: The CTL frontend might have provided a buffer * whose length (kern_data_len) is smaller than the * FirstBurstLength of unsolicited data. Treat those * as an empty transfer. */ xferlen = ctsio->kern_data_len; if (xferlen < first_burst || xferlen - first_burst < ci->ddp_threshold) { no_ddp: /* * No DDP for this transfer. Allocate a TTT (based on * the one passed in) that cannot be a valid hardware * DDP tag in the iSCSI region. */ ttt = *tttp & M_PPOD_TAG; ttt = V_PPOD_TAG(ttt) | pr->pr_invalid_bit; *tttp = htobe32(ttt); MPASS(io_to_ddp_state(io) == NULL); if (rc != 0) counter_u64_add( toep->ofld_rxq->rx_iscsi_ddp_setup_error, 1); return (0); } if (sg_entries == 0) { sgl = &sg_entry; sgl->len = xferlen; sgl->addr = (void *)ctsio->kern_data_ptr; sg_entries = 1; } else sgl = (void *)ctsio->kern_data_ptr; if (!ddp_sgl_check(sgl, sg_entries, xferlen)) goto no_ddp; /* * Reserve resources for DDP, update the ttt that should be used * in the PDU, and save DDP specific state for this I/O. */ MPASS(io_to_ddp_state(io) == NULL); ddp = malloc(sizeof(*ddp), M_CXGBEI, M_NOWAIT | M_ZERO); if (ddp == NULL) { rc = ENOMEM; goto no_ddp; } prsv = &ddp->prsv; rc = t4_alloc_page_pods_for_sgl(pr, sgl, sg_entries, prsv); if (rc != 0) { free(ddp, M_CXGBEI); goto no_ddp; } mbufq_init(&mq, INT_MAX); rc = t4_write_page_pods_for_sgl(sc, toep, prsv, sgl, sg_entries, xferlen, &mq); if (__predict_false(rc != 0)) { mbufq_drain(&mq); t4_free_page_pods(prsv); free(ddp, M_CXGBEI); goto no_ddp; } /* * Do not get inp from toep->inp as the toepcb might * have detached already. */ ICL_CONN_LOCK(ic); if (ic->ic_disconnecting || ic->ic_socket == NULL) { ICL_CONN_UNLOCK(ic); mbufq_drain(&mq); t4_free_page_pods(prsv); free(ddp, M_CXGBEI); return (ECONNRESET); } inp = sotoinpcb(ic->ic_socket); INP_WLOCK(inp); ICL_CONN_UNLOCK(ic); - if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) != 0) { + if ((inp->inp_flags & INP_DROPPED) != 0) { INP_WUNLOCK(inp); mbufq_drain(&mq); t4_free_page_pods(prsv); free(ddp, M_CXGBEI); return (ECONNRESET); } mbufq_concat(&toep->ulp_pduq, &mq); INP_WUNLOCK(inp); ddp->cmp.next_buffer_offset = ctsio->kern_rel_offset + first_burst; ddp->cmp.last_datasn = -1; cxgbei_insert_cmp(icc, &ddp->cmp, prsv->prsv_tag); *tttp = htobe32(prsv->prsv_tag); io_to_ddp_state(io) = ddp; *arg = ctsio; counter_u64_add(toep->ofld_rxq->rx_iscsi_ddp_setup_ok, 1); return (0); } /* * In the middle of an I/O. A non-NULL page pod reservation indicates * that a DDP buffer is being used for the I/O. */ ddp = io_to_ddp_state(ctsio); if (ddp == NULL) goto no_ddp; prsv = &ddp->prsv; alias = (prsv->prsv_tag & pr->pr_alias_mask) >> pr->pr_alias_shift; alias++; prsv->prsv_tag &= ~pr->pr_alias_mask; prsv->prsv_tag |= alias << pr->pr_alias_shift & pr->pr_alias_mask; ddp->cmp.last_datasn = -1; cxgbei_insert_cmp(icc, &ddp->cmp, prsv->prsv_tag); *tttp = htobe32(prsv->prsv_tag); *arg = ctsio; return (0); } void icl_cxgbei_conn_transfer_done(struct icl_conn *ic, void *arg) { struct ctl_scsiio *ctsio = arg; if (ctsio != NULL) { struct cxgbei_ddp_state *ddp; ddp = io_to_ddp_state(ctsio); MPASS(ddp != NULL); cxgbei_rm_cmp(ic_to_icc(ic), &ddp->cmp); if (ctsio->kern_data_len == ctsio->ext_data_filled || ic->ic_disconnecting) { t4_free_page_pods(&ddp->prsv); free(ddp, M_CXGBEI); io_to_ddp_state(ctsio) = NULL; } } } #ifdef COMPAT_FREEBSD13 static void cxgbei_limits(struct adapter *sc, void *arg) { struct icl_drv_limits *idl = arg; struct cxgbei_data *ci; int max_dsl; if (begin_synchronized_op(sc, NULL, HOLD_LOCK, "t4lims") != 0) return; if (uld_active(sc, ULD_ISCSI)) { ci = sc->iscsi_ulp_softc; MPASS(ci != NULL); max_dsl = ci->max_rx_data_len; if (idl->idl_max_recv_data_segment_length > max_dsl) idl->idl_max_recv_data_segment_length = max_dsl; max_dsl = ci->max_tx_data_len; if (idl->idl_max_send_data_segment_length > max_dsl) idl->idl_max_send_data_segment_length = max_dsl; } end_synchronized_op(sc, LOCK_HELD); } #endif static int cxgbei_limits_fd(struct icl_drv_limits *idl, int fd) { struct find_ofld_adapter_rr fa; struct file *fp; struct socket *so; struct adapter *sc; struct cxgbei_data *ci; cap_rights_t rights; int error; error = fget(curthread, fd, cap_rights_init_one(&rights, CAP_SOCK_CLIENT), &fp); if (error != 0) return (error); if (fp->f_type != DTYPE_SOCKET) { fdrop(fp, curthread); return (EINVAL); } so = fp->f_data; if (so->so_type != SOCK_STREAM || so->so_proto->pr_protocol != IPPROTO_TCP) { fdrop(fp, curthread); return (EINVAL); } /* Find the adapter offloading this socket. */ fa.sc = NULL; fa.so = so; t4_iterate(find_offload_adapter, &fa); if (fa.sc == NULL) { fdrop(fp, curthread); return (ENXIO); } fdrop(fp, curthread); sc = fa.sc; error = begin_synchronized_op(sc, NULL, HOLD_LOCK, "t4lims"); if (error != 0) return (error); if (uld_active(sc, ULD_ISCSI)) { ci = sc->iscsi_ulp_softc; MPASS(ci != NULL); idl->idl_max_recv_data_segment_length = ci->max_rx_data_len; idl->idl_max_send_data_segment_length = ci->max_tx_data_len; } else error = ENXIO; end_synchronized_op(sc, LOCK_HELD); return (error); } static int icl_cxgbei_limits(struct icl_drv_limits *idl, int socket) { /* Maximum allowed by the RFC. cxgbei_limits will clip them. */ idl->idl_max_recv_data_segment_length = (1 << 24) - 1; idl->idl_max_send_data_segment_length = (1 << 24) - 1; /* These are somewhat arbitrary. */ idl->idl_max_burst_length = max_burst_length; idl->idl_first_burst_length = first_burst_length; #ifdef COMPAT_FREEBSD13 if (socket == 0) { t4_iterate(cxgbei_limits, idl); return (0); } #endif return (cxgbei_limits_fd(idl, socket)); } int icl_cxgbei_mod_load(void) { int rc; refcount_init(&icl_cxgbei_ncons, 0); rc = icl_register("cxgbei", false, -100, icl_cxgbei_limits, icl_cxgbei_new_conn); return (rc); } int icl_cxgbei_mod_unload(void) { if (icl_cxgbei_ncons != 0) return (EBUSY); icl_unregister("cxgbei", false); return (0); } #endif diff --git a/sys/dev/cxgbe/iw_cxgbe/qp.c b/sys/dev/cxgbe/iw_cxgbe/qp.c index b2901f93988e..d3d4c0573a6f 100644 --- a/sys/dev/cxgbe/iw_cxgbe/qp.c +++ b/sys/dev/cxgbe/iw_cxgbe/qp.c @@ -1,1971 +1,1971 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2009-2013 Chelsio, Inc. All rights reserved. * * This software is available to you under a choice of one of two * licenses. You may choose to be licensed under the terms of the GNU * General Public License (GPL) Version 2, available from the file * COPYING in the main directory of this source tree, or the * OpenIB.org BSD license below: * * Redistribution and use in source and binary forms, with or * without modification, are permitted provided that the following * conditions are met: * * - Redistributions of source code must retain the above * copyright notice, this list of conditions and the following * disclaimer. * * - Redistributions in binary form must reproduce the above * copyright notice, this list of conditions and the following * disclaimer in the documentation and/or other materials * provided with the distribution. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #ifdef TCP_OFFLOAD #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include struct sge_iq; struct rss_header; struct cpl_set_tcb_rpl; #include #include "offload.h" #include "tom/t4_tom.h" #include "iw_cxgbe.h" #include "user.h" static int creds(struct toepcb *toep, struct inpcb *inp, size_t wrsize); static int max_fr_immd = T4_MAX_FR_IMMD;//SYSCTL parameter later... static int alloc_ird(struct c4iw_dev *dev, u32 ird) { int ret = 0; spin_lock_irq(&dev->lock); if (ird <= dev->avail_ird) dev->avail_ird -= ird; else ret = -ENOMEM; spin_unlock_irq(&dev->lock); if (ret) log(LOG_WARNING, "%s: device IRD resources exhausted\n", device_get_nameunit(dev->rdev.adap->dev)); return ret; } static void free_ird(struct c4iw_dev *dev, int ird) { spin_lock_irq(&dev->lock); dev->avail_ird += ird; spin_unlock_irq(&dev->lock); } static void set_state(struct c4iw_qp *qhp, enum c4iw_qp_state state) { unsigned long flag; spin_lock_irqsave(&qhp->lock, flag); qhp->attr.state = state; spin_unlock_irqrestore(&qhp->lock, flag); } static int destroy_qp(struct c4iw_rdev *rdev, struct t4_wq *wq, struct c4iw_dev_ucontext *uctx) { struct c4iw_dev *rhp = rdev_to_c4iw_dev(rdev); /* * uP clears EQ contexts when the connection exits rdma mode, * so no need to post a RESET WR for these EQs. */ dma_free_coherent(rhp->ibdev.dma_device, wq->rq.memsize, wq->rq.queue, dma_unmap_addr(&wq->rq, mapping)); dma_free_coherent(rhp->ibdev.dma_device, wq->sq.memsize, wq->sq.queue, dma_unmap_addr(&wq->sq, mapping)); c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size); kfree(wq->rq.sw_rq); kfree(wq->sq.sw_sq); c4iw_put_qpid(rdev, wq->rq.qid, uctx); c4iw_put_qpid(rdev, wq->sq.qid, uctx); return 0; } static int create_qp(struct c4iw_rdev *rdev, struct t4_wq *wq, struct t4_cq *rcq, struct t4_cq *scq, struct c4iw_dev_ucontext *uctx) { struct adapter *sc = rdev->adap; struct c4iw_dev *rhp = rdev_to_c4iw_dev(rdev); int user = (uctx != &rdev->uctx); struct fw_ri_res_wr *res_wr; struct fw_ri_res *res; int wr_len; struct c4iw_wr_wait wr_wait; int ret = 0; int eqsize; struct wrqe *wr; u64 sq_bar2_qoffset = 0, rq_bar2_qoffset = 0; wq->sq.qid = c4iw_get_qpid(rdev, uctx); if (!wq->sq.qid) return -ENOMEM; wq->rq.qid = c4iw_get_qpid(rdev, uctx); if (!wq->rq.qid) { ret = -ENOMEM; goto free_sq_qid; } if (!user) { wq->sq.sw_sq = kzalloc(wq->sq.size * sizeof *wq->sq.sw_sq, GFP_KERNEL); if (!wq->sq.sw_sq) { ret = -ENOMEM; goto free_rq_qid; } wq->rq.sw_rq = kzalloc(wq->rq.size * sizeof *wq->rq.sw_rq, GFP_KERNEL); if (!wq->rq.sw_rq) { ret = -ENOMEM; goto free_sw_sq; } } /* * RQT must be a power of 2 and at least 16 deep. */ wq->rq.rqt_size = roundup_pow_of_two(max_t(u16, wq->rq.size, 16)); wq->rq.rqt_hwaddr = c4iw_rqtpool_alloc(rdev, wq->rq.rqt_size); if (!wq->rq.rqt_hwaddr) { ret = -ENOMEM; goto free_sw_rq; } /*QP memory, allocate DMAable memory for Send & Receive Queues */ wq->sq.queue = dma_alloc_coherent(rhp->ibdev.dma_device, wq->sq.memsize, &(wq->sq.dma_addr), GFP_KERNEL); if (!wq->sq.queue) { ret = -ENOMEM; goto free_hwaddr; } wq->sq.phys_addr = vtophys(wq->sq.queue); dma_unmap_addr_set(&wq->sq, mapping, wq->sq.dma_addr); memset(wq->sq.queue, 0, wq->sq.memsize); wq->rq.queue = dma_alloc_coherent(rhp->ibdev.dma_device, wq->rq.memsize, &(wq->rq.dma_addr), GFP_KERNEL); if (!wq->rq.queue) { ret = -ENOMEM; goto free_sq_dma; } wq->rq.phys_addr = vtophys(wq->rq.queue); dma_unmap_addr_set(&wq->rq, mapping, wq->rq.dma_addr); memset(wq->rq.queue, 0, wq->rq.memsize); CTR5(KTR_IW_CXGBE, "%s QP sq base va 0x%p pa 0x%llx rq base va 0x%p pa 0x%llx", __func__, wq->sq.queue, (unsigned long long)wq->sq.phys_addr, wq->rq.queue, (unsigned long long)wq->rq.phys_addr); /* Doorbell/WC regions, determine the BAR2 queue offset and qid. */ t4_bar2_sge_qregs(rdev->adap, wq->sq.qid, T4_BAR2_QTYPE_EGRESS, user, &sq_bar2_qoffset, &wq->sq.bar2_qid); t4_bar2_sge_qregs(rdev->adap, wq->rq.qid, T4_BAR2_QTYPE_EGRESS, user, &rq_bar2_qoffset, &wq->rq.bar2_qid); if (user) { /* Compute BAR2 DB/WC physical address(page-aligned) for * Userspace mapping. */ wq->sq.bar2_pa = (rdev->bar2_pa + sq_bar2_qoffset) & PAGE_MASK; wq->rq.bar2_pa = (rdev->bar2_pa + rq_bar2_qoffset) & PAGE_MASK; CTR3(KTR_IW_CXGBE, "%s BAR2 DB/WC sq base pa 0x%llx rq base pa 0x%llx", __func__, (unsigned long long)wq->sq.bar2_pa, (unsigned long long)wq->rq.bar2_pa); } else { /* Compute BAR2 DB/WC virtual address to access in kernel. */ wq->sq.bar2_va = (void __iomem *)((u64)rdev->bar2_kva + sq_bar2_qoffset); wq->rq.bar2_va = (void __iomem *)((u64)rdev->bar2_kva + rq_bar2_qoffset); CTR3(KTR_IW_CXGBE, "%s BAR2 DB/WC sq base va %p rq base va %p", __func__, (unsigned long long)wq->sq.bar2_va, (unsigned long long)wq->rq.bar2_va); } wq->rdev = rdev; wq->rq.msn = 1; /* build fw_ri_res_wr */ wr_len = sizeof *res_wr + 2 * sizeof *res; wr = alloc_wrqe(wr_len, &sc->sge.ctrlq[0]); if (wr == NULL) { ret = -ENOMEM; goto free_rq_dma; } res_wr = wrtod(wr); memset(res_wr, 0, wr_len); res_wr->op_nres = cpu_to_be32( V_FW_WR_OP(FW_RI_RES_WR) | V_FW_RI_RES_WR_NRES(2) | F_FW_WR_COMPL); res_wr->len16_pkd = cpu_to_be32(DIV_ROUND_UP(wr_len, 16)); res_wr->cookie = (unsigned long) &wr_wait; res = res_wr->res; res->u.sqrq.restype = FW_RI_RES_TYPE_SQ; res->u.sqrq.op = FW_RI_RES_OP_WRITE; /* eqsize is the number of 64B entries plus the status page size. */ eqsize = wq->sq.size * T4_SQ_NUM_SLOTS + rdev->hw_queue.t4_eq_status_entries; res->u.sqrq.fetchszm_to_iqid = cpu_to_be32( V_FW_RI_RES_WR_HOSTFCMODE(0) | /* no host cidx updates */ V_FW_RI_RES_WR_CPRIO(0) | /* don't keep in chip cache */ V_FW_RI_RES_WR_PCIECHN(0) | /* set by uP at ri_init time */ V_FW_RI_RES_WR_IQID(scq->cqid)); res->u.sqrq.dcaen_to_eqsize = cpu_to_be32( V_FW_RI_RES_WR_DCAEN(0) | V_FW_RI_RES_WR_DCACPU(0) | V_FW_RI_RES_WR_FBMIN(chip_id(sc) <= CHELSIO_T5 ? X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) | V_FW_RI_RES_WR_FBMAX(3) | V_FW_RI_RES_WR_CIDXFTHRESHO(0) | V_FW_RI_RES_WR_CIDXFTHRESH(0) | V_FW_RI_RES_WR_EQSIZE(eqsize)); res->u.sqrq.eqid = cpu_to_be32(wq->sq.qid); res->u.sqrq.eqaddr = cpu_to_be64(wq->sq.dma_addr); res++; res->u.sqrq.restype = FW_RI_RES_TYPE_RQ; res->u.sqrq.op = FW_RI_RES_OP_WRITE; /* eqsize is the number of 64B entries plus the status page size. */ eqsize = wq->rq.size * T4_RQ_NUM_SLOTS + rdev->hw_queue.t4_eq_status_entries; res->u.sqrq.fetchszm_to_iqid = cpu_to_be32( V_FW_RI_RES_WR_HOSTFCMODE(0) | /* no host cidx updates */ V_FW_RI_RES_WR_CPRIO(0) | /* don't keep in chip cache */ V_FW_RI_RES_WR_PCIECHN(0) | /* set by uP at ri_init time */ V_FW_RI_RES_WR_IQID(rcq->cqid)); res->u.sqrq.dcaen_to_eqsize = cpu_to_be32( V_FW_RI_RES_WR_DCAEN(0) | V_FW_RI_RES_WR_DCACPU(0) | V_FW_RI_RES_WR_FBMIN(chip_id(sc) <= CHELSIO_T5 ? X_FETCHBURSTMIN_64B : X_FETCHBURSTMIN_64B_T6) | V_FW_RI_RES_WR_FBMAX(3) | V_FW_RI_RES_WR_CIDXFTHRESHO(0) | V_FW_RI_RES_WR_CIDXFTHRESH(0) | V_FW_RI_RES_WR_EQSIZE(eqsize)); res->u.sqrq.eqid = cpu_to_be32(wq->rq.qid); res->u.sqrq.eqaddr = cpu_to_be64(wq->rq.dma_addr); c4iw_init_wr_wait(&wr_wait); t4_wrq_tx(sc, wr); ret = c4iw_wait_for_reply(rdev, &wr_wait, 0, wq->sq.qid, NULL, __func__); if (ret) goto free_rq_dma; CTR5(KTR_IW_CXGBE, "%s sqid 0x%x rqid 0x%x kdb 0x%p squdb 0x%llx rqudb 0x%llx", __func__, wq->sq.qid, wq->rq.qid, (unsigned long long)wq->sq.bar2_va, (unsigned long long)wq->rq.bar2_va); return 0; free_rq_dma: dma_free_coherent(rhp->ibdev.dma_device, wq->rq.memsize, wq->rq.queue, dma_unmap_addr(&wq->rq, mapping)); free_sq_dma: dma_free_coherent(rhp->ibdev.dma_device, wq->sq.memsize, wq->sq.queue, dma_unmap_addr(&wq->sq, mapping)); free_hwaddr: c4iw_rqtpool_free(rdev, wq->rq.rqt_hwaddr, wq->rq.rqt_size); free_sw_rq: kfree(wq->rq.sw_rq); free_sw_sq: kfree(wq->sq.sw_sq); free_rq_qid: c4iw_put_qpid(rdev, wq->rq.qid, uctx); free_sq_qid: c4iw_put_qpid(rdev, wq->sq.qid, uctx); return ret; } static int build_immd(struct t4_sq *sq, struct fw_ri_immd *immdp, const struct ib_send_wr *wr, int max, u32 *plenp) { u8 *dstp, *srcp; u32 plen = 0; int i; int rem, len; dstp = (u8 *)immdp->data; for (i = 0; i < wr->num_sge; i++) { if ((plen + wr->sg_list[i].length) > max) return -EMSGSIZE; srcp = (u8 *)(unsigned long)wr->sg_list[i].addr; plen += wr->sg_list[i].length; rem = wr->sg_list[i].length; while (rem) { if (dstp == (u8 *)&sq->queue[sq->size]) dstp = (u8 *)sq->queue; if (rem <= (u8 *)&sq->queue[sq->size] - dstp) len = rem; else len = (u8 *)&sq->queue[sq->size] - dstp; memcpy(dstp, srcp, len); dstp += len; srcp += len; rem -= len; } } len = roundup(plen + sizeof *immdp, 16) - (plen + sizeof *immdp); if (len) memset(dstp, 0, len); immdp->op = FW_RI_DATA_IMMD; immdp->r1 = 0; immdp->r2 = 0; immdp->immdlen = cpu_to_be32(plen); *plenp = plen; return 0; } static int build_isgl(__be64 *queue_start, __be64 *queue_end, struct fw_ri_isgl *isglp, struct ib_sge *sg_list, int num_sge, u32 *plenp) { int i; u32 plen = 0; __be64 *flitp = (__be64 *)isglp->sge; for (i = 0; i < num_sge; i++) { if ((plen + sg_list[i].length) < plen) return -EMSGSIZE; plen += sg_list[i].length; *flitp = cpu_to_be64(((u64)sg_list[i].lkey << 32) | sg_list[i].length); if (++flitp == queue_end) flitp = queue_start; *flitp = cpu_to_be64(sg_list[i].addr); if (++flitp == queue_end) flitp = queue_start; } *flitp = (__force __be64)0; isglp->op = FW_RI_DATA_ISGL; isglp->r1 = 0; isglp->nsge = cpu_to_be16(num_sge); isglp->r2 = 0; if (plenp) *plenp = plen; return 0; } static int build_rdma_send(struct t4_sq *sq, union t4_wr *wqe, const struct ib_send_wr *wr, u8 *len16) { u32 plen; int size; int ret; if (wr->num_sge > T4_MAX_SEND_SGE) return -EINVAL; switch (wr->opcode) { case IB_WR_SEND: if (wr->send_flags & IB_SEND_SOLICITED) wqe->send.sendop_pkd = cpu_to_be32( V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND_WITH_SE)); else wqe->send.sendop_pkd = cpu_to_be32( V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND)); wqe->send.stag_inv = 0; break; case IB_WR_SEND_WITH_INV: if (wr->send_flags & IB_SEND_SOLICITED) wqe->send.sendop_pkd = cpu_to_be32( V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND_WITH_SE_INV)); else wqe->send.sendop_pkd = cpu_to_be32( V_FW_RI_SEND_WR_SENDOP(FW_RI_SEND_WITH_INV)); wqe->send.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey); break; default: return -EINVAL; } wqe->send.r3 = 0; wqe->send.r4 = 0; plen = 0; if (wr->num_sge) { if (wr->send_flags & IB_SEND_INLINE) { ret = build_immd(sq, wqe->send.u.immd_src, wr, T4_MAX_SEND_INLINE, &plen); if (ret) return ret; size = sizeof wqe->send + sizeof(struct fw_ri_immd) + plen; } else { ret = build_isgl((__be64 *)sq->queue, (__be64 *)&sq->queue[sq->size], wqe->send.u.isgl_src, wr->sg_list, wr->num_sge, &plen); if (ret) return ret; size = sizeof wqe->send + sizeof(struct fw_ri_isgl) + wr->num_sge * sizeof(struct fw_ri_sge); } } else { wqe->send.u.immd_src[0].op = FW_RI_DATA_IMMD; wqe->send.u.immd_src[0].r1 = 0; wqe->send.u.immd_src[0].r2 = 0; wqe->send.u.immd_src[0].immdlen = 0; size = sizeof wqe->send + sizeof(struct fw_ri_immd); plen = 0; } *len16 = DIV_ROUND_UP(size, 16); wqe->send.plen = cpu_to_be32(plen); return 0; } static int build_rdma_write(struct t4_sq *sq, union t4_wr *wqe, const struct ib_send_wr *wr, u8 *len16) { u32 plen; int size; int ret; if (wr->num_sge > T4_MAX_SEND_SGE) return -EINVAL; wqe->write.immd_data = 0; wqe->write.stag_sink = cpu_to_be32(rdma_wr(wr)->rkey); wqe->write.to_sink = cpu_to_be64(rdma_wr(wr)->remote_addr); if (wr->num_sge) { if (wr->send_flags & IB_SEND_INLINE) { ret = build_immd(sq, wqe->write.u.immd_src, wr, T4_MAX_WRITE_INLINE, &plen); if (ret) return ret; size = sizeof wqe->write + sizeof(struct fw_ri_immd) + plen; } else { ret = build_isgl((__be64 *)sq->queue, (__be64 *)&sq->queue[sq->size], wqe->write.u.isgl_src, wr->sg_list, wr->num_sge, &plen); if (ret) return ret; size = sizeof wqe->write + sizeof(struct fw_ri_isgl) + wr->num_sge * sizeof(struct fw_ri_sge); } } else { wqe->write.u.immd_src[0].op = FW_RI_DATA_IMMD; wqe->write.u.immd_src[0].r1 = 0; wqe->write.u.immd_src[0].r2 = 0; wqe->write.u.immd_src[0].immdlen = 0; size = sizeof wqe->write + sizeof(struct fw_ri_immd); plen = 0; } *len16 = DIV_ROUND_UP(size, 16); wqe->write.plen = cpu_to_be32(plen); return 0; } static int build_rdma_read(union t4_wr *wqe, const struct ib_send_wr *wr, u8 *len16) { if (wr->num_sge > 1) return -EINVAL; if (wr->num_sge && wr->sg_list[0].length) { wqe->read.stag_src = cpu_to_be32(rdma_wr(wr)->rkey); wqe->read.to_src_hi = cpu_to_be32((u32)(rdma_wr(wr)->remote_addr >> 32)); wqe->read.to_src_lo = cpu_to_be32((u32)rdma_wr(wr)->remote_addr); wqe->read.stag_sink = cpu_to_be32(wr->sg_list[0].lkey); wqe->read.plen = cpu_to_be32(wr->sg_list[0].length); wqe->read.to_sink_hi = cpu_to_be32((u32)(wr->sg_list[0].addr >> 32)); wqe->read.to_sink_lo = cpu_to_be32((u32)(wr->sg_list[0].addr)); } else { wqe->read.stag_src = cpu_to_be32(2); wqe->read.to_src_hi = 0; wqe->read.to_src_lo = 0; wqe->read.stag_sink = cpu_to_be32(2); wqe->read.plen = 0; wqe->read.to_sink_hi = 0; wqe->read.to_sink_lo = 0; } wqe->read.r2 = 0; wqe->read.r5 = 0; *len16 = DIV_ROUND_UP(sizeof wqe->read, 16); return 0; } static int build_rdma_recv(struct c4iw_qp *qhp, union t4_recv_wr *wqe, const struct ib_recv_wr *wr, u8 *len16) { int ret; ret = build_isgl((__be64 *)qhp->wq.rq.queue, (__be64 *)&qhp->wq.rq.queue[qhp->wq.rq.size], &wqe->recv.isgl, wr->sg_list, wr->num_sge, NULL); if (ret) return ret; *len16 = DIV_ROUND_UP(sizeof wqe->recv + wr->num_sge * sizeof(struct fw_ri_sge), 16); return 0; } static int build_inv_stag(union t4_wr *wqe, const struct ib_send_wr *wr, u8 *len16) { wqe->inv.stag_inv = cpu_to_be32(wr->ex.invalidate_rkey); wqe->inv.r2 = 0; *len16 = DIV_ROUND_UP(sizeof wqe->inv, 16); return 0; } static void free_qp_work(struct work_struct *work) { struct c4iw_ucontext *ucontext; struct c4iw_qp *qhp; struct c4iw_dev *rhp; qhp = container_of(work, struct c4iw_qp, free_work); ucontext = qhp->ucontext; rhp = qhp->rhp; CTR3(KTR_IW_CXGBE, "%s qhp %p ucontext %p", __func__, qhp, ucontext); destroy_qp(&rhp->rdev, &qhp->wq, ucontext ? &ucontext->uctx : &rhp->rdev.uctx); kfree(qhp); } static void queue_qp_free(struct kref *kref) { struct c4iw_qp *qhp; qhp = container_of(kref, struct c4iw_qp, kref); CTR2(KTR_IW_CXGBE, "%s qhp %p", __func__, qhp); queue_work(qhp->rhp->rdev.free_workq, &qhp->free_work); } void c4iw_qp_add_ref(struct ib_qp *qp) { CTR2(KTR_IW_CXGBE, "%s ib_qp %p", __func__, qp); kref_get(&to_c4iw_qp(qp)->kref); } void c4iw_qp_rem_ref(struct ib_qp *qp) { CTR2(KTR_IW_CXGBE, "%s ib_qp %p", __func__, qp); kref_put(&to_c4iw_qp(qp)->kref, queue_qp_free); } static void complete_sq_drain_wr(struct c4iw_qp *qhp, const struct ib_send_wr *wr) { struct t4_cqe cqe = {}; struct c4iw_cq *schp; unsigned long flag; struct t4_cq *cq; schp = to_c4iw_cq(qhp->ibqp.send_cq); cq = &schp->cq; PDBG("%s drain sq id %u\n", __func__, qhp->wq.sq.qid); cqe.u.drain_cookie = wr->wr_id; cqe.header = cpu_to_be32(V_CQE_STATUS(T4_ERR_SWFLUSH) | V_CQE_OPCODE(C4IW_DRAIN_OPCODE) | V_CQE_TYPE(1) | V_CQE_SWCQE(1) | V_CQE_QPID(qhp->wq.sq.qid)); spin_lock_irqsave(&schp->lock, flag); cqe.bits_type_ts = cpu_to_be64(V_CQE_GENBIT((u64)cq->gen)); cq->sw_queue[cq->sw_pidx] = cqe; t4_swcq_produce(cq); spin_unlock_irqrestore(&schp->lock, flag); spin_lock_irqsave(&schp->comp_handler_lock, flag); (*schp->ibcq.comp_handler)(&schp->ibcq, schp->ibcq.cq_context); spin_unlock_irqrestore(&schp->comp_handler_lock, flag); } static void complete_rq_drain_wr(struct c4iw_qp *qhp, const struct ib_recv_wr *wr) { struct t4_cqe cqe = {}; struct c4iw_cq *rchp; unsigned long flag; struct t4_cq *cq; rchp = to_c4iw_cq(qhp->ibqp.recv_cq); cq = &rchp->cq; PDBG("%s drain rq id %u\n", __func__, qhp->wq.sq.qid); cqe.u.drain_cookie = wr->wr_id; cqe.header = cpu_to_be32(V_CQE_STATUS(T4_ERR_SWFLUSH) | V_CQE_OPCODE(C4IW_DRAIN_OPCODE) | V_CQE_TYPE(0) | V_CQE_SWCQE(1) | V_CQE_QPID(qhp->wq.sq.qid)); spin_lock_irqsave(&rchp->lock, flag); cqe.bits_type_ts = cpu_to_be64(V_CQE_GENBIT((u64)cq->gen)); cq->sw_queue[cq->sw_pidx] = cqe; t4_swcq_produce(cq); spin_unlock_irqrestore(&rchp->lock, flag); spin_lock_irqsave(&rchp->comp_handler_lock, flag); (*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context); spin_unlock_irqrestore(&rchp->comp_handler_lock, flag); } static int build_tpte_memreg(struct fw_ri_fr_nsmr_tpte_wr *fr, const struct ib_reg_wr *wr, struct c4iw_mr *mhp, u8 *len16) { __be64 *p = (__be64 *)fr->pbl; if (wr->mr->page_size > C4IW_MAX_PAGE_SIZE) return -EINVAL; fr->r2 = cpu_to_be32(0); fr->stag = cpu_to_be32(mhp->ibmr.rkey); fr->tpte.valid_to_pdid = cpu_to_be32(F_FW_RI_TPTE_VALID | V_FW_RI_TPTE_STAGKEY((mhp->ibmr.rkey & M_FW_RI_TPTE_STAGKEY)) | V_FW_RI_TPTE_STAGSTATE(1) | V_FW_RI_TPTE_STAGTYPE(FW_RI_STAG_NSMR) | V_FW_RI_TPTE_PDID(mhp->attr.pdid)); fr->tpte.locread_to_qpid = cpu_to_be32( V_FW_RI_TPTE_PERM(c4iw_ib_to_tpt_access(wr->access)) | V_FW_RI_TPTE_ADDRTYPE(FW_RI_VA_BASED_TO) | V_FW_RI_TPTE_PS(ilog2(wr->mr->page_size) - 12)); fr->tpte.nosnoop_pbladdr = cpu_to_be32(V_FW_RI_TPTE_PBLADDR( PBL_OFF(&mhp->rhp->rdev, mhp->attr.pbl_addr)>>3)); fr->tpte.dca_mwbcnt_pstag = cpu_to_be32(0); fr->tpte.len_hi = cpu_to_be32(mhp->ibmr.length >> 32); fr->tpte.len_lo = cpu_to_be32(mhp->ibmr.length & 0xffffffff); fr->tpte.va_hi = cpu_to_be32(mhp->ibmr.iova >> 32); fr->tpte.va_lo_fbo = cpu_to_be32(mhp->ibmr.iova & 0xffffffff); p[0] = cpu_to_be64((u64)mhp->mpl[0]); p[1] = cpu_to_be64((u64)mhp->mpl[1]); *len16 = DIV_ROUND_UP(sizeof(*fr), 16); return 0; } static int build_memreg(struct t4_sq *sq, union t4_wr *wqe, const struct ib_reg_wr *wr, struct c4iw_mr *mhp, u8 *len16, bool dsgl_supported) { struct fw_ri_immd *imdp; __be64 *p; int i; int pbllen = roundup(mhp->mpl_len * sizeof(u64), 32); int rem; if (mhp->mpl_len > t4_max_fr_depth(&mhp->rhp->rdev, use_dsgl)) return -EINVAL; if (wr->mr->page_size > C4IW_MAX_PAGE_SIZE) return -EINVAL; wqe->fr.qpbinde_to_dcacpu = 0; wqe->fr.pgsz_shift = ilog2(wr->mr->page_size) - 12; wqe->fr.addr_type = FW_RI_VA_BASED_TO; wqe->fr.mem_perms = c4iw_ib_to_tpt_access(wr->access); wqe->fr.len_hi = cpu_to_be32(mhp->ibmr.length >> 32); wqe->fr.len_lo = cpu_to_be32(mhp->ibmr.length & 0xffffffff); wqe->fr.stag = cpu_to_be32(wr->key); wqe->fr.va_hi = cpu_to_be32(mhp->ibmr.iova >> 32); wqe->fr.va_lo_fbo = cpu_to_be32(mhp->ibmr.iova & 0xffffffff); if (dsgl_supported && use_dsgl && (pbllen > max_fr_immd)) { struct fw_ri_dsgl *sglp; for (i = 0; i < mhp->mpl_len; i++) mhp->mpl[i] = (__force u64)cpu_to_be64((u64)mhp->mpl[i]); sglp = (struct fw_ri_dsgl *)(&wqe->fr + 1); sglp->op = FW_RI_DATA_DSGL; sglp->r1 = 0; sglp->nsge = cpu_to_be16(1); sglp->addr0 = cpu_to_be64(mhp->mpl_addr); sglp->len0 = cpu_to_be32(pbllen); *len16 = DIV_ROUND_UP(sizeof(wqe->fr) + sizeof(*sglp), 16); } else { imdp = (struct fw_ri_immd *)(&wqe->fr + 1); imdp->op = FW_RI_DATA_IMMD; imdp->r1 = 0; imdp->r2 = 0; imdp->immdlen = cpu_to_be32(pbllen); p = (__be64 *)(imdp + 1); rem = pbllen; for (i = 0; i < mhp->mpl_len; i++) { *p = cpu_to_be64((u64)mhp->mpl[i]); rem -= sizeof(*p); if (++p == (__be64 *)&sq->queue[sq->size]) p = (__be64 *)sq->queue; } BUG_ON(rem < 0); while (rem) { *p = 0; rem -= sizeof(*p); if (++p == (__be64 *)&sq->queue[sq->size]) p = (__be64 *)sq->queue; } *len16 = DIV_ROUND_UP(sizeof(wqe->fr) + sizeof(*imdp) + pbllen, 16); } return 0; } int c4iw_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr, const struct ib_send_wr **bad_wr) { int err = 0; u8 len16 = 0; enum fw_wr_opcodes fw_opcode = 0; enum fw_ri_wr_flags fw_flags; struct c4iw_qp *qhp; union t4_wr *wqe = NULL; u32 num_wrs; struct t4_swsqe *swsqe; unsigned long flag; u16 idx = 0; struct c4iw_rdev *rdev; qhp = to_c4iw_qp(ibqp); rdev = &qhp->rhp->rdev; spin_lock_irqsave(&qhp->lock, flag); if (t4_wq_in_error(&qhp->wq)) { spin_unlock_irqrestore(&qhp->lock, flag); complete_sq_drain_wr(qhp, wr); return err; } num_wrs = t4_sq_avail(&qhp->wq); if (num_wrs == 0) { spin_unlock_irqrestore(&qhp->lock, flag); *bad_wr = wr; return -ENOMEM; } while (wr) { if (num_wrs == 0) { err = -ENOMEM; *bad_wr = wr; break; } wqe = (union t4_wr *)((u8 *)qhp->wq.sq.queue + qhp->wq.sq.wq_pidx * T4_EQ_ENTRY_SIZE); fw_flags = 0; if (wr->send_flags & IB_SEND_SOLICITED) fw_flags |= FW_RI_SOLICITED_EVENT_FLAG; if (wr->send_flags & IB_SEND_SIGNALED || qhp->sq_sig_all) fw_flags |= FW_RI_COMPLETION_FLAG; swsqe = &qhp->wq.sq.sw_sq[qhp->wq.sq.pidx]; switch (wr->opcode) { case IB_WR_SEND_WITH_INV: case IB_WR_SEND: if (wr->send_flags & IB_SEND_FENCE) fw_flags |= FW_RI_READ_FENCE_FLAG; fw_opcode = FW_RI_SEND_WR; if (wr->opcode == IB_WR_SEND) swsqe->opcode = FW_RI_SEND; else swsqe->opcode = FW_RI_SEND_WITH_INV; err = build_rdma_send(&qhp->wq.sq, wqe, wr, &len16); break; case IB_WR_RDMA_WRITE: fw_opcode = FW_RI_RDMA_WRITE_WR; swsqe->opcode = FW_RI_RDMA_WRITE; err = build_rdma_write(&qhp->wq.sq, wqe, wr, &len16); break; case IB_WR_RDMA_READ: case IB_WR_RDMA_READ_WITH_INV: fw_opcode = FW_RI_RDMA_READ_WR; swsqe->opcode = FW_RI_READ_REQ; if (wr->opcode == IB_WR_RDMA_READ_WITH_INV) { c4iw_invalidate_mr(qhp->rhp, wr->sg_list[0].lkey); fw_flags = FW_RI_RDMA_READ_INVALIDATE; } else { fw_flags = 0; } err = build_rdma_read(wqe, wr, &len16); if (err) break; swsqe->read_len = wr->sg_list[0].length; if (!qhp->wq.sq.oldest_read) qhp->wq.sq.oldest_read = swsqe; break; case IB_WR_REG_MR: { struct c4iw_mr *mhp = to_c4iw_mr(reg_wr(wr)->mr); swsqe->opcode = FW_RI_FAST_REGISTER; if (rdev->adap->params.fr_nsmr_tpte_wr_support && !mhp->attr.state && mhp->mpl_len <= 2) { fw_opcode = FW_RI_FR_NSMR_TPTE_WR; err = build_tpte_memreg(&wqe->fr_tpte, reg_wr(wr), mhp, &len16); } else { fw_opcode = FW_RI_FR_NSMR_WR; err = build_memreg(&qhp->wq.sq, wqe, reg_wr(wr), mhp, &len16, rdev->adap->params.ulptx_memwrite_dsgl); } if (err) break; mhp->attr.state = 1; break; } case IB_WR_LOCAL_INV: if (wr->send_flags & IB_SEND_FENCE) fw_flags |= FW_RI_LOCAL_FENCE_FLAG; fw_opcode = FW_RI_INV_LSTAG_WR; swsqe->opcode = FW_RI_LOCAL_INV; err = build_inv_stag(wqe, wr, &len16); c4iw_invalidate_mr(qhp->rhp, wr->ex.invalidate_rkey); break; default: CTR2(KTR_IW_CXGBE, "%s post of type =%d TBD!", __func__, wr->opcode); err = -EINVAL; } if (err) { *bad_wr = wr; break; } swsqe->idx = qhp->wq.sq.pidx; swsqe->complete = 0; swsqe->signaled = (wr->send_flags & IB_SEND_SIGNALED) || qhp->sq_sig_all; swsqe->flushed = 0; swsqe->wr_id = wr->wr_id; init_wr_hdr(wqe, qhp->wq.sq.pidx, fw_opcode, fw_flags, len16); CTR5(KTR_IW_CXGBE, "%s cookie 0x%llx pidx 0x%x opcode 0x%x read_len %u", __func__, (unsigned long long)wr->wr_id, qhp->wq.sq.pidx, swsqe->opcode, swsqe->read_len); wr = wr->next; num_wrs--; t4_sq_produce(&qhp->wq, len16); idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE); } t4_ring_sq_db(&qhp->wq, idx, wqe, rdev->adap->iwt.wc_en); spin_unlock_irqrestore(&qhp->lock, flag); return err; } int c4iw_post_receive(struct ib_qp *ibqp, const struct ib_recv_wr *wr, const struct ib_recv_wr **bad_wr) { int err = 0; struct c4iw_qp *qhp; union t4_recv_wr *wqe = NULL; u32 num_wrs; u8 len16 = 0; unsigned long flag; u16 idx = 0; qhp = to_c4iw_qp(ibqp); spin_lock_irqsave(&qhp->lock, flag); if (t4_wq_in_error(&qhp->wq)) { spin_unlock_irqrestore(&qhp->lock, flag); complete_rq_drain_wr(qhp, wr); return err; } num_wrs = t4_rq_avail(&qhp->wq); if (num_wrs == 0) { spin_unlock_irqrestore(&qhp->lock, flag); *bad_wr = wr; return -ENOMEM; } while (wr) { if (wr->num_sge > T4_MAX_RECV_SGE) { err = -EINVAL; *bad_wr = wr; break; } wqe = (union t4_recv_wr *)((u8 *)qhp->wq.rq.queue + qhp->wq.rq.wq_pidx * T4_EQ_ENTRY_SIZE); if (num_wrs) err = build_rdma_recv(qhp, wqe, wr, &len16); else err = -ENOMEM; if (err) { *bad_wr = wr; break; } qhp->wq.rq.sw_rq[qhp->wq.rq.pidx].wr_id = wr->wr_id; wqe->recv.opcode = FW_RI_RECV_WR; wqe->recv.r1 = 0; wqe->recv.wrid = qhp->wq.rq.pidx; wqe->recv.r2[0] = 0; wqe->recv.r2[1] = 0; wqe->recv.r2[2] = 0; wqe->recv.len16 = len16; CTR3(KTR_IW_CXGBE, "%s cookie 0x%llx pidx %u", __func__, (unsigned long long) wr->wr_id, qhp->wq.rq.pidx); t4_rq_produce(&qhp->wq, len16); idx += DIV_ROUND_UP(len16*16, T4_EQ_ENTRY_SIZE); wr = wr->next; num_wrs--; } t4_ring_rq_db(&qhp->wq, idx, wqe, qhp->rhp->rdev.adap->iwt.wc_en); spin_unlock_irqrestore(&qhp->lock, flag); return err; } static inline void build_term_codes(struct t4_cqe *err_cqe, u8 *layer_type, u8 *ecode) { int status; int tagged; int opcode; int rqtype; int send_inv; if (!err_cqe) { *layer_type = LAYER_RDMAP|DDP_LOCAL_CATA; *ecode = 0; return; } status = CQE_STATUS(err_cqe); opcode = CQE_OPCODE(err_cqe); rqtype = RQ_TYPE(err_cqe); send_inv = (opcode == FW_RI_SEND_WITH_INV) || (opcode == FW_RI_SEND_WITH_SE_INV); tagged = (opcode == FW_RI_RDMA_WRITE) || (rqtype && (opcode == FW_RI_READ_RESP)); switch (status) { case T4_ERR_STAG: if (send_inv) { *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP; *ecode = RDMAP_CANT_INV_STAG; } else { *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT; *ecode = RDMAP_INV_STAG; } break; case T4_ERR_PDID: *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT; if ((opcode == FW_RI_SEND_WITH_INV) || (opcode == FW_RI_SEND_WITH_SE_INV)) *ecode = RDMAP_CANT_INV_STAG; else *ecode = RDMAP_STAG_NOT_ASSOC; break; case T4_ERR_QPID: *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT; *ecode = RDMAP_STAG_NOT_ASSOC; break; case T4_ERR_ACCESS: *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT; *ecode = RDMAP_ACC_VIOL; break; case T4_ERR_WRAP: *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT; *ecode = RDMAP_TO_WRAP; break; case T4_ERR_BOUND: if (tagged) { *layer_type = LAYER_DDP|DDP_TAGGED_ERR; *ecode = DDPT_BASE_BOUNDS; } else { *layer_type = LAYER_RDMAP|RDMAP_REMOTE_PROT; *ecode = RDMAP_BASE_BOUNDS; } break; case T4_ERR_INVALIDATE_SHARED_MR: case T4_ERR_INVALIDATE_MR_WITH_MW_BOUND: *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP; *ecode = RDMAP_CANT_INV_STAG; break; case T4_ERR_ECC: case T4_ERR_ECC_PSTAG: case T4_ERR_INTERNAL_ERR: *layer_type = LAYER_RDMAP|RDMAP_LOCAL_CATA; *ecode = 0; break; case T4_ERR_OUT_OF_RQE: *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR; *ecode = DDPU_INV_MSN_NOBUF; break; case T4_ERR_PBL_ADDR_BOUND: *layer_type = LAYER_DDP|DDP_TAGGED_ERR; *ecode = DDPT_BASE_BOUNDS; break; case T4_ERR_CRC: *layer_type = LAYER_MPA|DDP_LLP; *ecode = MPA_CRC_ERR; break; case T4_ERR_MARKER: *layer_type = LAYER_MPA|DDP_LLP; *ecode = MPA_MARKER_ERR; break; case T4_ERR_PDU_LEN_ERR: *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR; *ecode = DDPU_MSG_TOOBIG; break; case T4_ERR_DDP_VERSION: if (tagged) { *layer_type = LAYER_DDP|DDP_TAGGED_ERR; *ecode = DDPT_INV_VERS; } else { *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR; *ecode = DDPU_INV_VERS; } break; case T4_ERR_RDMA_VERSION: *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP; *ecode = RDMAP_INV_VERS; break; case T4_ERR_OPCODE: *layer_type = LAYER_RDMAP|RDMAP_REMOTE_OP; *ecode = RDMAP_INV_OPCODE; break; case T4_ERR_DDP_QUEUE_NUM: *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR; *ecode = DDPU_INV_QN; break; case T4_ERR_MSN: case T4_ERR_MSN_GAP: case T4_ERR_MSN_RANGE: case T4_ERR_IRD_OVERFLOW: *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR; *ecode = DDPU_INV_MSN_RANGE; break; case T4_ERR_TBIT: *layer_type = LAYER_DDP|DDP_LOCAL_CATA; *ecode = 0; break; case T4_ERR_MO: *layer_type = LAYER_DDP|DDP_UNTAGGED_ERR; *ecode = DDPU_INV_MO; break; default: *layer_type = LAYER_RDMAP|DDP_LOCAL_CATA; *ecode = 0; break; } } static void post_terminate(struct c4iw_qp *qhp, struct t4_cqe *err_cqe, gfp_t gfp) { int ret; struct fw_ri_wr *wqe; struct terminate_message *term; struct wrqe *wr; struct socket *so = qhp->ep->com.so; struct inpcb *inp = sotoinpcb(so); struct tcpcb *tp = intotcpcb(inp); struct toepcb *toep = tp->t_toe; CTR4(KTR_IW_CXGBE, "%s qhp %p qid 0x%x tid %u", __func__, qhp, qhp->wq.sq.qid, qhp->ep->hwtid); wr = alloc_wrqe(sizeof(*wqe), &toep->ofld_txq->wrq); if (wr == NULL) return; wqe = wrtod(wr); memset(wqe, 0, sizeof *wqe); wqe->op_compl = cpu_to_be32(V_FW_WR_OP(FW_RI_WR)); wqe->flowid_len16 = cpu_to_be32( V_FW_WR_FLOWID(qhp->ep->hwtid) | V_FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16))); wqe->u.terminate.type = FW_RI_TYPE_TERMINATE; wqe->u.terminate.immdlen = cpu_to_be32(sizeof *term); term = (struct terminate_message *)wqe->u.terminate.termmsg; if (qhp->attr.layer_etype == (LAYER_MPA|DDP_LLP)) { term->layer_etype = qhp->attr.layer_etype; term->ecode = qhp->attr.ecode; } else build_term_codes(err_cqe, &term->layer_etype, &term->ecode); ret = creds(toep, inp, sizeof(*wqe)); if (ret) { free_wrqe(wr); return; } t4_wrq_tx(qhp->rhp->rdev.adap, wr); } /* Assumes qhp lock is held. */ static void __flush_qp(struct c4iw_qp *qhp, struct c4iw_cq *rchp, struct c4iw_cq *schp) { int count; int rq_flushed, sq_flushed; unsigned long flag; CTR4(KTR_IW_CXGBE, "%s qhp %p rchp %p schp %p", __func__, qhp, rchp, schp); /* locking hierarchy: cq lock first, then qp lock. */ spin_lock_irqsave(&rchp->lock, flag); spin_lock(&qhp->lock); if (qhp->wq.flushed) { spin_unlock(&qhp->lock); spin_unlock_irqrestore(&rchp->lock, flag); return; } qhp->wq.flushed = 1; c4iw_flush_hw_cq(rchp); c4iw_count_rcqes(&rchp->cq, &qhp->wq, &count); rq_flushed = c4iw_flush_rq(&qhp->wq, &rchp->cq, count); spin_unlock(&qhp->lock); spin_unlock_irqrestore(&rchp->lock, flag); /* locking hierarchy: cq lock first, then qp lock. */ spin_lock_irqsave(&schp->lock, flag); spin_lock(&qhp->lock); if (schp != rchp) c4iw_flush_hw_cq(schp); sq_flushed = c4iw_flush_sq(qhp); spin_unlock(&qhp->lock); spin_unlock_irqrestore(&schp->lock, flag); if (schp == rchp) { if (t4_clear_cq_armed(&rchp->cq) && (rq_flushed || sq_flushed)) { spin_lock_irqsave(&rchp->comp_handler_lock, flag); (*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context); spin_unlock_irqrestore(&rchp->comp_handler_lock, flag); } } else { if (t4_clear_cq_armed(&rchp->cq) && rq_flushed) { spin_lock_irqsave(&rchp->comp_handler_lock, flag); (*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context); spin_unlock_irqrestore(&rchp->comp_handler_lock, flag); } if (t4_clear_cq_armed(&schp->cq) && sq_flushed) { spin_lock_irqsave(&schp->comp_handler_lock, flag); (*schp->ibcq.comp_handler)(&schp->ibcq, schp->ibcq.cq_context); spin_unlock_irqrestore(&schp->comp_handler_lock, flag); } } } static void flush_qp(struct c4iw_qp *qhp) { struct c4iw_cq *rchp, *schp; unsigned long flag; rchp = to_c4iw_cq(qhp->ibqp.recv_cq); schp = to_c4iw_cq(qhp->ibqp.send_cq); t4_set_wq_in_error(&qhp->wq); if (qhp->ibqp.uobject) { t4_set_cq_in_error(&rchp->cq); spin_lock_irqsave(&rchp->comp_handler_lock, flag); (*rchp->ibcq.comp_handler)(&rchp->ibcq, rchp->ibcq.cq_context); spin_unlock_irqrestore(&rchp->comp_handler_lock, flag); if (schp != rchp) { t4_set_cq_in_error(&schp->cq); spin_lock_irqsave(&schp->comp_handler_lock, flag); (*schp->ibcq.comp_handler)(&schp->ibcq, schp->ibcq.cq_context); spin_unlock_irqrestore(&schp->comp_handler_lock, flag); } return; } __flush_qp(qhp, rchp, schp); } static int rdma_fini(struct c4iw_dev *rhp, struct c4iw_qp *qhp, struct c4iw_ep *ep) { struct c4iw_rdev *rdev = &rhp->rdev; struct adapter *sc = rdev->adap; struct fw_ri_wr *wqe; int ret; struct wrqe *wr; struct socket *so = ep->com.so; struct inpcb *inp = sotoinpcb(so); struct tcpcb *tp = intotcpcb(inp); struct toepcb *toep = tp->t_toe; KASSERT(rhp == qhp->rhp && ep == qhp->ep, ("%s: EDOOFUS", __func__)); CTR5(KTR_IW_CXGBE, "%s qhp %p qid 0x%x ep %p tid %u", __func__, qhp, qhp->wq.sq.qid, ep, ep->hwtid); wr = alloc_wrqe(sizeof(*wqe), &toep->ofld_txq->wrq); if (wr == NULL) return (0); wqe = wrtod(wr); memset(wqe, 0, sizeof *wqe); wqe->op_compl = cpu_to_be32(V_FW_WR_OP(FW_RI_WR) | F_FW_WR_COMPL); wqe->flowid_len16 = cpu_to_be32(V_FW_WR_FLOWID(ep->hwtid) | V_FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16))); wqe->cookie = (unsigned long) &ep->com.wr_wait; wqe->u.fini.type = FW_RI_TYPE_FINI; c4iw_init_wr_wait(&ep->com.wr_wait); ret = creds(toep, inp, sizeof(*wqe)); if (ret) { free_wrqe(wr); return ret; } t4_wrq_tx(sc, wr); ret = c4iw_wait_for_reply(rdev, &ep->com.wr_wait, ep->hwtid, qhp->wq.sq.qid, ep->com.so, __func__); return ret; } static void build_rtr_msg(u8 p2p_type, struct fw_ri_init *init) { CTR2(KTR_IW_CXGBE, "%s p2p_type = %d", __func__, p2p_type); memset(&init->u, 0, sizeof init->u); switch (p2p_type) { case FW_RI_INIT_P2PTYPE_RDMA_WRITE: init->u.write.opcode = FW_RI_RDMA_WRITE_WR; init->u.write.stag_sink = cpu_to_be32(1); init->u.write.to_sink = cpu_to_be64(1); init->u.write.u.immd_src[0].op = FW_RI_DATA_IMMD; init->u.write.len16 = DIV_ROUND_UP(sizeof init->u.write + sizeof(struct fw_ri_immd), 16); break; case FW_RI_INIT_P2PTYPE_READ_REQ: init->u.write.opcode = FW_RI_RDMA_READ_WR; init->u.read.stag_src = cpu_to_be32(1); init->u.read.to_src_lo = cpu_to_be32(1); init->u.read.stag_sink = cpu_to_be32(1); init->u.read.to_sink_lo = cpu_to_be32(1); init->u.read.len16 = DIV_ROUND_UP(sizeof init->u.read, 16); break; } } static int creds(struct toepcb *toep, struct inpcb *inp, size_t wrsize) { struct ofld_tx_sdesc *txsd; CTR3(KTR_IW_CXGBE, "%s:creB %p %u", __func__, toep , wrsize); INP_WLOCK(inp); - if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) != 0) { + if ((inp->inp_flags & INP_DROPPED) != 0) { INP_WUNLOCK(inp); return (EINVAL); } txsd = &toep->txsd[toep->txsd_pidx]; txsd->tx_credits = howmany(wrsize, 16); txsd->plen = 0; KASSERT(toep->tx_credits >= txsd->tx_credits && toep->txsd_avail > 0, ("%s: not enough credits (%d)", __func__, toep->tx_credits)); toep->tx_credits -= txsd->tx_credits; if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) toep->txsd_pidx = 0; toep->txsd_avail--; INP_WUNLOCK(inp); CTR5(KTR_IW_CXGBE, "%s:creE %p %u %u %u", __func__, toep , txsd->tx_credits, toep->tx_credits, toep->txsd_pidx); return (0); } static int rdma_init(struct c4iw_dev *rhp, struct c4iw_qp *qhp) { struct fw_ri_wr *wqe; int ret; struct wrqe *wr; struct c4iw_ep *ep = qhp->ep; struct c4iw_rdev *rdev = &qhp->rhp->rdev; struct adapter *sc = rdev->adap; struct socket *so = ep->com.so; struct inpcb *inp = sotoinpcb(so); struct tcpcb *tp = intotcpcb(inp); struct toepcb *toep = tp->t_toe; CTR5(KTR_IW_CXGBE, "%s qhp %p qid 0x%x ep %p tid %u", __func__, qhp, qhp->wq.sq.qid, ep, ep->hwtid); wr = alloc_wrqe(sizeof(*wqe), &toep->ofld_txq->wrq); if (wr == NULL) return (0); wqe = wrtod(wr); ret = alloc_ird(rhp, qhp->attr.max_ird); if (ret) { qhp->attr.max_ird = 0; free_wrqe(wr); return ret; } memset(wqe, 0, sizeof *wqe); wqe->op_compl = cpu_to_be32( V_FW_WR_OP(FW_RI_WR) | F_FW_WR_COMPL); wqe->flowid_len16 = cpu_to_be32(V_FW_WR_FLOWID(ep->hwtid) | V_FW_WR_LEN16(DIV_ROUND_UP(sizeof *wqe, 16))); wqe->cookie = (unsigned long) &ep->com.wr_wait; wqe->u.init.type = FW_RI_TYPE_INIT; wqe->u.init.mpareqbit_p2ptype = V_FW_RI_WR_MPAREQBIT(qhp->attr.mpa_attr.initiator) | V_FW_RI_WR_P2PTYPE(qhp->attr.mpa_attr.p2p_type); wqe->u.init.mpa_attrs = FW_RI_MPA_IETF_ENABLE; if (qhp->attr.mpa_attr.recv_marker_enabled) wqe->u.init.mpa_attrs |= FW_RI_MPA_RX_MARKER_ENABLE; if (qhp->attr.mpa_attr.xmit_marker_enabled) wqe->u.init.mpa_attrs |= FW_RI_MPA_TX_MARKER_ENABLE; if (qhp->attr.mpa_attr.crc_enabled) wqe->u.init.mpa_attrs |= FW_RI_MPA_CRC_ENABLE; wqe->u.init.qp_caps = FW_RI_QP_RDMA_READ_ENABLE | FW_RI_QP_RDMA_WRITE_ENABLE | FW_RI_QP_BIND_ENABLE; if (!qhp->ibqp.uobject) wqe->u.init.qp_caps |= FW_RI_QP_FAST_REGISTER_ENABLE | FW_RI_QP_STAG0_ENABLE; wqe->u.init.nrqe = cpu_to_be16(t4_rqes_posted(&qhp->wq)); wqe->u.init.pdid = cpu_to_be32(qhp->attr.pd); wqe->u.init.qpid = cpu_to_be32(qhp->wq.sq.qid); wqe->u.init.sq_eqid = cpu_to_be32(qhp->wq.sq.qid); wqe->u.init.rq_eqid = cpu_to_be32(qhp->wq.rq.qid); wqe->u.init.scqid = cpu_to_be32(qhp->attr.scq); wqe->u.init.rcqid = cpu_to_be32(qhp->attr.rcq); wqe->u.init.ord_max = cpu_to_be32(qhp->attr.max_ord); wqe->u.init.ird_max = cpu_to_be32(qhp->attr.max_ird); wqe->u.init.iss = cpu_to_be32(ep->snd_seq); wqe->u.init.irs = cpu_to_be32(ep->rcv_seq); wqe->u.init.hwrqsize = cpu_to_be32(qhp->wq.rq.rqt_size); wqe->u.init.hwrqaddr = cpu_to_be32(qhp->wq.rq.rqt_hwaddr - sc->vres.rq.start); if (qhp->attr.mpa_attr.initiator) build_rtr_msg(qhp->attr.mpa_attr.p2p_type, &wqe->u.init); c4iw_init_wr_wait(&ep->com.wr_wait); ret = creds(toep, inp, sizeof(*wqe)); if (ret) { free_wrqe(wr); free_ird(rhp, qhp->attr.max_ird); return ret; } t4_wrq_tx(sc, wr); ret = c4iw_wait_for_reply(rdev, &ep->com.wr_wait, ep->hwtid, qhp->wq.sq.qid, ep->com.so, __func__); toep->params.ulp_mode = ULP_MODE_RDMA; free_ird(rhp, qhp->attr.max_ird); return ret; } int c4iw_modify_qp(struct c4iw_dev *rhp, struct c4iw_qp *qhp, enum c4iw_qp_attr_mask mask, struct c4iw_qp_attributes *attrs, int internal) { int ret = 0; struct c4iw_qp_attributes newattr = qhp->attr; int disconnect = 0; int terminate = 0; int abort = 0; int free = 0; struct c4iw_ep *ep = NULL; CTR5(KTR_IW_CXGBE, "%s qhp %p sqid 0x%x rqid 0x%x ep %p", __func__, qhp, qhp->wq.sq.qid, qhp->wq.rq.qid, qhp->ep); CTR3(KTR_IW_CXGBE, "%s state %d -> %d", __func__, qhp->attr.state, (mask & C4IW_QP_ATTR_NEXT_STATE) ? attrs->next_state : -1); mutex_lock(&qhp->mutex); /* Process attr changes if in IDLE */ if (mask & C4IW_QP_ATTR_VALID_MODIFY) { if (qhp->attr.state != C4IW_QP_STATE_IDLE) { ret = -EIO; goto out; } if (mask & C4IW_QP_ATTR_ENABLE_RDMA_READ) newattr.enable_rdma_read = attrs->enable_rdma_read; if (mask & C4IW_QP_ATTR_ENABLE_RDMA_WRITE) newattr.enable_rdma_write = attrs->enable_rdma_write; if (mask & C4IW_QP_ATTR_ENABLE_RDMA_BIND) newattr.enable_bind = attrs->enable_bind; if (mask & C4IW_QP_ATTR_MAX_ORD) { if (attrs->max_ord > c4iw_max_read_depth) { ret = -EINVAL; goto out; } newattr.max_ord = attrs->max_ord; } if (mask & C4IW_QP_ATTR_MAX_IRD) { if (attrs->max_ird > cur_max_read_depth(rhp)) { ret = -EINVAL; goto out; } newattr.max_ird = attrs->max_ird; } qhp->attr = newattr; } if (!(mask & C4IW_QP_ATTR_NEXT_STATE)) goto out; if (qhp->attr.state == attrs->next_state) goto out; /* Return EINPROGRESS if QP is already in transition state. * Eg: CLOSING->IDLE transition or *->ERROR transition. * This can happen while connection is switching(due to rdma_fini) * from iWARP/RDDP to TOE mode and any inflight RDMA RX data will * reach TOE driver -> TCP stack -> iWARP driver. In this way * iWARP driver keep receiving inflight RDMA RX data until socket * is closed or aborted. And if iWARP CM is in FPDU sate, then * it tries to put QP in TERM state and disconnects endpoint. * But as QP is already in transition state, this event is ignored. */ if ((qhp->attr.state >= C4IW_QP_STATE_ERROR) && (attrs->next_state == C4IW_QP_STATE_TERMINATE)) { ret = -EINPROGRESS; goto out; } switch (qhp->attr.state) { case C4IW_QP_STATE_IDLE: switch (attrs->next_state) { case C4IW_QP_STATE_RTS: if (!(mask & C4IW_QP_ATTR_LLP_STREAM_HANDLE)) { ret = -EINVAL; goto out; } if (!(mask & C4IW_QP_ATTR_MPA_ATTR)) { ret = -EINVAL; goto out; } qhp->attr.mpa_attr = attrs->mpa_attr; qhp->attr.llp_stream_handle = attrs->llp_stream_handle; qhp->ep = qhp->attr.llp_stream_handle; set_state(qhp, C4IW_QP_STATE_RTS); /* * Ref the endpoint here and deref when we * disassociate the endpoint from the QP. This * happens in CLOSING->IDLE transition or *->ERROR * transition. */ c4iw_get_ep(&qhp->ep->com); ret = rdma_init(rhp, qhp); if (ret) goto err; break; case C4IW_QP_STATE_ERROR: set_state(qhp, C4IW_QP_STATE_ERROR); flush_qp(qhp); break; default: ret = -EINVAL; goto out; } break; case C4IW_QP_STATE_RTS: switch (attrs->next_state) { case C4IW_QP_STATE_CLOSING: BUG_ON(atomic_read(&qhp->ep->com.kref.refcount) < 2); t4_set_wq_in_error(&qhp->wq); set_state(qhp, C4IW_QP_STATE_CLOSING); ep = qhp->ep; if (!internal) { abort = 0; disconnect = 1; c4iw_get_ep(&qhp->ep->com); } ret = rdma_fini(rhp, qhp, ep); if (ret) goto err; break; case C4IW_QP_STATE_TERMINATE: t4_set_wq_in_error(&qhp->wq); set_state(qhp, C4IW_QP_STATE_TERMINATE); qhp->attr.layer_etype = attrs->layer_etype; qhp->attr.ecode = attrs->ecode; ep = qhp->ep; if (!internal) { c4iw_get_ep(&qhp->ep->com); terminate = 1; disconnect = 1; } else { terminate = qhp->attr.send_term; ret = rdma_fini(rhp, qhp, ep); if (ret) goto err; } break; case C4IW_QP_STATE_ERROR: t4_set_wq_in_error(&qhp->wq); set_state(qhp, C4IW_QP_STATE_ERROR); if (!internal) { abort = 1; disconnect = 1; ep = qhp->ep; c4iw_get_ep(&qhp->ep->com); } goto err; break; default: ret = -EINVAL; goto out; } break; case C4IW_QP_STATE_CLOSING: /* * Allow kernel users to move to ERROR for qp draining. */ if (!internal && (qhp->ibqp.uobject || attrs->next_state != C4IW_QP_STATE_ERROR)) { ret = -EINVAL; goto out; } switch (attrs->next_state) { case C4IW_QP_STATE_IDLE: flush_qp(qhp); set_state(qhp, C4IW_QP_STATE_IDLE); qhp->attr.llp_stream_handle = NULL; c4iw_put_ep(&qhp->ep->com); qhp->ep = NULL; wake_up(&qhp->wait); break; case C4IW_QP_STATE_ERROR: goto err; default: ret = -EINVAL; goto err; } break; case C4IW_QP_STATE_ERROR: if (attrs->next_state != C4IW_QP_STATE_IDLE) { ret = -EINVAL; goto out; } if (!t4_sq_empty(&qhp->wq) || !t4_rq_empty(&qhp->wq)) { ret = -EINVAL; goto out; } set_state(qhp, C4IW_QP_STATE_IDLE); break; case C4IW_QP_STATE_TERMINATE: if (!internal) { ret = -EINVAL; goto out; } goto err; break; default: printf("%s in a bad state %d\n", __func__, qhp->attr.state); ret = -EINVAL; goto err; break; } goto out; err: CTR3(KTR_IW_CXGBE, "%s disassociating ep %p qpid 0x%x", __func__, qhp->ep, qhp->wq.sq.qid); /* disassociate the LLP connection */ qhp->attr.llp_stream_handle = NULL; if (!ep) ep = qhp->ep; qhp->ep = NULL; set_state(qhp, C4IW_QP_STATE_ERROR); free = 1; abort = 1; BUG_ON(!ep); flush_qp(qhp); wake_up(&qhp->wait); out: mutex_unlock(&qhp->mutex); if (terminate) post_terminate(qhp, NULL, internal ? GFP_ATOMIC : GFP_KERNEL); /* * If disconnect is 1, then we need to initiate a disconnect * on the EP. This can be a normal close (RTS->CLOSING) or * an abnormal close (RTS/CLOSING->ERROR). */ if (disconnect) { __c4iw_ep_disconnect(ep, abort, internal ? GFP_ATOMIC : GFP_KERNEL); c4iw_put_ep(&ep->com); } /* * If free is 1, then we've disassociated the EP from the QP * and we need to dereference the EP. */ if (free) c4iw_put_ep(&ep->com); CTR2(KTR_IW_CXGBE, "%s exit state %d", __func__, qhp->attr.state); return ret; } int c4iw_destroy_qp(struct ib_qp *ib_qp, struct ib_udata *udata) { struct c4iw_dev *rhp; struct c4iw_qp *qhp; struct c4iw_qp_attributes attrs; CTR2(KTR_IW_CXGBE, "%s ib_qp %p", __func__, ib_qp); qhp = to_c4iw_qp(ib_qp); rhp = qhp->rhp; attrs.next_state = C4IW_QP_STATE_ERROR; if (qhp->attr.state == C4IW_QP_STATE_TERMINATE) c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 1); else c4iw_modify_qp(rhp, qhp, C4IW_QP_ATTR_NEXT_STATE, &attrs, 0); wait_event(qhp->wait, !qhp->ep); remove_handle(rhp, &rhp->qpidr, qhp->wq.sq.qid); free_ird(rhp, qhp->attr.max_ird); c4iw_qp_rem_ref(ib_qp); CTR3(KTR_IW_CXGBE, "%s ib_qp %p qpid 0x%0x", __func__, ib_qp, qhp->wq.sq.qid); return 0; } struct ib_qp * c4iw_create_qp(struct ib_pd *pd, struct ib_qp_init_attr *attrs, struct ib_udata *udata) { struct c4iw_dev *rhp; struct c4iw_qp *qhp; struct c4iw_pd *php; struct c4iw_cq *schp; struct c4iw_cq *rchp; struct c4iw_create_qp_resp uresp; unsigned int sqsize, rqsize; struct c4iw_ucontext *ucontext; int ret; struct c4iw_mm_entry *sq_key_mm = NULL, *rq_key_mm = NULL; struct c4iw_mm_entry *sq_db_key_mm = NULL, *rq_db_key_mm = NULL; CTR2(KTR_IW_CXGBE, "%s ib_pd %p", __func__, pd); if (attrs->qp_type != IB_QPT_RC) return ERR_PTR(-EINVAL); php = to_c4iw_pd(pd); rhp = php->rhp; schp = get_chp(rhp, ((struct c4iw_cq *)attrs->send_cq)->cq.cqid); rchp = get_chp(rhp, ((struct c4iw_cq *)attrs->recv_cq)->cq.cqid); if (!schp || !rchp) return ERR_PTR(-EINVAL); if (attrs->cap.max_inline_data > T4_MAX_SEND_INLINE) return ERR_PTR(-EINVAL); if (attrs->cap.max_recv_wr > rhp->rdev.hw_queue.t4_max_rq_size) return ERR_PTR(-E2BIG); rqsize = attrs->cap.max_recv_wr + 1; if (rqsize < 8) rqsize = 8; if (attrs->cap.max_send_wr > rhp->rdev.hw_queue.t4_max_sq_size) return ERR_PTR(-E2BIG); sqsize = attrs->cap.max_send_wr + 1; if (sqsize < 8) sqsize = 8; ucontext = pd->uobject ? to_c4iw_ucontext(pd->uobject->context) : NULL; qhp = kzalloc(sizeof(*qhp), GFP_KERNEL); if (!qhp) return ERR_PTR(-ENOMEM); qhp->wq.sq.size = sqsize; qhp->wq.sq.memsize = (sqsize + rhp->rdev.hw_queue.t4_eq_status_entries) * sizeof(*qhp->wq.sq.queue) + 16 * sizeof(__be64); qhp->wq.sq.flush_cidx = -1; qhp->wq.rq.size = rqsize; qhp->wq.rq.memsize = (rqsize + rhp->rdev.hw_queue.t4_eq_status_entries) * sizeof(*qhp->wq.rq.queue); if (ucontext) { qhp->wq.sq.memsize = roundup(qhp->wq.sq.memsize, PAGE_SIZE); qhp->wq.rq.memsize = roundup(qhp->wq.rq.memsize, PAGE_SIZE); } CTR5(KTR_IW_CXGBE, "%s sqsize %u sqmemsize %zu rqsize %u rqmemsize %zu", __func__, sqsize, qhp->wq.sq.memsize, rqsize, qhp->wq.rq.memsize); ret = create_qp(&rhp->rdev, &qhp->wq, &schp->cq, &rchp->cq, ucontext ? &ucontext->uctx : &rhp->rdev.uctx); if (ret) goto err1; attrs->cap.max_recv_wr = rqsize - 1; attrs->cap.max_send_wr = sqsize - 1; attrs->cap.max_inline_data = T4_MAX_SEND_INLINE; qhp->rhp = rhp; qhp->attr.pd = php->pdid; qhp->attr.scq = ((struct c4iw_cq *) attrs->send_cq)->cq.cqid; qhp->attr.rcq = ((struct c4iw_cq *) attrs->recv_cq)->cq.cqid; qhp->attr.sq_num_entries = attrs->cap.max_send_wr; qhp->attr.rq_num_entries = attrs->cap.max_recv_wr; qhp->attr.sq_max_sges = attrs->cap.max_send_sge; qhp->attr.sq_max_sges_rdma_write = attrs->cap.max_send_sge; qhp->attr.rq_max_sges = attrs->cap.max_recv_sge; qhp->attr.state = C4IW_QP_STATE_IDLE; qhp->attr.next_state = C4IW_QP_STATE_IDLE; qhp->attr.enable_rdma_read = 1; qhp->attr.enable_rdma_write = 1; qhp->attr.enable_bind = 1; qhp->attr.max_ord = 0; qhp->attr.max_ird = 0; qhp->sq_sig_all = attrs->sq_sig_type == IB_SIGNAL_ALL_WR; spin_lock_init(&qhp->lock); mutex_init(&qhp->mutex); init_waitqueue_head(&qhp->wait); kref_init(&qhp->kref); INIT_WORK(&qhp->free_work, free_qp_work); ret = insert_handle(rhp, &rhp->qpidr, qhp, qhp->wq.sq.qid); if (ret) goto err2; if (udata) { sq_key_mm = kmalloc(sizeof(*sq_key_mm), GFP_KERNEL); if (!sq_key_mm) { ret = -ENOMEM; goto err3; } rq_key_mm = kmalloc(sizeof(*rq_key_mm), GFP_KERNEL); if (!rq_key_mm) { ret = -ENOMEM; goto err4; } sq_db_key_mm = kmalloc(sizeof(*sq_db_key_mm), GFP_KERNEL); if (!sq_db_key_mm) { ret = -ENOMEM; goto err5; } rq_db_key_mm = kmalloc(sizeof(*rq_db_key_mm), GFP_KERNEL); if (!rq_db_key_mm) { ret = -ENOMEM; goto err6; } uresp.flags = 0; uresp.qid_mask = rhp->rdev.qpmask; uresp.sqid = qhp->wq.sq.qid; uresp.sq_size = qhp->wq.sq.size; uresp.sq_memsize = qhp->wq.sq.memsize; uresp.rqid = qhp->wq.rq.qid; uresp.rq_size = qhp->wq.rq.size; uresp.rq_memsize = qhp->wq.rq.memsize; spin_lock(&ucontext->mmap_lock); uresp.ma_sync_key = 0; uresp.sq_key = ucontext->key; ucontext->key += PAGE_SIZE; uresp.rq_key = ucontext->key; ucontext->key += PAGE_SIZE; uresp.sq_db_gts_key = ucontext->key; ucontext->key += PAGE_SIZE; uresp.rq_db_gts_key = ucontext->key; ucontext->key += PAGE_SIZE; spin_unlock(&ucontext->mmap_lock); ret = ib_copy_to_udata(udata, &uresp, sizeof uresp); if (ret) goto err7; sq_key_mm->key = uresp.sq_key; sq_key_mm->addr = qhp->wq.sq.phys_addr; sq_key_mm->len = PAGE_ALIGN(qhp->wq.sq.memsize); CTR4(KTR_IW_CXGBE, "%s sq_key_mm %x, %x, %d", __func__, sq_key_mm->key, sq_key_mm->addr, sq_key_mm->len); insert_mmap(ucontext, sq_key_mm); rq_key_mm->key = uresp.rq_key; rq_key_mm->addr = qhp->wq.rq.phys_addr; rq_key_mm->len = PAGE_ALIGN(qhp->wq.rq.memsize); CTR4(KTR_IW_CXGBE, "%s rq_key_mm %x, %x, %d", __func__, rq_key_mm->key, rq_key_mm->addr, rq_key_mm->len); insert_mmap(ucontext, rq_key_mm); sq_db_key_mm->key = uresp.sq_db_gts_key; sq_db_key_mm->addr = (u64)qhp->wq.sq.bar2_pa; sq_db_key_mm->len = PAGE_SIZE; CTR4(KTR_IW_CXGBE, "%s sq_db_key_mm %x, %x, %d", __func__, sq_db_key_mm->key, sq_db_key_mm->addr, sq_db_key_mm->len); insert_mmap(ucontext, sq_db_key_mm); rq_db_key_mm->key = uresp.rq_db_gts_key; rq_db_key_mm->addr = (u64)qhp->wq.rq.bar2_pa; rq_db_key_mm->len = PAGE_SIZE; CTR4(KTR_IW_CXGBE, "%s rq_db_key_mm %x, %x, %d", __func__, rq_db_key_mm->key, rq_db_key_mm->addr, rq_db_key_mm->len); insert_mmap(ucontext, rq_db_key_mm); qhp->ucontext = ucontext; } qhp->ibqp.qp_num = qhp->wq.sq.qid; init_timer(&(qhp->timer)); CTR5(KTR_IW_CXGBE, "%s sq id %u size %u memsize %zu num_entries %u", __func__, qhp->wq.sq.qid, qhp->wq.sq.size, qhp->wq.sq.memsize, attrs->cap.max_send_wr); CTR5(KTR_IW_CXGBE, "%s rq id %u size %u memsize %zu num_entries %u", __func__, qhp->wq.rq.qid, qhp->wq.rq.size, qhp->wq.rq.memsize, attrs->cap.max_recv_wr); return &qhp->ibqp; err7: kfree(rq_db_key_mm); err6: kfree(sq_db_key_mm); err5: kfree(rq_key_mm); err4: kfree(sq_key_mm); err3: remove_handle(rhp, &rhp->qpidr, qhp->wq.sq.qid); err2: destroy_qp(&rhp->rdev, &qhp->wq, ucontext ? &ucontext->uctx : &rhp->rdev.uctx); err1: kfree(qhp); return ERR_PTR(ret); } int c4iw_ib_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask, struct ib_udata *udata) { struct c4iw_dev *rhp; struct c4iw_qp *qhp; enum c4iw_qp_attr_mask mask = 0; struct c4iw_qp_attributes attrs; CTR2(KTR_IW_CXGBE, "%s ib_qp %p", __func__, ibqp); /* iwarp does not support the RTR state */ if ((attr_mask & IB_QP_STATE) && (attr->qp_state == IB_QPS_RTR)) attr_mask &= ~IB_QP_STATE; /* Make sure we still have something left to do */ if (!attr_mask) return 0; memset(&attrs, 0, sizeof attrs); qhp = to_c4iw_qp(ibqp); rhp = qhp->rhp; attrs.next_state = c4iw_convert_state(attr->qp_state); attrs.enable_rdma_read = (attr->qp_access_flags & IB_ACCESS_REMOTE_READ) ? 1 : 0; attrs.enable_rdma_write = (attr->qp_access_flags & IB_ACCESS_REMOTE_WRITE) ? 1 : 0; attrs.enable_bind = (attr->qp_access_flags & IB_ACCESS_MW_BIND) ? 1 : 0; mask |= (attr_mask & IB_QP_STATE) ? C4IW_QP_ATTR_NEXT_STATE : 0; mask |= (attr_mask & IB_QP_ACCESS_FLAGS) ? (C4IW_QP_ATTR_ENABLE_RDMA_READ | C4IW_QP_ATTR_ENABLE_RDMA_WRITE | C4IW_QP_ATTR_ENABLE_RDMA_BIND) : 0; return c4iw_modify_qp(rhp, qhp, mask, &attrs, 0); } struct ib_qp *c4iw_get_qp(struct ib_device *dev, int qpn) { CTR3(KTR_IW_CXGBE, "%s ib_dev %p qpn 0x%x", __func__, dev, qpn); return (struct ib_qp *)get_qhp(to_c4iw_dev(dev), qpn); } int c4iw_ib_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr, int attr_mask, struct ib_qp_init_attr *init_attr) { struct c4iw_qp *qhp = to_c4iw_qp(ibqp); memset(attr, 0, sizeof *attr); memset(init_attr, 0, sizeof *init_attr); attr->qp_state = to_ib_qp_state(qhp->attr.state); init_attr->cap.max_send_wr = qhp->attr.sq_num_entries; init_attr->cap.max_recv_wr = qhp->attr.rq_num_entries; init_attr->cap.max_send_sge = qhp->attr.sq_max_sges; init_attr->cap.max_recv_sge = qhp->attr.sq_max_sges; init_attr->cap.max_inline_data = T4_MAX_SEND_INLINE; init_attr->sq_sig_type = qhp->sq_sig_all ? IB_SIGNAL_ALL_WR : 0; return 0; } #endif diff --git a/sys/dev/cxgbe/tom/t4_cpl_io.c b/sys/dev/cxgbe/tom/t4_cpl_io.c index 59d1c367f94c..bfc9eb3b76f7 100644 --- a/sys/dev/cxgbe/tom/t4_cpl_io.c +++ b/sys/dev/cxgbe/tom/t4_cpl_io.c @@ -1,2470 +1,2470 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2012, 2015 Chelsio Communications, Inc. * All rights reserved. * Written by: Navdeep Parhar * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include "opt_kern_tls.h" #include "opt_ratelimit.h" #ifdef TCP_OFFLOAD #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #define TCPSTATES #include #include #include #include #include #include #include #include #include #include #include #include "common/common.h" #include "common/t4_msg.h" #include "common/t4_regs.h" #include "common/t4_tcb.h" #include "tom/t4_tom_l2t.h" #include "tom/t4_tom.h" static void t4_aiotx_cancel(struct kaiocb *job); static void t4_aiotx_queue_toep(struct socket *so, struct toepcb *toep); void send_flowc_wr(struct toepcb *toep, struct tcpcb *tp) { struct wrqe *wr; struct fw_flowc_wr *flowc; unsigned int nparams, flowclen, paramidx; struct vi_info *vi = toep->vi; struct port_info *pi = vi->pi; struct adapter *sc = pi->adapter; unsigned int pfvf = sc->pf << S_FW_VIID_PFN; struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx]; KASSERT(!(toep->flags & TPF_FLOWC_WR_SENT), ("%s: flowc for tid %u sent already", __func__, toep->tid)); if (tp != NULL) nparams = 8; else nparams = 6; if (ulp_mode(toep) == ULP_MODE_TLS) nparams++; if (toep->tls.fcplenmax != 0) nparams++; if (toep->params.tc_idx != -1) { MPASS(toep->params.tc_idx >= 0 && toep->params.tc_idx < sc->params.nsched_cls); nparams++; } flowclen = sizeof(*flowc) + nparams * sizeof(struct fw_flowc_mnemval); wr = alloc_wrqe(roundup2(flowclen, 16), &toep->ofld_txq->wrq); if (wr == NULL) { /* XXX */ panic("%s: allocation failure.", __func__); } flowc = wrtod(wr); memset(flowc, 0, wr->wr_len); flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) | V_FW_FLOWC_WR_NPARAMS(nparams)); flowc->flowid_len16 = htonl(V_FW_WR_LEN16(howmany(flowclen, 16)) | V_FW_WR_FLOWID(toep->tid)); #define FLOWC_PARAM(__m, __v) \ do { \ flowc->mnemval[paramidx].mnemonic = FW_FLOWC_MNEM_##__m; \ flowc->mnemval[paramidx].val = htobe32(__v); \ paramidx++; \ } while (0) paramidx = 0; FLOWC_PARAM(PFNVFN, pfvf); FLOWC_PARAM(CH, pi->tx_chan); FLOWC_PARAM(PORT, pi->tx_chan); FLOWC_PARAM(IQID, toep->ofld_rxq->iq.abs_id); FLOWC_PARAM(SNDBUF, toep->params.sndbuf); if (tp) { FLOWC_PARAM(MSS, toep->params.emss); FLOWC_PARAM(SNDNXT, tp->snd_nxt); FLOWC_PARAM(RCVNXT, tp->rcv_nxt); } else FLOWC_PARAM(MSS, 512); CTR6(KTR_CXGBE, "%s: tid %u, mss %u, sndbuf %u, snd_nxt 0x%x, rcv_nxt 0x%x", __func__, toep->tid, toep->params.emss, toep->params.sndbuf, tp ? tp->snd_nxt : 0, tp ? tp->rcv_nxt : 0); if (ulp_mode(toep) == ULP_MODE_TLS) FLOWC_PARAM(ULP_MODE, ulp_mode(toep)); if (toep->tls.fcplenmax != 0) FLOWC_PARAM(TXDATAPLEN_MAX, toep->tls.fcplenmax); if (toep->params.tc_idx != -1) FLOWC_PARAM(SCHEDCLASS, toep->params.tc_idx); #undef FLOWC_PARAM KASSERT(paramidx == nparams, ("nparams mismatch")); txsd->tx_credits = howmany(flowclen, 16); txsd->plen = 0; KASSERT(toep->tx_credits >= txsd->tx_credits && toep->txsd_avail > 0, ("%s: not enough credits (%d)", __func__, toep->tx_credits)); toep->tx_credits -= txsd->tx_credits; if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) toep->txsd_pidx = 0; toep->txsd_avail--; toep->flags |= TPF_FLOWC_WR_SENT; t4_wrq_tx(sc, wr); } #ifdef RATELIMIT /* * Input is Bytes/second (so_max_pacing_rate), chip counts in Kilobits/second. */ static int update_tx_rate_limit(struct adapter *sc, struct toepcb *toep, u_int Bps) { int tc_idx, rc; const u_int kbps = (u_int) (uint64_t)Bps * 8ULL / 1000; const int port_id = toep->vi->pi->port_id; CTR3(KTR_CXGBE, "%s: tid %u, rate %uKbps", __func__, toep->tid, kbps); if (kbps == 0) { /* unbind */ tc_idx = -1; } else { rc = t4_reserve_cl_rl_kbps(sc, port_id, kbps, &tc_idx); if (rc != 0) return (rc); MPASS(tc_idx >= 0 && tc_idx < sc->params.nsched_cls); } if (toep->params.tc_idx != tc_idx) { struct wrqe *wr; struct fw_flowc_wr *flowc; int nparams = 1, flowclen, flowclen16; struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx]; flowclen = sizeof(*flowc) + nparams * sizeof(struct fw_flowc_mnemval); flowclen16 = howmany(flowclen, 16); if (toep->tx_credits < flowclen16 || toep->txsd_avail == 0 || (wr = alloc_wrqe(roundup2(flowclen, 16), &toep->ofld_txq->wrq)) == NULL) { if (tc_idx >= 0) t4_release_cl_rl(sc, port_id, tc_idx); return (ENOMEM); } flowc = wrtod(wr); memset(flowc, 0, wr->wr_len); flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) | V_FW_FLOWC_WR_NPARAMS(nparams)); flowc->flowid_len16 = htonl(V_FW_WR_LEN16(flowclen16) | V_FW_WR_FLOWID(toep->tid)); flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_SCHEDCLASS; if (tc_idx == -1) flowc->mnemval[0].val = htobe32(0xff); else flowc->mnemval[0].val = htobe32(tc_idx); txsd->tx_credits = flowclen16; txsd->plen = 0; toep->tx_credits -= txsd->tx_credits; if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) toep->txsd_pidx = 0; toep->txsd_avail--; t4_wrq_tx(sc, wr); } if (toep->params.tc_idx >= 0) t4_release_cl_rl(sc, port_id, toep->params.tc_idx); toep->params.tc_idx = tc_idx; return (0); } #endif void send_reset(struct adapter *sc, struct toepcb *toep, uint32_t snd_nxt) { struct wrqe *wr; struct cpl_abort_req *req; int tid = toep->tid; struct inpcb *inp = toep->inp; struct tcpcb *tp = intotcpcb(inp); /* don't use if INP_DROPPED */ INP_WLOCK_ASSERT(inp); CTR6(KTR_CXGBE, "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x%s", __func__, toep->tid, inp->inp_flags & INP_DROPPED ? "inp dropped" : tcpstates[tp->t_state], toep->flags, inp->inp_flags, toep->flags & TPF_ABORT_SHUTDOWN ? " (abort already in progress)" : ""); if (toep->flags & TPF_ABORT_SHUTDOWN) return; /* abort already in progress */ toep->flags |= TPF_ABORT_SHUTDOWN; KASSERT(toep->flags & TPF_FLOWC_WR_SENT, ("%s: flowc_wr not sent for tid %d.", __func__, tid)); wr = alloc_wrqe(sizeof(*req), &toep->ofld_txq->wrq); if (wr == NULL) { /* XXX */ panic("%s: allocation failure.", __func__); } req = wrtod(wr); INIT_TP_WR_MIT_CPL(req, CPL_ABORT_REQ, tid); if (inp->inp_flags & INP_DROPPED) req->rsvd0 = htobe32(snd_nxt); else req->rsvd0 = htobe32(tp->snd_nxt); req->rsvd1 = !(toep->flags & TPF_TX_DATA_SENT); req->cmd = CPL_ABORT_SEND_RST; /* * XXX: What's the correct way to tell that the inp hasn't been detached * from its socket? Should I even be flushing the snd buffer here? */ - if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) { + if ((inp->inp_flags & INP_DROPPED) == 0) { struct socket *so = inp->inp_socket; if (so != NULL) /* because I'm not sure. See comment above */ sbflush(&so->so_snd); } t4_l2t_send(sc, wr, toep->l2te); } /* * Called when a connection is established to translate the TCP options * reported by HW to FreeBSD's native format. */ static void assign_rxopt(struct tcpcb *tp, uint16_t opt) { struct toepcb *toep = tp->t_toe; struct inpcb *inp = tp->t_inpcb; struct adapter *sc = td_adapter(toep->td); INP_LOCK_ASSERT(inp); toep->params.mtu_idx = G_TCPOPT_MSS(opt); tp->t_maxseg = sc->params.mtus[toep->params.mtu_idx]; if (inp->inp_inc.inc_flags & INC_ISIPV6) tp->t_maxseg -= sizeof(struct ip6_hdr) + sizeof(struct tcphdr); else tp->t_maxseg -= sizeof(struct ip) + sizeof(struct tcphdr); toep->params.emss = tp->t_maxseg; if (G_TCPOPT_TSTAMP(opt)) { toep->params.tstamp = 1; toep->params.emss -= TCPOLEN_TSTAMP_APPA; tp->t_flags |= TF_RCVD_TSTMP; /* timestamps ok */ tp->ts_recent = 0; /* hmmm */ tp->ts_recent_age = tcp_ts_getticks(); } else toep->params.tstamp = 0; if (G_TCPOPT_SACK(opt)) { toep->params.sack = 1; tp->t_flags |= TF_SACK_PERMIT; /* should already be set */ } else { toep->params.sack = 0; tp->t_flags &= ~TF_SACK_PERMIT; /* sack disallowed by peer */ } if (G_TCPOPT_WSCALE_OK(opt)) tp->t_flags |= TF_RCVD_SCALE; /* Doing window scaling? */ if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == (TF_RCVD_SCALE | TF_REQ_SCALE)) { tp->rcv_scale = tp->request_r_scale; tp->snd_scale = G_TCPOPT_SND_WSCALE(opt); } else toep->params.wscale = 0; CTR6(KTR_CXGBE, "assign_rxopt: tid %d, mtu_idx %u, emss %u, ts %u, sack %u, wscale %u", toep->tid, toep->params.mtu_idx, toep->params.emss, toep->params.tstamp, toep->params.sack, toep->params.wscale); } /* * Completes some final bits of initialization for just established connections * and changes their state to TCPS_ESTABLISHED. * * The ISNs are from the exchange of SYNs. */ void make_established(struct toepcb *toep, uint32_t iss, uint32_t irs, uint16_t opt) { struct inpcb *inp = toep->inp; struct socket *so = inp->inp_socket; struct tcpcb *tp = intotcpcb(inp); uint16_t tcpopt = be16toh(opt); INP_WLOCK_ASSERT(inp); KASSERT(tp->t_state == TCPS_SYN_SENT || tp->t_state == TCPS_SYN_RECEIVED, ("%s: TCP state %s", __func__, tcpstates[tp->t_state])); CTR6(KTR_CXGBE, "%s: tid %d, so %p, inp %p, tp %p, toep %p", __func__, toep->tid, so, inp, tp, toep); tcp_state_change(tp, TCPS_ESTABLISHED); tp->t_starttime = ticks; TCPSTAT_INC(tcps_connects); tp->irs = irs; tcp_rcvseqinit(tp); tp->rcv_wnd = (u_int)toep->params.opt0_bufsize << 10; tp->rcv_adv += tp->rcv_wnd; tp->last_ack_sent = tp->rcv_nxt; tp->iss = iss; tcp_sendseqinit(tp); tp->snd_una = iss + 1; tp->snd_nxt = iss + 1; tp->snd_max = iss + 1; assign_rxopt(tp, tcpopt); send_flowc_wr(toep, tp); soisconnected(so); if (ulp_mode(toep) == ULP_MODE_TLS) tls_establish(toep); } int send_rx_credits(struct adapter *sc, struct toepcb *toep, int credits) { struct wrqe *wr; struct cpl_rx_data_ack *req; uint32_t dack = F_RX_DACK_CHANGE | V_RX_DACK_MODE(1); KASSERT(credits >= 0, ("%s: %d credits", __func__, credits)); wr = alloc_wrqe(sizeof(*req), toep->ctrlq); if (wr == NULL) return (0); req = wrtod(wr); INIT_TP_WR_MIT_CPL(req, CPL_RX_DATA_ACK, toep->tid); req->credit_dack = htobe32(dack | V_RX_CREDITS(credits)); t4_wrq_tx(sc, wr); return (credits); } void send_rx_modulate(struct adapter *sc, struct toepcb *toep) { struct wrqe *wr; struct cpl_rx_data_ack *req; wr = alloc_wrqe(sizeof(*req), toep->ctrlq); if (wr == NULL) return; req = wrtod(wr); INIT_TP_WR_MIT_CPL(req, CPL_RX_DATA_ACK, toep->tid); req->credit_dack = htobe32(F_RX_MODULATE_RX); t4_wrq_tx(sc, wr); } void t4_rcvd_locked(struct toedev *tod, struct tcpcb *tp) { struct adapter *sc = tod->tod_softc; struct inpcb *inp = tp->t_inpcb; struct socket *so = inp->inp_socket; struct sockbuf *sb = &so->so_rcv; struct toepcb *toep = tp->t_toe; int rx_credits; INP_WLOCK_ASSERT(inp); SOCKBUF_LOCK_ASSERT(sb); rx_credits = sbspace(sb) > tp->rcv_wnd ? sbspace(sb) - tp->rcv_wnd : 0; if (rx_credits > 0 && (tp->rcv_wnd <= 32 * 1024 || rx_credits >= 64 * 1024 || (rx_credits >= 16 * 1024 && tp->rcv_wnd <= 128 * 1024) || sbused(sb) + tp->rcv_wnd < sb->sb_lowat)) { rx_credits = send_rx_credits(sc, toep, rx_credits); tp->rcv_wnd += rx_credits; tp->rcv_adv += rx_credits; } else if (toep->flags & TPF_FORCE_CREDITS) send_rx_modulate(sc, toep); } void t4_rcvd(struct toedev *tod, struct tcpcb *tp) { struct inpcb *inp = tp->t_inpcb; struct socket *so = inp->inp_socket; struct sockbuf *sb = &so->so_rcv; SOCKBUF_LOCK(sb); t4_rcvd_locked(tod, tp); SOCKBUF_UNLOCK(sb); } /* * Close a connection by sending a CPL_CLOSE_CON_REQ message. */ int t4_close_conn(struct adapter *sc, struct toepcb *toep) { struct wrqe *wr; struct cpl_close_con_req *req; unsigned int tid = toep->tid; CTR3(KTR_CXGBE, "%s: tid %u%s", __func__, toep->tid, toep->flags & TPF_FIN_SENT ? ", IGNORED" : ""); if (toep->flags & TPF_FIN_SENT) return (0); KASSERT(toep->flags & TPF_FLOWC_WR_SENT, ("%s: flowc_wr not sent for tid %u.", __func__, tid)); wr = alloc_wrqe(sizeof(*req), &toep->ofld_txq->wrq); if (wr == NULL) { /* XXX */ panic("%s: allocation failure.", __func__); } req = wrtod(wr); req->wr.wr_hi = htonl(V_FW_WR_OP(FW_TP_WR) | V_FW_WR_IMMDLEN(sizeof(*req) - sizeof(req->wr))); req->wr.wr_mid = htonl(V_FW_WR_LEN16(howmany(sizeof(*req), 16)) | V_FW_WR_FLOWID(tid)); req->wr.wr_lo = cpu_to_be64(0); OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_CLOSE_CON_REQ, tid)); req->rsvd = 0; toep->flags |= TPF_FIN_SENT; toep->flags &= ~TPF_SEND_FIN; t4_l2t_send(sc, wr, toep->l2te); return (0); } #define MAX_OFLD_TX_CREDITS (SGE_MAX_WR_LEN / 16) #define MIN_OFLD_TX_CREDITS (howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16)) #define MIN_ISO_TX_CREDITS (howmany(sizeof(struct cpl_tx_data_iso), 16)) #define MIN_TX_CREDITS(iso) \ (MIN_OFLD_TX_CREDITS + ((iso) ? MIN_ISO_TX_CREDITS : 0)) /* Maximum amount of immediate data we could stuff in a WR */ static inline int max_imm_payload(int tx_credits, int iso) { const int iso_cpl_size = iso ? sizeof(struct cpl_tx_data_iso) : 0; const int n = 1; /* Use no more than one desc for imm. data WR */ KASSERT(tx_credits >= 0 && tx_credits <= MAX_OFLD_TX_CREDITS, ("%s: %d credits", __func__, tx_credits)); if (tx_credits < MIN_TX_CREDITS(iso)) return (0); if (tx_credits >= (n * EQ_ESIZE) / 16) return ((n * EQ_ESIZE) - sizeof(struct fw_ofld_tx_data_wr) - iso_cpl_size); else return (tx_credits * 16 - sizeof(struct fw_ofld_tx_data_wr) - iso_cpl_size); } /* Maximum number of SGL entries we could stuff in a WR */ static inline int max_dsgl_nsegs(int tx_credits, int iso) { int nseg = 1; /* ulptx_sgl has room for 1, rest ulp_tx_sge_pair */ int sge_pair_credits = tx_credits - MIN_TX_CREDITS(iso); KASSERT(tx_credits >= 0 && tx_credits <= MAX_OFLD_TX_CREDITS, ("%s: %d credits", __func__, tx_credits)); if (tx_credits < MIN_TX_CREDITS(iso)) return (0); nseg += 2 * (sge_pair_credits * 16 / 24); if ((sge_pair_credits * 16) % 24 == 16) nseg++; return (nseg); } static inline void write_tx_wr(void *dst, struct toepcb *toep, int fw_wr_opcode, unsigned int immdlen, unsigned int plen, uint8_t credits, int shove, int ulp_submode) { struct fw_ofld_tx_data_wr *txwr = dst; txwr->op_to_immdlen = htobe32(V_WR_OP(fw_wr_opcode) | V_FW_WR_IMMDLEN(immdlen)); txwr->flowid_len16 = htobe32(V_FW_WR_FLOWID(toep->tid) | V_FW_WR_LEN16(credits)); txwr->lsodisable_to_flags = htobe32(V_TX_ULP_MODE(ulp_mode(toep)) | V_TX_ULP_SUBMODE(ulp_submode) | V_TX_URG(0) | V_TX_SHOVE(shove)); txwr->plen = htobe32(plen); if (toep->params.tx_align > 0) { if (plen < 2 * toep->params.emss) txwr->lsodisable_to_flags |= htobe32(F_FW_OFLD_TX_DATA_WR_LSODISABLE); else txwr->lsodisable_to_flags |= htobe32(F_FW_OFLD_TX_DATA_WR_ALIGNPLD | (toep->params.nagle == 0 ? 0 : F_FW_OFLD_TX_DATA_WR_ALIGNPLDSHOVE)); } } /* * Generate a DSGL from a starting mbuf. The total number of segments and the * maximum segments in any one mbuf are provided. */ static void write_tx_sgl(void *dst, struct mbuf *start, struct mbuf *stop, int nsegs, int n) { struct mbuf *m; struct ulptx_sgl *usgl = dst; int i, j, rc; struct sglist sg; struct sglist_seg segs[n]; KASSERT(nsegs > 0, ("%s: nsegs 0", __func__)); sglist_init(&sg, n, segs); usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) | V_ULPTX_NSGE(nsegs)); i = -1; for (m = start; m != stop; m = m->m_next) { if (m->m_flags & M_EXTPG) rc = sglist_append_mbuf_epg(&sg, m, mtod(m, vm_offset_t), m->m_len); else rc = sglist_append(&sg, mtod(m, void *), m->m_len); if (__predict_false(rc != 0)) panic("%s: sglist_append %d", __func__, rc); for (j = 0; j < sg.sg_nseg; i++, j++) { if (i < 0) { usgl->len0 = htobe32(segs[j].ss_len); usgl->addr0 = htobe64(segs[j].ss_paddr); } else { usgl->sge[i / 2].len[i & 1] = htobe32(segs[j].ss_len); usgl->sge[i / 2].addr[i & 1] = htobe64(segs[j].ss_paddr); } #ifdef INVARIANTS nsegs--; #endif } sglist_reset(&sg); } if (i & 1) usgl->sge[i / 2].len[1] = htobe32(0); KASSERT(nsegs == 0, ("%s: nsegs %d, start %p, stop %p", __func__, nsegs, start, stop)); } /* * Max number of SGL entries an offload tx work request can have. This is 41 * (1 + 40) for a full 512B work request. * fw_ofld_tx_data_wr(16B) + ulptx_sgl(16B, 1) + ulptx_sge_pair(480B, 40) */ #define OFLD_SGL_LEN (41) /* * Send data and/or a FIN to the peer. * * The socket's so_snd buffer consists of a stream of data starting with sb_mb * and linked together with m_next. sb_sndptr, if set, is the last mbuf that * was transmitted. * * drop indicates the number of bytes that should be dropped from the head of * the send buffer. It is an optimization that lets do_fw4_ack avoid creating * contention on the send buffer lock (before this change it used to do * sowwakeup and then t4_push_frames right after that when recovering from tx * stalls). When drop is set this function MUST drop the bytes and wake up any * writers. */ void t4_push_frames(struct adapter *sc, struct toepcb *toep, int drop) { struct mbuf *sndptr, *m, *sb_sndptr; struct fw_ofld_tx_data_wr *txwr; struct wrqe *wr; u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf; struct inpcb *inp = toep->inp; struct tcpcb *tp = intotcpcb(inp); struct socket *so = inp->inp_socket; struct sockbuf *sb = &so->so_snd; int tx_credits, shove, compl, sowwakeup; struct ofld_tx_sdesc *txsd; bool nomap_mbuf_seen; INP_WLOCK_ASSERT(inp); KASSERT(toep->flags & TPF_FLOWC_WR_SENT, ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid)); KASSERT(ulp_mode(toep) == ULP_MODE_NONE || ulp_mode(toep) == ULP_MODE_TCPDDP || ulp_mode(toep) == ULP_MODE_TLS || ulp_mode(toep) == ULP_MODE_RDMA, ("%s: ulp_mode %u for toep %p", __func__, ulp_mode(toep), toep)); #ifdef VERBOSE_TRACES CTR5(KTR_CXGBE, "%s: tid %d toep flags %#x tp flags %#x drop %d", __func__, toep->tid, toep->flags, tp->t_flags, drop); #endif if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN)) return; #ifdef RATELIMIT if (__predict_false(inp->inp_flags2 & INP_RATE_LIMIT_CHANGED) && (update_tx_rate_limit(sc, toep, so->so_max_pacing_rate) == 0)) { inp->inp_flags2 &= ~INP_RATE_LIMIT_CHANGED; } #endif /* * This function doesn't resume by itself. Someone else must clear the * flag and call this function. */ if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) { KASSERT(drop == 0, ("%s: drop (%d) != 0 but tx is suspended", __func__, drop)); return; } txsd = &toep->txsd[toep->txsd_pidx]; do { tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS); max_imm = max_imm_payload(tx_credits, 0); max_nsegs = max_dsgl_nsegs(tx_credits, 0); SOCKBUF_LOCK(sb); sowwakeup = drop; if (drop) { sbdrop_locked(sb, drop); drop = 0; } sb_sndptr = sb->sb_sndptr; sndptr = sb_sndptr ? sb_sndptr->m_next : sb->sb_mb; plen = 0; nsegs = 0; max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */ nomap_mbuf_seen = false; for (m = sndptr; m != NULL; m = m->m_next) { int n; if ((m->m_flags & M_NOTAVAIL) != 0) break; if (m->m_flags & M_EXTPG) { #ifdef KERN_TLS if (m->m_epg_tls != NULL) { toep->flags |= TPF_KTLS; if (plen == 0) { SOCKBUF_UNLOCK(sb); t4_push_ktls(sc, toep, 0); return; } break; } #endif n = sglist_count_mbuf_epg(m, mtod(m, vm_offset_t), m->m_len); } else n = sglist_count(mtod(m, void *), m->m_len); nsegs += n; plen += m->m_len; /* This mbuf sent us _over_ the nsegs limit, back out */ if (plen > max_imm && nsegs > max_nsegs) { nsegs -= n; plen -= m->m_len; if (plen == 0) { /* Too few credits */ toep->flags |= TPF_TX_SUSPENDED; if (sowwakeup) { if (!TAILQ_EMPTY( &toep->aiotx_jobq)) t4_aiotx_queue_toep(so, toep); sowwakeup_locked(so); } else SOCKBUF_UNLOCK(sb); SOCKBUF_UNLOCK_ASSERT(sb); return; } break; } if (m->m_flags & M_EXTPG) nomap_mbuf_seen = true; if (max_nsegs_1mbuf < n) max_nsegs_1mbuf = n; sb_sndptr = m; /* new sb->sb_sndptr if all goes well */ /* This mbuf put us right at the max_nsegs limit */ if (plen > max_imm && nsegs == max_nsegs) { m = m->m_next; break; } } if (sbused(sb) > sb->sb_hiwat * 5 / 8 && toep->plen_nocompl + plen >= sb->sb_hiwat / 4) compl = 1; else compl = 0; if (sb->sb_flags & SB_AUTOSIZE && V_tcp_do_autosndbuf && sb->sb_hiwat < V_tcp_autosndbuf_max && sbused(sb) >= sb->sb_hiwat * 7 / 8) { int newsize = min(sb->sb_hiwat + V_tcp_autosndbuf_inc, V_tcp_autosndbuf_max); if (!sbreserve_locked(so, SO_SND, newsize, NULL)) sb->sb_flags &= ~SB_AUTOSIZE; else sowwakeup = 1; /* room available */ } if (sowwakeup) { if (!TAILQ_EMPTY(&toep->aiotx_jobq)) t4_aiotx_queue_toep(so, toep); sowwakeup_locked(so); } else SOCKBUF_UNLOCK(sb); SOCKBUF_UNLOCK_ASSERT(sb); /* nothing to send */ if (plen == 0) { KASSERT(m == NULL || (m->m_flags & M_NOTAVAIL) != 0, ("%s: nothing to send, but m != NULL is ready", __func__)); break; } if (__predict_false(toep->flags & TPF_FIN_SENT)) panic("%s: excess tx.", __func__); shove = m == NULL && !(tp->t_flags & TF_MORETOCOME); if (plen <= max_imm && !nomap_mbuf_seen) { /* Immediate data tx */ wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16), &toep->ofld_txq->wrq); if (wr == NULL) { /* XXX: how will we recover from this? */ toep->flags |= TPF_TX_SUSPENDED; return; } txwr = wrtod(wr); credits = howmany(wr->wr_len, 16); write_tx_wr(txwr, toep, FW_OFLD_TX_DATA_WR, plen, plen, credits, shove, 0); m_copydata(sndptr, 0, plen, (void *)(txwr + 1)); nsegs = 0; } else { int wr_len; /* DSGL tx */ wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) + ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8; wr = alloc_wrqe(roundup2(wr_len, 16), &toep->ofld_txq->wrq); if (wr == NULL) { /* XXX: how will we recover from this? */ toep->flags |= TPF_TX_SUSPENDED; return; } txwr = wrtod(wr); credits = howmany(wr_len, 16); write_tx_wr(txwr, toep, FW_OFLD_TX_DATA_WR, 0, plen, credits, shove, 0); write_tx_sgl(txwr + 1, sndptr, m, nsegs, max_nsegs_1mbuf); if (wr_len & 0xf) { uint64_t *pad = (uint64_t *) ((uintptr_t)txwr + wr_len); *pad = 0; } } KASSERT(toep->tx_credits >= credits, ("%s: not enough credits", __func__)); toep->tx_credits -= credits; toep->tx_nocompl += credits; toep->plen_nocompl += plen; if (toep->tx_credits <= toep->tx_total * 3 / 8 && toep->tx_nocompl >= toep->tx_total / 4) compl = 1; if (compl || ulp_mode(toep) == ULP_MODE_RDMA) { txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL); toep->tx_nocompl = 0; toep->plen_nocompl = 0; } tp->snd_nxt += plen; tp->snd_max += plen; SOCKBUF_LOCK(sb); KASSERT(sb_sndptr, ("%s: sb_sndptr is NULL", __func__)); sb->sb_sndptr = sb_sndptr; SOCKBUF_UNLOCK(sb); toep->flags |= TPF_TX_DATA_SENT; if (toep->tx_credits < MIN_OFLD_TX_CREDITS) toep->flags |= TPF_TX_SUSPENDED; KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__)); txsd->plen = plen; txsd->tx_credits = credits; txsd++; if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) { toep->txsd_pidx = 0; txsd = &toep->txsd[0]; } toep->txsd_avail--; t4_l2t_send(sc, wr, toep->l2te); } while (m != NULL && (m->m_flags & M_NOTAVAIL) == 0); /* Send a FIN if requested, but only if there's no more data to send */ if (m == NULL && toep->flags & TPF_SEND_FIN) t4_close_conn(sc, toep); } static inline void rqdrop_locked(struct mbufq *q, int plen) { struct mbuf *m; while (plen > 0) { m = mbufq_dequeue(q); /* Too many credits. */ MPASS(m != NULL); M_ASSERTPKTHDR(m); /* Partial credits. */ MPASS(plen >= m->m_pkthdr.len); plen -= m->m_pkthdr.len; m_freem(m); } } /* * Not a bit in the TCB, but is a bit in the ulp_submode field of the * CPL_TX_DATA flags field in FW_ISCSI_TX_DATA_WR. */ #define ULP_ISO G_TX_ULP_SUBMODE(F_FW_ISCSI_TX_DATA_WR_ULPSUBMODE_ISO) static void write_tx_data_iso(void *dst, u_int ulp_submode, uint8_t flags, uint16_t mss, int len, int npdu) { struct cpl_tx_data_iso *cpl; unsigned int burst_size; unsigned int last; /* * The firmware will set the 'F' bit on the last PDU when * either condition is true: * * - this large PDU is marked as the "last" slice * * - the amount of data payload bytes equals the burst_size * * The strategy used here is to always set the burst_size * artificially high (len includes the size of the template * BHS) and only set the "last" flag if the original PDU had * 'F' set. */ burst_size = len; last = !!(flags & CXGBE_ISO_F); cpl = (struct cpl_tx_data_iso *)dst; cpl->op_to_scsi = htonl(V_CPL_TX_DATA_ISO_OP(CPL_TX_DATA_ISO) | V_CPL_TX_DATA_ISO_FIRST(1) | V_CPL_TX_DATA_ISO_LAST(last) | V_CPL_TX_DATA_ISO_CPLHDRLEN(0) | V_CPL_TX_DATA_ISO_HDRCRC(!!(ulp_submode & ULP_CRC_HEADER)) | V_CPL_TX_DATA_ISO_PLDCRC(!!(ulp_submode & ULP_CRC_DATA)) | V_CPL_TX_DATA_ISO_IMMEDIATE(0) | V_CPL_TX_DATA_ISO_SCSI(CXGBE_ISO_TYPE(flags))); cpl->ahs_len = 0; cpl->mpdu = htons(DIV_ROUND_UP(mss, 4)); cpl->burst_size = htonl(DIV_ROUND_UP(burst_size, 4)); cpl->len = htonl(len); cpl->reserved2_seglen_offset = htonl(0); cpl->datasn_offset = htonl(0); cpl->buffer_offset = htonl(0); cpl->reserved3 = 0; } static struct wrqe * write_iscsi_mbuf_wr(struct toepcb *toep, struct mbuf *sndptr) { struct mbuf *m; struct fw_ofld_tx_data_wr *txwr; struct cpl_tx_data_iso *cpl_iso; void *p; struct wrqe *wr; u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf; u_int adjusted_plen, imm_data, ulp_submode; struct inpcb *inp = toep->inp; struct tcpcb *tp = intotcpcb(inp); int tx_credits, shove, npdu, wr_len; uint16_t iso_mss; static const u_int ulp_extra_len[] = {0, 4, 4, 8}; bool iso, nomap_mbuf_seen; M_ASSERTPKTHDR(sndptr); tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS); if (mbuf_raw_wr(sndptr)) { plen = sndptr->m_pkthdr.len; KASSERT(plen <= SGE_MAX_WR_LEN, ("raw WR len %u is greater than max WR len", plen)); if (plen > tx_credits * 16) return (NULL); wr = alloc_wrqe(roundup2(plen, 16), &toep->ofld_txq->wrq); if (__predict_false(wr == NULL)) return (NULL); m_copydata(sndptr, 0, plen, wrtod(wr)); return (wr); } iso = mbuf_iscsi_iso(sndptr); max_imm = max_imm_payload(tx_credits, iso); max_nsegs = max_dsgl_nsegs(tx_credits, iso); iso_mss = mbuf_iscsi_iso_mss(sndptr); plen = 0; nsegs = 0; max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */ nomap_mbuf_seen = false; for (m = sndptr; m != NULL; m = m->m_next) { int n; if (m->m_flags & M_EXTPG) n = sglist_count_mbuf_epg(m, mtod(m, vm_offset_t), m->m_len); else n = sglist_count(mtod(m, void *), m->m_len); nsegs += n; plen += m->m_len; /* * This mbuf would send us _over_ the nsegs limit. * Suspend tx because the PDU can't be sent out. */ if ((nomap_mbuf_seen || plen > max_imm) && nsegs > max_nsegs) return (NULL); if (m->m_flags & M_EXTPG) nomap_mbuf_seen = true; if (max_nsegs_1mbuf < n) max_nsegs_1mbuf = n; } if (__predict_false(toep->flags & TPF_FIN_SENT)) panic("%s: excess tx.", __func__); /* * We have a PDU to send. All of it goes out in one WR so 'm' * is NULL. A PDU's length is always a multiple of 4. */ MPASS(m == NULL); MPASS((plen & 3) == 0); MPASS(sndptr->m_pkthdr.len == plen); shove = !(tp->t_flags & TF_MORETOCOME); /* * plen doesn't include header and data digests, which are * generated and inserted in the right places by the TOE, but * they do occupy TCP sequence space and need to be accounted * for. */ ulp_submode = mbuf_ulp_submode(sndptr); MPASS(ulp_submode < nitems(ulp_extra_len)); npdu = iso ? howmany(plen - ISCSI_BHS_SIZE, iso_mss) : 1; adjusted_plen = plen + ulp_extra_len[ulp_submode] * npdu; if (iso) adjusted_plen += ISCSI_BHS_SIZE * (npdu - 1); wr_len = sizeof(*txwr); if (iso) wr_len += sizeof(struct cpl_tx_data_iso); if (plen <= max_imm && !nomap_mbuf_seen) { /* Immediate data tx */ imm_data = plen; wr_len += plen; nsegs = 0; } else { /* DSGL tx */ imm_data = 0; wr_len += sizeof(struct ulptx_sgl) + ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8; } wr = alloc_wrqe(roundup2(wr_len, 16), &toep->ofld_txq->wrq); if (wr == NULL) { /* XXX: how will we recover from this? */ return (NULL); } txwr = wrtod(wr); credits = howmany(wr->wr_len, 16); if (iso) { write_tx_wr(txwr, toep, FW_ISCSI_TX_DATA_WR, imm_data + sizeof(struct cpl_tx_data_iso), adjusted_plen, credits, shove, ulp_submode | ULP_ISO); cpl_iso = (struct cpl_tx_data_iso *)(txwr + 1); MPASS(plen == sndptr->m_pkthdr.len); write_tx_data_iso(cpl_iso, ulp_submode, mbuf_iscsi_iso_flags(sndptr), iso_mss, plen, npdu); p = cpl_iso + 1; } else { write_tx_wr(txwr, toep, FW_OFLD_TX_DATA_WR, imm_data, adjusted_plen, credits, shove, ulp_submode); p = txwr + 1; } if (imm_data != 0) { m_copydata(sndptr, 0, plen, p); } else { write_tx_sgl(p, sndptr, m, nsegs, max_nsegs_1mbuf); if (wr_len & 0xf) { uint64_t *pad = (uint64_t *)((uintptr_t)txwr + wr_len); *pad = 0; } } KASSERT(toep->tx_credits >= credits, ("%s: not enough credits: credits %u " "toep->tx_credits %u tx_credits %u nsegs %u " "max_nsegs %u iso %d", __func__, credits, toep->tx_credits, tx_credits, nsegs, max_nsegs, iso)); tp->snd_nxt += adjusted_plen; tp->snd_max += adjusted_plen; counter_u64_add(toep->ofld_txq->tx_iscsi_pdus, npdu); counter_u64_add(toep->ofld_txq->tx_iscsi_octets, plen); if (iso) counter_u64_add(toep->ofld_txq->tx_iscsi_iso_wrs, 1); return (wr); } void t4_push_pdus(struct adapter *sc, struct toepcb *toep, int drop) { struct mbuf *sndptr, *m; struct fw_wr_hdr *wrhdr; struct wrqe *wr; u_int plen, credits; struct inpcb *inp = toep->inp; struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx]; struct mbufq *pduq = &toep->ulp_pduq; INP_WLOCK_ASSERT(inp); KASSERT(toep->flags & TPF_FLOWC_WR_SENT, ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid)); KASSERT(ulp_mode(toep) == ULP_MODE_ISCSI, ("%s: ulp_mode %u for toep %p", __func__, ulp_mode(toep), toep)); if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN)) return; /* * This function doesn't resume by itself. Someone else must clear the * flag and call this function. */ if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) { KASSERT(drop == 0, ("%s: drop (%d) != 0 but tx is suspended", __func__, drop)); return; } if (drop) { struct socket *so = inp->inp_socket; struct sockbuf *sb = &so->so_snd; int sbu; /* * An unlocked read is ok here as the data should only * transition from a non-zero value to either another * non-zero value or zero. Once it is zero it should * stay zero. */ if (__predict_false(sbused(sb)) > 0) { SOCKBUF_LOCK(sb); sbu = sbused(sb); if (sbu > 0) { /* * The data transmitted before the * tid's ULP mode changed to ISCSI is * still in so_snd. Incoming credits * should account for so_snd first. */ sbdrop_locked(sb, min(sbu, drop)); drop -= min(sbu, drop); } sowwakeup_locked(so); /* unlocks so_snd */ } rqdrop_locked(&toep->ulp_pdu_reclaimq, drop); } while ((sndptr = mbufq_first(pduq)) != NULL) { wr = write_iscsi_mbuf_wr(toep, sndptr); if (wr == NULL) { toep->flags |= TPF_TX_SUSPENDED; return; } plen = sndptr->m_pkthdr.len; credits = howmany(wr->wr_len, 16); KASSERT(toep->tx_credits >= credits, ("%s: not enough credits", __func__)); m = mbufq_dequeue(pduq); MPASS(m == sndptr); mbufq_enqueue(&toep->ulp_pdu_reclaimq, m); toep->tx_credits -= credits; toep->tx_nocompl += credits; toep->plen_nocompl += plen; /* * Ensure there are enough credits for a full-sized WR * as page pod WRs can be full-sized. */ if (toep->tx_credits <= SGE_MAX_WR_LEN * 5 / 4 && toep->tx_nocompl >= toep->tx_total / 4) { wrhdr = wrtod(wr); wrhdr->hi |= htobe32(F_FW_WR_COMPL); toep->tx_nocompl = 0; toep->plen_nocompl = 0; } toep->flags |= TPF_TX_DATA_SENT; if (toep->tx_credits < MIN_OFLD_TX_CREDITS) toep->flags |= TPF_TX_SUSPENDED; KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__)); txsd->plen = plen; txsd->tx_credits = credits; txsd++; if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) { toep->txsd_pidx = 0; txsd = &toep->txsd[0]; } toep->txsd_avail--; t4_l2t_send(sc, wr, toep->l2te); } /* Send a FIN if requested, but only if there are no more PDUs to send */ if (mbufq_first(pduq) == NULL && toep->flags & TPF_SEND_FIN) t4_close_conn(sc, toep); } static inline void t4_push_data(struct adapter *sc, struct toepcb *toep, int drop) { if (ulp_mode(toep) == ULP_MODE_ISCSI) t4_push_pdus(sc, toep, drop); else if (toep->flags & TPF_KTLS) t4_push_ktls(sc, toep, drop); else t4_push_frames(sc, toep, drop); } int t4_tod_output(struct toedev *tod, struct tcpcb *tp) { struct adapter *sc = tod->tod_softc; #ifdef INVARIANTS struct inpcb *inp = tp->t_inpcb; #endif struct toepcb *toep = tp->t_toe; INP_WLOCK_ASSERT(inp); KASSERT((inp->inp_flags & INP_DROPPED) == 0, ("%s: inp %p dropped.", __func__, inp)); KASSERT(toep != NULL, ("%s: toep is NULL", __func__)); t4_push_data(sc, toep, 0); return (0); } int t4_send_fin(struct toedev *tod, struct tcpcb *tp) { struct adapter *sc = tod->tod_softc; #ifdef INVARIANTS struct inpcb *inp = tp->t_inpcb; #endif struct toepcb *toep = tp->t_toe; INP_WLOCK_ASSERT(inp); KASSERT((inp->inp_flags & INP_DROPPED) == 0, ("%s: inp %p dropped.", __func__, inp)); KASSERT(toep != NULL, ("%s: toep is NULL", __func__)); toep->flags |= TPF_SEND_FIN; if (tp->t_state >= TCPS_ESTABLISHED) t4_push_data(sc, toep, 0); return (0); } int t4_send_rst(struct toedev *tod, struct tcpcb *tp) { struct adapter *sc = tod->tod_softc; #if defined(INVARIANTS) struct inpcb *inp = tp->t_inpcb; #endif struct toepcb *toep = tp->t_toe; INP_WLOCK_ASSERT(inp); KASSERT((inp->inp_flags & INP_DROPPED) == 0, ("%s: inp %p dropped.", __func__, inp)); KASSERT(toep != NULL, ("%s: toep is NULL", __func__)); /* hmmmm */ KASSERT(toep->flags & TPF_FLOWC_WR_SENT, ("%s: flowc for tid %u [%s] not sent already", __func__, toep->tid, tcpstates[tp->t_state])); send_reset(sc, toep, 0); return (0); } /* * Peer has sent us a FIN. */ static int do_peer_close(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { struct adapter *sc = iq->adapter; const struct cpl_peer_close *cpl = (const void *)(rss + 1); unsigned int tid = GET_TID(cpl); struct toepcb *toep = lookup_tid(sc, tid); struct inpcb *inp = toep->inp; struct tcpcb *tp = NULL; struct socket *so; struct epoch_tracker et; #ifdef INVARIANTS unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); #endif KASSERT(opcode == CPL_PEER_CLOSE, ("%s: unexpected opcode 0x%x", __func__, opcode)); KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); if (__predict_false(toep->flags & TPF_SYNQE)) { /* * do_pass_establish must have run before do_peer_close and if * this is still a synqe instead of a toepcb then the connection * must be getting aborted. */ MPASS(toep->flags & TPF_ABORT_SHUTDOWN); CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid, toep, toep->flags); return (0); } KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__)); CURVNET_SET(toep->vnet); NET_EPOCH_ENTER(et); INP_WLOCK(inp); tp = intotcpcb(inp); CTR6(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x, ddp_flags 0x%x, inp %p", __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags, toep->ddp.flags, inp); if (toep->flags & TPF_ABORT_SHUTDOWN) goto done; so = inp->inp_socket; socantrcvmore(so); if (ulp_mode(toep) == ULP_MODE_TCPDDP) { DDP_LOCK(toep); if (__predict_false(toep->ddp.flags & (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE))) handle_ddp_close(toep, tp, cpl->rcv_nxt); DDP_UNLOCK(toep); } if (ulp_mode(toep) == ULP_MODE_RDMA || (ulp_mode(toep) == ULP_MODE_ISCSI && chip_id(sc) >= CHELSIO_T6)) { /* * There might be data received via DDP before the FIN * not reported to the driver. Just assume the * sequence number in the CPL is correct as the * sequence number of the FIN. */ } else { KASSERT(tp->rcv_nxt + 1 == be32toh(cpl->rcv_nxt), ("%s: rcv_nxt mismatch: %u %u", __func__, tp->rcv_nxt, be32toh(cpl->rcv_nxt))); } tp->rcv_nxt = be32toh(cpl->rcv_nxt); switch (tp->t_state) { case TCPS_SYN_RECEIVED: tp->t_starttime = ticks; /* FALLTHROUGH */ case TCPS_ESTABLISHED: tcp_state_change(tp, TCPS_CLOSE_WAIT); break; case TCPS_FIN_WAIT_1: tcp_state_change(tp, TCPS_CLOSING); break; case TCPS_FIN_WAIT_2: restore_so_proto(so, inp->inp_vflag & INP_IPV6); tcp_twstart(tp); INP_UNLOCK_ASSERT(inp); /* safe, we have a ref on the inp */ NET_EPOCH_EXIT(et); CURVNET_RESTORE(); INP_WLOCK(inp); final_cpl_received(toep); return (0); default: log(LOG_ERR, "%s: TID %u received CPL_PEER_CLOSE in state %d\n", __func__, tid, tp->t_state); } done: INP_WUNLOCK(inp); NET_EPOCH_EXIT(et); CURVNET_RESTORE(); return (0); } /* * Peer has ACK'd our FIN. */ static int do_close_con_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { struct adapter *sc = iq->adapter; const struct cpl_close_con_rpl *cpl = (const void *)(rss + 1); unsigned int tid = GET_TID(cpl); struct toepcb *toep = lookup_tid(sc, tid); struct inpcb *inp = toep->inp; struct tcpcb *tp = NULL; struct socket *so = NULL; struct epoch_tracker et; #ifdef INVARIANTS unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); #endif KASSERT(opcode == CPL_CLOSE_CON_RPL, ("%s: unexpected opcode 0x%x", __func__, opcode)); KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__)); CURVNET_SET(toep->vnet); NET_EPOCH_ENTER(et); INP_WLOCK(inp); tp = intotcpcb(inp); CTR4(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x", __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags); if (toep->flags & TPF_ABORT_SHUTDOWN) goto done; so = inp->inp_socket; tp->snd_una = be32toh(cpl->snd_nxt) - 1; /* exclude FIN */ switch (tp->t_state) { case TCPS_CLOSING: /* see TCPS_FIN_WAIT_2 in do_peer_close too */ restore_so_proto(so, inp->inp_vflag & INP_IPV6); tcp_twstart(tp); release: INP_UNLOCK_ASSERT(inp); /* safe, we have a ref on the inp */ NET_EPOCH_EXIT(et); CURVNET_RESTORE(); INP_WLOCK(inp); final_cpl_received(toep); /* no more CPLs expected */ return (0); case TCPS_LAST_ACK: if (tcp_close(tp)) INP_WUNLOCK(inp); goto release; case TCPS_FIN_WAIT_1: if (so->so_rcv.sb_state & SBS_CANTRCVMORE) soisdisconnected(so); tcp_state_change(tp, TCPS_FIN_WAIT_2); break; default: log(LOG_ERR, "%s: TID %u received CPL_CLOSE_CON_RPL in state %s\n", __func__, tid, tcpstates[tp->t_state]); } done: INP_WUNLOCK(inp); NET_EPOCH_EXIT(et); CURVNET_RESTORE(); return (0); } void send_abort_rpl(struct adapter *sc, struct sge_ofld_txq *ofld_txq, int tid, int rst_status) { struct wrqe *wr; struct cpl_abort_rpl *cpl; wr = alloc_wrqe(sizeof(*cpl), &ofld_txq->wrq); if (wr == NULL) { /* XXX */ panic("%s: allocation failure.", __func__); } cpl = wrtod(wr); INIT_TP_WR_MIT_CPL(cpl, CPL_ABORT_RPL, tid); cpl->cmd = rst_status; t4_wrq_tx(sc, wr); } static int abort_status_to_errno(struct tcpcb *tp, unsigned int abort_reason) { switch (abort_reason) { case CPL_ERR_BAD_SYN: case CPL_ERR_CONN_RESET: return (tp->t_state == TCPS_CLOSE_WAIT ? EPIPE : ECONNRESET); case CPL_ERR_XMIT_TIMEDOUT: case CPL_ERR_PERSIST_TIMEDOUT: case CPL_ERR_FINWAIT2_TIMEDOUT: case CPL_ERR_KEEPALIVE_TIMEDOUT: return (ETIMEDOUT); default: return (EIO); } } /* * TCP RST from the peer, timeout, or some other such critical error. */ static int do_abort_req(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { struct adapter *sc = iq->adapter; const struct cpl_abort_req_rss *cpl = (const void *)(rss + 1); unsigned int tid = GET_TID(cpl); struct toepcb *toep = lookup_tid(sc, tid); struct sge_ofld_txq *ofld_txq = toep->ofld_txq; struct inpcb *inp; struct tcpcb *tp; struct epoch_tracker et; #ifdef INVARIANTS unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); #endif KASSERT(opcode == CPL_ABORT_REQ_RSS, ("%s: unexpected opcode 0x%x", __func__, opcode)); KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); if (toep->flags & TPF_SYNQE) return (do_abort_req_synqe(iq, rss, m)); KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__)); if (negative_advice(cpl->status)) { CTR4(KTR_CXGBE, "%s: negative advice %d for tid %d (0x%x)", __func__, cpl->status, tid, toep->flags); return (0); /* Ignore negative advice */ } inp = toep->inp; CURVNET_SET(toep->vnet); NET_EPOCH_ENTER(et); /* for tcp_close */ INP_WLOCK(inp); tp = intotcpcb(inp); CTR6(KTR_CXGBE, "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x, status %d", __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags, inp->inp_flags, cpl->status); /* * If we'd initiated an abort earlier the reply to it is responsible for * cleaning up resources. Otherwise we tear everything down right here * right now. We owe the T4 a CPL_ABORT_RPL no matter what. */ if (toep->flags & TPF_ABORT_SHUTDOWN) { INP_WUNLOCK(inp); goto done; } toep->flags |= TPF_ABORT_SHUTDOWN; - if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) { + if ((inp->inp_flags & INP_DROPPED) == 0) { struct socket *so = inp->inp_socket; if (so != NULL) so_error_set(so, abort_status_to_errno(tp, cpl->status)); tp = tcp_close(tp); if (tp == NULL) INP_WLOCK(inp); /* re-acquire */ } final_cpl_received(toep); done: NET_EPOCH_EXIT(et); CURVNET_RESTORE(); send_abort_rpl(sc, ofld_txq, tid, CPL_ABORT_NO_RST); return (0); } /* * Reply to the CPL_ABORT_REQ (send_reset) */ static int do_abort_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { struct adapter *sc = iq->adapter; const struct cpl_abort_rpl_rss *cpl = (const void *)(rss + 1); unsigned int tid = GET_TID(cpl); struct toepcb *toep = lookup_tid(sc, tid); struct inpcb *inp = toep->inp; #ifdef INVARIANTS unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); #endif KASSERT(opcode == CPL_ABORT_RPL_RSS, ("%s: unexpected opcode 0x%x", __func__, opcode)); KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); if (toep->flags & TPF_SYNQE) return (do_abort_rpl_synqe(iq, rss, m)); KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__)); CTR5(KTR_CXGBE, "%s: tid %u, toep %p, inp %p, status %d", __func__, tid, toep, inp, cpl->status); KASSERT(toep->flags & TPF_ABORT_SHUTDOWN, ("%s: wasn't expecting abort reply", __func__)); INP_WLOCK(inp); final_cpl_received(toep); return (0); } static int do_rx_data(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { struct adapter *sc = iq->adapter; const struct cpl_rx_data *cpl = mtod(m, const void *); unsigned int tid = GET_TID(cpl); struct toepcb *toep = lookup_tid(sc, tid); struct inpcb *inp = toep->inp; struct tcpcb *tp; struct socket *so; struct sockbuf *sb; struct epoch_tracker et; int len, rx_credits; uint32_t ddp_placed = 0; if (__predict_false(toep->flags & TPF_SYNQE)) { /* * do_pass_establish must have run before do_rx_data and if this * is still a synqe instead of a toepcb then the connection must * be getting aborted. */ MPASS(toep->flags & TPF_ABORT_SHUTDOWN); CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid, toep, toep->flags); m_freem(m); return (0); } KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__)); /* strip off CPL header */ m_adj(m, sizeof(*cpl)); len = m->m_pkthdr.len; INP_WLOCK(inp); - if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) { + if (inp->inp_flags & INP_DROPPED) { CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x", __func__, tid, len, inp->inp_flags); INP_WUNLOCK(inp); m_freem(m); return (0); } tp = intotcpcb(inp); if (__predict_false(ulp_mode(toep) == ULP_MODE_TLS && toep->flags & TPF_TLS_RECEIVE)) { /* Received "raw" data on a TLS socket. */ CTR3(KTR_CXGBE, "%s: tid %u, raw TLS data (%d bytes)", __func__, tid, len); do_rx_data_tls(cpl, toep, m); return (0); } if (__predict_false(tp->rcv_nxt != be32toh(cpl->seq))) ddp_placed = be32toh(cpl->seq) - tp->rcv_nxt; tp->rcv_nxt += len; if (tp->rcv_wnd < len) { KASSERT(ulp_mode(toep) == ULP_MODE_RDMA, ("%s: negative window size", __func__)); } tp->rcv_wnd -= len; tp->t_rcvtime = ticks; if (ulp_mode(toep) == ULP_MODE_TCPDDP) DDP_LOCK(toep); so = inp_inpcbtosocket(inp); sb = &so->so_rcv; SOCKBUF_LOCK(sb); if (__predict_false(sb->sb_state & SBS_CANTRCVMORE)) { CTR3(KTR_CXGBE, "%s: tid %u, excess rx (%d bytes)", __func__, tid, len); m_freem(m); SOCKBUF_UNLOCK(sb); if (ulp_mode(toep) == ULP_MODE_TCPDDP) DDP_UNLOCK(toep); INP_WUNLOCK(inp); CURVNET_SET(toep->vnet); NET_EPOCH_ENTER(et); INP_WLOCK(inp); tp = tcp_drop(tp, ECONNRESET); if (tp) INP_WUNLOCK(inp); NET_EPOCH_EXIT(et); CURVNET_RESTORE(); return (0); } /* receive buffer autosize */ MPASS(toep->vnet == so->so_vnet); CURVNET_SET(toep->vnet); if (sb->sb_flags & SB_AUTOSIZE && V_tcp_do_autorcvbuf && sb->sb_hiwat < V_tcp_autorcvbuf_max && len > (sbspace(sb) / 8 * 7)) { unsigned int hiwat = sb->sb_hiwat; unsigned int newsize = min(hiwat + sc->tt.autorcvbuf_inc, V_tcp_autorcvbuf_max); if (!sbreserve_locked(so, SO_RCV, newsize, NULL)) sb->sb_flags &= ~SB_AUTOSIZE; } if (ulp_mode(toep) == ULP_MODE_TCPDDP) { int changed = !(toep->ddp.flags & DDP_ON) ^ cpl->ddp_off; if (toep->ddp.waiting_count != 0 || toep->ddp.active_count != 0) CTR3(KTR_CXGBE, "%s: tid %u, non-ddp rx (%d bytes)", __func__, tid, len); if (changed) { if (toep->ddp.flags & DDP_SC_REQ) toep->ddp.flags ^= DDP_ON | DDP_SC_REQ; else { KASSERT(cpl->ddp_off == 1, ("%s: DDP switched on by itself.", __func__)); /* Fell out of DDP mode */ toep->ddp.flags &= ~DDP_ON; CTR1(KTR_CXGBE, "%s: fell out of DDP mode", __func__); insert_ddp_data(toep, ddp_placed); } } if (toep->ddp.flags & DDP_ON) { /* * CPL_RX_DATA with DDP on can only be an indicate. * Start posting queued AIO requests via DDP. The * payload that arrived in this indicate is appended * to the socket buffer as usual. */ handle_ddp_indicate(toep); } } sbappendstream_locked(sb, m, 0); rx_credits = sbspace(sb) > tp->rcv_wnd ? sbspace(sb) - tp->rcv_wnd : 0; if (rx_credits > 0 && sbused(sb) + tp->rcv_wnd < sb->sb_lowat) { rx_credits = send_rx_credits(sc, toep, rx_credits); tp->rcv_wnd += rx_credits; tp->rcv_adv += rx_credits; } if (ulp_mode(toep) == ULP_MODE_TCPDDP && toep->ddp.waiting_count > 0 && sbavail(sb) != 0) { CTR2(KTR_CXGBE, "%s: tid %u queueing AIO task", __func__, tid); ddp_queue_toep(toep); } sorwakeup_locked(so); SOCKBUF_UNLOCK_ASSERT(sb); if (ulp_mode(toep) == ULP_MODE_TCPDDP) DDP_UNLOCK(toep); INP_WUNLOCK(inp); CURVNET_RESTORE(); return (0); } static int do_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { struct adapter *sc = iq->adapter; const struct cpl_fw4_ack *cpl = (const void *)(rss + 1); unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl))); struct toepcb *toep = lookup_tid(sc, tid); struct inpcb *inp; struct tcpcb *tp; struct socket *so; uint8_t credits = cpl->credits; struct ofld_tx_sdesc *txsd; int plen; #ifdef INVARIANTS unsigned int opcode = G_CPL_FW4_ACK_OPCODE(be32toh(OPCODE_TID(cpl))); #endif /* * Very unusual case: we'd sent a flowc + abort_req for a synq entry and * now this comes back carrying the credits for the flowc. */ if (__predict_false(toep->flags & TPF_SYNQE)) { KASSERT(toep->flags & TPF_ABORT_SHUTDOWN, ("%s: credits for a synq entry %p", __func__, toep)); return (0); } inp = toep->inp; KASSERT(opcode == CPL_FW4_ACK, ("%s: unexpected opcode 0x%x", __func__, opcode)); KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__)); INP_WLOCK(inp); if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN)) { INP_WUNLOCK(inp); return (0); } - KASSERT((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) == 0, + KASSERT((inp->inp_flags & INP_DROPPED) == 0, ("%s: inp_flags 0x%x", __func__, inp->inp_flags)); tp = intotcpcb(inp); if (cpl->flags & CPL_FW4_ACK_FLAGS_SEQVAL) { tcp_seq snd_una = be32toh(cpl->snd_una); #ifdef INVARIANTS if (__predict_false(SEQ_LT(snd_una, tp->snd_una))) { log(LOG_ERR, "%s: unexpected seq# %x for TID %u, snd_una %x\n", __func__, snd_una, toep->tid, tp->snd_una); } #endif if (tp->snd_una != snd_una) { tp->snd_una = snd_una; tp->ts_recent_age = tcp_ts_getticks(); } } #ifdef VERBOSE_TRACES CTR3(KTR_CXGBE, "%s: tid %d credits %u", __func__, tid, credits); #endif so = inp->inp_socket; txsd = &toep->txsd[toep->txsd_cidx]; plen = 0; while (credits) { KASSERT(credits >= txsd->tx_credits, ("%s: too many (or partial) credits", __func__)); credits -= txsd->tx_credits; toep->tx_credits += txsd->tx_credits; plen += txsd->plen; txsd++; toep->txsd_avail++; KASSERT(toep->txsd_avail <= toep->txsd_total, ("%s: txsd avail > total", __func__)); if (__predict_false(++toep->txsd_cidx == toep->txsd_total)) { txsd = &toep->txsd[0]; toep->txsd_cidx = 0; } } if (toep->tx_credits == toep->tx_total) { toep->tx_nocompl = 0; toep->plen_nocompl = 0; } if (toep->flags & TPF_TX_SUSPENDED && toep->tx_credits >= toep->tx_total / 4) { #ifdef VERBOSE_TRACES CTR2(KTR_CXGBE, "%s: tid %d calling t4_push_frames", __func__, tid); #endif toep->flags &= ~TPF_TX_SUSPENDED; CURVNET_SET(toep->vnet); t4_push_data(sc, toep, plen); CURVNET_RESTORE(); } else if (plen > 0) { struct sockbuf *sb = &so->so_snd; int sbu; SOCKBUF_LOCK(sb); sbu = sbused(sb); if (ulp_mode(toep) == ULP_MODE_ISCSI) { if (__predict_false(sbu > 0)) { /* * The data transmitted before the * tid's ULP mode changed to ISCSI is * still in so_snd. Incoming credits * should account for so_snd first. */ sbdrop_locked(sb, min(sbu, plen)); plen -= min(sbu, plen); } sowwakeup_locked(so); /* unlocks so_snd */ rqdrop_locked(&toep->ulp_pdu_reclaimq, plen); } else { #ifdef VERBOSE_TRACES CTR3(KTR_CXGBE, "%s: tid %d dropped %d bytes", __func__, tid, plen); #endif sbdrop_locked(sb, plen); if (!TAILQ_EMPTY(&toep->aiotx_jobq)) t4_aiotx_queue_toep(so, toep); sowwakeup_locked(so); /* unlocks so_snd */ } SOCKBUF_UNLOCK_ASSERT(sb); } INP_WUNLOCK(inp); return (0); } void t4_set_tcb_field(struct adapter *sc, struct sge_wrq *wrq, struct toepcb *toep, uint16_t word, uint64_t mask, uint64_t val, int reply, int cookie) { struct wrqe *wr; struct cpl_set_tcb_field *req; struct ofld_tx_sdesc *txsd; MPASS((cookie & ~M_COOKIE) == 0); if (reply) { MPASS(cookie != CPL_COOKIE_RESERVED); } wr = alloc_wrqe(sizeof(*req), wrq); if (wr == NULL) { /* XXX */ panic("%s: allocation failure.", __func__); } req = wrtod(wr); INIT_TP_WR_MIT_CPL(req, CPL_SET_TCB_FIELD, toep->tid); req->reply_ctrl = htobe16(V_QUEUENO(toep->ofld_rxq->iq.abs_id)); if (reply == 0) req->reply_ctrl |= htobe16(F_NO_REPLY); req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(cookie)); req->mask = htobe64(mask); req->val = htobe64(val); if (wrq->eq.type == EQ_OFLD) { txsd = &toep->txsd[toep->txsd_pidx]; txsd->tx_credits = howmany(sizeof(*req), 16); txsd->plen = 0; KASSERT(toep->tx_credits >= txsd->tx_credits && toep->txsd_avail > 0, ("%s: not enough credits (%d)", __func__, toep->tx_credits)); toep->tx_credits -= txsd->tx_credits; if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) toep->txsd_pidx = 0; toep->txsd_avail--; } t4_wrq_tx(sc, wr); } void t4_init_cpl_io_handlers(void) { t4_register_cpl_handler(CPL_PEER_CLOSE, do_peer_close); t4_register_cpl_handler(CPL_CLOSE_CON_RPL, do_close_con_rpl); t4_register_cpl_handler(CPL_ABORT_REQ_RSS, do_abort_req); t4_register_shared_cpl_handler(CPL_ABORT_RPL_RSS, do_abort_rpl, CPL_COOKIE_TOM); t4_register_cpl_handler(CPL_RX_DATA, do_rx_data); t4_register_shared_cpl_handler(CPL_FW4_ACK, do_fw4_ack, CPL_COOKIE_TOM); } void t4_uninit_cpl_io_handlers(void) { t4_register_cpl_handler(CPL_PEER_CLOSE, NULL); t4_register_cpl_handler(CPL_CLOSE_CON_RPL, NULL); t4_register_cpl_handler(CPL_ABORT_REQ_RSS, NULL); t4_register_shared_cpl_handler(CPL_ABORT_RPL_RSS, NULL, CPL_COOKIE_TOM); t4_register_cpl_handler(CPL_RX_DATA, NULL); t4_register_shared_cpl_handler(CPL_FW4_ACK, NULL, CPL_COOKIE_TOM); } /* * Use the 'backend1' field in AIO jobs to hold an error that should * be reported when the job is completed, the 'backend3' field to * store the amount of data sent by the AIO job so far, and the * 'backend4' field to hold a reference count on the job. * * Each unmapped mbuf holds a reference on the job as does the queue * so long as the job is queued. */ #define aio_error backend1 #define aio_sent backend3 #define aio_refs backend4 #define jobtotid(job) \ (((struct toepcb *)(so_sototcpcb((job)->fd_file->f_data)->t_toe))->tid) static void aiotx_free_job(struct kaiocb *job) { long status; int error; if (refcount_release(&job->aio_refs) == 0) return; error = (intptr_t)job->aio_error; status = job->aio_sent; #ifdef VERBOSE_TRACES CTR5(KTR_CXGBE, "%s: tid %d completed %p len %ld, error %d", __func__, jobtotid(job), job, status, error); #endif if (error != 0 && status != 0) error = 0; if (error == ECANCELED) aio_cancel(job); else if (error) aio_complete(job, -1, error); else { job->msgsnd = 1; aio_complete(job, status, 0); } } static void aiotx_free_pgs(struct mbuf *m) { struct kaiocb *job; vm_page_t pg; M_ASSERTEXTPG(m); job = m->m_ext.ext_arg1; #ifdef VERBOSE_TRACES CTR3(KTR_CXGBE, "%s: completed %d bytes for tid %d", __func__, m->m_len, jobtotid(job)); #endif for (int i = 0; i < m->m_epg_npgs; i++) { pg = PHYS_TO_VM_PAGE(m->m_epg_pa[i]); vm_page_unwire(pg, PQ_ACTIVE); } aiotx_free_job(job); } /* * Allocate a chain of unmapped mbufs describing the next 'len' bytes * of an AIO job. */ static struct mbuf * alloc_aiotx_mbuf(struct kaiocb *job, int len) { struct vmspace *vm; vm_page_t pgs[MBUF_PEXT_MAX_PGS]; struct mbuf *m, *top, *last; vm_map_t map; vm_offset_t start; int i, mlen, npages, pgoff; KASSERT(job->aio_sent + len <= job->uaiocb.aio_nbytes, ("%s(%p, %d): request to send beyond end of buffer", __func__, job, len)); /* * The AIO subsystem will cancel and drain all requests before * permitting a process to exit or exec, so p_vmspace should * be stable here. */ vm = job->userproc->p_vmspace; map = &vm->vm_map; start = (uintptr_t)job->uaiocb.aio_buf + job->aio_sent; pgoff = start & PAGE_MASK; top = NULL; last = NULL; while (len > 0) { mlen = imin(len, MBUF_PEXT_MAX_PGS * PAGE_SIZE - pgoff); KASSERT(mlen == len || ((start + mlen) & PAGE_MASK) == 0, ("%s: next start (%#jx + %#x) is not page aligned", __func__, (uintmax_t)start, mlen)); npages = vm_fault_quick_hold_pages(map, start, mlen, VM_PROT_WRITE, pgs, nitems(pgs)); if (npages < 0) break; m = mb_alloc_ext_pgs(M_WAITOK, aiotx_free_pgs); if (m == NULL) { vm_page_unhold_pages(pgs, npages); break; } m->m_epg_1st_off = pgoff; m->m_epg_npgs = npages; if (npages == 1) { KASSERT(mlen + pgoff <= PAGE_SIZE, ("%s: single page is too large (off %d len %d)", __func__, pgoff, mlen)); m->m_epg_last_len = mlen; } else { m->m_epg_last_len = mlen - (PAGE_SIZE - pgoff) - (npages - 2) * PAGE_SIZE; } for (i = 0; i < npages; i++) m->m_epg_pa[i] = VM_PAGE_TO_PHYS(pgs[i]); m->m_len = mlen; m->m_ext.ext_size = npages * PAGE_SIZE; m->m_ext.ext_arg1 = job; refcount_acquire(&job->aio_refs); #ifdef VERBOSE_TRACES CTR5(KTR_CXGBE, "%s: tid %d, new mbuf %p for job %p, npages %d", __func__, jobtotid(job), m, job, npages); #endif if (top == NULL) top = m; else last->m_next = m; last = m; len -= mlen; start += mlen; pgoff = 0; } return (top); } static void t4_aiotx_process_job(struct toepcb *toep, struct socket *so, struct kaiocb *job) { struct sockbuf *sb; struct inpcb *inp; struct tcpcb *tp; struct mbuf *m; int error, len; bool moretocome, sendmore; sb = &so->so_snd; SOCKBUF_UNLOCK(sb); m = NULL; #ifdef MAC error = mac_socket_check_send(job->fd_file->f_cred, so); if (error != 0) goto out; #endif /* Inline sosend_generic(). */ error = SOCK_IO_SEND_LOCK(so, SBL_WAIT); MPASS(error == 0); sendanother: SOCKBUF_LOCK(sb); if (so->so_snd.sb_state & SBS_CANTSENDMORE) { SOCKBUF_UNLOCK(sb); SOCK_IO_SEND_UNLOCK(so); if ((so->so_options & SO_NOSIGPIPE) == 0) { PROC_LOCK(job->userproc); kern_psignal(job->userproc, SIGPIPE); PROC_UNLOCK(job->userproc); } error = EPIPE; goto out; } if (so->so_error) { error = so->so_error; so->so_error = 0; SOCKBUF_UNLOCK(sb); SOCK_IO_SEND_UNLOCK(so); goto out; } if ((so->so_state & SS_ISCONNECTED) == 0) { SOCKBUF_UNLOCK(sb); SOCK_IO_SEND_UNLOCK(so); error = ENOTCONN; goto out; } if (sbspace(sb) < sb->sb_lowat) { MPASS(job->aio_sent == 0 || !(so->so_state & SS_NBIO)); /* * Don't block if there is too little room in the socket * buffer. Instead, requeue the request. */ if (!aio_set_cancel_function(job, t4_aiotx_cancel)) { SOCKBUF_UNLOCK(sb); SOCK_IO_SEND_UNLOCK(so); error = ECANCELED; goto out; } TAILQ_INSERT_HEAD(&toep->aiotx_jobq, job, list); SOCKBUF_UNLOCK(sb); SOCK_IO_SEND_UNLOCK(so); goto out; } /* * Write as much data as the socket permits, but no more than a * a single sndbuf at a time. */ len = sbspace(sb); if (len > job->uaiocb.aio_nbytes - job->aio_sent) { len = job->uaiocb.aio_nbytes - job->aio_sent; moretocome = false; } else moretocome = true; if (len > toep->params.sndbuf) { len = toep->params.sndbuf; sendmore = true; } else sendmore = false; if (!TAILQ_EMPTY(&toep->aiotx_jobq)) moretocome = true; SOCKBUF_UNLOCK(sb); MPASS(len != 0); m = alloc_aiotx_mbuf(job, len); if (m == NULL) { SOCK_IO_SEND_UNLOCK(so); error = EFAULT; goto out; } /* Inlined tcp_usr_send(). */ inp = toep->inp; INP_WLOCK(inp); - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { INP_WUNLOCK(inp); SOCK_IO_SEND_UNLOCK(so); error = ECONNRESET; goto out; } job->aio_sent += m_length(m, NULL); sbappendstream(sb, m, 0); m = NULL; if (!(inp->inp_flags & INP_DROPPED)) { tp = intotcpcb(inp); if (moretocome) tp->t_flags |= TF_MORETOCOME; error = tcp_output(tp); if (error < 0) { INP_UNLOCK_ASSERT(inp); SOCK_IO_SEND_UNLOCK(so); error = -error; goto out; } if (moretocome) tp->t_flags &= ~TF_MORETOCOME; } INP_WUNLOCK(inp); if (sendmore) goto sendanother; SOCK_IO_SEND_UNLOCK(so); if (error) goto out; /* * If this is a blocking socket and the request has not been * fully completed, requeue it until the socket is ready * again. */ if (job->aio_sent < job->uaiocb.aio_nbytes && !(so->so_state & SS_NBIO)) { SOCKBUF_LOCK(sb); if (!aio_set_cancel_function(job, t4_aiotx_cancel)) { SOCKBUF_UNLOCK(sb); error = ECANCELED; goto out; } TAILQ_INSERT_HEAD(&toep->aiotx_jobq, job, list); return; } /* * If the request will not be requeued, drop the queue's * reference to the job. Any mbufs in flight should still * hold a reference, but this drops the reference that the * queue owns while it is waiting to queue mbufs to the * socket. */ aiotx_free_job(job); out: if (error) { job->aio_error = (void *)(intptr_t)error; aiotx_free_job(job); } m_freem(m); SOCKBUF_LOCK(sb); } static void t4_aiotx_task(void *context, int pending) { struct toepcb *toep = context; struct socket *so; struct kaiocb *job; struct epoch_tracker et; so = toep->aiotx_so; CURVNET_SET(toep->vnet); NET_EPOCH_ENTER(et); SOCKBUF_LOCK(&so->so_snd); while (!TAILQ_EMPTY(&toep->aiotx_jobq) && sowriteable(so)) { job = TAILQ_FIRST(&toep->aiotx_jobq); TAILQ_REMOVE(&toep->aiotx_jobq, job, list); if (!aio_clear_cancel_function(job)) continue; t4_aiotx_process_job(toep, so, job); } toep->aiotx_so = NULL; SOCKBUF_UNLOCK(&so->so_snd); NET_EPOCH_EXIT(et); free_toepcb(toep); sorele(so); CURVNET_RESTORE(); } static void t4_aiotx_queue_toep(struct socket *so, struct toepcb *toep) { SOCKBUF_LOCK_ASSERT(&toep->inp->inp_socket->so_snd); #ifdef VERBOSE_TRACES CTR3(KTR_CXGBE, "%s: queueing aiotx task for tid %d, active = %s", __func__, toep->tid, toep->aiotx_so != NULL ? "true" : "false"); #endif if (toep->aiotx_so != NULL) return; soref(so); toep->aiotx_so = so; hold_toepcb(toep); soaio_enqueue(&toep->aiotx_task); } static void t4_aiotx_cancel(struct kaiocb *job) { struct socket *so; struct sockbuf *sb; struct tcpcb *tp; struct toepcb *toep; so = job->fd_file->f_data; tp = so_sototcpcb(so); toep = tp->t_toe; MPASS(job->uaiocb.aio_lio_opcode == LIO_WRITE); sb = &so->so_snd; SOCKBUF_LOCK(sb); if (!aio_cancel_cleared(job)) TAILQ_REMOVE(&toep->aiotx_jobq, job, list); SOCKBUF_UNLOCK(sb); job->aio_error = (void *)(intptr_t)ECANCELED; aiotx_free_job(job); } int t4_aio_queue_aiotx(struct socket *so, struct kaiocb *job) { struct tcpcb *tp = so_sototcpcb(so); struct toepcb *toep = tp->t_toe; struct adapter *sc = td_adapter(toep->td); /* This only handles writes. */ if (job->uaiocb.aio_lio_opcode != LIO_WRITE) return (EOPNOTSUPP); if (!sc->tt.tx_zcopy) return (EOPNOTSUPP); if (tls_tx_key(toep)) return (EOPNOTSUPP); SOCKBUF_LOCK(&so->so_snd); #ifdef VERBOSE_TRACES CTR3(KTR_CXGBE, "%s: queueing %p for tid %u", __func__, job, toep->tid); #endif if (!aio_set_cancel_function(job, t4_aiotx_cancel)) panic("new job was cancelled"); refcount_init(&job->aio_refs, 1); TAILQ_INSERT_TAIL(&toep->aiotx_jobq, job, list); if (sowriteable(so)) t4_aiotx_queue_toep(so, toep); SOCKBUF_UNLOCK(&so->so_snd); return (0); } void aiotx_init_toep(struct toepcb *toep) { TAILQ_INIT(&toep->aiotx_jobq); TASK_INIT(&toep->aiotx_task, 0, t4_aiotx_task, toep); } #endif diff --git a/sys/dev/cxgbe/tom/t4_ddp.c b/sys/dev/cxgbe/tom/t4_ddp.c index 11fea91b060e..605157286bb2 100644 --- a/sys/dev/cxgbe/tom/t4_ddp.c +++ b/sys/dev/cxgbe/tom/t4_ddp.c @@ -1,2262 +1,2262 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2012 Chelsio Communications, Inc. * All rights reserved. * Written by: Navdeep Parhar * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #define TCPSTATES #include #include #include #include #include #include #include #include #include #include #include #ifdef TCP_OFFLOAD #include "common/common.h" #include "common/t4_msg.h" #include "common/t4_regs.h" #include "common/t4_tcb.h" #include "tom/t4_tom.h" /* * Use the 'backend3' field in AIO jobs to store the amount of data * received by the AIO job so far. */ #define aio_received backend3 static void aio_ddp_requeue_task(void *context, int pending); static void ddp_complete_all(struct toepcb *toep, int error); static void t4_aio_cancel_active(struct kaiocb *job); static void t4_aio_cancel_queued(struct kaiocb *job); static TAILQ_HEAD(, pageset) ddp_orphan_pagesets; static struct mtx ddp_orphan_pagesets_lock; static struct task ddp_orphan_task; #define MAX_DDP_BUFFER_SIZE (M_TCB_RX_DDP_BUF0_LEN) /* * A page set holds information about a buffer used for DDP. The page * set holds resources such as the VM pages backing the buffer (either * held or wired) and the page pods associated with the buffer. * Recently used page sets are cached to allow for efficient reuse of * buffers (avoiding the need to re-fault in pages, hold them, etc.). * Note that cached page sets keep the backing pages wired. The * number of wired pages is capped by only allowing for two wired * pagesets per connection. This is not a perfect cap, but is a * trade-off for performance. * * If an application ping-pongs two buffers for a connection via * aio_read(2) then those buffers should remain wired and expensive VM * fault lookups should be avoided after each buffer has been used * once. If an application uses more than two buffers then this will * fall back to doing expensive VM fault lookups for each operation. */ static void free_pageset(struct tom_data *td, struct pageset *ps) { vm_page_t p; int i; if (ps->prsv.prsv_nppods > 0) t4_free_page_pods(&ps->prsv); for (i = 0; i < ps->npages; i++) { p = ps->pages[i]; vm_page_unwire(p, PQ_INACTIVE); } mtx_lock(&ddp_orphan_pagesets_lock); TAILQ_INSERT_TAIL(&ddp_orphan_pagesets, ps, link); taskqueue_enqueue(taskqueue_thread, &ddp_orphan_task); mtx_unlock(&ddp_orphan_pagesets_lock); } static void ddp_free_orphan_pagesets(void *context, int pending) { struct pageset *ps; mtx_lock(&ddp_orphan_pagesets_lock); while (!TAILQ_EMPTY(&ddp_orphan_pagesets)) { ps = TAILQ_FIRST(&ddp_orphan_pagesets); TAILQ_REMOVE(&ddp_orphan_pagesets, ps, link); mtx_unlock(&ddp_orphan_pagesets_lock); if (ps->vm) vmspace_free(ps->vm); free(ps, M_CXGBE); mtx_lock(&ddp_orphan_pagesets_lock); } mtx_unlock(&ddp_orphan_pagesets_lock); } static void recycle_pageset(struct toepcb *toep, struct pageset *ps) { DDP_ASSERT_LOCKED(toep); if (!(toep->ddp.flags & DDP_DEAD)) { KASSERT(toep->ddp.cached_count + toep->ddp.active_count < nitems(toep->ddp.db), ("too many wired pagesets")); TAILQ_INSERT_HEAD(&toep->ddp.cached_pagesets, ps, link); toep->ddp.cached_count++; } else free_pageset(toep->td, ps); } static void ddp_complete_one(struct kaiocb *job, int error) { long copied; /* * If this job had copied data out of the socket buffer before * it was cancelled, report it as a short read rather than an * error. */ copied = job->aio_received; if (copied != 0 || error == 0) aio_complete(job, copied, 0); else aio_complete(job, -1, error); } static void free_ddp_buffer(struct tom_data *td, struct ddp_buffer *db) { if (db->job) { /* * XXX: If we are un-offloading the socket then we * should requeue these on the socket somehow. If we * got a FIN from the remote end, then this completes * any remaining requests with an EOF read. */ if (!aio_clear_cancel_function(db->job)) ddp_complete_one(db->job, 0); } if (db->ps) free_pageset(td, db->ps); } void ddp_init_toep(struct toepcb *toep) { TAILQ_INIT(&toep->ddp.aiojobq); TASK_INIT(&toep->ddp.requeue_task, 0, aio_ddp_requeue_task, toep); toep->ddp.flags = DDP_OK; toep->ddp.active_id = -1; mtx_init(&toep->ddp.lock, "t4 ddp", NULL, MTX_DEF); } void ddp_uninit_toep(struct toepcb *toep) { mtx_destroy(&toep->ddp.lock); } void release_ddp_resources(struct toepcb *toep) { struct pageset *ps; int i; DDP_LOCK(toep); toep->ddp.flags |= DDP_DEAD; for (i = 0; i < nitems(toep->ddp.db); i++) { free_ddp_buffer(toep->td, &toep->ddp.db[i]); } while ((ps = TAILQ_FIRST(&toep->ddp.cached_pagesets)) != NULL) { TAILQ_REMOVE(&toep->ddp.cached_pagesets, ps, link); free_pageset(toep->td, ps); } ddp_complete_all(toep, 0); DDP_UNLOCK(toep); } #ifdef INVARIANTS void ddp_assert_empty(struct toepcb *toep) { int i; MPASS(!(toep->ddp.flags & DDP_TASK_ACTIVE)); for (i = 0; i < nitems(toep->ddp.db); i++) { MPASS(toep->ddp.db[i].job == NULL); MPASS(toep->ddp.db[i].ps == NULL); } MPASS(TAILQ_EMPTY(&toep->ddp.cached_pagesets)); MPASS(TAILQ_EMPTY(&toep->ddp.aiojobq)); } #endif static void complete_ddp_buffer(struct toepcb *toep, struct ddp_buffer *db, unsigned int db_idx) { unsigned int db_flag; toep->ddp.active_count--; if (toep->ddp.active_id == db_idx) { if (toep->ddp.active_count == 0) { KASSERT(toep->ddp.db[db_idx ^ 1].job == NULL, ("%s: active_count mismatch", __func__)); toep->ddp.active_id = -1; } else toep->ddp.active_id ^= 1; #ifdef VERBOSE_TRACES CTR3(KTR_CXGBE, "%s: tid %u, ddp_active_id = %d", __func__, toep->tid, toep->ddp.active_id); #endif } else { KASSERT(toep->ddp.active_count != 0 && toep->ddp.active_id != -1, ("%s: active count mismatch", __func__)); } db->cancel_pending = 0; db->job = NULL; recycle_pageset(toep, db->ps); db->ps = NULL; db_flag = db_idx == 1 ? DDP_BUF1_ACTIVE : DDP_BUF0_ACTIVE; KASSERT(toep->ddp.flags & db_flag, ("%s: DDP buffer not active. toep %p, ddp_flags 0x%x", __func__, toep, toep->ddp.flags)); toep->ddp.flags &= ~db_flag; } /* XXX: handle_ddp_data code duplication */ void insert_ddp_data(struct toepcb *toep, uint32_t n) { struct inpcb *inp = toep->inp; struct tcpcb *tp = intotcpcb(inp); struct ddp_buffer *db; struct kaiocb *job; size_t placed; long copied; unsigned int db_idx; #ifdef INVARIANTS unsigned int db_flag; #endif INP_WLOCK_ASSERT(inp); DDP_ASSERT_LOCKED(toep); tp->rcv_nxt += n; #ifndef USE_DDP_RX_FLOW_CONTROL KASSERT(tp->rcv_wnd >= n, ("%s: negative window size", __func__)); tp->rcv_wnd -= n; #endif CTR2(KTR_CXGBE, "%s: placed %u bytes before falling out of DDP", __func__, n); while (toep->ddp.active_count > 0) { MPASS(toep->ddp.active_id != -1); db_idx = toep->ddp.active_id; #ifdef INVARIANTS db_flag = db_idx == 1 ? DDP_BUF1_ACTIVE : DDP_BUF0_ACTIVE; #endif MPASS((toep->ddp.flags & db_flag) != 0); db = &toep->ddp.db[db_idx]; job = db->job; copied = job->aio_received; placed = n; if (placed > job->uaiocb.aio_nbytes - copied) placed = job->uaiocb.aio_nbytes - copied; if (placed > 0) job->msgrcv = 1; if (!aio_clear_cancel_function(job)) { /* * Update the copied length for when * t4_aio_cancel_active() completes this * request. */ job->aio_received += placed; } else if (copied + placed != 0) { CTR4(KTR_CXGBE, "%s: completing %p (copied %ld, placed %lu)", __func__, job, copied, placed); /* XXX: This always completes if there is some data. */ aio_complete(job, copied + placed, 0); } else if (aio_set_cancel_function(job, t4_aio_cancel_queued)) { TAILQ_INSERT_HEAD(&toep->ddp.aiojobq, job, list); toep->ddp.waiting_count++; } else aio_cancel(job); n -= placed; complete_ddp_buffer(toep, db, db_idx); } MPASS(n == 0); } /* SET_TCB_FIELD sent as a ULP command looks like this */ #define LEN__SET_TCB_FIELD_ULP (sizeof(struct ulp_txpkt) + \ sizeof(struct ulptx_idata) + sizeof(struct cpl_set_tcb_field_core)) /* RX_DATA_ACK sent as a ULP command looks like this */ #define LEN__RX_DATA_ACK_ULP (sizeof(struct ulp_txpkt) + \ sizeof(struct ulptx_idata) + sizeof(struct cpl_rx_data_ack_core)) static inline void * mk_set_tcb_field_ulp(struct ulp_txpkt *ulpmc, struct toepcb *toep, uint64_t word, uint64_t mask, uint64_t val) { struct ulptx_idata *ulpsc; struct cpl_set_tcb_field_core *req; ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0)); ulpmc->len = htobe32(howmany(LEN__SET_TCB_FIELD_ULP, 16)); ulpsc = (struct ulptx_idata *)(ulpmc + 1); ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM)); ulpsc->len = htobe32(sizeof(*req)); req = (struct cpl_set_tcb_field_core *)(ulpsc + 1); OPCODE_TID(req) = htobe32(MK_OPCODE_TID(CPL_SET_TCB_FIELD, toep->tid)); req->reply_ctrl = htobe16(V_NO_REPLY(1) | V_QUEUENO(toep->ofld_rxq->iq.abs_id)); req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(0)); req->mask = htobe64(mask); req->val = htobe64(val); ulpsc = (struct ulptx_idata *)(req + 1); if (LEN__SET_TCB_FIELD_ULP % 16) { ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_NOOP)); ulpsc->len = htobe32(0); return (ulpsc + 1); } return (ulpsc); } static inline void * mk_rx_data_ack_ulp(struct ulp_txpkt *ulpmc, struct toepcb *toep) { struct ulptx_idata *ulpsc; struct cpl_rx_data_ack_core *req; ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0)); ulpmc->len = htobe32(howmany(LEN__RX_DATA_ACK_ULP, 16)); ulpsc = (struct ulptx_idata *)(ulpmc + 1); ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM)); ulpsc->len = htobe32(sizeof(*req)); req = (struct cpl_rx_data_ack_core *)(ulpsc + 1); OPCODE_TID(req) = htobe32(MK_OPCODE_TID(CPL_RX_DATA_ACK, toep->tid)); req->credit_dack = htobe32(F_RX_MODULATE_RX); ulpsc = (struct ulptx_idata *)(req + 1); if (LEN__RX_DATA_ACK_ULP % 16) { ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_NOOP)); ulpsc->len = htobe32(0); return (ulpsc + 1); } return (ulpsc); } static struct wrqe * mk_update_tcb_for_ddp(struct adapter *sc, struct toepcb *toep, int db_idx, struct pageset *ps, int offset, uint64_t ddp_flags, uint64_t ddp_flags_mask) { struct wrqe *wr; struct work_request_hdr *wrh; struct ulp_txpkt *ulpmc; int len; KASSERT(db_idx == 0 || db_idx == 1, ("%s: bad DDP buffer index %d", __func__, db_idx)); /* * We'll send a compound work request that has 3 SET_TCB_FIELDs and an * RX_DATA_ACK (with RX_MODULATE to speed up delivery). * * The work request header is 16B and always ends at a 16B boundary. * The ULPTX master commands that follow must all end at 16B boundaries * too so we round up the size to 16. */ len = sizeof(*wrh) + 3 * roundup2(LEN__SET_TCB_FIELD_ULP, 16) + roundup2(LEN__RX_DATA_ACK_ULP, 16); wr = alloc_wrqe(len, toep->ctrlq); if (wr == NULL) return (NULL); wrh = wrtod(wr); INIT_ULPTX_WRH(wrh, len, 1, 0); /* atomic */ ulpmc = (struct ulp_txpkt *)(wrh + 1); /* Write the buffer's tag */ ulpmc = mk_set_tcb_field_ulp(ulpmc, toep, W_TCB_RX_DDP_BUF0_TAG + db_idx, V_TCB_RX_DDP_BUF0_TAG(M_TCB_RX_DDP_BUF0_TAG), V_TCB_RX_DDP_BUF0_TAG(ps->prsv.prsv_tag)); /* Update the current offset in the DDP buffer and its total length */ if (db_idx == 0) ulpmc = mk_set_tcb_field_ulp(ulpmc, toep, W_TCB_RX_DDP_BUF0_OFFSET, V_TCB_RX_DDP_BUF0_OFFSET(M_TCB_RX_DDP_BUF0_OFFSET) | V_TCB_RX_DDP_BUF0_LEN(M_TCB_RX_DDP_BUF0_LEN), V_TCB_RX_DDP_BUF0_OFFSET(offset) | V_TCB_RX_DDP_BUF0_LEN(ps->len)); else ulpmc = mk_set_tcb_field_ulp(ulpmc, toep, W_TCB_RX_DDP_BUF1_OFFSET, V_TCB_RX_DDP_BUF1_OFFSET(M_TCB_RX_DDP_BUF1_OFFSET) | V_TCB_RX_DDP_BUF1_LEN((u64)M_TCB_RX_DDP_BUF1_LEN << 32), V_TCB_RX_DDP_BUF1_OFFSET(offset) | V_TCB_RX_DDP_BUF1_LEN((u64)ps->len << 32)); /* Update DDP flags */ ulpmc = mk_set_tcb_field_ulp(ulpmc, toep, W_TCB_RX_DDP_FLAGS, ddp_flags_mask, ddp_flags); /* Gratuitous RX_DATA_ACK with RX_MODULATE set to speed up delivery. */ ulpmc = mk_rx_data_ack_ulp(ulpmc, toep); return (wr); } static int handle_ddp_data(struct toepcb *toep, __be32 ddp_report, __be32 rcv_nxt, int len) { uint32_t report = be32toh(ddp_report); unsigned int db_idx; struct inpcb *inp = toep->inp; struct ddp_buffer *db; struct tcpcb *tp; struct socket *so; struct sockbuf *sb; struct kaiocb *job; long copied; db_idx = report & F_DDP_BUF_IDX ? 1 : 0; if (__predict_false(!(report & F_DDP_INV))) CXGBE_UNIMPLEMENTED("DDP buffer still valid"); INP_WLOCK(inp); so = inp_inpcbtosocket(inp); sb = &so->so_rcv; DDP_LOCK(toep); KASSERT(toep->ddp.active_id == db_idx, ("completed DDP buffer (%d) != active_id (%d) for tid %d", db_idx, toep->ddp.active_id, toep->tid)); db = &toep->ddp.db[db_idx]; job = db->job; - if (__predict_false(inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT))) { + if (__predict_false(inp->inp_flags & INP_DROPPED)) { /* * This can happen due to an administrative tcpdrop(8). * Just fail the request with ECONNRESET. */ CTR5(KTR_CXGBE, "%s: tid %u, seq 0x%x, len %d, inp_flags 0x%x", __func__, toep->tid, be32toh(rcv_nxt), len, inp->inp_flags); if (aio_clear_cancel_function(job)) ddp_complete_one(job, ECONNRESET); goto completed; } tp = intotcpcb(inp); /* * For RX_DDP_COMPLETE, len will be zero and rcv_nxt is the * sequence number of the next byte to receive. The length of * the data received for this message must be computed by * comparing the new and old values of rcv_nxt. * * For RX_DATA_DDP, len might be non-zero, but it is only the * length of the most recent DMA. It does not include the * total length of the data received since the previous update * for this DDP buffer. rcv_nxt is the sequence number of the * first received byte from the most recent DMA. */ len += be32toh(rcv_nxt) - tp->rcv_nxt; tp->rcv_nxt += len; tp->t_rcvtime = ticks; #ifndef USE_DDP_RX_FLOW_CONTROL KASSERT(tp->rcv_wnd >= len, ("%s: negative window size", __func__)); tp->rcv_wnd -= len; #endif #ifdef VERBOSE_TRACES CTR5(KTR_CXGBE, "%s: tid %u, DDP[%d] placed %d bytes (%#x)", __func__, toep->tid, db_idx, len, report); #endif /* receive buffer autosize */ MPASS(toep->vnet == so->so_vnet); CURVNET_SET(toep->vnet); SOCKBUF_LOCK(sb); if (sb->sb_flags & SB_AUTOSIZE && V_tcp_do_autorcvbuf && sb->sb_hiwat < V_tcp_autorcvbuf_max && len > (sbspace(sb) / 8 * 7)) { struct adapter *sc = td_adapter(toep->td); unsigned int hiwat = sb->sb_hiwat; unsigned int newsize = min(hiwat + sc->tt.autorcvbuf_inc, V_tcp_autorcvbuf_max); if (!sbreserve_locked(so, SO_RCV, newsize, NULL)) sb->sb_flags &= ~SB_AUTOSIZE; } SOCKBUF_UNLOCK(sb); CURVNET_RESTORE(); job->msgrcv = 1; if (db->cancel_pending) { /* * Update the job's length but defer completion to the * TCB_RPL callback. */ job->aio_received += len; goto out; } else if (!aio_clear_cancel_function(job)) { /* * Update the copied length for when * t4_aio_cancel_active() completes this request. */ job->aio_received += len; } else { copied = job->aio_received; #ifdef VERBOSE_TRACES CTR5(KTR_CXGBE, "%s: tid %u, completing %p (copied %ld, placed %d)", __func__, toep->tid, job, copied, len); #endif aio_complete(job, copied + len, 0); t4_rcvd(&toep->td->tod, tp); } completed: complete_ddp_buffer(toep, db, db_idx); if (toep->ddp.waiting_count > 0) ddp_queue_toep(toep); out: DDP_UNLOCK(toep); INP_WUNLOCK(inp); return (0); } void handle_ddp_indicate(struct toepcb *toep) { DDP_ASSERT_LOCKED(toep); MPASS(toep->ddp.active_count == 0); MPASS((toep->ddp.flags & (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE)) == 0); if (toep->ddp.waiting_count == 0) { /* * The pending requests that triggered the request for an * an indicate were cancelled. Those cancels should have * already disabled DDP. Just ignore this as the data is * going into the socket buffer anyway. */ return; } CTR3(KTR_CXGBE, "%s: tid %d indicated (%d waiting)", __func__, toep->tid, toep->ddp.waiting_count); ddp_queue_toep(toep); } CTASSERT(CPL_COOKIE_DDP0 + 1 == CPL_COOKIE_DDP1); static int do_ddp_tcb_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { struct adapter *sc = iq->adapter; const struct cpl_set_tcb_rpl *cpl = (const void *)(rss + 1); unsigned int tid = GET_TID(cpl); unsigned int db_idx; struct toepcb *toep; struct inpcb *inp; struct ddp_buffer *db; struct kaiocb *job; long copied; if (cpl->status != CPL_ERR_NONE) panic("XXX: tcp_rpl failed: %d", cpl->status); toep = lookup_tid(sc, tid); inp = toep->inp; switch (cpl->cookie) { case V_WORD(W_TCB_RX_DDP_FLAGS) | V_COOKIE(CPL_COOKIE_DDP0): case V_WORD(W_TCB_RX_DDP_FLAGS) | V_COOKIE(CPL_COOKIE_DDP1): /* * XXX: This duplicates a lot of code with handle_ddp_data(). */ db_idx = G_COOKIE(cpl->cookie) - CPL_COOKIE_DDP0; MPASS(db_idx < nitems(toep->ddp.db)); INP_WLOCK(inp); DDP_LOCK(toep); db = &toep->ddp.db[db_idx]; /* * handle_ddp_data() should leave the job around until * this callback runs once a cancel is pending. */ MPASS(db != NULL); MPASS(db->job != NULL); MPASS(db->cancel_pending); /* * XXX: It's not clear what happens if there is data * placed when the buffer is invalidated. I suspect we * need to read the TCB to see how much data was placed. * * For now this just pretends like nothing was placed. * * XXX: Note that if we did check the PCB we would need to * also take care of updating the tp, etc. */ job = db->job; copied = job->aio_received; if (copied == 0) { CTR2(KTR_CXGBE, "%s: cancelling %p", __func__, job); aio_cancel(job); } else { CTR3(KTR_CXGBE, "%s: completing %p (copied %ld)", __func__, job, copied); aio_complete(job, copied, 0); t4_rcvd(&toep->td->tod, intotcpcb(inp)); } complete_ddp_buffer(toep, db, db_idx); if (toep->ddp.waiting_count > 0) ddp_queue_toep(toep); DDP_UNLOCK(toep); INP_WUNLOCK(inp); break; default: panic("XXX: unknown tcb_rpl offset %#x, cookie %#x", G_WORD(cpl->cookie), G_COOKIE(cpl->cookie)); } return (0); } void handle_ddp_close(struct toepcb *toep, struct tcpcb *tp, __be32 rcv_nxt) { struct ddp_buffer *db; struct kaiocb *job; long copied; unsigned int db_idx; #ifdef INVARIANTS unsigned int db_flag; #endif int len, placed; INP_WLOCK_ASSERT(toep->inp); DDP_ASSERT_LOCKED(toep); /* - 1 is to ignore the byte for FIN */ len = be32toh(rcv_nxt) - tp->rcv_nxt - 1; tp->rcv_nxt += len; while (toep->ddp.active_count > 0) { MPASS(toep->ddp.active_id != -1); db_idx = toep->ddp.active_id; #ifdef INVARIANTS db_flag = db_idx == 1 ? DDP_BUF1_ACTIVE : DDP_BUF0_ACTIVE; #endif MPASS((toep->ddp.flags & db_flag) != 0); db = &toep->ddp.db[db_idx]; job = db->job; copied = job->aio_received; placed = len; if (placed > job->uaiocb.aio_nbytes - copied) placed = job->uaiocb.aio_nbytes - copied; if (placed > 0) job->msgrcv = 1; if (!aio_clear_cancel_function(job)) { /* * Update the copied length for when * t4_aio_cancel_active() completes this * request. */ job->aio_received += placed; } else { CTR4(KTR_CXGBE, "%s: tid %d completed buf %d len %d", __func__, toep->tid, db_idx, placed); aio_complete(job, copied + placed, 0); } len -= placed; complete_ddp_buffer(toep, db, db_idx); } MPASS(len == 0); ddp_complete_all(toep, 0); } #define DDP_ERR (F_DDP_PPOD_MISMATCH | F_DDP_LLIMIT_ERR | F_DDP_ULIMIT_ERR |\ F_DDP_PPOD_PARITY_ERR | F_DDP_PADDING_ERR | F_DDP_OFFSET_ERR |\ F_DDP_INVALID_TAG | F_DDP_COLOR_ERR | F_DDP_TID_MISMATCH |\ F_DDP_INVALID_PPOD | F_DDP_HDRCRC_ERR | F_DDP_DATACRC_ERR) extern cpl_handler_t t4_cpl_handler[]; static int do_rx_data_ddp(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { struct adapter *sc = iq->adapter; const struct cpl_rx_data_ddp *cpl = (const void *)(rss + 1); unsigned int tid = GET_TID(cpl); uint32_t vld; struct toepcb *toep = lookup_tid(sc, tid); KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); KASSERT(toep->tid == tid, ("%s: toep tid/atid mismatch", __func__)); KASSERT(!(toep->flags & TPF_SYNQE), ("%s: toep %p claims to be a synq entry", __func__, toep)); vld = be32toh(cpl->ddpvld); if (__predict_false(vld & DDP_ERR)) { panic("%s: DDP error 0x%x (tid %d, toep %p)", __func__, vld, tid, toep); } if (ulp_mode(toep) == ULP_MODE_ISCSI) { t4_cpl_handler[CPL_RX_ISCSI_DDP](iq, rss, m); return (0); } handle_ddp_data(toep, cpl->u.ddp_report, cpl->seq, be16toh(cpl->len)); return (0); } static int do_rx_ddp_complete(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { struct adapter *sc = iq->adapter; const struct cpl_rx_ddp_complete *cpl = (const void *)(rss + 1); unsigned int tid = GET_TID(cpl); struct toepcb *toep = lookup_tid(sc, tid); KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); KASSERT(toep->tid == tid, ("%s: toep tid/atid mismatch", __func__)); KASSERT(!(toep->flags & TPF_SYNQE), ("%s: toep %p claims to be a synq entry", __func__, toep)); handle_ddp_data(toep, cpl->ddp_report, cpl->rcv_nxt, 0); return (0); } static void enable_ddp(struct adapter *sc, struct toepcb *toep) { KASSERT((toep->ddp.flags & (DDP_ON | DDP_OK | DDP_SC_REQ)) == DDP_OK, ("%s: toep %p has bad ddp_flags 0x%x", __func__, toep, toep->ddp.flags)); CTR3(KTR_CXGBE, "%s: tid %u (time %u)", __func__, toep->tid, time_uptime); DDP_ASSERT_LOCKED(toep); toep->ddp.flags |= DDP_SC_REQ; t4_set_tcb_field(sc, toep->ctrlq, toep, W_TCB_RX_DDP_FLAGS, V_TF_DDP_OFF(1) | V_TF_DDP_INDICATE_OUT(1) | V_TF_DDP_BUF0_INDICATE(1) | V_TF_DDP_BUF1_INDICATE(1) | V_TF_DDP_BUF0_VALID(1) | V_TF_DDP_BUF1_VALID(1), V_TF_DDP_BUF0_INDICATE(1) | V_TF_DDP_BUF1_INDICATE(1), 0, 0); t4_set_tcb_field(sc, toep->ctrlq, toep, W_TCB_T_FLAGS, V_TF_RCV_COALESCE_ENABLE(1), 0, 0, 0); } static int calculate_hcf(int n1, int n2) { int a, b, t; if (n1 <= n2) { a = n1; b = n2; } else { a = n2; b = n1; } while (a != 0) { t = a; a = b % a; b = t; } return (b); } static inline int pages_to_nppods(int npages, int ddp_page_shift) { MPASS(ddp_page_shift >= PAGE_SHIFT); return (howmany(npages >> (ddp_page_shift - PAGE_SHIFT), PPOD_PAGES)); } static int alloc_page_pods(struct ppod_region *pr, u_int nppods, u_int pgsz_idx, struct ppod_reservation *prsv) { vmem_addr_t addr; /* relative to start of region */ if (vmem_alloc(pr->pr_arena, PPOD_SZ(nppods), M_NOWAIT | M_FIRSTFIT, &addr) != 0) return (ENOMEM); #ifdef VERBOSE_TRACES CTR5(KTR_CXGBE, "%-17s arena %p, addr 0x%08x, nppods %d, pgsz %d", __func__, pr->pr_arena, (uint32_t)addr & pr->pr_tag_mask, nppods, 1 << pr->pr_page_shift[pgsz_idx]); #endif /* * The hardware tagmask includes an extra invalid bit but the arena was * seeded with valid values only. An allocation out of this arena will * fit inside the tagmask but won't have the invalid bit set. */ MPASS((addr & pr->pr_tag_mask) == addr); MPASS((addr & pr->pr_invalid_bit) == 0); prsv->prsv_pr = pr; prsv->prsv_tag = V_PPOD_PGSZ(pgsz_idx) | addr; prsv->prsv_nppods = nppods; return (0); } static int t4_alloc_page_pods_for_vmpages(struct ppod_region *pr, vm_page_t *pages, int npages, struct ppod_reservation *prsv) { int i, hcf, seglen, idx, nppods; /* * The DDP page size is unrelated to the VM page size. We combine * contiguous physical pages into larger segments to get the best DDP * page size possible. This is the largest of the four sizes in * A_ULP_RX_TDDP_PSZ that evenly divides the HCF of the segment sizes in * the page list. */ hcf = 0; for (i = 0; i < npages; i++) { seglen = PAGE_SIZE; while (i < npages - 1 && VM_PAGE_TO_PHYS(pages[i]) + PAGE_SIZE == VM_PAGE_TO_PHYS(pages[i + 1])) { seglen += PAGE_SIZE; i++; } hcf = calculate_hcf(hcf, seglen); if (hcf < (1 << pr->pr_page_shift[1])) { idx = 0; goto have_pgsz; /* give up, short circuit */ } } #define PR_PAGE_MASK(x) ((1 << pr->pr_page_shift[(x)]) - 1) MPASS((hcf & PR_PAGE_MASK(0)) == 0); /* PAGE_SIZE is >= 4K everywhere */ for (idx = nitems(pr->pr_page_shift) - 1; idx > 0; idx--) { if ((hcf & PR_PAGE_MASK(idx)) == 0) break; } #undef PR_PAGE_MASK have_pgsz: MPASS(idx <= M_PPOD_PGSZ); nppods = pages_to_nppods(npages, pr->pr_page_shift[idx]); if (alloc_page_pods(pr, nppods, idx, prsv) != 0) return (ENOMEM); MPASS(prsv->prsv_nppods > 0); return (0); } int t4_alloc_page_pods_for_ps(struct ppod_region *pr, struct pageset *ps) { struct ppod_reservation *prsv = &ps->prsv; KASSERT(prsv->prsv_nppods == 0, ("%s: page pods already allocated", __func__)); return (t4_alloc_page_pods_for_vmpages(pr, ps->pages, ps->npages, prsv)); } int t4_alloc_page_pods_for_bio(struct ppod_region *pr, struct bio *bp, struct ppod_reservation *prsv) { MPASS(bp->bio_flags & BIO_UNMAPPED); return (t4_alloc_page_pods_for_vmpages(pr, bp->bio_ma, bp->bio_ma_n, prsv)); } int t4_alloc_page_pods_for_buf(struct ppod_region *pr, vm_offset_t buf, int len, struct ppod_reservation *prsv) { int hcf, seglen, idx, npages, nppods; uintptr_t start_pva, end_pva, pva, p1; MPASS(buf > 0); MPASS(len > 0); /* * The DDP page size is unrelated to the VM page size. We combine * contiguous physical pages into larger segments to get the best DDP * page size possible. This is the largest of the four sizes in * A_ULP_RX_ISCSI_PSZ that evenly divides the HCF of the segment sizes * in the page list. */ hcf = 0; start_pva = trunc_page(buf); end_pva = trunc_page(buf + len - 1); pva = start_pva; while (pva <= end_pva) { seglen = PAGE_SIZE; p1 = pmap_kextract(pva); pva += PAGE_SIZE; while (pva <= end_pva && p1 + seglen == pmap_kextract(pva)) { seglen += PAGE_SIZE; pva += PAGE_SIZE; } hcf = calculate_hcf(hcf, seglen); if (hcf < (1 << pr->pr_page_shift[1])) { idx = 0; goto have_pgsz; /* give up, short circuit */ } } #define PR_PAGE_MASK(x) ((1 << pr->pr_page_shift[(x)]) - 1) MPASS((hcf & PR_PAGE_MASK(0)) == 0); /* PAGE_SIZE is >= 4K everywhere */ for (idx = nitems(pr->pr_page_shift) - 1; idx > 0; idx--) { if ((hcf & PR_PAGE_MASK(idx)) == 0) break; } #undef PR_PAGE_MASK have_pgsz: MPASS(idx <= M_PPOD_PGSZ); npages = 1; npages += (end_pva - start_pva) >> pr->pr_page_shift[idx]; nppods = howmany(npages, PPOD_PAGES); if (alloc_page_pods(pr, nppods, idx, prsv) != 0) return (ENOMEM); MPASS(prsv->prsv_nppods > 0); return (0); } int t4_alloc_page_pods_for_sgl(struct ppod_region *pr, struct ctl_sg_entry *sgl, int entries, struct ppod_reservation *prsv) { int hcf, seglen, idx = 0, npages, nppods, i, len; uintptr_t start_pva, end_pva, pva, p1 ; vm_offset_t buf; struct ctl_sg_entry *sge; MPASS(entries > 0); MPASS(sgl); /* * The DDP page size is unrelated to the VM page size. We combine * contiguous physical pages into larger segments to get the best DDP * page size possible. This is the largest of the four sizes in * A_ULP_RX_ISCSI_PSZ that evenly divides the HCF of the segment sizes * in the page list. */ hcf = 0; for (i = entries - 1; i >= 0; i--) { sge = sgl + i; buf = (vm_offset_t)sge->addr; len = sge->len; start_pva = trunc_page(buf); end_pva = trunc_page(buf + len - 1); pva = start_pva; while (pva <= end_pva) { seglen = PAGE_SIZE; p1 = pmap_kextract(pva); pva += PAGE_SIZE; while (pva <= end_pva && p1 + seglen == pmap_kextract(pva)) { seglen += PAGE_SIZE; pva += PAGE_SIZE; } hcf = calculate_hcf(hcf, seglen); if (hcf < (1 << pr->pr_page_shift[1])) { idx = 0; goto have_pgsz; /* give up, short circuit */ } } } #define PR_PAGE_MASK(x) ((1 << pr->pr_page_shift[(x)]) - 1) MPASS((hcf & PR_PAGE_MASK(0)) == 0); /* PAGE_SIZE is >= 4K everywhere */ for (idx = nitems(pr->pr_page_shift) - 1; idx > 0; idx--) { if ((hcf & PR_PAGE_MASK(idx)) == 0) break; } #undef PR_PAGE_MASK have_pgsz: MPASS(idx <= M_PPOD_PGSZ); npages = 0; while (entries--) { npages++; start_pva = trunc_page((vm_offset_t)sgl->addr); end_pva = trunc_page((vm_offset_t)sgl->addr + sgl->len - 1); npages += (end_pva - start_pva) >> pr->pr_page_shift[idx]; sgl = sgl + 1; } nppods = howmany(npages, PPOD_PAGES); if (alloc_page_pods(pr, nppods, idx, prsv) != 0) return (ENOMEM); MPASS(prsv->prsv_nppods > 0); return (0); } void t4_free_page_pods(struct ppod_reservation *prsv) { struct ppod_region *pr = prsv->prsv_pr; vmem_addr_t addr; MPASS(prsv != NULL); MPASS(prsv->prsv_nppods != 0); addr = prsv->prsv_tag & pr->pr_tag_mask; MPASS((addr & pr->pr_invalid_bit) == 0); #ifdef VERBOSE_TRACES CTR4(KTR_CXGBE, "%-17s arena %p, addr 0x%08x, nppods %d", __func__, pr->pr_arena, addr, prsv->prsv_nppods); #endif vmem_free(pr->pr_arena, addr, PPOD_SZ(prsv->prsv_nppods)); prsv->prsv_nppods = 0; } #define NUM_ULP_TX_SC_IMM_PPODS (256 / PPOD_SIZE) int t4_write_page_pods_for_ps(struct adapter *sc, struct sge_wrq *wrq, int tid, struct pageset *ps) { struct wrqe *wr; struct ulp_mem_io *ulpmc; struct ulptx_idata *ulpsc; struct pagepod *ppod; int i, j, k, n, chunk, len, ddp_pgsz, idx; u_int ppod_addr; uint32_t cmd; struct ppod_reservation *prsv = &ps->prsv; struct ppod_region *pr = prsv->prsv_pr; vm_paddr_t pa; KASSERT(!(ps->flags & PS_PPODS_WRITTEN), ("%s: page pods already written", __func__)); MPASS(prsv->prsv_nppods > 0); cmd = htobe32(V_ULPTX_CMD(ULP_TX_MEM_WRITE)); if (is_t4(sc)) cmd |= htobe32(F_ULP_MEMIO_ORDER); else cmd |= htobe32(F_T5_ULP_MEMIO_IMM); ddp_pgsz = 1 << pr->pr_page_shift[G_PPOD_PGSZ(prsv->prsv_tag)]; ppod_addr = pr->pr_start + (prsv->prsv_tag & pr->pr_tag_mask); for (i = 0; i < prsv->prsv_nppods; ppod_addr += chunk) { /* How many page pods are we writing in this cycle */ n = min(prsv->prsv_nppods - i, NUM_ULP_TX_SC_IMM_PPODS); chunk = PPOD_SZ(n); len = roundup2(sizeof(*ulpmc) + sizeof(*ulpsc) + chunk, 16); wr = alloc_wrqe(len, wrq); if (wr == NULL) return (ENOMEM); /* ok to just bail out */ ulpmc = wrtod(wr); INIT_ULPTX_WR(ulpmc, len, 0, 0); ulpmc->cmd = cmd; ulpmc->dlen = htobe32(V_ULP_MEMIO_DATA_LEN(chunk / 32)); ulpmc->len16 = htobe32(howmany(len - sizeof(ulpmc->wr), 16)); ulpmc->lock_addr = htobe32(V_ULP_MEMIO_ADDR(ppod_addr >> 5)); ulpsc = (struct ulptx_idata *)(ulpmc + 1); ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM)); ulpsc->len = htobe32(chunk); ppod = (struct pagepod *)(ulpsc + 1); for (j = 0; j < n; i++, j++, ppod++) { ppod->vld_tid_pgsz_tag_color = htobe64(F_PPOD_VALID | V_PPOD_TID(tid) | prsv->prsv_tag); ppod->len_offset = htobe64(V_PPOD_LEN(ps->len) | V_PPOD_OFST(ps->offset)); ppod->rsvd = 0; idx = i * PPOD_PAGES * (ddp_pgsz / PAGE_SIZE); for (k = 0; k < nitems(ppod->addr); k++) { if (idx < ps->npages) { pa = VM_PAGE_TO_PHYS(ps->pages[idx]); ppod->addr[k] = htobe64(pa); idx += ddp_pgsz / PAGE_SIZE; } else ppod->addr[k] = 0; #if 0 CTR5(KTR_CXGBE, "%s: tid %d ppod[%d]->addr[%d] = %p", __func__, tid, i, k, be64toh(ppod->addr[k])); #endif } } t4_wrq_tx(sc, wr); } ps->flags |= PS_PPODS_WRITTEN; return (0); } static struct mbuf * alloc_raw_wr_mbuf(int len) { struct mbuf *m; if (len <= MHLEN) m = m_gethdr(M_NOWAIT, MT_DATA); else if (len <= MCLBYTES) m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); else m = NULL; if (m == NULL) return (NULL); m->m_pkthdr.len = len; m->m_len = len; set_mbuf_raw_wr(m, true); return (m); } int t4_write_page_pods_for_bio(struct adapter *sc, struct toepcb *toep, struct ppod_reservation *prsv, struct bio *bp, struct mbufq *wrq) { struct ulp_mem_io *ulpmc; struct ulptx_idata *ulpsc; struct pagepod *ppod; int i, j, k, n, chunk, len, ddp_pgsz, idx; u_int ppod_addr; uint32_t cmd; struct ppod_region *pr = prsv->prsv_pr; vm_paddr_t pa; struct mbuf *m; MPASS(bp->bio_flags & BIO_UNMAPPED); cmd = htobe32(V_ULPTX_CMD(ULP_TX_MEM_WRITE)); if (is_t4(sc)) cmd |= htobe32(F_ULP_MEMIO_ORDER); else cmd |= htobe32(F_T5_ULP_MEMIO_IMM); ddp_pgsz = 1 << pr->pr_page_shift[G_PPOD_PGSZ(prsv->prsv_tag)]; ppod_addr = pr->pr_start + (prsv->prsv_tag & pr->pr_tag_mask); for (i = 0; i < prsv->prsv_nppods; ppod_addr += chunk) { /* How many page pods are we writing in this cycle */ n = min(prsv->prsv_nppods - i, NUM_ULP_TX_SC_IMM_PPODS); MPASS(n > 0); chunk = PPOD_SZ(n); len = roundup2(sizeof(*ulpmc) + sizeof(*ulpsc) + chunk, 16); m = alloc_raw_wr_mbuf(len); if (m == NULL) return (ENOMEM); ulpmc = mtod(m, struct ulp_mem_io *); INIT_ULPTX_WR(ulpmc, len, 0, toep->tid); ulpmc->cmd = cmd; ulpmc->dlen = htobe32(V_ULP_MEMIO_DATA_LEN(chunk / 32)); ulpmc->len16 = htobe32(howmany(len - sizeof(ulpmc->wr), 16)); ulpmc->lock_addr = htobe32(V_ULP_MEMIO_ADDR(ppod_addr >> 5)); ulpsc = (struct ulptx_idata *)(ulpmc + 1); ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM)); ulpsc->len = htobe32(chunk); ppod = (struct pagepod *)(ulpsc + 1); for (j = 0; j < n; i++, j++, ppod++) { ppod->vld_tid_pgsz_tag_color = htobe64(F_PPOD_VALID | V_PPOD_TID(toep->tid) | (prsv->prsv_tag & ~V_PPOD_PGSZ(M_PPOD_PGSZ))); ppod->len_offset = htobe64(V_PPOD_LEN(bp->bio_bcount) | V_PPOD_OFST(bp->bio_ma_offset)); ppod->rsvd = 0; idx = i * PPOD_PAGES * (ddp_pgsz / PAGE_SIZE); for (k = 0; k < nitems(ppod->addr); k++) { if (idx < bp->bio_ma_n) { pa = VM_PAGE_TO_PHYS(bp->bio_ma[idx]); ppod->addr[k] = htobe64(pa); idx += ddp_pgsz / PAGE_SIZE; } else ppod->addr[k] = 0; #if 0 CTR5(KTR_CXGBE, "%s: tid %d ppod[%d]->addr[%d] = %p", __func__, toep->tid, i, k, be64toh(ppod->addr[k])); #endif } } mbufq_enqueue(wrq, m); } return (0); } int t4_write_page_pods_for_buf(struct adapter *sc, struct toepcb *toep, struct ppod_reservation *prsv, vm_offset_t buf, int buflen, struct mbufq *wrq) { struct ulp_mem_io *ulpmc; struct ulptx_idata *ulpsc; struct pagepod *ppod; int i, j, k, n, chunk, len, ddp_pgsz; u_int ppod_addr, offset; uint32_t cmd; struct ppod_region *pr = prsv->prsv_pr; uintptr_t end_pva, pva; vm_paddr_t pa; struct mbuf *m; cmd = htobe32(V_ULPTX_CMD(ULP_TX_MEM_WRITE)); if (is_t4(sc)) cmd |= htobe32(F_ULP_MEMIO_ORDER); else cmd |= htobe32(F_T5_ULP_MEMIO_IMM); ddp_pgsz = 1 << pr->pr_page_shift[G_PPOD_PGSZ(prsv->prsv_tag)]; offset = buf & PAGE_MASK; ppod_addr = pr->pr_start + (prsv->prsv_tag & pr->pr_tag_mask); pva = trunc_page(buf); end_pva = trunc_page(buf + buflen - 1); for (i = 0; i < prsv->prsv_nppods; ppod_addr += chunk) { /* How many page pods are we writing in this cycle */ n = min(prsv->prsv_nppods - i, NUM_ULP_TX_SC_IMM_PPODS); MPASS(n > 0); chunk = PPOD_SZ(n); len = roundup2(sizeof(*ulpmc) + sizeof(*ulpsc) + chunk, 16); m = alloc_raw_wr_mbuf(len); if (m == NULL) return (ENOMEM); ulpmc = mtod(m, struct ulp_mem_io *); INIT_ULPTX_WR(ulpmc, len, 0, toep->tid); ulpmc->cmd = cmd; ulpmc->dlen = htobe32(V_ULP_MEMIO_DATA_LEN(chunk / 32)); ulpmc->len16 = htobe32(howmany(len - sizeof(ulpmc->wr), 16)); ulpmc->lock_addr = htobe32(V_ULP_MEMIO_ADDR(ppod_addr >> 5)); ulpsc = (struct ulptx_idata *)(ulpmc + 1); ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM)); ulpsc->len = htobe32(chunk); ppod = (struct pagepod *)(ulpsc + 1); for (j = 0; j < n; i++, j++, ppod++) { ppod->vld_tid_pgsz_tag_color = htobe64(F_PPOD_VALID | V_PPOD_TID(toep->tid) | (prsv->prsv_tag & ~V_PPOD_PGSZ(M_PPOD_PGSZ))); ppod->len_offset = htobe64(V_PPOD_LEN(buflen) | V_PPOD_OFST(offset)); ppod->rsvd = 0; for (k = 0; k < nitems(ppod->addr); k++) { if (pva > end_pva) ppod->addr[k] = 0; else { pa = pmap_kextract(pva); ppod->addr[k] = htobe64(pa); pva += ddp_pgsz; } #if 0 CTR5(KTR_CXGBE, "%s: tid %d ppod[%d]->addr[%d] = %p", __func__, toep->tid, i, k, be64toh(ppod->addr[k])); #endif } /* * Walk back 1 segment so that the first address in the * next pod is the same as the last one in the current * pod. */ pva -= ddp_pgsz; } mbufq_enqueue(wrq, m); } MPASS(pva <= end_pva); return (0); } int t4_write_page_pods_for_sgl(struct adapter *sc, struct toepcb *toep, struct ppod_reservation *prsv, struct ctl_sg_entry *sgl, int entries, int xferlen, struct mbufq *wrq) { struct ulp_mem_io *ulpmc; struct ulptx_idata *ulpsc; struct pagepod *ppod; int i, j, k, n, chunk, len, ddp_pgsz; u_int ppod_addr, offset, sg_offset = 0; uint32_t cmd; struct ppod_region *pr = prsv->prsv_pr; uintptr_t pva; vm_paddr_t pa; struct mbuf *m; MPASS(sgl != NULL); MPASS(entries > 0); cmd = htobe32(V_ULPTX_CMD(ULP_TX_MEM_WRITE)); if (is_t4(sc)) cmd |= htobe32(F_ULP_MEMIO_ORDER); else cmd |= htobe32(F_T5_ULP_MEMIO_IMM); ddp_pgsz = 1 << pr->pr_page_shift[G_PPOD_PGSZ(prsv->prsv_tag)]; offset = (vm_offset_t)sgl->addr & PAGE_MASK; ppod_addr = pr->pr_start + (prsv->prsv_tag & pr->pr_tag_mask); pva = trunc_page((vm_offset_t)sgl->addr); for (i = 0; i < prsv->prsv_nppods; ppod_addr += chunk) { /* How many page pods are we writing in this cycle */ n = min(prsv->prsv_nppods - i, NUM_ULP_TX_SC_IMM_PPODS); MPASS(n > 0); chunk = PPOD_SZ(n); len = roundup2(sizeof(*ulpmc) + sizeof(*ulpsc) + chunk, 16); m = alloc_raw_wr_mbuf(len); if (m == NULL) return (ENOMEM); ulpmc = mtod(m, struct ulp_mem_io *); INIT_ULPTX_WR(ulpmc, len, 0, toep->tid); ulpmc->cmd = cmd; ulpmc->dlen = htobe32(V_ULP_MEMIO_DATA_LEN(chunk / 32)); ulpmc->len16 = htobe32(howmany(len - sizeof(ulpmc->wr), 16)); ulpmc->lock_addr = htobe32(V_ULP_MEMIO_ADDR(ppod_addr >> 5)); ulpsc = (struct ulptx_idata *)(ulpmc + 1); ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM)); ulpsc->len = htobe32(chunk); ppod = (struct pagepod *)(ulpsc + 1); for (j = 0; j < n; i++, j++, ppod++) { ppod->vld_tid_pgsz_tag_color = htobe64(F_PPOD_VALID | V_PPOD_TID(toep->tid) | (prsv->prsv_tag & ~V_PPOD_PGSZ(M_PPOD_PGSZ))); ppod->len_offset = htobe64(V_PPOD_LEN(xferlen) | V_PPOD_OFST(offset)); ppod->rsvd = 0; for (k = 0; k < nitems(ppod->addr); k++) { if (entries != 0) { pa = pmap_kextract(pva + sg_offset); ppod->addr[k] = htobe64(pa); } else ppod->addr[k] = 0; #if 0 CTR5(KTR_CXGBE, "%s: tid %d ppod[%d]->addr[%d] = %p", __func__, toep->tid, i, k, be64toh(ppod->addr[k])); #endif /* * If this is the last entry in a pod, * reuse the same entry for first address * in the next pod. */ if (k + 1 == nitems(ppod->addr)) break; /* * Don't move to the next DDP page if the * sgl is already finished. */ if (entries == 0) continue; sg_offset += ddp_pgsz; if (sg_offset == sgl->len) { /* * This sgl entry is done. Go * to the next. */ entries--; sgl++; sg_offset = 0; if (entries != 0) pva = trunc_page( (vm_offset_t)sgl->addr); } } } mbufq_enqueue(wrq, m); } return (0); } /* * Prepare a pageset for DDP. This sets up page pods. */ static int prep_pageset(struct adapter *sc, struct toepcb *toep, struct pageset *ps) { struct tom_data *td = sc->tom_softc; if (ps->prsv.prsv_nppods == 0 && t4_alloc_page_pods_for_ps(&td->pr, ps) != 0) { return (0); } if (!(ps->flags & PS_PPODS_WRITTEN) && t4_write_page_pods_for_ps(sc, toep->ctrlq, toep->tid, ps) != 0) { return (0); } return (1); } int t4_init_ppod_region(struct ppod_region *pr, struct t4_range *r, u_int psz, const char *name) { int i; MPASS(pr != NULL); MPASS(r->size > 0); pr->pr_start = r->start; pr->pr_len = r->size; pr->pr_page_shift[0] = 12 + G_HPZ0(psz); pr->pr_page_shift[1] = 12 + G_HPZ1(psz); pr->pr_page_shift[2] = 12 + G_HPZ2(psz); pr->pr_page_shift[3] = 12 + G_HPZ3(psz); /* The SGL -> page pod algorithm requires the sizes to be in order. */ for (i = 1; i < nitems(pr->pr_page_shift); i++) { if (pr->pr_page_shift[i] <= pr->pr_page_shift[i - 1]) return (ENXIO); } pr->pr_tag_mask = ((1 << fls(r->size)) - 1) & V_PPOD_TAG(M_PPOD_TAG); pr->pr_alias_mask = V_PPOD_TAG(M_PPOD_TAG) & ~pr->pr_tag_mask; if (pr->pr_tag_mask == 0 || pr->pr_alias_mask == 0) return (ENXIO); pr->pr_alias_shift = fls(pr->pr_tag_mask); pr->pr_invalid_bit = 1 << (pr->pr_alias_shift - 1); pr->pr_arena = vmem_create(name, 0, pr->pr_len, PPOD_SIZE, 0, M_FIRSTFIT | M_NOWAIT); if (pr->pr_arena == NULL) return (ENOMEM); return (0); } void t4_free_ppod_region(struct ppod_region *pr) { MPASS(pr != NULL); if (pr->pr_arena) vmem_destroy(pr->pr_arena); bzero(pr, sizeof(*pr)); } static int pscmp(struct pageset *ps, struct vmspace *vm, vm_offset_t start, int npages, int pgoff, int len) { if (ps->start != start || ps->npages != npages || ps->offset != pgoff || ps->len != len) return (1); return (ps->vm != vm || ps->vm_timestamp != vm->vm_map.timestamp); } static int hold_aio(struct toepcb *toep, struct kaiocb *job, struct pageset **pps) { struct vmspace *vm; vm_map_t map; vm_offset_t start, end, pgoff; struct pageset *ps; int n; DDP_ASSERT_LOCKED(toep); /* * The AIO subsystem will cancel and drain all requests before * permitting a process to exit or exec, so p_vmspace should * be stable here. */ vm = job->userproc->p_vmspace; map = &vm->vm_map; start = (uintptr_t)job->uaiocb.aio_buf; pgoff = start & PAGE_MASK; end = round_page(start + job->uaiocb.aio_nbytes); start = trunc_page(start); if (end - start > MAX_DDP_BUFFER_SIZE) { /* * Truncate the request to a short read. * Alternatively, we could DDP in chunks to the larger * buffer, but that would be quite a bit more work. * * When truncating, round the request down to avoid * crossing a cache line on the final transaction. */ end = rounddown2(start + MAX_DDP_BUFFER_SIZE, CACHE_LINE_SIZE); #ifdef VERBOSE_TRACES CTR4(KTR_CXGBE, "%s: tid %d, truncating size from %lu to %lu", __func__, toep->tid, (unsigned long)job->uaiocb.aio_nbytes, (unsigned long)(end - (start + pgoff))); job->uaiocb.aio_nbytes = end - (start + pgoff); #endif end = round_page(end); } n = atop(end - start); /* * Try to reuse a cached pageset. */ TAILQ_FOREACH(ps, &toep->ddp.cached_pagesets, link) { if (pscmp(ps, vm, start, n, pgoff, job->uaiocb.aio_nbytes) == 0) { TAILQ_REMOVE(&toep->ddp.cached_pagesets, ps, link); toep->ddp.cached_count--; *pps = ps; return (0); } } /* * If there are too many cached pagesets to create a new one, * free a pageset before creating a new one. */ KASSERT(toep->ddp.active_count + toep->ddp.cached_count <= nitems(toep->ddp.db), ("%s: too many wired pagesets", __func__)); if (toep->ddp.active_count + toep->ddp.cached_count == nitems(toep->ddp.db)) { KASSERT(toep->ddp.cached_count > 0, ("no cached pageset to free")); ps = TAILQ_LAST(&toep->ddp.cached_pagesets, pagesetq); TAILQ_REMOVE(&toep->ddp.cached_pagesets, ps, link); toep->ddp.cached_count--; free_pageset(toep->td, ps); } DDP_UNLOCK(toep); /* Create a new pageset. */ ps = malloc(sizeof(*ps) + n * sizeof(vm_page_t), M_CXGBE, M_WAITOK | M_ZERO); ps->pages = (vm_page_t *)(ps + 1); ps->vm_timestamp = map->timestamp; ps->npages = vm_fault_quick_hold_pages(map, start, end - start, VM_PROT_WRITE, ps->pages, n); DDP_LOCK(toep); if (ps->npages < 0) { free(ps, M_CXGBE); return (EFAULT); } KASSERT(ps->npages == n, ("hold_aio: page count mismatch: %d vs %d", ps->npages, n)); ps->offset = pgoff; ps->len = job->uaiocb.aio_nbytes; refcount_acquire(&vm->vm_refcnt); ps->vm = vm; ps->start = start; CTR5(KTR_CXGBE, "%s: tid %d, new pageset %p for job %p, npages %d", __func__, toep->tid, ps, job, ps->npages); *pps = ps; return (0); } static void ddp_complete_all(struct toepcb *toep, int error) { struct kaiocb *job; DDP_ASSERT_LOCKED(toep); while (!TAILQ_EMPTY(&toep->ddp.aiojobq)) { job = TAILQ_FIRST(&toep->ddp.aiojobq); TAILQ_REMOVE(&toep->ddp.aiojobq, job, list); toep->ddp.waiting_count--; if (aio_clear_cancel_function(job)) ddp_complete_one(job, error); } } static void aio_ddp_cancel_one(struct kaiocb *job) { long copied; /* * If this job had copied data out of the socket buffer before * it was cancelled, report it as a short read rather than an * error. */ copied = job->aio_received; if (copied != 0) aio_complete(job, copied, 0); else aio_cancel(job); } /* * Called when the main loop wants to requeue a job to retry it later. * Deals with the race of the job being cancelled while it was being * examined. */ static void aio_ddp_requeue_one(struct toepcb *toep, struct kaiocb *job) { DDP_ASSERT_LOCKED(toep); if (!(toep->ddp.flags & DDP_DEAD) && aio_set_cancel_function(job, t4_aio_cancel_queued)) { TAILQ_INSERT_HEAD(&toep->ddp.aiojobq, job, list); toep->ddp.waiting_count++; } else aio_ddp_cancel_one(job); } static void aio_ddp_requeue(struct toepcb *toep) { struct adapter *sc = td_adapter(toep->td); struct socket *so; struct sockbuf *sb; struct inpcb *inp; struct kaiocb *job; struct ddp_buffer *db; size_t copied, offset, resid; struct pageset *ps; struct mbuf *m; uint64_t ddp_flags, ddp_flags_mask; struct wrqe *wr; int buf_flag, db_idx, error; DDP_ASSERT_LOCKED(toep); restart: if (toep->ddp.flags & DDP_DEAD) { MPASS(toep->ddp.waiting_count == 0); MPASS(toep->ddp.active_count == 0); return; } if (toep->ddp.waiting_count == 0 || toep->ddp.active_count == nitems(toep->ddp.db)) { return; } job = TAILQ_FIRST(&toep->ddp.aiojobq); so = job->fd_file->f_data; sb = &so->so_rcv; SOCKBUF_LOCK(sb); /* We will never get anything unless we are or were connected. */ if (!(so->so_state & (SS_ISCONNECTED|SS_ISDISCONNECTED))) { SOCKBUF_UNLOCK(sb); ddp_complete_all(toep, ENOTCONN); return; } KASSERT(toep->ddp.active_count == 0 || sbavail(sb) == 0, ("%s: pending sockbuf data and DDP is active", __func__)); /* Abort if socket has reported problems. */ /* XXX: Wait for any queued DDP's to finish and/or flush them? */ if (so->so_error && sbavail(sb) == 0) { toep->ddp.waiting_count--; TAILQ_REMOVE(&toep->ddp.aiojobq, job, list); if (!aio_clear_cancel_function(job)) { SOCKBUF_UNLOCK(sb); goto restart; } /* * If this job has previously copied some data, report * a short read and leave the error to be reported by * a future request. */ copied = job->aio_received; if (copied != 0) { SOCKBUF_UNLOCK(sb); aio_complete(job, copied, 0); goto restart; } error = so->so_error; so->so_error = 0; SOCKBUF_UNLOCK(sb); aio_complete(job, -1, error); goto restart; } /* * Door is closed. If there is pending data in the socket buffer, * deliver it. If there are pending DDP requests, wait for those * to complete. Once they have completed, return EOF reads. */ if (sb->sb_state & SBS_CANTRCVMORE && sbavail(sb) == 0) { SOCKBUF_UNLOCK(sb); if (toep->ddp.active_count != 0) return; ddp_complete_all(toep, 0); return; } /* * If DDP is not enabled and there is no pending socket buffer * data, try to enable DDP. */ if (sbavail(sb) == 0 && (toep->ddp.flags & DDP_ON) == 0) { SOCKBUF_UNLOCK(sb); /* * Wait for the card to ACK that DDP is enabled before * queueing any buffers. Currently this waits for an * indicate to arrive. This could use a TCB_SET_FIELD_RPL * message to know that DDP was enabled instead of waiting * for the indicate which would avoid copying the indicate * if no data is pending. * * XXX: Might want to limit the indicate size to the size * of the first queued request. */ if ((toep->ddp.flags & DDP_SC_REQ) == 0) enable_ddp(sc, toep); return; } SOCKBUF_UNLOCK(sb); /* * If another thread is queueing a buffer for DDP, let it * drain any work and return. */ if (toep->ddp.queueing != NULL) return; /* Take the next job to prep it for DDP. */ toep->ddp.waiting_count--; TAILQ_REMOVE(&toep->ddp.aiojobq, job, list); if (!aio_clear_cancel_function(job)) goto restart; toep->ddp.queueing = job; /* NB: This drops DDP_LOCK while it holds the backing VM pages. */ error = hold_aio(toep, job, &ps); if (error != 0) { ddp_complete_one(job, error); toep->ddp.queueing = NULL; goto restart; } SOCKBUF_LOCK(sb); if (so->so_error && sbavail(sb) == 0) { copied = job->aio_received; if (copied != 0) { SOCKBUF_UNLOCK(sb); recycle_pageset(toep, ps); aio_complete(job, copied, 0); toep->ddp.queueing = NULL; goto restart; } error = so->so_error; so->so_error = 0; SOCKBUF_UNLOCK(sb); recycle_pageset(toep, ps); aio_complete(job, -1, error); toep->ddp.queueing = NULL; goto restart; } if (sb->sb_state & SBS_CANTRCVMORE && sbavail(sb) == 0) { SOCKBUF_UNLOCK(sb); recycle_pageset(toep, ps); if (toep->ddp.active_count != 0) { /* * The door is closed, but there are still pending * DDP buffers. Requeue. These jobs will all be * completed once those buffers drain. */ aio_ddp_requeue_one(toep, job); toep->ddp.queueing = NULL; return; } ddp_complete_one(job, 0); ddp_complete_all(toep, 0); toep->ddp.queueing = NULL; return; } sbcopy: /* * If the toep is dead, there shouldn't be any data in the socket * buffer, so the above case should have handled this. */ MPASS(!(toep->ddp.flags & DDP_DEAD)); /* * If there is pending data in the socket buffer (either * from before the requests were queued or a DDP indicate), * copy those mbufs out directly. */ copied = 0; offset = ps->offset + job->aio_received; MPASS(job->aio_received <= job->uaiocb.aio_nbytes); resid = job->uaiocb.aio_nbytes - job->aio_received; m = sb->sb_mb; KASSERT(m == NULL || toep->ddp.active_count == 0, ("%s: sockbuf data with active DDP", __func__)); while (m != NULL && resid > 0) { struct iovec iov[1]; struct uio uio; #ifdef INVARIANTS int error; #endif iov[0].iov_base = mtod(m, void *); iov[0].iov_len = m->m_len; if (iov[0].iov_len > resid) iov[0].iov_len = resid; uio.uio_iov = iov; uio.uio_iovcnt = 1; uio.uio_offset = 0; uio.uio_resid = iov[0].iov_len; uio.uio_segflg = UIO_SYSSPACE; uio.uio_rw = UIO_WRITE; #ifdef INVARIANTS error = uiomove_fromphys(ps->pages, offset + copied, uio.uio_resid, &uio); #else uiomove_fromphys(ps->pages, offset + copied, uio.uio_resid, &uio); #endif MPASS(error == 0 && uio.uio_resid == 0); copied += uio.uio_offset; resid -= uio.uio_offset; m = m->m_next; } if (copied != 0) { sbdrop_locked(sb, copied); job->aio_received += copied; job->msgrcv = 1; copied = job->aio_received; inp = sotoinpcb(so); if (!INP_TRY_WLOCK(inp)) { /* * The reference on the socket file descriptor in * the AIO job should keep 'sb' and 'inp' stable. * Our caller has a reference on the 'toep' that * keeps it stable. */ SOCKBUF_UNLOCK(sb); DDP_UNLOCK(toep); INP_WLOCK(inp); DDP_LOCK(toep); SOCKBUF_LOCK(sb); /* * If the socket has been closed, we should detect * that and complete this request if needed on * the next trip around the loop. */ } t4_rcvd_locked(&toep->td->tod, intotcpcb(inp)); INP_WUNLOCK(inp); if (resid == 0 || toep->ddp.flags & DDP_DEAD) { /* * We filled the entire buffer with socket * data, DDP is not being used, or the socket * is being shut down, so complete the * request. */ SOCKBUF_UNLOCK(sb); recycle_pageset(toep, ps); aio_complete(job, copied, 0); toep->ddp.queueing = NULL; goto restart; } /* * If DDP is not enabled, requeue this request and restart. * This will either enable DDP or wait for more data to * arrive on the socket buffer. */ if ((toep->ddp.flags & (DDP_ON | DDP_SC_REQ)) != DDP_ON) { SOCKBUF_UNLOCK(sb); recycle_pageset(toep, ps); aio_ddp_requeue_one(toep, job); toep->ddp.queueing = NULL; goto restart; } /* * An indicate might have arrived and been added to * the socket buffer while it was unlocked after the * copy to lock the INP. If so, restart the copy. */ if (sbavail(sb) != 0) goto sbcopy; } SOCKBUF_UNLOCK(sb); if (prep_pageset(sc, toep, ps) == 0) { recycle_pageset(toep, ps); aio_ddp_requeue_one(toep, job); toep->ddp.queueing = NULL; /* * XXX: Need to retry this later. Mostly need a trigger * when page pods are freed up. */ printf("%s: prep_pageset failed\n", __func__); return; } /* Determine which DDP buffer to use. */ if (toep->ddp.db[0].job == NULL) { db_idx = 0; } else { MPASS(toep->ddp.db[1].job == NULL); db_idx = 1; } ddp_flags = 0; ddp_flags_mask = 0; if (db_idx == 0) { ddp_flags |= V_TF_DDP_BUF0_VALID(1); if (so->so_state & SS_NBIO) ddp_flags |= V_TF_DDP_BUF0_FLUSH(1); ddp_flags_mask |= V_TF_DDP_PSH_NO_INVALIDATE0(1) | V_TF_DDP_PUSH_DISABLE_0(1) | V_TF_DDP_PSHF_ENABLE_0(1) | V_TF_DDP_BUF0_FLUSH(1) | V_TF_DDP_BUF0_VALID(1); buf_flag = DDP_BUF0_ACTIVE; } else { ddp_flags |= V_TF_DDP_BUF1_VALID(1); if (so->so_state & SS_NBIO) ddp_flags |= V_TF_DDP_BUF1_FLUSH(1); ddp_flags_mask |= V_TF_DDP_PSH_NO_INVALIDATE1(1) | V_TF_DDP_PUSH_DISABLE_1(1) | V_TF_DDP_PSHF_ENABLE_1(1) | V_TF_DDP_BUF1_FLUSH(1) | V_TF_DDP_BUF1_VALID(1); buf_flag = DDP_BUF1_ACTIVE; } MPASS((toep->ddp.flags & buf_flag) == 0); if ((toep->ddp.flags & (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE)) == 0) { MPASS(db_idx == 0); MPASS(toep->ddp.active_id == -1); MPASS(toep->ddp.active_count == 0); ddp_flags_mask |= V_TF_DDP_ACTIVE_BUF(1); } /* * The TID for this connection should still be valid. If DDP_DEAD * is set, SBS_CANTRCVMORE should be set, so we shouldn't be * this far anyway. Even if the socket is closing on the other * end, the AIO job holds a reference on this end of the socket * which will keep it open and keep the TCP PCB attached until * after the job is completed. */ wr = mk_update_tcb_for_ddp(sc, toep, db_idx, ps, job->aio_received, ddp_flags, ddp_flags_mask); if (wr == NULL) { recycle_pageset(toep, ps); aio_ddp_requeue_one(toep, job); toep->ddp.queueing = NULL; /* * XXX: Need a way to kick a retry here. * * XXX: We know the fixed size needed and could * preallocate this using a blocking request at the * start of the task to avoid having to handle this * edge case. */ printf("%s: mk_update_tcb_for_ddp failed\n", __func__); return; } if (!aio_set_cancel_function(job, t4_aio_cancel_active)) { free_wrqe(wr); recycle_pageset(toep, ps); aio_ddp_cancel_one(job); toep->ddp.queueing = NULL; goto restart; } #ifdef VERBOSE_TRACES CTR6(KTR_CXGBE, "%s: tid %u, scheduling %p for DDP[%d] (flags %#lx/%#lx)", __func__, toep->tid, job, db_idx, ddp_flags, ddp_flags_mask); #endif /* Give the chip the go-ahead. */ t4_wrq_tx(sc, wr); db = &toep->ddp.db[db_idx]; db->cancel_pending = 0; db->job = job; db->ps = ps; toep->ddp.queueing = NULL; toep->ddp.flags |= buf_flag; toep->ddp.active_count++; if (toep->ddp.active_count == 1) { MPASS(toep->ddp.active_id == -1); toep->ddp.active_id = db_idx; CTR2(KTR_CXGBE, "%s: ddp_active_id = %d", __func__, toep->ddp.active_id); } goto restart; } void ddp_queue_toep(struct toepcb *toep) { DDP_ASSERT_LOCKED(toep); if (toep->ddp.flags & DDP_TASK_ACTIVE) return; toep->ddp.flags |= DDP_TASK_ACTIVE; hold_toepcb(toep); soaio_enqueue(&toep->ddp.requeue_task); } static void aio_ddp_requeue_task(void *context, int pending) { struct toepcb *toep = context; DDP_LOCK(toep); aio_ddp_requeue(toep); toep->ddp.flags &= ~DDP_TASK_ACTIVE; DDP_UNLOCK(toep); free_toepcb(toep); } static void t4_aio_cancel_active(struct kaiocb *job) { struct socket *so = job->fd_file->f_data; struct tcpcb *tp = so_sototcpcb(so); struct toepcb *toep = tp->t_toe; struct adapter *sc = td_adapter(toep->td); uint64_t valid_flag; int i; DDP_LOCK(toep); if (aio_cancel_cleared(job)) { DDP_UNLOCK(toep); aio_ddp_cancel_one(job); return; } for (i = 0; i < nitems(toep->ddp.db); i++) { if (toep->ddp.db[i].job == job) { /* Should only ever get one cancel request for a job. */ MPASS(toep->ddp.db[i].cancel_pending == 0); /* * Invalidate this buffer. It will be * cancelled or partially completed once the * card ACKs the invalidate. */ valid_flag = i == 0 ? V_TF_DDP_BUF0_VALID(1) : V_TF_DDP_BUF1_VALID(1); t4_set_tcb_field(sc, toep->ctrlq, toep, W_TCB_RX_DDP_FLAGS, valid_flag, 0, 1, CPL_COOKIE_DDP0 + i); toep->ddp.db[i].cancel_pending = 1; CTR2(KTR_CXGBE, "%s: request %p marked pending", __func__, job); break; } } DDP_UNLOCK(toep); } static void t4_aio_cancel_queued(struct kaiocb *job) { struct socket *so = job->fd_file->f_data; struct tcpcb *tp = so_sototcpcb(so); struct toepcb *toep = tp->t_toe; DDP_LOCK(toep); if (!aio_cancel_cleared(job)) { TAILQ_REMOVE(&toep->ddp.aiojobq, job, list); toep->ddp.waiting_count--; if (toep->ddp.waiting_count == 0) ddp_queue_toep(toep); } CTR2(KTR_CXGBE, "%s: request %p cancelled", __func__, job); DDP_UNLOCK(toep); aio_ddp_cancel_one(job); } int t4_aio_queue_ddp(struct socket *so, struct kaiocb *job) { struct tcpcb *tp = so_sototcpcb(so); struct toepcb *toep = tp->t_toe; /* Ignore writes. */ if (job->uaiocb.aio_lio_opcode != LIO_READ) return (EOPNOTSUPP); DDP_LOCK(toep); /* * XXX: Think about possibly returning errors for ENOTCONN, * etc. Perhaps the caller would only queue the request * if it failed with EOPNOTSUPP? */ #ifdef VERBOSE_TRACES CTR3(KTR_CXGBE, "%s: queueing %p for tid %u", __func__, job, toep->tid); #endif if (!aio_set_cancel_function(job, t4_aio_cancel_queued)) panic("new job was cancelled"); TAILQ_INSERT_TAIL(&toep->ddp.aiojobq, job, list); toep->ddp.waiting_count++; toep->ddp.flags |= DDP_OK; /* * Try to handle this request synchronously. If this has * to block because the task is running, it will just bail * and let the task handle it instead. */ aio_ddp_requeue(toep); DDP_UNLOCK(toep); return (0); } void t4_ddp_mod_load(void) { t4_register_shared_cpl_handler(CPL_SET_TCB_RPL, do_ddp_tcb_rpl, CPL_COOKIE_DDP0); t4_register_shared_cpl_handler(CPL_SET_TCB_RPL, do_ddp_tcb_rpl, CPL_COOKIE_DDP1); t4_register_cpl_handler(CPL_RX_DATA_DDP, do_rx_data_ddp); t4_register_cpl_handler(CPL_RX_DDP_COMPLETE, do_rx_ddp_complete); TAILQ_INIT(&ddp_orphan_pagesets); mtx_init(&ddp_orphan_pagesets_lock, "ddp orphans", NULL, MTX_DEF); TASK_INIT(&ddp_orphan_task, 0, ddp_free_orphan_pagesets, NULL); } void t4_ddp_mod_unload(void) { taskqueue_drain(taskqueue_thread, &ddp_orphan_task); MPASS(TAILQ_EMPTY(&ddp_orphan_pagesets)); mtx_destroy(&ddp_orphan_pagesets_lock); t4_register_shared_cpl_handler(CPL_SET_TCB_RPL, NULL, CPL_COOKIE_DDP0); t4_register_shared_cpl_handler(CPL_SET_TCB_RPL, NULL, CPL_COOKIE_DDP1); t4_register_cpl_handler(CPL_RX_DATA_DDP, NULL); t4_register_cpl_handler(CPL_RX_DDP_COMPLETE, NULL); } #endif diff --git a/sys/dev/cxgbe/tom/t4_tls.c b/sys/dev/cxgbe/tom/t4_tls.c index 447943b5b9db..9f9d4a48de93 100644 --- a/sys/dev/cxgbe/tom/t4_tls.c +++ b/sys/dev/cxgbe/tom/t4_tls.c @@ -1,1242 +1,1242 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2017-2018 Chelsio Communications, Inc. * All rights reserved. * Written by: John Baldwin * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include "opt_inet.h" #include "opt_kern_tls.h" #include __FBSDID("$FreeBSD$"); #ifdef KERN_TLS #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef TCP_OFFLOAD #include "common/common.h" #include "common/t4_tcb.h" #include "crypto/t4_crypto.h" #include "tom/t4_tom_l2t.h" #include "tom/t4_tom.h" /* * The TCP sequence number of a CPL_TLS_DATA mbuf is saved here while * the mbuf is in the ulp_pdu_reclaimq. */ #define tls_tcp_seq PH_loc.thirtytwo[0] static void t4_set_tls_tcb_field(struct toepcb *toep, uint16_t word, uint64_t mask, uint64_t val) { struct adapter *sc = td_adapter(toep->td); t4_set_tcb_field(sc, &toep->ofld_txq->wrq, toep, word, mask, val, 0, 0); } /* TLS and DTLS common routines */ bool can_tls_offload(struct adapter *sc) { return (sc->tt.tls && sc->cryptocaps & FW_CAPS_CONFIG_TLSKEYS); } int tls_tx_key(struct toepcb *toep) { struct tls_ofld_info *tls_ofld = &toep->tls; return (tls_ofld->tx_key_addr >= 0); } /* Set TLS Key-Id in TCB */ static void t4_set_tls_keyid(struct toepcb *toep, unsigned int key_id) { t4_set_tls_tcb_field(toep, W_TCB_RX_TLS_KEY_TAG, V_TCB_RX_TLS_KEY_TAG(M_TCB_RX_TLS_BUF_TAG), V_TCB_RX_TLS_KEY_TAG(key_id)); } /* Clear TF_RX_QUIESCE to re-enable receive. */ static void t4_clear_rx_quiesce(struct toepcb *toep) { t4_set_tls_tcb_field(toep, W_TCB_T_FLAGS, V_TF_RX_QUIESCE(1), 0); } static void tls_clr_ofld_mode(struct toepcb *toep) { tls_stop_handshake_timer(toep); KASSERT(toep->tls.rx_key_addr == -1, ("%s: tid %d has RX key", __func__, toep->tid)); /* Switch to plain TOE mode. */ t4_set_tls_tcb_field(toep, W_TCB_ULP_RAW, V_TCB_ULP_RAW(V_TF_TLS_ENABLE(1)), V_TCB_ULP_RAW(V_TF_TLS_ENABLE(0))); t4_set_tls_tcb_field(toep, W_TCB_ULP_TYPE, V_TCB_ULP_TYPE(M_TCB_ULP_TYPE), V_TCB_ULP_TYPE(ULP_MODE_NONE)); t4_clear_rx_quiesce(toep); toep->flags &= ~(TPF_FORCE_CREDITS | TPF_TLS_ESTABLISHED); toep->params.ulp_mode = ULP_MODE_NONE; } /* TLS/DTLS content type for CPL SFO */ static inline unsigned char tls_content_type(unsigned char content_type) { switch (content_type) { case CONTENT_TYPE_CCS: return CPL_TX_TLS_SFO_TYPE_CCS; case CONTENT_TYPE_ALERT: return CPL_TX_TLS_SFO_TYPE_ALERT; case CONTENT_TYPE_HANDSHAKE: return CPL_TX_TLS_SFO_TYPE_HANDSHAKE; case CONTENT_TYPE_APP_DATA: return CPL_TX_TLS_SFO_TYPE_DATA; default: return CPL_TX_TLS_SFO_TYPE_CUSTOM; } } /* TLS Key memory management */ static void clear_tls_keyid(struct toepcb *toep) { struct tls_ofld_info *tls_ofld = &toep->tls; struct adapter *sc = td_adapter(toep->td); if (tls_ofld->rx_key_addr >= 0) { t4_free_tls_keyid(sc, tls_ofld->rx_key_addr); tls_ofld->rx_key_addr = -1; } if (tls_ofld->tx_key_addr >= 0) { t4_free_tls_keyid(sc, tls_ofld->tx_key_addr); tls_ofld->tx_key_addr = -1; } } static int get_tp_plen_max(struct ktls_session *tls) { int plen = ((min(3*4096, TP_TX_PG_SZ))/1448) * 1448; return (tls->params.max_frame_len <= 8192 ? plen : FC_TP_PLEN_MAX); } /* Send request to get the key-id */ static int tls_program_key_id(struct toepcb *toep, struct ktls_session *tls, int direction) { struct tls_ofld_info *tls_ofld = &toep->tls; struct adapter *sc = td_adapter(toep->td); struct ofld_tx_sdesc *txsd; int keyid; struct wrqe *wr; struct tls_key_req *kwr; struct tls_keyctx *kctx; #ifdef INVARIANTS int kwrlen, kctxlen, len; kwrlen = sizeof(*kwr); kctxlen = roundup2(sizeof(*kctx), 32); len = roundup2(kwrlen + kctxlen, 16); MPASS(TLS_KEY_WR_SZ == len); #endif if (toep->txsd_avail == 0) return (EAGAIN); if ((keyid = t4_alloc_tls_keyid(sc)) < 0) { return (ENOSPC); } wr = alloc_wrqe(TLS_KEY_WR_SZ, &toep->ofld_txq->wrq); if (wr == NULL) { t4_free_tls_keyid(sc, keyid); return (ENOMEM); } kwr = wrtod(wr); memset(kwr, 0, TLS_KEY_WR_SZ); t4_write_tlskey_wr(tls, direction, toep->tid, F_FW_WR_COMPL, keyid, kwr); kctx = (struct tls_keyctx *)(kwr + 1); if (direction == KTLS_TX) tls_ofld->tx_key_addr = keyid; else tls_ofld->rx_key_addr = keyid; t4_tls_key_ctx(tls, direction, kctx); txsd = &toep->txsd[toep->txsd_pidx]; txsd->tx_credits = DIV_ROUND_UP(TLS_KEY_WR_SZ, 16); txsd->plen = 0; toep->tx_credits -= txsd->tx_credits; if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) toep->txsd_pidx = 0; toep->txsd_avail--; t4_wrq_tx(sc, wr); return (0); } /* * In some cases a client connection can hang without sending the * ServerHelloDone message from the NIC to the host. Send a dummy * RX_DATA_ACK with RX_MODULATE to unstick the connection. */ static void tls_send_handshake_ack(void *arg) { struct toepcb *toep = arg; struct tls_ofld_info *tls_ofld = &toep->tls; struct adapter *sc = td_adapter(toep->td); /* Bail without rescheduling if the connection has closed. */ if ((toep->flags & (TPF_FIN_SENT | TPF_ABORT_SHUTDOWN)) != 0) return; /* * If this connection has timed out without receiving more * data, downgrade to plain TOE mode and don't re-arm the * timer. */ if (sc->tt.tls_rx_timeout != 0) { struct inpcb *inp; struct tcpcb *tp; inp = toep->inp; tp = intotcpcb(inp); if ((ticks - tp->t_rcvtime) >= sc->tt.tls_rx_timeout) { CTR2(KTR_CXGBE, "%s: tid %d clr_ofld_mode", __func__, toep->tid); tls_clr_ofld_mode(toep); return; } } /* * XXX: Does not have the t4_get_tcb() checks to refine the * workaround. */ callout_schedule(&tls_ofld->handshake_timer, TLS_SRV_HELLO_RD_TM * hz); CTR2(KTR_CXGBE, "%s: tid %d sending RX_DATA_ACK", __func__, toep->tid); send_rx_modulate(sc, toep); } static void tls_start_handshake_timer(struct toepcb *toep) { struct tls_ofld_info *tls_ofld = &toep->tls; INP_WLOCK_ASSERT(toep->inp); callout_reset(&tls_ofld->handshake_timer, TLS_SRV_HELLO_BKOFF_TM * hz, tls_send_handshake_ack, toep); } void tls_stop_handshake_timer(struct toepcb *toep) { struct tls_ofld_info *tls_ofld = &toep->tls; INP_WLOCK_ASSERT(toep->inp); callout_stop(&tls_ofld->handshake_timer); } int tls_alloc_ktls(struct toepcb *toep, struct ktls_session *tls, int direction) { struct adapter *sc = td_adapter(toep->td); int error, explicit_iv_size, key_offset, mac_first; if (!can_tls_offload(td_adapter(toep->td))) return (EINVAL); switch (ulp_mode(toep)) { case ULP_MODE_TLS: break; case ULP_MODE_NONE: case ULP_MODE_TCPDDP: if (direction != KTLS_TX) return (EINVAL); break; default: return (EINVAL); } switch (tls->params.cipher_algorithm) { case CRYPTO_AES_CBC: /* XXX: Explicitly ignore any provided IV. */ switch (tls->params.cipher_key_len) { case 128 / 8: case 192 / 8: case 256 / 8: break; default: error = EINVAL; goto clr_ofld; } switch (tls->params.auth_algorithm) { case CRYPTO_SHA1_HMAC: case CRYPTO_SHA2_256_HMAC: case CRYPTO_SHA2_384_HMAC: break; default: error = EPROTONOSUPPORT; goto clr_ofld; } explicit_iv_size = AES_BLOCK_LEN; mac_first = 1; break; case CRYPTO_AES_NIST_GCM_16: if (tls->params.iv_len != SALT_SIZE) { error = EINVAL; goto clr_ofld; } switch (tls->params.cipher_key_len) { case 128 / 8: case 192 / 8: case 256 / 8: break; default: error = EINVAL; goto clr_ofld; } explicit_iv_size = 8; mac_first = 0; break; default: error = EPROTONOSUPPORT; goto clr_ofld; } /* Only TLS 1.1 and TLS 1.2 are currently supported. */ if (tls->params.tls_vmajor != TLS_MAJOR_VER_ONE || tls->params.tls_vminor < TLS_MINOR_VER_ONE || tls->params.tls_vminor > TLS_MINOR_VER_TWO) { error = EPROTONOSUPPORT; goto clr_ofld; } /* Bail if we already have a key. */ if (direction == KTLS_TX) { if (toep->tls.tx_key_addr != -1) return (EOPNOTSUPP); } else { if (toep->tls.rx_key_addr != -1) return (EOPNOTSUPP); } error = tls_program_key_id(toep, tls, direction); if (error) { if (direction == KTLS_RX) goto clr_ofld; return (error); } if (direction == KTLS_TX) { toep->tls.scmd0.seqno_numivs = (V_SCMD_SEQ_NO_CTRL(3) | V_SCMD_PROTO_VERSION(t4_tls_proto_ver(tls)) | V_SCMD_ENC_DEC_CTRL(SCMD_ENCDECCTRL_ENCRYPT) | V_SCMD_CIPH_AUTH_SEQ_CTRL((mac_first == 0)) | V_SCMD_CIPH_MODE(t4_tls_cipher_mode(tls)) | V_SCMD_AUTH_MODE(t4_tls_auth_mode(tls)) | V_SCMD_HMAC_CTRL(t4_tls_hmac_ctrl(tls)) | V_SCMD_IV_SIZE(explicit_iv_size / 2)); toep->tls.scmd0.ivgen_hdrlen = (V_SCMD_IV_GEN_CTRL(1) | V_SCMD_KEY_CTX_INLINE(0) | V_SCMD_TLS_FRAG_ENABLE(1)); toep->tls.iv_len = explicit_iv_size; toep->tls.frag_size = tls->params.max_frame_len; toep->tls.fcplenmax = get_tp_plen_max(tls); toep->tls.expn_per_ulp = tls->params.tls_hlen + tls->params.tls_tlen; toep->tls.pdus_per_ulp = 1; toep->tls.adjusted_plen = toep->tls.expn_per_ulp + tls->params.max_frame_len; toep->tls.tx_key_info_size = t4_tls_key_info_size(tls); } else { /* Stop timer on handshake completion */ tls_stop_handshake_timer(toep); toep->flags &= ~TPF_FORCE_CREDITS; toep->flags |= TPF_TLS_RECEIVE; toep->tls.rx_version = tls->params.tls_vmajor << 8 | tls->params.tls_vminor; /* * RX key tags are an index into the key portion of MA * memory stored as an offset from the base address in * units of 64 bytes. */ key_offset = toep->tls.rx_key_addr - sc->vres.key.start; t4_set_tls_keyid(toep, key_offset / 64); t4_set_tls_tcb_field(toep, W_TCB_ULP_RAW, V_TCB_ULP_RAW(M_TCB_ULP_RAW), V_TCB_ULP_RAW((V_TF_TLS_KEY_SIZE(3) | V_TF_TLS_CONTROL(1) | V_TF_TLS_ACTIVE(1) | V_TF_TLS_ENABLE(1)))); t4_set_tls_tcb_field(toep, W_TCB_TLS_SEQ, V_TCB_TLS_SEQ(M_TCB_TLS_SEQ), V_TCB_TLS_SEQ(0)); t4_clear_rx_quiesce(toep); } return (0); clr_ofld: if (ulp_mode(toep) == ULP_MODE_TLS) { CTR2(KTR_CXGBE, "%s: tid %d clr_ofld_mode", __func__, toep->tid); tls_clr_ofld_mode(toep); } return (error); } void tls_init_toep(struct toepcb *toep) { struct tls_ofld_info *tls_ofld = &toep->tls; tls_ofld->rx_key_addr = -1; tls_ofld->tx_key_addr = -1; } void tls_establish(struct toepcb *toep) { /* * Enable PDU extraction. * * XXX: Supposedly this should be done by the firmware when * the ULP_MODE FLOWC parameter is set in send_flowc_wr(), but * in practice this seems to be required. */ CTR2(KTR_CXGBE, "%s: tid %d setting TLS_ENABLE", __func__, toep->tid); t4_set_tls_tcb_field(toep, W_TCB_ULP_RAW, V_TCB_ULP_RAW(M_TCB_ULP_RAW), V_TCB_ULP_RAW(V_TF_TLS_ENABLE(1))); toep->flags |= TPF_FORCE_CREDITS | TPF_TLS_ESTABLISHED; callout_init_rw(&toep->tls.handshake_timer, &toep->inp->inp_lock, 0); tls_start_handshake_timer(toep); } void tls_detach(struct toepcb *toep) { if (toep->flags & TPF_TLS_ESTABLISHED) { tls_stop_handshake_timer(toep); toep->flags &= ~TPF_TLS_ESTABLISHED; } } void tls_uninit_toep(struct toepcb *toep) { MPASS((toep->flags & TPF_TLS_ESTABLISHED) == 0); clear_tls_keyid(toep); } #define MAX_OFLD_TX_CREDITS (SGE_MAX_WR_LEN / 16) #define MIN_OFLD_TLSTX_CREDITS(toep) \ (howmany(sizeof(struct fw_tlstx_data_wr) + \ sizeof(struct cpl_tx_tls_sfo) + sizeof(struct ulptx_idata) + \ sizeof(struct ulptx_sc_memrd) + \ AES_BLOCK_LEN + 1, 16)) static void write_tlstx_wr(struct fw_tlstx_data_wr *txwr, struct toepcb *toep, unsigned int plen, unsigned int expn, uint8_t credits, int shove) { struct tls_ofld_info *tls_ofld = &toep->tls; unsigned int len = plen + expn; txwr->op_to_immdlen = htobe32(V_WR_OP(FW_TLSTX_DATA_WR) | V_FW_TLSTX_DATA_WR_COMPL(1) | V_FW_TLSTX_DATA_WR_IMMDLEN(0)); txwr->flowid_len16 = htobe32(V_FW_TLSTX_DATA_WR_FLOWID(toep->tid) | V_FW_TLSTX_DATA_WR_LEN16(credits)); txwr->plen = htobe32(len); txwr->lsodisable_to_flags = htobe32(V_TX_ULP_MODE(ULP_MODE_TLS) | V_TX_URG(0) | /* F_T6_TX_FORCE | */ V_TX_SHOVE(shove)); txwr->ctxloc_to_exp = htobe32(V_FW_TLSTX_DATA_WR_NUMIVS(1) | V_FW_TLSTX_DATA_WR_EXP(expn) | V_FW_TLSTX_DATA_WR_CTXLOC(TLS_SFO_WR_CONTEXTLOC_DDR) | V_FW_TLSTX_DATA_WR_IVDSGL(0) | V_FW_TLSTX_DATA_WR_KEYSIZE(tls_ofld->tx_key_info_size >> 4)); txwr->mfs = htobe16(tls_ofld->frag_size); txwr->adjustedplen_pkd = htobe16( V_FW_TLSTX_DATA_WR_ADJUSTEDPLEN(tls_ofld->adjusted_plen)); txwr->expinplenmax_pkd = htobe16( V_FW_TLSTX_DATA_WR_EXPINPLENMAX(tls_ofld->expn_per_ulp)); txwr->pdusinplenmax_pkd = V_FW_TLSTX_DATA_WR_PDUSINPLENMAX(tls_ofld->pdus_per_ulp); } static void write_tlstx_cpl(struct cpl_tx_tls_sfo *cpl, struct toepcb *toep, struct tls_hdr *tls_hdr, unsigned int plen, uint64_t seqno) { struct tls_ofld_info *tls_ofld = &toep->tls; int data_type, seglen; seglen = plen; data_type = tls_content_type(tls_hdr->type); cpl->op_to_seg_len = htobe32(V_CPL_TX_TLS_SFO_OPCODE(CPL_TX_TLS_SFO) | V_CPL_TX_TLS_SFO_DATA_TYPE(data_type) | V_CPL_TX_TLS_SFO_CPL_LEN(2) | V_CPL_TX_TLS_SFO_SEG_LEN(seglen)); cpl->pld_len = htobe32(plen); if (data_type == CPL_TX_TLS_SFO_TYPE_CUSTOM) cpl->type_protover = htobe32( V_CPL_TX_TLS_SFO_TYPE(tls_hdr->type)); cpl->seqno_numivs = htobe32(tls_ofld->scmd0.seqno_numivs | V_SCMD_NUM_IVS(1)); cpl->ivgen_hdrlen = htobe32(tls_ofld->scmd0.ivgen_hdrlen); cpl->scmd1 = htobe64(seqno); } static int count_ext_pgs_segs(struct mbuf *m) { vm_paddr_t nextpa; u_int i, nsegs; MPASS(m->m_epg_npgs > 0); nsegs = 1; nextpa = m->m_epg_pa[0] + PAGE_SIZE; for (i = 1; i < m->m_epg_npgs; i++) { if (nextpa != m->m_epg_pa[i]) nsegs++; nextpa = m->m_epg_pa[i] + PAGE_SIZE; } return (nsegs); } static void write_ktlstx_sgl(void *dst, struct mbuf *m, int nsegs) { struct ulptx_sgl *usgl = dst; vm_paddr_t pa; uint32_t len; int i, j; KASSERT(nsegs > 0, ("%s: nsegs 0", __func__)); usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) | V_ULPTX_NSGE(nsegs)); /* Figure out the first S/G length. */ pa = m->m_epg_pa[0] + m->m_epg_1st_off; usgl->addr0 = htobe64(pa); len = m_epg_pagelen(m, 0, m->m_epg_1st_off); pa += len; for (i = 1; i < m->m_epg_npgs; i++) { if (m->m_epg_pa[i] != pa) break; len += m_epg_pagelen(m, i, 0); pa += m_epg_pagelen(m, i, 0); } usgl->len0 = htobe32(len); #ifdef INVARIANTS nsegs--; #endif j = -1; for (; i < m->m_epg_npgs; i++) { if (j == -1 || m->m_epg_pa[i] != pa) { if (j >= 0) usgl->sge[j / 2].len[j & 1] = htobe32(len); j++; #ifdef INVARIANTS nsegs--; #endif pa = m->m_epg_pa[i]; usgl->sge[j / 2].addr[j & 1] = htobe64(pa); len = m_epg_pagelen(m, i, 0); pa += len; } else { len += m_epg_pagelen(m, i, 0); pa += m_epg_pagelen(m, i, 0); } } if (j >= 0) { usgl->sge[j / 2].len[j & 1] = htobe32(len); if ((j & 1) == 0) usgl->sge[j / 2].len[1] = htobe32(0); } KASSERT(nsegs == 0, ("%s: nsegs %d, m %p", __func__, nsegs, m)); } /* * Similar to t4_push_frames() but handles sockets that contain TLS * record mbufs. */ void t4_push_ktls(struct adapter *sc, struct toepcb *toep, int drop) { struct tls_hdr *thdr; struct fw_tlstx_data_wr *txwr; struct cpl_tx_tls_sfo *cpl; struct ulptx_idata *idata; struct ulptx_sc_memrd *memrd; struct wrqe *wr; struct mbuf *m; u_int nsegs, credits, wr_len; u_int expn_size; struct inpcb *inp = toep->inp; struct tcpcb *tp = intotcpcb(inp); struct socket *so = inp->inp_socket; struct sockbuf *sb = &so->so_snd; int tls_size, tx_credits, shove, sowwakeup; struct ofld_tx_sdesc *txsd; char *buf; INP_WLOCK_ASSERT(inp); KASSERT(toep->flags & TPF_FLOWC_WR_SENT, ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid)); KASSERT(ulp_mode(toep) == ULP_MODE_NONE || ulp_mode(toep) == ULP_MODE_TCPDDP || ulp_mode(toep) == ULP_MODE_TLS, ("%s: ulp_mode %u for toep %p", __func__, ulp_mode(toep), toep)); KASSERT(tls_tx_key(toep), ("%s: TX key not set for toep %p", __func__, toep)); #ifdef VERBOSE_TRACES CTR4(KTR_CXGBE, "%s: tid %d toep flags %#x tp flags %#x drop %d", __func__, toep->tid, toep->flags, tp->t_flags); #endif if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN)) return; #ifdef RATELIMIT if (__predict_false(inp->inp_flags2 & INP_RATE_LIMIT_CHANGED) && (update_tx_rate_limit(sc, toep, so->so_max_pacing_rate) == 0)) { inp->inp_flags2 &= ~INP_RATE_LIMIT_CHANGED; } #endif /* * This function doesn't resume by itself. Someone else must clear the * flag and call this function. */ if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) { KASSERT(drop == 0, ("%s: drop (%d) != 0 but tx is suspended", __func__, drop)); return; } txsd = &toep->txsd[toep->txsd_pidx]; for (;;) { tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS); SOCKBUF_LOCK(sb); sowwakeup = drop; if (drop) { sbdrop_locked(sb, drop); drop = 0; } m = sb->sb_sndptr != NULL ? sb->sb_sndptr->m_next : sb->sb_mb; /* * Send a FIN if requested, but only if there's no * more data to send. */ if (m == NULL && toep->flags & TPF_SEND_FIN) { if (sowwakeup) sowwakeup_locked(so); else SOCKBUF_UNLOCK(sb); SOCKBUF_UNLOCK_ASSERT(sb); t4_close_conn(sc, toep); return; } /* * If there is no ready data to send, wait until more * data arrives. */ if (m == NULL || (m->m_flags & M_NOTAVAIL) != 0) { if (sowwakeup) sowwakeup_locked(so); else SOCKBUF_UNLOCK(sb); SOCKBUF_UNLOCK_ASSERT(sb); #ifdef VERBOSE_TRACES CTR2(KTR_CXGBE, "%s: tid %d no ready data to send", __func__, toep->tid); #endif return; } KASSERT(m->m_flags & M_EXTPG, ("%s: mbuf %p is not NOMAP", __func__, m)); KASSERT(m->m_epg_tls != NULL, ("%s: mbuf %p doesn't have TLS session", __func__, m)); /* Calculate WR length. */ wr_len = sizeof(struct fw_tlstx_data_wr) + sizeof(struct cpl_tx_tls_sfo) + sizeof(struct ulptx_idata) + sizeof(struct ulptx_sc_memrd); /* Explicit IVs for AES-CBC and AES-GCM are <= 16. */ MPASS(toep->tls.iv_len <= AES_BLOCK_LEN); wr_len += AES_BLOCK_LEN; /* Account for SGL in work request length. */ nsegs = count_ext_pgs_segs(m); wr_len += sizeof(struct ulptx_sgl) + ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8; /* Not enough credits for this work request. */ if (howmany(wr_len, 16) > tx_credits) { if (sowwakeup) sowwakeup_locked(so); else SOCKBUF_UNLOCK(sb); SOCKBUF_UNLOCK_ASSERT(sb); #ifdef VERBOSE_TRACES CTR5(KTR_CXGBE, "%s: tid %d mbuf %p requires %d credits, but only %d available", __func__, toep->tid, m, howmany(wr_len, 16), tx_credits); #endif toep->flags |= TPF_TX_SUSPENDED; return; } /* Shove if there is no additional data pending. */ shove = ((m->m_next == NULL || (m->m_next->m_flags & M_NOTAVAIL) != 0)) && (tp->t_flags & TF_MORETOCOME) == 0; if (sb->sb_flags & SB_AUTOSIZE && V_tcp_do_autosndbuf && sb->sb_hiwat < V_tcp_autosndbuf_max && sbused(sb) >= sb->sb_hiwat * 7 / 8) { int newsize = min(sb->sb_hiwat + V_tcp_autosndbuf_inc, V_tcp_autosndbuf_max); if (!sbreserve_locked(so, SO_SND, newsize, NULL)) sb->sb_flags &= ~SB_AUTOSIZE; else sowwakeup = 1; /* room available */ } if (sowwakeup) sowwakeup_locked(so); else SOCKBUF_UNLOCK(sb); SOCKBUF_UNLOCK_ASSERT(sb); if (__predict_false(toep->flags & TPF_FIN_SENT)) panic("%s: excess tx.", __func__); wr = alloc_wrqe(roundup2(wr_len, 16), &toep->ofld_txq->wrq); if (wr == NULL) { /* XXX: how will we recover from this? */ toep->flags |= TPF_TX_SUSPENDED; return; } thdr = (struct tls_hdr *)&m->m_epg_hdr; #ifdef VERBOSE_TRACES CTR5(KTR_CXGBE, "%s: tid %d TLS record %ju type %d len %#x", __func__, toep->tid, m->m_epg_seqno, thdr->type, m->m_len); #endif txwr = wrtod(wr); cpl = (struct cpl_tx_tls_sfo *)(txwr + 1); memset(txwr, 0, roundup2(wr_len, 16)); credits = howmany(wr_len, 16); expn_size = m->m_epg_hdrlen + m->m_epg_trllen; tls_size = m->m_len - expn_size; write_tlstx_wr(txwr, toep, tls_size, expn_size, credits, shove); write_tlstx_cpl(cpl, toep, thdr, tls_size, m->m_epg_seqno); idata = (struct ulptx_idata *)(cpl + 1); idata->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_NOOP)); idata->len = htobe32(0); memrd = (struct ulptx_sc_memrd *)(idata + 1); memrd->cmd_to_len = htobe32(V_ULPTX_CMD(ULP_TX_SC_MEMRD) | V_ULP_TX_SC_MORE(1) | V_ULPTX_LEN16(toep->tls.tx_key_info_size >> 4)); memrd->addr = htobe32(toep->tls.tx_key_addr >> 5); /* Copy IV. */ buf = (char *)(memrd + 1); memcpy(buf, thdr + 1, toep->tls.iv_len); buf += AES_BLOCK_LEN; write_ktlstx_sgl(buf, m, nsegs); KASSERT(toep->tx_credits >= credits, ("%s: not enough credits", __func__)); toep->tx_credits -= credits; tp->snd_nxt += m->m_len; tp->snd_max += m->m_len; SOCKBUF_LOCK(sb); sb->sb_sndptr = m; SOCKBUF_UNLOCK(sb); toep->flags |= TPF_TX_DATA_SENT; if (toep->tx_credits < MIN_OFLD_TLSTX_CREDITS(toep)) toep->flags |= TPF_TX_SUSPENDED; KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__)); txsd->plen = m->m_len; txsd->tx_credits = credits; txsd++; if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) { toep->txsd_pidx = 0; txsd = &toep->txsd[0]; } toep->txsd_avail--; counter_u64_add(toep->ofld_txq->tx_toe_tls_records, 1); counter_u64_add(toep->ofld_txq->tx_toe_tls_octets, m->m_len); t4_l2t_send(sc, wr, toep->l2te); } } /* * For TLS data we place received mbufs received via CPL_TLS_DATA into * an mbufq in the TLS offload state. When CPL_RX_TLS_CMP is * received, the completed PDUs are placed into the socket receive * buffer. * * The TLS code reuses the ulp_pdu_reclaimq to hold the pending mbufs. */ static int do_tls_data(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { struct adapter *sc = iq->adapter; const struct cpl_tls_data *cpl = mtod(m, const void *); unsigned int tid = GET_TID(cpl); struct toepcb *toep = lookup_tid(sc, tid); struct inpcb *inp = toep->inp; struct tcpcb *tp; int len; /* XXX: Should this match do_rx_data instead? */ KASSERT(!(toep->flags & TPF_SYNQE), ("%s: toep %p claims to be a synq entry", __func__, toep)); KASSERT(toep->tid == tid, ("%s: toep tid/atid mismatch", __func__)); /* strip off CPL header */ m_adj(m, sizeof(*cpl)); len = m->m_pkthdr.len; toep->ofld_rxq->rx_toe_tls_octets += len; KASSERT(len == G_CPL_TLS_DATA_LENGTH(be32toh(cpl->length_pkd)), ("%s: payload length mismatch", __func__)); INP_WLOCK(inp); - if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) { + if (inp->inp_flags & INP_DROPPED) { CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x", __func__, tid, len, inp->inp_flags); INP_WUNLOCK(inp); m_freem(m); return (0); } /* Save TCP sequence number. */ m->m_pkthdr.tls_tcp_seq = be32toh(cpl->seq); if (mbufq_enqueue(&toep->ulp_pdu_reclaimq, m)) { #ifdef INVARIANTS panic("Failed to queue TLS data packet"); #else printf("%s: Failed to queue TLS data packet\n", __func__); INP_WUNLOCK(inp); m_freem(m); return (0); #endif } tp = intotcpcb(inp); tp->t_rcvtime = ticks; #ifdef VERBOSE_TRACES CTR4(KTR_CXGBE, "%s: tid %u len %d seq %u", __func__, tid, len, be32toh(cpl->seq)); #endif INP_WUNLOCK(inp); return (0); } static int do_rx_tls_cmp(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { struct adapter *sc = iq->adapter; const struct cpl_rx_tls_cmp *cpl = mtod(m, const void *); struct tlsrx_hdr_pkt *tls_hdr_pkt; unsigned int tid = GET_TID(cpl); struct toepcb *toep = lookup_tid(sc, tid); struct inpcb *inp = toep->inp; struct tcpcb *tp; struct socket *so; struct sockbuf *sb; struct mbuf *tls_data; struct tls_get_record *tgr; struct mbuf *control; int pdu_length, rx_credits; #if defined(KTR) || defined(INVARIANTS) int len; #endif KASSERT(toep->tid == tid, ("%s: toep tid/atid mismatch", __func__)); KASSERT(!(toep->flags & TPF_SYNQE), ("%s: toep %p claims to be a synq entry", __func__, toep)); /* strip off CPL header */ m_adj(m, sizeof(*cpl)); #if defined(KTR) || defined(INVARIANTS) len = m->m_pkthdr.len; #endif toep->ofld_rxq->rx_toe_tls_records++; KASSERT(len == G_CPL_RX_TLS_CMP_LENGTH(be32toh(cpl->pdulength_length)), ("%s: payload length mismatch", __func__)); INP_WLOCK(inp); - if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) { + if (inp->inp_flags & INP_DROPPED) { CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x", __func__, tid, len, inp->inp_flags); INP_WUNLOCK(inp); m_freem(m); return (0); } pdu_length = G_CPL_RX_TLS_CMP_PDULENGTH(be32toh(cpl->pdulength_length)); so = inp_inpcbtosocket(inp); tp = intotcpcb(inp); #ifdef VERBOSE_TRACES CTR6(KTR_CXGBE, "%s: tid %u PDU len %d len %d seq %u, rcv_nxt %u", __func__, tid, pdu_length, len, be32toh(cpl->seq), tp->rcv_nxt); #endif tp->rcv_nxt += pdu_length; KASSERT(tp->rcv_wnd >= pdu_length, ("%s: negative window size", __func__)); tp->rcv_wnd -= pdu_length; /* XXX: Not sure what to do about urgent data. */ /* * The payload of this CPL is the TLS header followed by * additional fields. */ KASSERT(m->m_len >= sizeof(*tls_hdr_pkt), ("%s: payload too small", __func__)); tls_hdr_pkt = mtod(m, void *); tls_data = mbufq_dequeue(&toep->ulp_pdu_reclaimq); if (tls_data != NULL) { KASSERT(be32toh(cpl->seq) == tls_data->m_pkthdr.tls_tcp_seq, ("%s: sequence mismatch", __func__)); } /* Report decryption errors as EBADMSG. */ if ((tls_hdr_pkt->res_to_mac_error & M_TLSRX_HDR_PKT_ERROR) != 0) { m_freem(m); m_freem(tls_data); CURVNET_SET(toep->vnet); so->so_error = EBADMSG; sorwakeup(so); INP_WUNLOCK(inp); CURVNET_RESTORE(); return (0); } /* Allocate the control message mbuf. */ control = sbcreatecontrol(NULL, sizeof(*tgr), TLS_GET_RECORD, IPPROTO_TCP, M_NOWAIT); if (control == NULL) { m_freem(m); m_freem(tls_data); CURVNET_SET(toep->vnet); so->so_error = ENOBUFS; sorwakeup(so); INP_WUNLOCK(inp); CURVNET_RESTORE(); return (0); } tgr = (struct tls_get_record *) CMSG_DATA(mtod(control, struct cmsghdr *)); memset(tgr, 0, sizeof(*tgr)); tgr->tls_type = tls_hdr_pkt->type; tgr->tls_vmajor = be16toh(tls_hdr_pkt->version) >> 8; tgr->tls_vminor = be16toh(tls_hdr_pkt->version) & 0xff; m_freem(m); if (tls_data != NULL) { m_last(tls_data)->m_flags |= M_EOR; tgr->tls_length = htobe16(tls_data->m_pkthdr.len); } else tgr->tls_length = 0; m = tls_data; sb = &so->so_rcv; SOCKBUF_LOCK(sb); if (__predict_false(sb->sb_state & SBS_CANTRCVMORE)) { struct epoch_tracker et; CTR3(KTR_CXGBE, "%s: tid %u, excess rx (%d bytes)", __func__, tid, pdu_length); m_freem(m); m_freem(control); SOCKBUF_UNLOCK(sb); INP_WUNLOCK(inp); CURVNET_SET(toep->vnet); NET_EPOCH_ENTER(et); INP_WLOCK(inp); tp = tcp_drop(tp, ECONNRESET); if (tp) INP_WUNLOCK(inp); NET_EPOCH_EXIT(et); CURVNET_RESTORE(); return (0); } /* * Not all of the bytes on the wire are included in the socket buffer * (e.g. the MAC of the TLS record). However, those bytes are included * in the TCP sequence space. */ /* receive buffer autosize */ MPASS(toep->vnet == so->so_vnet); CURVNET_SET(toep->vnet); if (sb->sb_flags & SB_AUTOSIZE && V_tcp_do_autorcvbuf && sb->sb_hiwat < V_tcp_autorcvbuf_max && m->m_pkthdr.len > (sbspace(sb) / 8 * 7)) { unsigned int hiwat = sb->sb_hiwat; unsigned int newsize = min(hiwat + sc->tt.autorcvbuf_inc, V_tcp_autorcvbuf_max); if (!sbreserve_locked(so, SO_RCV, newsize, NULL)) sb->sb_flags &= ~SB_AUTOSIZE; } sbappendcontrol_locked(sb, m, control, 0); rx_credits = sbspace(sb) > tp->rcv_wnd ? sbspace(sb) - tp->rcv_wnd : 0; #ifdef VERBOSE_TRACES CTR4(KTR_CXGBE, "%s: tid %u rx_credits %u rcv_wnd %u", __func__, tid, rx_credits, tp->rcv_wnd); #endif if (rx_credits > 0 && sbused(sb) + tp->rcv_wnd < sb->sb_lowat) { rx_credits = send_rx_credits(sc, toep, rx_credits); tp->rcv_wnd += rx_credits; tp->rcv_adv += rx_credits; } sorwakeup_locked(so); SOCKBUF_UNLOCK_ASSERT(sb); INP_WUNLOCK(inp); CURVNET_RESTORE(); return (0); } void do_rx_data_tls(const struct cpl_rx_data *cpl, struct toepcb *toep, struct mbuf *m) { struct inpcb *inp = toep->inp; struct tls_ofld_info *tls_ofld = &toep->tls; struct tls_hdr *hdr; struct tcpcb *tp; struct socket *so; struct sockbuf *sb; int len, rx_credits; len = m->m_pkthdr.len; INP_WLOCK_ASSERT(inp); so = inp_inpcbtosocket(inp); tp = intotcpcb(inp); sb = &so->so_rcv; SOCKBUF_LOCK(sb); CURVNET_SET(toep->vnet); tp->rcv_nxt += len; KASSERT(tp->rcv_wnd >= len, ("%s: negative window size", __func__)); tp->rcv_wnd -= len; /* Do we have a full TLS header? */ if (len < sizeof(*hdr)) { CTR3(KTR_CXGBE, "%s: tid %u len %d: too short for a TLS header", __func__, toep->tid, len); so->so_error = EMSGSIZE; goto out; } hdr = mtod(m, struct tls_hdr *); /* Is the header valid? */ if (be16toh(hdr->version) != tls_ofld->rx_version) { CTR3(KTR_CXGBE, "%s: tid %u invalid version %04x", __func__, toep->tid, be16toh(hdr->version)); so->so_error = EINVAL; goto out; } if (be16toh(hdr->length) < sizeof(*hdr)) { CTR3(KTR_CXGBE, "%s: tid %u invalid length %u", __func__, toep->tid, be16toh(hdr->length)); so->so_error = EBADMSG; goto out; } /* Did we get a truncated record? */ if (len < be16toh(hdr->length)) { CTR4(KTR_CXGBE, "%s: tid %u truncated TLS record (%d vs %u)", __func__, toep->tid, len, be16toh(hdr->length)); so->so_error = EMSGSIZE; goto out; } /* Is the header type unknown? */ switch (hdr->type) { case CONTENT_TYPE_CCS: case CONTENT_TYPE_ALERT: case CONTENT_TYPE_APP_DATA: case CONTENT_TYPE_HANDSHAKE: break; default: CTR3(KTR_CXGBE, "%s: tid %u invalid TLS record type %u", __func__, toep->tid, hdr->type); so->so_error = EBADMSG; goto out; } /* * Just punt. Although this could fall back to software * decryption, this case should never really happen. */ CTR4(KTR_CXGBE, "%s: tid %u dropping TLS record type %u, length %u", __func__, toep->tid, hdr->type, be16toh(hdr->length)); so->so_error = EBADMSG; out: /* * This connection is going to die anyway, so probably don't * need to bother with returning credits. */ rx_credits = sbspace(sb) > tp->rcv_wnd ? sbspace(sb) - tp->rcv_wnd : 0; #ifdef VERBOSE_TRACES CTR4(KTR_CXGBE, "%s: tid %u rx_credits %u rcv_wnd %u", __func__, toep->tid, rx_credits, tp->rcv_wnd); #endif if (rx_credits > 0 && sbused(sb) + tp->rcv_wnd < sb->sb_lowat) { rx_credits = send_rx_credits(toep->vi->adapter, toep, rx_credits); tp->rcv_wnd += rx_credits; tp->rcv_adv += rx_credits; } sorwakeup_locked(so); SOCKBUF_UNLOCK_ASSERT(sb); INP_WUNLOCK(inp); CURVNET_RESTORE(); m_freem(m); } void t4_tls_mod_load(void) { t4_register_cpl_handler(CPL_TLS_DATA, do_tls_data); t4_register_cpl_handler(CPL_RX_TLS_CMP, do_rx_tls_cmp); } void t4_tls_mod_unload(void) { t4_register_cpl_handler(CPL_TLS_DATA, NULL); t4_register_cpl_handler(CPL_RX_TLS_CMP, NULL); } #endif /* TCP_OFFLOAD */ #endif /* KERN_TLS */ diff --git a/sys/kern/uipc_ktls.c b/sys/kern/uipc_ktls.c index bf91b774c0dc..83fcc9e70546 100644 --- a/sys/kern/uipc_ktls.c +++ b/sys/kern/uipc_ktls.c @@ -1,3254 +1,3253 @@ /*- * SPDX-License-Identifier: BSD-2-Clause * * Copyright (c) 2014-2019 Netflix Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include "opt_kern_tls.h" #include "opt_ratelimit.h" #include "opt_rss.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__) #include #endif #include #include #include #ifdef RSS #include #include #endif #include #include #if defined(INET) || defined(INET6) #include #include #endif #include #ifdef TCP_OFFLOAD #include #endif #include #include #include #include #include #include #include struct ktls_wq { struct mtx mtx; STAILQ_HEAD(, mbuf) m_head; STAILQ_HEAD(, socket) so_head; bool running; int lastallocfail; } __aligned(CACHE_LINE_SIZE); struct ktls_alloc_thread { uint64_t wakeups; uint64_t allocs; struct thread *td; int running; }; struct ktls_domain_info { int count; int cpu[MAXCPU]; struct ktls_alloc_thread alloc_td; }; struct ktls_domain_info ktls_domains[MAXMEMDOM]; static struct ktls_wq *ktls_wq; static struct proc *ktls_proc; static uma_zone_t ktls_session_zone; static uma_zone_t ktls_buffer_zone; static uint16_t ktls_cpuid_lookup[MAXCPU]; static int ktls_init_state; static struct sx ktls_init_lock; SX_SYSINIT(ktls_init_lock, &ktls_init_lock, "ktls init"); SYSCTL_NODE(_kern_ipc, OID_AUTO, tls, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Kernel TLS offload"); SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Kernel TLS offload stats"); #ifdef RSS static int ktls_bind_threads = 1; #else static int ktls_bind_threads; #endif SYSCTL_INT(_kern_ipc_tls, OID_AUTO, bind_threads, CTLFLAG_RDTUN, &ktls_bind_threads, 0, "Bind crypto threads to cores (1) or cores and domains (2) at boot"); static u_int ktls_maxlen = 16384; SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, maxlen, CTLFLAG_RDTUN, &ktls_maxlen, 0, "Maximum TLS record size"); static int ktls_number_threads; SYSCTL_INT(_kern_ipc_tls_stats, OID_AUTO, threads, CTLFLAG_RD, &ktls_number_threads, 0, "Number of TLS threads in thread-pool"); unsigned int ktls_ifnet_max_rexmit_pct = 2; SYSCTL_UINT(_kern_ipc_tls, OID_AUTO, ifnet_max_rexmit_pct, CTLFLAG_RWTUN, &ktls_ifnet_max_rexmit_pct, 2, "Max percent bytes retransmitted before ifnet TLS is disabled"); static bool ktls_offload_enable; SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, enable, CTLFLAG_RWTUN, &ktls_offload_enable, 0, "Enable support for kernel TLS offload"); static bool ktls_cbc_enable = true; SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, cbc_enable, CTLFLAG_RWTUN, &ktls_cbc_enable, 1, "Enable Support of AES-CBC crypto for kernel TLS"); static bool ktls_sw_buffer_cache = true; SYSCTL_BOOL(_kern_ipc_tls, OID_AUTO, sw_buffer_cache, CTLFLAG_RDTUN, &ktls_sw_buffer_cache, 1, "Enable caching of output buffers for SW encryption"); static int ktls_max_alloc = 128; SYSCTL_INT(_kern_ipc_tls, OID_AUTO, max_alloc, CTLFLAG_RWTUN, &ktls_max_alloc, 128, "Max number of 16k buffers to allocate in thread context"); static COUNTER_U64_DEFINE_EARLY(ktls_tasks_active); SYSCTL_COUNTER_U64(_kern_ipc_tls, OID_AUTO, tasks_active, CTLFLAG_RD, &ktls_tasks_active, "Number of active tasks"); static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_pending); SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_pending, CTLFLAG_RD, &ktls_cnt_tx_pending, "Number of TLS 1.0 records waiting for earlier TLS records"); static COUNTER_U64_DEFINE_EARLY(ktls_cnt_tx_queued); SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_tx_inqueue, CTLFLAG_RD, &ktls_cnt_tx_queued, "Number of TLS records in queue to tasks for SW encryption"); static COUNTER_U64_DEFINE_EARLY(ktls_cnt_rx_queued); SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, sw_rx_inqueue, CTLFLAG_RD, &ktls_cnt_rx_queued, "Number of TLS sockets in queue to tasks for SW decryption"); static COUNTER_U64_DEFINE_EARLY(ktls_offload_total); SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, offload_total, CTLFLAG_RD, &ktls_offload_total, "Total successful TLS setups (parameters set)"); static COUNTER_U64_DEFINE_EARLY(ktls_offload_enable_calls); SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, enable_calls, CTLFLAG_RD, &ktls_offload_enable_calls, "Total number of TLS enable calls made"); static COUNTER_U64_DEFINE_EARLY(ktls_offload_active); SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, active, CTLFLAG_RD, &ktls_offload_active, "Total Active TLS sessions"); static COUNTER_U64_DEFINE_EARLY(ktls_offload_corrupted_records); SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, corrupted_records, CTLFLAG_RD, &ktls_offload_corrupted_records, "Total corrupted TLS records received"); static COUNTER_U64_DEFINE_EARLY(ktls_offload_failed_crypto); SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, failed_crypto, CTLFLAG_RD, &ktls_offload_failed_crypto, "Total TLS crypto failures"); static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_ifnet); SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_ifnet, CTLFLAG_RD, &ktls_switch_to_ifnet, "TLS sessions switched from SW to ifnet"); static COUNTER_U64_DEFINE_EARLY(ktls_switch_to_sw); SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_to_sw, CTLFLAG_RD, &ktls_switch_to_sw, "TLS sessions switched from ifnet to SW"); static COUNTER_U64_DEFINE_EARLY(ktls_switch_failed); SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, switch_failed, CTLFLAG_RD, &ktls_switch_failed, "TLS sessions unable to switch between SW and ifnet"); static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_fail); SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_failed, CTLFLAG_RD, &ktls_ifnet_disable_fail, "TLS sessions unable to switch to SW from ifnet"); static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_disable_ok); SYSCTL_COUNTER_U64(_kern_ipc_tls_stats, OID_AUTO, ifnet_disable_ok, CTLFLAG_RD, &ktls_ifnet_disable_ok, "TLS sessions able to switch to SW from ifnet"); SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, sw, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "Software TLS session stats"); SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, ifnet, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "Hardware (ifnet) TLS session stats"); #ifdef TCP_OFFLOAD SYSCTL_NODE(_kern_ipc_tls, OID_AUTO, toe, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "TOE TLS session stats"); #endif static COUNTER_U64_DEFINE_EARLY(ktls_sw_cbc); SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, cbc, CTLFLAG_RD, &ktls_sw_cbc, "Active number of software TLS sessions using AES-CBC"); static COUNTER_U64_DEFINE_EARLY(ktls_sw_gcm); SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, gcm, CTLFLAG_RD, &ktls_sw_gcm, "Active number of software TLS sessions using AES-GCM"); static COUNTER_U64_DEFINE_EARLY(ktls_sw_chacha20); SYSCTL_COUNTER_U64(_kern_ipc_tls_sw, OID_AUTO, chacha20, CTLFLAG_RD, &ktls_sw_chacha20, "Active number of software TLS sessions using Chacha20-Poly1305"); static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_cbc); SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, cbc, CTLFLAG_RD, &ktls_ifnet_cbc, "Active number of ifnet TLS sessions using AES-CBC"); static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_gcm); SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, gcm, CTLFLAG_RD, &ktls_ifnet_gcm, "Active number of ifnet TLS sessions using AES-GCM"); static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_chacha20); SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, chacha20, CTLFLAG_RD, &ktls_ifnet_chacha20, "Active number of ifnet TLS sessions using Chacha20-Poly1305"); static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset); SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset, CTLFLAG_RD, &ktls_ifnet_reset, "TLS sessions updated to a new ifnet send tag"); static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_dropped); SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_dropped, CTLFLAG_RD, &ktls_ifnet_reset_dropped, "TLS sessions dropped after failing to update ifnet send tag"); static COUNTER_U64_DEFINE_EARLY(ktls_ifnet_reset_failed); SYSCTL_COUNTER_U64(_kern_ipc_tls_ifnet, OID_AUTO, reset_failed, CTLFLAG_RD, &ktls_ifnet_reset_failed, "TLS sessions that failed to allocate a new ifnet send tag"); static int ktls_ifnet_permitted; SYSCTL_UINT(_kern_ipc_tls_ifnet, OID_AUTO, permitted, CTLFLAG_RWTUN, &ktls_ifnet_permitted, 1, "Whether to permit hardware (ifnet) TLS sessions"); #ifdef TCP_OFFLOAD static COUNTER_U64_DEFINE_EARLY(ktls_toe_cbc); SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, cbc, CTLFLAG_RD, &ktls_toe_cbc, "Active number of TOE TLS sessions using AES-CBC"); static COUNTER_U64_DEFINE_EARLY(ktls_toe_gcm); SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, gcm, CTLFLAG_RD, &ktls_toe_gcm, "Active number of TOE TLS sessions using AES-GCM"); static COUNTER_U64_DEFINE_EARLY(ktls_toe_chacha20); SYSCTL_COUNTER_U64(_kern_ipc_tls_toe, OID_AUTO, chacha20, CTLFLAG_RD, &ktls_toe_chacha20, "Active number of TOE TLS sessions using Chacha20-Poly1305"); #endif static MALLOC_DEFINE(M_KTLS, "ktls", "Kernel TLS"); static void ktls_cleanup(struct ktls_session *tls); #if defined(INET) || defined(INET6) static void ktls_reset_receive_tag(void *context, int pending); static void ktls_reset_send_tag(void *context, int pending); #endif static void ktls_work_thread(void *ctx); static void ktls_alloc_thread(void *ctx); #if defined(INET) || defined(INET6) static u_int ktls_get_cpu(struct socket *so) { struct inpcb *inp; #ifdef NUMA struct ktls_domain_info *di; #endif u_int cpuid; inp = sotoinpcb(so); #ifdef RSS cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype); if (cpuid != NETISR_CPUID_NONE) return (cpuid); #endif /* * Just use the flowid to shard connections in a repeatable * fashion. Note that TLS 1.0 sessions rely on the * serialization provided by having the same connection use * the same queue. */ #ifdef NUMA if (ktls_bind_threads > 1 && inp->inp_numa_domain != M_NODOM) { di = &ktls_domains[inp->inp_numa_domain]; cpuid = di->cpu[inp->inp_flowid % di->count]; } else #endif cpuid = ktls_cpuid_lookup[inp->inp_flowid % ktls_number_threads]; return (cpuid); } #endif static int ktls_buffer_import(void *arg, void **store, int count, int domain, int flags) { vm_page_t m; int i, req; KASSERT((ktls_maxlen & PAGE_MASK) == 0, ("%s: ktls max length %d is not page size-aligned", __func__, ktls_maxlen)); req = VM_ALLOC_WIRED | VM_ALLOC_NODUMP | malloc2vm_flags(flags); for (i = 0; i < count; i++) { m = vm_page_alloc_noobj_contig_domain(domain, req, atop(ktls_maxlen), 0, ~0ul, PAGE_SIZE, 0, VM_MEMATTR_DEFAULT); if (m == NULL) break; store[i] = (void *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); } return (i); } static void ktls_buffer_release(void *arg __unused, void **store, int count) { vm_page_t m; int i, j; for (i = 0; i < count; i++) { m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)store[i])); for (j = 0; j < atop(ktls_maxlen); j++) { (void)vm_page_unwire_noq(m + j); vm_page_free(m + j); } } } static void ktls_free_mext_contig(struct mbuf *m) { M_ASSERTEXTPG(m); uma_zfree(ktls_buffer_zone, (void *)PHYS_TO_DMAP(m->m_epg_pa[0])); } static int ktls_init(void) { struct thread *td; struct pcpu *pc; int count, domain, error, i; ktls_wq = malloc(sizeof(*ktls_wq) * (mp_maxid + 1), M_KTLS, M_WAITOK | M_ZERO); ktls_session_zone = uma_zcreate("ktls_session", sizeof(struct ktls_session), NULL, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0); if (ktls_sw_buffer_cache) { ktls_buffer_zone = uma_zcache_create("ktls_buffers", roundup2(ktls_maxlen, PAGE_SIZE), NULL, NULL, NULL, NULL, ktls_buffer_import, ktls_buffer_release, NULL, UMA_ZONE_FIRSTTOUCH); } /* * Initialize the workqueues to run the TLS work. We create a * work queue for each CPU. */ CPU_FOREACH(i) { STAILQ_INIT(&ktls_wq[i].m_head); STAILQ_INIT(&ktls_wq[i].so_head); mtx_init(&ktls_wq[i].mtx, "ktls work queue", NULL, MTX_DEF); if (ktls_bind_threads > 1) { pc = pcpu_find(i); domain = pc->pc_domain; count = ktls_domains[domain].count; ktls_domains[domain].cpu[count] = i; ktls_domains[domain].count++; } ktls_cpuid_lookup[ktls_number_threads] = i; ktls_number_threads++; } /* * If we somehow have an empty domain, fall back to choosing * among all KTLS threads. */ if (ktls_bind_threads > 1) { for (i = 0; i < vm_ndomains; i++) { if (ktls_domains[i].count == 0) { ktls_bind_threads = 1; break; } } } /* Start kthreads for each workqueue. */ CPU_FOREACH(i) { error = kproc_kthread_add(ktls_work_thread, &ktls_wq[i], &ktls_proc, &td, 0, 0, "KTLS", "thr_%d", i); if (error) { printf("Can't add KTLS thread %d error %d\n", i, error); return (error); } } /* * Start an allocation thread per-domain to perform blocking allocations * of 16k physically contiguous TLS crypto destination buffers. */ if (ktls_sw_buffer_cache) { for (domain = 0; domain < vm_ndomains; domain++) { if (VM_DOMAIN_EMPTY(domain)) continue; if (CPU_EMPTY(&cpuset_domain[domain])) continue; error = kproc_kthread_add(ktls_alloc_thread, &ktls_domains[domain], &ktls_proc, &ktls_domains[domain].alloc_td.td, 0, 0, "KTLS", "alloc_%d", domain); if (error) { printf("Can't add KTLS alloc thread %d error %d\n", domain, error); return (error); } } } if (bootverbose) printf("KTLS: Initialized %d threads\n", ktls_number_threads); return (0); } static int ktls_start_kthreads(void) { int error, state; start: state = atomic_load_acq_int(&ktls_init_state); if (__predict_true(state > 0)) return (0); if (state < 0) return (ENXIO); sx_xlock(&ktls_init_lock); if (ktls_init_state != 0) { sx_xunlock(&ktls_init_lock); goto start; } error = ktls_init(); if (error == 0) state = 1; else state = -1; atomic_store_rel_int(&ktls_init_state, state); sx_xunlock(&ktls_init_lock); return (error); } #if defined(INET) || defined(INET6) static int ktls_create_session(struct socket *so, struct tls_enable *en, struct ktls_session **tlsp, int direction) { struct ktls_session *tls; int error; /* Only TLS 1.0 - 1.3 are supported. */ if (en->tls_vmajor != TLS_MAJOR_VER_ONE) return (EINVAL); if (en->tls_vminor < TLS_MINOR_VER_ZERO || en->tls_vminor > TLS_MINOR_VER_THREE) return (EINVAL); if (en->auth_key_len < 0 || en->auth_key_len > TLS_MAX_PARAM_SIZE) return (EINVAL); if (en->cipher_key_len < 0 || en->cipher_key_len > TLS_MAX_PARAM_SIZE) return (EINVAL); if (en->iv_len < 0 || en->iv_len > sizeof(tls->params.iv)) return (EINVAL); /* All supported algorithms require a cipher key. */ if (en->cipher_key_len == 0) return (EINVAL); /* No flags are currently supported. */ if (en->flags != 0) return (EINVAL); /* Common checks for supported algorithms. */ switch (en->cipher_algorithm) { case CRYPTO_AES_NIST_GCM_16: /* * auth_algorithm isn't used, but permit GMAC values * for compatibility. */ switch (en->auth_algorithm) { case 0: #ifdef COMPAT_FREEBSD12 /* XXX: Really 13.0-current COMPAT. */ case CRYPTO_AES_128_NIST_GMAC: case CRYPTO_AES_192_NIST_GMAC: case CRYPTO_AES_256_NIST_GMAC: #endif break; default: return (EINVAL); } if (en->auth_key_len != 0) return (EINVAL); switch (en->tls_vminor) { case TLS_MINOR_VER_TWO: if (en->iv_len != TLS_AEAD_GCM_LEN) return (EINVAL); break; case TLS_MINOR_VER_THREE: if (en->iv_len != TLS_1_3_GCM_IV_LEN) return (EINVAL); break; default: return (EINVAL); } break; case CRYPTO_AES_CBC: switch (en->auth_algorithm) { case CRYPTO_SHA1_HMAC: break; case CRYPTO_SHA2_256_HMAC: case CRYPTO_SHA2_384_HMAC: if (en->tls_vminor != TLS_MINOR_VER_TWO) return (EINVAL); break; default: return (EINVAL); } if (en->auth_key_len == 0) return (EINVAL); /* * TLS 1.0 requires an implicit IV. TLS 1.1 and 1.2 * use explicit IVs. */ switch (en->tls_vminor) { case TLS_MINOR_VER_ZERO: if (en->iv_len != TLS_CBC_IMPLICIT_IV_LEN) return (EINVAL); break; case TLS_MINOR_VER_ONE: case TLS_MINOR_VER_TWO: /* Ignore any supplied IV. */ en->iv_len = 0; break; default: return (EINVAL); } break; case CRYPTO_CHACHA20_POLY1305: if (en->auth_algorithm != 0 || en->auth_key_len != 0) return (EINVAL); if (en->tls_vminor != TLS_MINOR_VER_TWO && en->tls_vminor != TLS_MINOR_VER_THREE) return (EINVAL); if (en->iv_len != TLS_CHACHA20_IV_LEN) return (EINVAL); break; default: return (EINVAL); } error = ktls_start_kthreads(); if (error != 0) return (error); tls = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO); counter_u64_add(ktls_offload_active, 1); refcount_init(&tls->refcount, 1); if (direction == KTLS_RX) TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_receive_tag, tls); else TASK_INIT(&tls->reset_tag_task, 0, ktls_reset_send_tag, tls); tls->wq_index = ktls_get_cpu(so); tls->params.cipher_algorithm = en->cipher_algorithm; tls->params.auth_algorithm = en->auth_algorithm; tls->params.tls_vmajor = en->tls_vmajor; tls->params.tls_vminor = en->tls_vminor; tls->params.flags = en->flags; tls->params.max_frame_len = min(TLS_MAX_MSG_SIZE_V10_2, ktls_maxlen); /* Set the header and trailer lengths. */ tls->params.tls_hlen = sizeof(struct tls_record_layer); switch (en->cipher_algorithm) { case CRYPTO_AES_NIST_GCM_16: /* * TLS 1.2 uses a 4 byte implicit IV with an explicit 8 byte * nonce. TLS 1.3 uses a 12 byte implicit IV. */ if (en->tls_vminor < TLS_MINOR_VER_THREE) tls->params.tls_hlen += sizeof(uint64_t); tls->params.tls_tlen = AES_GMAC_HASH_LEN; tls->params.tls_bs = 1; break; case CRYPTO_AES_CBC: switch (en->auth_algorithm) { case CRYPTO_SHA1_HMAC: if (en->tls_vminor == TLS_MINOR_VER_ZERO) { /* Implicit IV, no nonce. */ tls->sequential_records = true; tls->next_seqno = be64dec(en->rec_seq); STAILQ_INIT(&tls->pending_records); } else { tls->params.tls_hlen += AES_BLOCK_LEN; } tls->params.tls_tlen = AES_BLOCK_LEN + SHA1_HASH_LEN; break; case CRYPTO_SHA2_256_HMAC: tls->params.tls_hlen += AES_BLOCK_LEN; tls->params.tls_tlen = AES_BLOCK_LEN + SHA2_256_HASH_LEN; break; case CRYPTO_SHA2_384_HMAC: tls->params.tls_hlen += AES_BLOCK_LEN; tls->params.tls_tlen = AES_BLOCK_LEN + SHA2_384_HASH_LEN; break; default: panic("invalid hmac"); } tls->params.tls_bs = AES_BLOCK_LEN; break; case CRYPTO_CHACHA20_POLY1305: /* * Chacha20 uses a 12 byte implicit IV. */ tls->params.tls_tlen = POLY1305_HASH_LEN; tls->params.tls_bs = 1; break; default: panic("invalid cipher"); } /* * TLS 1.3 includes optional padding which we do not support, * and also puts the "real" record type at the end of the * encrypted data. */ if (en->tls_vminor == TLS_MINOR_VER_THREE) tls->params.tls_tlen += sizeof(uint8_t); KASSERT(tls->params.tls_hlen <= MBUF_PEXT_HDR_LEN, ("TLS header length too long: %d", tls->params.tls_hlen)); KASSERT(tls->params.tls_tlen <= MBUF_PEXT_TRAIL_LEN, ("TLS trailer length too long: %d", tls->params.tls_tlen)); if (en->auth_key_len != 0) { tls->params.auth_key_len = en->auth_key_len; tls->params.auth_key = malloc(en->auth_key_len, M_KTLS, M_WAITOK); error = copyin(en->auth_key, tls->params.auth_key, en->auth_key_len); if (error) goto out; } tls->params.cipher_key_len = en->cipher_key_len; tls->params.cipher_key = malloc(en->cipher_key_len, M_KTLS, M_WAITOK); error = copyin(en->cipher_key, tls->params.cipher_key, en->cipher_key_len); if (error) goto out; /* * This holds the implicit portion of the nonce for AEAD * ciphers and the initial implicit IV for TLS 1.0. The * explicit portions of the IV are generated in ktls_frame(). */ if (en->iv_len != 0) { tls->params.iv_len = en->iv_len; error = copyin(en->iv, tls->params.iv, en->iv_len); if (error) goto out; /* * For TLS 1.2 with GCM, generate an 8-byte nonce as a * counter to generate unique explicit IVs. * * Store this counter in the last 8 bytes of the IV * array so that it is 8-byte aligned. */ if (en->cipher_algorithm == CRYPTO_AES_NIST_GCM_16 && en->tls_vminor == TLS_MINOR_VER_TWO) arc4rand(tls->params.iv + 8, sizeof(uint64_t), 0); } *tlsp = tls; return (0); out: ktls_cleanup(tls); return (error); } static struct ktls_session * ktls_clone_session(struct ktls_session *tls, int direction) { struct ktls_session *tls_new; tls_new = uma_zalloc(ktls_session_zone, M_WAITOK | M_ZERO); counter_u64_add(ktls_offload_active, 1); refcount_init(&tls_new->refcount, 1); if (direction == KTLS_RX) TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_receive_tag, tls_new); else TASK_INIT(&tls_new->reset_tag_task, 0, ktls_reset_send_tag, tls_new); /* Copy fields from existing session. */ tls_new->params = tls->params; tls_new->wq_index = tls->wq_index; /* Deep copy keys. */ if (tls_new->params.auth_key != NULL) { tls_new->params.auth_key = malloc(tls->params.auth_key_len, M_KTLS, M_WAITOK); memcpy(tls_new->params.auth_key, tls->params.auth_key, tls->params.auth_key_len); } tls_new->params.cipher_key = malloc(tls->params.cipher_key_len, M_KTLS, M_WAITOK); memcpy(tls_new->params.cipher_key, tls->params.cipher_key, tls->params.cipher_key_len); return (tls_new); } #endif static void ktls_cleanup(struct ktls_session *tls) { counter_u64_add(ktls_offload_active, -1); switch (tls->mode) { case TCP_TLS_MODE_SW: switch (tls->params.cipher_algorithm) { case CRYPTO_AES_CBC: counter_u64_add(ktls_sw_cbc, -1); break; case CRYPTO_AES_NIST_GCM_16: counter_u64_add(ktls_sw_gcm, -1); break; case CRYPTO_CHACHA20_POLY1305: counter_u64_add(ktls_sw_chacha20, -1); break; } break; case TCP_TLS_MODE_IFNET: switch (tls->params.cipher_algorithm) { case CRYPTO_AES_CBC: counter_u64_add(ktls_ifnet_cbc, -1); break; case CRYPTO_AES_NIST_GCM_16: counter_u64_add(ktls_ifnet_gcm, -1); break; case CRYPTO_CHACHA20_POLY1305: counter_u64_add(ktls_ifnet_chacha20, -1); break; } if (tls->snd_tag != NULL) m_snd_tag_rele(tls->snd_tag); if (tls->rx_ifp != NULL) if_rele(tls->rx_ifp); break; #ifdef TCP_OFFLOAD case TCP_TLS_MODE_TOE: switch (tls->params.cipher_algorithm) { case CRYPTO_AES_CBC: counter_u64_add(ktls_toe_cbc, -1); break; case CRYPTO_AES_NIST_GCM_16: counter_u64_add(ktls_toe_gcm, -1); break; case CRYPTO_CHACHA20_POLY1305: counter_u64_add(ktls_toe_chacha20, -1); break; } break; #endif } if (tls->ocf_session != NULL) ktls_ocf_free(tls); if (tls->params.auth_key != NULL) { zfree(tls->params.auth_key, M_KTLS); tls->params.auth_key = NULL; tls->params.auth_key_len = 0; } if (tls->params.cipher_key != NULL) { zfree(tls->params.cipher_key, M_KTLS); tls->params.cipher_key = NULL; tls->params.cipher_key_len = 0; } explicit_bzero(tls->params.iv, sizeof(tls->params.iv)); } #if defined(INET) || defined(INET6) #ifdef TCP_OFFLOAD static int ktls_try_toe(struct socket *so, struct ktls_session *tls, int direction) { struct inpcb *inp; struct tcpcb *tp; int error; inp = so->so_pcb; INP_WLOCK(inp); - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { INP_WUNLOCK(inp); return (ECONNRESET); } if (inp->inp_socket == NULL) { INP_WUNLOCK(inp); return (ECONNRESET); } tp = intotcpcb(inp); if (!(tp->t_flags & TF_TOE)) { INP_WUNLOCK(inp); return (EOPNOTSUPP); } error = tcp_offload_alloc_tls_session(tp, tls, direction); INP_WUNLOCK(inp); if (error == 0) { tls->mode = TCP_TLS_MODE_TOE; switch (tls->params.cipher_algorithm) { case CRYPTO_AES_CBC: counter_u64_add(ktls_toe_cbc, 1); break; case CRYPTO_AES_NIST_GCM_16: counter_u64_add(ktls_toe_gcm, 1); break; case CRYPTO_CHACHA20_POLY1305: counter_u64_add(ktls_toe_chacha20, 1); break; } } return (error); } #endif /* * Common code used when first enabling ifnet TLS on a connection or * when allocating a new ifnet TLS session due to a routing change. * This function allocates a new TLS send tag on whatever interface * the connection is currently routed over. */ static int ktls_alloc_snd_tag(struct inpcb *inp, struct ktls_session *tls, bool force, struct m_snd_tag **mstp) { union if_snd_tag_alloc_params params; struct ifnet *ifp; struct nhop_object *nh; struct tcpcb *tp; int error; INP_RLOCK(inp); - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { INP_RUNLOCK(inp); return (ECONNRESET); } if (inp->inp_socket == NULL) { INP_RUNLOCK(inp); return (ECONNRESET); } tp = intotcpcb(inp); /* * Check administrative controls on ifnet TLS to determine if * ifnet TLS should be denied. * * - Always permit 'force' requests. * - ktls_ifnet_permitted == 0: always deny. */ if (!force && ktls_ifnet_permitted == 0) { INP_RUNLOCK(inp); return (ENXIO); } /* * XXX: Use the cached route in the inpcb to find the * interface. This should perhaps instead use * rtalloc1_fib(dst, 0, 0, fibnum). Since KTLS is only * enabled after a connection has completed key negotiation in * userland, the cached route will be present in practice. */ nh = inp->inp_route.ro_nh; if (nh == NULL) { INP_RUNLOCK(inp); return (ENXIO); } ifp = nh->nh_ifp; if_ref(ifp); /* * Allocate a TLS + ratelimit tag if the connection has an * existing pacing rate. */ if (tp->t_pacing_rate != -1 && (ifp->if_capenable & IFCAP_TXTLS_RTLMT) != 0) { params.hdr.type = IF_SND_TAG_TYPE_TLS_RATE_LIMIT; params.tls_rate_limit.inp = inp; params.tls_rate_limit.tls = tls; params.tls_rate_limit.max_rate = tp->t_pacing_rate; } else { params.hdr.type = IF_SND_TAG_TYPE_TLS; params.tls.inp = inp; params.tls.tls = tls; } params.hdr.flowid = inp->inp_flowid; params.hdr.flowtype = inp->inp_flowtype; params.hdr.numa_domain = inp->inp_numa_domain; INP_RUNLOCK(inp); if ((ifp->if_capenable & IFCAP_MEXTPG) == 0) { error = EOPNOTSUPP; goto out; } if (inp->inp_vflag & INP_IPV6) { if ((ifp->if_capenable & IFCAP_TXTLS6) == 0) { error = EOPNOTSUPP; goto out; } } else { if ((ifp->if_capenable & IFCAP_TXTLS4) == 0) { error = EOPNOTSUPP; goto out; } } error = m_snd_tag_alloc(ifp, ¶ms, mstp); out: if_rele(ifp); return (error); } /* * Allocate an initial TLS receive tag for doing HW decryption of TLS * data. * * This function allocates a new TLS receive tag on whatever interface * the connection is currently routed over. If the connection ends up * using a different interface for receive this will get fixed up via * ktls_input_ifp_mismatch as future packets arrive. */ static int ktls_alloc_rcv_tag(struct inpcb *inp, struct ktls_session *tls, struct m_snd_tag **mstp) { union if_snd_tag_alloc_params params; struct ifnet *ifp; struct nhop_object *nh; int error; if (!ktls_ocf_recrypt_supported(tls)) return (ENXIO); INP_RLOCK(inp); - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { INP_RUNLOCK(inp); return (ECONNRESET); } if (inp->inp_socket == NULL) { INP_RUNLOCK(inp); return (ECONNRESET); } /* * Check administrative controls on ifnet TLS to determine if * ifnet TLS should be denied. */ if (ktls_ifnet_permitted == 0) { INP_RUNLOCK(inp); return (ENXIO); } /* * XXX: As with ktls_alloc_snd_tag, use the cached route in * the inpcb to find the interface. */ nh = inp->inp_route.ro_nh; if (nh == NULL) { INP_RUNLOCK(inp); return (ENXIO); } ifp = nh->nh_ifp; if_ref(ifp); tls->rx_ifp = ifp; params.hdr.type = IF_SND_TAG_TYPE_TLS_RX; params.hdr.flowid = inp->inp_flowid; params.hdr.flowtype = inp->inp_flowtype; params.hdr.numa_domain = inp->inp_numa_domain; params.tls_rx.inp = inp; params.tls_rx.tls = tls; params.tls_rx.vlan_id = 0; INP_RUNLOCK(inp); if (inp->inp_vflag & INP_IPV6) { if ((ifp->if_capenable2 & IFCAP2_RXTLS6) == 0) { error = EOPNOTSUPP; goto out; } } else { if ((ifp->if_capenable2 & IFCAP2_RXTLS4) == 0) { error = EOPNOTSUPP; goto out; } } error = m_snd_tag_alloc(ifp, ¶ms, mstp); /* * If this connection is over a vlan, vlan_snd_tag_alloc * rewrites vlan_id with the saved interface. Save the VLAN * ID for use in ktls_reset_receive_tag which allocates new * receive tags directly from the leaf interface bypassing * if_vlan. */ if (error == 0) tls->rx_vlan_id = params.tls_rx.vlan_id; out: return (error); } static int ktls_try_ifnet(struct socket *so, struct ktls_session *tls, int direction, bool force) { struct m_snd_tag *mst; int error; switch (direction) { case KTLS_TX: error = ktls_alloc_snd_tag(so->so_pcb, tls, force, &mst); if (__predict_false(error != 0)) goto done; break; case KTLS_RX: KASSERT(!force, ("%s: forced receive tag", __func__)); error = ktls_alloc_rcv_tag(so->so_pcb, tls, &mst); if (__predict_false(error != 0)) goto done; break; default: __assert_unreachable(); } tls->mode = TCP_TLS_MODE_IFNET; tls->snd_tag = mst; switch (tls->params.cipher_algorithm) { case CRYPTO_AES_CBC: counter_u64_add(ktls_ifnet_cbc, 1); break; case CRYPTO_AES_NIST_GCM_16: counter_u64_add(ktls_ifnet_gcm, 1); break; case CRYPTO_CHACHA20_POLY1305: counter_u64_add(ktls_ifnet_chacha20, 1); break; default: break; } done: return (error); } static void ktls_use_sw(struct ktls_session *tls) { tls->mode = TCP_TLS_MODE_SW; switch (tls->params.cipher_algorithm) { case CRYPTO_AES_CBC: counter_u64_add(ktls_sw_cbc, 1); break; case CRYPTO_AES_NIST_GCM_16: counter_u64_add(ktls_sw_gcm, 1); break; case CRYPTO_CHACHA20_POLY1305: counter_u64_add(ktls_sw_chacha20, 1); break; } } static int ktls_try_sw(struct socket *so, struct ktls_session *tls, int direction) { int error; error = ktls_ocf_try(so, tls, direction); if (error) return (error); ktls_use_sw(tls); return (0); } /* * KTLS RX stores data in the socket buffer as a list of TLS records, * where each record is stored as a control message containg the TLS * header followed by data mbufs containing the decrypted data. This * is different from KTLS TX which always uses an mb_ext_pgs mbuf for * both encrypted and decrypted data. TLS records decrypted by a NIC * should be queued to the socket buffer as records, but encrypted * data which needs to be decrypted by software arrives as a stream of * regular mbufs which need to be converted. In addition, there may * already be pending encrypted data in the socket buffer when KTLS RX * is enabled. * * To manage not-yet-decrypted data for KTLS RX, the following scheme * is used: * * - A single chain of NOTREADY mbufs is hung off of sb_mtls. * * - ktls_check_rx checks this chain of mbufs reading the TLS header * from the first mbuf. Once all of the data for that TLS record is * queued, the socket is queued to a worker thread. * * - The worker thread calls ktls_decrypt to decrypt TLS records in * the TLS chain. Each TLS record is detached from the TLS chain, * decrypted, and inserted into the regular socket buffer chain as * record starting with a control message holding the TLS header and * a chain of mbufs holding the encrypted data. */ static void sb_mark_notready(struct sockbuf *sb) { struct mbuf *m; m = sb->sb_mb; sb->sb_mtls = m; sb->sb_mb = NULL; sb->sb_mbtail = NULL; sb->sb_lastrecord = NULL; for (; m != NULL; m = m->m_next) { KASSERT(m->m_nextpkt == NULL, ("%s: m_nextpkt != NULL", __func__)); KASSERT((m->m_flags & M_NOTAVAIL) == 0, ("%s: mbuf not avail", __func__)); KASSERT(sb->sb_acc >= m->m_len, ("%s: sb_acc < m->m_len", __func__)); m->m_flags |= M_NOTREADY; sb->sb_acc -= m->m_len; sb->sb_tlscc += m->m_len; sb->sb_mtlstail = m; } KASSERT(sb->sb_acc == 0 && sb->sb_tlscc == sb->sb_ccc, ("%s: acc %u tlscc %u ccc %u", __func__, sb->sb_acc, sb->sb_tlscc, sb->sb_ccc)); } /* * Return information about the pending TLS data in a socket * buffer. On return, 'seqno' is set to the sequence number * of the next TLS record to be received, 'resid' is set to * the amount of bytes still needed for the last pending * record. The function returns 'false' if the last pending * record contains a partial TLS header. In that case, 'resid' * is the number of bytes needed to complete the TLS header. */ bool ktls_pending_rx_info(struct sockbuf *sb, uint64_t *seqnop, size_t *residp) { struct tls_record_layer hdr; struct mbuf *m; uint64_t seqno; size_t resid; u_int offset, record_len; SOCKBUF_LOCK_ASSERT(sb); MPASS(sb->sb_flags & SB_TLS_RX); seqno = sb->sb_tls_seqno; resid = sb->sb_tlscc; m = sb->sb_mtls; offset = 0; if (resid == 0) { *seqnop = seqno; *residp = 0; return (true); } for (;;) { seqno++; if (resid < sizeof(hdr)) { *seqnop = seqno; *residp = sizeof(hdr) - resid; return (false); } m_copydata(m, offset, sizeof(hdr), (void *)&hdr); record_len = sizeof(hdr) + ntohs(hdr.tls_length); if (resid <= record_len) { *seqnop = seqno; *residp = record_len - resid; return (true); } resid -= record_len; while (record_len != 0) { if (m->m_len - offset > record_len) { offset += record_len; break; } record_len -= (m->m_len - offset); offset = 0; m = m->m_next; } } } int ktls_enable_rx(struct socket *so, struct tls_enable *en) { struct ktls_session *tls; int error; if (!ktls_offload_enable) return (ENOTSUP); if (SOLISTENING(so)) return (EINVAL); counter_u64_add(ktls_offload_enable_calls, 1); /* * This should always be true since only the TCP socket option * invokes this function. */ if (so->so_proto->pr_protocol != IPPROTO_TCP) return (EINVAL); /* * XXX: Don't overwrite existing sessions. We should permit * this to support rekeying in the future. */ if (so->so_rcv.sb_tls_info != NULL) return (EALREADY); if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable) return (ENOTSUP); error = ktls_create_session(so, en, &tls, KTLS_RX); if (error) return (error); error = ktls_ocf_try(so, tls, KTLS_RX); if (error) { ktls_cleanup(tls); return (error); } /* Mark the socket as using TLS offload. */ SOCKBUF_LOCK(&so->so_rcv); so->so_rcv.sb_tls_seqno = be64dec(en->rec_seq); so->so_rcv.sb_tls_info = tls; so->so_rcv.sb_flags |= SB_TLS_RX; /* Mark existing data as not ready until it can be decrypted. */ sb_mark_notready(&so->so_rcv); ktls_check_rx(&so->so_rcv); SOCKBUF_UNLOCK(&so->so_rcv); /* Prefer TOE -> ifnet TLS -> software TLS. */ #ifdef TCP_OFFLOAD error = ktls_try_toe(so, tls, KTLS_RX); if (error) #endif error = ktls_try_ifnet(so, tls, KTLS_RX, false); if (error) ktls_use_sw(tls); counter_u64_add(ktls_offload_total, 1); return (0); } int ktls_enable_tx(struct socket *so, struct tls_enable *en) { struct ktls_session *tls; struct inpcb *inp; int error; if (!ktls_offload_enable) return (ENOTSUP); if (SOLISTENING(so)) return (EINVAL); counter_u64_add(ktls_offload_enable_calls, 1); /* * This should always be true since only the TCP socket option * invokes this function. */ if (so->so_proto->pr_protocol != IPPROTO_TCP) return (EINVAL); /* * XXX: Don't overwrite existing sessions. We should permit * this to support rekeying in the future. */ if (so->so_snd.sb_tls_info != NULL) return (EALREADY); if (en->cipher_algorithm == CRYPTO_AES_CBC && !ktls_cbc_enable) return (ENOTSUP); /* TLS requires ext pgs */ if (mb_use_ext_pgs == 0) return (ENXIO); error = ktls_create_session(so, en, &tls, KTLS_TX); if (error) return (error); /* Prefer TOE -> ifnet TLS -> software TLS. */ #ifdef TCP_OFFLOAD error = ktls_try_toe(so, tls, KTLS_TX); if (error) #endif error = ktls_try_ifnet(so, tls, KTLS_TX, false); if (error) error = ktls_try_sw(so, tls, KTLS_TX); if (error) { ktls_cleanup(tls); return (error); } error = SOCK_IO_SEND_LOCK(so, SBL_WAIT); if (error) { ktls_cleanup(tls); return (error); } /* * Write lock the INP when setting sb_tls_info so that * routines in tcp_ratelimit.c can read sb_tls_info while * holding the INP lock. */ inp = so->so_pcb; INP_WLOCK(inp); SOCKBUF_LOCK(&so->so_snd); so->so_snd.sb_tls_seqno = be64dec(en->rec_seq); so->so_snd.sb_tls_info = tls; if (tls->mode != TCP_TLS_MODE_SW) so->so_snd.sb_flags |= SB_TLS_IFNET; SOCKBUF_UNLOCK(&so->so_snd); INP_WUNLOCK(inp); SOCK_IO_SEND_UNLOCK(so); counter_u64_add(ktls_offload_total, 1); return (0); } int ktls_get_rx_mode(struct socket *so, int *modep) { struct ktls_session *tls; struct inpcb *inp __diagused; if (SOLISTENING(so)) return (EINVAL); inp = so->so_pcb; INP_WLOCK_ASSERT(inp); SOCK_RECVBUF_LOCK(so); tls = so->so_rcv.sb_tls_info; if (tls == NULL) *modep = TCP_TLS_MODE_NONE; else *modep = tls->mode; SOCK_RECVBUF_UNLOCK(so); return (0); } /* * ktls_get_rx_sequence - get the next TCP- and TLS- sequence number. * * This function gets information about the next TCP- and TLS- * sequence number to be processed by the TLS receive worker * thread. The information is extracted from the given "inpcb" * structure. The values are stored in host endian format at the two * given output pointer locations. The TCP sequence number points to * the beginning of the TLS header. * * This function returns zero on success, else a non-zero error code * is returned. */ int ktls_get_rx_sequence(struct inpcb *inp, uint32_t *tcpseq, uint64_t *tlsseq) { struct socket *so; struct tcpcb *tp; INP_RLOCK(inp); so = inp->inp_socket; if (__predict_false(so == NULL)) { INP_RUNLOCK(inp); return (EINVAL); } - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { INP_RUNLOCK(inp); return (ECONNRESET); } tp = intotcpcb(inp); MPASS(tp != NULL); SOCKBUF_LOCK(&so->so_rcv); *tcpseq = tp->rcv_nxt - so->so_rcv.sb_tlscc; *tlsseq = so->so_rcv.sb_tls_seqno; SOCKBUF_UNLOCK(&so->so_rcv); INP_RUNLOCK(inp); return (0); } int ktls_get_tx_mode(struct socket *so, int *modep) { struct ktls_session *tls; struct inpcb *inp __diagused; if (SOLISTENING(so)) return (EINVAL); inp = so->so_pcb; INP_WLOCK_ASSERT(inp); SOCK_SENDBUF_LOCK(so); tls = so->so_snd.sb_tls_info; if (tls == NULL) *modep = TCP_TLS_MODE_NONE; else *modep = tls->mode; SOCK_SENDBUF_UNLOCK(so); return (0); } /* * Switch between SW and ifnet TLS sessions as requested. */ int ktls_set_tx_mode(struct socket *so, int mode) { struct ktls_session *tls, *tls_new; struct inpcb *inp; int error; if (SOLISTENING(so)) return (EINVAL); switch (mode) { case TCP_TLS_MODE_SW: case TCP_TLS_MODE_IFNET: break; default: return (EINVAL); } inp = so->so_pcb; INP_WLOCK_ASSERT(inp); SOCKBUF_LOCK(&so->so_snd); tls = so->so_snd.sb_tls_info; if (tls == NULL) { SOCKBUF_UNLOCK(&so->so_snd); return (0); } if (tls->mode == mode) { SOCKBUF_UNLOCK(&so->so_snd); return (0); } tls = ktls_hold(tls); SOCKBUF_UNLOCK(&so->so_snd); INP_WUNLOCK(inp); tls_new = ktls_clone_session(tls, KTLS_TX); if (mode == TCP_TLS_MODE_IFNET) error = ktls_try_ifnet(so, tls_new, KTLS_TX, true); else error = ktls_try_sw(so, tls_new, KTLS_TX); if (error) { counter_u64_add(ktls_switch_failed, 1); ktls_free(tls_new); ktls_free(tls); INP_WLOCK(inp); return (error); } error = SOCK_IO_SEND_LOCK(so, SBL_WAIT); if (error) { counter_u64_add(ktls_switch_failed, 1); ktls_free(tls_new); ktls_free(tls); INP_WLOCK(inp); return (error); } /* * If we raced with another session change, keep the existing * session. */ if (tls != so->so_snd.sb_tls_info) { counter_u64_add(ktls_switch_failed, 1); SOCK_IO_SEND_UNLOCK(so); ktls_free(tls_new); ktls_free(tls); INP_WLOCK(inp); return (EBUSY); } INP_WLOCK(inp); SOCKBUF_LOCK(&so->so_snd); so->so_snd.sb_tls_info = tls_new; if (tls_new->mode != TCP_TLS_MODE_SW) so->so_snd.sb_flags |= SB_TLS_IFNET; SOCKBUF_UNLOCK(&so->so_snd); SOCK_IO_SEND_UNLOCK(so); /* * Drop two references on 'tls'. The first is for the * ktls_hold() above. The second drops the reference from the * socket buffer. */ KASSERT(tls->refcount >= 2, ("too few references on old session")); ktls_free(tls); ktls_free(tls); if (mode == TCP_TLS_MODE_IFNET) counter_u64_add(ktls_switch_to_ifnet, 1); else counter_u64_add(ktls_switch_to_sw, 1); return (0); } /* * Try to allocate a new TLS receive tag. This task is scheduled when * sbappend_ktls_rx detects an input path change. If a new tag is * allocated, replace the tag in the TLS session. If a new tag cannot * be allocated, let the session fall back to software decryption. */ static void ktls_reset_receive_tag(void *context, int pending) { union if_snd_tag_alloc_params params; struct ktls_session *tls; struct m_snd_tag *mst; struct inpcb *inp; struct ifnet *ifp; struct socket *so; int error; MPASS(pending == 1); tls = context; so = tls->so; inp = so->so_pcb; ifp = NULL; INP_RLOCK(inp); - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { INP_RUNLOCK(inp); goto out; } SOCKBUF_LOCK(&so->so_rcv); mst = tls->snd_tag; tls->snd_tag = NULL; if (mst != NULL) m_snd_tag_rele(mst); ifp = tls->rx_ifp; if_ref(ifp); SOCKBUF_UNLOCK(&so->so_rcv); params.hdr.type = IF_SND_TAG_TYPE_TLS_RX; params.hdr.flowid = inp->inp_flowid; params.hdr.flowtype = inp->inp_flowtype; params.hdr.numa_domain = inp->inp_numa_domain; params.tls_rx.inp = inp; params.tls_rx.tls = tls; params.tls_rx.vlan_id = tls->rx_vlan_id; INP_RUNLOCK(inp); if (inp->inp_vflag & INP_IPV6) { if ((ifp->if_capenable2 & IFCAP2_RXTLS6) == 0) goto out; } else { if ((ifp->if_capenable2 & IFCAP2_RXTLS4) == 0) goto out; } error = m_snd_tag_alloc(ifp, ¶ms, &mst); if (error == 0) { SOCKBUF_LOCK(&so->so_rcv); tls->snd_tag = mst; SOCKBUF_UNLOCK(&so->so_rcv); counter_u64_add(ktls_ifnet_reset, 1); } else { /* * Just fall back to software decryption if a tag * cannot be allocated leaving the connection intact. * If a future input path change switches to another * interface this connection will resume ifnet TLS. */ counter_u64_add(ktls_ifnet_reset_failed, 1); } out: mtx_pool_lock(mtxpool_sleep, tls); tls->reset_pending = false; mtx_pool_unlock(mtxpool_sleep, tls); if (ifp != NULL) if_rele(ifp); sorele(so); ktls_free(tls); } /* * Try to allocate a new TLS send tag. This task is scheduled when * ip_output detects a route change while trying to transmit a packet * holding a TLS record. If a new tag is allocated, replace the tag * in the TLS session. Subsequent packets on the connection will use * the new tag. If a new tag cannot be allocated, drop the * connection. */ static void ktls_reset_send_tag(void *context, int pending) { struct epoch_tracker et; struct ktls_session *tls; struct m_snd_tag *old, *new; struct inpcb *inp; struct tcpcb *tp; int error; MPASS(pending == 1); tls = context; inp = tls->inp; /* * Free the old tag first before allocating a new one. * ip[6]_output_send() will treat a NULL send tag the same as * an ifp mismatch and drop packets until a new tag is * allocated. * * Write-lock the INP when changing tls->snd_tag since * ip[6]_output_send() holds a read-lock when reading the * pointer. */ INP_WLOCK(inp); old = tls->snd_tag; tls->snd_tag = NULL; INP_WUNLOCK(inp); if (old != NULL) m_snd_tag_rele(old); error = ktls_alloc_snd_tag(inp, tls, true, &new); if (error == 0) { INP_WLOCK(inp); tls->snd_tag = new; mtx_pool_lock(mtxpool_sleep, tls); tls->reset_pending = false; mtx_pool_unlock(mtxpool_sleep, tls); if (!in_pcbrele_wlocked(inp)) INP_WUNLOCK(inp); counter_u64_add(ktls_ifnet_reset, 1); /* * XXX: Should we kick tcp_output explicitly now that * the send tag is fixed or just rely on timers? */ } else { NET_EPOCH_ENTER(et); INP_WLOCK(inp); if (!in_pcbrele_wlocked(inp)) { - if (!(inp->inp_flags & INP_TIMEWAIT) && - !(inp->inp_flags & INP_DROPPED)) { + if (!(inp->inp_flags & INP_DROPPED)) { tp = intotcpcb(inp); CURVNET_SET(tp->t_vnet); tp = tcp_drop(tp, ECONNABORTED); CURVNET_RESTORE(); if (tp != NULL) INP_WUNLOCK(inp); counter_u64_add(ktls_ifnet_reset_dropped, 1); } else INP_WUNLOCK(inp); } NET_EPOCH_EXIT(et); counter_u64_add(ktls_ifnet_reset_failed, 1); /* * Leave reset_pending true to avoid future tasks while * the socket goes away. */ } ktls_free(tls); } void ktls_input_ifp_mismatch(struct sockbuf *sb, struct ifnet *ifp) { struct ktls_session *tls; struct socket *so; SOCKBUF_LOCK_ASSERT(sb); KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX", __func__, sb)); so = __containerof(sb, struct socket, so_rcv); tls = sb->sb_tls_info; if_rele(tls->rx_ifp); if_ref(ifp); tls->rx_ifp = ifp; /* * See if we should schedule a task to update the receive tag for * this session. */ mtx_pool_lock(mtxpool_sleep, tls); if (!tls->reset_pending) { (void) ktls_hold(tls); soref(so); tls->so = so; tls->reset_pending = true; taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task); } mtx_pool_unlock(mtxpool_sleep, tls); } int ktls_output_eagain(struct inpcb *inp, struct ktls_session *tls) { if (inp == NULL) return (ENOBUFS); INP_LOCK_ASSERT(inp); /* * See if we should schedule a task to update the send tag for * this session. */ mtx_pool_lock(mtxpool_sleep, tls); if (!tls->reset_pending) { (void) ktls_hold(tls); in_pcbref(inp); tls->inp = inp; tls->reset_pending = true; taskqueue_enqueue(taskqueue_thread, &tls->reset_tag_task); } mtx_pool_unlock(mtxpool_sleep, tls); return (ENOBUFS); } #ifdef RATELIMIT int ktls_modify_txrtlmt(struct ktls_session *tls, uint64_t max_pacing_rate) { union if_snd_tag_modify_params params = { .rate_limit.max_rate = max_pacing_rate, .rate_limit.flags = M_NOWAIT, }; struct m_snd_tag *mst; /* Can't get to the inp, but it should be locked. */ /* INP_LOCK_ASSERT(inp); */ MPASS(tls->mode == TCP_TLS_MODE_IFNET); if (tls->snd_tag == NULL) { /* * Resetting send tag, ignore this change. The * pending reset may or may not see this updated rate * in the tcpcb. If it doesn't, we will just lose * this rate change. */ return (0); } mst = tls->snd_tag; MPASS(mst != NULL); MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RATE_LIMIT); return (mst->sw->snd_tag_modify(mst, ¶ms)); } #endif #endif void ktls_destroy(struct ktls_session *tls) { if (tls->sequential_records) { struct mbuf *m, *n; int page_count; STAILQ_FOREACH_SAFE(m, &tls->pending_records, m_epg_stailq, n) { page_count = m->m_epg_enc_cnt; while (page_count > 0) { KASSERT(page_count >= m->m_epg_nrdy, ("%s: too few pages", __func__)); page_count -= m->m_epg_nrdy; m = m_free(m); } } } ktls_cleanup(tls); uma_zfree(ktls_session_zone, tls); } void ktls_seq(struct sockbuf *sb, struct mbuf *m) { for (; m != NULL; m = m->m_next) { KASSERT((m->m_flags & M_EXTPG) != 0, ("ktls_seq: mapped mbuf %p", m)); m->m_epg_seqno = sb->sb_tls_seqno; sb->sb_tls_seqno++; } } /* * Add TLS framing (headers and trailers) to a chain of mbufs. Each * mbuf in the chain must be an unmapped mbuf. The payload of the * mbuf must be populated with the payload of each TLS record. * * The record_type argument specifies the TLS record type used when * populating the TLS header. * * The enq_count argument on return is set to the number of pages of * payload data for this entire chain that need to be encrypted via SW * encryption. The returned value should be passed to ktls_enqueue * when scheduling encryption of this chain of mbufs. To handle the * special case of empty fragments for TLS 1.0 sessions, an empty * fragment counts as one page. */ void ktls_frame(struct mbuf *top, struct ktls_session *tls, int *enq_cnt, uint8_t record_type) { struct tls_record_layer *tlshdr; struct mbuf *m; uint64_t *noncep; uint16_t tls_len; int maxlen __diagused; maxlen = tls->params.max_frame_len; *enq_cnt = 0; for (m = top; m != NULL; m = m->m_next) { /* * All mbufs in the chain should be TLS records whose * payload does not exceed the maximum frame length. * * Empty TLS 1.0 records are permitted when using CBC. */ KASSERT(m->m_len <= maxlen && m->m_len >= 0 && (m->m_len > 0 || ktls_permit_empty_frames(tls)), ("ktls_frame: m %p len %d", m, m->m_len)); /* * TLS frames require unmapped mbufs to store session * info. */ KASSERT((m->m_flags & M_EXTPG) != 0, ("ktls_frame: mapped mbuf %p (top = %p)", m, top)); tls_len = m->m_len; /* Save a reference to the session. */ m->m_epg_tls = ktls_hold(tls); m->m_epg_hdrlen = tls->params.tls_hlen; m->m_epg_trllen = tls->params.tls_tlen; if (tls->params.cipher_algorithm == CRYPTO_AES_CBC) { int bs, delta; /* * AES-CBC pads messages to a multiple of the * block size. Note that the padding is * applied after the digest and the encryption * is done on the "plaintext || mac || padding". * At least one byte of padding is always * present. * * Compute the final trailer length assuming * at most one block of padding. * tls->params.tls_tlen is the maximum * possible trailer length (padding + digest). * delta holds the number of excess padding * bytes if the maximum were used. Those * extra bytes are removed. */ bs = tls->params.tls_bs; delta = (tls_len + tls->params.tls_tlen) & (bs - 1); m->m_epg_trllen -= delta; } m->m_len += m->m_epg_hdrlen + m->m_epg_trllen; /* Populate the TLS header. */ tlshdr = (void *)m->m_epg_hdr; tlshdr->tls_vmajor = tls->params.tls_vmajor; /* * TLS 1.3 masquarades as TLS 1.2 with a record type * of TLS_RLTYPE_APP. */ if (tls->params.tls_vminor == TLS_MINOR_VER_THREE && tls->params.tls_vmajor == TLS_MAJOR_VER_ONE) { tlshdr->tls_vminor = TLS_MINOR_VER_TWO; tlshdr->tls_type = TLS_RLTYPE_APP; /* save the real record type for later */ m->m_epg_record_type = record_type; m->m_epg_trail[0] = record_type; } else { tlshdr->tls_vminor = tls->params.tls_vminor; tlshdr->tls_type = record_type; } tlshdr->tls_length = htons(m->m_len - sizeof(*tlshdr)); /* * Store nonces / explicit IVs after the end of the * TLS header. * * For GCM with TLS 1.2, an 8 byte nonce is copied * from the end of the IV. The nonce is then * incremented for use by the next record. * * For CBC, a random nonce is inserted for TLS 1.1+. */ if (tls->params.cipher_algorithm == CRYPTO_AES_NIST_GCM_16 && tls->params.tls_vminor == TLS_MINOR_VER_TWO) { noncep = (uint64_t *)(tls->params.iv + 8); be64enc(tlshdr + 1, *noncep); (*noncep)++; } else if (tls->params.cipher_algorithm == CRYPTO_AES_CBC && tls->params.tls_vminor >= TLS_MINOR_VER_ONE) arc4rand(tlshdr + 1, AES_BLOCK_LEN, 0); /* * When using SW encryption, mark the mbuf not ready. * It will be marked ready via sbready() after the * record has been encrypted. * * When using ifnet TLS, unencrypted TLS records are * sent down the stack to the NIC. */ if (tls->mode == TCP_TLS_MODE_SW) { m->m_flags |= M_NOTREADY; if (__predict_false(tls_len == 0)) { /* TLS 1.0 empty fragment. */ m->m_epg_nrdy = 1; } else m->m_epg_nrdy = m->m_epg_npgs; *enq_cnt += m->m_epg_nrdy; } } } bool ktls_permit_empty_frames(struct ktls_session *tls) { return (tls->params.cipher_algorithm == CRYPTO_AES_CBC && tls->params.tls_vminor == TLS_MINOR_VER_ZERO); } void ktls_check_rx(struct sockbuf *sb) { struct tls_record_layer hdr; struct ktls_wq *wq; struct socket *so; bool running; SOCKBUF_LOCK_ASSERT(sb); KASSERT(sb->sb_flags & SB_TLS_RX, ("%s: sockbuf %p isn't TLS RX", __func__, sb)); so = __containerof(sb, struct socket, so_rcv); if (sb->sb_flags & SB_TLS_RX_RUNNING) return; /* Is there enough queued for a TLS header? */ if (sb->sb_tlscc < sizeof(hdr)) { if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc != 0) so->so_error = EMSGSIZE; return; } m_copydata(sb->sb_mtls, 0, sizeof(hdr), (void *)&hdr); /* Is the entire record queued? */ if (sb->sb_tlscc < sizeof(hdr) + ntohs(hdr.tls_length)) { if ((sb->sb_state & SBS_CANTRCVMORE) != 0) so->so_error = EMSGSIZE; return; } sb->sb_flags |= SB_TLS_RX_RUNNING; soref(so); wq = &ktls_wq[so->so_rcv.sb_tls_info->wq_index]; mtx_lock(&wq->mtx); STAILQ_INSERT_TAIL(&wq->so_head, so, so_ktls_rx_list); running = wq->running; mtx_unlock(&wq->mtx); if (!running) wakeup(wq); counter_u64_add(ktls_cnt_rx_queued, 1); } static struct mbuf * ktls_detach_record(struct sockbuf *sb, int len) { struct mbuf *m, *n, *top; int remain; SOCKBUF_LOCK_ASSERT(sb); MPASS(len <= sb->sb_tlscc); /* * If TLS chain is the exact size of the record, * just grab the whole record. */ top = sb->sb_mtls; if (sb->sb_tlscc == len) { sb->sb_mtls = NULL; sb->sb_mtlstail = NULL; goto out; } /* * While it would be nice to use m_split() here, we need * to know exactly what m_split() allocates to update the * accounting, so do it inline instead. */ remain = len; for (m = top; remain > m->m_len; m = m->m_next) remain -= m->m_len; /* Easy case: don't have to split 'm'. */ if (remain == m->m_len) { sb->sb_mtls = m->m_next; if (sb->sb_mtls == NULL) sb->sb_mtlstail = NULL; m->m_next = NULL; goto out; } /* * Need to allocate an mbuf to hold the remainder of 'm'. Try * with M_NOWAIT first. */ n = m_get(M_NOWAIT, MT_DATA); if (n == NULL) { /* * Use M_WAITOK with socket buffer unlocked. If * 'sb_mtls' changes while the lock is dropped, return * NULL to force the caller to retry. */ SOCKBUF_UNLOCK(sb); n = m_get(M_WAITOK, MT_DATA); SOCKBUF_LOCK(sb); if (sb->sb_mtls != top) { m_free(n); return (NULL); } } n->m_flags |= (m->m_flags & (M_NOTREADY | M_DECRYPTED)); /* Store remainder in 'n'. */ n->m_len = m->m_len - remain; if (m->m_flags & M_EXT) { n->m_data = m->m_data + remain; mb_dupcl(n, m); } else { bcopy(mtod(m, caddr_t) + remain, mtod(n, caddr_t), n->m_len); } /* Trim 'm' and update accounting. */ m->m_len -= n->m_len; sb->sb_tlscc -= n->m_len; sb->sb_ccc -= n->m_len; /* Account for 'n'. */ sballoc_ktls_rx(sb, n); /* Insert 'n' into the TLS chain. */ sb->sb_mtls = n; n->m_next = m->m_next; if (sb->sb_mtlstail == m) sb->sb_mtlstail = n; /* Detach the record from the TLS chain. */ m->m_next = NULL; out: MPASS(m_length(top, NULL) == len); for (m = top; m != NULL; m = m->m_next) sbfree_ktls_rx(sb, m); sb->sb_tlsdcc = len; sb->sb_ccc += len; SBCHECK(sb); return (top); } /* * Determine the length of the trailing zero padding and find the real * record type in the byte before the padding. * * Walking the mbuf chain backwards is clumsy, so another option would * be to scan forwards remembering the last non-zero byte before the * trailer. However, it would be expensive to scan the entire record. * Instead, find the last non-zero byte of each mbuf in the chain * keeping track of the relative offset of that nonzero byte. * * trail_len is the size of the MAC/tag on input and is set to the * size of the full trailer including padding and the record type on * return. */ static int tls13_find_record_type(struct ktls_session *tls, struct mbuf *m, int tls_len, int *trailer_len, uint8_t *record_typep) { char *cp; u_int digest_start, last_offset, m_len, offset; uint8_t record_type; digest_start = tls_len - *trailer_len; last_offset = 0; offset = 0; for (; m != NULL && offset < digest_start; offset += m->m_len, m = m->m_next) { /* Don't look for padding in the tag. */ m_len = min(digest_start - offset, m->m_len); cp = mtod(m, char *); /* Find last non-zero byte in this mbuf. */ while (m_len > 0 && cp[m_len - 1] == 0) m_len--; if (m_len > 0) { record_type = cp[m_len - 1]; last_offset = offset + m_len; } } if (last_offset < tls->params.tls_hlen) return (EBADMSG); *record_typep = record_type; *trailer_len = tls_len - last_offset + 1; return (0); } /* * Check if a mbuf chain is fully decrypted at the given offset and * length. Returns KTLS_MBUF_CRYPTO_ST_DECRYPTED if all data is * decrypted. KTLS_MBUF_CRYPTO_ST_MIXED if there is a mix of encrypted * and decrypted data. Else KTLS_MBUF_CRYPTO_ST_ENCRYPTED if all data * is encrypted. */ ktls_mbuf_crypto_st_t ktls_mbuf_crypto_state(struct mbuf *mb, int offset, int len) { int m_flags_ored = 0; int m_flags_anded = -1; for (; mb != NULL; mb = mb->m_next) { if (offset < mb->m_len) break; offset -= mb->m_len; } offset += len; for (; mb != NULL; mb = mb->m_next) { m_flags_ored |= mb->m_flags; m_flags_anded &= mb->m_flags; if (offset <= mb->m_len) break; offset -= mb->m_len; } MPASS(mb != NULL || offset == 0); if ((m_flags_ored ^ m_flags_anded) & M_DECRYPTED) return (KTLS_MBUF_CRYPTO_ST_MIXED); else return ((m_flags_ored & M_DECRYPTED) ? KTLS_MBUF_CRYPTO_ST_DECRYPTED : KTLS_MBUF_CRYPTO_ST_ENCRYPTED); } /* * ktls_resync_ifnet - get HW TLS RX back on track after packet loss */ static int ktls_resync_ifnet(struct socket *so, uint32_t tls_len, uint64_t tls_rcd_num) { union if_snd_tag_modify_params params; struct m_snd_tag *mst; struct inpcb *inp; struct tcpcb *tp; mst = so->so_rcv.sb_tls_info->snd_tag; if (__predict_false(mst == NULL)) return (EINVAL); inp = sotoinpcb(so); if (__predict_false(inp == NULL)) return (EINVAL); INP_RLOCK(inp); - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { INP_RUNLOCK(inp); return (ECONNRESET); } tp = intotcpcb(inp); MPASS(tp != NULL); /* Get the TCP sequence number of the next valid TLS header. */ SOCKBUF_LOCK(&so->so_rcv); params.tls_rx.tls_hdr_tcp_sn = tp->rcv_nxt - so->so_rcv.sb_tlscc - tls_len; params.tls_rx.tls_rec_length = tls_len; params.tls_rx.tls_seq_number = tls_rcd_num; SOCKBUF_UNLOCK(&so->so_rcv); INP_RUNLOCK(inp); MPASS(mst->sw->type == IF_SND_TAG_TYPE_TLS_RX); return (mst->sw->snd_tag_modify(mst, ¶ms)); } static void ktls_decrypt(struct socket *so) { char tls_header[MBUF_PEXT_HDR_LEN]; struct ktls_session *tls; struct sockbuf *sb; struct tls_record_layer *hdr; struct tls_get_record tgr; struct mbuf *control, *data, *m; ktls_mbuf_crypto_st_t state; uint64_t seqno; int error, remain, tls_len, trail_len; bool tls13; uint8_t vminor, record_type; hdr = (struct tls_record_layer *)tls_header; sb = &so->so_rcv; SOCKBUF_LOCK(sb); KASSERT(sb->sb_flags & SB_TLS_RX_RUNNING, ("%s: socket %p not running", __func__, so)); tls = sb->sb_tls_info; MPASS(tls != NULL); tls13 = (tls->params.tls_vminor == TLS_MINOR_VER_THREE); if (tls13) vminor = TLS_MINOR_VER_TWO; else vminor = tls->params.tls_vminor; for (;;) { /* Is there enough queued for a TLS header? */ if (sb->sb_tlscc < tls->params.tls_hlen) break; m_copydata(sb->sb_mtls, 0, tls->params.tls_hlen, tls_header); tls_len = sizeof(*hdr) + ntohs(hdr->tls_length); if (hdr->tls_vmajor != tls->params.tls_vmajor || hdr->tls_vminor != vminor) error = EINVAL; else if (tls13 && hdr->tls_type != TLS_RLTYPE_APP) error = EINVAL; else if (tls_len < tls->params.tls_hlen || tls_len > tls->params.tls_hlen + TLS_MAX_MSG_SIZE_V10_2 + tls->params.tls_tlen) error = EMSGSIZE; else error = 0; if (__predict_false(error != 0)) { /* * We have a corrupted record and are likely * out of sync. The connection isn't * recoverable at this point, so abort it. */ SOCKBUF_UNLOCK(sb); counter_u64_add(ktls_offload_corrupted_records, 1); CURVNET_SET(so->so_vnet); so->so_proto->pr_abort(so); so->so_error = error; CURVNET_RESTORE(); goto deref; } /* Is the entire record queued? */ if (sb->sb_tlscc < tls_len) break; /* * Split out the portion of the mbuf chain containing * this TLS record. */ data = ktls_detach_record(sb, tls_len); if (data == NULL) continue; MPASS(sb->sb_tlsdcc == tls_len); seqno = sb->sb_tls_seqno; sb->sb_tls_seqno++; SBCHECK(sb); SOCKBUF_UNLOCK(sb); /* get crypto state for this TLS record */ state = ktls_mbuf_crypto_state(data, 0, tls_len); switch (state) { case KTLS_MBUF_CRYPTO_ST_MIXED: error = ktls_ocf_recrypt(tls, hdr, data, seqno); if (error) break; /* FALLTHROUGH */ case KTLS_MBUF_CRYPTO_ST_ENCRYPTED: error = ktls_ocf_decrypt(tls, hdr, data, seqno, &trail_len); if (__predict_true(error == 0)) { if (tls13) { error = tls13_find_record_type(tls, data, tls_len, &trail_len, &record_type); } else { record_type = hdr->tls_type; } } break; case KTLS_MBUF_CRYPTO_ST_DECRYPTED: /* * NIC TLS is only supported for AEAD * ciphersuites which used a fixed sized * trailer. */ if (tls13) { trail_len = tls->params.tls_tlen - 1; error = tls13_find_record_type(tls, data, tls_len, &trail_len, &record_type); } else { trail_len = tls->params.tls_tlen; error = 0; record_type = hdr->tls_type; } break; default: error = EINVAL; break; } if (error) { counter_u64_add(ktls_offload_failed_crypto, 1); SOCKBUF_LOCK(sb); if (sb->sb_tlsdcc == 0) { /* * sbcut/drop/flush discarded these * mbufs. */ m_freem(data); break; } /* * Drop this TLS record's data, but keep * decrypting subsequent records. */ sb->sb_ccc -= tls_len; sb->sb_tlsdcc = 0; CURVNET_SET(so->so_vnet); so->so_error = EBADMSG; sorwakeup_locked(so); CURVNET_RESTORE(); m_freem(data); SOCKBUF_LOCK(sb); continue; } /* Allocate the control mbuf. */ memset(&tgr, 0, sizeof(tgr)); tgr.tls_type = record_type; tgr.tls_vmajor = hdr->tls_vmajor; tgr.tls_vminor = hdr->tls_vminor; tgr.tls_length = htobe16(tls_len - tls->params.tls_hlen - trail_len); control = sbcreatecontrol(&tgr, sizeof(tgr), TLS_GET_RECORD, IPPROTO_TCP, M_WAITOK); SOCKBUF_LOCK(sb); if (sb->sb_tlsdcc == 0) { /* sbcut/drop/flush discarded these mbufs. */ MPASS(sb->sb_tlscc == 0); m_freem(data); m_freem(control); break; } /* * Clear the 'dcc' accounting in preparation for * adding the decrypted record. */ sb->sb_ccc -= tls_len; sb->sb_tlsdcc = 0; SBCHECK(sb); /* If there is no payload, drop all of the data. */ if (tgr.tls_length == htobe16(0)) { m_freem(data); data = NULL; } else { /* Trim header. */ remain = tls->params.tls_hlen; while (remain > 0) { if (data->m_len > remain) { data->m_data += remain; data->m_len -= remain; break; } remain -= data->m_len; data = m_free(data); } /* Trim trailer and clear M_NOTREADY. */ remain = be16toh(tgr.tls_length); m = data; for (m = data; remain > m->m_len; m = m->m_next) { m->m_flags &= ~(M_NOTREADY | M_DECRYPTED); remain -= m->m_len; } m->m_len = remain; m_freem(m->m_next); m->m_next = NULL; m->m_flags &= ~(M_NOTREADY | M_DECRYPTED); /* Set EOR on the final mbuf. */ m->m_flags |= M_EOR; } sbappendcontrol_locked(sb, data, control, 0); if (__predict_false(state != KTLS_MBUF_CRYPTO_ST_DECRYPTED)) { sb->sb_flags |= SB_TLS_RX_RESYNC; SOCKBUF_UNLOCK(sb); ktls_resync_ifnet(so, tls_len, seqno); SOCKBUF_LOCK(sb); } else if (__predict_false(sb->sb_flags & SB_TLS_RX_RESYNC)) { sb->sb_flags &= ~SB_TLS_RX_RESYNC; SOCKBUF_UNLOCK(sb); ktls_resync_ifnet(so, 0, seqno); SOCKBUF_LOCK(sb); } } sb->sb_flags &= ~SB_TLS_RX_RUNNING; if ((sb->sb_state & SBS_CANTRCVMORE) != 0 && sb->sb_tlscc > 0) so->so_error = EMSGSIZE; sorwakeup_locked(so); deref: SOCKBUF_UNLOCK_ASSERT(sb); CURVNET_SET(so->so_vnet); sorele(so); CURVNET_RESTORE(); } void ktls_enqueue_to_free(struct mbuf *m) { struct ktls_wq *wq; bool running; /* Mark it for freeing. */ m->m_epg_flags |= EPG_FLAG_2FREE; wq = &ktls_wq[m->m_epg_tls->wq_index]; mtx_lock(&wq->mtx); STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq); running = wq->running; mtx_unlock(&wq->mtx); if (!running) wakeup(wq); } static void * ktls_buffer_alloc(struct ktls_wq *wq, struct mbuf *m) { void *buf; int domain, running; if (m->m_epg_npgs <= 2) return (NULL); if (ktls_buffer_zone == NULL) return (NULL); if ((u_int)(ticks - wq->lastallocfail) < hz) { /* * Rate-limit allocation attempts after a failure. * ktls_buffer_import() will acquire a per-domain mutex to check * the free page queues and may fail consistently if memory is * fragmented. */ return (NULL); } buf = uma_zalloc(ktls_buffer_zone, M_NOWAIT | M_NORECLAIM); if (buf == NULL) { domain = PCPU_GET(domain); wq->lastallocfail = ticks; /* * Note that this check is "racy", but the races are * harmless, and are either a spurious wakeup if * multiple threads fail allocations before the alloc * thread wakes, or waiting an extra second in case we * see an old value of running == true. */ if (!VM_DOMAIN_EMPTY(domain)) { running = atomic_load_int(&ktls_domains[domain].alloc_td.running); if (!running) wakeup(&ktls_domains[domain].alloc_td); } } return (buf); } static int ktls_encrypt_record(struct ktls_wq *wq, struct mbuf *m, struct ktls_session *tls, struct ktls_ocf_encrypt_state *state) { vm_page_t pg; int error, i, len, off; KASSERT((m->m_flags & (M_EXTPG | M_NOTREADY)) == (M_EXTPG | M_NOTREADY), ("%p not unready & nomap mbuf\n", m)); KASSERT(ptoa(m->m_epg_npgs) <= ktls_maxlen, ("page count %d larger than maximum frame length %d", m->m_epg_npgs, ktls_maxlen)); /* Anonymous mbufs are encrypted in place. */ if ((m->m_epg_flags & EPG_FLAG_ANON) != 0) return (ktls_ocf_encrypt(state, tls, m, NULL, 0)); /* * For file-backed mbufs (from sendfile), anonymous wired * pages are allocated and used as the encryption destination. */ if ((state->cbuf = ktls_buffer_alloc(wq, m)) != NULL) { len = ptoa(m->m_epg_npgs - 1) + m->m_epg_last_len - m->m_epg_1st_off; state->dst_iov[0].iov_base = (char *)state->cbuf + m->m_epg_1st_off; state->dst_iov[0].iov_len = len; state->parray[0] = DMAP_TO_PHYS((vm_offset_t)state->cbuf); i = 1; } else { off = m->m_epg_1st_off; for (i = 0; i < m->m_epg_npgs; i++, off = 0) { pg = vm_page_alloc_noobj(VM_ALLOC_NODUMP | VM_ALLOC_WIRED | VM_ALLOC_WAITOK); len = m_epg_pagelen(m, i, off); state->parray[i] = VM_PAGE_TO_PHYS(pg); state->dst_iov[i].iov_base = (char *)PHYS_TO_DMAP(state->parray[i]) + off; state->dst_iov[i].iov_len = len; } } KASSERT(i + 1 <= nitems(state->dst_iov), ("dst_iov is too small")); state->dst_iov[i].iov_base = m->m_epg_trail; state->dst_iov[i].iov_len = m->m_epg_trllen; error = ktls_ocf_encrypt(state, tls, m, state->dst_iov, i + 1); if (__predict_false(error != 0)) { /* Free the anonymous pages. */ if (state->cbuf != NULL) uma_zfree(ktls_buffer_zone, state->cbuf); else { for (i = 0; i < m->m_epg_npgs; i++) { pg = PHYS_TO_VM_PAGE(state->parray[i]); (void)vm_page_unwire_noq(pg); vm_page_free(pg); } } } return (error); } /* Number of TLS records in a batch passed to ktls_enqueue(). */ static u_int ktls_batched_records(struct mbuf *m) { int page_count, records; records = 0; page_count = m->m_epg_enc_cnt; while (page_count > 0) { records++; page_count -= m->m_epg_nrdy; m = m->m_next; } KASSERT(page_count == 0, ("%s: mismatched page count", __func__)); return (records); } void ktls_enqueue(struct mbuf *m, struct socket *so, int page_count) { struct ktls_session *tls; struct ktls_wq *wq; int queued; bool running; KASSERT(((m->m_flags & (M_EXTPG | M_NOTREADY)) == (M_EXTPG | M_NOTREADY)), ("ktls_enqueue: %p not unready & nomap mbuf\n", m)); KASSERT(page_count != 0, ("enqueueing TLS mbuf with zero page count")); KASSERT(m->m_epg_tls->mode == TCP_TLS_MODE_SW, ("!SW TLS mbuf")); m->m_epg_enc_cnt = page_count; /* * Save a pointer to the socket. The caller is responsible * for taking an additional reference via soref(). */ m->m_epg_so = so; queued = 1; tls = m->m_epg_tls; wq = &ktls_wq[tls->wq_index]; mtx_lock(&wq->mtx); if (__predict_false(tls->sequential_records)) { /* * For TLS 1.0, records must be encrypted * sequentially. For a given connection, all records * queued to the associated work queue are processed * sequentially. However, sendfile(2) might complete * I/O requests spanning multiple TLS records out of * order. Here we ensure TLS records are enqueued to * the work queue in FIFO order. * * tls->next_seqno holds the sequence number of the * next TLS record that should be enqueued to the work * queue. If this next record is not tls->next_seqno, * it must be a future record, so insert it, sorted by * TLS sequence number, into tls->pending_records and * return. * * If this TLS record matches tls->next_seqno, place * it in the work queue and then check * tls->pending_records to see if any * previously-queued records are now ready for * encryption. */ if (m->m_epg_seqno != tls->next_seqno) { struct mbuf *n, *p; p = NULL; STAILQ_FOREACH(n, &tls->pending_records, m_epg_stailq) { if (n->m_epg_seqno > m->m_epg_seqno) break; p = n; } if (n == NULL) STAILQ_INSERT_TAIL(&tls->pending_records, m, m_epg_stailq); else if (p == NULL) STAILQ_INSERT_HEAD(&tls->pending_records, m, m_epg_stailq); else STAILQ_INSERT_AFTER(&tls->pending_records, p, m, m_epg_stailq); mtx_unlock(&wq->mtx); counter_u64_add(ktls_cnt_tx_pending, 1); return; } tls->next_seqno += ktls_batched_records(m); STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq); while (!STAILQ_EMPTY(&tls->pending_records)) { struct mbuf *n; n = STAILQ_FIRST(&tls->pending_records); if (n->m_epg_seqno != tls->next_seqno) break; queued++; STAILQ_REMOVE_HEAD(&tls->pending_records, m_epg_stailq); tls->next_seqno += ktls_batched_records(n); STAILQ_INSERT_TAIL(&wq->m_head, n, m_epg_stailq); } counter_u64_add(ktls_cnt_tx_pending, -(queued - 1)); } else STAILQ_INSERT_TAIL(&wq->m_head, m, m_epg_stailq); running = wq->running; mtx_unlock(&wq->mtx); if (!running) wakeup(wq); counter_u64_add(ktls_cnt_tx_queued, queued); } /* * Once a file-backed mbuf (from sendfile) has been encrypted, free * the pages from the file and replace them with the anonymous pages * allocated in ktls_encrypt_record(). */ static void ktls_finish_nonanon(struct mbuf *m, struct ktls_ocf_encrypt_state *state) { int i; MPASS((m->m_epg_flags & EPG_FLAG_ANON) == 0); /* Free the old pages. */ m->m_ext.ext_free(m); /* Replace them with the new pages. */ if (state->cbuf != NULL) { for (i = 0; i < m->m_epg_npgs; i++) m->m_epg_pa[i] = state->parray[0] + ptoa(i); /* Contig pages should go back to the cache. */ m->m_ext.ext_free = ktls_free_mext_contig; } else { for (i = 0; i < m->m_epg_npgs; i++) m->m_epg_pa[i] = state->parray[i]; /* Use the basic free routine. */ m->m_ext.ext_free = mb_free_mext_pgs; } /* Pages are now writable. */ m->m_epg_flags |= EPG_FLAG_ANON; } static __noinline void ktls_encrypt(struct ktls_wq *wq, struct mbuf *top) { struct ktls_ocf_encrypt_state state; struct ktls_session *tls; struct socket *so; struct mbuf *m; int error, npages, total_pages; so = top->m_epg_so; tls = top->m_epg_tls; KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top)); KASSERT(so != NULL, ("so = NULL, top = %p\n", top)); #ifdef INVARIANTS top->m_epg_so = NULL; #endif total_pages = top->m_epg_enc_cnt; npages = 0; /* * Encrypt the TLS records in the chain of mbufs starting with * 'top'. 'total_pages' gives us a total count of pages and is * used to know when we have finished encrypting the TLS * records originally queued with 'top'. * * NB: These mbufs are queued in the socket buffer and * 'm_next' is traversing the mbufs in the socket buffer. The * socket buffer lock is not held while traversing this chain. * Since the mbufs are all marked M_NOTREADY their 'm_next' * pointers should be stable. However, the 'm_next' of the * last mbuf encrypted is not necessarily NULL. It can point * to other mbufs appended while 'top' was on the TLS work * queue. * * Each mbuf holds an entire TLS record. */ error = 0; for (m = top; npages != total_pages; m = m->m_next) { KASSERT(m->m_epg_tls == tls, ("different TLS sessions in a single mbuf chain: %p vs %p", tls, m->m_epg_tls)); KASSERT(npages + m->m_epg_npgs <= total_pages, ("page count mismatch: top %p, total_pages %d, m %p", top, total_pages, m)); error = ktls_encrypt_record(wq, m, tls, &state); if (error) { counter_u64_add(ktls_offload_failed_crypto, 1); break; } if ((m->m_epg_flags & EPG_FLAG_ANON) == 0) ktls_finish_nonanon(m, &state); npages += m->m_epg_nrdy; /* * Drop a reference to the session now that it is no * longer needed. Existing code depends on encrypted * records having no associated session vs * yet-to-be-encrypted records having an associated * session. */ m->m_epg_tls = NULL; ktls_free(tls); } CURVNET_SET(so->so_vnet); if (error == 0) { (void)so->so_proto->pr_ready(so, top, npages); } else { so->so_proto->pr_abort(so); so->so_error = EIO; mb_free_notready(top, total_pages); } sorele(so); CURVNET_RESTORE(); } void ktls_encrypt_cb(struct ktls_ocf_encrypt_state *state, int error) { struct ktls_session *tls; struct socket *so; struct mbuf *m; int npages; m = state->m; if ((m->m_epg_flags & EPG_FLAG_ANON) == 0) ktls_finish_nonanon(m, state); so = state->so; free(state, M_KTLS); /* * Drop a reference to the session now that it is no longer * needed. Existing code depends on encrypted records having * no associated session vs yet-to-be-encrypted records having * an associated session. */ tls = m->m_epg_tls; m->m_epg_tls = NULL; ktls_free(tls); if (error != 0) counter_u64_add(ktls_offload_failed_crypto, 1); CURVNET_SET(so->so_vnet); npages = m->m_epg_nrdy; if (error == 0) { (void)so->so_proto->pr_ready(so, m, npages); } else { so->so_proto->pr_abort(so); so->so_error = EIO; mb_free_notready(m, npages); } sorele(so); CURVNET_RESTORE(); } /* * Similar to ktls_encrypt, but used with asynchronous OCF backends * (coprocessors) where encryption does not use host CPU resources and * it can be beneficial to queue more requests than CPUs. */ static __noinline void ktls_encrypt_async(struct ktls_wq *wq, struct mbuf *top) { struct ktls_ocf_encrypt_state *state; struct ktls_session *tls; struct socket *so; struct mbuf *m, *n; int error, mpages, npages, total_pages; so = top->m_epg_so; tls = top->m_epg_tls; KASSERT(tls != NULL, ("tls = NULL, top = %p\n", top)); KASSERT(so != NULL, ("so = NULL, top = %p\n", top)); #ifdef INVARIANTS top->m_epg_so = NULL; #endif total_pages = top->m_epg_enc_cnt; npages = 0; error = 0; for (m = top; npages != total_pages; m = n) { KASSERT(m->m_epg_tls == tls, ("different TLS sessions in a single mbuf chain: %p vs %p", tls, m->m_epg_tls)); KASSERT(npages + m->m_epg_npgs <= total_pages, ("page count mismatch: top %p, total_pages %d, m %p", top, total_pages, m)); state = malloc(sizeof(*state), M_KTLS, M_WAITOK | M_ZERO); soref(so); state->so = so; state->m = m; mpages = m->m_epg_nrdy; n = m->m_next; error = ktls_encrypt_record(wq, m, tls, state); if (error) { counter_u64_add(ktls_offload_failed_crypto, 1); free(state, M_KTLS); CURVNET_SET(so->so_vnet); sorele(so); CURVNET_RESTORE(); break; } npages += mpages; } CURVNET_SET(so->so_vnet); if (error != 0) { so->so_proto->pr_abort(so); so->so_error = EIO; mb_free_notready(m, total_pages - npages); } sorele(so); CURVNET_RESTORE(); } static int ktls_bind_domain(int domain) { int error; error = cpuset_setthread(curthread->td_tid, &cpuset_domain[domain]); if (error != 0) return (error); curthread->td_domain.dr_policy = DOMAINSET_PREF(domain); return (0); } static void ktls_alloc_thread(void *ctx) { struct ktls_domain_info *ktls_domain = ctx; struct ktls_alloc_thread *sc = &ktls_domain->alloc_td; void **buf; struct sysctl_oid *oid; char name[80]; int domain, error, i, nbufs; domain = ktls_domain - ktls_domains; if (bootverbose) printf("Starting KTLS alloc thread for domain %d\n", domain); error = ktls_bind_domain(domain); if (error) printf("Unable to bind KTLS alloc thread for domain %d: error %d\n", domain, error); snprintf(name, sizeof(name), "domain%d", domain); oid = SYSCTL_ADD_NODE(NULL, SYSCTL_STATIC_CHILDREN(_kern_ipc_tls), OID_AUTO, name, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, ""); SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "allocs", CTLFLAG_RD, &sc->allocs, 0, "buffers allocated"); SYSCTL_ADD_U64(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "wakeups", CTLFLAG_RD, &sc->wakeups, 0, "thread wakeups"); SYSCTL_ADD_INT(NULL, SYSCTL_CHILDREN(oid), OID_AUTO, "running", CTLFLAG_RD, &sc->running, 0, "thread running"); buf = NULL; nbufs = 0; for (;;) { atomic_store_int(&sc->running, 0); tsleep(sc, PZERO | PNOLOCK, "-", 0); atomic_store_int(&sc->running, 1); sc->wakeups++; if (nbufs != ktls_max_alloc) { free(buf, M_KTLS); nbufs = atomic_load_int(&ktls_max_alloc); buf = malloc(sizeof(void *) * nbufs, M_KTLS, M_WAITOK | M_ZERO); } /* * Below we allocate nbufs with different allocation * flags than we use when allocating normally during * encryption in the ktls worker thread. We specify * M_NORECLAIM in the worker thread. However, we omit * that flag here and add M_WAITOK so that the VM * system is permitted to perform expensive work to * defragment memory. We do this here, as it does not * matter if this thread blocks. If we block a ktls * worker thread, we risk developing backlogs of * buffers to be encrypted, leading to surges of * traffic and potential NIC output drops. */ for (i = 0; i < nbufs; i++) { buf[i] = uma_zalloc(ktls_buffer_zone, M_WAITOK); sc->allocs++; } for (i = 0; i < nbufs; i++) { uma_zfree(ktls_buffer_zone, buf[i]); buf[i] = NULL; } } } static void ktls_work_thread(void *ctx) { struct ktls_wq *wq = ctx; struct mbuf *m, *n; struct socket *so, *son; STAILQ_HEAD(, mbuf) local_m_head; STAILQ_HEAD(, socket) local_so_head; int cpu; cpu = wq - ktls_wq; if (bootverbose) printf("Starting KTLS worker thread for CPU %d\n", cpu); /* * Bind to a core. If ktls_bind_threads is > 1, then * we bind to the NUMA domain instead. */ if (ktls_bind_threads) { int error; if (ktls_bind_threads > 1) { struct pcpu *pc = pcpu_find(cpu); error = ktls_bind_domain(pc->pc_domain); } else { cpuset_t mask; CPU_SETOF(cpu, &mask); error = cpuset_setthread(curthread->td_tid, &mask); } if (error) printf("Unable to bind KTLS worker thread for CPU %d: error %d\n", cpu, error); } #if defined(__aarch64__) || defined(__amd64__) || defined(__i386__) fpu_kern_thread(0); #endif for (;;) { mtx_lock(&wq->mtx); while (STAILQ_EMPTY(&wq->m_head) && STAILQ_EMPTY(&wq->so_head)) { wq->running = false; mtx_sleep(wq, &wq->mtx, 0, "-", 0); wq->running = true; } STAILQ_INIT(&local_m_head); STAILQ_CONCAT(&local_m_head, &wq->m_head); STAILQ_INIT(&local_so_head); STAILQ_CONCAT(&local_so_head, &wq->so_head); mtx_unlock(&wq->mtx); STAILQ_FOREACH_SAFE(m, &local_m_head, m_epg_stailq, n) { if (m->m_epg_flags & EPG_FLAG_2FREE) { ktls_free(m->m_epg_tls); m_free_raw(m); } else { if (m->m_epg_tls->sync_dispatch) ktls_encrypt(wq, m); else ktls_encrypt_async(wq, m); counter_u64_add(ktls_cnt_tx_queued, -1); } } STAILQ_FOREACH_SAFE(so, &local_so_head, so_ktls_rx_list, son) { ktls_decrypt(so); counter_u64_add(ktls_cnt_rx_queued, -1); } } } #if defined(INET) || defined(INET6) static void ktls_disable_ifnet_help(void *context, int pending __unused) { struct ktls_session *tls; struct inpcb *inp; struct tcpcb *tp; struct socket *so; int err; tls = context; inp = tls->inp; if (inp == NULL) return; INP_WLOCK(inp); so = inp->inp_socket; MPASS(so != NULL); - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { goto out; } if (so->so_snd.sb_tls_info != NULL) err = ktls_set_tx_mode(so, TCP_TLS_MODE_SW); else err = ENXIO; if (err == 0) { counter_u64_add(ktls_ifnet_disable_ok, 1); /* ktls_set_tx_mode() drops inp wlock, so recheck flags */ - if ((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) == 0 && + if ((inp->inp_flags & INP_DROPPED) == 0 && (tp = intotcpcb(inp)) != NULL && tp->t_fb->tfb_hwtls_change != NULL) (*tp->t_fb->tfb_hwtls_change)(tp, 0); } else { counter_u64_add(ktls_ifnet_disable_fail, 1); } out: sorele(so); if (!in_pcbrele_wlocked(inp)) INP_WUNLOCK(inp); ktls_free(tls); } /* * Called when re-transmits are becoming a substantial portion of the * sends on this connection. When this happens, we transition the * connection to software TLS. This is needed because most inline TLS * NICs keep crypto state only for in-order transmits. This means * that to handle a TCP rexmit (which is out-of-order), the NIC must * re-DMA the entire TLS record up to and including the current * segment. This means that when re-transmitting the last ~1448 byte * segment of a 16KB TLS record, we could wind up re-DMA'ing an order * of magnitude more data than we are sending. This can cause the * PCIe link to saturate well before the network, which can cause * output drops, and a general loss of capacity. */ void ktls_disable_ifnet(void *arg) { struct tcpcb *tp; struct inpcb *inp; struct socket *so; struct ktls_session *tls; tp = arg; inp = tp->t_inpcb; INP_WLOCK_ASSERT(inp); so = inp->inp_socket; SOCK_LOCK(so); tls = so->so_snd.sb_tls_info; if (tls->disable_ifnet_pending) { SOCK_UNLOCK(so); return; } /* * note that disable_ifnet_pending is never cleared; disabling * ifnet can only be done once per session, so we never want * to do it again */ (void)ktls_hold(tls); in_pcbref(inp); soref(so); tls->disable_ifnet_pending = true; tls->inp = inp; SOCK_UNLOCK(so); TASK_INIT(&tls->disable_ifnet_task, 0, ktls_disable_ifnet_help, tls); (void)taskqueue_enqueue(taskqueue_thread, &tls->disable_ifnet_task); } #endif diff --git a/sys/netinet/in_pcb.c b/sys/netinet/in_pcb.c index 9dd6d3d019ca..75889b2011ac 100644 --- a/sys/netinet/in_pcb.c +++ b/sys/netinet/in_pcb.c @@ -1,3432 +1,3427 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1982, 1986, 1991, 1993, 1995 * The Regents of the University of California. * Copyright (c) 2007-2009 Robert N. M. Watson * Copyright (c) 2010-2011 Juniper Networks, Inc. * All rights reserved. * * Portions of this software were developed by Robert N. M. Watson under * contract to Juniper Networks, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)in_pcb.c 8.4 (Berkeley) 5/24/95 */ #include __FBSDID("$FreeBSD$"); #include "opt_ddb.h" #include "opt_ipsec.h" #include "opt_inet.h" #include "opt_inet6.h" #include "opt_ratelimit.h" #include "opt_route.h" #include "opt_rss.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef DDB #include #endif #include #include #include #include #include #include #include #include #include #if defined(INET) || defined(INET6) #include #include #include #ifdef INET #include #include #endif #include #include #ifdef TCPHPTS #include #endif #include #include #ifdef INET6 #include #include #include #include #endif /* INET6 */ #include #endif #include #include #define INPCBLBGROUP_SIZMIN 8 #define INPCBLBGROUP_SIZMAX 256 #define INP_FREED 0x00000200 /* See in_pcb.h. */ static struct callout ipport_tick_callout; /* * These configure the range of local port addresses assigned to * "unspecified" outgoing connections/packets/whatever. */ VNET_DEFINE(int, ipport_lowfirstauto) = IPPORT_RESERVED - 1; /* 1023 */ VNET_DEFINE(int, ipport_lowlastauto) = IPPORT_RESERVEDSTART; /* 600 */ VNET_DEFINE(int, ipport_firstauto) = IPPORT_EPHEMERALFIRST; /* 10000 */ VNET_DEFINE(int, ipport_lastauto) = IPPORT_EPHEMERALLAST; /* 65535 */ VNET_DEFINE(int, ipport_hifirstauto) = IPPORT_HIFIRSTAUTO; /* 49152 */ VNET_DEFINE(int, ipport_hilastauto) = IPPORT_HILASTAUTO; /* 65535 */ /* * Reserved ports accessible only to root. There are significant * security considerations that must be accounted for when changing these, * but the security benefits can be great. Please be careful. */ VNET_DEFINE(int, ipport_reservedhigh) = IPPORT_RESERVED - 1; /* 1023 */ VNET_DEFINE(int, ipport_reservedlow); /* Variables dealing with random ephemeral port allocation. */ VNET_DEFINE(int, ipport_randomized) = 1; /* user controlled via sysctl */ VNET_DEFINE(int, ipport_randomcps) = 10; /* user controlled via sysctl */ VNET_DEFINE(int, ipport_randomtime) = 45; /* user controlled via sysctl */ VNET_DEFINE(int, ipport_stoprandom); /* toggled by ipport_tick */ VNET_DEFINE(int, ipport_tcpallocs); VNET_DEFINE_STATIC(int, ipport_tcplastcount); #define V_ipport_tcplastcount VNET(ipport_tcplastcount) #ifdef INET static struct inpcb *in_pcblookup_hash_locked(struct inpcbinfo *pcbinfo, struct in_addr faddr, u_int fport_arg, struct in_addr laddr, u_int lport_arg, int lookupflags, struct ifnet *ifp, uint8_t numa_domain); #define RANGECHK(var, min, max) \ if ((var) < (min)) { (var) = (min); } \ else if ((var) > (max)) { (var) = (max); } static int sysctl_net_ipport_check(SYSCTL_HANDLER_ARGS) { int error; error = sysctl_handle_int(oidp, arg1, arg2, req); if (error == 0) { RANGECHK(V_ipport_lowfirstauto, 1, IPPORT_RESERVED - 1); RANGECHK(V_ipport_lowlastauto, 1, IPPORT_RESERVED - 1); RANGECHK(V_ipport_firstauto, IPPORT_RESERVED, IPPORT_MAX); RANGECHK(V_ipport_lastauto, IPPORT_RESERVED, IPPORT_MAX); RANGECHK(V_ipport_hifirstauto, IPPORT_RESERVED, IPPORT_MAX); RANGECHK(V_ipport_hilastauto, IPPORT_RESERVED, IPPORT_MAX); } return (error); } #undef RANGECHK static SYSCTL_NODE(_net_inet_ip, IPPROTO_IP, portrange, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "IP Ports"); SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, lowfirst, CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &VNET_NAME(ipport_lowfirstauto), 0, &sysctl_net_ipport_check, "I", ""); SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, lowlast, CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &VNET_NAME(ipport_lowlastauto), 0, &sysctl_net_ipport_check, "I", ""); SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, first, CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &VNET_NAME(ipport_firstauto), 0, &sysctl_net_ipport_check, "I", ""); SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, last, CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &VNET_NAME(ipport_lastauto), 0, &sysctl_net_ipport_check, "I", ""); SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, hifirst, CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &VNET_NAME(ipport_hifirstauto), 0, &sysctl_net_ipport_check, "I", ""); SYSCTL_PROC(_net_inet_ip_portrange, OID_AUTO, hilast, CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &VNET_NAME(ipport_hilastauto), 0, &sysctl_net_ipport_check, "I", ""); SYSCTL_INT(_net_inet_ip_portrange, OID_AUTO, reservedhigh, CTLFLAG_VNET | CTLFLAG_RW | CTLFLAG_SECURE, &VNET_NAME(ipport_reservedhigh), 0, ""); SYSCTL_INT(_net_inet_ip_portrange, OID_AUTO, reservedlow, CTLFLAG_RW|CTLFLAG_SECURE, &VNET_NAME(ipport_reservedlow), 0, ""); SYSCTL_INT(_net_inet_ip_portrange, OID_AUTO, randomized, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(ipport_randomized), 0, "Enable random port allocation"); SYSCTL_INT(_net_inet_ip_portrange, OID_AUTO, randomcps, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(ipport_randomcps), 0, "Maximum number of random port " "allocations before switching to a sequential one"); SYSCTL_INT(_net_inet_ip_portrange, OID_AUTO, randomtime, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(ipport_randomtime), 0, "Minimum time to keep sequential port " "allocation before switching to a random one"); #ifdef RATELIMIT counter_u64_t rate_limit_new; counter_u64_t rate_limit_chg; counter_u64_t rate_limit_active; counter_u64_t rate_limit_alloc_fail; counter_u64_t rate_limit_set_ok; static SYSCTL_NODE(_net_inet_ip, OID_AUTO, rl, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "IP Rate Limiting"); SYSCTL_COUNTER_U64(_net_inet_ip_rl, OID_AUTO, active, CTLFLAG_RD, &rate_limit_active, "Active rate limited connections"); SYSCTL_COUNTER_U64(_net_inet_ip_rl, OID_AUTO, alloc_fail, CTLFLAG_RD, &rate_limit_alloc_fail, "Rate limited connection failures"); SYSCTL_COUNTER_U64(_net_inet_ip_rl, OID_AUTO, set_ok, CTLFLAG_RD, &rate_limit_set_ok, "Rate limited setting succeeded"); SYSCTL_COUNTER_U64(_net_inet_ip_rl, OID_AUTO, newrl, CTLFLAG_RD, &rate_limit_new, "Total Rate limit new attempts"); SYSCTL_COUNTER_U64(_net_inet_ip_rl, OID_AUTO, chgrl, CTLFLAG_RD, &rate_limit_chg, "Total Rate limited change attempts"); #endif /* RATELIMIT */ #endif /* INET */ VNET_DEFINE(uint32_t, in_pcbhashseed); static void in_pcbhashseed_init(void) { V_in_pcbhashseed = arc4random(); } VNET_SYSINIT(in_pcbhashseed_init, SI_SUB_PROTO_DOMAIN, SI_ORDER_FIRST, in_pcbhashseed_init, 0); /* * in_pcb.c: manage the Protocol Control Blocks. * * NOTE: It is assumed that most of these functions will be called with * the pcbinfo lock held, and often, the inpcb lock held, as these utility * functions often modify hash chains or addresses in pcbs. */ static struct inpcblbgroup * in_pcblbgroup_alloc(struct inpcblbgrouphead *hdr, u_char vflag, uint16_t port, const union in_dependaddr *addr, int size, uint8_t numa_domain) { struct inpcblbgroup *grp; size_t bytes; bytes = __offsetof(struct inpcblbgroup, il_inp[size]); grp = malloc(bytes, M_PCB, M_ZERO | M_NOWAIT); if (!grp) return (NULL); grp->il_vflag = vflag; grp->il_lport = port; grp->il_numa_domain = numa_domain; grp->il_dependladdr = *addr; grp->il_inpsiz = size; CK_LIST_INSERT_HEAD(hdr, grp, il_list); return (grp); } static void in_pcblbgroup_free_deferred(epoch_context_t ctx) { struct inpcblbgroup *grp; grp = __containerof(ctx, struct inpcblbgroup, il_epoch_ctx); free(grp, M_PCB); } static void in_pcblbgroup_free(struct inpcblbgroup *grp) { CK_LIST_REMOVE(grp, il_list); NET_EPOCH_CALL(in_pcblbgroup_free_deferred, &grp->il_epoch_ctx); } static struct inpcblbgroup * in_pcblbgroup_resize(struct inpcblbgrouphead *hdr, struct inpcblbgroup *old_grp, int size) { struct inpcblbgroup *grp; int i; grp = in_pcblbgroup_alloc(hdr, old_grp->il_vflag, old_grp->il_lport, &old_grp->il_dependladdr, size, old_grp->il_numa_domain); if (grp == NULL) return (NULL); KASSERT(old_grp->il_inpcnt < grp->il_inpsiz, ("invalid new local group size %d and old local group count %d", grp->il_inpsiz, old_grp->il_inpcnt)); for (i = 0; i < old_grp->il_inpcnt; ++i) grp->il_inp[i] = old_grp->il_inp[i]; grp->il_inpcnt = old_grp->il_inpcnt; in_pcblbgroup_free(old_grp); return (grp); } /* * PCB at index 'i' is removed from the group. Pull up the ones below il_inp[i] * and shrink group if possible. */ static void in_pcblbgroup_reorder(struct inpcblbgrouphead *hdr, struct inpcblbgroup **grpp, int i) { struct inpcblbgroup *grp, *new_grp; grp = *grpp; for (; i + 1 < grp->il_inpcnt; ++i) grp->il_inp[i] = grp->il_inp[i + 1]; grp->il_inpcnt--; if (grp->il_inpsiz > INPCBLBGROUP_SIZMIN && grp->il_inpcnt <= grp->il_inpsiz / 4) { /* Shrink this group. */ new_grp = in_pcblbgroup_resize(hdr, grp, grp->il_inpsiz / 2); if (new_grp != NULL) *grpp = new_grp; } } /* * Add PCB to load balance group for SO_REUSEPORT_LB option. */ static int in_pcbinslbgrouphash(struct inpcb *inp, uint8_t numa_domain) { const static struct timeval interval = { 60, 0 }; static struct timeval lastprint; struct inpcbinfo *pcbinfo; struct inpcblbgrouphead *hdr; struct inpcblbgroup *grp; uint32_t idx; pcbinfo = inp->inp_pcbinfo; INP_WLOCK_ASSERT(inp); INP_HASH_WLOCK_ASSERT(pcbinfo); /* * Don't allow jailed socket to join local group. */ if (inp->inp_socket != NULL && jailed(inp->inp_socket->so_cred)) return (0); #ifdef INET6 /* * Don't allow IPv4 mapped INET6 wild socket. */ if ((inp->inp_vflag & INP_IPV4) && inp->inp_laddr.s_addr == INADDR_ANY && INP_CHECK_SOCKAF(inp->inp_socket, AF_INET6)) { return (0); } #endif idx = INP_PCBPORTHASH(inp->inp_lport, pcbinfo->ipi_lbgrouphashmask); hdr = &pcbinfo->ipi_lbgrouphashbase[idx]; CK_LIST_FOREACH(grp, hdr, il_list) { if (grp->il_vflag == inp->inp_vflag && grp->il_lport == inp->inp_lport && grp->il_numa_domain == numa_domain && memcmp(&grp->il_dependladdr, &inp->inp_inc.inc_ie.ie_dependladdr, sizeof(grp->il_dependladdr)) == 0) break; } if (grp == NULL) { /* Create new load balance group. */ grp = in_pcblbgroup_alloc(hdr, inp->inp_vflag, inp->inp_lport, &inp->inp_inc.inc_ie.ie_dependladdr, INPCBLBGROUP_SIZMIN, numa_domain); if (grp == NULL) return (ENOBUFS); } else if (grp->il_inpcnt == grp->il_inpsiz) { if (grp->il_inpsiz >= INPCBLBGROUP_SIZMAX) { if (ratecheck(&lastprint, &interval)) printf("lb group port %d, limit reached\n", ntohs(grp->il_lport)); return (0); } /* Expand this local group. */ grp = in_pcblbgroup_resize(hdr, grp, grp->il_inpsiz * 2); if (grp == NULL) return (ENOBUFS); } KASSERT(grp->il_inpcnt < grp->il_inpsiz, ("invalid local group size %d and count %d", grp->il_inpsiz, grp->il_inpcnt)); grp->il_inp[grp->il_inpcnt] = inp; grp->il_inpcnt++; return (0); } /* * Remove PCB from load balance group. */ static void in_pcbremlbgrouphash(struct inpcb *inp) { struct inpcbinfo *pcbinfo; struct inpcblbgrouphead *hdr; struct inpcblbgroup *grp; int i; pcbinfo = inp->inp_pcbinfo; INP_WLOCK_ASSERT(inp); INP_HASH_WLOCK_ASSERT(pcbinfo); hdr = &pcbinfo->ipi_lbgrouphashbase[ INP_PCBPORTHASH(inp->inp_lport, pcbinfo->ipi_lbgrouphashmask)]; CK_LIST_FOREACH(grp, hdr, il_list) { for (i = 0; i < grp->il_inpcnt; ++i) { if (grp->il_inp[i] != inp) continue; if (grp->il_inpcnt == 1) { /* We are the last, free this local group. */ in_pcblbgroup_free(grp); } else { /* Pull up inpcbs, shrink group if possible. */ in_pcblbgroup_reorder(hdr, &grp, i); } return; } } } int in_pcblbgroup_numa(struct inpcb *inp, int arg) { struct inpcbinfo *pcbinfo; struct inpcblbgrouphead *hdr; struct inpcblbgroup *grp; int err, i; uint8_t numa_domain; switch (arg) { case TCP_REUSPORT_LB_NUMA_NODOM: numa_domain = M_NODOM; break; case TCP_REUSPORT_LB_NUMA_CURDOM: numa_domain = PCPU_GET(domain); break; default: if (arg < 0 || arg >= vm_ndomains) return (EINVAL); numa_domain = arg; } err = 0; pcbinfo = inp->inp_pcbinfo; INP_WLOCK_ASSERT(inp); INP_HASH_WLOCK(pcbinfo); hdr = &pcbinfo->ipi_lbgrouphashbase[ INP_PCBPORTHASH(inp->inp_lport, pcbinfo->ipi_lbgrouphashmask)]; CK_LIST_FOREACH(grp, hdr, il_list) { for (i = 0; i < grp->il_inpcnt; ++i) { if (grp->il_inp[i] != inp) continue; if (grp->il_numa_domain == numa_domain) { goto abort_with_hash_wlock; } /* Remove it from the old group. */ in_pcbremlbgrouphash(inp); /* Add it to the new group based on numa domain. */ in_pcbinslbgrouphash(inp, numa_domain); goto abort_with_hash_wlock; } } err = ENOENT; abort_with_hash_wlock: INP_HASH_WUNLOCK(pcbinfo); return (err); } /* Make sure it is safe to use hashinit(9) on CK_LIST. */ CTASSERT(sizeof(struct inpcbhead) == sizeof(LIST_HEAD(, inpcb))); /* * Initialize an inpcbinfo - a per-VNET instance of connections db. */ void in_pcbinfo_init(struct inpcbinfo *pcbinfo, struct inpcbstorage *pcbstor, u_int hash_nelements, u_int porthash_nelements) { mtx_init(&pcbinfo->ipi_lock, pcbstor->ips_infolock_name, NULL, MTX_DEF); mtx_init(&pcbinfo->ipi_hash_lock, pcbstor->ips_hashlock_name, NULL, MTX_DEF); #ifdef VIMAGE pcbinfo->ipi_vnet = curvnet; #endif CK_LIST_INIT(&pcbinfo->ipi_listhead); pcbinfo->ipi_count = 0; pcbinfo->ipi_hashbase = hashinit(hash_nelements, M_PCB, &pcbinfo->ipi_hashmask); porthash_nelements = imin(porthash_nelements, IPPORT_MAX + 1); pcbinfo->ipi_porthashbase = hashinit(porthash_nelements, M_PCB, &pcbinfo->ipi_porthashmask); pcbinfo->ipi_lbgrouphashbase = hashinit(porthash_nelements, M_PCB, &pcbinfo->ipi_lbgrouphashmask); pcbinfo->ipi_zone = pcbstor->ips_zone; pcbinfo->ipi_portzone = pcbstor->ips_portzone; pcbinfo->ipi_smr = uma_zone_get_smr(pcbinfo->ipi_zone); } /* * Destroy an inpcbinfo. */ void in_pcbinfo_destroy(struct inpcbinfo *pcbinfo) { KASSERT(pcbinfo->ipi_count == 0, ("%s: ipi_count = %u", __func__, pcbinfo->ipi_count)); hashdestroy(pcbinfo->ipi_hashbase, M_PCB, pcbinfo->ipi_hashmask); hashdestroy(pcbinfo->ipi_porthashbase, M_PCB, pcbinfo->ipi_porthashmask); hashdestroy(pcbinfo->ipi_lbgrouphashbase, M_PCB, pcbinfo->ipi_lbgrouphashmask); mtx_destroy(&pcbinfo->ipi_hash_lock); mtx_destroy(&pcbinfo->ipi_lock); } /* * Initialize a pcbstorage - per protocol zones to allocate inpcbs. */ static void inpcb_dtor(void *, int, void *); static void inpcb_fini(void *, int); void in_pcbstorage_init(void *arg) { struct inpcbstorage *pcbstor = arg; pcbstor->ips_zone = uma_zcreate(pcbstor->ips_zone_name, sizeof(struct inpcb), NULL, inpcb_dtor, pcbstor->ips_pcbinit, inpcb_fini, UMA_ALIGN_PTR, UMA_ZONE_SMR); pcbstor->ips_portzone = uma_zcreate(pcbstor->ips_portzone_name, sizeof(struct inpcbport), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); uma_zone_set_smr(pcbstor->ips_portzone, uma_zone_get_smr(pcbstor->ips_zone)); } /* * Destroy a pcbstorage - used by unloadable protocols. */ void in_pcbstorage_destroy(void *arg) { struct inpcbstorage *pcbstor = arg; uma_zdestroy(pcbstor->ips_zone); uma_zdestroy(pcbstor->ips_portzone); } /* * Allocate a PCB and associate it with the socket. * On success return with the PCB locked. */ int in_pcballoc(struct socket *so, struct inpcbinfo *pcbinfo) { struct inpcb *inp; #if defined(IPSEC) || defined(IPSEC_SUPPORT) || defined(MAC) int error; #endif inp = uma_zalloc_smr(pcbinfo->ipi_zone, M_NOWAIT); if (inp == NULL) return (ENOBUFS); bzero(&inp->inp_start_zero, inp_zero_size); #ifdef NUMA inp->inp_numa_domain = M_NODOM; #endif inp->inp_pcbinfo = pcbinfo; inp->inp_socket = so; inp->inp_cred = crhold(so->so_cred); inp->inp_inc.inc_fibnum = so->so_fibnum; #ifdef MAC error = mac_inpcb_init(inp, M_NOWAIT); if (error != 0) goto out; mac_inpcb_create(so, inp); #endif #if defined(IPSEC) || defined(IPSEC_SUPPORT) error = ipsec_init_pcbpolicy(inp); if (error != 0) { #ifdef MAC mac_inpcb_destroy(inp); #endif goto out; } #endif /*IPSEC*/ #ifdef INET6 if (INP_SOCKAF(so) == AF_INET6) { inp->inp_vflag |= INP_IPV6PROTO | INP_IPV6; if (V_ip6_v6only) inp->inp_flags |= IN6P_IPV6_V6ONLY; #ifdef INET else inp->inp_vflag |= INP_IPV4; #endif if (V_ip6_auto_flowlabel) inp->inp_flags |= IN6P_AUTOFLOWLABEL; inp->in6p_hops = -1; /* use kernel default */ } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET inp->inp_vflag |= INP_IPV4; #endif /* * Routes in inpcb's can cache L2 as well; they are guaranteed * to be cleaned up. */ inp->inp_route.ro_flags = RT_LLE_CACHE; #ifdef TCPHPTS /* * If using hpts lets drop a random number in so * not all new connections fall on the same CPU. */ inp->inp_hpts_cpu = hpts_random_cpu(inp); #endif refcount_init(&inp->inp_refcount, 1); /* Reference from socket. */ INP_WLOCK(inp); INP_INFO_WLOCK(pcbinfo); pcbinfo->ipi_count++; inp->inp_gencnt = ++pcbinfo->ipi_gencnt; CK_LIST_INSERT_HEAD(&pcbinfo->ipi_listhead, inp, inp_list); INP_INFO_WUNLOCK(pcbinfo); so->so_pcb = inp; return (0); #if defined(IPSEC) || defined(IPSEC_SUPPORT) || defined(MAC) out: uma_zfree_smr(pcbinfo->ipi_zone, inp); return (error); #endif } #ifdef INET int in_pcbbind(struct inpcb *inp, struct sockaddr *nam, struct ucred *cred) { int anonport, error; KASSERT(nam == NULL || nam->sa_family == AF_INET, ("%s: invalid address family for %p", __func__, nam)); KASSERT(nam == NULL || nam->sa_len == sizeof(struct sockaddr_in), ("%s: invalid address length for %p", __func__, nam)); INP_WLOCK_ASSERT(inp); INP_HASH_WLOCK_ASSERT(inp->inp_pcbinfo); if (inp->inp_lport != 0 || inp->inp_laddr.s_addr != INADDR_ANY) return (EINVAL); anonport = nam == NULL || ((struct sockaddr_in *)nam)->sin_port == 0; error = in_pcbbind_setup(inp, nam, &inp->inp_laddr.s_addr, &inp->inp_lport, cred); if (error) return (error); if (in_pcbinshash(inp) != 0) { inp->inp_laddr.s_addr = INADDR_ANY; inp->inp_lport = 0; return (EAGAIN); } if (anonport) inp->inp_flags |= INP_ANONPORT; return (0); } #endif #if defined(INET) || defined(INET6) /* * Assign a local port like in_pcb_lport(), but also used with connect() * and a foreign address and port. If fsa is non-NULL, choose a local port * that is unused with those, otherwise one that is completely unused. * lsa can be NULL for IPv6. */ int in_pcb_lport_dest(struct inpcb *inp, struct sockaddr *lsa, u_short *lportp, struct sockaddr *fsa, u_short fport, struct ucred *cred, int lookupflags) { struct inpcbinfo *pcbinfo; struct inpcb *tmpinp; unsigned short *lastport; int count, dorandom, error; u_short aux, first, last, lport; #ifdef INET struct in_addr laddr, faddr; #endif #ifdef INET6 struct in6_addr *laddr6, *faddr6; #endif pcbinfo = inp->inp_pcbinfo; /* * Because no actual state changes occur here, a global write lock on * the pcbinfo isn't required. */ INP_LOCK_ASSERT(inp); INP_HASH_LOCK_ASSERT(pcbinfo); if (inp->inp_flags & INP_HIGHPORT) { first = V_ipport_hifirstauto; /* sysctl */ last = V_ipport_hilastauto; lastport = &pcbinfo->ipi_lasthi; } else if (inp->inp_flags & INP_LOWPORT) { error = priv_check_cred(cred, PRIV_NETINET_RESERVEDPORT); if (error) return (error); first = V_ipport_lowfirstauto; /* 1023 */ last = V_ipport_lowlastauto; /* 600 */ lastport = &pcbinfo->ipi_lastlow; } else { first = V_ipport_firstauto; /* sysctl */ last = V_ipport_lastauto; lastport = &pcbinfo->ipi_lastport; } /* * For UDP(-Lite), use random port allocation as long as the user * allows it. For TCP (and as of yet unknown) connections, * use random port allocation only if the user allows it AND * ipport_tick() allows it. */ if (V_ipport_randomized && (!V_ipport_stoprandom || pcbinfo == &V_udbinfo || pcbinfo == &V_ulitecbinfo)) dorandom = 1; else dorandom = 0; /* * It makes no sense to do random port allocation if * we have the only port available. */ if (first == last) dorandom = 0; /* Make sure to not include UDP(-Lite) packets in the count. */ if (pcbinfo != &V_udbinfo && pcbinfo != &V_ulitecbinfo) V_ipport_tcpallocs++; /* * Instead of having two loops further down counting up or down * make sure that first is always <= last and go with only one * code path implementing all logic. */ if (first > last) { aux = first; first = last; last = aux; } #ifdef INET laddr.s_addr = INADDR_ANY; /* used by INET6+INET below too */ if ((inp->inp_vflag & (INP_IPV4|INP_IPV6)) == INP_IPV4) { if (lsa != NULL) laddr = ((struct sockaddr_in *)lsa)->sin_addr; if (fsa != NULL) faddr = ((struct sockaddr_in *)fsa)->sin_addr; } #endif #ifdef INET6 laddr6 = NULL; if ((inp->inp_vflag & INP_IPV6) != 0) { if (lsa != NULL) laddr6 = &((struct sockaddr_in6 *)lsa)->sin6_addr; if (fsa != NULL) faddr6 = &((struct sockaddr_in6 *)fsa)->sin6_addr; } #endif tmpinp = NULL; lport = *lportp; if (dorandom) *lastport = first + (arc4random() % (last - first)); count = last - first; do { if (count-- < 0) /* completely used? */ return (EADDRNOTAVAIL); ++*lastport; if (*lastport < first || *lastport > last) *lastport = first; lport = htons(*lastport); if (fsa != NULL) { #ifdef INET if (lsa->sa_family == AF_INET) { tmpinp = in_pcblookup_hash_locked(pcbinfo, faddr, fport, laddr, lport, lookupflags, NULL, M_NODOM); } #endif #ifdef INET6 if (lsa->sa_family == AF_INET6) { tmpinp = in6_pcblookup_hash_locked(pcbinfo, faddr6, fport, laddr6, lport, lookupflags, NULL, M_NODOM); } #endif } else { #ifdef INET6 if ((inp->inp_vflag & INP_IPV6) != 0) { tmpinp = in6_pcblookup_local(pcbinfo, &inp->in6p_laddr, lport, lookupflags, cred); #ifdef INET if (tmpinp == NULL && (inp->inp_vflag & INP_IPV4)) tmpinp = in_pcblookup_local(pcbinfo, laddr, lport, lookupflags, cred); #endif } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET tmpinp = in_pcblookup_local(pcbinfo, laddr, lport, lookupflags, cred); #endif } } while (tmpinp != NULL); *lportp = lport; return (0); } /* * Select a local port (number) to use. */ int in_pcb_lport(struct inpcb *inp, struct in_addr *laddrp, u_short *lportp, struct ucred *cred, int lookupflags) { struct sockaddr_in laddr; if (laddrp) { bzero(&laddr, sizeof(laddr)); laddr.sin_family = AF_INET; laddr.sin_addr = *laddrp; } return (in_pcb_lport_dest(inp, laddrp ? (struct sockaddr *) &laddr : NULL, lportp, NULL, 0, cred, lookupflags)); } /* * Return cached socket options. */ int inp_so_options(const struct inpcb *inp) { int so_options; so_options = 0; if ((inp->inp_flags2 & INP_REUSEPORT_LB) != 0) so_options |= SO_REUSEPORT_LB; if ((inp->inp_flags2 & INP_REUSEPORT) != 0) so_options |= SO_REUSEPORT; if ((inp->inp_flags2 & INP_REUSEADDR) != 0) so_options |= SO_REUSEADDR; return (so_options); } #endif /* INET || INET6 */ /* * Check if a new BINDMULTI socket is allowed to be created. * * ni points to the new inp. * oi points to the existing inp. * * This checks whether the existing inp also has BINDMULTI and * whether the credentials match. */ int in_pcbbind_check_bindmulti(const struct inpcb *ni, const struct inpcb *oi) { /* Check permissions match */ if ((ni->inp_flags2 & INP_BINDMULTI) && (ni->inp_cred->cr_uid != oi->inp_cred->cr_uid)) return (0); /* Check the existing inp has BINDMULTI set */ if ((ni->inp_flags2 & INP_BINDMULTI) && ((oi->inp_flags2 & INP_BINDMULTI) == 0)) return (0); /* * We're okay - either INP_BINDMULTI isn't set on ni, or * it is and it matches the checks. */ return (1); } #ifdef INET /* * Set up a bind operation on a PCB, performing port allocation * as required, but do not actually modify the PCB. Callers can * either complete the bind by setting inp_laddr/inp_lport and * calling in_pcbinshash(), or they can just use the resulting * port and address to authorise the sending of a once-off packet. * * On error, the values of *laddrp and *lportp are not changed. */ int in_pcbbind_setup(struct inpcb *inp, struct sockaddr *nam, in_addr_t *laddrp, u_short *lportp, struct ucred *cred) { struct socket *so = inp->inp_socket; struct sockaddr_in *sin; struct inpcbinfo *pcbinfo = inp->inp_pcbinfo; struct in_addr laddr; u_short lport = 0; int lookupflags = 0, reuseport = (so->so_options & SO_REUSEPORT); int error; /* * XXX: Maybe we could let SO_REUSEPORT_LB set SO_REUSEPORT bit here * so that we don't have to add to the (already messy) code below. */ int reuseport_lb = (so->so_options & SO_REUSEPORT_LB); /* * No state changes, so read locks are sufficient here. */ INP_LOCK_ASSERT(inp); INP_HASH_LOCK_ASSERT(pcbinfo); laddr.s_addr = *laddrp; if (nam != NULL && laddr.s_addr != INADDR_ANY) return (EINVAL); if ((so->so_options & (SO_REUSEADDR|SO_REUSEPORT|SO_REUSEPORT_LB)) == 0) lookupflags = INPLOOKUP_WILDCARD; if (nam == NULL) { if ((error = prison_local_ip4(cred, &laddr)) != 0) return (error); } else { sin = (struct sockaddr_in *)nam; KASSERT(sin->sin_family == AF_INET, ("%s: invalid family for address %p", __func__, sin)); KASSERT(sin->sin_len == sizeof(*sin), ("%s: invalid length for address %p", __func__, sin)); error = prison_local_ip4(cred, &sin->sin_addr); if (error) return (error); if (sin->sin_port != *lportp) { /* Don't allow the port to change. */ if (*lportp != 0) return (EINVAL); lport = sin->sin_port; } /* NB: lport is left as 0 if the port isn't being changed. */ if (IN_MULTICAST(ntohl(sin->sin_addr.s_addr))) { /* * Treat SO_REUSEADDR as SO_REUSEPORT for multicast; * allow complete duplication of binding if * SO_REUSEPORT is set, or if SO_REUSEADDR is set * and a multicast address is bound on both * new and duplicated sockets. */ if ((so->so_options & (SO_REUSEADDR|SO_REUSEPORT)) != 0) reuseport = SO_REUSEADDR|SO_REUSEPORT; /* * XXX: How to deal with SO_REUSEPORT_LB here? * Treat same as SO_REUSEPORT for now. */ if ((so->so_options & (SO_REUSEADDR|SO_REUSEPORT_LB)) != 0) reuseport_lb = SO_REUSEADDR|SO_REUSEPORT_LB; } else if (sin->sin_addr.s_addr != INADDR_ANY) { sin->sin_port = 0; /* yech... */ bzero(&sin->sin_zero, sizeof(sin->sin_zero)); /* * Is the address a local IP address? * If INP_BINDANY is set, then the socket may be bound * to any endpoint address, local or not. */ if ((inp->inp_flags & INP_BINDANY) == 0 && ifa_ifwithaddr_check((struct sockaddr *)sin) == 0) return (EADDRNOTAVAIL); } laddr = sin->sin_addr; if (lport) { struct inpcb *t; /* GROSS */ if (ntohs(lport) <= V_ipport_reservedhigh && ntohs(lport) >= V_ipport_reservedlow && priv_check_cred(cred, PRIV_NETINET_RESERVEDPORT)) return (EACCES); if (!IN_MULTICAST(ntohl(sin->sin_addr.s_addr)) && priv_check_cred(inp->inp_cred, PRIV_NETINET_REUSEPORT) != 0) { t = in_pcblookup_local(pcbinfo, sin->sin_addr, lport, INPLOOKUP_WILDCARD, cred); /* * XXX * This entire block sorely needs a rewrite. */ if (t && ((inp->inp_flags2 & INP_BINDMULTI) == 0) && - ((t->inp_flags & INP_TIMEWAIT) == 0) && (so->so_type != SOCK_STREAM || ntohl(t->inp_faddr.s_addr) == INADDR_ANY) && (ntohl(sin->sin_addr.s_addr) != INADDR_ANY || ntohl(t->inp_laddr.s_addr) != INADDR_ANY || (t->inp_flags2 & INP_REUSEPORT) || (t->inp_flags2 & INP_REUSEPORT_LB) == 0) && (inp->inp_cred->cr_uid != t->inp_cred->cr_uid)) return (EADDRINUSE); /* * If the socket is a BINDMULTI socket, then * the credentials need to match and the * original socket also has to have been bound * with BINDMULTI. */ if (t && (! in_pcbbind_check_bindmulti(inp, t))) return (EADDRINUSE); } t = in_pcblookup_local(pcbinfo, sin->sin_addr, lport, lookupflags, cred); if (t && ((inp->inp_flags2 & INP_BINDMULTI) == 0) && (reuseport & inp_so_options(t)) == 0 && (reuseport_lb & inp_so_options(t)) == 0) { #ifdef INET6 if (ntohl(sin->sin_addr.s_addr) != INADDR_ANY || ntohl(t->inp_laddr.s_addr) != INADDR_ANY || (inp->inp_vflag & INP_IPV6PROTO) == 0 || (t->inp_vflag & INP_IPV6PROTO) == 0) #endif return (EADDRINUSE); if (t && (! in_pcbbind_check_bindmulti(inp, t))) return (EADDRINUSE); } } } if (*lportp != 0) lport = *lportp; if (lport == 0) { error = in_pcb_lport(inp, &laddr, &lport, cred, lookupflags); if (error != 0) return (error); } *laddrp = laddr.s_addr; *lportp = lport; return (0); } /* * Connect from a socket to a specified address. * Both address and port must be specified in argument sin. * If don't have a local address for this socket yet, * then pick one. */ int in_pcbconnect(struct inpcb *inp, struct sockaddr *nam, struct ucred *cred, bool rehash) { u_short lport, fport; in_addr_t laddr, faddr; int anonport, error; INP_WLOCK_ASSERT(inp); INP_HASH_WLOCK_ASSERT(inp->inp_pcbinfo); lport = inp->inp_lport; laddr = inp->inp_laddr.s_addr; anonport = (lport == 0); error = in_pcbconnect_setup(inp, nam, &laddr, &lport, &faddr, &fport, NULL, cred); if (error) return (error); /* Do the initial binding of the local address if required. */ if (inp->inp_laddr.s_addr == INADDR_ANY && inp->inp_lport == 0) { KASSERT(rehash == true, ("Rehashing required for unbound inps")); inp->inp_lport = lport; inp->inp_laddr.s_addr = laddr; if (in_pcbinshash(inp) != 0) { inp->inp_laddr.s_addr = INADDR_ANY; inp->inp_lport = 0; return (EAGAIN); } } /* Commit the remaining changes. */ inp->inp_lport = lport; inp->inp_laddr.s_addr = laddr; inp->inp_faddr.s_addr = faddr; inp->inp_fport = fport; if (rehash) { in_pcbrehash(inp); } else { in_pcbinshash(inp); } if (anonport) inp->inp_flags |= INP_ANONPORT; return (0); } /* * Do proper source address selection on an unbound socket in case * of connect. Take jails into account as well. */ int in_pcbladdr(struct inpcb *inp, struct in_addr *faddr, struct in_addr *laddr, struct ucred *cred) { struct ifaddr *ifa; struct sockaddr *sa; struct sockaddr_in *sin, dst; struct nhop_object *nh; int error; NET_EPOCH_ASSERT(); KASSERT(laddr != NULL, ("%s: laddr NULL", __func__)); /* * Bypass source address selection and use the primary jail IP * if requested. */ if (cred != NULL && !prison_saddrsel_ip4(cred, laddr)) return (0); error = 0; nh = NULL; bzero(&dst, sizeof(dst)); sin = &dst; sin->sin_family = AF_INET; sin->sin_len = sizeof(struct sockaddr_in); sin->sin_addr.s_addr = faddr->s_addr; /* * If route is known our src addr is taken from the i/f, * else punt. * * Find out route to destination. */ if ((inp->inp_socket->so_options & SO_DONTROUTE) == 0) nh = fib4_lookup(inp->inp_inc.inc_fibnum, *faddr, 0, NHR_NONE, 0); /* * If we found a route, use the address corresponding to * the outgoing interface. * * Otherwise assume faddr is reachable on a directly connected * network and try to find a corresponding interface to take * the source address from. */ if (nh == NULL || nh->nh_ifp == NULL) { struct in_ifaddr *ia; struct ifnet *ifp; ia = ifatoia(ifa_ifwithdstaddr((struct sockaddr *)sin, inp->inp_socket->so_fibnum)); if (ia == NULL) { ia = ifatoia(ifa_ifwithnet((struct sockaddr *)sin, 0, inp->inp_socket->so_fibnum)); } if (ia == NULL) { error = ENETUNREACH; goto done; } if (cred == NULL || !prison_flag(cred, PR_IP4)) { laddr->s_addr = ia->ia_addr.sin_addr.s_addr; goto done; } ifp = ia->ia_ifp; ia = NULL; CK_STAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) { sa = ifa->ifa_addr; if (sa->sa_family != AF_INET) continue; sin = (struct sockaddr_in *)sa; if (prison_check_ip4(cred, &sin->sin_addr) == 0) { ia = (struct in_ifaddr *)ifa; break; } } if (ia != NULL) { laddr->s_addr = ia->ia_addr.sin_addr.s_addr; goto done; } /* 3. As a last resort return the 'default' jail address. */ error = prison_get_ip4(cred, laddr); goto done; } /* * If the outgoing interface on the route found is not * a loopback interface, use the address from that interface. * In case of jails do those three steps: * 1. check if the interface address belongs to the jail. If so use it. * 2. check if we have any address on the outgoing interface * belonging to this jail. If so use it. * 3. as a last resort return the 'default' jail address. */ if ((nh->nh_ifp->if_flags & IFF_LOOPBACK) == 0) { struct in_ifaddr *ia; struct ifnet *ifp; /* If not jailed, use the default returned. */ if (cred == NULL || !prison_flag(cred, PR_IP4)) { ia = (struct in_ifaddr *)nh->nh_ifa; laddr->s_addr = ia->ia_addr.sin_addr.s_addr; goto done; } /* Jailed. */ /* 1. Check if the iface address belongs to the jail. */ sin = (struct sockaddr_in *)nh->nh_ifa->ifa_addr; if (prison_check_ip4(cred, &sin->sin_addr) == 0) { ia = (struct in_ifaddr *)nh->nh_ifa; laddr->s_addr = ia->ia_addr.sin_addr.s_addr; goto done; } /* * 2. Check if we have any address on the outgoing interface * belonging to this jail. */ ia = NULL; ifp = nh->nh_ifp; CK_STAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) { sa = ifa->ifa_addr; if (sa->sa_family != AF_INET) continue; sin = (struct sockaddr_in *)sa; if (prison_check_ip4(cred, &sin->sin_addr) == 0) { ia = (struct in_ifaddr *)ifa; break; } } if (ia != NULL) { laddr->s_addr = ia->ia_addr.sin_addr.s_addr; goto done; } /* 3. As a last resort return the 'default' jail address. */ error = prison_get_ip4(cred, laddr); goto done; } /* * The outgoing interface is marked with 'loopback net', so a route * to ourselves is here. * Try to find the interface of the destination address and then * take the address from there. That interface is not necessarily * a loopback interface. * In case of jails, check that it is an address of the jail * and if we cannot find, fall back to the 'default' jail address. */ if ((nh->nh_ifp->if_flags & IFF_LOOPBACK) != 0) { struct in_ifaddr *ia; ia = ifatoia(ifa_ifwithdstaddr(sintosa(&dst), inp->inp_socket->so_fibnum)); if (ia == NULL) ia = ifatoia(ifa_ifwithnet(sintosa(&dst), 0, inp->inp_socket->so_fibnum)); if (ia == NULL) ia = ifatoia(ifa_ifwithaddr(sintosa(&dst))); if (cred == NULL || !prison_flag(cred, PR_IP4)) { if (ia == NULL) { error = ENETUNREACH; goto done; } laddr->s_addr = ia->ia_addr.sin_addr.s_addr; goto done; } /* Jailed. */ if (ia != NULL) { struct ifnet *ifp; ifp = ia->ia_ifp; ia = NULL; CK_STAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) { sa = ifa->ifa_addr; if (sa->sa_family != AF_INET) continue; sin = (struct sockaddr_in *)sa; if (prison_check_ip4(cred, &sin->sin_addr) == 0) { ia = (struct in_ifaddr *)ifa; break; } } if (ia != NULL) { laddr->s_addr = ia->ia_addr.sin_addr.s_addr; goto done; } } /* 3. As a last resort return the 'default' jail address. */ error = prison_get_ip4(cred, laddr); goto done; } done: return (error); } /* * Set up for a connect from a socket to the specified address. * On entry, *laddrp and *lportp should contain the current local * address and port for the PCB; these are updated to the values * that should be placed in inp_laddr and inp_lport to complete * the connect. * * On success, *faddrp and *fportp will be set to the remote address * and port. These are not updated in the error case. * * If the operation fails because the connection already exists, * *oinpp will be set to the PCB of that connection so that the * caller can decide to override it. In all other cases, *oinpp * is set to NULL. */ int in_pcbconnect_setup(struct inpcb *inp, struct sockaddr *nam, in_addr_t *laddrp, u_short *lportp, in_addr_t *faddrp, u_short *fportp, struct inpcb **oinpp, struct ucred *cred) { struct sockaddr_in *sin = (struct sockaddr_in *)nam; struct in_ifaddr *ia; struct inpcb *oinp; struct in_addr laddr, faddr; u_short lport, fport; int error; KASSERT(sin->sin_family == AF_INET, ("%s: invalid address family for %p", __func__, sin)); KASSERT(sin->sin_len == sizeof(*sin), ("%s: invalid address length for %p", __func__, sin)); /* * Because a global state change doesn't actually occur here, a read * lock is sufficient. */ NET_EPOCH_ASSERT(); INP_LOCK_ASSERT(inp); INP_HASH_LOCK_ASSERT(inp->inp_pcbinfo); if (oinpp != NULL) *oinpp = NULL; if (sin->sin_port == 0) return (EADDRNOTAVAIL); laddr.s_addr = *laddrp; lport = *lportp; faddr = sin->sin_addr; fport = sin->sin_port; #ifdef ROUTE_MPATH if (CALC_FLOWID_OUTBOUND) { uint32_t hash_val, hash_type; hash_val = fib4_calc_software_hash(laddr, faddr, 0, fport, inp->inp_socket->so_proto->pr_protocol, &hash_type); inp->inp_flowid = hash_val; inp->inp_flowtype = hash_type; } #endif if (!CK_STAILQ_EMPTY(&V_in_ifaddrhead)) { /* * If the destination address is INADDR_ANY, * use the primary local address. * If the supplied address is INADDR_BROADCAST, * and the primary interface supports broadcast, * choose the broadcast address for that interface. */ if (faddr.s_addr == INADDR_ANY) { faddr = IA_SIN(CK_STAILQ_FIRST(&V_in_ifaddrhead))->sin_addr; if (cred != NULL && (error = prison_get_ip4(cred, &faddr)) != 0) return (error); } else if (faddr.s_addr == (u_long)INADDR_BROADCAST) { if (CK_STAILQ_FIRST(&V_in_ifaddrhead)->ia_ifp->if_flags & IFF_BROADCAST) faddr = satosin(&CK_STAILQ_FIRST( &V_in_ifaddrhead)->ia_broadaddr)->sin_addr; } } if (laddr.s_addr == INADDR_ANY) { error = in_pcbladdr(inp, &faddr, &laddr, cred); /* * If the destination address is multicast and an outgoing * interface has been set as a multicast option, prefer the * address of that interface as our source address. */ if (IN_MULTICAST(ntohl(faddr.s_addr)) && inp->inp_moptions != NULL) { struct ip_moptions *imo; struct ifnet *ifp; imo = inp->inp_moptions; if (imo->imo_multicast_ifp != NULL) { ifp = imo->imo_multicast_ifp; CK_STAILQ_FOREACH(ia, &V_in_ifaddrhead, ia_link) { if ((ia->ia_ifp == ifp) && (cred == NULL || prison_check_ip4(cred, &ia->ia_addr.sin_addr) == 0)) break; } if (ia == NULL) error = EADDRNOTAVAIL; else { laddr = ia->ia_addr.sin_addr; error = 0; } } } if (error) return (error); } if (lport != 0) { oinp = in_pcblookup_hash_locked(inp->inp_pcbinfo, faddr, fport, laddr, lport, 0, NULL, M_NODOM); if (oinp != NULL) { if (oinpp != NULL) *oinpp = oinp; return (EADDRINUSE); } } else { struct sockaddr_in lsin, fsin; bzero(&lsin, sizeof(lsin)); bzero(&fsin, sizeof(fsin)); lsin.sin_family = AF_INET; lsin.sin_addr = laddr; fsin.sin_family = AF_INET; fsin.sin_addr = faddr; error = in_pcb_lport_dest(inp, (struct sockaddr *) &lsin, &lport, (struct sockaddr *)& fsin, fport, cred, INPLOOKUP_WILDCARD); if (error) return (error); } *laddrp = laddr.s_addr; *lportp = lport; *faddrp = faddr.s_addr; *fportp = fport; return (0); } void in_pcbdisconnect(struct inpcb *inp) { INP_WLOCK_ASSERT(inp); INP_HASH_WLOCK_ASSERT(inp->inp_pcbinfo); inp->inp_faddr.s_addr = INADDR_ANY; inp->inp_fport = 0; in_pcbrehash(inp); } #endif /* INET */ /* * in_pcbdetach() is responsibe for disassociating a socket from an inpcb. * For most protocols, this will be invoked immediately prior to calling * in_pcbfree(). However, with TCP the inpcb may significantly outlive the * socket, in which case in_pcbfree() is deferred. */ void in_pcbdetach(struct inpcb *inp) { KASSERT(inp->inp_socket != NULL, ("%s: inp_socket == NULL", __func__)); #ifdef RATELIMIT if (inp->inp_snd_tag != NULL) in_pcbdetach_txrtlmt(inp); #endif inp->inp_socket->so_pcb = NULL; inp->inp_socket = NULL; } /* * inpcb hash lookups are protected by SMR section. * * Once desired pcb has been found, switching from SMR section to a pcb * lock is performed with inp_smr_lock(). We can not use INP_(W|R)LOCK * here because SMR is a critical section. * In 99%+ cases inp_smr_lock() would obtain the lock immediately. */ static inline void inp_lock(struct inpcb *inp, const inp_lookup_t lock) { lock == INPLOOKUP_RLOCKPCB ? rw_rlock(&inp->inp_lock) : rw_wlock(&inp->inp_lock); } static inline void inp_unlock(struct inpcb *inp, const inp_lookup_t lock) { lock == INPLOOKUP_RLOCKPCB ? rw_runlock(&inp->inp_lock) : rw_wunlock(&inp->inp_lock); } static inline int inp_trylock(struct inpcb *inp, const inp_lookup_t lock) { return (lock == INPLOOKUP_RLOCKPCB ? rw_try_rlock(&inp->inp_lock) : rw_try_wlock(&inp->inp_lock)); } static inline bool in_pcbrele(struct inpcb *inp, const inp_lookup_t lock) { return (lock == INPLOOKUP_RLOCKPCB ? in_pcbrele_rlocked(inp) : in_pcbrele_wlocked(inp)); } bool inp_smr_lock(struct inpcb *inp, const inp_lookup_t lock) { MPASS(lock == INPLOOKUP_RLOCKPCB || lock == INPLOOKUP_WLOCKPCB); SMR_ASSERT_ENTERED(inp->inp_pcbinfo->ipi_smr); if (__predict_true(inp_trylock(inp, lock))) { if (__predict_false(inp->inp_flags & INP_FREED)) { smr_exit(inp->inp_pcbinfo->ipi_smr); inp_unlock(inp, lock); return (false); } smr_exit(inp->inp_pcbinfo->ipi_smr); return (true); } if (__predict_true(refcount_acquire_if_not_zero(&inp->inp_refcount))) { smr_exit(inp->inp_pcbinfo->ipi_smr); inp_lock(inp, lock); if (__predict_false(in_pcbrele(inp, lock))) return (false); /* * inp acquired through refcount & lock for sure didn't went * through uma_zfree(). However, it may have already went * through in_pcbfree() and has another reference, that * prevented its release by our in_pcbrele(). */ if (__predict_false(inp->inp_flags & INP_FREED)) { inp_unlock(inp, lock); return (false); } return (true); } else { smr_exit(inp->inp_pcbinfo->ipi_smr); return (false); } } /* * inp_next() - inpcb hash/list traversal iterator * * Requires initialized struct inpcb_iterator for context. * The structure can be initialized with INP_ITERATOR() or INP_ALL_ITERATOR(). * * - Iterator can have either write-lock or read-lock semantics, that can not * be changed later. * - Iterator can iterate either over all pcbs list (INP_ALL_LIST), or through * a single hash slot. Note: only rip_input() does the latter. * - Iterator may have optional bool matching function. The matching function * will be executed for each inpcb in the SMR context, so it can not acquire * locks and can safely access only immutable fields of inpcb. * * A fresh initialized iterator has NULL inpcb in its context and that * means that inp_next() call would return the very first inpcb on the list * locked with desired semantic. In all following calls the context pointer * shall hold the current inpcb pointer. The KPI user is not supposed to * unlock the current inpcb! Upon end of traversal inp_next() will return NULL * and write NULL to its context. After end of traversal an iterator can be * reused. * * List traversals have the following features/constraints: * - New entries won't be seen, as they are always added to the head of a list. * - Removed entries won't stop traversal as long as they are not added to * a different list. This is violated by in_pcbrehash(). */ #define II_LIST_FIRST(ipi, hash) \ (((hash) == INP_ALL_LIST) ? \ CK_LIST_FIRST(&(ipi)->ipi_listhead) : \ CK_LIST_FIRST(&(ipi)->ipi_hashbase[(hash)])) #define II_LIST_NEXT(inp, hash) \ (((hash) == INP_ALL_LIST) ? \ CK_LIST_NEXT((inp), inp_list) : \ CK_LIST_NEXT((inp), inp_hash)) #define II_LOCK_ASSERT(inp, lock) \ rw_assert(&(inp)->inp_lock, \ (lock) == INPLOOKUP_RLOCKPCB ? RA_RLOCKED : RA_WLOCKED ) struct inpcb * inp_next(struct inpcb_iterator *ii) { const struct inpcbinfo *ipi = ii->ipi; inp_match_t *match = ii->match; void *ctx = ii->ctx; inp_lookup_t lock = ii->lock; int hash = ii->hash; struct inpcb *inp; if (ii->inp == NULL) { /* First call. */ smr_enter(ipi->ipi_smr); /* This is unrolled CK_LIST_FOREACH(). */ for (inp = II_LIST_FIRST(ipi, hash); inp != NULL; inp = II_LIST_NEXT(inp, hash)) { if (match != NULL && (match)(inp, ctx) == false) continue; if (__predict_true(inp_smr_lock(inp, lock))) break; else { smr_enter(ipi->ipi_smr); MPASS(inp != II_LIST_FIRST(ipi, hash)); inp = II_LIST_FIRST(ipi, hash); if (inp == NULL) break; } } if (inp == NULL) smr_exit(ipi->ipi_smr); else ii->inp = inp; return (inp); } /* Not a first call. */ smr_enter(ipi->ipi_smr); restart: inp = ii->inp; II_LOCK_ASSERT(inp, lock); next: inp = II_LIST_NEXT(inp, hash); if (inp == NULL) { smr_exit(ipi->ipi_smr); goto found; } if (match != NULL && (match)(inp, ctx) == false) goto next; if (__predict_true(inp_trylock(inp, lock))) { if (__predict_false(inp->inp_flags & INP_FREED)) { /* * Entries are never inserted in middle of a list, thus * as long as we are in SMR, we can continue traversal. * Jump to 'restart' should yield in the same result, * but could produce unnecessary looping. Could this * looping be unbound? */ inp_unlock(inp, lock); goto next; } else { smr_exit(ipi->ipi_smr); goto found; } } /* * Can't obtain lock immediately, thus going hard. Once we exit the * SMR section we can no longer jump to 'next', and our only stable * anchoring point is ii->inp, which we keep locked for this case, so * we jump to 'restart'. */ if (__predict_true(refcount_acquire_if_not_zero(&inp->inp_refcount))) { smr_exit(ipi->ipi_smr); inp_lock(inp, lock); if (__predict_false(in_pcbrele(inp, lock))) { smr_enter(ipi->ipi_smr); goto restart; } /* * See comment in inp_smr_lock(). */ if (__predict_false(inp->inp_flags & INP_FREED)) { inp_unlock(inp, lock); smr_enter(ipi->ipi_smr); goto restart; } } else goto next; found: inp_unlock(ii->inp, lock); ii->inp = inp; return (ii->inp); } /* * in_pcbref() bumps the reference count on an inpcb in order to maintain * stability of an inpcb pointer despite the inpcb lock being released or * SMR section exited. * * To free a reference later in_pcbrele_(r|w)locked() must be performed. */ void in_pcbref(struct inpcb *inp) { u_int old __diagused; old = refcount_acquire(&inp->inp_refcount); KASSERT(old > 0, ("%s: refcount 0", __func__)); } /* * Drop a refcount on an inpcb elevated using in_pcbref(), potentially * freeing the pcb, if the reference was very last. */ bool in_pcbrele_rlocked(struct inpcb *inp) { INP_RLOCK_ASSERT(inp); if (refcount_release(&inp->inp_refcount) == 0) return (false); MPASS(inp->inp_flags & INP_FREED); MPASS(inp->inp_socket == NULL); MPASS(inp->inp_in_hpts == 0); INP_RUNLOCK(inp); uma_zfree_smr(inp->inp_pcbinfo->ipi_zone, inp); return (true); } bool in_pcbrele_wlocked(struct inpcb *inp) { INP_WLOCK_ASSERT(inp); if (refcount_release(&inp->inp_refcount) == 0) return (false); MPASS(inp->inp_flags & INP_FREED); MPASS(inp->inp_socket == NULL); MPASS(inp->inp_in_hpts == 0); INP_WUNLOCK(inp); uma_zfree_smr(inp->inp_pcbinfo->ipi_zone, inp); return (true); } /* * Unconditionally schedule an inpcb to be freed by decrementing its * reference count, which should occur only after the inpcb has been detached * from its socket. If another thread holds a temporary reference (acquired * using in_pcbref()) then the free is deferred until that reference is * released using in_pcbrele_(r|w)locked(), but the inpcb is still unlocked. * Almost all work, including removal from global lists, is done in this * context, where the pcbinfo lock is held. */ void in_pcbfree(struct inpcb *inp) { struct inpcbinfo *pcbinfo = inp->inp_pcbinfo; #ifdef INET struct ip_moptions *imo; #endif #ifdef INET6 struct ip6_moptions *im6o; #endif INP_WLOCK_ASSERT(inp); KASSERT(inp->inp_socket == NULL, ("%s: inp_socket != NULL", __func__)); KASSERT((inp->inp_flags & INP_FREED) == 0, ("%s: called twice for pcb %p", __func__, inp)); inp->inp_flags |= INP_FREED; INP_INFO_WLOCK(pcbinfo); inp->inp_gencnt = ++pcbinfo->ipi_gencnt; pcbinfo->ipi_count--; CK_LIST_REMOVE(inp, inp_list); INP_INFO_WUNLOCK(pcbinfo); if (inp->inp_flags & INP_INHASHLIST) { struct inpcbport *phd = inp->inp_phd; INP_HASH_WLOCK(pcbinfo); /* XXX: Only do if SO_REUSEPORT_LB set? */ in_pcbremlbgrouphash(inp); CK_LIST_REMOVE(inp, inp_hash); CK_LIST_REMOVE(inp, inp_portlist); if (CK_LIST_FIRST(&phd->phd_pcblist) == NULL) { CK_LIST_REMOVE(phd, phd_hash); uma_zfree_smr(pcbinfo->ipi_portzone, phd); } INP_HASH_WUNLOCK(pcbinfo); inp->inp_flags &= ~INP_INHASHLIST; } RO_INVALIDATE_CACHE(&inp->inp_route); #ifdef MAC mac_inpcb_destroy(inp); #endif #if defined(IPSEC) || defined(IPSEC_SUPPORT) if (inp->inp_sp != NULL) ipsec_delete_pcbpolicy(inp); #endif #ifdef INET if (inp->inp_options) (void)m_free(inp->inp_options); imo = inp->inp_moptions; #endif #ifdef INET6 if (inp->inp_vflag & INP_IPV6PROTO) { ip6_freepcbopts(inp->in6p_outputopts); im6o = inp->in6p_moptions; } else im6o = NULL; #endif if (__predict_false(in_pcbrele_wlocked(inp) == false)) { INP_WUNLOCK(inp); } #ifdef INET6 ip6_freemoptions(im6o); #endif #ifdef INET inp_freemoptions(imo); #endif /* Destruction is finalized in inpcb_dtor(). */ } static void inpcb_dtor(void *mem, int size, void *arg) { struct inpcb *inp = mem; crfree(inp->inp_cred); #ifdef INVARIANTS inp->inp_cred = NULL; #endif } /* * Different protocols initialize their inpcbs differently - giving * different name to the lock. But they all are disposed the same. */ static void inpcb_fini(void *mem, int size) { struct inpcb *inp = mem; INP_LOCK_DESTROY(inp); } /* * in_pcbdrop() removes an inpcb from hashed lists, releasing its address and * port reservation, and preventing it from being returned by inpcb lookups. * * It is used by TCP to mark an inpcb as unused and avoid future packet * delivery or event notification when a socket remains open but TCP has * closed. This might occur as a result of a shutdown()-initiated TCP close * or a RST on the wire, and allows the port binding to be reused while still * maintaining the invariant that so_pcb always points to a valid inpcb until * in_pcbdetach(). * * XXXRW: Possibly in_pcbdrop() should also prevent future notifications by * in_pcbnotifyall() and in_pcbpurgeif0()? */ void in_pcbdrop(struct inpcb *inp) { INP_WLOCK_ASSERT(inp); #ifdef INVARIANTS if (inp->inp_socket != NULL && inp->inp_ppcb != NULL) MPASS(inp->inp_refcount > 1); #endif /* * XXXRW: Possibly we should protect the setting of INP_DROPPED with * the hash lock...? */ inp->inp_flags |= INP_DROPPED; if (inp->inp_flags & INP_INHASHLIST) { struct inpcbport *phd = inp->inp_phd; INP_HASH_WLOCK(inp->inp_pcbinfo); in_pcbremlbgrouphash(inp); CK_LIST_REMOVE(inp, inp_hash); CK_LIST_REMOVE(inp, inp_portlist); if (CK_LIST_FIRST(&phd->phd_pcblist) == NULL) { CK_LIST_REMOVE(phd, phd_hash); uma_zfree_smr(inp->inp_pcbinfo->ipi_portzone, phd); } INP_HASH_WUNLOCK(inp->inp_pcbinfo); inp->inp_flags &= ~INP_INHASHLIST; } } #ifdef INET /* * Common routines to return the socket addresses associated with inpcbs. */ struct sockaddr * in_sockaddr(in_port_t port, struct in_addr *addr_p) { struct sockaddr_in *sin; sin = malloc(sizeof *sin, M_SONAME, M_WAITOK | M_ZERO); sin->sin_family = AF_INET; sin->sin_len = sizeof(*sin); sin->sin_addr = *addr_p; sin->sin_port = port; return (struct sockaddr *)sin; } int in_getsockaddr(struct socket *so, struct sockaddr **nam) { struct inpcb *inp; struct in_addr addr; in_port_t port; inp = sotoinpcb(so); KASSERT(inp != NULL, ("in_getsockaddr: inp == NULL")); INP_RLOCK(inp); port = inp->inp_lport; addr = inp->inp_laddr; INP_RUNLOCK(inp); *nam = in_sockaddr(port, &addr); return 0; } int in_getpeeraddr(struct socket *so, struct sockaddr **nam) { struct inpcb *inp; struct in_addr addr; in_port_t port; inp = sotoinpcb(so); KASSERT(inp != NULL, ("in_getpeeraddr: inp == NULL")); INP_RLOCK(inp); port = inp->inp_fport; addr = inp->inp_faddr; INP_RUNLOCK(inp); *nam = in_sockaddr(port, &addr); return 0; } void in_pcbnotifyall(struct inpcbinfo *pcbinfo, struct in_addr faddr, int errno, struct inpcb *(*notify)(struct inpcb *, int)) { struct inpcb *inp, *inp_temp; INP_INFO_WLOCK(pcbinfo); CK_LIST_FOREACH_SAFE(inp, &pcbinfo->ipi_listhead, inp_list, inp_temp) { INP_WLOCK(inp); #ifdef INET6 if ((inp->inp_vflag & INP_IPV4) == 0) { INP_WUNLOCK(inp); continue; } #endif if (inp->inp_faddr.s_addr != faddr.s_addr || inp->inp_socket == NULL) { INP_WUNLOCK(inp); continue; } if ((*notify)(inp, errno)) INP_WUNLOCK(inp); } INP_INFO_WUNLOCK(pcbinfo); } static bool inp_v4_multi_match(const struct inpcb *inp, void *v __unused) { if ((inp->inp_vflag & INP_IPV4) && inp->inp_moptions != NULL) return (true); else return (false); } void in_pcbpurgeif0(struct inpcbinfo *pcbinfo, struct ifnet *ifp) { struct inpcb_iterator inpi = INP_ITERATOR(pcbinfo, INPLOOKUP_WLOCKPCB, inp_v4_multi_match, NULL); struct inpcb *inp; struct in_multi *inm; struct in_mfilter *imf; struct ip_moptions *imo; IN_MULTI_LOCK_ASSERT(); while ((inp = inp_next(&inpi)) != NULL) { INP_WLOCK_ASSERT(inp); imo = inp->inp_moptions; /* * Unselect the outgoing interface if it is being * detached. */ if (imo->imo_multicast_ifp == ifp) imo->imo_multicast_ifp = NULL; /* * Drop multicast group membership if we joined * through the interface being detached. * * XXX This can all be deferred to an epoch_call */ restart: IP_MFILTER_FOREACH(imf, &imo->imo_head) { if ((inm = imf->imf_inm) == NULL) continue; if (inm->inm_ifp != ifp) continue; ip_mfilter_remove(&imo->imo_head, imf); in_leavegroup_locked(inm, NULL); ip_mfilter_free(imf); goto restart; } } } /* * Lookup a PCB based on the local address and port. Caller must hold the * hash lock. No inpcb locks or references are acquired. */ #define INP_LOOKUP_MAPPED_PCB_COST 3 struct inpcb * in_pcblookup_local(struct inpcbinfo *pcbinfo, struct in_addr laddr, u_short lport, int lookupflags, struct ucred *cred) { struct inpcb *inp; #ifdef INET6 int matchwild = 3 + INP_LOOKUP_MAPPED_PCB_COST; #else int matchwild = 3; #endif int wildcard; KASSERT((lookupflags & ~(INPLOOKUP_WILDCARD)) == 0, ("%s: invalid lookup flags %d", __func__, lookupflags)); INP_HASH_LOCK_ASSERT(pcbinfo); if ((lookupflags & INPLOOKUP_WILDCARD) == 0) { struct inpcbhead *head; /* * Look for an unconnected (wildcard foreign addr) PCB that * matches the local address and port we're looking for. */ head = &pcbinfo->ipi_hashbase[INP_PCBHASH_WILD(lport, pcbinfo->ipi_hashmask)]; CK_LIST_FOREACH(inp, head, inp_hash) { #ifdef INET6 /* XXX inp locking */ if ((inp->inp_vflag & INP_IPV4) == 0) continue; #endif if (inp->inp_faddr.s_addr == INADDR_ANY && inp->inp_laddr.s_addr == laddr.s_addr && inp->inp_lport == lport) { /* * Found? */ if (cred == NULL || prison_equal_ip4(cred->cr_prison, inp->inp_cred->cr_prison)) return (inp); } } /* * Not found. */ return (NULL); } else { struct inpcbporthead *porthash; struct inpcbport *phd; struct inpcb *match = NULL; /* * Best fit PCB lookup. * * First see if this local port is in use by looking on the * port hash list. */ porthash = &pcbinfo->ipi_porthashbase[INP_PCBPORTHASH(lport, pcbinfo->ipi_porthashmask)]; CK_LIST_FOREACH(phd, porthash, phd_hash) { if (phd->phd_port == lport) break; } if (phd != NULL) { /* * Port is in use by one or more PCBs. Look for best * fit. */ CK_LIST_FOREACH(inp, &phd->phd_pcblist, inp_portlist) { wildcard = 0; if (cred != NULL && !prison_equal_ip4(inp->inp_cred->cr_prison, cred->cr_prison)) continue; #ifdef INET6 /* XXX inp locking */ if ((inp->inp_vflag & INP_IPV4) == 0) continue; /* * We never select the PCB that has * INP_IPV6 flag and is bound to :: if * we have another PCB which is bound * to 0.0.0.0. If a PCB has the * INP_IPV6 flag, then we set its cost * higher than IPv4 only PCBs. * * Note that the case only happens * when a socket is bound to ::, under * the condition that the use of the * mapped address is allowed. */ if ((inp->inp_vflag & INP_IPV6) != 0) wildcard += INP_LOOKUP_MAPPED_PCB_COST; #endif if (inp->inp_faddr.s_addr != INADDR_ANY) wildcard++; if (inp->inp_laddr.s_addr != INADDR_ANY) { if (laddr.s_addr == INADDR_ANY) wildcard++; else if (inp->inp_laddr.s_addr != laddr.s_addr) continue; } else { if (laddr.s_addr != INADDR_ANY) wildcard++; } if (wildcard < matchwild) { match = inp; matchwild = wildcard; if (matchwild == 0) break; } } } return (match); } } #undef INP_LOOKUP_MAPPED_PCB_COST static struct inpcb * in_pcblookup_lbgroup(const struct inpcbinfo *pcbinfo, const struct in_addr *laddr, uint16_t lport, const struct in_addr *faddr, uint16_t fport, int lookupflags, int numa_domain) { struct inpcb *local_wild, *numa_wild; const struct inpcblbgrouphead *hdr; struct inpcblbgroup *grp; uint32_t idx; INP_HASH_LOCK_ASSERT(pcbinfo); hdr = &pcbinfo->ipi_lbgrouphashbase[ INP_PCBPORTHASH(lport, pcbinfo->ipi_lbgrouphashmask)]; /* * Order of socket selection: * 1. non-wild. * 2. wild (if lookupflags contains INPLOOKUP_WILDCARD). * * NOTE: * - Load balanced group does not contain jailed sockets * - Load balanced group does not contain IPv4 mapped INET6 wild sockets */ local_wild = NULL; numa_wild = NULL; CK_LIST_FOREACH(grp, hdr, il_list) { #ifdef INET6 if (!(grp->il_vflag & INP_IPV4)) continue; #endif if (grp->il_lport != lport) continue; idx = INP_PCBLBGROUP_PKTHASH(faddr, lport, fport) % grp->il_inpcnt; if (grp->il_laddr.s_addr == laddr->s_addr) { if (numa_domain == M_NODOM || grp->il_numa_domain == numa_domain) { return (grp->il_inp[idx]); } else { numa_wild = grp->il_inp[idx]; } } if (grp->il_laddr.s_addr == INADDR_ANY && (lookupflags & INPLOOKUP_WILDCARD) != 0 && (local_wild == NULL || numa_domain == M_NODOM || grp->il_numa_domain == numa_domain)) { local_wild = grp->il_inp[idx]; } } if (numa_wild != NULL) return (numa_wild); return (local_wild); } /* * Lookup PCB in hash list, using pcbinfo tables. This variation assumes * that the caller has either locked the hash list, which usually happens * for bind(2) operations, or is in SMR section, which happens when sorting * out incoming packets. */ static struct inpcb * in_pcblookup_hash_locked(struct inpcbinfo *pcbinfo, struct in_addr faddr, u_int fport_arg, struct in_addr laddr, u_int lport_arg, int lookupflags, struct ifnet *ifp, uint8_t numa_domain) { struct inpcbhead *head; struct inpcb *inp, *tmpinp; u_short fport = fport_arg, lport = lport_arg; KASSERT((lookupflags & ~(INPLOOKUP_WILDCARD)) == 0, ("%s: invalid lookup flags %d", __func__, lookupflags)); INP_HASH_LOCK_ASSERT(pcbinfo); /* * First look for an exact match. */ tmpinp = NULL; head = &pcbinfo->ipi_hashbase[INP_PCBHASH(&faddr, lport, fport, pcbinfo->ipi_hashmask)]; CK_LIST_FOREACH(inp, head, inp_hash) { #ifdef INET6 /* XXX inp locking */ if ((inp->inp_vflag & INP_IPV4) == 0) continue; #endif if (inp->inp_faddr.s_addr == faddr.s_addr && inp->inp_laddr.s_addr == laddr.s_addr && inp->inp_fport == fport && inp->inp_lport == lport) { /* * XXX We should be able to directly return * the inp here, without any checks. * Well unless both bound with SO_REUSEPORT? */ if (prison_flag(inp->inp_cred, PR_IP4)) return (inp); if (tmpinp == NULL) tmpinp = inp; } } if (tmpinp != NULL) return (tmpinp); /* * Then look in lb group (for wildcard match). */ if ((lookupflags & INPLOOKUP_WILDCARD) != 0) { inp = in_pcblookup_lbgroup(pcbinfo, &laddr, lport, &faddr, fport, lookupflags, numa_domain); if (inp != NULL) return (inp); } /* * Then look for a wildcard match, if requested. */ if ((lookupflags & INPLOOKUP_WILDCARD) != 0) { struct inpcb *local_wild = NULL, *local_exact = NULL; #ifdef INET6 struct inpcb *local_wild_mapped = NULL; #endif struct inpcb *jail_wild = NULL; int injail; /* * Order of socket selection - we always prefer jails. * 1. jailed, non-wild. * 2. jailed, wild. * 3. non-jailed, non-wild. * 4. non-jailed, wild. */ head = &pcbinfo->ipi_hashbase[INP_PCBHASH_WILD(lport, pcbinfo->ipi_hashmask)]; CK_LIST_FOREACH(inp, head, inp_hash) { #ifdef INET6 /* XXX inp locking */ if ((inp->inp_vflag & INP_IPV4) == 0) continue; #endif if (inp->inp_faddr.s_addr != INADDR_ANY || inp->inp_lport != lport) continue; injail = prison_flag(inp->inp_cred, PR_IP4); if (injail) { if (prison_check_ip4_locked( inp->inp_cred->cr_prison, &laddr) != 0) continue; } else { if (local_exact != NULL) continue; } if (inp->inp_laddr.s_addr == laddr.s_addr) { if (injail) return (inp); else local_exact = inp; } else if (inp->inp_laddr.s_addr == INADDR_ANY) { #ifdef INET6 /* XXX inp locking, NULL check */ if (inp->inp_vflag & INP_IPV6PROTO) local_wild_mapped = inp; else #endif if (injail) jail_wild = inp; else local_wild = inp; } } /* LIST_FOREACH */ if (jail_wild != NULL) return (jail_wild); if (local_exact != NULL) return (local_exact); if (local_wild != NULL) return (local_wild); #ifdef INET6 if (local_wild_mapped != NULL) return (local_wild_mapped); #endif } /* if ((lookupflags & INPLOOKUP_WILDCARD) != 0) */ return (NULL); } /* * Lookup PCB in hash list, using pcbinfo tables. This variation locks the * hash list lock, and will return the inpcb locked (i.e., requires * INPLOOKUP_LOCKPCB). */ static struct inpcb * in_pcblookup_hash(struct inpcbinfo *pcbinfo, struct in_addr faddr, u_int fport, struct in_addr laddr, u_int lport, int lookupflags, struct ifnet *ifp, uint8_t numa_domain) { struct inpcb *inp; smr_enter(pcbinfo->ipi_smr); inp = in_pcblookup_hash_locked(pcbinfo, faddr, fport, laddr, lport, lookupflags & INPLOOKUP_WILDCARD, ifp, numa_domain); if (inp != NULL) { if (__predict_false(inp_smr_lock(inp, (lookupflags & INPLOOKUP_LOCKMASK)) == false)) inp = NULL; } else smr_exit(pcbinfo->ipi_smr); return (inp); } /* * Public inpcb lookup routines, accepting a 4-tuple, and optionally, an mbuf * from which a pre-calculated hash value may be extracted. */ struct inpcb * in_pcblookup(struct inpcbinfo *pcbinfo, struct in_addr faddr, u_int fport, struct in_addr laddr, u_int lport, int lookupflags, struct ifnet *ifp) { KASSERT((lookupflags & ~INPLOOKUP_MASK) == 0, ("%s: invalid lookup flags %d", __func__, lookupflags)); KASSERT((lookupflags & (INPLOOKUP_RLOCKPCB | INPLOOKUP_WLOCKPCB)) != 0, ("%s: LOCKPCB not set", __func__)); return (in_pcblookup_hash(pcbinfo, faddr, fport, laddr, lport, lookupflags, ifp, M_NODOM)); } struct inpcb * in_pcblookup_mbuf(struct inpcbinfo *pcbinfo, struct in_addr faddr, u_int fport, struct in_addr laddr, u_int lport, int lookupflags, struct ifnet *ifp, struct mbuf *m) { KASSERT((lookupflags & ~INPLOOKUP_MASK) == 0, ("%s: invalid lookup flags %d", __func__, lookupflags)); KASSERT((lookupflags & (INPLOOKUP_RLOCKPCB | INPLOOKUP_WLOCKPCB)) != 0, ("%s: LOCKPCB not set", __func__)); return (in_pcblookup_hash(pcbinfo, faddr, fport, laddr, lport, lookupflags, ifp, m->m_pkthdr.numa_domain)); } #endif /* INET */ /* * Insert PCB onto various hash lists. */ int in_pcbinshash(struct inpcb *inp) { struct inpcbhead *pcbhash; struct inpcbporthead *pcbporthash; struct inpcbinfo *pcbinfo = inp->inp_pcbinfo; struct inpcbport *phd; int so_options; INP_WLOCK_ASSERT(inp); INP_HASH_WLOCK_ASSERT(pcbinfo); KASSERT((inp->inp_flags & INP_INHASHLIST) == 0, ("in_pcbinshash: INP_INHASHLIST")); #ifdef INET6 if (inp->inp_vflag & INP_IPV6) pcbhash = &pcbinfo->ipi_hashbase[INP6_PCBHASH(&inp->in6p_faddr, inp->inp_lport, inp->inp_fport, pcbinfo->ipi_hashmask)]; else #endif pcbhash = &pcbinfo->ipi_hashbase[INP_PCBHASH(&inp->inp_faddr, inp->inp_lport, inp->inp_fport, pcbinfo->ipi_hashmask)]; pcbporthash = &pcbinfo->ipi_porthashbase[ INP_PCBPORTHASH(inp->inp_lport, pcbinfo->ipi_porthashmask)]; /* * Add entry to load balance group. * Only do this if SO_REUSEPORT_LB is set. */ so_options = inp_so_options(inp); if (so_options & SO_REUSEPORT_LB) { int ret = in_pcbinslbgrouphash(inp, M_NODOM); if (ret) { /* pcb lb group malloc fail (ret=ENOBUFS). */ return (ret); } } /* * Go through port list and look for a head for this lport. */ CK_LIST_FOREACH(phd, pcbporthash, phd_hash) { if (phd->phd_port == inp->inp_lport) break; } /* * If none exists, malloc one and tack it on. */ if (phd == NULL) { phd = uma_zalloc_smr(pcbinfo->ipi_portzone, M_NOWAIT); if (phd == NULL) { return (ENOBUFS); /* XXX */ } phd->phd_port = inp->inp_lport; CK_LIST_INIT(&phd->phd_pcblist); CK_LIST_INSERT_HEAD(pcbporthash, phd, phd_hash); } inp->inp_phd = phd; CK_LIST_INSERT_HEAD(&phd->phd_pcblist, inp, inp_portlist); CK_LIST_INSERT_HEAD(pcbhash, inp, inp_hash); inp->inp_flags |= INP_INHASHLIST; return (0); } /* * Move PCB to the proper hash bucket when { faddr, fport } have been * changed. NOTE: This does not handle the case of the lport changing (the * hashed port list would have to be updated as well), so the lport must * not change after in_pcbinshash() has been called. * * XXXGL: a race between this function and SMR-protected hash iterator * will lead to iterator traversing a possibly wrong hash list. However, * this race should have been here since change from rwlock to epoch. */ void in_pcbrehash(struct inpcb *inp) { struct inpcbinfo *pcbinfo = inp->inp_pcbinfo; struct inpcbhead *head; INP_WLOCK_ASSERT(inp); INP_HASH_WLOCK_ASSERT(pcbinfo); KASSERT(inp->inp_flags & INP_INHASHLIST, ("in_pcbrehash: !INP_INHASHLIST")); #ifdef INET6 if (inp->inp_vflag & INP_IPV6) head = &pcbinfo->ipi_hashbase[INP6_PCBHASH(&inp->in6p_faddr, inp->inp_lport, inp->inp_fport, pcbinfo->ipi_hashmask)]; else #endif head = &pcbinfo->ipi_hashbase[INP_PCBHASH(&inp->inp_faddr, inp->inp_lport, inp->inp_fport, pcbinfo->ipi_hashmask)]; CK_LIST_REMOVE(inp, inp_hash); CK_LIST_INSERT_HEAD(head, inp, inp_hash); } /* * Check for alternatives when higher level complains * about service problems. For now, invalidate cached * routing information. If the route was created dynamically * (by a redirect), time to try a default gateway again. */ void in_losing(struct inpcb *inp) { RO_INVALIDATE_CACHE(&inp->inp_route); return; } /* * A set label operation has occurred at the socket layer, propagate the * label change into the in_pcb for the socket. */ void in_pcbsosetlabel(struct socket *so) { #ifdef MAC struct inpcb *inp; inp = sotoinpcb(so); KASSERT(inp != NULL, ("in_pcbsosetlabel: so->so_pcb == NULL")); INP_WLOCK(inp); SOCK_LOCK(so); mac_inpcb_sosetlabel(so, inp); SOCK_UNLOCK(so); INP_WUNLOCK(inp); #endif } /* * ipport_tick runs once per second, determining if random port allocation * should be continued. If more than ipport_randomcps ports have been * allocated in the last second, then we return to sequential port * allocation. We return to random allocation only once we drop below * ipport_randomcps for at least ipport_randomtime seconds. */ static void ipport_tick(void *xtp) { VNET_ITERATOR_DECL(vnet_iter); VNET_LIST_RLOCK_NOSLEEP(); VNET_FOREACH(vnet_iter) { CURVNET_SET(vnet_iter); /* XXX appease INVARIANTS here */ if (V_ipport_tcpallocs - V_ipport_tcplastcount <= V_ipport_randomcps) { if (V_ipport_stoprandom > 0) V_ipport_stoprandom--; } else V_ipport_stoprandom = V_ipport_randomtime; V_ipport_tcplastcount = V_ipport_tcpallocs; CURVNET_RESTORE(); } VNET_LIST_RUNLOCK_NOSLEEP(); callout_reset(&ipport_tick_callout, hz, ipport_tick, NULL); } static void ip_fini(void *xtp) { callout_stop(&ipport_tick_callout); } /* * The ipport_callout should start running at about the time we attach the * inet or inet6 domains. */ static void ipport_tick_init(const void *unused __unused) { /* Start ipport_tick. */ callout_init(&ipport_tick_callout, 1); callout_reset(&ipport_tick_callout, 1, ipport_tick, NULL); EVENTHANDLER_REGISTER(shutdown_pre_sync, ip_fini, NULL, SHUTDOWN_PRI_DEFAULT); } SYSINIT(ipport_tick_init, SI_SUB_PROTO_DOMAIN, SI_ORDER_MIDDLE, ipport_tick_init, NULL); void inp_wlock(struct inpcb *inp) { INP_WLOCK(inp); } void inp_wunlock(struct inpcb *inp) { INP_WUNLOCK(inp); } void inp_rlock(struct inpcb *inp) { INP_RLOCK(inp); } void inp_runlock(struct inpcb *inp) { INP_RUNLOCK(inp); } #ifdef INVARIANT_SUPPORT void inp_lock_assert(struct inpcb *inp) { INP_WLOCK_ASSERT(inp); } void inp_unlock_assert(struct inpcb *inp) { INP_UNLOCK_ASSERT(inp); } #endif void inp_apply_all(void (*func)(struct inpcb *, void *), void *arg) { struct inpcb_iterator inpi = INP_ALL_ITERATOR(&V_tcbinfo, INPLOOKUP_WLOCKPCB); struct inpcb *inp; while ((inp = inp_next(&inpi)) != NULL) func(inp, arg); } struct socket * inp_inpcbtosocket(struct inpcb *inp) { INP_WLOCK_ASSERT(inp); return (inp->inp_socket); } struct tcpcb * inp_inpcbtotcpcb(struct inpcb *inp) { INP_WLOCK_ASSERT(inp); return ((struct tcpcb *)inp->inp_ppcb); } int inp_ip_tos_get(const struct inpcb *inp) { return (inp->inp_ip_tos); } void inp_ip_tos_set(struct inpcb *inp, int val) { inp->inp_ip_tos = val; } void inp_4tuple_get(struct inpcb *inp, uint32_t *laddr, uint16_t *lp, uint32_t *faddr, uint16_t *fp) { INP_LOCK_ASSERT(inp); *laddr = inp->inp_laddr.s_addr; *faddr = inp->inp_faddr.s_addr; *lp = inp->inp_lport; *fp = inp->inp_fport; } struct inpcb * so_sotoinpcb(struct socket *so) { return (sotoinpcb(so)); } struct tcpcb * so_sototcpcb(struct socket *so) { return (sototcpcb(so)); } /* * Create an external-format (``xinpcb'') structure using the information in * the kernel-format in_pcb structure pointed to by inp. This is done to * reduce the spew of irrelevant information over this interface, to isolate * user code from changes in the kernel structure, and potentially to provide * information-hiding if we decide that some of this information should be * hidden from users. */ void in_pcbtoxinpcb(const struct inpcb *inp, struct xinpcb *xi) { bzero(xi, sizeof(*xi)); xi->xi_len = sizeof(struct xinpcb); if (inp->inp_socket) sotoxsocket(inp->inp_socket, &xi->xi_socket); bcopy(&inp->inp_inc, &xi->inp_inc, sizeof(struct in_conninfo)); xi->inp_gencnt = inp->inp_gencnt; xi->inp_ppcb = (uintptr_t)inp->inp_ppcb; xi->inp_flow = inp->inp_flow; xi->inp_flowid = inp->inp_flowid; xi->inp_flowtype = inp->inp_flowtype; xi->inp_flags = inp->inp_flags; xi->inp_flags2 = inp->inp_flags2; xi->inp_rss_listen_bucket = inp->inp_rss_listen_bucket; xi->in6p_cksum = inp->in6p_cksum; xi->in6p_hops = inp->in6p_hops; xi->inp_ip_tos = inp->inp_ip_tos; xi->inp_vflag = inp->inp_vflag; xi->inp_ip_ttl = inp->inp_ip_ttl; xi->inp_ip_p = inp->inp_ip_p; xi->inp_ip_minttl = inp->inp_ip_minttl; } int sysctl_setsockopt(SYSCTL_HANDLER_ARGS, struct inpcbinfo *pcbinfo, int (*ctloutput_set)(struct inpcb *, struct sockopt *)) { struct sockopt sopt; struct inpcb_iterator inpi = INP_ALL_ITERATOR(pcbinfo, INPLOOKUP_WLOCKPCB); struct inpcb *inp; struct sockopt_parameters *params; struct socket *so; int error; char buf[1024]; if (req->oldptr != NULL || req->oldlen != 0) return (EINVAL); if (req->newptr == NULL) return (EPERM); if (req->newlen > sizeof(buf)) return (ENOMEM); error = SYSCTL_IN(req, buf, req->newlen); if (error != 0) return (error); if (req->newlen < sizeof(struct sockopt_parameters)) return (EINVAL); params = (struct sockopt_parameters *)buf; sopt.sopt_level = params->sop_level; sopt.sopt_name = params->sop_optname; sopt.sopt_dir = SOPT_SET; sopt.sopt_val = params->sop_optval; sopt.sopt_valsize = req->newlen - sizeof(struct sockopt_parameters); sopt.sopt_td = NULL; #ifdef INET6 if (params->sop_inc.inc_flags & INC_ISIPV6) { if (IN6_IS_SCOPE_LINKLOCAL(¶ms->sop_inc.inc6_laddr)) params->sop_inc.inc6_laddr.s6_addr16[1] = htons(params->sop_inc.inc6_zoneid & 0xffff); if (IN6_IS_SCOPE_LINKLOCAL(¶ms->sop_inc.inc6_faddr)) params->sop_inc.inc6_faddr.s6_addr16[1] = htons(params->sop_inc.inc6_zoneid & 0xffff); } #endif if (params->sop_inc.inc_lport != htons(0)) { if (params->sop_inc.inc_fport == htons(0)) inpi.hash = INP_PCBHASH_WILD(params->sop_inc.inc_lport, pcbinfo->ipi_hashmask); else #ifdef INET6 if (params->sop_inc.inc_flags & INC_ISIPV6) inpi.hash = INP6_PCBHASH( ¶ms->sop_inc.inc6_faddr, params->sop_inc.inc_lport, params->sop_inc.inc_fport, pcbinfo->ipi_hashmask); else #endif inpi.hash = INP_PCBHASH( ¶ms->sop_inc.inc_faddr, params->sop_inc.inc_lport, params->sop_inc.inc_fport, pcbinfo->ipi_hashmask); } while ((inp = inp_next(&inpi)) != NULL) if (inp->inp_gencnt == params->sop_id) { - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { INP_WUNLOCK(inp); return (ECONNRESET); } so = inp->inp_socket; KASSERT(so != NULL, ("inp_socket == NULL")); soref(so); error = (*ctloutput_set)(inp, &sopt); sorele(so); break; } if (inp == NULL) error = ESRCH; return (error); } #ifdef DDB static void db_print_indent(int indent) { int i; for (i = 0; i < indent; i++) db_printf(" "); } static void db_print_inconninfo(struct in_conninfo *inc, const char *name, int indent) { char faddr_str[48], laddr_str[48]; db_print_indent(indent); db_printf("%s at %p\n", name, inc); indent += 2; #ifdef INET6 if (inc->inc_flags & INC_ISIPV6) { /* IPv6. */ ip6_sprintf(laddr_str, &inc->inc6_laddr); ip6_sprintf(faddr_str, &inc->inc6_faddr); } else #endif { /* IPv4. */ inet_ntoa_r(inc->inc_laddr, laddr_str); inet_ntoa_r(inc->inc_faddr, faddr_str); } db_print_indent(indent); db_printf("inc_laddr %s inc_lport %u\n", laddr_str, ntohs(inc->inc_lport)); db_print_indent(indent); db_printf("inc_faddr %s inc_fport %u\n", faddr_str, ntohs(inc->inc_fport)); } static void db_print_inpflags(int inp_flags) { int comma; comma = 0; if (inp_flags & INP_RECVOPTS) { db_printf("%sINP_RECVOPTS", comma ? ", " : ""); comma = 1; } if (inp_flags & INP_RECVRETOPTS) { db_printf("%sINP_RECVRETOPTS", comma ? ", " : ""); comma = 1; } if (inp_flags & INP_RECVDSTADDR) { db_printf("%sINP_RECVDSTADDR", comma ? ", " : ""); comma = 1; } if (inp_flags & INP_ORIGDSTADDR) { db_printf("%sINP_ORIGDSTADDR", comma ? ", " : ""); comma = 1; } if (inp_flags & INP_HDRINCL) { db_printf("%sINP_HDRINCL", comma ? ", " : ""); comma = 1; } if (inp_flags & INP_HIGHPORT) { db_printf("%sINP_HIGHPORT", comma ? ", " : ""); comma = 1; } if (inp_flags & INP_LOWPORT) { db_printf("%sINP_LOWPORT", comma ? ", " : ""); comma = 1; } if (inp_flags & INP_ANONPORT) { db_printf("%sINP_ANONPORT", comma ? ", " : ""); comma = 1; } if (inp_flags & INP_RECVIF) { db_printf("%sINP_RECVIF", comma ? ", " : ""); comma = 1; } if (inp_flags & INP_MTUDISC) { db_printf("%sINP_MTUDISC", comma ? ", " : ""); comma = 1; } if (inp_flags & INP_RECVTTL) { db_printf("%sINP_RECVTTL", comma ? ", " : ""); comma = 1; } if (inp_flags & INP_DONTFRAG) { db_printf("%sINP_DONTFRAG", comma ? ", " : ""); comma = 1; } if (inp_flags & INP_RECVTOS) { db_printf("%sINP_RECVTOS", comma ? ", " : ""); comma = 1; } if (inp_flags & IN6P_IPV6_V6ONLY) { db_printf("%sIN6P_IPV6_V6ONLY", comma ? ", " : ""); comma = 1; } if (inp_flags & IN6P_PKTINFO) { db_printf("%sIN6P_PKTINFO", comma ? ", " : ""); comma = 1; } if (inp_flags & IN6P_HOPLIMIT) { db_printf("%sIN6P_HOPLIMIT", comma ? ", " : ""); comma = 1; } if (inp_flags & IN6P_HOPOPTS) { db_printf("%sIN6P_HOPOPTS", comma ? ", " : ""); comma = 1; } if (inp_flags & IN6P_DSTOPTS) { db_printf("%sIN6P_DSTOPTS", comma ? ", " : ""); comma = 1; } if (inp_flags & IN6P_RTHDR) { db_printf("%sIN6P_RTHDR", comma ? ", " : ""); comma = 1; } if (inp_flags & IN6P_RTHDRDSTOPTS) { db_printf("%sIN6P_RTHDRDSTOPTS", comma ? ", " : ""); comma = 1; } if (inp_flags & IN6P_TCLASS) { db_printf("%sIN6P_TCLASS", comma ? ", " : ""); comma = 1; } if (inp_flags & IN6P_AUTOFLOWLABEL) { db_printf("%sIN6P_AUTOFLOWLABEL", comma ? ", " : ""); comma = 1; } - if (inp_flags & INP_TIMEWAIT) { - db_printf("%sINP_TIMEWAIT", comma ? ", " : ""); - comma = 1; - } if (inp_flags & INP_ONESBCAST) { db_printf("%sINP_ONESBCAST", comma ? ", " : ""); comma = 1; } if (inp_flags & INP_DROPPED) { db_printf("%sINP_DROPPED", comma ? ", " : ""); comma = 1; } if (inp_flags & INP_SOCKREF) { db_printf("%sINP_SOCKREF", comma ? ", " : ""); comma = 1; } if (inp_flags & IN6P_RFC2292) { db_printf("%sIN6P_RFC2292", comma ? ", " : ""); comma = 1; } if (inp_flags & IN6P_MTU) { db_printf("IN6P_MTU%s", comma ? ", " : ""); comma = 1; } } static void db_print_inpvflag(u_char inp_vflag) { int comma; comma = 0; if (inp_vflag & INP_IPV4) { db_printf("%sINP_IPV4", comma ? ", " : ""); comma = 1; } if (inp_vflag & INP_IPV6) { db_printf("%sINP_IPV6", comma ? ", " : ""); comma = 1; } if (inp_vflag & INP_IPV6PROTO) { db_printf("%sINP_IPV6PROTO", comma ? ", " : ""); comma = 1; } } static void db_print_inpcb(struct inpcb *inp, const char *name, int indent) { db_print_indent(indent); db_printf("%s at %p\n", name, inp); indent += 2; db_print_indent(indent); db_printf("inp_flow: 0x%x\n", inp->inp_flow); db_print_inconninfo(&inp->inp_inc, "inp_conninfo", indent); db_print_indent(indent); db_printf("inp_ppcb: %p inp_pcbinfo: %p inp_socket: %p\n", inp->inp_ppcb, inp->inp_pcbinfo, inp->inp_socket); db_print_indent(indent); db_printf("inp_label: %p inp_flags: 0x%x (", inp->inp_label, inp->inp_flags); db_print_inpflags(inp->inp_flags); db_printf(")\n"); db_print_indent(indent); db_printf("inp_sp: %p inp_vflag: 0x%x (", inp->inp_sp, inp->inp_vflag); db_print_inpvflag(inp->inp_vflag); db_printf(")\n"); db_print_indent(indent); db_printf("inp_ip_ttl: %d inp_ip_p: %d inp_ip_minttl: %d\n", inp->inp_ip_ttl, inp->inp_ip_p, inp->inp_ip_minttl); db_print_indent(indent); #ifdef INET6 if (inp->inp_vflag & INP_IPV6) { db_printf("in6p_options: %p in6p_outputopts: %p " "in6p_moptions: %p\n", inp->in6p_options, inp->in6p_outputopts, inp->in6p_moptions); db_printf("in6p_icmp6filt: %p in6p_cksum %d " "in6p_hops %u\n", inp->in6p_icmp6filt, inp->in6p_cksum, inp->in6p_hops); } else #endif { db_printf("inp_ip_tos: %d inp_ip_options: %p " "inp_ip_moptions: %p\n", inp->inp_ip_tos, inp->inp_options, inp->inp_moptions); } db_print_indent(indent); db_printf("inp_phd: %p inp_gencnt: %ju\n", inp->inp_phd, (uintmax_t)inp->inp_gencnt); } DB_SHOW_COMMAND(inpcb, db_show_inpcb) { struct inpcb *inp; if (!have_addr) { db_printf("usage: show inpcb \n"); return; } inp = (struct inpcb *)addr; db_print_inpcb(inp, "inpcb", 0); } #endif /* DDB */ #ifdef RATELIMIT /* * Modify TX rate limit based on the existing "inp->inp_snd_tag", * if any. */ int in_pcbmodify_txrtlmt(struct inpcb *inp, uint32_t max_pacing_rate) { union if_snd_tag_modify_params params = { .rate_limit.max_rate = max_pacing_rate, .rate_limit.flags = M_NOWAIT, }; struct m_snd_tag *mst; int error; mst = inp->inp_snd_tag; if (mst == NULL) return (EINVAL); if (mst->sw->snd_tag_modify == NULL) { error = EOPNOTSUPP; } else { error = mst->sw->snd_tag_modify(mst, ¶ms); } return (error); } /* * Query existing TX rate limit based on the existing * "inp->inp_snd_tag", if any. */ int in_pcbquery_txrtlmt(struct inpcb *inp, uint32_t *p_max_pacing_rate) { union if_snd_tag_query_params params = { }; struct m_snd_tag *mst; int error; mst = inp->inp_snd_tag; if (mst == NULL) return (EINVAL); if (mst->sw->snd_tag_query == NULL) { error = EOPNOTSUPP; } else { error = mst->sw->snd_tag_query(mst, ¶ms); if (error == 0 && p_max_pacing_rate != NULL) *p_max_pacing_rate = params.rate_limit.max_rate; } return (error); } /* * Query existing TX queue level based on the existing * "inp->inp_snd_tag", if any. */ int in_pcbquery_txrlevel(struct inpcb *inp, uint32_t *p_txqueue_level) { union if_snd_tag_query_params params = { }; struct m_snd_tag *mst; int error; mst = inp->inp_snd_tag; if (mst == NULL) return (EINVAL); if (mst->sw->snd_tag_query == NULL) return (EOPNOTSUPP); error = mst->sw->snd_tag_query(mst, ¶ms); if (error == 0 && p_txqueue_level != NULL) *p_txqueue_level = params.rate_limit.queue_level; return (error); } /* * Allocate a new TX rate limit send tag from the network interface * given by the "ifp" argument and save it in "inp->inp_snd_tag": */ int in_pcbattach_txrtlmt(struct inpcb *inp, struct ifnet *ifp, uint32_t flowtype, uint32_t flowid, uint32_t max_pacing_rate, struct m_snd_tag **st) { union if_snd_tag_alloc_params params = { .rate_limit.hdr.type = (max_pacing_rate == -1U) ? IF_SND_TAG_TYPE_UNLIMITED : IF_SND_TAG_TYPE_RATE_LIMIT, .rate_limit.hdr.flowid = flowid, .rate_limit.hdr.flowtype = flowtype, .rate_limit.hdr.numa_domain = inp->inp_numa_domain, .rate_limit.max_rate = max_pacing_rate, .rate_limit.flags = M_NOWAIT, }; int error; INP_WLOCK_ASSERT(inp); /* * If there is already a send tag, or the INP is being torn * down, allocating a new send tag is not allowed. Else send * tags may leak. */ - if (*st != NULL || (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) != 0) + if (*st != NULL || (inp->inp_flags & INP_DROPPED) != 0) return (EINVAL); error = m_snd_tag_alloc(ifp, ¶ms, st); #ifdef INET if (error == 0) { counter_u64_add(rate_limit_set_ok, 1); counter_u64_add(rate_limit_active, 1); } else if (error != EOPNOTSUPP) counter_u64_add(rate_limit_alloc_fail, 1); #endif return (error); } void in_pcbdetach_tag(struct m_snd_tag *mst) { m_snd_tag_rele(mst); #ifdef INET counter_u64_add(rate_limit_active, -1); #endif } /* * Free an existing TX rate limit tag based on the "inp->inp_snd_tag", * if any: */ void in_pcbdetach_txrtlmt(struct inpcb *inp) { struct m_snd_tag *mst; INP_WLOCK_ASSERT(inp); mst = inp->inp_snd_tag; inp->inp_snd_tag = NULL; if (mst == NULL) return; m_snd_tag_rele(mst); #ifdef INET counter_u64_add(rate_limit_active, -1); #endif } int in_pcboutput_txrtlmt_locked(struct inpcb *inp, struct ifnet *ifp, struct mbuf *mb, uint32_t max_pacing_rate) { int error; /* * If the existing send tag is for the wrong interface due to * a route change, first drop the existing tag. Set the * CHANGED flag so that we will keep trying to allocate a new * tag if we fail to allocate one this time. */ if (inp->inp_snd_tag != NULL && inp->inp_snd_tag->ifp != ifp) { in_pcbdetach_txrtlmt(inp); inp->inp_flags2 |= INP_RATE_LIMIT_CHANGED; } /* * NOTE: When attaching to a network interface a reference is * made to ensure the network interface doesn't go away until * all ratelimit connections are gone. The network interface * pointers compared below represent valid network interfaces, * except when comparing towards NULL. */ if (max_pacing_rate == 0 && inp->inp_snd_tag == NULL) { error = 0; } else if (!(ifp->if_capenable & IFCAP_TXRTLMT)) { if (inp->inp_snd_tag != NULL) in_pcbdetach_txrtlmt(inp); error = 0; } else if (inp->inp_snd_tag == NULL) { /* * In order to utilize packet pacing with RSS, we need * to wait until there is a valid RSS hash before we * can proceed: */ if (M_HASHTYPE_GET(mb) == M_HASHTYPE_NONE) { error = EAGAIN; } else { error = in_pcbattach_txrtlmt(inp, ifp, M_HASHTYPE_GET(mb), mb->m_pkthdr.flowid, max_pacing_rate, &inp->inp_snd_tag); } } else { error = in_pcbmodify_txrtlmt(inp, max_pacing_rate); } if (error == 0 || error == EOPNOTSUPP) inp->inp_flags2 &= ~INP_RATE_LIMIT_CHANGED; return (error); } /* * This function should be called when the INP_RATE_LIMIT_CHANGED flag * is set in the fast path and will attach/detach/modify the TX rate * limit send tag based on the socket's so_max_pacing_rate value. */ void in_pcboutput_txrtlmt(struct inpcb *inp, struct ifnet *ifp, struct mbuf *mb) { struct socket *socket; uint32_t max_pacing_rate; bool did_upgrade; if (inp == NULL) return; socket = inp->inp_socket; if (socket == NULL) return; if (!INP_WLOCKED(inp)) { /* * NOTE: If the write locking fails, we need to bail * out and use the non-ratelimited ring for the * transmit until there is a new chance to get the * write lock. */ if (!INP_TRY_UPGRADE(inp)) return; did_upgrade = 1; } else { did_upgrade = 0; } /* * NOTE: The so_max_pacing_rate value is read unlocked, * because atomic updates are not required since the variable * is checked at every mbuf we send. It is assumed that the * variable read itself will be atomic. */ max_pacing_rate = socket->so_max_pacing_rate; in_pcboutput_txrtlmt_locked(inp, ifp, mb, max_pacing_rate); if (did_upgrade) INP_DOWNGRADE(inp); } /* * Track route changes for TX rate limiting. */ void in_pcboutput_eagain(struct inpcb *inp) { bool did_upgrade; if (inp == NULL) return; if (inp->inp_snd_tag == NULL) return; if (!INP_WLOCKED(inp)) { /* * NOTE: If the write locking fails, we need to bail * out and use the non-ratelimited ring for the * transmit until there is a new chance to get the * write lock. */ if (!INP_TRY_UPGRADE(inp)) return; did_upgrade = 1; } else { did_upgrade = 0; } /* detach rate limiting */ in_pcbdetach_txrtlmt(inp); /* make sure new mbuf send tag allocation is made */ inp->inp_flags2 |= INP_RATE_LIMIT_CHANGED; if (did_upgrade) INP_DOWNGRADE(inp); } #ifdef INET static void rl_init(void *st) { rate_limit_new = counter_u64_alloc(M_WAITOK); rate_limit_chg = counter_u64_alloc(M_WAITOK); rate_limit_active = counter_u64_alloc(M_WAITOK); rate_limit_alloc_fail = counter_u64_alloc(M_WAITOK); rate_limit_set_ok = counter_u64_alloc(M_WAITOK); } SYSINIT(rl, SI_SUB_PROTO_DOMAININIT, SI_ORDER_ANY, rl_init, NULL); #endif #endif /* RATELIMIT */ diff --git a/sys/netinet/in_pcb.h b/sys/netinet/in_pcb.h index 331474999163..a72ae5742d4e 100644 --- a/sys/netinet/in_pcb.h +++ b/sys/netinet/in_pcb.h @@ -1,826 +1,825 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1982, 1986, 1990, 1993 * The Regents of the University of California. * Copyright (c) 2010-2011 Juniper Networks, Inc. * All rights reserved. * * Portions of this software were developed by Robert N. M. Watson under * contract to Juniper Networks, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)in_pcb.h 8.1 (Berkeley) 6/10/93 * $FreeBSD$ */ #ifndef _NETINET_IN_PCB_H_ #define _NETINET_IN_PCB_H_ #include #include #include #include #include #include #ifdef _KERNEL #include #include #include #include #include #include #include #endif #include /* * struct inpcb is the common protocol control block structure used in most * IP transport protocols. * * Pointers to local and foreign host table entries, local and foreign socket * numbers, and pointers up (to a socket structure) and down (to a * protocol-specific control block) are stored here. */ CK_LIST_HEAD(inpcbhead, inpcb); CK_LIST_HEAD(inpcbporthead, inpcbport); CK_LIST_HEAD(inpcblbgrouphead, inpcblbgroup); typedef uint64_t inp_gen_t; /* * PCB with AF_INET6 null bind'ed laddr can receive AF_INET input packet. * So, AF_INET6 null laddr is also used as AF_INET null laddr, by utilizing * the following structure. This requires padding always be zeroed out, * which is done right after inpcb allocation and stays through its lifetime. */ struct in_addr_4in6 { u_int32_t ia46_pad32[3]; struct in_addr ia46_addr4; }; union in_dependaddr { struct in_addr_4in6 id46_addr; struct in6_addr id6_addr; }; /* * NOTE: ipv6 addrs should be 64-bit aligned, per RFC 2553. in_conninfo has * some extra padding to accomplish this. * NOTE 2: tcp_syncache.c uses first 5 32-bit words, which identify fport, * lport, faddr to generate hash, so these fields shouldn't be moved. */ struct in_endpoints { u_int16_t ie_fport; /* foreign port */ u_int16_t ie_lport; /* local port */ /* protocol dependent part, local and foreign addr */ union in_dependaddr ie_dependfaddr; /* foreign host table entry */ union in_dependaddr ie_dependladdr; /* local host table entry */ #define ie_faddr ie_dependfaddr.id46_addr.ia46_addr4 #define ie_laddr ie_dependladdr.id46_addr.ia46_addr4 #define ie6_faddr ie_dependfaddr.id6_addr #define ie6_laddr ie_dependladdr.id6_addr u_int32_t ie6_zoneid; /* scope zone id */ }; /* * XXX The defines for inc_* are hacks and should be changed to direct * references. */ struct in_conninfo { u_int8_t inc_flags; u_int8_t inc_len; u_int16_t inc_fibnum; /* XXX was pad, 16 bits is plenty */ /* protocol dependent part */ struct in_endpoints inc_ie; }; /* * Flags for inc_flags. */ #define INC_ISIPV6 0x01 #define INC_IPV6MINMTU 0x02 #define inc_fport inc_ie.ie_fport #define inc_lport inc_ie.ie_lport #define inc_faddr inc_ie.ie_faddr #define inc_laddr inc_ie.ie_laddr #define inc6_faddr inc_ie.ie6_faddr #define inc6_laddr inc_ie.ie6_laddr #define inc6_zoneid inc_ie.ie6_zoneid #if defined(_KERNEL) || defined(_WANT_INPCB) /* * struct inpcb captures the network layer state for TCP, UDP, and raw IPv4 and * IPv6 sockets. In the case of TCP and UDP, further per-connection state is * hung off of inp_ppcb most of the time. Almost all fields of struct inpcb * are static after creation or protected by a per-inpcb rwlock, inp_lock. * * A inpcb database is indexed by addresses/ports hash as well as list of * all pcbs that belong to a certain proto. Database lookups or list traversals * are be performed inside SMR section. Once desired PCB is found its own * lock is to be obtained and SMR section exited. * * Key: * (b) - Protected by the hpts lock. * (c) - Constant after initialization * (e) - Protected by the SMR section * (i) - Protected by the inpcb lock * (p) - Protected by the pcbinfo lock for the inpcb * (h) - Protected by the pcbhash lock for the inpcb * (s) - Protected by another subsystem's locks * (x) - Undefined locking * * Notes on the tcp_hpts: * * First Hpts lock order is * 1) INP_WLOCK() * 2) HPTS_LOCK() i.e. hpts->pmtx * * To insert a TCB on the hpts you *must* be holding the INP_WLOCK(). * You may check the inp->inp_in_hpts flag without the hpts lock. * The hpts is the only one that will clear this flag holding * only the hpts lock. This means that in your tcp_output() * routine when you test for the inp_in_hpts flag to be 1 * it may be transitioning to 0 (by the hpts). * That's ok since that will just mean an extra call to tcp_output * that most likely will find the call you executed * (when the mis-match occurred) will have put the TCB back * on the hpts and it will return. If your * call did not add the inp back to the hpts then you will either * over-send or the cwnd will block you from sending more. * * Note you should also be holding the INP_WLOCK() when you * call the remove from the hpts as well. Though usually * you are either doing this from a timer, where you need and have * the INP_WLOCK() or from destroying your TCB where again * you should already have the INP_WLOCK(). * * The inp_hpts_cpu, inp_hpts_cpu_set, inp_input_cpu and * inp_input_cpu_set fields are controlled completely by * the hpts. Do not ever set these. The inp_hpts_cpu_set * and inp_input_cpu_set fields indicate if the hpts has * setup the respective cpu field. It is advised if this * field is 0, to enqueue the packet with the appropriate * hpts_immediate() call. If the _set field is 1, then * you may compare the inp_*_cpu field to the curcpu and * may want to again insert onto the hpts if these fields * are not equal (i.e. you are not on the expected CPU). * * A note on inp_hpts_calls and inp_input_calls, these * flags are set when the hpts calls either the output * or do_segment routines respectively. If the routine * being called wants to use this, then it needs to * clear the flag before returning. The hpts will not * clear the flag. The flags can be used to tell if * the hpts is the function calling the respective * routine. * * A few other notes: * * When a read lock is held, stability of the field is guaranteed; to write * to a field, a write lock must generally be held. * * netinet/netinet6-layer code should not assume that the inp_socket pointer - * is safe to dereference without inp_lock being held, even for protocols - * other than TCP (where the inpcb persists during TIMEWAIT even after the - * socket has been freed), or there may be close(2)-related races. + * is safe to dereference without inp_lock being held, there may be + * close(2)-related races. * * The inp_vflag field is overloaded, and would otherwise ideally be (c). */ struct icmp6_filter; struct inpcbpolicy; struct m_snd_tag; struct inpcb { /* Cache line #1 (amd64) */ CK_LIST_ENTRY(inpcb) inp_hash; /* (w:h/r:e) hash list */ struct rwlock inp_lock; /* Cache line #2 (amd64) */ #define inp_start_zero inp_hpts #define inp_zero_size (sizeof(struct inpcb) - \ offsetof(struct inpcb, inp_start_zero)) TAILQ_ENTRY(inpcb) inp_hpts; /* pacing out queue next lock(b) */ uint32_t inp_hpts_gencnt; /* XXXGL */ uint32_t inp_hpts_request; /* Current hpts request, zero if * fits in the pacing window (i&b). */ /* * Note the next fields are protected by a * different lock (hpts-lock). This means that * they must correspond in size to the smallest * protectable bit field (uint8_t on x86, and * other platfomrs potentially uint32_t?). Also * since CPU switches can occur at different times the two * fields can *not* be collapsed into a signal bit field. */ #if defined(__amd64__) || defined(__i386__) uint8_t inp_in_hpts; /* on output hpts (lock b) */ #else uint32_t inp_in_hpts; /* on output hpts (lock b) */ #endif volatile uint16_t inp_hpts_cpu; /* Lock (i) */ volatile uint16_t inp_irq_cpu; /* Set by LRO in behalf of or the driver */ u_int inp_refcount; /* (i) refcount */ int inp_flags; /* (i) generic IP/datagram flags */ int inp_flags2; /* (i) generic IP/datagram flags #2*/ uint8_t inp_hpts_cpu_set :1, /* on output hpts (i) */ inp_hpts_calls :1, /* (i) from output hpts */ inp_irq_cpu_set :1, /* (i) from LRO/Driver */ inp_spare_bits2 : 3; uint8_t inp_numa_domain; /* numa domain */ void *inp_ppcb; /* (i) pointer to per-protocol pcb */ struct socket *inp_socket; /* (i) back pointer to socket */ int32_t inp_hptsslot; /* Hpts wheel slot this tcb is Lock(i&b) */ uint32_t inp_hpts_drop_reas; /* reason we are dropping the PCB (lock i&b) */ struct inpcbinfo *inp_pcbinfo; /* (c) PCB list info */ struct ucred *inp_cred; /* (c) cache of socket cred */ u_int32_t inp_flow; /* (i) IPv6 flow information */ u_char inp_vflag; /* (i) IP version flag (v4/v6) */ u_char inp_ip_ttl; /* (i) time to live proto */ u_char inp_ip_p; /* (c) protocol proto */ u_char inp_ip_minttl; /* (i) minimum TTL or drop */ uint32_t inp_flowid; /* (x) flow id / queue id */ struct m_snd_tag *inp_snd_tag; /* (i) send tag for outgoing mbufs */ uint32_t inp_flowtype; /* (x) M_HASHTYPE value */ uint32_t inp_rss_listen_bucket; /* (x) overridden RSS listen bucket */ /* Local and foreign ports, local and foreign addr. */ struct in_conninfo inp_inc; /* (i) list for PCB's local port */ /* MAC and IPSEC policy information. */ struct label *inp_label; /* (i) MAC label */ struct inpcbpolicy *inp_sp; /* (s) for IPSEC */ /* Protocol-dependent part; options. */ struct { u_char inp_ip_tos; /* (i) type of service proto */ struct mbuf *inp_options; /* (i) IP options */ struct ip_moptions *inp_moptions; /* (i) mcast options */ }; struct { /* (i) IP options */ struct mbuf *in6p_options; /* (i) IP6 options for outgoing packets */ struct ip6_pktopts *in6p_outputopts; /* (i) IP multicast options */ struct ip6_moptions *in6p_moptions; /* (i) ICMPv6 code type filter */ struct icmp6_filter *in6p_icmp6filt; /* (i) IPV6_CHECKSUM setsockopt */ int in6p_cksum; short in6p_hops; }; CK_LIST_ENTRY(inpcb) inp_portlist; /* (r:e/w:h) port list */ struct inpcbport *inp_phd; /* (r:e/w:h) head of this list */ inp_gen_t inp_gencnt; /* (c) generation count */ void *spare_ptr; /* Spare pointer. */ rt_gen_t inp_rt_cookie; /* generation for route entry */ union { /* cached L3 information */ struct route inp_route; struct route_in6 inp_route6; }; CK_LIST_ENTRY(inpcb) inp_list; /* (r:e/w:p) all PCBs for proto */ }; #endif /* _KERNEL */ #define inp_fport inp_inc.inc_fport #define inp_lport inp_inc.inc_lport #define inp_faddr inp_inc.inc_faddr #define inp_laddr inp_inc.inc_laddr #define in6p_faddr inp_inc.inc6_faddr #define in6p_laddr inp_inc.inc6_laddr #define in6p_zoneid inp_inc.inc6_zoneid #define inp_vnet inp_pcbinfo->ipi_vnet /* * The range of the generation count, as used in this implementation, is 9e19. * We would have to create 300 billion connections per second for this number * to roll over in a year. This seems sufficiently unlikely that we simply * don't concern ourselves with that possibility. */ /* * Interface exported to userland by various protocols which use inpcbs. Hack * alert -- only define if struct xsocket is in scope. * Fields prefixed with "xi_" are unique to this structure, and the rest * match fields in the struct inpcb, to ease coding and porting. * * Legend: * (s) - used by userland utilities in src * (p) - used by utilities in ports * (3) - is known to be used by third party software not in ports * (n) - no known usage */ #ifdef _SYS_SOCKETVAR_H_ struct xinpcb { ksize_t xi_len; /* length of this structure */ struct xsocket xi_socket; /* (s,p) */ struct in_conninfo inp_inc; /* (s,p) */ uint64_t inp_gencnt; /* (s,p) */ kvaddr_t inp_ppcb; /* (s) netstat(1) */ int64_t inp_spare64[4]; uint32_t inp_flow; /* (s) */ uint32_t inp_flowid; /* (s) */ uint32_t inp_flowtype; /* (s) */ int32_t inp_flags; /* (s,p) */ int32_t inp_flags2; /* (s) */ int32_t inp_rss_listen_bucket; /* (n) */ int32_t in6p_cksum; /* (n) */ int32_t inp_spare32[4]; uint16_t in6p_hops; /* (n) */ uint8_t inp_ip_tos; /* (n) */ int8_t pad8; uint8_t inp_vflag; /* (s,p) */ uint8_t inp_ip_ttl; /* (n) */ uint8_t inp_ip_p; /* (n) */ uint8_t inp_ip_minttl; /* (n) */ int8_t inp_spare8[4]; } __aligned(8); struct xinpgen { ksize_t xig_len; /* length of this structure */ u_int xig_count; /* number of PCBs at this time */ uint32_t _xig_spare32; inp_gen_t xig_gen; /* generation count at this time */ so_gen_t xig_sogen; /* socket generation count this time */ uint64_t _xig_spare64[4]; } __aligned(8); struct sockopt_parameters { struct in_conninfo sop_inc; uint64_t sop_id; int sop_level; int sop_optname; char sop_optval[]; }; #ifdef _KERNEL int sysctl_setsockopt(SYSCTL_HANDLER_ARGS, struct inpcbinfo *pcbinfo, int (*ctloutput_set)(struct inpcb *, struct sockopt *)); void in_pcbtoxinpcb(const struct inpcb *, struct xinpcb *); #endif #endif /* _SYS_SOCKETVAR_H_ */ #ifdef _KERNEL /* * Per-VNET pcb database for each high-level protocol (UDP, TCP, ...) in both * IPv4 and IPv6. * * The pcbs are protected with SMR section and thus all lists in inpcbinfo * are CK-lists. Locking is required to insert a pcb into database. Two * locks are provided: one for the hash and one for the global list of pcbs, * as well as overall count and generation count. * * Locking key: * * (c) Constant or nearly constant after initialisation * (e) Protected by SMR section * (g) Locked by ipi_lock * (h) Locked by ipi_hash_lock */ struct inpcbinfo { /* * Global lock protecting inpcb list modification */ struct mtx ipi_lock; struct inpcbhead ipi_listhead; /* (r:e/w:g) */ u_int ipi_count; /* (g) */ /* * Generation count -- incremented each time a connection is allocated * or freed. */ u_quad_t ipi_gencnt; /* (g) */ /* * Fields associated with port lookup and allocation. */ u_short ipi_lastport; /* (h) */ u_short ipi_lastlow; /* (h) */ u_short ipi_lasthi; /* (h) */ /* * UMA zone from which inpcbs are allocated for this protocol. */ uma_zone_t ipi_zone; /* (c) */ uma_zone_t ipi_portzone; /* (c) */ smr_t ipi_smr; /* (c) */ /* * Global hash of inpcbs, hashed by local and foreign addresses and * port numbers. */ struct mtx ipi_hash_lock; struct inpcbhead *ipi_hashbase; /* (r:e/w:h) */ u_long ipi_hashmask; /* (c) */ /* * Global hash of inpcbs, hashed by only local port number. */ struct inpcbporthead *ipi_porthashbase; /* (h) */ u_long ipi_porthashmask; /* (h) */ /* * Load balance groups used for the SO_REUSEPORT_LB option, * hashed by local port. */ struct inpcblbgrouphead *ipi_lbgrouphashbase; /* (r:e/w:h) */ u_long ipi_lbgrouphashmask; /* (h) */ /* * Pointer to network stack instance */ struct vnet *ipi_vnet; /* (c) */ }; /* * Global allocation storage for each high-level protocol (UDP, TCP, ...). * Each corresponding per-VNET inpcbinfo points into this one. */ struct inpcbstorage { uma_zone_t ips_zone; uma_zone_t ips_portzone; uma_init ips_pcbinit; const char * ips_zone_name; const char * ips_portzone_name; const char * ips_infolock_name; const char * ips_hashlock_name; }; #define INPCBSTORAGE_DEFINE(prot, lname, zname, iname, hname) \ static int \ prot##_inpcb_init(void *mem, int size __unused, int flags __unused) \ { \ struct inpcb *inp = mem; \ \ rw_init_flags(&inp->inp_lock, lname, RW_RECURSE | RW_DUPOK); \ return (0); \ } \ static struct inpcbstorage prot = { \ .ips_pcbinit = prot##_inpcb_init, \ .ips_zone_name = zname, \ .ips_portzone_name = zname " ports", \ .ips_infolock_name = iname, \ .ips_hashlock_name = hname, \ }; \ SYSINIT(prot##_inpcbstorage_init, SI_SUB_PROTO_DOMAIN, \ SI_ORDER_SECOND, in_pcbstorage_init, &prot); \ SYSUNINIT(prot##_inpcbstorage_uninit, SI_SUB_PROTO_DOMAIN, \ SI_ORDER_SECOND, in_pcbstorage_destroy, &prot) /* * Load balance groups used for the SO_REUSEPORT_LB socket option. Each group * (or unique address:port combination) can be re-used at most * INPCBLBGROUP_SIZMAX (256) times. The inpcbs are stored in il_inp which * is dynamically resized as processes bind/unbind to that specific group. */ struct inpcblbgroup { CK_LIST_ENTRY(inpcblbgroup) il_list; struct epoch_context il_epoch_ctx; uint16_t il_lport; /* (c) */ u_char il_vflag; /* (c) */ u_int8_t il_numa_domain; uint32_t il_pad2; union in_dependaddr il_dependladdr; /* (c) */ #define il_laddr il_dependladdr.id46_addr.ia46_addr4 #define il6_laddr il_dependladdr.id6_addr uint32_t il_inpsiz; /* max count in il_inp[] (h) */ uint32_t il_inpcnt; /* cur count in il_inp[] (h) */ struct inpcb *il_inp[]; /* (h) */ }; #define INP_LOCK_DESTROY(inp) rw_destroy(&(inp)->inp_lock) #define INP_RLOCK(inp) rw_rlock(&(inp)->inp_lock) #define INP_WLOCK(inp) rw_wlock(&(inp)->inp_lock) #define INP_TRY_RLOCK(inp) rw_try_rlock(&(inp)->inp_lock) #define INP_TRY_WLOCK(inp) rw_try_wlock(&(inp)->inp_lock) #define INP_RUNLOCK(inp) rw_runlock(&(inp)->inp_lock) #define INP_WUNLOCK(inp) rw_wunlock(&(inp)->inp_lock) #define INP_UNLOCK(inp) rw_unlock(&(inp)->inp_lock) #define INP_TRY_UPGRADE(inp) rw_try_upgrade(&(inp)->inp_lock) #define INP_DOWNGRADE(inp) rw_downgrade(&(inp)->inp_lock) #define INP_WLOCKED(inp) rw_wowned(&(inp)->inp_lock) #define INP_LOCK_ASSERT(inp) rw_assert(&(inp)->inp_lock, RA_LOCKED) #define INP_RLOCK_ASSERT(inp) rw_assert(&(inp)->inp_lock, RA_RLOCKED) #define INP_WLOCK_ASSERT(inp) rw_assert(&(inp)->inp_lock, RA_WLOCKED) #define INP_UNLOCK_ASSERT(inp) rw_assert(&(inp)->inp_lock, RA_UNLOCKED) /* * These locking functions are for inpcb consumers outside of sys/netinet, * more specifically, they were added for the benefit of TOE drivers. The * macros are reserved for use by the stack. */ void inp_wlock(struct inpcb *); void inp_wunlock(struct inpcb *); void inp_rlock(struct inpcb *); void inp_runlock(struct inpcb *); #ifdef INVARIANT_SUPPORT void inp_lock_assert(struct inpcb *); void inp_unlock_assert(struct inpcb *); #else #define inp_lock_assert(inp) do {} while (0) #define inp_unlock_assert(inp) do {} while (0) #endif void inp_apply_all(void (*func)(struct inpcb *, void *), void *arg); int inp_ip_tos_get(const struct inpcb *inp); void inp_ip_tos_set(struct inpcb *inp, int val); struct socket * inp_inpcbtosocket(struct inpcb *inp); struct tcpcb * inp_inpcbtotcpcb(struct inpcb *inp); void inp_4tuple_get(struct inpcb *inp, uint32_t *laddr, uint16_t *lp, uint32_t *faddr, uint16_t *fp); int inp_so_options(const struct inpcb *inp); #endif /* _KERNEL */ #define INP_INFO_WLOCK(ipi) mtx_lock(&(ipi)->ipi_lock) #define INP_INFO_WLOCKED(ipi) mtx_owned(&(ipi)->ipi_lock) #define INP_INFO_WUNLOCK(ipi) mtx_unlock(&(ipi)->ipi_lock) #define INP_INFO_LOCK_ASSERT(ipi) MPASS(SMR_ENTERED((ipi)->ipi_smr) || \ mtx_owned(&(ipi)->ipi_lock)) #define INP_INFO_WLOCK_ASSERT(ipi) mtx_assert(&(ipi)->ipi_lock, MA_OWNED) #define INP_INFO_WUNLOCK_ASSERT(ipi) \ mtx_assert(&(ipi)->ipi_lock, MA_NOTOWNED) #define INP_HASH_WLOCK(ipi) mtx_lock(&(ipi)->ipi_hash_lock) #define INP_HASH_WUNLOCK(ipi) mtx_unlock(&(ipi)->ipi_hash_lock) #define INP_HASH_LOCK_ASSERT(ipi) MPASS(SMR_ENTERED((ipi)->ipi_smr) || \ mtx_owned(&(ipi)->ipi_hash_lock)) #define INP_HASH_WLOCK_ASSERT(ipi) mtx_assert(&(ipi)->ipi_hash_lock, \ MA_OWNED) /* * Wildcard matching hash is not just a microoptimisation! The hash for * wildcard IPv4 and wildcard IPv6 must be the same, otherwise AF_INET6 * wildcard bound pcb won't be able to receive AF_INET connections, while: * jenkins_hash(&zeroes, 1, s) != jenkins_hash(&zeroes, 4, s) * See also comment above struct in_addr_4in6. */ #define IN_ADDR_JHASH32(addr) \ ((addr)->s_addr == INADDR_ANY ? V_in_pcbhashseed : \ jenkins_hash32((&(addr)->s_addr), 1, V_in_pcbhashseed)) #define IN6_ADDR_JHASH32(addr) \ (memcmp((addr), &in6addr_any, sizeof(in6addr_any)) == 0 ? \ V_in_pcbhashseed : \ jenkins_hash32((addr)->__u6_addr.__u6_addr32, \ nitems((addr)->__u6_addr.__u6_addr32), V_in_pcbhashseed)) #define INP_PCBHASH(faddr, lport, fport, mask) \ ((IN_ADDR_JHASH32(faddr) ^ ntohs((lport) ^ (fport))) & (mask)) #define INP6_PCBHASH(faddr, lport, fport, mask) \ ((IN6_ADDR_JHASH32(faddr) ^ ntohs((lport) ^ (fport))) & (mask)) #define INP_PCBHASH_WILD(lport, mask) \ ((V_in_pcbhashseed ^ ntohs(lport)) & (mask)) #define INP_PCBLBGROUP_PKTHASH(faddr, lport, fport) \ (IN_ADDR_JHASH32(faddr) ^ ntohs((lport) ^ (fport))) #define INP6_PCBLBGROUP_PKTHASH(faddr, lport, fport) \ (IN6_ADDR_JHASH32(faddr) ^ ntohs((lport) ^ (fport))) #define INP_PCBPORTHASH(lport, mask) (ntohs((lport)) & (mask)) /* * Flags for inp_vflags -- historically version flags only */ #define INP_IPV4 0x1 #define INP_IPV6 0x2 #define INP_IPV6PROTO 0x4 /* opened under IPv6 protocol */ /* * Flags for inp_flags. */ #define INP_RECVOPTS 0x00000001 /* receive incoming IP options */ #define INP_RECVRETOPTS 0x00000002 /* receive IP options for reply */ #define INP_RECVDSTADDR 0x00000004 /* receive IP dst address */ #define INP_HDRINCL 0x00000008 /* user supplies entire IP header */ #define INP_HIGHPORT 0x00000010 /* user wants "high" port binding */ #define INP_LOWPORT 0x00000020 /* user wants "low" port binding */ #define INP_ANONPORT 0x00000040 /* port chosen for user */ #define INP_RECVIF 0x00000080 /* receive incoming interface */ #define INP_MTUDISC 0x00000100 /* user can do MTU discovery */ /* INP_FREED 0x00000200 private to in_pcb.c */ #define INP_RECVTTL 0x00000400 /* receive incoming IP TTL */ #define INP_DONTFRAG 0x00000800 /* don't fragment packet */ #define INP_BINDANY 0x00001000 /* allow bind to any address */ #define INP_INHASHLIST 0x00002000 /* in_pcbinshash() has been called */ #define INP_RECVTOS 0x00004000 /* receive incoming IP TOS */ #define IN6P_IPV6_V6ONLY 0x00008000 /* restrict AF_INET6 socket for v6 */ #define IN6P_PKTINFO 0x00010000 /* receive IP6 dst and I/F */ #define IN6P_HOPLIMIT 0x00020000 /* receive hoplimit */ #define IN6P_HOPOPTS 0x00040000 /* receive hop-by-hop options */ #define IN6P_DSTOPTS 0x00080000 /* receive dst options after rthdr */ #define IN6P_RTHDR 0x00100000 /* receive routing header */ #define IN6P_RTHDRDSTOPTS 0x00200000 /* receive dstoptions before rthdr */ #define IN6P_TCLASS 0x00400000 /* receive traffic class value */ #define IN6P_AUTOFLOWLABEL 0x00800000 /* attach flowlabel automatically */ -#define INP_TIMEWAIT 0x01000000 /* in TIMEWAIT, ppcb is tcptw */ +/* was INP_TIMEWAIT 0x01000000 */ #define INP_ONESBCAST 0x02000000 /* send all-ones broadcast */ #define INP_DROPPED 0x04000000 /* protocol drop flag */ #define INP_SOCKREF 0x08000000 /* strong socket reference */ #define INP_RESERVED_0 0x10000000 /* reserved field */ #define INP_RESERVED_1 0x20000000 /* reserved field */ #define IN6P_RFC2292 0x40000000 /* used RFC2292 API on the socket */ #define IN6P_MTU 0x80000000 /* receive path MTU */ #define INP_CONTROLOPTS (INP_RECVOPTS|INP_RECVRETOPTS|INP_RECVDSTADDR|\ INP_RECVIF|INP_RECVTTL|INP_RECVTOS|\ IN6P_PKTINFO|IN6P_HOPLIMIT|IN6P_HOPOPTS|\ IN6P_DSTOPTS|IN6P_RTHDR|IN6P_RTHDRDSTOPTS|\ IN6P_TCLASS|IN6P_AUTOFLOWLABEL|IN6P_RFC2292|\ IN6P_MTU) /* * Flags for inp_flags2. */ #define INP_MBUF_L_ACKS 0x00000001 /* We need large mbufs for ack compression */ #define INP_MBUF_ACKCMP 0x00000002 /* TCP mbuf ack compression ok */ /* 0x00000004 */ #define INP_REUSEPORT 0x00000008 /* SO_REUSEPORT option is set */ /* 0x00000010 */ #define INP_REUSEADDR 0x00000020 /* SO_REUSEADDR option is set */ #define INP_BINDMULTI 0x00000040 /* IP_BINDMULTI option is set */ #define INP_RSS_BUCKET_SET 0x00000080 /* IP_RSS_LISTEN_BUCKET is set */ #define INP_RECVFLOWID 0x00000100 /* populate recv datagram with flow info */ #define INP_RECVRSSBUCKETID 0x00000200 /* populate recv datagram with bucket id */ #define INP_RATE_LIMIT_CHANGED 0x00000400 /* rate limit needs attention */ #define INP_ORIGDSTADDR 0x00000800 /* receive IP dst address/port */ #define INP_CANNOT_DO_ECN 0x00001000 /* The stack does not do ECN */ #define INP_REUSEPORT_LB 0x00002000 /* SO_REUSEPORT_LB option is set */ #define INP_SUPPORTS_MBUFQ 0x00004000 /* Supports the mbuf queue method of LRO */ #define INP_MBUF_QUEUE_READY 0x00008000 /* The transport is pacing, inputs can be queued */ #define INP_DONT_SACK_QUEUE 0x00010000 /* If a sack arrives do not wake me */ #define INP_2PCP_SET 0x00020000 /* If the Eth PCP should be set explicitly */ #define INP_2PCP_BIT0 0x00040000 /* Eth PCP Bit 0 */ #define INP_2PCP_BIT1 0x00080000 /* Eth PCP Bit 1 */ #define INP_2PCP_BIT2 0x00100000 /* Eth PCP Bit 2 */ #define INP_2PCP_BASE INP_2PCP_BIT0 #define INP_2PCP_MASK (INP_2PCP_BIT0 | INP_2PCP_BIT1 | INP_2PCP_BIT2) #define INP_2PCP_SHIFT 18 /* shift PCP field in/out of inp_flags2 */ /* * Flags passed to in_pcblookup*(), inp_smr_lock() and inp_next(). */ typedef enum { INPLOOKUP_WILDCARD = 0x00000001, /* Allow wildcard sockets. */ INPLOOKUP_RLOCKPCB = 0x00000002, /* Return inpcb read-locked. */ INPLOOKUP_WLOCKPCB = 0x00000004, /* Return inpcb write-locked. */ } inp_lookup_t; #define INPLOOKUP_MASK (INPLOOKUP_WILDCARD | INPLOOKUP_RLOCKPCB | \ INPLOOKUP_WLOCKPCB) #define INPLOOKUP_LOCKMASK (INPLOOKUP_RLOCKPCB | INPLOOKUP_WLOCKPCB) #define sotoinpcb(so) ((struct inpcb *)(so)->so_pcb) #define INP_SOCKAF(so) so->so_proto->pr_domain->dom_family #define INP_CHECK_SOCKAF(so, af) (INP_SOCKAF(so) == af) #ifdef _KERNEL VNET_DECLARE(int, ipport_reservedhigh); VNET_DECLARE(int, ipport_reservedlow); VNET_DECLARE(int, ipport_lowfirstauto); VNET_DECLARE(int, ipport_lowlastauto); VNET_DECLARE(int, ipport_firstauto); VNET_DECLARE(int, ipport_lastauto); VNET_DECLARE(int, ipport_hifirstauto); VNET_DECLARE(int, ipport_hilastauto); VNET_DECLARE(int, ipport_randomized); VNET_DECLARE(int, ipport_randomcps); VNET_DECLARE(int, ipport_randomtime); VNET_DECLARE(int, ipport_stoprandom); VNET_DECLARE(int, ipport_tcpallocs); #define V_ipport_reservedhigh VNET(ipport_reservedhigh) #define V_ipport_reservedlow VNET(ipport_reservedlow) #define V_ipport_lowfirstauto VNET(ipport_lowfirstauto) #define V_ipport_lowlastauto VNET(ipport_lowlastauto) #define V_ipport_firstauto VNET(ipport_firstauto) #define V_ipport_lastauto VNET(ipport_lastauto) #define V_ipport_hifirstauto VNET(ipport_hifirstauto) #define V_ipport_hilastauto VNET(ipport_hilastauto) #define V_ipport_randomized VNET(ipport_randomized) #define V_ipport_randomcps VNET(ipport_randomcps) #define V_ipport_randomtime VNET(ipport_randomtime) #define V_ipport_stoprandom VNET(ipport_stoprandom) #define V_ipport_tcpallocs VNET(ipport_tcpallocs) void in_pcbinfo_init(struct inpcbinfo *, struct inpcbstorage *, u_int, u_int); void in_pcbinfo_destroy(struct inpcbinfo *); void in_pcbstorage_init(void *); void in_pcbstorage_destroy(void *); int in_pcbbind_check_bindmulti(const struct inpcb *ni, const struct inpcb *oi); void in_pcbpurgeif0(struct inpcbinfo *, struct ifnet *); int in_pcballoc(struct socket *, struct inpcbinfo *); int in_pcbbind(struct inpcb *, struct sockaddr *, struct ucred *); int in_pcbbind_setup(struct inpcb *, struct sockaddr *, in_addr_t *, u_short *, struct ucred *); int in_pcbconnect(struct inpcb *, struct sockaddr *, struct ucred *, bool); int in_pcbconnect_setup(struct inpcb *, struct sockaddr *, in_addr_t *, u_short *, in_addr_t *, u_short *, struct inpcb **, struct ucred *); void in_pcbdetach(struct inpcb *); void in_pcbdisconnect(struct inpcb *); void in_pcbdrop(struct inpcb *); void in_pcbfree(struct inpcb *); int in_pcbinshash(struct inpcb *); int in_pcbladdr(struct inpcb *, struct in_addr *, struct in_addr *, struct ucred *); int in_pcblbgroup_numa(struct inpcb *, int arg); struct inpcb * in_pcblookup(struct inpcbinfo *, struct in_addr, u_int, struct in_addr, u_int, int, struct ifnet *); struct inpcb * in_pcblookup_mbuf(struct inpcbinfo *, struct in_addr, u_int, struct in_addr, u_int, int, struct ifnet *, struct mbuf *); void in_pcbnotifyall(struct inpcbinfo *pcbinfo, struct in_addr, int, struct inpcb *(*)(struct inpcb *, int)); void in_pcbref(struct inpcb *); void in_pcbrehash(struct inpcb *); bool in_pcbrele_rlocked(struct inpcb *); bool in_pcbrele_wlocked(struct inpcb *); typedef bool inp_match_t(const struct inpcb *, void *); struct inpcb_iterator { const struct inpcbinfo *ipi; struct inpcb *inp; inp_match_t *match; void *ctx; int hash; #define INP_ALL_LIST -1 const inp_lookup_t lock; }; /* Note: sparse initializers guarantee .inp = NULL. */ #define INP_ITERATOR(_ipi, _lock, _match, _ctx) \ { \ .ipi = (_ipi), \ .lock = (_lock), \ .hash = INP_ALL_LIST, \ .match = (_match), \ .ctx = (_ctx), \ } #define INP_ALL_ITERATOR(_ipi, _lock) \ { \ .ipi = (_ipi), \ .lock = (_lock), \ .hash = INP_ALL_LIST, \ } struct inpcb *inp_next(struct inpcb_iterator *); void in_losing(struct inpcb *); void in_pcbsetsolabel(struct socket *so); int in_getpeeraddr(struct socket *so, struct sockaddr **nam); int in_getsockaddr(struct socket *so, struct sockaddr **nam); struct sockaddr * in_sockaddr(in_port_t port, struct in_addr *addr); void in_pcbsosetlabel(struct socket *so); #ifdef RATELIMIT int in_pcboutput_txrtlmt_locked(struct inpcb *, struct ifnet *, struct mbuf *, uint32_t); int in_pcbattach_txrtlmt(struct inpcb *, struct ifnet *, uint32_t, uint32_t, uint32_t, struct m_snd_tag **); void in_pcbdetach_txrtlmt(struct inpcb *); void in_pcbdetach_tag(struct m_snd_tag *); int in_pcbmodify_txrtlmt(struct inpcb *, uint32_t); int in_pcbquery_txrtlmt(struct inpcb *, uint32_t *); int in_pcbquery_txrlevel(struct inpcb *, uint32_t *); void in_pcboutput_txrtlmt(struct inpcb *, struct ifnet *, struct mbuf *); void in_pcboutput_eagain(struct inpcb *); #endif #endif /* _KERNEL */ #endif /* !_NETINET_IN_PCB_H_ */ diff --git a/sys/netinet/siftr.c b/sys/netinet/siftr.c index b326218a8729..c025c06d7f32 100644 --- a/sys/netinet/siftr.c +++ b/sys/netinet/siftr.c @@ -1,1599 +1,1597 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2007-2009 * Swinburne University of Technology, Melbourne, Australia. * Copyright (c) 2009-2010, The FreeBSD Foundation * All rights reserved. * * Portions of this software were developed at the Centre for Advanced * Internet Architectures, Swinburne University of Technology, Melbourne, * Australia by Lawrence Stewart under sponsorship from the FreeBSD Foundation. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHORS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHORS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ /****************************************************** * Statistical Information For TCP Research (SIFTR) * * A FreeBSD kernel module that adds very basic intrumentation to the * TCP stack, allowing internal stats to be recorded to a log file * for experimental, debugging and performance analysis purposes. * * SIFTR was first released in 2007 by James Healy and Lawrence Stewart whilst * working on the NewTCP research project at Swinburne University of * Technology's Centre for Advanced Internet Architectures, Melbourne, * Australia, which was made possible in part by a grant from the Cisco * University Research Program Fund at Community Foundation Silicon Valley. * More details are available at: * http://caia.swin.edu.au/urp/newtcp/ * * Work on SIFTR v1.2.x was sponsored by the FreeBSD Foundation as part of * the "Enhancing the FreeBSD TCP Implementation" project 2008-2009. * More details are available at: * http://www.freebsdfoundation.org/ * http://caia.swin.edu.au/freebsd/etcp09/ * * Lawrence Stewart is the current maintainer, and all contact regarding * SIFTR should be directed to him via email: lastewart@swin.edu.au * * Initial release date: June 2007 * Most recent update: September 2010 ******************************************************/ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef SIFTR_IPV6 #include #include #include #endif /* SIFTR_IPV6 */ #include /* * Three digit version number refers to X.Y.Z where: * X is the major version number * Y is bumped to mark backwards incompatible changes * Z is bumped to mark backwards compatible changes */ #define V_MAJOR 1 #define V_BACKBREAK 2 #define V_BACKCOMPAT 4 #define MODVERSION __CONCAT(V_MAJOR, __CONCAT(V_BACKBREAK, V_BACKCOMPAT)) #define MODVERSION_STR __XSTRING(V_MAJOR) "." __XSTRING(V_BACKBREAK) "." \ __XSTRING(V_BACKCOMPAT) #define HOOK 0 #define UNHOOK 1 #define SIFTR_EXPECTED_MAX_TCP_FLOWS 65536 #define SYS_NAME "FreeBSD" #define PACKET_TAG_SIFTR 100 #define PACKET_COOKIE_SIFTR 21749576 #define SIFTR_LOG_FILE_MODE 0644 #define SIFTR_DISABLE 0 #define SIFTR_ENABLE 1 /* * Hard upper limit on the length of log messages. Bump this up if you add new * data fields such that the line length could exceed the below value. */ #define MAX_LOG_MSG_LEN 300 /* XXX: Make this a sysctl tunable. */ #define SIFTR_ALQ_BUFLEN (1000*MAX_LOG_MSG_LEN) /* * 1 byte for IP version * IPv4: src/dst IP (4+4) + src/dst port (2+2) = 12 bytes * IPv6: src/dst IP (16+16) + src/dst port (2+2) = 36 bytes */ #ifdef SIFTR_IPV6 #define FLOW_KEY_LEN 37 #else #define FLOW_KEY_LEN 13 #endif #ifdef SIFTR_IPV6 #define SIFTR_IPMODE 6 #else #define SIFTR_IPMODE 4 #endif /* useful macros */ #define UPPER_SHORT(X) (((X) & 0xFFFF0000) >> 16) #define LOWER_SHORT(X) ((X) & 0x0000FFFF) #define FIRST_OCTET(X) (((X) & 0xFF000000) >> 24) #define SECOND_OCTET(X) (((X) & 0x00FF0000) >> 16) #define THIRD_OCTET(X) (((X) & 0x0000FF00) >> 8) #define FOURTH_OCTET(X) ((X) & 0x000000FF) static MALLOC_DEFINE(M_SIFTR, "siftr", "dynamic memory used by SIFTR"); static MALLOC_DEFINE(M_SIFTR_PKTNODE, "siftr_pktnode", "SIFTR pkt_node struct"); static MALLOC_DEFINE(M_SIFTR_HASHNODE, "siftr_hashnode", "SIFTR flow_hash_node struct"); /* Used as links in the pkt manager queue. */ struct pkt_node { /* Timestamp of pkt as noted in the pfil hook. */ struct timeval tval; /* Direction pkt is travelling. */ enum { DIR_IN = 0, DIR_OUT = 1, } direction; /* IP version pkt_node relates to; either INP_IPV4 or INP_IPV6. */ uint8_t ipver; /* Hash of the pkt which triggered the log message. */ uint32_t hash; /* Local/foreign IP address. */ #ifdef SIFTR_IPV6 uint32_t ip_laddr[4]; uint32_t ip_faddr[4]; #else uint8_t ip_laddr[4]; uint8_t ip_faddr[4]; #endif /* Local TCP port. */ uint16_t tcp_localport; /* Foreign TCP port. */ uint16_t tcp_foreignport; /* Congestion Window (bytes). */ uint32_t snd_cwnd; /* Sending Window (bytes). */ uint32_t snd_wnd; /* Receive Window (bytes). */ uint32_t rcv_wnd; /* More tcpcb flags storage */ uint32_t t_flags2; /* Slow Start Threshold (bytes). */ uint32_t snd_ssthresh; /* Current state of the TCP FSM. */ int conn_state; /* Max Segment Size (bytes). */ u_int max_seg_size; /* * Smoothed RTT stored as found in the TCP control block * in units of (TCP_RTT_SCALE*hz). */ int smoothed_rtt; /* Is SACK enabled? */ u_char sack_enabled; /* Window scaling for snd window. */ u_char snd_scale; /* Window scaling for recv window. */ u_char rcv_scale; /* TCP control block flags. */ u_int flags; /* Retransmit timeout length. */ int rxt_length; /* Size of the TCP send buffer in bytes. */ u_int snd_buf_hiwater; /* Current num bytes in the send socket buffer. */ u_int snd_buf_cc; /* Size of the TCP receive buffer in bytes. */ u_int rcv_buf_hiwater; /* Current num bytes in the receive socket buffer. */ u_int rcv_buf_cc; /* Number of bytes inflight that we are waiting on ACKs for. */ u_int sent_inflight_bytes; /* Number of segments currently in the reassembly queue. */ int t_segqlen; /* Flowid for the connection. */ u_int flowid; /* Flow type for the connection. */ u_int flowtype; /* Link to next pkt_node in the list. */ STAILQ_ENTRY(pkt_node) nodes; }; struct flow_hash_node { uint16_t counter; uint8_t key[FLOW_KEY_LEN]; LIST_ENTRY(flow_hash_node) nodes; }; struct siftr_stats { /* # TCP pkts seen by the SIFTR PFIL hooks, including any skipped. */ uint64_t n_in; uint64_t n_out; /* # pkts skipped due to failed malloc calls. */ uint32_t nskip_in_malloc; uint32_t nskip_out_malloc; /* # pkts skipped due to failed mtx acquisition. */ uint32_t nskip_in_mtx; uint32_t nskip_out_mtx; /* # pkts skipped due to failed inpcb lookups. */ uint32_t nskip_in_inpcb; uint32_t nskip_out_inpcb; /* # pkts skipped due to failed tcpcb lookups. */ uint32_t nskip_in_tcpcb; uint32_t nskip_out_tcpcb; /* # pkts skipped due to stack reinjection. */ uint32_t nskip_in_dejavu; uint32_t nskip_out_dejavu; }; DPCPU_DEFINE_STATIC(struct siftr_stats, ss); static volatile unsigned int siftr_exit_pkt_manager_thread = 0; static unsigned int siftr_enabled = 0; static unsigned int siftr_pkts_per_log = 1; static unsigned int siftr_generate_hashes = 0; static uint16_t siftr_port_filter = 0; /* static unsigned int siftr_binary_log = 0; */ static char siftr_logfile[PATH_MAX] = "/var/log/siftr.log"; static char siftr_logfile_shadow[PATH_MAX] = "/var/log/siftr.log"; static u_long siftr_hashmask; STAILQ_HEAD(pkthead, pkt_node) pkt_queue = STAILQ_HEAD_INITIALIZER(pkt_queue); LIST_HEAD(listhead, flow_hash_node) *counter_hash; static int wait_for_pkt; static struct alq *siftr_alq = NULL; static struct mtx siftr_pkt_queue_mtx; static struct mtx siftr_pkt_mgr_mtx; static struct thread *siftr_pkt_manager_thr = NULL; static char direction[2] = {'i','o'}; /* Required function prototypes. */ static int siftr_sysctl_enabled_handler(SYSCTL_HANDLER_ARGS); static int siftr_sysctl_logfile_name_handler(SYSCTL_HANDLER_ARGS); /* Declare the net.inet.siftr sysctl tree and populate it. */ SYSCTL_DECL(_net_inet_siftr); SYSCTL_NODE(_net_inet, OID_AUTO, siftr, CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, "siftr related settings"); SYSCTL_PROC(_net_inet_siftr, OID_AUTO, enabled, CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &siftr_enabled, 0, &siftr_sysctl_enabled_handler, "IU", "switch siftr module operations on/off"); SYSCTL_PROC(_net_inet_siftr, OID_AUTO, logfile, CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &siftr_logfile_shadow, sizeof(siftr_logfile_shadow), &siftr_sysctl_logfile_name_handler, "A", "file to save siftr log messages to"); SYSCTL_UINT(_net_inet_siftr, OID_AUTO, ppl, CTLFLAG_RW, &siftr_pkts_per_log, 1, "number of packets between generating a log message"); SYSCTL_UINT(_net_inet_siftr, OID_AUTO, genhashes, CTLFLAG_RW, &siftr_generate_hashes, 0, "enable packet hash generation"); SYSCTL_U16(_net_inet_siftr, OID_AUTO, port_filter, CTLFLAG_RW, &siftr_port_filter, 0, "enable packet filter on a TCP port"); /* XXX: TODO SYSCTL_UINT(_net_inet_siftr, OID_AUTO, binary, CTLFLAG_RW, &siftr_binary_log, 0, "write log files in binary instead of ascii"); */ /* Begin functions. */ static void siftr_process_pkt(struct pkt_node * pkt_node) { struct flow_hash_node *hash_node; struct listhead *counter_list; struct siftr_stats *ss; struct ale *log_buf; uint8_t key[FLOW_KEY_LEN]; uint8_t found_match, key_offset; hash_node = NULL; ss = DPCPU_PTR(ss); found_match = 0; key_offset = 1; /* * Create the key that will be used to create a hash index * into our hash table. Our key consists of: * ipversion, localip, localport, foreignip, foreignport */ key[0] = pkt_node->ipver; memcpy(key + key_offset, &pkt_node->ip_laddr, sizeof(pkt_node->ip_laddr)); key_offset += sizeof(pkt_node->ip_laddr); memcpy(key + key_offset, &pkt_node->tcp_localport, sizeof(pkt_node->tcp_localport)); key_offset += sizeof(pkt_node->tcp_localport); memcpy(key + key_offset, &pkt_node->ip_faddr, sizeof(pkt_node->ip_faddr)); key_offset += sizeof(pkt_node->ip_faddr); memcpy(key + key_offset, &pkt_node->tcp_foreignport, sizeof(pkt_node->tcp_foreignport)); counter_list = counter_hash + (hash32_buf(key, sizeof(key), 0) & siftr_hashmask); /* * If the list is not empty i.e. the hash index has * been used by another flow previously. */ if (LIST_FIRST(counter_list) != NULL) { /* * Loop through the hash nodes in the list. * There should normally only be 1 hash node in the list, * except if there have been collisions at the hash index * computed by hash32_buf(). */ LIST_FOREACH(hash_node, counter_list, nodes) { /* * Check if the key for the pkt we are currently * processing is the same as the key stored in the * hash node we are currently processing. * If they are the same, then we've found the * hash node that stores the counter for the flow * the pkt belongs to. */ if (memcmp(hash_node->key, key, sizeof(key)) == 0) { found_match = 1; break; } } } /* If this flow hash hasn't been seen before or we have a collision. */ if (hash_node == NULL || !found_match) { /* Create a new hash node to store the flow's counter. */ hash_node = malloc(sizeof(struct flow_hash_node), M_SIFTR_HASHNODE, M_WAITOK); if (hash_node != NULL) { /* Initialise our new hash node list entry. */ hash_node->counter = 0; memcpy(hash_node->key, key, sizeof(key)); LIST_INSERT_HEAD(counter_list, hash_node, nodes); } else { /* Malloc failed. */ if (pkt_node->direction == DIR_IN) ss->nskip_in_malloc++; else ss->nskip_out_malloc++; return; } } else if (siftr_pkts_per_log > 1) { /* * Taking the remainder of the counter divided * by the current value of siftr_pkts_per_log * and storing that in counter provides a neat * way to modulate the frequency of log * messages being written to the log file. */ hash_node->counter = (hash_node->counter + 1) % siftr_pkts_per_log; /* * If we have not seen enough packets since the last time * we wrote a log message for this connection, return. */ if (hash_node->counter > 0) return; } log_buf = alq_getn(siftr_alq, MAX_LOG_MSG_LEN, ALQ_WAITOK); if (log_buf == NULL) return; /* Should only happen if the ALQ is shutting down. */ #ifdef SIFTR_IPV6 pkt_node->ip_laddr[3] = ntohl(pkt_node->ip_laddr[3]); pkt_node->ip_faddr[3] = ntohl(pkt_node->ip_faddr[3]); if (pkt_node->ipver == INP_IPV6) { /* IPv6 packet */ pkt_node->ip_laddr[0] = ntohl(pkt_node->ip_laddr[0]); pkt_node->ip_laddr[1] = ntohl(pkt_node->ip_laddr[1]); pkt_node->ip_laddr[2] = ntohl(pkt_node->ip_laddr[2]); pkt_node->ip_faddr[0] = ntohl(pkt_node->ip_faddr[0]); pkt_node->ip_faddr[1] = ntohl(pkt_node->ip_faddr[1]); pkt_node->ip_faddr[2] = ntohl(pkt_node->ip_faddr[2]); /* Construct an IPv6 log message. */ log_buf->ae_bytesused = snprintf(log_buf->ae_data, MAX_LOG_MSG_LEN, "%c,0x%08x,%zd.%06ld,%x:%x:%x:%x:%x:%x:%x:%x,%u,%x:%x:%x:" "%x:%x:%x:%x:%x,%u,%u,%u,%u,%u,%u,%u,%u,%u,%u,%u,%u," "%u,%d,%u,%u,%u,%u,%u,%u,%u,%u\n", direction[pkt_node->direction], pkt_node->hash, pkt_node->tval.tv_sec, pkt_node->tval.tv_usec, UPPER_SHORT(pkt_node->ip_laddr[0]), LOWER_SHORT(pkt_node->ip_laddr[0]), UPPER_SHORT(pkt_node->ip_laddr[1]), LOWER_SHORT(pkt_node->ip_laddr[1]), UPPER_SHORT(pkt_node->ip_laddr[2]), LOWER_SHORT(pkt_node->ip_laddr[2]), UPPER_SHORT(pkt_node->ip_laddr[3]), LOWER_SHORT(pkt_node->ip_laddr[3]), ntohs(pkt_node->tcp_localport), UPPER_SHORT(pkt_node->ip_faddr[0]), LOWER_SHORT(pkt_node->ip_faddr[0]), UPPER_SHORT(pkt_node->ip_faddr[1]), LOWER_SHORT(pkt_node->ip_faddr[1]), UPPER_SHORT(pkt_node->ip_faddr[2]), LOWER_SHORT(pkt_node->ip_faddr[2]), UPPER_SHORT(pkt_node->ip_faddr[3]), LOWER_SHORT(pkt_node->ip_faddr[3]), ntohs(pkt_node->tcp_foreignport), pkt_node->snd_ssthresh, pkt_node->snd_cwnd, pkt_node->t_flags2, pkt_node->snd_wnd, pkt_node->rcv_wnd, pkt_node->snd_scale, pkt_node->rcv_scale, pkt_node->conn_state, pkt_node->max_seg_size, pkt_node->smoothed_rtt, pkt_node->sack_enabled, pkt_node->flags, pkt_node->rxt_length, pkt_node->snd_buf_hiwater, pkt_node->snd_buf_cc, pkt_node->rcv_buf_hiwater, pkt_node->rcv_buf_cc, pkt_node->sent_inflight_bytes, pkt_node->t_segqlen, pkt_node->flowid, pkt_node->flowtype); } else { /* IPv4 packet */ pkt_node->ip_laddr[0] = FIRST_OCTET(pkt_node->ip_laddr[3]); pkt_node->ip_laddr[1] = SECOND_OCTET(pkt_node->ip_laddr[3]); pkt_node->ip_laddr[2] = THIRD_OCTET(pkt_node->ip_laddr[3]); pkt_node->ip_laddr[3] = FOURTH_OCTET(pkt_node->ip_laddr[3]); pkt_node->ip_faddr[0] = FIRST_OCTET(pkt_node->ip_faddr[3]); pkt_node->ip_faddr[1] = SECOND_OCTET(pkt_node->ip_faddr[3]); pkt_node->ip_faddr[2] = THIRD_OCTET(pkt_node->ip_faddr[3]); pkt_node->ip_faddr[3] = FOURTH_OCTET(pkt_node->ip_faddr[3]); #endif /* SIFTR_IPV6 */ /* Construct an IPv4 log message. */ log_buf->ae_bytesused = snprintf(log_buf->ae_data, MAX_LOG_MSG_LEN, "%c,0x%08x,%jd.%06ld,%u.%u.%u.%u,%u,%u.%u.%u.%u,%u,%u,%u," "%u,%u,%u,%u,%u,%u,%u,%u,%u,%u,%d,%u,%u,%u,%u,%u,%u,%u,%u\n", direction[pkt_node->direction], pkt_node->hash, (intmax_t)pkt_node->tval.tv_sec, pkt_node->tval.tv_usec, pkt_node->ip_laddr[0], pkt_node->ip_laddr[1], pkt_node->ip_laddr[2], pkt_node->ip_laddr[3], ntohs(pkt_node->tcp_localport), pkt_node->ip_faddr[0], pkt_node->ip_faddr[1], pkt_node->ip_faddr[2], pkt_node->ip_faddr[3], ntohs(pkt_node->tcp_foreignport), pkt_node->snd_ssthresh, pkt_node->snd_cwnd, pkt_node->t_flags2, pkt_node->snd_wnd, pkt_node->rcv_wnd, pkt_node->snd_scale, pkt_node->rcv_scale, pkt_node->conn_state, pkt_node->max_seg_size, pkt_node->smoothed_rtt, pkt_node->sack_enabled, pkt_node->flags, pkt_node->rxt_length, pkt_node->snd_buf_hiwater, pkt_node->snd_buf_cc, pkt_node->rcv_buf_hiwater, pkt_node->rcv_buf_cc, pkt_node->sent_inflight_bytes, pkt_node->t_segqlen, pkt_node->flowid, pkt_node->flowtype); #ifdef SIFTR_IPV6 } #endif alq_post_flags(siftr_alq, log_buf, 0); } static void siftr_pkt_manager_thread(void *arg) { STAILQ_HEAD(pkthead, pkt_node) tmp_pkt_queue = STAILQ_HEAD_INITIALIZER(tmp_pkt_queue); struct pkt_node *pkt_node, *pkt_node_temp; uint8_t draining; draining = 2; mtx_lock(&siftr_pkt_mgr_mtx); /* draining == 0 when queue has been flushed and it's safe to exit. */ while (draining) { /* * Sleep until we are signalled to wake because thread has * been told to exit or until 1 tick has passed. */ mtx_sleep(&wait_for_pkt, &siftr_pkt_mgr_mtx, PWAIT, "pktwait", 1); /* Gain exclusive access to the pkt_node queue. */ mtx_lock(&siftr_pkt_queue_mtx); /* * Move pkt_queue to tmp_pkt_queue, which leaves * pkt_queue empty and ready to receive more pkt_nodes. */ STAILQ_CONCAT(&tmp_pkt_queue, &pkt_queue); /* * We've finished making changes to the list. Unlock it * so the pfil hooks can continue queuing pkt_nodes. */ mtx_unlock(&siftr_pkt_queue_mtx); /* * We can't hold a mutex whilst calling siftr_process_pkt * because ALQ might sleep waiting for buffer space. */ mtx_unlock(&siftr_pkt_mgr_mtx); /* Flush all pkt_nodes to the log file. */ STAILQ_FOREACH_SAFE(pkt_node, &tmp_pkt_queue, nodes, pkt_node_temp) { siftr_process_pkt(pkt_node); STAILQ_REMOVE_HEAD(&tmp_pkt_queue, nodes); free(pkt_node, M_SIFTR_PKTNODE); } KASSERT(STAILQ_EMPTY(&tmp_pkt_queue), ("SIFTR tmp_pkt_queue not empty after flush")); mtx_lock(&siftr_pkt_mgr_mtx); /* * If siftr_exit_pkt_manager_thread gets set during the window * where we are draining the tmp_pkt_queue above, there might * still be pkts in pkt_queue that need to be drained. * Allow one further iteration to occur after * siftr_exit_pkt_manager_thread has been set to ensure * pkt_queue is completely empty before we kill the thread. * * siftr_exit_pkt_manager_thread is set only after the pfil * hooks have been removed, so only 1 extra iteration * is needed to drain the queue. */ if (siftr_exit_pkt_manager_thread) draining--; } mtx_unlock(&siftr_pkt_mgr_mtx); /* Calls wakeup on this thread's struct thread ptr. */ kthread_exit(); } static uint32_t hash_pkt(struct mbuf *m, uint32_t offset) { uint32_t hash; hash = 0; while (m != NULL && offset > m->m_len) { /* * The IP packet payload does not start in this mbuf, so * need to figure out which mbuf it starts in and what offset * into the mbuf's data region the payload starts at. */ offset -= m->m_len; m = m->m_next; } while (m != NULL) { /* Ensure there is data in the mbuf */ if ((m->m_len - offset) > 0) hash = hash32_buf(m->m_data + offset, m->m_len - offset, hash); m = m->m_next; offset = 0; } return (hash); } /* * Check if a given mbuf has the SIFTR mbuf tag. If it does, log the fact that * it's a reinjected packet and return. If it doesn't, tag the mbuf and return. * Return value >0 means the caller should skip processing this mbuf. */ static inline int siftr_chkreinject(struct mbuf *m, int dir, struct siftr_stats *ss) { if (m_tag_locate(m, PACKET_COOKIE_SIFTR, PACKET_TAG_SIFTR, NULL) != NULL) { if (dir == PFIL_IN) ss->nskip_in_dejavu++; else ss->nskip_out_dejavu++; return (1); } else { struct m_tag *tag = m_tag_alloc(PACKET_COOKIE_SIFTR, PACKET_TAG_SIFTR, 0, M_NOWAIT); if (tag == NULL) { if (dir == PFIL_IN) ss->nskip_in_malloc++; else ss->nskip_out_malloc++; return (1); } m_tag_prepend(m, tag); } return (0); } /* * Look up an inpcb for a packet. Return the inpcb pointer if found, or NULL * otherwise. */ static inline struct inpcb * siftr_findinpcb(int ipver, struct ip *ip, struct mbuf *m, uint16_t sport, uint16_t dport, int dir, struct siftr_stats *ss) { struct inpcb *inp; /* We need the tcbinfo lock. */ INP_INFO_WUNLOCK_ASSERT(&V_tcbinfo); if (dir == PFIL_IN) inp = (ipver == INP_IPV4 ? in_pcblookup(&V_tcbinfo, ip->ip_src, sport, ip->ip_dst, dport, INPLOOKUP_RLOCKPCB, m->m_pkthdr.rcvif) : #ifdef SIFTR_IPV6 in6_pcblookup(&V_tcbinfo, &((struct ip6_hdr *)ip)->ip6_src, sport, &((struct ip6_hdr *)ip)->ip6_dst, dport, INPLOOKUP_RLOCKPCB, m->m_pkthdr.rcvif) #else NULL #endif ); else inp = (ipver == INP_IPV4 ? in_pcblookup(&V_tcbinfo, ip->ip_dst, dport, ip->ip_src, sport, INPLOOKUP_RLOCKPCB, m->m_pkthdr.rcvif) : #ifdef SIFTR_IPV6 in6_pcblookup(&V_tcbinfo, &((struct ip6_hdr *)ip)->ip6_dst, dport, &((struct ip6_hdr *)ip)->ip6_src, sport, INPLOOKUP_RLOCKPCB, m->m_pkthdr.rcvif) #else NULL #endif ); /* If we can't find the inpcb, bail. */ if (inp == NULL) { if (dir == PFIL_IN) ss->nskip_in_inpcb++; else ss->nskip_out_inpcb++; } return (inp); } static inline void siftr_siftdata(struct pkt_node *pn, struct inpcb *inp, struct tcpcb *tp, int ipver, int dir, int inp_locally_locked) { #ifdef SIFTR_IPV6 if (ipver == INP_IPV4) { pn->ip_laddr[3] = inp->inp_laddr.s_addr; pn->ip_faddr[3] = inp->inp_faddr.s_addr; #else *((uint32_t *)pn->ip_laddr) = inp->inp_laddr.s_addr; *((uint32_t *)pn->ip_faddr) = inp->inp_faddr.s_addr; #endif #ifdef SIFTR_IPV6 } else { pn->ip_laddr[0] = inp->in6p_laddr.s6_addr32[0]; pn->ip_laddr[1] = inp->in6p_laddr.s6_addr32[1]; pn->ip_laddr[2] = inp->in6p_laddr.s6_addr32[2]; pn->ip_laddr[3] = inp->in6p_laddr.s6_addr32[3]; pn->ip_faddr[0] = inp->in6p_faddr.s6_addr32[0]; pn->ip_faddr[1] = inp->in6p_faddr.s6_addr32[1]; pn->ip_faddr[2] = inp->in6p_faddr.s6_addr32[2]; pn->ip_faddr[3] = inp->in6p_faddr.s6_addr32[3]; } #endif pn->tcp_localport = inp->inp_lport; pn->tcp_foreignport = inp->inp_fport; pn->snd_cwnd = tp->snd_cwnd; pn->snd_wnd = tp->snd_wnd; pn->rcv_wnd = tp->rcv_wnd; pn->t_flags2 = tp->t_flags2; pn->snd_ssthresh = tp->snd_ssthresh; pn->snd_scale = tp->snd_scale; pn->rcv_scale = tp->rcv_scale; pn->conn_state = tp->t_state; pn->max_seg_size = tp->t_maxseg; pn->smoothed_rtt = tp->t_srtt; pn->sack_enabled = (tp->t_flags & TF_SACK_PERMIT) != 0; pn->flags = tp->t_flags; pn->rxt_length = tp->t_rxtcur; pn->snd_buf_hiwater = inp->inp_socket->so_snd.sb_hiwat; pn->snd_buf_cc = sbused(&inp->inp_socket->so_snd); pn->rcv_buf_hiwater = inp->inp_socket->so_rcv.sb_hiwat; pn->rcv_buf_cc = sbused(&inp->inp_socket->so_rcv); pn->sent_inflight_bytes = tp->snd_max - tp->snd_una; pn->t_segqlen = tp->t_segqlen; pn->flowid = inp->inp_flowid; pn->flowtype = inp->inp_flowtype; /* We've finished accessing the tcb so release the lock. */ if (inp_locally_locked) INP_RUNLOCK(inp); pn->ipver = ipver; pn->direction = (dir == PFIL_IN ? DIR_IN : DIR_OUT); /* * Significantly more accurate than using getmicrotime(), but slower! * Gives true microsecond resolution at the expense of a hit to * maximum pps throughput processing when SIFTR is loaded and enabled. */ microtime(&pn->tval); TCP_PROBE1(siftr, &pn); } /* * pfil hook that is called for each IPv4 packet making its way through the * stack in either direction. * The pfil subsystem holds a non-sleepable mutex somewhere when * calling our hook function, so we can't sleep at all. * It's very important to use the M_NOWAIT flag with all function calls * that support it so that they won't sleep, otherwise you get a panic. */ static pfil_return_t siftr_chkpkt(struct mbuf **m, struct ifnet *ifp, int flags, void *ruleset __unused, struct inpcb *inp) { struct pkt_node *pn; struct ip *ip; struct tcphdr *th; struct tcpcb *tp; struct siftr_stats *ss; unsigned int ip_hl; int inp_locally_locked, dir; inp_locally_locked = 0; dir = PFIL_DIR(flags); ss = DPCPU_PTR(ss); /* * m_pullup is not required here because ip_{input|output} * already do the heavy lifting for us. */ ip = mtod(*m, struct ip *); /* Only continue processing if the packet is TCP. */ if (ip->ip_p != IPPROTO_TCP) goto ret; /* * If a kernel subsystem reinjects packets into the stack, our pfil * hook will be called multiple times for the same packet. * Make sure we only process unique packets. */ if (siftr_chkreinject(*m, dir, ss)) goto ret; if (dir == PFIL_IN) ss->n_in++; else ss->n_out++; /* * Create a tcphdr struct starting at the correct offset * in the IP packet. ip->ip_hl gives the ip header length * in 4-byte words, so multiply it to get the size in bytes. */ ip_hl = (ip->ip_hl << 2); th = (struct tcphdr *)((caddr_t)ip + ip_hl); /* * If the pfil hooks don't provide a pointer to the * inpcb, we need to find it ourselves and lock it. */ if (!inp) { /* Find the corresponding inpcb for this pkt. */ inp = siftr_findinpcb(INP_IPV4, ip, *m, th->th_sport, th->th_dport, dir, ss); if (inp == NULL) goto ret; else inp_locally_locked = 1; } INP_LOCK_ASSERT(inp); /* Find the TCP control block that corresponds with this packet */ tp = intotcpcb(inp); /* * If we can't find the TCP control block (happens occasionaly for a - * packet sent during the shutdown phase of a TCP connection), - * or we're in the timewait state, bail + * packet sent during the shutdown phase of a TCP connection), bail */ - if (tp == NULL || inp->inp_flags & INP_TIMEWAIT) { + if (tp == NULL) { if (dir == PFIL_IN) ss->nskip_in_tcpcb++; else ss->nskip_out_tcpcb++; goto inp_unlock; } /* * Only pkts selected by the tcp port filter * can be inserted into the pkt_queue */ if ((siftr_port_filter != 0) && (siftr_port_filter != ntohs(inp->inp_lport)) && (siftr_port_filter != ntohs(inp->inp_fport))) { goto inp_unlock; } pn = malloc(sizeof(struct pkt_node), M_SIFTR_PKTNODE, M_NOWAIT|M_ZERO); if (pn == NULL) { if (dir == PFIL_IN) ss->nskip_in_malloc++; else ss->nskip_out_malloc++; goto inp_unlock; } siftr_siftdata(pn, inp, tp, INP_IPV4, dir, inp_locally_locked); if (siftr_generate_hashes) { if ((*m)->m_pkthdr.csum_flags & CSUM_TCP) { /* * For outbound packets, the TCP checksum isn't * calculated yet. This is a problem for our packet * hashing as the receiver will calc a different hash * to ours if we don't include the correct TCP checksum * in the bytes being hashed. To work around this * problem, we manually calc the TCP checksum here in * software. We unset the CSUM_TCP flag so the lower * layers don't recalc it. */ (*m)->m_pkthdr.csum_flags &= ~CSUM_TCP; /* * Calculate the TCP checksum in software and assign * to correct TCP header field, which will follow the * packet mbuf down the stack. The trick here is that * tcp_output() sets th->th_sum to the checksum of the * pseudo header for us already. Because of the nature * of the checksumming algorithm, we can sum over the * entire IP payload (i.e. TCP header and data), which * will include the already calculated pseduo header * checksum, thus giving us the complete TCP checksum. * * To put it in simple terms, if checksum(1,2,3,4)=10, * then checksum(1,2,3,4,5) == checksum(10,5). * This property is what allows us to "cheat" and * checksum only the IP payload which has the TCP * th_sum field populated with the pseudo header's * checksum, and not need to futz around checksumming * pseudo header bytes and TCP header/data in one hit. * Refer to RFC 1071 for more info. * * NB: in_cksum_skip(struct mbuf *m, int len, int skip) * in_cksum_skip 2nd argument is NOT the number of * bytes to read from the mbuf at "skip" bytes offset * from the start of the mbuf (very counter intuitive!). * The number of bytes to read is calculated internally * by the function as len-skip i.e. to sum over the IP * payload (TCP header + data) bytes, it is INCORRECT * to call the function like this: * in_cksum_skip(at, ip->ip_len - offset, offset) * Rather, it should be called like this: * in_cksum_skip(at, ip->ip_len, offset) * which means read "ip->ip_len - offset" bytes from * the mbuf cluster "at" at offset "offset" bytes from * the beginning of the "at" mbuf's data pointer. */ th->th_sum = in_cksum_skip(*m, ntohs(ip->ip_len), ip_hl); } /* * XXX: Having to calculate the checksum in software and then * hash over all bytes is really inefficient. Would be nice to * find a way to create the hash and checksum in the same pass * over the bytes. */ pn->hash = hash_pkt(*m, ip_hl); } mtx_lock(&siftr_pkt_queue_mtx); STAILQ_INSERT_TAIL(&pkt_queue, pn, nodes); mtx_unlock(&siftr_pkt_queue_mtx); goto ret; inp_unlock: if (inp_locally_locked) INP_RUNLOCK(inp); ret: return (PFIL_PASS); } #ifdef SIFTR_IPV6 static pfil_return_t siftr_chkpkt6(struct mbuf **m, struct ifnet *ifp, int flags, void *ruleset __unused, struct inpcb *inp) { struct pkt_node *pn; struct ip6_hdr *ip6; struct tcphdr *th; struct tcpcb *tp; struct siftr_stats *ss; unsigned int ip6_hl; int inp_locally_locked, dir; inp_locally_locked = 0; dir = PFIL_DIR(flags); ss = DPCPU_PTR(ss); /* * m_pullup is not required here because ip6_{input|output} * already do the heavy lifting for us. */ ip6 = mtod(*m, struct ip6_hdr *); /* * Only continue processing if the packet is TCP * XXX: We should follow the next header fields * as shown on Pg 6 RFC 2460, but right now we'll * only check pkts that have no extension headers. */ if (ip6->ip6_nxt != IPPROTO_TCP) goto ret6; /* * If a kernel subsystem reinjects packets into the stack, our pfil * hook will be called multiple times for the same packet. * Make sure we only process unique packets. */ if (siftr_chkreinject(*m, dir, ss)) goto ret6; if (dir == PFIL_IN) ss->n_in++; else ss->n_out++; ip6_hl = sizeof(struct ip6_hdr); /* * Create a tcphdr struct starting at the correct offset * in the ipv6 packet. ip->ip_hl gives the ip header length * in 4-byte words, so multiply it to get the size in bytes. */ th = (struct tcphdr *)((caddr_t)ip6 + ip6_hl); /* * For inbound packets, the pfil hooks don't provide a pointer to the * inpcb, so we need to find it ourselves and lock it. */ if (!inp) { /* Find the corresponding inpcb for this pkt. */ inp = siftr_findinpcb(INP_IPV6, (struct ip *)ip6, *m, th->th_sport, th->th_dport, dir, ss); if (inp == NULL) goto ret6; else inp_locally_locked = 1; } /* Find the TCP control block that corresponds with this packet. */ tp = intotcpcb(inp); /* * If we can't find the TCP control block (happens occasionaly for a - * packet sent during the shutdown phase of a TCP connection), - * or we're in the timewait state, bail. + * packet sent during the shutdown phase of a TCP connection), bail */ - if (tp == NULL || inp->inp_flags & INP_TIMEWAIT) { + if (tp == NULL) { if (dir == PFIL_IN) ss->nskip_in_tcpcb++; else ss->nskip_out_tcpcb++; goto inp_unlock6; } /* * Only pkts selected by the tcp port filter * can be inserted into the pkt_queue */ if ((siftr_port_filter != 0) && (siftr_port_filter != ntohs(inp->inp_lport)) && (siftr_port_filter != ntohs(inp->inp_fport))) { goto inp_unlock6; } pn = malloc(sizeof(struct pkt_node), M_SIFTR_PKTNODE, M_NOWAIT|M_ZERO); if (pn == NULL) { if (dir == PFIL_IN) ss->nskip_in_malloc++; else ss->nskip_out_malloc++; goto inp_unlock6; } siftr_siftdata(pn, inp, tp, INP_IPV6, dir, inp_locally_locked); /* XXX: Figure out how to generate hashes for IPv6 packets. */ mtx_lock(&siftr_pkt_queue_mtx); STAILQ_INSERT_TAIL(&pkt_queue, pn, nodes); mtx_unlock(&siftr_pkt_queue_mtx); goto ret6; inp_unlock6: if (inp_locally_locked) INP_RUNLOCK(inp); ret6: return (PFIL_PASS); } #endif /* #ifdef SIFTR_IPV6 */ VNET_DEFINE_STATIC(pfil_hook_t, siftr_inet_hook); #define V_siftr_inet_hook VNET(siftr_inet_hook) #ifdef SIFTR_IPV6 VNET_DEFINE_STATIC(pfil_hook_t, siftr_inet6_hook); #define V_siftr_inet6_hook VNET(siftr_inet6_hook) #endif static int siftr_pfil(int action) { struct pfil_hook_args pha; struct pfil_link_args pla; pha.pa_version = PFIL_VERSION; pha.pa_flags = PFIL_IN | PFIL_OUT; pha.pa_modname = "siftr"; pha.pa_ruleset = NULL; pha.pa_rulname = "default"; pla.pa_version = PFIL_VERSION; pla.pa_flags = PFIL_IN | PFIL_OUT | PFIL_HEADPTR | PFIL_HOOKPTR; VNET_ITERATOR_DECL(vnet_iter); VNET_LIST_RLOCK(); VNET_FOREACH(vnet_iter) { CURVNET_SET(vnet_iter); if (action == HOOK) { pha.pa_func = siftr_chkpkt; pha.pa_type = PFIL_TYPE_IP4; V_siftr_inet_hook = pfil_add_hook(&pha); pla.pa_hook = V_siftr_inet_hook; pla.pa_head = V_inet_pfil_head; (void)pfil_link(&pla); #ifdef SIFTR_IPV6 pha.pa_func = siftr_chkpkt6; pha.pa_type = PFIL_TYPE_IP6; V_siftr_inet6_hook = pfil_add_hook(&pha); pla.pa_hook = V_siftr_inet6_hook; pla.pa_head = V_inet6_pfil_head; (void)pfil_link(&pla); #endif } else if (action == UNHOOK) { pfil_remove_hook(V_siftr_inet_hook); #ifdef SIFTR_IPV6 pfil_remove_hook(V_siftr_inet6_hook); #endif } CURVNET_RESTORE(); } VNET_LIST_RUNLOCK(); return (0); } static int siftr_sysctl_logfile_name_handler(SYSCTL_HANDLER_ARGS) { struct alq *new_alq; int error; error = sysctl_handle_string(oidp, arg1, arg2, req); /* Check for error or same filename */ if (error != 0 || req->newptr == NULL || strncmp(siftr_logfile, arg1, arg2) == 0) goto done; /* Filname changed */ error = alq_open(&new_alq, arg1, curthread->td_ucred, SIFTR_LOG_FILE_MODE, SIFTR_ALQ_BUFLEN, 0); if (error != 0) goto done; /* * If disabled, siftr_alq == NULL so we simply close * the alq as we've proved it can be opened. * If enabled, close the existing alq and switch the old * for the new. */ if (siftr_alq == NULL) { alq_close(new_alq); } else { alq_close(siftr_alq); siftr_alq = new_alq; } /* Update filename upon success */ strlcpy(siftr_logfile, arg1, arg2); done: return (error); } static int siftr_manage_ops(uint8_t action) { struct siftr_stats totalss; struct timeval tval; struct flow_hash_node *counter, *tmp_counter; struct sbuf *s; int i, key_index, error; uint32_t bytes_to_write, total_skipped_pkts; uint16_t lport, fport; uint8_t *key, ipver __unused; #ifdef SIFTR_IPV6 uint32_t laddr[4]; uint32_t faddr[4]; #else uint8_t laddr[4]; uint8_t faddr[4]; #endif error = 0; total_skipped_pkts = 0; /* Init an autosizing sbuf that initially holds 200 chars. */ if ((s = sbuf_new(NULL, NULL, 200, SBUF_AUTOEXTEND)) == NULL) return (-1); if (action == SIFTR_ENABLE && siftr_pkt_manager_thr == NULL) { /* * Create our alq * XXX: We should abort if alq_open fails! */ alq_open(&siftr_alq, siftr_logfile, curthread->td_ucred, SIFTR_LOG_FILE_MODE, SIFTR_ALQ_BUFLEN, 0); STAILQ_INIT(&pkt_queue); DPCPU_ZERO(ss); siftr_exit_pkt_manager_thread = 0; kthread_add(&siftr_pkt_manager_thread, NULL, NULL, &siftr_pkt_manager_thr, RFNOWAIT, 0, "siftr_pkt_manager_thr"); siftr_pfil(HOOK); microtime(&tval); sbuf_printf(s, "enable_time_secs=%jd\tenable_time_usecs=%06ld\t" "siftrver=%s\thz=%u\ttcp_rtt_scale=%u\tsysname=%s\t" "sysver=%u\tipmode=%u\n", (intmax_t)tval.tv_sec, tval.tv_usec, MODVERSION_STR, hz, TCP_RTT_SCALE, SYS_NAME, __FreeBSD_version, SIFTR_IPMODE); sbuf_finish(s); alq_writen(siftr_alq, sbuf_data(s), sbuf_len(s), ALQ_WAITOK); } else if (action == SIFTR_DISABLE && siftr_pkt_manager_thr != NULL) { /* * Remove the pfil hook functions. All threads currently in * the hook functions are allowed to exit before siftr_pfil() * returns. */ siftr_pfil(UNHOOK); /* This will block until the pkt manager thread unlocks it. */ mtx_lock(&siftr_pkt_mgr_mtx); /* Tell the pkt manager thread that it should exit now. */ siftr_exit_pkt_manager_thread = 1; /* * Wake the pkt_manager thread so it realises that * siftr_exit_pkt_manager_thread == 1 and exits gracefully. * The wakeup won't be delivered until we unlock * siftr_pkt_mgr_mtx so this isn't racy. */ wakeup(&wait_for_pkt); /* Wait for the pkt_manager thread to exit. */ mtx_sleep(siftr_pkt_manager_thr, &siftr_pkt_mgr_mtx, PWAIT, "thrwait", 0); siftr_pkt_manager_thr = NULL; mtx_unlock(&siftr_pkt_mgr_mtx); totalss.n_in = DPCPU_VARSUM(ss, n_in); totalss.n_out = DPCPU_VARSUM(ss, n_out); totalss.nskip_in_malloc = DPCPU_VARSUM(ss, nskip_in_malloc); totalss.nskip_out_malloc = DPCPU_VARSUM(ss, nskip_out_malloc); totalss.nskip_in_mtx = DPCPU_VARSUM(ss, nskip_in_mtx); totalss.nskip_out_mtx = DPCPU_VARSUM(ss, nskip_out_mtx); totalss.nskip_in_tcpcb = DPCPU_VARSUM(ss, nskip_in_tcpcb); totalss.nskip_out_tcpcb = DPCPU_VARSUM(ss, nskip_out_tcpcb); totalss.nskip_in_inpcb = DPCPU_VARSUM(ss, nskip_in_inpcb); totalss.nskip_out_inpcb = DPCPU_VARSUM(ss, nskip_out_inpcb); total_skipped_pkts = totalss.nskip_in_malloc + totalss.nskip_out_malloc + totalss.nskip_in_mtx + totalss.nskip_out_mtx + totalss.nskip_in_tcpcb + totalss.nskip_out_tcpcb + totalss.nskip_in_inpcb + totalss.nskip_out_inpcb; microtime(&tval); sbuf_printf(s, "disable_time_secs=%jd\tdisable_time_usecs=%06ld\t" "num_inbound_tcp_pkts=%ju\tnum_outbound_tcp_pkts=%ju\t" "total_tcp_pkts=%ju\tnum_inbound_skipped_pkts_malloc=%u\t" "num_outbound_skipped_pkts_malloc=%u\t" "num_inbound_skipped_pkts_mtx=%u\t" "num_outbound_skipped_pkts_mtx=%u\t" "num_inbound_skipped_pkts_tcpcb=%u\t" "num_outbound_skipped_pkts_tcpcb=%u\t" "num_inbound_skipped_pkts_inpcb=%u\t" "num_outbound_skipped_pkts_inpcb=%u\t" "total_skipped_tcp_pkts=%u\tflow_list=", (intmax_t)tval.tv_sec, tval.tv_usec, (uintmax_t)totalss.n_in, (uintmax_t)totalss.n_out, (uintmax_t)(totalss.n_in + totalss.n_out), totalss.nskip_in_malloc, totalss.nskip_out_malloc, totalss.nskip_in_mtx, totalss.nskip_out_mtx, totalss.nskip_in_tcpcb, totalss.nskip_out_tcpcb, totalss.nskip_in_inpcb, totalss.nskip_out_inpcb, total_skipped_pkts); /* * Iterate over the flow hash, printing a summary of each * flow seen and freeing any malloc'd memory. * The hash consists of an array of LISTs (man 3 queue). */ for (i = 0; i <= siftr_hashmask; i++) { LIST_FOREACH_SAFE(counter, counter_hash + i, nodes, tmp_counter) { key = counter->key; key_index = 1; ipver = key[0]; memcpy(laddr, key + key_index, sizeof(laddr)); key_index += sizeof(laddr); memcpy(&lport, key + key_index, sizeof(lport)); key_index += sizeof(lport); memcpy(faddr, key + key_index, sizeof(faddr)); key_index += sizeof(faddr); memcpy(&fport, key + key_index, sizeof(fport)); #ifdef SIFTR_IPV6 laddr[3] = ntohl(laddr[3]); faddr[3] = ntohl(faddr[3]); if (ipver == INP_IPV6) { laddr[0] = ntohl(laddr[0]); laddr[1] = ntohl(laddr[1]); laddr[2] = ntohl(laddr[2]); faddr[0] = ntohl(faddr[0]); faddr[1] = ntohl(faddr[1]); faddr[2] = ntohl(faddr[2]); sbuf_printf(s, "%x:%x:%x:%x:%x:%x:%x:%x;%u-" "%x:%x:%x:%x:%x:%x:%x:%x;%u,", UPPER_SHORT(laddr[0]), LOWER_SHORT(laddr[0]), UPPER_SHORT(laddr[1]), LOWER_SHORT(laddr[1]), UPPER_SHORT(laddr[2]), LOWER_SHORT(laddr[2]), UPPER_SHORT(laddr[3]), LOWER_SHORT(laddr[3]), ntohs(lport), UPPER_SHORT(faddr[0]), LOWER_SHORT(faddr[0]), UPPER_SHORT(faddr[1]), LOWER_SHORT(faddr[1]), UPPER_SHORT(faddr[2]), LOWER_SHORT(faddr[2]), UPPER_SHORT(faddr[3]), LOWER_SHORT(faddr[3]), ntohs(fport)); } else { laddr[0] = FIRST_OCTET(laddr[3]); laddr[1] = SECOND_OCTET(laddr[3]); laddr[2] = THIRD_OCTET(laddr[3]); laddr[3] = FOURTH_OCTET(laddr[3]); faddr[0] = FIRST_OCTET(faddr[3]); faddr[1] = SECOND_OCTET(faddr[3]); faddr[2] = THIRD_OCTET(faddr[3]); faddr[3] = FOURTH_OCTET(faddr[3]); #endif sbuf_printf(s, "%u.%u.%u.%u;%u-%u.%u.%u.%u;%u,", laddr[0], laddr[1], laddr[2], laddr[3], ntohs(lport), faddr[0], faddr[1], faddr[2], faddr[3], ntohs(fport)); #ifdef SIFTR_IPV6 } #endif free(counter, M_SIFTR_HASHNODE); } LIST_INIT(counter_hash + i); } sbuf_printf(s, "\n"); sbuf_finish(s); i = 0; do { bytes_to_write = min(SIFTR_ALQ_BUFLEN, sbuf_len(s)-i); alq_writen(siftr_alq, sbuf_data(s)+i, bytes_to_write, ALQ_WAITOK); i += bytes_to_write; } while (i < sbuf_len(s)); alq_close(siftr_alq); siftr_alq = NULL; } else error = EINVAL; sbuf_delete(s); /* * XXX: Should be using ret to check if any functions fail * and set error appropriately */ return (error); } static int siftr_sysctl_enabled_handler(SYSCTL_HANDLER_ARGS) { int error; uint32_t new; new = siftr_enabled; error = sysctl_handle_int(oidp, &new, 0, req); if (error == 0 && req->newptr != NULL) { if (new > 1) return (EINVAL); else if (new != siftr_enabled) { if ((error = siftr_manage_ops(new)) == 0) { siftr_enabled = new; } else { siftr_manage_ops(SIFTR_DISABLE); } } } return (error); } static void siftr_shutdown_handler(void *arg) { if (siftr_enabled == 1) { siftr_manage_ops(SIFTR_DISABLE); } } /* * Module is being unloaded or machine is shutting down. Take care of cleanup. */ static int deinit_siftr(void) { /* Cleanup. */ siftr_manage_ops(SIFTR_DISABLE); hashdestroy(counter_hash, M_SIFTR, siftr_hashmask); mtx_destroy(&siftr_pkt_queue_mtx); mtx_destroy(&siftr_pkt_mgr_mtx); return (0); } /* * Module has just been loaded into the kernel. */ static int init_siftr(void) { EVENTHANDLER_REGISTER(shutdown_pre_sync, siftr_shutdown_handler, NULL, SHUTDOWN_PRI_FIRST); /* Initialise our flow counter hash table. */ counter_hash = hashinit(SIFTR_EXPECTED_MAX_TCP_FLOWS, M_SIFTR, &siftr_hashmask); mtx_init(&siftr_pkt_queue_mtx, "siftr_pkt_queue_mtx", NULL, MTX_DEF); mtx_init(&siftr_pkt_mgr_mtx, "siftr_pkt_mgr_mtx", NULL, MTX_DEF); /* Print message to the user's current terminal. */ uprintf("\nStatistical Information For TCP Research (SIFTR) %s\n" " http://caia.swin.edu.au/urp/newtcp\n\n", MODVERSION_STR); return (0); } /* * This is the function that is called to load and unload the module. * When the module is loaded, this function is called once with * "what" == MOD_LOAD * When the module is unloaded, this function is called twice with * "what" = MOD_QUIESCE first, followed by "what" = MOD_UNLOAD second * When the system is shut down e.g. CTRL-ALT-DEL or using the shutdown command, * this function is called once with "what" = MOD_SHUTDOWN * When the system is shut down, the handler isn't called until the very end * of the shutdown sequence i.e. after the disks have been synced. */ static int siftr_load_handler(module_t mod, int what, void *arg) { int ret; switch (what) { case MOD_LOAD: ret = init_siftr(); break; case MOD_QUIESCE: case MOD_SHUTDOWN: ret = deinit_siftr(); break; case MOD_UNLOAD: ret = 0; break; default: ret = EINVAL; break; } return (ret); } static moduledata_t siftr_mod = { .name = "siftr", .evhand = siftr_load_handler, }; /* * Param 1: name of the kernel module * Param 2: moduledata_t struct containing info about the kernel module * and the execution entry point for the module * Param 3: From sysinit_sub_id enumeration in /usr/include/sys/kernel.h * Defines the module initialisation order * Param 4: From sysinit_elem_order enumeration in /usr/include/sys/kernel.h * Defines the initialisation order of this kld relative to others * within the same subsystem as defined by param 3 */ DECLARE_MODULE(siftr, siftr_mod, SI_SUB_LAST, SI_ORDER_ANY); MODULE_DEPEND(siftr, alq, 1, 1, 1); MODULE_VERSION(siftr, MODVERSION); diff --git a/sys/netinet/tcp_hpts.c b/sys/netinet/tcp_hpts.c index 828ce16e8dad..3ed0c1e27c58 100644 --- a/sys/netinet/tcp_hpts.c +++ b/sys/netinet/tcp_hpts.c @@ -1,2043 +1,2043 @@ /*- * Copyright (c) 2016-2018 Netflix, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include "opt_rss.h" #include "opt_tcpdebug.h" /** * Some notes about usage. * * The tcp_hpts system is designed to provide a high precision timer * system for tcp. Its main purpose is to provide a mechanism for * pacing packets out onto the wire. It can be used in two ways * by a given TCP stack (and those two methods can be used simultaneously). * * First, and probably the main thing its used by Rack and BBR, it can * be used to call tcp_output() of a transport stack at some time in the future. * The normal way this is done is that tcp_output() of the stack schedules * itself to be called again by calling tcp_hpts_insert(tcpcb, slot). The * slot is the time from now that the stack wants to be called but it * must be converted to tcp_hpts's notion of slot. This is done with * one of the macros HPTS_MS_TO_SLOTS or HPTS_USEC_TO_SLOTS. So a typical * call from the tcp_output() routine might look like: * * tcp_hpts_insert(tp, HPTS_USEC_TO_SLOTS(550)); * * The above would schedule tcp_ouput() to be called in 550 useconds. * Note that if using this mechanism the stack will want to add near * its top a check to prevent unwanted calls (from user land or the * arrival of incoming ack's). So it would add something like: * * if (tcp_in_hpts(inp)) * return; * * to prevent output processing until the time alotted has gone by. * Of course this is a bare bones example and the stack will probably * have more consideration then just the above. * * In order to run input queued segments from the HPTS context the * tcp stack must define an input function for * tfb_do_queued_segments(). This function understands * how to dequeue a array of packets that were input and * knows how to call the correct processing routine. * * Locking in this is important as well so most likely the * stack will need to define the tfb_do_segment_nounlock() * splitting tfb_do_segment() into two parts. The main processing * part that does not unlock the INP and returns a value of 1 or 0. * It returns 0 if all is well and the lock was not released. It * returns 1 if we had to destroy the TCB (a reset received etc). * The remains of tfb_do_segment() then become just a simple call * to the tfb_do_segment_nounlock() function and check the return * code and possibly unlock. * * The stack must also set the flag on the INP that it supports this * feature i.e. INP_SUPPORTS_MBUFQ. The LRO code recoginizes * this flag as well and will queue packets when it is set. * There are other flags as well INP_MBUF_QUEUE_READY and * INP_DONT_SACK_QUEUE. The first flag tells the LRO code * that we are in the pacer for output so there is no * need to wake up the hpts system to get immediate * input. The second tells the LRO code that its okay * if a SACK arrives you can still defer input and let * the current hpts timer run (this is usually set when * a rack timer is up so we know SACK's are happening * on the connection already and don't want to wakeup yet). * * There is a common functions within the rack_bbr_common code * version i.e. ctf_do_queued_segments(). This function * knows how to take the input queue of packets from * tp->t_in_pkts and process them digging out * all the arguments, calling any bpf tap and * calling into tfb_do_segment_nounlock(). The common * function (ctf_do_queued_segments()) requires that * you have defined the tfb_do_segment_nounlock() as * described above. */ #include #include #include #include #include #include #include #include #include /* for proc0 declaration */ #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef RSS #include #include #endif #define TCPSTATES /* for logging */ #include #include #include #include #include /* required for icmp_var.h */ #include /* for ICMP_BANDLIM */ #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef tcpdebug #include #endif /* tcpdebug */ #ifdef tcp_offload #include #endif /* * The hpts uses a 102400 wheel. The wheel * defines the time in 10 usec increments (102400 x 10). * This gives a range of 10usec - 1024ms to place * an entry within. If the user requests more than * 1.024 second, a remaineder is attached and the hpts * when seeing the remainder will re-insert the * inpcb forward in time from where it is until * the remainder is zero. */ #define NUM_OF_HPTSI_SLOTS 102400 /* Each hpts has its own p_mtx which is used for locking */ #define HPTS_MTX_ASSERT(hpts) mtx_assert(&(hpts)->p_mtx, MA_OWNED) #define HPTS_LOCK(hpts) mtx_lock(&(hpts)->p_mtx) #define HPTS_UNLOCK(hpts) mtx_unlock(&(hpts)->p_mtx) struct tcp_hpts_entry { /* Cache line 0x00 */ struct mtx p_mtx; /* Mutex for hpts */ struct timeval p_mysleep; /* Our min sleep time */ uint64_t syscall_cnt; uint64_t sleeping; /* What the actual sleep was (if sleeping) */ uint16_t p_hpts_active; /* Flag that says hpts is awake */ uint8_t p_wheel_complete; /* have we completed the wheel arc walk? */ uint32_t p_curtick; /* Tick in 10 us the hpts is going to */ uint32_t p_runningslot; /* Current tick we are at if we are running */ uint32_t p_prev_slot; /* Previous slot we were on */ uint32_t p_cur_slot; /* Current slot in wheel hpts is draining */ uint32_t p_nxt_slot; /* The next slot outside the current range of * slots that the hpts is running on. */ int32_t p_on_queue_cnt; /* Count on queue in this hpts */ uint32_t p_lasttick; /* Last tick before the current one */ uint8_t p_direct_wake :1, /* boolean */ p_on_min_sleep:1, /* boolean */ p_hpts_wake_scheduled:1, /* boolean */ p_avail:5; uint8_t p_fill[3]; /* Fill to 32 bits */ /* Cache line 0x40 */ struct hptsh { TAILQ_HEAD(, inpcb) head; uint32_t count; uint32_t gencnt; } *p_hptss; /* Hptsi wheel */ uint32_t p_hpts_sleep_time; /* Current sleep interval having a max * of 255ms */ uint32_t overidden_sleep; /* what was overrided by min-sleep for logging */ uint32_t saved_lasttick; /* for logging */ uint32_t saved_curtick; /* for logging */ uint32_t saved_curslot; /* for logging */ uint32_t saved_prev_slot; /* for logging */ uint32_t p_delayed_by; /* How much were we delayed by */ /* Cache line 0x80 */ struct sysctl_ctx_list hpts_ctx; struct sysctl_oid *hpts_root; struct intr_event *ie; void *ie_cookie; uint16_t p_num; /* The hpts number one per cpu */ uint16_t p_cpu; /* The hpts CPU */ /* There is extra space in here */ /* Cache line 0x100 */ struct callout co __aligned(CACHE_LINE_SIZE); } __aligned(CACHE_LINE_SIZE); static struct tcp_hptsi { struct cpu_group **grps; struct tcp_hpts_entry **rp_ent; /* Array of hptss */ uint32_t *cts_last_ran; uint32_t grp_cnt; uint32_t rp_num_hptss; /* Number of hpts threads */ } tcp_pace; MALLOC_DEFINE(M_TCPHPTS, "tcp_hpts", "TCP hpts"); #ifdef RSS static int tcp_bind_threads = 1; #else static int tcp_bind_threads = 2; #endif static int tcp_use_irq_cpu = 0; static uint32_t *cts_last_ran; static int hpts_does_tp_logging = 0; static int32_t tcp_hptsi(struct tcp_hpts_entry *hpts, int from_callout); static void tcp_hpts_thread(void *ctx); static void tcp_init_hptsi(void *st); int32_t tcp_min_hptsi_time = DEFAULT_MIN_SLEEP; static int conn_cnt_thresh = DEFAULT_CONNECTION_THESHOLD; static int32_t dynamic_min_sleep = DYNAMIC_MIN_SLEEP; static int32_t dynamic_max_sleep = DYNAMIC_MAX_SLEEP; SYSCTL_NODE(_net_inet_tcp, OID_AUTO, hpts, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "TCP Hpts controls"); SYSCTL_NODE(_net_inet_tcp_hpts, OID_AUTO, stats, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "TCP Hpts statistics"); #define timersub(tvp, uvp, vvp) \ do { \ (vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec; \ (vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec; \ if ((vvp)->tv_usec < 0) { \ (vvp)->tv_sec--; \ (vvp)->tv_usec += 1000000; \ } \ } while (0) static int32_t tcp_hpts_precision = 120; static struct hpts_domain_info { int count; int cpu[MAXCPU]; } hpts_domains[MAXMEMDOM]; enum { IHPTS_NONE = 0, IHPTS_ONQUEUE, IHPTS_MOVING, }; counter_u64_t hpts_hopelessly_behind; SYSCTL_COUNTER_U64(_net_inet_tcp_hpts_stats, OID_AUTO, hopeless, CTLFLAG_RD, &hpts_hopelessly_behind, "Number of times hpts could not catch up and was behind hopelessly"); counter_u64_t hpts_loops; SYSCTL_COUNTER_U64(_net_inet_tcp_hpts_stats, OID_AUTO, loops, CTLFLAG_RD, &hpts_loops, "Number of times hpts had to loop to catch up"); counter_u64_t back_tosleep; SYSCTL_COUNTER_U64(_net_inet_tcp_hpts_stats, OID_AUTO, no_tcbsfound, CTLFLAG_RD, &back_tosleep, "Number of times hpts found no tcbs"); counter_u64_t combined_wheel_wrap; SYSCTL_COUNTER_U64(_net_inet_tcp_hpts_stats, OID_AUTO, comb_wheel_wrap, CTLFLAG_RD, &combined_wheel_wrap, "Number of times the wheel lagged enough to have an insert see wrap"); counter_u64_t wheel_wrap; SYSCTL_COUNTER_U64(_net_inet_tcp_hpts_stats, OID_AUTO, wheel_wrap, CTLFLAG_RD, &wheel_wrap, "Number of times the wheel lagged enough to have an insert see wrap"); counter_u64_t hpts_direct_call; SYSCTL_COUNTER_U64(_net_inet_tcp_hpts_stats, OID_AUTO, direct_call, CTLFLAG_RD, &hpts_direct_call, "Number of times hpts was called by syscall/trap or other entry"); counter_u64_t hpts_wake_timeout; SYSCTL_COUNTER_U64(_net_inet_tcp_hpts_stats, OID_AUTO, timeout_wakeup, CTLFLAG_RD, &hpts_wake_timeout, "Number of times hpts threads woke up via the callout expiring"); counter_u64_t hpts_direct_awakening; SYSCTL_COUNTER_U64(_net_inet_tcp_hpts_stats, OID_AUTO, direct_awakening, CTLFLAG_RD, &hpts_direct_awakening, "Number of times hpts threads woke up via the callout expiring"); counter_u64_t hpts_back_tosleep; SYSCTL_COUNTER_U64(_net_inet_tcp_hpts_stats, OID_AUTO, back_tosleep, CTLFLAG_RD, &hpts_back_tosleep, "Number of times hpts threads woke up via the callout expiring and went back to sleep no work"); counter_u64_t cpu_uses_flowid; counter_u64_t cpu_uses_random; SYSCTL_COUNTER_U64(_net_inet_tcp_hpts_stats, OID_AUTO, cpusel_flowid, CTLFLAG_RD, &cpu_uses_flowid, "Number of times when setting cpuid we used the flowid field"); SYSCTL_COUNTER_U64(_net_inet_tcp_hpts_stats, OID_AUTO, cpusel_random, CTLFLAG_RD, &cpu_uses_random, "Number of times when setting cpuid we used the a random value"); TUNABLE_INT("net.inet.tcp.bind_hptss", &tcp_bind_threads); TUNABLE_INT("net.inet.tcp.use_irq", &tcp_use_irq_cpu); SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, bind_hptss, CTLFLAG_RD, &tcp_bind_threads, 2, "Thread Binding tunable"); SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, use_irq, CTLFLAG_RD, &tcp_use_irq_cpu, 0, "Use of irq CPU tunable"); SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, precision, CTLFLAG_RW, &tcp_hpts_precision, 120, "Value for PRE() precision of callout"); SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, cnt_thresh, CTLFLAG_RW, &conn_cnt_thresh, 0, "How many connections (below) make us use the callout based mechanism"); SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, logging, CTLFLAG_RW, &hpts_does_tp_logging, 0, "Do we add to any tp that has logging on pacer logs"); SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, dyn_minsleep, CTLFLAG_RW, &dynamic_min_sleep, 250, "What is the dynamic minsleep value?"); SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, dyn_maxsleep, CTLFLAG_RW, &dynamic_max_sleep, 5000, "What is the dynamic maxsleep value?"); static int32_t max_pacer_loops = 10; SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, loopmax, CTLFLAG_RW, &max_pacer_loops, 10, "What is the maximum number of times the pacer will loop trying to catch up"); #define HPTS_MAX_SLEEP_ALLOWED (NUM_OF_HPTSI_SLOTS/2) static uint32_t hpts_sleep_max = HPTS_MAX_SLEEP_ALLOWED; static int sysctl_net_inet_tcp_hpts_max_sleep(SYSCTL_HANDLER_ARGS) { int error; uint32_t new; new = hpts_sleep_max; error = sysctl_handle_int(oidp, &new, 0, req); if (error == 0 && req->newptr) { if ((new < (dynamic_min_sleep/HPTS_TICKS_PER_SLOT)) || (new > HPTS_MAX_SLEEP_ALLOWED)) error = EINVAL; else hpts_sleep_max = new; } return (error); } static int sysctl_net_inet_tcp_hpts_min_sleep(SYSCTL_HANDLER_ARGS) { int error; uint32_t new; new = tcp_min_hptsi_time; error = sysctl_handle_int(oidp, &new, 0, req); if (error == 0 && req->newptr) { if (new < LOWEST_SLEEP_ALLOWED) error = EINVAL; else tcp_min_hptsi_time = new; } return (error); } SYSCTL_PROC(_net_inet_tcp_hpts, OID_AUTO, maxsleep, CTLTYPE_UINT | CTLFLAG_RW, &hpts_sleep_max, 0, &sysctl_net_inet_tcp_hpts_max_sleep, "IU", "Maximum time hpts will sleep in slots"); SYSCTL_PROC(_net_inet_tcp_hpts, OID_AUTO, minsleep, CTLTYPE_UINT | CTLFLAG_RW, &tcp_min_hptsi_time, 0, &sysctl_net_inet_tcp_hpts_min_sleep, "IU", "The minimum time the hpts must sleep before processing more slots"); static int ticks_indicate_more_sleep = TICKS_INDICATE_MORE_SLEEP; static int ticks_indicate_less_sleep = TICKS_INDICATE_LESS_SLEEP; static int tcp_hpts_no_wake_over_thresh = 1; SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, more_sleep, CTLFLAG_RW, &ticks_indicate_more_sleep, 0, "If we only process this many or less on a timeout, we need longer sleep on the next callout"); SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, less_sleep, CTLFLAG_RW, &ticks_indicate_less_sleep, 0, "If we process this many or more on a timeout, we need less sleep on the next callout"); SYSCTL_INT(_net_inet_tcp_hpts, OID_AUTO, nowake_over_thresh, CTLFLAG_RW, &tcp_hpts_no_wake_over_thresh, 0, "When we are over the threshold on the pacer do we prohibit wakeups?"); static void tcp_hpts_log(struct tcp_hpts_entry *hpts, struct tcpcb *tp, struct timeval *tv, int slots_to_run, int idx, int from_callout) { union tcp_log_stackspecific log; /* * Unused logs are * 64 bit - delRate, rttProp, bw_inuse * 16 bit - cwnd_gain * 8 bit - bbr_state, bbr_substate, inhpts; */ memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.flex1 = hpts->p_nxt_slot; log.u_bbr.flex2 = hpts->p_cur_slot; log.u_bbr.flex3 = hpts->p_prev_slot; log.u_bbr.flex4 = idx; log.u_bbr.flex5 = hpts->p_curtick; log.u_bbr.flex6 = hpts->p_on_queue_cnt; log.u_bbr.flex7 = hpts->p_cpu; log.u_bbr.flex8 = (uint8_t)from_callout; log.u_bbr.inflight = slots_to_run; log.u_bbr.applimited = hpts->overidden_sleep; log.u_bbr.delivered = hpts->saved_curtick; log.u_bbr.timeStamp = tcp_tv_to_usectick(tv); log.u_bbr.epoch = hpts->saved_curslot; log.u_bbr.lt_epoch = hpts->saved_prev_slot; log.u_bbr.pkts_out = hpts->p_delayed_by; log.u_bbr.lost = hpts->p_hpts_sleep_time; log.u_bbr.pacing_gain = hpts->p_cpu; log.u_bbr.pkt_epoch = hpts->p_runningslot; log.u_bbr.use_lt_bw = 1; TCP_LOG_EVENTP(tp, NULL, &tp->t_inpcb->inp_socket->so_rcv, &tp->t_inpcb->inp_socket->so_snd, BBR_LOG_HPTSDIAG, 0, 0, &log, false, tv); } static void tcp_wakehpts(struct tcp_hpts_entry *hpts) { HPTS_MTX_ASSERT(hpts); if (tcp_hpts_no_wake_over_thresh && (hpts->p_on_queue_cnt >= conn_cnt_thresh)) { hpts->p_direct_wake = 0; return; } if (hpts->p_hpts_wake_scheduled == 0) { hpts->p_hpts_wake_scheduled = 1; swi_sched(hpts->ie_cookie, 0); } } static void hpts_timeout_swi(void *arg) { struct tcp_hpts_entry *hpts; hpts = (struct tcp_hpts_entry *)arg; swi_sched(hpts->ie_cookie, 0); } static void inp_hpts_insert(struct inpcb *inp, struct tcp_hpts_entry *hpts) { struct hptsh *hptsh; INP_WLOCK_ASSERT(inp); HPTS_MTX_ASSERT(hpts); MPASS(hpts->p_cpu == inp->inp_hpts_cpu); - MPASS(!(inp->inp_flags & (INP_DROPPED|INP_TIMEWAIT))); + MPASS(!(inp->inp_flags & INP_DROPPED)); hptsh = &hpts->p_hptss[inp->inp_hptsslot]; if (inp->inp_in_hpts == IHPTS_NONE) { inp->inp_in_hpts = IHPTS_ONQUEUE; in_pcbref(inp); } else if (inp->inp_in_hpts == IHPTS_MOVING) { inp->inp_in_hpts = IHPTS_ONQUEUE; } else MPASS(inp->inp_in_hpts == IHPTS_ONQUEUE); inp->inp_hpts_gencnt = hptsh->gencnt; TAILQ_INSERT_TAIL(&hptsh->head, inp, inp_hpts); hptsh->count++; hpts->p_on_queue_cnt++; } static struct tcp_hpts_entry * tcp_hpts_lock(struct inpcb *inp) { struct tcp_hpts_entry *hpts; INP_LOCK_ASSERT(inp); hpts = tcp_pace.rp_ent[inp->inp_hpts_cpu]; HPTS_LOCK(hpts); return (hpts); } static void inp_hpts_release(struct inpcb *inp) { bool released __diagused; inp->inp_in_hpts = IHPTS_NONE; released = in_pcbrele_wlocked(inp); MPASS(released == false); } /* * Called normally with the INP_LOCKED but it * does not matter, the hpts lock is the key * but the lock order allows us to hold the * INP lock and then get the hpts lock. */ void tcp_hpts_remove(struct inpcb *inp) { struct tcp_hpts_entry *hpts; struct hptsh *hptsh; INP_WLOCK_ASSERT(inp); hpts = tcp_hpts_lock(inp); if (inp->inp_in_hpts == IHPTS_ONQUEUE) { hptsh = &hpts->p_hptss[inp->inp_hptsslot]; inp->inp_hpts_request = 0; if (__predict_true(inp->inp_hpts_gencnt == hptsh->gencnt)) { TAILQ_REMOVE(&hptsh->head, inp, inp_hpts); MPASS(hptsh->count > 0); hptsh->count--; MPASS(hpts->p_on_queue_cnt > 0); hpts->p_on_queue_cnt--; inp_hpts_release(inp); } else { /* * tcp_hptsi() now owns the TAILQ head of this inp. * Can't TAILQ_REMOVE, just mark it. */ #ifdef INVARIANTS struct inpcb *tmp; TAILQ_FOREACH(tmp, &hptsh->head, inp_hpts) MPASS(tmp != inp); #endif inp->inp_in_hpts = IHPTS_MOVING; inp->inp_hptsslot = -1; } } else if (inp->inp_in_hpts == IHPTS_MOVING) { /* * Handle a special race condition: * tcp_hptsi() moves inpcb to detached tailq * tcp_hpts_remove() marks as IHPTS_MOVING, slot = -1 * tcp_hpts_insert() sets slot to a meaningful value * tcp_hpts_remove() again (we are here!), then in_pcbdrop() * tcp_hptsi() finds pcb with meaningful slot and INP_DROPPED */ inp->inp_hptsslot = -1; } HPTS_UNLOCK(hpts); } bool tcp_in_hpts(struct inpcb *inp) { return (inp->inp_in_hpts == IHPTS_ONQUEUE); } static inline int hpts_slot(uint32_t wheel_slot, uint32_t plus) { /* * Given a slot on the wheel, what slot * is that plus ticks out? */ KASSERT(wheel_slot < NUM_OF_HPTSI_SLOTS, ("Invalid tick %u not on wheel", wheel_slot)); return ((wheel_slot + plus) % NUM_OF_HPTSI_SLOTS); } static inline int tick_to_wheel(uint32_t cts_in_wticks) { /* * Given a timestamp in ticks (so by * default to get it to a real time one * would multiply by 10.. i.e the number * of ticks in a slot) map it to our limited * space wheel. */ return (cts_in_wticks % NUM_OF_HPTSI_SLOTS); } static inline int hpts_slots_diff(int prev_slot, int slot_now) { /* * Given two slots that are someplace * on our wheel. How far are they apart? */ if (slot_now > prev_slot) return (slot_now - prev_slot); else if (slot_now == prev_slot) /* * Special case, same means we can go all of our * wheel less one slot. */ return (NUM_OF_HPTSI_SLOTS - 1); else return ((NUM_OF_HPTSI_SLOTS - prev_slot) + slot_now); } /* * Given a slot on the wheel that is the current time * mapped to the wheel (wheel_slot), what is the maximum * distance forward that can be obtained without * wrapping past either prev_slot or running_slot * depending on the htps state? Also if passed * a uint32_t *, fill it with the slot location. * * Note if you do not give this function the current * time (that you think it is) mapped to the wheel slot * then the results will not be what you expect and * could lead to invalid inserts. */ static inline int32_t max_slots_available(struct tcp_hpts_entry *hpts, uint32_t wheel_slot, uint32_t *target_slot) { uint32_t dis_to_travel, end_slot, pacer_to_now, avail_on_wheel; if ((hpts->p_hpts_active == 1) && (hpts->p_wheel_complete == 0)) { end_slot = hpts->p_runningslot; /* Back up one tick */ if (end_slot == 0) end_slot = NUM_OF_HPTSI_SLOTS - 1; else end_slot--; if (target_slot) *target_slot = end_slot; } else { /* * For the case where we are * not active, or we have * completed the pass over * the wheel, we can use the * prev tick and subtract one from it. This puts us * as far out as possible on the wheel. */ end_slot = hpts->p_prev_slot; if (end_slot == 0) end_slot = NUM_OF_HPTSI_SLOTS - 1; else end_slot--; if (target_slot) *target_slot = end_slot; /* * Now we have close to the full wheel left minus the * time it has been since the pacer went to sleep. Note * that wheel_tick, passed in, should be the current time * from the perspective of the caller, mapped to the wheel. */ if (hpts->p_prev_slot != wheel_slot) dis_to_travel = hpts_slots_diff(hpts->p_prev_slot, wheel_slot); else dis_to_travel = 1; /* * dis_to_travel in this case is the space from when the * pacer stopped (p_prev_slot) and where our wheel_slot * is now. To know how many slots we can put it in we * subtract from the wheel size. We would not want * to place something after p_prev_slot or it will * get ran too soon. */ return (NUM_OF_HPTSI_SLOTS - dis_to_travel); } /* * So how many slots are open between p_runningslot -> p_cur_slot * that is what is currently un-available for insertion. Special * case when we are at the last slot, this gets 1, so that * the answer to how many slots are available is all but 1. */ if (hpts->p_runningslot == hpts->p_cur_slot) dis_to_travel = 1; else dis_to_travel = hpts_slots_diff(hpts->p_runningslot, hpts->p_cur_slot); /* * How long has the pacer been running? */ if (hpts->p_cur_slot != wheel_slot) { /* The pacer is a bit late */ pacer_to_now = hpts_slots_diff(hpts->p_cur_slot, wheel_slot); } else { /* The pacer is right on time, now == pacers start time */ pacer_to_now = 0; } /* * To get the number left we can insert into we simply * subract the distance the pacer has to run from how * many slots there are. */ avail_on_wheel = NUM_OF_HPTSI_SLOTS - dis_to_travel; /* * Now how many of those we will eat due to the pacer's * time (p_cur_slot) of start being behind the * real time (wheel_slot)? */ if (avail_on_wheel <= pacer_to_now) { /* * Wheel wrap, we can't fit on the wheel, that * is unusual the system must be way overloaded! * Insert into the assured slot, and return special * "0". */ counter_u64_add(combined_wheel_wrap, 1); *target_slot = hpts->p_nxt_slot; return (0); } else { /* * We know how many slots are open * on the wheel (the reverse of what * is left to run. Take away the time * the pacer started to now (wheel_slot) * and that tells you how many slots are * open that can be inserted into that won't * be touched by the pacer until later. */ return (avail_on_wheel - pacer_to_now); } } #ifdef INVARIANTS static void check_if_slot_would_be_wrong(struct tcp_hpts_entry *hpts, struct inpcb *inp, uint32_t inp_hptsslot, int line) { /* * Sanity checks for the pacer with invariants * on insert. */ KASSERT(inp_hptsslot < NUM_OF_HPTSI_SLOTS, ("hpts:%p inp:%p slot:%d > max", hpts, inp, inp_hptsslot)); if ((hpts->p_hpts_active) && (hpts->p_wheel_complete == 0)) { /* * If the pacer is processing a arc * of the wheel, we need to make * sure we are not inserting within * that arc. */ int distance, yet_to_run; distance = hpts_slots_diff(hpts->p_runningslot, inp_hptsslot); if (hpts->p_runningslot != hpts->p_cur_slot) yet_to_run = hpts_slots_diff(hpts->p_runningslot, hpts->p_cur_slot); else yet_to_run = 0; /* processing last slot */ KASSERT(yet_to_run <= distance, ("hpts:%p inp:%p slot:%d distance:%d yet_to_run:%d rs:%d cs:%d", hpts, inp, inp_hptsslot, distance, yet_to_run, hpts->p_runningslot, hpts->p_cur_slot)); } } #endif uint32_t tcp_hpts_insert_diag(struct inpcb *inp, uint32_t slot, int32_t line, struct hpts_diag *diag) { struct tcp_hpts_entry *hpts; struct timeval tv; uint32_t slot_on, wheel_cts, last_slot, need_new_to = 0; int32_t wheel_slot, maxslots; bool need_wakeup = false; INP_WLOCK_ASSERT(inp); MPASS(!tcp_in_hpts(inp)); - MPASS(!(inp->inp_flags & (INP_DROPPED|INP_TIMEWAIT))); + MPASS(!(inp->inp_flags & INP_DROPPED)); /* * We now return the next-slot the hpts will be on, beyond its * current run (if up) or where it was when it stopped if it is * sleeping. */ hpts = tcp_hpts_lock(inp); microuptime(&tv); if (diag) { memset(diag, 0, sizeof(struct hpts_diag)); diag->p_hpts_active = hpts->p_hpts_active; diag->p_prev_slot = hpts->p_prev_slot; diag->p_runningslot = hpts->p_runningslot; diag->p_nxt_slot = hpts->p_nxt_slot; diag->p_cur_slot = hpts->p_cur_slot; diag->p_curtick = hpts->p_curtick; diag->p_lasttick = hpts->p_lasttick; diag->slot_req = slot; diag->p_on_min_sleep = hpts->p_on_min_sleep; diag->hpts_sleep_time = hpts->p_hpts_sleep_time; } if (slot == 0) { /* Ok we need to set it on the hpts in the current slot */ inp->inp_hpts_request = 0; if ((hpts->p_hpts_active == 0) || (hpts->p_wheel_complete)) { /* * A sleeping hpts we want in next slot to run * note that in this state p_prev_slot == p_cur_slot */ inp->inp_hptsslot = hpts_slot(hpts->p_prev_slot, 1); if ((hpts->p_on_min_sleep == 0) && (hpts->p_hpts_active == 0)) need_wakeup = true; } else inp->inp_hptsslot = hpts->p_runningslot; if (__predict_true(inp->inp_in_hpts != IHPTS_MOVING)) inp_hpts_insert(inp, hpts); if (need_wakeup) { /* * Activate the hpts if it is sleeping and its * timeout is not 1. */ hpts->p_direct_wake = 1; tcp_wakehpts(hpts); } slot_on = hpts->p_nxt_slot; HPTS_UNLOCK(hpts); return (slot_on); } /* Get the current time relative to the wheel */ wheel_cts = tcp_tv_to_hptstick(&tv); /* Map it onto the wheel */ wheel_slot = tick_to_wheel(wheel_cts); /* Now what's the max we can place it at? */ maxslots = max_slots_available(hpts, wheel_slot, &last_slot); if (diag) { diag->wheel_slot = wheel_slot; diag->maxslots = maxslots; diag->wheel_cts = wheel_cts; } if (maxslots == 0) { /* The pacer is in a wheel wrap behind, yikes! */ if (slot > 1) { /* * Reduce by 1 to prevent a forever loop in * case something else is wrong. Note this * probably does not hurt because the pacer * if its true is so far behind we will be * > 1second late calling anyway. */ slot--; } inp->inp_hptsslot = last_slot; inp->inp_hpts_request = slot; } else if (maxslots >= slot) { /* It all fits on the wheel */ inp->inp_hpts_request = 0; inp->inp_hptsslot = hpts_slot(wheel_slot, slot); } else { /* It does not fit */ inp->inp_hpts_request = slot - maxslots; inp->inp_hptsslot = last_slot; } if (diag) { diag->slot_remaining = inp->inp_hpts_request; diag->inp_hptsslot = inp->inp_hptsslot; } #ifdef INVARIANTS check_if_slot_would_be_wrong(hpts, inp, inp->inp_hptsslot, line); #endif if (__predict_true(inp->inp_in_hpts != IHPTS_MOVING)) inp_hpts_insert(inp, hpts); if ((hpts->p_hpts_active == 0) && (inp->inp_hpts_request == 0) && (hpts->p_on_min_sleep == 0)) { /* * The hpts is sleeping and NOT on a minimum * sleep time, we need to figure out where * it will wake up at and if we need to reschedule * its time-out. */ uint32_t have_slept, yet_to_sleep; /* Now do we need to restart the hpts's timer? */ have_slept = hpts_slots_diff(hpts->p_prev_slot, wheel_slot); if (have_slept < hpts->p_hpts_sleep_time) yet_to_sleep = hpts->p_hpts_sleep_time - have_slept; else { /* We are over-due */ yet_to_sleep = 0; need_wakeup = 1; } if (diag) { diag->have_slept = have_slept; diag->yet_to_sleep = yet_to_sleep; } if (yet_to_sleep && (yet_to_sleep > slot)) { /* * We need to reschedule the hpts's time-out. */ hpts->p_hpts_sleep_time = slot; need_new_to = slot * HPTS_TICKS_PER_SLOT; } } /* * Now how far is the hpts sleeping to? if active is 1, its * up and ticking we do nothing, otherwise we may need to * reschedule its callout if need_new_to is set from above. */ if (need_wakeup) { hpts->p_direct_wake = 1; tcp_wakehpts(hpts); if (diag) { diag->need_new_to = 0; diag->co_ret = 0xffff0000; } } else if (need_new_to) { int32_t co_ret; struct timeval tv; sbintime_t sb; tv.tv_sec = 0; tv.tv_usec = 0; while (need_new_to > HPTS_USEC_IN_SEC) { tv.tv_sec++; need_new_to -= HPTS_USEC_IN_SEC; } tv.tv_usec = need_new_to; sb = tvtosbt(tv); co_ret = callout_reset_sbt_on(&hpts->co, sb, 0, hpts_timeout_swi, hpts, hpts->p_cpu, (C_DIRECT_EXEC | C_PREL(tcp_hpts_precision))); if (diag) { diag->need_new_to = need_new_to; diag->co_ret = co_ret; } } slot_on = hpts->p_nxt_slot; HPTS_UNLOCK(hpts); return (slot_on); } uint16_t hpts_random_cpu(struct inpcb *inp){ /* * No flow type set distribute the load randomly. */ uint16_t cpuid; uint32_t ran; /* * Shortcut if it is already set. XXXGL: does it happen? */ if (inp->inp_hpts_cpu_set) { return (inp->inp_hpts_cpu); } /* Nothing set use a random number */ ran = arc4random(); cpuid = (((ran & 0xffff) % mp_ncpus) % tcp_pace.rp_num_hptss); return (cpuid); } static uint16_t hpts_cpuid(struct inpcb *inp, int *failed) { u_int cpuid; #ifdef NUMA struct hpts_domain_info *di; #endif *failed = 0; if (inp->inp_hpts_cpu_set) { return (inp->inp_hpts_cpu); } /* * If we are using the irq cpu set by LRO or * the driver then it overrides all other domains. */ if (tcp_use_irq_cpu) { if (inp->inp_irq_cpu_set == 0) { *failed = 1; return(0); } return(inp->inp_irq_cpu); } /* If one is set the other must be the same */ #ifdef RSS cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype); if (cpuid == NETISR_CPUID_NONE) return (hpts_random_cpu(inp)); else return (cpuid); #endif /* * We don't have a flowid -> cpuid mapping, so cheat and just map * unknown cpuids to curcpu. Not the best, but apparently better * than defaulting to swi 0. */ if (inp->inp_flowtype == M_HASHTYPE_NONE) { counter_u64_add(cpu_uses_random, 1); return (hpts_random_cpu(inp)); } /* * Hash to a thread based on the flowid. If we are using numa, * then restrict the hash to the numa domain where the inp lives. */ #ifdef NUMA if ((vm_ndomains == 1) || (inp->inp_numa_domain == M_NODOM)) { #endif cpuid = inp->inp_flowid % mp_ncpus; #ifdef NUMA } else { /* Hash into the cpu's that use that domain */ di = &hpts_domains[inp->inp_numa_domain]; cpuid = di->cpu[inp->inp_flowid % di->count]; } #endif counter_u64_add(cpu_uses_flowid, 1); return (cpuid); } #ifdef not_longer_used_gleb static void tcp_drop_in_pkts(struct tcpcb *tp) { struct mbuf *m, *n; m = tp->t_in_pkt; if (m) n = m->m_nextpkt; else n = NULL; tp->t_in_pkt = NULL; while (m) { m_freem(m); m = n; if (m) n = m->m_nextpkt; } } #endif static void tcp_hpts_set_max_sleep(struct tcp_hpts_entry *hpts, int wrap_loop_cnt) { uint32_t t = 0, i; if ((hpts->p_on_queue_cnt) && (wrap_loop_cnt < 2)) { /* * Find next slot that is occupied and use that to * be the sleep time. */ for (i = 0, t = hpts_slot(hpts->p_cur_slot, 1); i < NUM_OF_HPTSI_SLOTS; i++) { if (TAILQ_EMPTY(&hpts->p_hptss[t].head) == 0) { break; } t = (t + 1) % NUM_OF_HPTSI_SLOTS; } KASSERT((i != NUM_OF_HPTSI_SLOTS), ("Hpts:%p cnt:%d but none found", hpts, hpts->p_on_queue_cnt)); hpts->p_hpts_sleep_time = min((i + 1), hpts_sleep_max); } else { /* No one on the wheel sleep for all but 400 slots or sleep max */ hpts->p_hpts_sleep_time = hpts_sleep_max; } } static int32_t tcp_hptsi(struct tcp_hpts_entry *hpts, int from_callout) { struct tcpcb *tp; struct inpcb *inp; struct timeval tv; int32_t slots_to_run, i, error; int32_t loop_cnt = 0; int32_t did_prefetch = 0; int32_t prefetch_ninp = 0; int32_t prefetch_tp = 0; int32_t wrap_loop_cnt = 0; int32_t slot_pos_of_endpoint = 0; int32_t orig_exit_slot; int8_t completed_measure = 0, seen_endpoint = 0; HPTS_MTX_ASSERT(hpts); NET_EPOCH_ASSERT(); /* record previous info for any logging */ hpts->saved_lasttick = hpts->p_lasttick; hpts->saved_curtick = hpts->p_curtick; hpts->saved_curslot = hpts->p_cur_slot; hpts->saved_prev_slot = hpts->p_prev_slot; hpts->p_lasttick = hpts->p_curtick; hpts->p_curtick = tcp_gethptstick(&tv); cts_last_ran[hpts->p_num] = tcp_tv_to_usectick(&tv); orig_exit_slot = hpts->p_cur_slot = tick_to_wheel(hpts->p_curtick); if ((hpts->p_on_queue_cnt == 0) || (hpts->p_lasttick == hpts->p_curtick)) { /* * No time has yet passed, * or nothing to do. */ hpts->p_prev_slot = hpts->p_cur_slot; hpts->p_lasttick = hpts->p_curtick; goto no_run; } again: hpts->p_wheel_complete = 0; HPTS_MTX_ASSERT(hpts); slots_to_run = hpts_slots_diff(hpts->p_prev_slot, hpts->p_cur_slot); if (((hpts->p_curtick - hpts->p_lasttick) > ((NUM_OF_HPTSI_SLOTS-1) * HPTS_TICKS_PER_SLOT)) && (hpts->p_on_queue_cnt != 0)) { /* * Wheel wrap is occuring, basically we * are behind and the distance between * run's has spread so much it has exceeded * the time on the wheel (1.024 seconds). This * is ugly and should NOT be happening. We * need to run the entire wheel. We last processed * p_prev_slot, so that needs to be the last slot * we run. The next slot after that should be our * reserved first slot for new, and then starts * the running position. Now the problem is the * reserved "not to yet" place does not exist * and there may be inp's in there that need * running. We can merge those into the * first slot at the head. */ wrap_loop_cnt++; hpts->p_nxt_slot = hpts_slot(hpts->p_prev_slot, 1); hpts->p_runningslot = hpts_slot(hpts->p_prev_slot, 2); /* * Adjust p_cur_slot to be where we are starting from * hopefully we will catch up (fat chance if something * is broken this bad :( ) */ hpts->p_cur_slot = hpts->p_prev_slot; /* * The next slot has guys to run too, and that would * be where we would normally start, lets move them into * the next slot (p_prev_slot + 2) so that we will * run them, the extra 10usecs of late (by being * put behind) does not really matter in this situation. */ TAILQ_FOREACH(inp, &hpts->p_hptss[hpts->p_nxt_slot].head, inp_hpts) { MPASS(inp->inp_hptsslot == hpts->p_nxt_slot); MPASS(inp->inp_hpts_gencnt == hpts->p_hptss[hpts->p_nxt_slot].gencnt); MPASS(inp->inp_in_hpts == IHPTS_ONQUEUE); /* * Update gencnt and nextslot accordingly to match * the new location. This is safe since it takes both * the INP lock and the pacer mutex to change the * inp_hptsslot and inp_hpts_gencnt. */ inp->inp_hpts_gencnt = hpts->p_hptss[hpts->p_runningslot].gencnt; inp->inp_hptsslot = hpts->p_runningslot; } TAILQ_CONCAT(&hpts->p_hptss[hpts->p_runningslot].head, &hpts->p_hptss[hpts->p_nxt_slot].head, inp_hpts); hpts->p_hptss[hpts->p_runningslot].count += hpts->p_hptss[hpts->p_nxt_slot].count; hpts->p_hptss[hpts->p_nxt_slot].count = 0; hpts->p_hptss[hpts->p_nxt_slot].gencnt++; slots_to_run = NUM_OF_HPTSI_SLOTS - 1; counter_u64_add(wheel_wrap, 1); } else { /* * Nxt slot is always one after p_runningslot though * its not used usually unless we are doing wheel wrap. */ hpts->p_nxt_slot = hpts->p_prev_slot; hpts->p_runningslot = hpts_slot(hpts->p_prev_slot, 1); } if (hpts->p_on_queue_cnt == 0) { goto no_one; } for (i = 0; i < slots_to_run; i++) { struct inpcb *inp, *ninp; TAILQ_HEAD(, inpcb) head = TAILQ_HEAD_INITIALIZER(head); struct hptsh *hptsh; uint32_t runningslot; /* * Calculate our delay, if there are no extra ticks there * was not any (i.e. if slots_to_run == 1, no delay). */ hpts->p_delayed_by = (slots_to_run - (i + 1)) * HPTS_TICKS_PER_SLOT; runningslot = hpts->p_runningslot; hptsh = &hpts->p_hptss[runningslot]; TAILQ_SWAP(&head, &hptsh->head, inpcb, inp_hpts); hpts->p_on_queue_cnt -= hptsh->count; hptsh->count = 0; hptsh->gencnt++; HPTS_UNLOCK(hpts); TAILQ_FOREACH_SAFE(inp, &head, inp_hpts, ninp) { bool set_cpu; if (ninp != NULL) { /* We prefetch the next inp if possible */ kern_prefetch(ninp, &prefetch_ninp); prefetch_ninp = 1; } /* For debugging */ if (seen_endpoint == 0) { seen_endpoint = 1; orig_exit_slot = slot_pos_of_endpoint = runningslot; } else if (completed_measure == 0) { /* Record the new position */ orig_exit_slot = runningslot; } INP_WLOCK(inp); if (inp->inp_hpts_cpu_set == 0) { set_cpu = true; } else { set_cpu = false; } if (__predict_false(inp->inp_in_hpts == IHPTS_MOVING)) { if (inp->inp_hptsslot == -1) { inp->inp_in_hpts = IHPTS_NONE; if (in_pcbrele_wlocked(inp) == false) INP_WUNLOCK(inp); } else { HPTS_LOCK(hpts); inp_hpts_insert(inp, hpts); HPTS_UNLOCK(hpts); INP_WUNLOCK(inp); } continue; } MPASS(inp->inp_in_hpts == IHPTS_ONQUEUE); - MPASS(!(inp->inp_flags & (INP_DROPPED|INP_TIMEWAIT))); + MPASS(!(inp->inp_flags & INP_DROPPED)); KASSERT(runningslot == inp->inp_hptsslot, ("Hpts:%p inp:%p slot mis-aligned %u vs %u", hpts, inp, runningslot, inp->inp_hptsslot)); if (inp->inp_hpts_request) { /* * This guy is deferred out further in time * then our wheel had available on it. * Push him back on the wheel or run it * depending. */ uint32_t maxslots, last_slot, remaining_slots; remaining_slots = slots_to_run - (i + 1); if (inp->inp_hpts_request > remaining_slots) { HPTS_LOCK(hpts); /* * How far out can we go? */ maxslots = max_slots_available(hpts, hpts->p_cur_slot, &last_slot); if (maxslots >= inp->inp_hpts_request) { /* We can place it finally to * be processed. */ inp->inp_hptsslot = hpts_slot( hpts->p_runningslot, inp->inp_hpts_request); inp->inp_hpts_request = 0; } else { /* Work off some more time */ inp->inp_hptsslot = last_slot; inp->inp_hpts_request -= maxslots; } inp_hpts_insert(inp, hpts); HPTS_UNLOCK(hpts); INP_WUNLOCK(inp); continue; } inp->inp_hpts_request = 0; /* Fall through we will so do it now */ } inp_hpts_release(inp); tp = intotcpcb(inp); MPASS(tp); if (set_cpu) { /* * Setup so the next time we will move to * the right CPU. This should be a rare * event. It will sometimes happens when we * are the client side (usually not the * server). Somehow tcp_output() gets called * before the tcp_do_segment() sets the * intial state. This means the r_cpu and * r_hpts_cpu is 0. We get on the hpts, and * then tcp_input() gets called setting up * the r_cpu to the correct value. The hpts * goes off and sees the mis-match. We * simply correct it here and the CPU will * switch to the new hpts nextime the tcb * gets added to the hpts (not this one) * :-) */ tcp_set_hpts(inp); } CURVNET_SET(inp->inp_vnet); /* Lets do any logging that we might want to */ if (hpts_does_tp_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) { tcp_hpts_log(hpts, tp, &tv, slots_to_run, i, from_callout); } if (tp->t_fb_ptr != NULL) { kern_prefetch(tp->t_fb_ptr, &did_prefetch); did_prefetch = 1; } if ((inp->inp_flags2 & INP_SUPPORTS_MBUFQ) && tp->t_in_pkt) { error = (*tp->t_fb->tfb_do_queued_segments)(inp->inp_socket, tp, 0); if (error) { /* The input killed the connection */ goto skip_pacing; } } inp->inp_hpts_calls = 1; error = tcp_output(tp); if (error < 0) goto skip_pacing; inp->inp_hpts_calls = 0; if (ninp && ninp->inp_ppcb) { /* * If we have a nxt inp, see if we can * prefetch its ppcb. Note this may seem * "risky" since we have no locks (other * than the previous inp) and there no * assurance that ninp was not pulled while * we were processing inp and freed. If this * occurred it could mean that either: * * a) Its NULL (which is fine we won't go * here) b) Its valid (which is cool we * will prefetch it) c) The inp got * freed back to the slab which was * reallocated. Then the piece of memory was * re-used and something else (not an * address) is in inp_ppcb. If that occurs * we don't crash, but take a TLB shootdown * performance hit (same as if it was NULL * and we tried to pre-fetch it). * * Considering that the likelyhood of is * quite rare we will take a risk on doing * this. If performance drops after testing * we can always take this out. NB: the * kern_prefetch on amd64 actually has * protection against a bad address now via * the DMAP_() tests. This will prevent the * TLB hit, and instead if occurs just * cause us to load cache with a useless * address (to us). */ kern_prefetch(ninp->inp_ppcb, &prefetch_tp); prefetch_tp = 1; } INP_WUNLOCK(inp); skip_pacing: CURVNET_RESTORE(); } if (seen_endpoint) { /* * We now have a accurate distance between * slot_pos_of_endpoint <-> orig_exit_slot * to tell us how late we were, orig_exit_slot * is where we calculated the end of our cycle to * be when we first entered. */ completed_measure = 1; } HPTS_LOCK(hpts); hpts->p_runningslot++; if (hpts->p_runningslot >= NUM_OF_HPTSI_SLOTS) { hpts->p_runningslot = 0; } } no_one: HPTS_MTX_ASSERT(hpts); hpts->p_delayed_by = 0; /* * Check to see if we took an excess amount of time and need to run * more ticks (if we did not hit eno-bufs). */ hpts->p_prev_slot = hpts->p_cur_slot; hpts->p_lasttick = hpts->p_curtick; if ((from_callout == 0) || (loop_cnt > max_pacer_loops)) { /* * Something is serious slow we have * looped through processing the wheel * and by the time we cleared the * needs to run max_pacer_loops time * we still needed to run. That means * the system is hopelessly behind and * can never catch up :( * * We will just lie to this thread * and let it thing p_curtick is * correct. When it next awakens * it will find itself further behind. */ if (from_callout) counter_u64_add(hpts_hopelessly_behind, 1); goto no_run; } hpts->p_curtick = tcp_gethptstick(&tv); hpts->p_cur_slot = tick_to_wheel(hpts->p_curtick); if (seen_endpoint == 0) { /* We saw no endpoint but we may be looping */ orig_exit_slot = hpts->p_cur_slot; } if ((wrap_loop_cnt < 2) && (hpts->p_lasttick != hpts->p_curtick)) { counter_u64_add(hpts_loops, 1); loop_cnt++; goto again; } no_run: cts_last_ran[hpts->p_num] = tcp_tv_to_usectick(&tv); /* * Set flag to tell that we are done for * any slot input that happens during * input. */ hpts->p_wheel_complete = 1; /* * Now did we spend too long running input and need to run more ticks? * Note that if wrap_loop_cnt < 2 then we should have the conditions * in the KASSERT's true. But if the wheel is behind i.e. wrap_loop_cnt * is greater than 2, then the condtion most likely are *not* true. * Also if we are called not from the callout, we don't run the wheel * multiple times so the slots may not align either. */ KASSERT(((hpts->p_prev_slot == hpts->p_cur_slot) || (wrap_loop_cnt >= 2) || (from_callout == 0)), ("H:%p p_prev_slot:%u not equal to p_cur_slot:%u", hpts, hpts->p_prev_slot, hpts->p_cur_slot)); KASSERT(((hpts->p_lasttick == hpts->p_curtick) || (wrap_loop_cnt >= 2) || (from_callout == 0)), ("H:%p p_lasttick:%u not equal to p_curtick:%u", hpts, hpts->p_lasttick, hpts->p_curtick)); if (from_callout && (hpts->p_lasttick != hpts->p_curtick)) { hpts->p_curtick = tcp_gethptstick(&tv); counter_u64_add(hpts_loops, 1); hpts->p_cur_slot = tick_to_wheel(hpts->p_curtick); goto again; } if (from_callout){ tcp_hpts_set_max_sleep(hpts, wrap_loop_cnt); } if (seen_endpoint) return(hpts_slots_diff(slot_pos_of_endpoint, orig_exit_slot)); else return (0); } void __tcp_set_hpts(struct inpcb *inp, int32_t line) { struct tcp_hpts_entry *hpts; int failed; INP_WLOCK_ASSERT(inp); hpts = tcp_hpts_lock(inp); if ((inp->inp_in_hpts == 0) && (inp->inp_hpts_cpu_set == 0)) { inp->inp_hpts_cpu = hpts_cpuid(inp, &failed); if (failed == 0) inp->inp_hpts_cpu_set = 1; } mtx_unlock(&hpts->p_mtx); } static void __tcp_run_hpts(struct tcp_hpts_entry *hpts) { int ticks_ran; if (hpts->p_hpts_active) { /* Already active */ return; } if (mtx_trylock(&hpts->p_mtx) == 0) { /* Someone else got the lock */ return; } if (hpts->p_hpts_active) goto out_with_mtx; hpts->syscall_cnt++; counter_u64_add(hpts_direct_call, 1); hpts->p_hpts_active = 1; ticks_ran = tcp_hptsi(hpts, 0); /* We may want to adjust the sleep values here */ if (hpts->p_on_queue_cnt >= conn_cnt_thresh) { if (ticks_ran > ticks_indicate_less_sleep) { struct timeval tv; sbintime_t sb; hpts->p_mysleep.tv_usec /= 2; if (hpts->p_mysleep.tv_usec < dynamic_min_sleep) hpts->p_mysleep.tv_usec = dynamic_min_sleep; /* Reschedule with new to value */ tcp_hpts_set_max_sleep(hpts, 0); tv.tv_usec = hpts->p_hpts_sleep_time * HPTS_TICKS_PER_SLOT; /* Validate its in the right ranges */ if (tv.tv_usec < hpts->p_mysleep.tv_usec) { hpts->overidden_sleep = tv.tv_usec; tv.tv_usec = hpts->p_mysleep.tv_usec; } else if (tv.tv_usec > dynamic_max_sleep) { /* Lets not let sleep get above this value */ hpts->overidden_sleep = tv.tv_usec; tv.tv_usec = dynamic_max_sleep; } /* * In this mode the timer is a backstop to * all the userret/lro_flushes so we use * the dynamic value and set the on_min_sleep * flag so we will not be awoken. */ sb = tvtosbt(tv); /* Store off to make visible the actual sleep time */ hpts->sleeping = tv.tv_usec; callout_reset_sbt_on(&hpts->co, sb, 0, hpts_timeout_swi, hpts, hpts->p_cpu, (C_DIRECT_EXEC | C_PREL(tcp_hpts_precision))); } else if (ticks_ran < ticks_indicate_more_sleep) { /* For the further sleep, don't reschedule hpts */ hpts->p_mysleep.tv_usec *= 2; if (hpts->p_mysleep.tv_usec > dynamic_max_sleep) hpts->p_mysleep.tv_usec = dynamic_max_sleep; } hpts->p_on_min_sleep = 1; } hpts->p_hpts_active = 0; out_with_mtx: HPTS_MTX_ASSERT(hpts); mtx_unlock(&hpts->p_mtx); } static struct tcp_hpts_entry * tcp_choose_hpts_to_run(void) { int i, oldest_idx, start, end; uint32_t cts, time_since_ran, calc; cts = tcp_get_usecs(NULL); time_since_ran = 0; /* Default is all one group */ start = 0; end = tcp_pace.rp_num_hptss; /* * If we have more than one L3 group figure out which one * this CPU is in. */ if (tcp_pace.grp_cnt > 1) { for (i = 0; i < tcp_pace.grp_cnt; i++) { if (CPU_ISSET(curcpu, &tcp_pace.grps[i]->cg_mask)) { start = tcp_pace.grps[i]->cg_first; end = (tcp_pace.grps[i]->cg_last + 1); break; } } } oldest_idx = -1; for (i = start; i < end; i++) { if (TSTMP_GT(cts, cts_last_ran[i])) calc = cts - cts_last_ran[i]; else calc = 0; if (calc > time_since_ran) { oldest_idx = i; time_since_ran = calc; } } if (oldest_idx >= 0) return(tcp_pace.rp_ent[oldest_idx]); else return(tcp_pace.rp_ent[(curcpu % tcp_pace.rp_num_hptss)]); } void tcp_run_hpts(void) { static struct tcp_hpts_entry *hpts; struct epoch_tracker et; NET_EPOCH_ENTER(et); hpts = tcp_choose_hpts_to_run(); __tcp_run_hpts(hpts); NET_EPOCH_EXIT(et); } static void tcp_hpts_thread(void *ctx) { struct tcp_hpts_entry *hpts; struct epoch_tracker et; struct timeval tv; sbintime_t sb; int ticks_ran; hpts = (struct tcp_hpts_entry *)ctx; mtx_lock(&hpts->p_mtx); if (hpts->p_direct_wake) { /* Signaled by input or output with low occupancy count. */ callout_stop(&hpts->co); counter_u64_add(hpts_direct_awakening, 1); } else { /* Timed out, the normal case. */ counter_u64_add(hpts_wake_timeout, 1); if (callout_pending(&hpts->co) || !callout_active(&hpts->co)) { mtx_unlock(&hpts->p_mtx); return; } } callout_deactivate(&hpts->co); hpts->p_hpts_wake_scheduled = 0; NET_EPOCH_ENTER(et); if (hpts->p_hpts_active) { /* * We are active already. This means that a syscall * trap or LRO is running in behalf of hpts. In that case * we need to double our timeout since there seems to be * enough activity in the system that we don't need to * run as often (if we were not directly woken). */ if (hpts->p_direct_wake == 0) { counter_u64_add(hpts_back_tosleep, 1); if (hpts->p_on_queue_cnt >= conn_cnt_thresh) { hpts->p_mysleep.tv_usec *= 2; if (hpts->p_mysleep.tv_usec > dynamic_max_sleep) hpts->p_mysleep.tv_usec = dynamic_max_sleep; tv.tv_usec = hpts->p_mysleep.tv_usec; hpts->p_on_min_sleep = 1; } else { /* * Here we have low count on the wheel, but * somehow we still collided with one of the * connections. Lets go back to sleep for a * min sleep time, but clear the flag so we * can be awoken by insert. */ hpts->p_on_min_sleep = 0; tv.tv_usec = tcp_min_hptsi_time; } } else { /* * Directly woken most likely to reset the * callout time. */ tv.tv_sec = 0; tv.tv_usec = hpts->p_mysleep.tv_usec; } goto back_to_sleep; } hpts->sleeping = 0; hpts->p_hpts_active = 1; ticks_ran = tcp_hptsi(hpts, 1); tv.tv_sec = 0; tv.tv_usec = hpts->p_hpts_sleep_time * HPTS_TICKS_PER_SLOT; if (hpts->p_on_queue_cnt >= conn_cnt_thresh) { if(hpts->p_direct_wake == 0) { /* * Only adjust sleep time if we were * called from the callout i.e. direct_wake == 0. */ if (ticks_ran < ticks_indicate_more_sleep) { hpts->p_mysleep.tv_usec *= 2; if (hpts->p_mysleep.tv_usec > dynamic_max_sleep) hpts->p_mysleep.tv_usec = dynamic_max_sleep; } else if (ticks_ran > ticks_indicate_less_sleep) { hpts->p_mysleep.tv_usec /= 2; if (hpts->p_mysleep.tv_usec < dynamic_min_sleep) hpts->p_mysleep.tv_usec = dynamic_min_sleep; } } if (tv.tv_usec < hpts->p_mysleep.tv_usec) { hpts->overidden_sleep = tv.tv_usec; tv.tv_usec = hpts->p_mysleep.tv_usec; } else if (tv.tv_usec > dynamic_max_sleep) { /* Lets not let sleep get above this value */ hpts->overidden_sleep = tv.tv_usec; tv.tv_usec = dynamic_max_sleep; } /* * In this mode the timer is a backstop to * all the userret/lro_flushes so we use * the dynamic value and set the on_min_sleep * flag so we will not be awoken. */ hpts->p_on_min_sleep = 1; } else if (hpts->p_on_queue_cnt == 0) { /* * No one on the wheel, please wake us up * if you insert on the wheel. */ hpts->p_on_min_sleep = 0; hpts->overidden_sleep = 0; } else { /* * We hit here when we have a low number of * clients on the wheel (our else clause). * We may need to go on min sleep, if we set * the flag we will not be awoken if someone * is inserted ahead of us. Clearing the flag * means we can be awoken. This is "old mode" * where the timer is what runs hpts mainly. */ if (tv.tv_usec < tcp_min_hptsi_time) { /* * Yes on min sleep, which means * we cannot be awoken. */ hpts->overidden_sleep = tv.tv_usec; tv.tv_usec = tcp_min_hptsi_time; hpts->p_on_min_sleep = 1; } else { /* Clear the min sleep flag */ hpts->overidden_sleep = 0; hpts->p_on_min_sleep = 0; } } HPTS_MTX_ASSERT(hpts); hpts->p_hpts_active = 0; back_to_sleep: hpts->p_direct_wake = 0; sb = tvtosbt(tv); /* Store off to make visible the actual sleep time */ hpts->sleeping = tv.tv_usec; callout_reset_sbt_on(&hpts->co, sb, 0, hpts_timeout_swi, hpts, hpts->p_cpu, (C_DIRECT_EXEC | C_PREL(tcp_hpts_precision))); NET_EPOCH_EXIT(et); mtx_unlock(&hpts->p_mtx); } #undef timersub static int32_t hpts_count_level(struct cpu_group *cg) { int32_t count_l3, i; count_l3 = 0; if (cg->cg_level == CG_SHARE_L3) count_l3++; /* Walk all the children looking for L3 */ for (i = 0; i < cg->cg_children; i++) { count_l3 += hpts_count_level(&cg->cg_child[i]); } return (count_l3); } static void hpts_gather_grps(struct cpu_group **grps, int32_t *at, int32_t max, struct cpu_group *cg) { int32_t idx, i; idx = *at; if (cg->cg_level == CG_SHARE_L3) { grps[idx] = cg; idx++; if (idx == max) { *at = idx; return; } } *at = idx; /* Walk all the children looking for L3 */ for (i = 0; i < cg->cg_children; i++) { hpts_gather_grps(grps, at, max, &cg->cg_child[i]); } } static void tcp_init_hptsi(void *st) { struct cpu_group *cpu_top; int32_t error __diagused; int32_t i, j, bound = 0, created = 0; size_t sz, asz; struct timeval tv; sbintime_t sb; struct tcp_hpts_entry *hpts; struct pcpu *pc; char unit[16]; uint32_t ncpus = mp_ncpus ? mp_ncpus : MAXCPU; int count, domain; #ifdef SMP cpu_top = smp_topo(); #else cpu_top = NULL; #endif tcp_pace.rp_num_hptss = ncpus; hpts_hopelessly_behind = counter_u64_alloc(M_WAITOK); hpts_loops = counter_u64_alloc(M_WAITOK); back_tosleep = counter_u64_alloc(M_WAITOK); combined_wheel_wrap = counter_u64_alloc(M_WAITOK); wheel_wrap = counter_u64_alloc(M_WAITOK); hpts_wake_timeout = counter_u64_alloc(M_WAITOK); hpts_direct_awakening = counter_u64_alloc(M_WAITOK); hpts_back_tosleep = counter_u64_alloc(M_WAITOK); hpts_direct_call = counter_u64_alloc(M_WAITOK); cpu_uses_flowid = counter_u64_alloc(M_WAITOK); cpu_uses_random = counter_u64_alloc(M_WAITOK); sz = (tcp_pace.rp_num_hptss * sizeof(struct tcp_hpts_entry *)); tcp_pace.rp_ent = malloc(sz, M_TCPHPTS, M_WAITOK | M_ZERO); sz = (sizeof(uint32_t) * tcp_pace.rp_num_hptss); cts_last_ran = malloc(sz, M_TCPHPTS, M_WAITOK); tcp_pace.grp_cnt = 0; if (cpu_top == NULL) { tcp_pace.grp_cnt = 1; } else { /* Find out how many cache level 3 domains we have */ count = 0; tcp_pace.grp_cnt = hpts_count_level(cpu_top); if (tcp_pace.grp_cnt == 0) { tcp_pace.grp_cnt = 1; } sz = (tcp_pace.grp_cnt * sizeof(struct cpu_group *)); tcp_pace.grps = malloc(sz, M_TCPHPTS, M_WAITOK); /* Now populate the groups */ if (tcp_pace.grp_cnt == 1) { /* * All we need is the top level all cpu's are in * the same cache so when we use grp[0]->cg_mask * with the cg_first <-> cg_last it will include * all cpu's in it. The level here is probably * zero which is ok. */ tcp_pace.grps[0] = cpu_top; } else { /* * Here we must find all the level three cache domains * and setup our pointers to them. */ count = 0; hpts_gather_grps(tcp_pace.grps, &count, tcp_pace.grp_cnt, cpu_top); } } asz = sizeof(struct hptsh) * NUM_OF_HPTSI_SLOTS; for (i = 0; i < tcp_pace.rp_num_hptss; i++) { tcp_pace.rp_ent[i] = malloc(sizeof(struct tcp_hpts_entry), M_TCPHPTS, M_WAITOK | M_ZERO); tcp_pace.rp_ent[i]->p_hptss = malloc(asz, M_TCPHPTS, M_WAITOK); hpts = tcp_pace.rp_ent[i]; /* * Init all the hpts structures that are not specifically * zero'd by the allocations. Also lets attach them to the * appropriate sysctl block as well. */ mtx_init(&hpts->p_mtx, "tcp_hpts_lck", "hpts", MTX_DEF | MTX_DUPOK); for (j = 0; j < NUM_OF_HPTSI_SLOTS; j++) { TAILQ_INIT(&hpts->p_hptss[j].head); hpts->p_hptss[j].count = 0; hpts->p_hptss[j].gencnt = 0; } sysctl_ctx_init(&hpts->hpts_ctx); sprintf(unit, "%d", i); hpts->hpts_root = SYSCTL_ADD_NODE(&hpts->hpts_ctx, SYSCTL_STATIC_CHILDREN(_net_inet_tcp_hpts), OID_AUTO, unit, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, ""); SYSCTL_ADD_INT(&hpts->hpts_ctx, SYSCTL_CHILDREN(hpts->hpts_root), OID_AUTO, "out_qcnt", CTLFLAG_RD, &hpts->p_on_queue_cnt, 0, "Count TCB's awaiting output processing"); SYSCTL_ADD_U16(&hpts->hpts_ctx, SYSCTL_CHILDREN(hpts->hpts_root), OID_AUTO, "active", CTLFLAG_RD, &hpts->p_hpts_active, 0, "Is the hpts active"); SYSCTL_ADD_UINT(&hpts->hpts_ctx, SYSCTL_CHILDREN(hpts->hpts_root), OID_AUTO, "curslot", CTLFLAG_RD, &hpts->p_cur_slot, 0, "What the current running pacers goal"); SYSCTL_ADD_UINT(&hpts->hpts_ctx, SYSCTL_CHILDREN(hpts->hpts_root), OID_AUTO, "runtick", CTLFLAG_RD, &hpts->p_runningslot, 0, "What the running pacers current slot is"); SYSCTL_ADD_UINT(&hpts->hpts_ctx, SYSCTL_CHILDREN(hpts->hpts_root), OID_AUTO, "curtick", CTLFLAG_RD, &hpts->p_curtick, 0, "What the running pacers last tick mapped to the wheel was"); SYSCTL_ADD_UINT(&hpts->hpts_ctx, SYSCTL_CHILDREN(hpts->hpts_root), OID_AUTO, "lastran", CTLFLAG_RD, &cts_last_ran[i], 0, "The last usec tick that this hpts ran"); SYSCTL_ADD_LONG(&hpts->hpts_ctx, SYSCTL_CHILDREN(hpts->hpts_root), OID_AUTO, "cur_min_sleep", CTLFLAG_RD, &hpts->p_mysleep.tv_usec, "What the running pacers is using for p_mysleep.tv_usec"); SYSCTL_ADD_U64(&hpts->hpts_ctx, SYSCTL_CHILDREN(hpts->hpts_root), OID_AUTO, "now_sleeping", CTLFLAG_RD, &hpts->sleeping, 0, "What the running pacers is actually sleeping for"); SYSCTL_ADD_U64(&hpts->hpts_ctx, SYSCTL_CHILDREN(hpts->hpts_root), OID_AUTO, "syscall_cnt", CTLFLAG_RD, &hpts->syscall_cnt, 0, "How many times we had syscalls on this hpts"); hpts->p_hpts_sleep_time = hpts_sleep_max; hpts->p_num = i; hpts->p_curtick = tcp_gethptstick(&tv); cts_last_ran[i] = tcp_tv_to_usectick(&tv); hpts->p_prev_slot = hpts->p_cur_slot = tick_to_wheel(hpts->p_curtick); hpts->p_cpu = 0xffff; hpts->p_nxt_slot = hpts_slot(hpts->p_cur_slot, 1); callout_init(&hpts->co, 1); } /* Don't try to bind to NUMA domains if we don't have any */ if (vm_ndomains == 1 && tcp_bind_threads == 2) tcp_bind_threads = 0; /* * Now lets start ithreads to handle the hptss. */ for (i = 0; i < tcp_pace.rp_num_hptss; i++) { hpts = tcp_pace.rp_ent[i]; hpts->p_cpu = i; error = swi_add(&hpts->ie, "hpts", tcp_hpts_thread, (void *)hpts, SWI_NET, INTR_MPSAFE, &hpts->ie_cookie); KASSERT(error == 0, ("Can't add hpts:%p i:%d err:%d", hpts, i, error)); created++; hpts->p_mysleep.tv_sec = 0; hpts->p_mysleep.tv_usec = tcp_min_hptsi_time; if (tcp_bind_threads == 1) { if (intr_event_bind(hpts->ie, i) == 0) bound++; } else if (tcp_bind_threads == 2) { /* Find the group for this CPU (i) and bind into it */ for (j = 0; j < tcp_pace.grp_cnt; j++) { if (CPU_ISSET(i, &tcp_pace.grps[j]->cg_mask)) { if (intr_event_bind_ithread_cpuset(hpts->ie, &tcp_pace.grps[j]->cg_mask) == 0) { bound++; pc = pcpu_find(i); domain = pc->pc_domain; count = hpts_domains[domain].count; hpts_domains[domain].cpu[count] = i; hpts_domains[domain].count++; break; } } } } tv.tv_sec = 0; tv.tv_usec = hpts->p_hpts_sleep_time * HPTS_TICKS_PER_SLOT; hpts->sleeping = tv.tv_usec; sb = tvtosbt(tv); callout_reset_sbt_on(&hpts->co, sb, 0, hpts_timeout_swi, hpts, hpts->p_cpu, (C_DIRECT_EXEC | C_PREL(tcp_hpts_precision))); } /* * If we somehow have an empty domain, fall back to choosing * among all htps threads. */ for (i = 0; i < vm_ndomains; i++) { if (hpts_domains[i].count == 0) { tcp_bind_threads = 0; break; } } printf("TCP Hpts created %d swi interrupt threads and bound %d to %s\n", created, bound, tcp_bind_threads == 2 ? "NUMA domains" : "cpus"); #ifdef INVARIANTS printf("HPTS is in INVARIANT mode!!\n"); #endif } SYSINIT(tcphptsi, SI_SUB_SOFTINTR, SI_ORDER_ANY, tcp_init_hptsi, NULL); MODULE_VERSION(tcphpts, 1); diff --git a/sys/netinet/tcp_log_buf.c b/sys/netinet/tcp_log_buf.c index 5ec4acf367d5..6baf1bce4623 100644 --- a/sys/netinet/tcp_log_buf.c +++ b/sys/netinet/tcp_log_buf.c @@ -1,2639 +1,2639 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2016-2018 Netflix, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include /* Must come after qmath.h and tree.h */ #include #include #include #include #include #include #include #include #include #include /* Default expiry time */ #define TCP_LOG_EXPIRE_TIME ((sbintime_t)60 * SBT_1S) /* Max interval at which to run the expiry timer */ #define TCP_LOG_EXPIRE_INTVL ((sbintime_t)5 * SBT_1S) bool tcp_log_verbose; static uma_zone_t tcp_log_id_bucket_zone, tcp_log_id_node_zone, tcp_log_zone; static int tcp_log_session_limit = TCP_LOG_BUF_DEFAULT_SESSION_LIMIT; static uint32_t tcp_log_version = TCP_LOG_BUF_VER; RB_HEAD(tcp_log_id_tree, tcp_log_id_bucket); static struct tcp_log_id_tree tcp_log_id_head; static STAILQ_HEAD(, tcp_log_id_node) tcp_log_expireq_head = STAILQ_HEAD_INITIALIZER(tcp_log_expireq_head); static struct mtx tcp_log_expireq_mtx; static struct callout tcp_log_expireq_callout; static u_long tcp_log_auto_ratio = 0; static volatile u_long tcp_log_auto_ratio_cur = 0; static uint32_t tcp_log_auto_mode = TCP_LOG_STATE_TAIL; static bool tcp_log_auto_all = false; static uint32_t tcp_disable_all_bb_logs = 0; RB_PROTOTYPE_STATIC(tcp_log_id_tree, tcp_log_id_bucket, tlb_rb, tcp_log_id_cmp) SYSCTL_NODE(_net_inet_tcp, OID_AUTO, bb, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "TCP Black Box controls"); SYSCTL_BOOL(_net_inet_tcp_bb, OID_AUTO, log_verbose, CTLFLAG_RW, &tcp_log_verbose, 0, "Force verbose logging for TCP traces"); SYSCTL_INT(_net_inet_tcp_bb, OID_AUTO, log_session_limit, CTLFLAG_RW, &tcp_log_session_limit, 0, "Maximum number of events maintained for each TCP session"); SYSCTL_UMA_MAX(_net_inet_tcp_bb, OID_AUTO, log_global_limit, CTLFLAG_RW, &tcp_log_zone, "Maximum number of events maintained for all TCP sessions"); SYSCTL_UMA_CUR(_net_inet_tcp_bb, OID_AUTO, log_global_entries, CTLFLAG_RD, &tcp_log_zone, "Current number of events maintained for all TCP sessions"); SYSCTL_UMA_MAX(_net_inet_tcp_bb, OID_AUTO, log_id_limit, CTLFLAG_RW, &tcp_log_id_bucket_zone, "Maximum number of log IDs"); SYSCTL_UMA_CUR(_net_inet_tcp_bb, OID_AUTO, log_id_entries, CTLFLAG_RD, &tcp_log_id_bucket_zone, "Current number of log IDs"); SYSCTL_UMA_MAX(_net_inet_tcp_bb, OID_AUTO, log_id_tcpcb_limit, CTLFLAG_RW, &tcp_log_id_node_zone, "Maximum number of tcpcbs with log IDs"); SYSCTL_UMA_CUR(_net_inet_tcp_bb, OID_AUTO, log_id_tcpcb_entries, CTLFLAG_RD, &tcp_log_id_node_zone, "Current number of tcpcbs with log IDs"); SYSCTL_U32(_net_inet_tcp_bb, OID_AUTO, log_version, CTLFLAG_RD, &tcp_log_version, 0, "Version of log formats exported"); SYSCTL_U32(_net_inet_tcp_bb, OID_AUTO, disable_all, CTLFLAG_RW, &tcp_disable_all_bb_logs, TCP_LOG_STATE_HEAD_AUTO, "Disable all BB logging for all connections"); SYSCTL_ULONG(_net_inet_tcp_bb, OID_AUTO, log_auto_ratio, CTLFLAG_RW, &tcp_log_auto_ratio, 0, "Do auto capturing for 1 out of N sessions"); SYSCTL_U32(_net_inet_tcp_bb, OID_AUTO, log_auto_mode, CTLFLAG_RW, &tcp_log_auto_mode, TCP_LOG_STATE_HEAD_AUTO, "Logging mode for auto-selected sessions (default is TCP_LOG_STATE_HEAD_AUTO)"); SYSCTL_BOOL(_net_inet_tcp_bb, OID_AUTO, log_auto_all, CTLFLAG_RW, &tcp_log_auto_all, false, "Auto-select from all sessions (rather than just those with IDs)"); #ifdef TCPLOG_DEBUG_COUNTERS counter_u64_t tcp_log_queued; counter_u64_t tcp_log_que_fail1; counter_u64_t tcp_log_que_fail2; counter_u64_t tcp_log_que_fail3; counter_u64_t tcp_log_que_fail4; counter_u64_t tcp_log_que_fail5; counter_u64_t tcp_log_que_copyout; counter_u64_t tcp_log_que_read; counter_u64_t tcp_log_que_freed; SYSCTL_COUNTER_U64(_net_inet_tcp_bb, OID_AUTO, queued, CTLFLAG_RD, &tcp_log_queued, "Number of entries queued"); SYSCTL_COUNTER_U64(_net_inet_tcp_bb, OID_AUTO, fail1, CTLFLAG_RD, &tcp_log_que_fail1, "Number of entries queued but fail 1"); SYSCTL_COUNTER_U64(_net_inet_tcp_bb, OID_AUTO, fail2, CTLFLAG_RD, &tcp_log_que_fail2, "Number of entries queued but fail 2"); SYSCTL_COUNTER_U64(_net_inet_tcp_bb, OID_AUTO, fail3, CTLFLAG_RD, &tcp_log_que_fail3, "Number of entries queued but fail 3"); SYSCTL_COUNTER_U64(_net_inet_tcp_bb, OID_AUTO, fail4, CTLFLAG_RD, &tcp_log_que_fail4, "Number of entries queued but fail 4"); SYSCTL_COUNTER_U64(_net_inet_tcp_bb, OID_AUTO, fail5, CTLFLAG_RD, &tcp_log_que_fail5, "Number of entries queued but fail 4"); SYSCTL_COUNTER_U64(_net_inet_tcp_bb, OID_AUTO, copyout, CTLFLAG_RD, &tcp_log_que_copyout, "Number of entries copied out"); SYSCTL_COUNTER_U64(_net_inet_tcp_bb, OID_AUTO, read, CTLFLAG_RD, &tcp_log_que_read, "Number of entries read from the queue"); SYSCTL_COUNTER_U64(_net_inet_tcp_bb, OID_AUTO, freed, CTLFLAG_RD, &tcp_log_que_freed, "Number of entries freed after reading"); #endif #ifdef INVARIANTS #define TCPLOG_DEBUG_RINGBUF #endif /* Number of requests to consider a PBCID "active". */ #define ACTIVE_REQUEST_COUNT 10 /* Statistic tracking for "active" PBCIDs. */ static counter_u64_t tcp_log_pcb_ids_cur; static counter_u64_t tcp_log_pcb_ids_tot; SYSCTL_COUNTER_U64(_net_inet_tcp_bb, OID_AUTO, pcb_ids_cur, CTLFLAG_RD, &tcp_log_pcb_ids_cur, "Number of pcb IDs allocated in the system"); SYSCTL_COUNTER_U64(_net_inet_tcp_bb, OID_AUTO, pcb_ids_tot, CTLFLAG_RD, &tcp_log_pcb_ids_tot, "Total number of pcb IDs that have been allocated"); struct tcp_log_mem { STAILQ_ENTRY(tcp_log_mem) tlm_queue; struct tcp_log_buffer tlm_buf; struct tcp_log_verbose tlm_v; #ifdef TCPLOG_DEBUG_RINGBUF volatile int tlm_refcnt; #endif }; /* 60 bytes for the header, + 16 bytes for padding */ static uint8_t zerobuf[76]; /* * Lock order: * 1. TCPID_TREE * 2. TCPID_BUCKET * 3. INP * * Rules: * A. You need a lock on the Tree to add/remove buckets. * B. You need a lock on the bucket to add/remove nodes from the bucket. * C. To change information in a node, you need the INP lock if the tln_closed * field is false. Otherwise, you need the bucket lock. (Note that the * tln_closed field can change at any point, so you need to recheck the * entry after acquiring the INP lock.) * D. To remove a node from the bucket, you must have that entry locked, * according to the criteria of Rule C. Also, the node must not be on * the expiry queue. * E. The exception to C is the expiry queue fields, which are locked by * the TCPLOG_EXPIREQ lock. * * Buckets have a reference count. Each node is a reference. Further, * other callers may add reference counts to keep a bucket from disappearing. * You can add a reference as long as you own a lock sufficient to keep the * bucket from disappearing. For example, a common use is: * a. Have a locked INP, but need to lock the TCPID_BUCKET. * b. Add a refcount on the bucket. (Safe because the INP lock prevents * the TCPID_BUCKET from going away.) * c. Drop the INP lock. * d. Acquire a lock on the TCPID_BUCKET. * e. Acquire a lock on the INP. * f. Drop the refcount on the bucket. * (At this point, the bucket may disappear.) * * Expire queue lock: * You can acquire this with either the bucket or INP lock. Don't reverse it. * When the expire code has committed to freeing a node, it resets the expiry * time to SBT_MAX. That is the signal to everyone else that they should * leave that node alone. */ static struct rwlock tcp_id_tree_lock; #define TCPID_TREE_WLOCK() rw_wlock(&tcp_id_tree_lock) #define TCPID_TREE_RLOCK() rw_rlock(&tcp_id_tree_lock) #define TCPID_TREE_UPGRADE() rw_try_upgrade(&tcp_id_tree_lock) #define TCPID_TREE_WUNLOCK() rw_wunlock(&tcp_id_tree_lock) #define TCPID_TREE_RUNLOCK() rw_runlock(&tcp_id_tree_lock) #define TCPID_TREE_WLOCK_ASSERT() rw_assert(&tcp_id_tree_lock, RA_WLOCKED) #define TCPID_TREE_RLOCK_ASSERT() rw_assert(&tcp_id_tree_lock, RA_RLOCKED) #define TCPID_TREE_UNLOCK_ASSERT() rw_assert(&tcp_id_tree_lock, RA_UNLOCKED) #define TCPID_BUCKET_LOCK_INIT(tlb) mtx_init(&((tlb)->tlb_mtx), "tcp log id bucket", NULL, MTX_DEF) #define TCPID_BUCKET_LOCK_DESTROY(tlb) mtx_destroy(&((tlb)->tlb_mtx)) #define TCPID_BUCKET_LOCK(tlb) mtx_lock(&((tlb)->tlb_mtx)) #define TCPID_BUCKET_UNLOCK(tlb) mtx_unlock(&((tlb)->tlb_mtx)) #define TCPID_BUCKET_LOCK_ASSERT(tlb) mtx_assert(&((tlb)->tlb_mtx), MA_OWNED) #define TCPID_BUCKET_UNLOCK_ASSERT(tlb) mtx_assert(&((tlb)->tlb_mtx), MA_NOTOWNED) #define TCPID_BUCKET_REF(tlb) refcount_acquire(&((tlb)->tlb_refcnt)) #define TCPID_BUCKET_UNREF(tlb) refcount_release(&((tlb)->tlb_refcnt)) #define TCPLOG_EXPIREQ_LOCK() mtx_lock(&tcp_log_expireq_mtx) #define TCPLOG_EXPIREQ_UNLOCK() mtx_unlock(&tcp_log_expireq_mtx) SLIST_HEAD(tcp_log_id_head, tcp_log_id_node); struct tcp_log_id_bucket { /* * tlb_id must be first. This lets us use strcmp on * (struct tcp_log_id_bucket *) and (char *) interchangeably. */ char tlb_id[TCP_LOG_ID_LEN]; char tlb_tag[TCP_LOG_TAG_LEN]; RB_ENTRY(tcp_log_id_bucket) tlb_rb; struct tcp_log_id_head tlb_head; struct mtx tlb_mtx; volatile u_int tlb_refcnt; volatile u_int tlb_reqcnt; uint32_t tlb_loglimit; uint8_t tlb_logstate; }; struct tcp_log_id_node { SLIST_ENTRY(tcp_log_id_node) tln_list; STAILQ_ENTRY(tcp_log_id_node) tln_expireq; /* Locked by the expireq lock */ sbintime_t tln_expiretime; /* Locked by the expireq lock */ /* * If INP is NULL, that means the connection has closed. We've * saved the connection endpoint information and the log entries * in the tln_ie and tln_entries members. We've also saved a pointer * to the enclosing bucket here. If INP is not NULL, the information is * in the PCB and not here. */ struct inpcb *tln_inp; struct tcpcb *tln_tp; struct tcp_log_id_bucket *tln_bucket; struct in_endpoints tln_ie; struct tcp_log_stailq tln_entries; int tln_count; volatile int tln_closed; uint8_t tln_af; }; enum tree_lock_state { TREE_UNLOCKED = 0, TREE_RLOCKED, TREE_WLOCKED, }; /* Do we want to select this session for auto-logging? */ static __inline bool tcp_log_selectauto(void) { /* * If we are doing auto-capturing, figure out whether we will capture * this session. */ if (tcp_log_auto_ratio && (tcp_disable_all_bb_logs == 0) && (atomic_fetchadd_long(&tcp_log_auto_ratio_cur, 1) % tcp_log_auto_ratio) == 0) return (true); return (false); } static __inline int tcp_log_id_cmp(struct tcp_log_id_bucket *a, struct tcp_log_id_bucket *b) { KASSERT(a != NULL, ("tcp_log_id_cmp: argument a is unexpectedly NULL")); KASSERT(b != NULL, ("tcp_log_id_cmp: argument b is unexpectedly NULL")); return strncmp(a->tlb_id, b->tlb_id, TCP_LOG_ID_LEN); } RB_GENERATE_STATIC(tcp_log_id_tree, tcp_log_id_bucket, tlb_rb, tcp_log_id_cmp) static __inline void tcp_log_id_validate_tree_lock(int tree_locked) { #ifdef INVARIANTS switch (tree_locked) { case TREE_WLOCKED: TCPID_TREE_WLOCK_ASSERT(); break; case TREE_RLOCKED: TCPID_TREE_RLOCK_ASSERT(); break; case TREE_UNLOCKED: TCPID_TREE_UNLOCK_ASSERT(); break; default: kassert_panic("%s:%d: unknown tree lock state", __func__, __LINE__); } #endif } static __inline void tcp_log_remove_bucket(struct tcp_log_id_bucket *tlb) { TCPID_TREE_WLOCK_ASSERT(); KASSERT(SLIST_EMPTY(&tlb->tlb_head), ("%s: Attempt to remove non-empty bucket", __func__)); if (RB_REMOVE(tcp_log_id_tree, &tcp_log_id_head, tlb) == NULL) { #ifdef INVARIANTS kassert_panic("%s:%d: error removing element from tree", __func__, __LINE__); #endif } TCPID_BUCKET_LOCK_DESTROY(tlb); counter_u64_add(tcp_log_pcb_ids_cur, (int64_t)-1); uma_zfree(tcp_log_id_bucket_zone, tlb); } /* * Call with a referenced and locked bucket. * Will return true if the bucket was freed; otherwise, false. * tlb: The bucket to unreference. * tree_locked: A pointer to the state of the tree lock. If the tree lock * state changes, the function will update it. * inp: If not NULL and the function needs to drop the inp lock to relock the * tree, it will do so. (The caller must ensure inp will not become invalid, * probably by holding a reference to it.) */ static bool tcp_log_unref_bucket(struct tcp_log_id_bucket *tlb, int *tree_locked, struct inpcb *inp) { KASSERT(tlb != NULL, ("%s: called with NULL tlb", __func__)); KASSERT(tree_locked != NULL, ("%s: called with NULL tree_locked", __func__)); tcp_log_id_validate_tree_lock(*tree_locked); /* * Did we hold the last reference on the tlb? If so, we may need * to free it. (Note that we can realistically only execute the * loop twice: once without a write lock and once with a write * lock.) */ while (TCPID_BUCKET_UNREF(tlb)) { /* * We need a write lock on the tree to free this. * If we can upgrade the tree lock, this is "easy". If we * can't upgrade the tree lock, we need to do this the * "hard" way: unwind all our locks and relock everything. * In the meantime, anything could have changed. We even * need to validate that we still need to free the bucket. */ if (*tree_locked == TREE_RLOCKED && TCPID_TREE_UPGRADE()) *tree_locked = TREE_WLOCKED; else if (*tree_locked != TREE_WLOCKED) { TCPID_BUCKET_REF(tlb); if (inp != NULL) INP_WUNLOCK(inp); TCPID_BUCKET_UNLOCK(tlb); if (*tree_locked == TREE_RLOCKED) TCPID_TREE_RUNLOCK(); TCPID_TREE_WLOCK(); *tree_locked = TREE_WLOCKED; TCPID_BUCKET_LOCK(tlb); if (inp != NULL) INP_WLOCK(inp); continue; } /* * We have an empty bucket and a write lock on the tree. * Remove the empty bucket. */ tcp_log_remove_bucket(tlb); return (true); } return (false); } /* * Call with a locked bucket. This function will release the lock on the * bucket before returning. * * The caller is responsible for freeing the tp->t_lin/tln node! * * Note: one of tp or both tlb and tln must be supplied. * * inp: A pointer to the inp. If the function needs to drop the inp lock to * acquire the tree write lock, it will do so. (The caller must ensure inp * will not become invalid, probably by holding a reference to it.) * tp: A pointer to the tcpcb. (optional; if specified, tlb and tln are ignored) * tlb: A pointer to the bucket. (optional; ignored if tp is specified) * tln: A pointer to the node. (optional; ignored if tp is specified) * tree_locked: A pointer to the state of the tree lock. If the tree lock * state changes, the function will update it. * * Will return true if the INP lock was reacquired; otherwise, false. */ static bool tcp_log_remove_id_node(struct inpcb *inp, struct tcpcb *tp, struct tcp_log_id_bucket *tlb, struct tcp_log_id_node *tln, int *tree_locked) { int orig_tree_locked; KASSERT(tp != NULL || (tlb != NULL && tln != NULL), ("%s: called with tp=%p, tlb=%p, tln=%p", __func__, tp, tlb, tln)); KASSERT(tree_locked != NULL, ("%s: called with NULL tree_locked", __func__)); if (tp != NULL) { tlb = tp->t_lib; tln = tp->t_lin; KASSERT(tlb != NULL, ("%s: unexpectedly NULL tlb", __func__)); KASSERT(tln != NULL, ("%s: unexpectedly NULL tln", __func__)); } tcp_log_id_validate_tree_lock(*tree_locked); TCPID_BUCKET_LOCK_ASSERT(tlb); /* * Remove the node, clear the log bucket and node from the TCPCB, and * decrement the bucket refcount. In the process, if this is the * last reference, the bucket will be freed. */ SLIST_REMOVE(&tlb->tlb_head, tln, tcp_log_id_node, tln_list); if (tp != NULL) { tp->t_lib = NULL; tp->t_lin = NULL; } orig_tree_locked = *tree_locked; if (!tcp_log_unref_bucket(tlb, tree_locked, inp)) TCPID_BUCKET_UNLOCK(tlb); return (*tree_locked != orig_tree_locked); } #define RECHECK_INP_CLEAN(cleanup) do { \ - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { \ + if (inp->inp_flags & INP_DROPPED) { \ rv = ECONNRESET; \ cleanup; \ goto done; \ } \ tp = intotcpcb(inp); \ } while (0) #define RECHECK_INP() RECHECK_INP_CLEAN(/* noop */) static void tcp_log_grow_tlb(char *tlb_id, struct tcpcb *tp) { INP_WLOCK_ASSERT(tp->t_inpcb); #ifdef STATS if (V_tcp_perconn_stats_enable == 2 && tp->t_stats == NULL) (void)tcp_stats_sample_rollthedice(tp, tlb_id, strlen(tlb_id)); #endif } static void tcp_log_increment_reqcnt(struct tcp_log_id_bucket *tlb) { atomic_fetchadd_int(&tlb->tlb_reqcnt, 1); } /* * Associate the specified tag with a particular TCP log ID. * Called with INPCB locked. Returns with it unlocked. * Returns 0 on success or EOPNOTSUPP if the connection has no TCP log ID. */ int tcp_log_set_tag(struct tcpcb *tp, char *tag) { struct tcp_log_id_bucket *tlb; int tree_locked; INP_WLOCK_ASSERT(tp->t_inpcb); tree_locked = TREE_UNLOCKED; tlb = tp->t_lib; if (tlb == NULL) { INP_WUNLOCK(tp->t_inpcb); return (EOPNOTSUPP); } TCPID_BUCKET_REF(tlb); INP_WUNLOCK(tp->t_inpcb); TCPID_BUCKET_LOCK(tlb); strlcpy(tlb->tlb_tag, tag, TCP_LOG_TAG_LEN); if (!tcp_log_unref_bucket(tlb, &tree_locked, NULL)) TCPID_BUCKET_UNLOCK(tlb); if (tree_locked == TREE_WLOCKED) { TCPID_TREE_WLOCK_ASSERT(); TCPID_TREE_WUNLOCK(); } else if (tree_locked == TREE_RLOCKED) { TCPID_TREE_RLOCK_ASSERT(); TCPID_TREE_RUNLOCK(); } else TCPID_TREE_UNLOCK_ASSERT(); return (0); } /* * Set the TCP log ID for a TCPCB. * Called with INPCB locked. Returns with it unlocked. */ int tcp_log_set_id(struct tcpcb *tp, char *id) { struct tcp_log_id_bucket *tlb, *tmp_tlb; struct tcp_log_id_node *tln; struct inpcb *inp; int tree_locked, rv; bool bucket_locked; tlb = NULL; tln = NULL; inp = tp->t_inpcb; tree_locked = TREE_UNLOCKED; bucket_locked = false; restart: INP_WLOCK_ASSERT(inp); /* See if the ID is unchanged. */ if ((tp->t_lib != NULL && !strcmp(tp->t_lib->tlb_id, id)) || (tp->t_lib == NULL && *id == 0)) { if (tp->t_lib != NULL) { tcp_log_increment_reqcnt(tp->t_lib); if ((tp->t_lib->tlb_logstate) && (tp->t_log_state_set == 0)) { /* Clone in any logging */ tp->t_logstate = tp->t_lib->tlb_logstate; } if ((tp->t_lib->tlb_loglimit) && (tp->t_log_state_set == 0)) { /* We also have a limit set */ tp->t_loglimit = tp->t_lib->tlb_loglimit; } } rv = 0; goto done; } /* * If the TCPCB had a previous ID, we need to extricate it from * the previous list. * * Drop the TCPCB lock and lock the tree and the bucket. * Because this is called in the socket context, we (theoretically) * don't need to worry about the INPCB completely going away * while we are gone. */ if (tp->t_lib != NULL) { tlb = tp->t_lib; TCPID_BUCKET_REF(tlb); INP_WUNLOCK(inp); if (tree_locked == TREE_UNLOCKED) { TCPID_TREE_RLOCK(); tree_locked = TREE_RLOCKED; } TCPID_BUCKET_LOCK(tlb); bucket_locked = true; INP_WLOCK(inp); /* * Unreference the bucket. If our bucket went away, it is no * longer locked or valid. */ if (tcp_log_unref_bucket(tlb, &tree_locked, inp)) { bucket_locked = false; tlb = NULL; } /* Validate the INP. */ RECHECK_INP(); /* * Evaluate whether the bucket changed while we were unlocked. * * Possible scenarios here: * 1. Bucket is unchanged and the same one we started with. * 2. The TCPCB no longer has a bucket and our bucket was * freed. * 3. The TCPCB has a new bucket, whether ours was freed. * 4. The TCPCB no longer has a bucket and our bucket was * not freed. * * In cases 2-4, we will start over. In case 1, we will * proceed here to remove the bucket. */ if (tlb == NULL || tp->t_lib != tlb) { KASSERT(bucket_locked || tlb == NULL, ("%s: bucket_locked (%d) and tlb (%p) are " "inconsistent", __func__, bucket_locked, tlb)); if (bucket_locked) { TCPID_BUCKET_UNLOCK(tlb); bucket_locked = false; tlb = NULL; } goto restart; } /* * Store the (struct tcp_log_id_node) for reuse. Then, remove * it from the bucket. In the process, we may end up relocking. * If so, we need to validate that the INP is still valid, and * the TCPCB entries match we expect. * * We will clear tlb and change the bucket_locked state just * before calling tcp_log_remove_id_node(), since that function * will unlock the bucket. */ if (tln != NULL) uma_zfree(tcp_log_id_node_zone, tln); tln = tp->t_lin; tlb = NULL; bucket_locked = false; if (tcp_log_remove_id_node(inp, tp, NULL, NULL, &tree_locked)) { RECHECK_INP(); /* * If the TCPCB moved to a new bucket while we had * dropped the lock, restart. */ if (tp->t_lib != NULL || tp->t_lin != NULL) goto restart; } /* * Yay! We successfully removed the TCPCB from its old * bucket. Phew! * * On to bigger and better things... */ } /* At this point, the TCPCB should not be in any bucket. */ KASSERT(tp->t_lib == NULL, ("%s: tp->t_lib is not NULL", __func__)); /* * If the new ID is not empty, we need to now assign this TCPCB to a * new bucket. */ if (*id) { /* Get a new tln, if we don't already have one to reuse. */ if (tln == NULL) { tln = uma_zalloc(tcp_log_id_node_zone, M_NOWAIT | M_ZERO); if (tln == NULL) { rv = ENOBUFS; goto done; } tln->tln_inp = inp; tln->tln_tp = tp; } /* * Drop the INP lock for a bit. We don't need it, and dropping * it prevents lock order reversals. */ INP_WUNLOCK(inp); /* Make sure we have at least a read lock on the tree. */ tcp_log_id_validate_tree_lock(tree_locked); if (tree_locked == TREE_UNLOCKED) { TCPID_TREE_RLOCK(); tree_locked = TREE_RLOCKED; } refind: /* * Remember that we constructed (struct tcp_log_id_node) so * we can safely cast the id to it for the purposes of finding. */ KASSERT(tlb == NULL, ("%s:%d tlb unexpectedly non-NULL", __func__, __LINE__)); tmp_tlb = RB_FIND(tcp_log_id_tree, &tcp_log_id_head, (struct tcp_log_id_bucket *) id); /* * If we didn't find a matching bucket, we need to add a new * one. This requires a write lock. But, of course, we will * need to recheck some things when we re-acquire the lock. */ if (tmp_tlb == NULL && tree_locked != TREE_WLOCKED) { tree_locked = TREE_WLOCKED; if (!TCPID_TREE_UPGRADE()) { TCPID_TREE_RUNLOCK(); TCPID_TREE_WLOCK(); /* * The tree may have changed while we were * unlocked. */ goto refind; } } /* If we need to add a new bucket, do it now. */ if (tmp_tlb == NULL) { /* Allocate new bucket. */ tlb = uma_zalloc(tcp_log_id_bucket_zone, M_NOWAIT); if (tlb == NULL) { rv = ENOBUFS; goto done_noinp; } counter_u64_add(tcp_log_pcb_ids_cur, 1); counter_u64_add(tcp_log_pcb_ids_tot, 1); if ((tcp_log_auto_all == false) && tcp_log_auto_mode && tcp_log_selectauto()) { /* Save off the log state */ tlb->tlb_logstate = tcp_log_auto_mode; } else tlb->tlb_logstate = TCP_LOG_STATE_OFF; tlb->tlb_loglimit = 0; tlb->tlb_tag[0] = '\0'; /* Default to an empty tag. */ /* * Copy the ID to the bucket. * NB: Don't use strlcpy() unless you are sure * we've always validated NULL termination. * * TODO: When I'm done writing this, see if we * we have correctly validated NULL termination and * can use strlcpy(). :-) */ strncpy(tlb->tlb_id, id, TCP_LOG_ID_LEN - 1); tlb->tlb_id[TCP_LOG_ID_LEN - 1] = '\0'; /* * Take the refcount for the first node and go ahead * and lock this. Note that we zero the tlb_mtx * structure, since 0xdeadc0de flips the right bits * for the code to think that this mutex has already * been initialized. :-( */ SLIST_INIT(&tlb->tlb_head); refcount_init(&tlb->tlb_refcnt, 1); tlb->tlb_reqcnt = 1; memset(&tlb->tlb_mtx, 0, sizeof(struct mtx)); TCPID_BUCKET_LOCK_INIT(tlb); TCPID_BUCKET_LOCK(tlb); bucket_locked = true; #define FREE_NEW_TLB() do { \ TCPID_BUCKET_LOCK_DESTROY(tlb); \ uma_zfree(tcp_log_id_bucket_zone, tlb); \ counter_u64_add(tcp_log_pcb_ids_cur, (int64_t)-1); \ counter_u64_add(tcp_log_pcb_ids_tot, (int64_t)-1); \ bucket_locked = false; \ tlb = NULL; \ } while (0) /* * Relock the INP and make sure we are still * unassigned. */ INP_WLOCK(inp); RECHECK_INP_CLEAN(FREE_NEW_TLB()); if (tp->t_lib != NULL) { FREE_NEW_TLB(); goto restart; } /* Add the new bucket to the tree. */ tmp_tlb = RB_INSERT(tcp_log_id_tree, &tcp_log_id_head, tlb); KASSERT(tmp_tlb == NULL, ("%s: Unexpected conflicting bucket (%p) while " "adding new bucket (%p)", __func__, tmp_tlb, tlb)); /* * If we found a conflicting bucket, free the new * one we made and fall through to use the existing * bucket. */ if (tmp_tlb != NULL) { FREE_NEW_TLB(); INP_WUNLOCK(inp); } #undef FREE_NEW_TLB } /* If we found an existing bucket, use it. */ if (tmp_tlb != NULL) { tlb = tmp_tlb; TCPID_BUCKET_LOCK(tlb); bucket_locked = true; /* * Relock the INP and make sure we are still * unassigned. */ INP_UNLOCK_ASSERT(inp); INP_WLOCK(inp); RECHECK_INP(); if (tp->t_lib != NULL) { TCPID_BUCKET_UNLOCK(tlb); bucket_locked = false; tlb = NULL; goto restart; } /* Take a reference on the bucket. */ TCPID_BUCKET_REF(tlb); /* Record the request. */ tcp_log_increment_reqcnt(tlb); } tcp_log_grow_tlb(tlb->tlb_id, tp); /* Add the new node to the list. */ SLIST_INSERT_HEAD(&tlb->tlb_head, tln, tln_list); tp->t_lib = tlb; tp->t_lin = tln; if (tp->t_lib->tlb_logstate) { /* Clone in any logging */ tp->t_logstate = tp->t_lib->tlb_logstate; } if (tp->t_lib->tlb_loglimit) { /* The loglimit too */ tp->t_loglimit = tp->t_lib->tlb_loglimit; } tln = NULL; } rv = 0; done: /* Unlock things, as needed, and return. */ INP_WUNLOCK(inp); done_noinp: INP_UNLOCK_ASSERT(inp); if (bucket_locked) { TCPID_BUCKET_LOCK_ASSERT(tlb); TCPID_BUCKET_UNLOCK(tlb); } else if (tlb != NULL) TCPID_BUCKET_UNLOCK_ASSERT(tlb); if (tree_locked == TREE_WLOCKED) { TCPID_TREE_WLOCK_ASSERT(); TCPID_TREE_WUNLOCK(); } else if (tree_locked == TREE_RLOCKED) { TCPID_TREE_RLOCK_ASSERT(); TCPID_TREE_RUNLOCK(); } else TCPID_TREE_UNLOCK_ASSERT(); if (tln != NULL) uma_zfree(tcp_log_id_node_zone, tln); return (rv); } /* * Get the TCP log ID for a TCPCB. * Called with INPCB locked. * 'buf' must point to a buffer that is at least TCP_LOG_ID_LEN bytes long. * Returns number of bytes copied. */ size_t tcp_log_get_id(struct tcpcb *tp, char *buf) { size_t len; INP_LOCK_ASSERT(tp->t_inpcb); if (tp->t_lib != NULL) { len = strlcpy(buf, tp->t_lib->tlb_id, TCP_LOG_ID_LEN); KASSERT(len < TCP_LOG_ID_LEN, ("%s:%d: tp->t_lib->tlb_id too long (%zu)", __func__, __LINE__, len)); } else { *buf = '\0'; len = 0; } return (len); } /* * Get the tag associated with the TCPCB's log ID. * Called with INPCB locked. Returns with it unlocked. * 'buf' must point to a buffer that is at least TCP_LOG_TAG_LEN bytes long. * Returns number of bytes copied. */ size_t tcp_log_get_tag(struct tcpcb *tp, char *buf) { struct tcp_log_id_bucket *tlb; size_t len; int tree_locked; INP_WLOCK_ASSERT(tp->t_inpcb); tree_locked = TREE_UNLOCKED; tlb = tp->t_lib; if (tlb != NULL) { TCPID_BUCKET_REF(tlb); INP_WUNLOCK(tp->t_inpcb); TCPID_BUCKET_LOCK(tlb); len = strlcpy(buf, tlb->tlb_tag, TCP_LOG_TAG_LEN); KASSERT(len < TCP_LOG_TAG_LEN, ("%s:%d: tp->t_lib->tlb_tag too long (%zu)", __func__, __LINE__, len)); if (!tcp_log_unref_bucket(tlb, &tree_locked, NULL)) TCPID_BUCKET_UNLOCK(tlb); if (tree_locked == TREE_WLOCKED) { TCPID_TREE_WLOCK_ASSERT(); TCPID_TREE_WUNLOCK(); } else if (tree_locked == TREE_RLOCKED) { TCPID_TREE_RLOCK_ASSERT(); TCPID_TREE_RUNLOCK(); } else TCPID_TREE_UNLOCK_ASSERT(); } else { INP_WUNLOCK(tp->t_inpcb); *buf = '\0'; len = 0; } return (len); } /* * Get number of connections with the same log ID. * Log ID is taken from given TCPCB. * Called with INPCB locked. */ u_int tcp_log_get_id_cnt(struct tcpcb *tp) { INP_WLOCK_ASSERT(tp->t_inpcb); return ((tp->t_lib == NULL) ? 0 : tp->t_lib->tlb_refcnt); } #ifdef TCPLOG_DEBUG_RINGBUF /* * Functions/macros to increment/decrement reference count for a log * entry. This should catch when we do a double-free/double-remove or * a double-add. */ static inline void _tcp_log_entry_refcnt_add(struct tcp_log_mem *log_entry, const char *func, int line) { int refcnt; refcnt = atomic_fetchadd_int(&log_entry->tlm_refcnt, 1); if (refcnt != 0) panic("%s:%d: log_entry(%p)->tlm_refcnt is %d (expected 0)", func, line, log_entry, refcnt); } #define tcp_log_entry_refcnt_add(l) \ _tcp_log_entry_refcnt_add((l), __func__, __LINE__) static inline void _tcp_log_entry_refcnt_rem(struct tcp_log_mem *log_entry, const char *func, int line) { int refcnt; refcnt = atomic_fetchadd_int(&log_entry->tlm_refcnt, -1); if (refcnt != 1) panic("%s:%d: log_entry(%p)->tlm_refcnt is %d (expected 1)", func, line, log_entry, refcnt); } #define tcp_log_entry_refcnt_rem(l) \ _tcp_log_entry_refcnt_rem((l), __func__, __LINE__) #else /* !TCPLOG_DEBUG_RINGBUF */ #define tcp_log_entry_refcnt_add(l) #define tcp_log_entry_refcnt_rem(l) #endif /* * Cleanup after removing a log entry, but only decrement the count if we * are running INVARIANTS. */ static inline void tcp_log_free_log_common(struct tcp_log_mem *log_entry, int *count __unused) { uma_zfree(tcp_log_zone, log_entry); #ifdef INVARIANTS (*count)--; KASSERT(*count >= 0, ("%s: count unexpectedly negative", __func__)); #endif } static void tcp_log_free_entries(struct tcp_log_stailq *head, int *count) { struct tcp_log_mem *log_entry; /* Free the entries. */ while ((log_entry = STAILQ_FIRST(head)) != NULL) { STAILQ_REMOVE_HEAD(head, tlm_queue); tcp_log_entry_refcnt_rem(log_entry); tcp_log_free_log_common(log_entry, count); } } /* Cleanup after removing a log entry. */ static inline void tcp_log_remove_log_cleanup(struct tcpcb *tp, struct tcp_log_mem *log_entry) { uma_zfree(tcp_log_zone, log_entry); tp->t_lognum--; KASSERT(tp->t_lognum >= 0, ("%s: tp->t_lognum unexpectedly negative", __func__)); } /* Remove a log entry from the head of a list. */ static inline void tcp_log_remove_log_head(struct tcpcb *tp, struct tcp_log_mem *log_entry) { KASSERT(log_entry == STAILQ_FIRST(&tp->t_logs), ("%s: attempt to remove non-HEAD log entry", __func__)); STAILQ_REMOVE_HEAD(&tp->t_logs, tlm_queue); tcp_log_entry_refcnt_rem(log_entry); tcp_log_remove_log_cleanup(tp, log_entry); } #ifdef TCPLOG_DEBUG_RINGBUF /* * Initialize the log entry's reference count, which we want to * survive allocations. */ static int tcp_log_zone_init(void *mem, int size, int flags __unused) { struct tcp_log_mem *tlm; KASSERT(size >= sizeof(struct tcp_log_mem), ("%s: unexpectedly short (%d) allocation", __func__, size)); tlm = (struct tcp_log_mem *)mem; tlm->tlm_refcnt = 0; return (0); } /* * Double check that the refcnt is zero on allocation and return. */ static int tcp_log_zone_ctor(void *mem, int size, void *args __unused, int flags __unused) { struct tcp_log_mem *tlm; KASSERT(size >= sizeof(struct tcp_log_mem), ("%s: unexpectedly short (%d) allocation", __func__, size)); tlm = (struct tcp_log_mem *)mem; if (tlm->tlm_refcnt != 0) panic("%s:%d: tlm(%p)->tlm_refcnt is %d (expected 0)", __func__, __LINE__, tlm, tlm->tlm_refcnt); return (0); } static void tcp_log_zone_dtor(void *mem, int size, void *args __unused) { struct tcp_log_mem *tlm; KASSERT(size >= sizeof(struct tcp_log_mem), ("%s: unexpectedly short (%d) allocation", __func__, size)); tlm = (struct tcp_log_mem *)mem; if (tlm->tlm_refcnt != 0) panic("%s:%d: tlm(%p)->tlm_refcnt is %d (expected 0)", __func__, __LINE__, tlm, tlm->tlm_refcnt); } #endif /* TCPLOG_DEBUG_RINGBUF */ /* Do global initialization. */ void tcp_log_init(void) { tcp_log_zone = uma_zcreate("tcp_log", sizeof(struct tcp_log_mem), #ifdef TCPLOG_DEBUG_RINGBUF tcp_log_zone_ctor, tcp_log_zone_dtor, tcp_log_zone_init, #else NULL, NULL, NULL, #endif NULL, UMA_ALIGN_PTR, 0); (void)uma_zone_set_max(tcp_log_zone, TCP_LOG_BUF_DEFAULT_GLOBAL_LIMIT); tcp_log_id_bucket_zone = uma_zcreate("tcp_log_id_bucket", sizeof(struct tcp_log_id_bucket), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); tcp_log_id_node_zone = uma_zcreate("tcp_log_id_node", sizeof(struct tcp_log_id_node), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); #ifdef TCPLOG_DEBUG_COUNTERS tcp_log_queued = counter_u64_alloc(M_WAITOK); tcp_log_que_fail1 = counter_u64_alloc(M_WAITOK); tcp_log_que_fail2 = counter_u64_alloc(M_WAITOK); tcp_log_que_fail3 = counter_u64_alloc(M_WAITOK); tcp_log_que_fail4 = counter_u64_alloc(M_WAITOK); tcp_log_que_fail5 = counter_u64_alloc(M_WAITOK); tcp_log_que_copyout = counter_u64_alloc(M_WAITOK); tcp_log_que_read = counter_u64_alloc(M_WAITOK); tcp_log_que_freed = counter_u64_alloc(M_WAITOK); #endif tcp_log_pcb_ids_cur = counter_u64_alloc(M_WAITOK); tcp_log_pcb_ids_tot = counter_u64_alloc(M_WAITOK); rw_init_flags(&tcp_id_tree_lock, "TCP ID tree", RW_NEW); mtx_init(&tcp_log_expireq_mtx, "TCP log expireq", NULL, MTX_DEF); callout_init(&tcp_log_expireq_callout, 1); } /* Do per-TCPCB initialization. */ void tcp_log_tcpcbinit(struct tcpcb *tp) { /* A new TCPCB should start out zero-initialized. */ STAILQ_INIT(&tp->t_logs); /* * If we are doing auto-capturing, figure out whether we will capture * this session. */ tp->t_loglimit = tcp_log_session_limit; if ((tcp_log_auto_all == true) && tcp_log_auto_mode && tcp_log_selectauto()) { tp->t_logstate = tcp_log_auto_mode; tp->t_flags2 |= TF2_LOG_AUTO; } } /* Remove entries */ static void tcp_log_expire(void *unused __unused) { struct tcp_log_id_bucket *tlb; struct tcp_log_id_node *tln; sbintime_t expiry_limit; int tree_locked; TCPLOG_EXPIREQ_LOCK(); if (callout_pending(&tcp_log_expireq_callout)) { /* Callout was reset. */ TCPLOG_EXPIREQ_UNLOCK(); return; } /* * Process entries until we reach one that expires too far in the * future. Look one second in the future. */ expiry_limit = getsbinuptime() + SBT_1S; tree_locked = TREE_UNLOCKED; while ((tln = STAILQ_FIRST(&tcp_log_expireq_head)) != NULL && tln->tln_expiretime <= expiry_limit) { if (!callout_active(&tcp_log_expireq_callout)) { /* * Callout was stopped. I guess we should * just quit at this point. */ TCPLOG_EXPIREQ_UNLOCK(); return; } /* * Remove the node from the head of the list and unlock * the list. Change the expiry time to SBT_MAX as a signal * to other threads that we now own this. */ STAILQ_REMOVE_HEAD(&tcp_log_expireq_head, tln_expireq); tln->tln_expiretime = SBT_MAX; TCPLOG_EXPIREQ_UNLOCK(); /* * Remove the node from the bucket. */ tlb = tln->tln_bucket; TCPID_BUCKET_LOCK(tlb); if (tcp_log_remove_id_node(NULL, NULL, tlb, tln, &tree_locked)) { tcp_log_id_validate_tree_lock(tree_locked); if (tree_locked == TREE_WLOCKED) TCPID_TREE_WUNLOCK(); else TCPID_TREE_RUNLOCK(); tree_locked = TREE_UNLOCKED; } /* Drop the INP reference. */ INP_WLOCK(tln->tln_inp); if (!in_pcbrele_wlocked(tln->tln_inp)) INP_WUNLOCK(tln->tln_inp); /* Free the log records. */ tcp_log_free_entries(&tln->tln_entries, &tln->tln_count); /* Free the node. */ uma_zfree(tcp_log_id_node_zone, tln); /* Relock the expiry queue. */ TCPLOG_EXPIREQ_LOCK(); } /* * We've expired all the entries we can. Do we need to reschedule * ourselves? */ callout_deactivate(&tcp_log_expireq_callout); if (tln != NULL) { /* * Get max(now + TCP_LOG_EXPIRE_INTVL, tln->tln_expiretime) and * set the next callout to that. (This helps ensure we generally * run the callout no more often than desired.) */ expiry_limit = getsbinuptime() + TCP_LOG_EXPIRE_INTVL; if (expiry_limit < tln->tln_expiretime) expiry_limit = tln->tln_expiretime; callout_reset_sbt(&tcp_log_expireq_callout, expiry_limit, SBT_1S, tcp_log_expire, NULL, C_ABSOLUTE); } /* We're done. */ TCPLOG_EXPIREQ_UNLOCK(); return; } /* * Move log data from the TCPCB to a new node. This will reset the TCPCB log * entries and log count; however, it will not touch other things from the * TCPCB (e.g. t_lin, t_lib). * * NOTE: Must hold a lock on the INP. */ static void tcp_log_move_tp_to_node(struct tcpcb *tp, struct tcp_log_id_node *tln) { INP_WLOCK_ASSERT(tp->t_inpcb); tln->tln_ie = tp->t_inpcb->inp_inc.inc_ie; if (tp->t_inpcb->inp_inc.inc_flags & INC_ISIPV6) tln->tln_af = AF_INET6; else tln->tln_af = AF_INET; tln->tln_entries = tp->t_logs; tln->tln_count = tp->t_lognum; tln->tln_bucket = tp->t_lib; /* Clear information from the PCB. */ STAILQ_INIT(&tp->t_logs); tp->t_lognum = 0; } /* Do per-TCPCB cleanup */ void tcp_log_tcpcbfini(struct tcpcb *tp) { struct tcp_log_id_node *tln, *tln_first; struct tcp_log_mem *log_entry; sbintime_t callouttime; INP_WLOCK_ASSERT(tp->t_inpcb); TCP_LOG_EVENT(tp, NULL, NULL, NULL, TCP_LOG_CONNEND, 0, 0, NULL, false); /* * If we were gathering packets to be automatically dumped, try to do * it now. If this succeeds, the log information in the TCPCB will be * cleared. Otherwise, we'll handle the log information as we do * for other states. */ switch(tp->t_logstate) { case TCP_LOG_STATE_HEAD_AUTO: (void)tcp_log_dump_tp_logbuf(tp, "auto-dumped from head", M_NOWAIT, false); break; case TCP_LOG_STATE_TAIL_AUTO: (void)tcp_log_dump_tp_logbuf(tp, "auto-dumped from tail", M_NOWAIT, false); break; case TCP_LOG_STATE_CONTINUAL: (void)tcp_log_dump_tp_logbuf(tp, "auto-dumped from continual", M_NOWAIT, false); break; } /* * There are two ways we could keep logs: per-socket or per-ID. If * we are tracking logs with an ID, then the logs survive the * destruction of the TCPCB. * * If the TCPCB is associated with an ID node, move the logs from the * TCPCB to the ID node. In theory, this is safe, for reasons which I * will now explain for my own benefit when I next need to figure out * this code. :-) * * We own the INP lock. Therefore, no one else can change the contents * of this node (Rule C). Further, no one can remove this node from * the bucket while we hold the lock (Rule D). Basically, no one can * mess with this node. That leaves two states in which we could be: * * 1. Another thread is currently waiting to acquire the INP lock, with * plans to do something with this node. When we drop the INP lock, * they will have a chance to do that. They will recheck the * tln_closed field (see note to Rule C) and then acquire the * bucket lock before proceeding further. * * 2. Another thread will try to acquire a lock at some point in the * future. If they try to acquire a lock before we set the * tln_closed field, they will follow state #1. If they try to * acquire a lock after we set the tln_closed field, they will be * able to make changes to the node, at will, following Rule C. * * Therefore, we currently own this node and can make any changes * we want. But, as soon as we set the tln_closed field to true, we * have effectively dropped our lock on the node. (For this reason, we * also need to make sure our writes are ordered correctly. An atomic * operation with "release" semantics should be sufficient.) */ if (tp->t_lin != NULL) { /* Copy the relevant information to the log entry. */ tln = tp->t_lin; KASSERT(tln->tln_inp == tp->t_inpcb, ("%s: Mismatched inp (tln->tln_inp=%p, tp->t_inpcb=%p)", __func__, tln->tln_inp, tp->t_inpcb)); tcp_log_move_tp_to_node(tp, tln); /* Clear information from the PCB. */ tp->t_lin = NULL; tp->t_lib = NULL; /* * Take a reference on the INP. This ensures that the INP * remains valid while the node is on the expiry queue. This * ensures the INP is valid for other threads that may be * racing to lock this node when we move it to the expire * queue. */ in_pcbref(tp->t_inpcb); /* * Store the entry on the expiry list. The exact behavior * depends on whether we have entries to keep. If so, we * put the entry at the tail of the list and expire in * TCP_LOG_EXPIRE_TIME. Otherwise, we expire "now" and put * the entry at the head of the list. (Handling the cleanup * via the expiry timer lets us avoid locking messy-ness here.) */ tln->tln_expiretime = getsbinuptime(); TCPLOG_EXPIREQ_LOCK(); if (tln->tln_count) { tln->tln_expiretime += TCP_LOG_EXPIRE_TIME; if (STAILQ_EMPTY(&tcp_log_expireq_head) && !callout_active(&tcp_log_expireq_callout)) { /* * We are adding the first entry and a callout * is not currently scheduled; therefore, we * need to schedule one. */ callout_reset_sbt(&tcp_log_expireq_callout, tln->tln_expiretime, SBT_1S, tcp_log_expire, NULL, C_ABSOLUTE); } STAILQ_INSERT_TAIL(&tcp_log_expireq_head, tln, tln_expireq); } else { callouttime = tln->tln_expiretime + TCP_LOG_EXPIRE_INTVL; tln_first = STAILQ_FIRST(&tcp_log_expireq_head); if ((tln_first == NULL || callouttime < tln_first->tln_expiretime) && (callout_pending(&tcp_log_expireq_callout) || !callout_active(&tcp_log_expireq_callout))) { /* * The list is empty, or we want to run the * expire code before the first entry's timer * fires. Also, we are in a case where a callout * is not actively running. We want to reset * the callout to occur sooner. */ callout_reset_sbt(&tcp_log_expireq_callout, callouttime, SBT_1S, tcp_log_expire, NULL, C_ABSOLUTE); } /* * Insert to the head, or just after the head, as * appropriate. (This might result in small * mis-orderings as a bunch of "expire now" entries * gather at the start of the list, but that should * not produce big problems, since the expire timer * will walk through all of them.) */ if (tln_first == NULL || tln->tln_expiretime < tln_first->tln_expiretime) STAILQ_INSERT_HEAD(&tcp_log_expireq_head, tln, tln_expireq); else STAILQ_INSERT_AFTER(&tcp_log_expireq_head, tln_first, tln, tln_expireq); } TCPLOG_EXPIREQ_UNLOCK(); /* * We are done messing with the tln. After this point, we * can't touch it. (Note that the "release" semantics should * be included with the TCPLOG_EXPIREQ_UNLOCK() call above. * Therefore, they should be unnecessary here. However, it * seems like a good idea to include them anyway, since we * really are releasing a lock here.) */ atomic_store_rel_int(&tln->tln_closed, 1); } else { /* Remove log entries. */ while ((log_entry = STAILQ_FIRST(&tp->t_logs)) != NULL) tcp_log_remove_log_head(tp, log_entry); KASSERT(tp->t_lognum == 0, ("%s: After freeing entries, tp->t_lognum=%d (expected 0)", __func__, tp->t_lognum)); } /* * Change the log state to off (just in case anything tries to sneak * in a last-minute log). */ tp->t_logstate = TCP_LOG_STATE_OFF; } static void tcp_log_purge_tp_logbuf(struct tcpcb *tp) { struct tcp_log_mem *log_entry; struct inpcb *inp __diagused; inp = tp->t_inpcb; INP_WLOCK_ASSERT(inp); if (tp->t_lognum == 0) return; while ((log_entry = STAILQ_FIRST(&tp->t_logs)) != NULL) tcp_log_remove_log_head(tp, log_entry); KASSERT(tp->t_lognum == 0, ("%s: After freeing entries, tp->t_lognum=%d (expected 0)", __func__, tp->t_lognum)); tp->t_logstate = TCP_LOG_STATE_OFF; } /* * This logs an event for a TCP socket. Normally, this is called via * TCP_LOG_EVENT or TCP_LOG_EVENT_VERBOSE. See the documentation for * TCP_LOG_EVENT(). */ struct tcp_log_buffer * tcp_log_event_(struct tcpcb *tp, struct tcphdr *th, struct sockbuf *rxbuf, struct sockbuf *txbuf, uint8_t eventid, int errornum, uint32_t len, union tcp_log_stackspecific *stackinfo, int th_hostorder, const char *output_caller, const char *func, int line, const struct timeval *itv) { struct tcp_log_mem *log_entry; struct tcp_log_buffer *log_buf; int attempt_count = 0; struct tcp_log_verbose *log_verbose; uint32_t logsn; KASSERT((func == NULL && line == 0) || (func != NULL && line > 0), ("%s called with inconsistent func (%p) and line (%d) arguments", __func__, func, line)); INP_WLOCK_ASSERT(tp->t_inpcb); if (tcp_disable_all_bb_logs) { /* * The global shutdown logging * switch has been thrown. Call * the purge function that frees * purges out the logs and * turns off logging. */ tcp_log_purge_tp_logbuf(tp); return (NULL); } KASSERT(tp->t_logstate == TCP_LOG_STATE_HEAD || tp->t_logstate == TCP_LOG_STATE_TAIL || tp->t_logstate == TCP_LOG_STATE_CONTINUAL || tp->t_logstate == TCP_LOG_STATE_HEAD_AUTO || tp->t_logstate == TCP_LOG_STATE_TAIL_AUTO, ("%s called with unexpected tp->t_logstate (%d)", __func__, tp->t_logstate)); /* * Get the serial number. We do this early so it will * increment even if we end up skipping the log entry for some * reason. */ logsn = tp->t_logsn++; /* * Can we get a new log entry? If so, increment the lognum counter * here. */ retry: if (tp->t_lognum < tp->t_loglimit) { if ((log_entry = uma_zalloc(tcp_log_zone, M_NOWAIT)) != NULL) tp->t_lognum++; } else log_entry = NULL; /* Do we need to try to reuse? */ if (log_entry == NULL) { /* * Sacrifice auto-logged sessions without a log ID if * tcp_log_auto_all is false. (If they don't have a log * ID by now, it is probable that either they won't get one * or we are resource-constrained.) */ if (tp->t_lib == NULL && (tp->t_flags2 & TF2_LOG_AUTO) && !tcp_log_auto_all) { if (tcp_log_state_change(tp, TCP_LOG_STATE_CLEAR)) { #ifdef INVARIANTS panic("%s:%d: tcp_log_state_change() failed " "to set tp %p to TCP_LOG_STATE_CLEAR", __func__, __LINE__, tp); #endif tp->t_logstate = TCP_LOG_STATE_OFF; } return (NULL); } /* * If we are in TCP_LOG_STATE_HEAD_AUTO state, try to dump * the buffers. If successful, deactivate tracing. Otherwise, * leave it active so we will retry. */ if (tp->t_logstate == TCP_LOG_STATE_HEAD_AUTO && !tcp_log_dump_tp_logbuf(tp, "auto-dumped from head", M_NOWAIT, false)) { tp->t_logstate = TCP_LOG_STATE_OFF; return(NULL); } else if ((tp->t_logstate == TCP_LOG_STATE_CONTINUAL) && !tcp_log_dump_tp_logbuf(tp, "auto-dumped from continual", M_NOWAIT, false)) { if (attempt_count == 0) { attempt_count++; goto retry; } #ifdef TCPLOG_DEBUG_COUNTERS counter_u64_add(tcp_log_que_fail4, 1); #endif return(NULL); } else if (tp->t_logstate == TCP_LOG_STATE_HEAD_AUTO) return(NULL); /* If in HEAD state, just deactivate the tracing and return. */ if (tp->t_logstate == TCP_LOG_STATE_HEAD) { tp->t_logstate = TCP_LOG_STATE_OFF; return(NULL); } /* * Get a buffer to reuse. If that fails, just give up. * (We can't log anything without a buffer in which to * put it.) * * Note that we don't change the t_lognum counter * here. Because we are re-using the buffer, the total * number won't change. */ if ((log_entry = STAILQ_FIRST(&tp->t_logs)) == NULL) return(NULL); STAILQ_REMOVE_HEAD(&tp->t_logs, tlm_queue); tcp_log_entry_refcnt_rem(log_entry); } KASSERT(log_entry != NULL, ("%s: log_entry unexpectedly NULL", __func__)); /* Extract the log buffer and verbose buffer pointers. */ log_buf = &log_entry->tlm_buf; log_verbose = &log_entry->tlm_v; /* Basic entries. */ if (itv == NULL) getmicrouptime(&log_buf->tlb_tv); else memcpy(&log_buf->tlb_tv, itv, sizeof(struct timeval)); log_buf->tlb_ticks = ticks; log_buf->tlb_sn = logsn; log_buf->tlb_stackid = tp->t_fb->tfb_id; log_buf->tlb_eventid = eventid; log_buf->tlb_eventflags = 0; log_buf->tlb_errno = errornum; /* Socket buffers */ if (rxbuf != NULL) { log_buf->tlb_eventflags |= TLB_FLAG_RXBUF; log_buf->tlb_rxbuf.tls_sb_acc = rxbuf->sb_acc; log_buf->tlb_rxbuf.tls_sb_ccc = rxbuf->sb_ccc; log_buf->tlb_rxbuf.tls_sb_spare = 0; } if (txbuf != NULL) { log_buf->tlb_eventflags |= TLB_FLAG_TXBUF; log_buf->tlb_txbuf.tls_sb_acc = txbuf->sb_acc; log_buf->tlb_txbuf.tls_sb_ccc = txbuf->sb_ccc; log_buf->tlb_txbuf.tls_sb_spare = 0; } /* Copy values from tp to the log entry. */ #define COPY_STAT(f) log_buf->tlb_ ## f = tp->f #define COPY_STAT_T(f) log_buf->tlb_ ## f = tp->t_ ## f COPY_STAT_T(state); COPY_STAT_T(starttime); COPY_STAT(iss); COPY_STAT_T(flags); COPY_STAT(snd_una); COPY_STAT(snd_max); COPY_STAT(snd_cwnd); COPY_STAT(snd_nxt); COPY_STAT(snd_recover); COPY_STAT(snd_wnd); COPY_STAT(snd_ssthresh); COPY_STAT_T(srtt); COPY_STAT_T(rttvar); COPY_STAT(rcv_up); COPY_STAT(rcv_adv); COPY_STAT(rcv_nxt); COPY_STAT(rcv_wnd); COPY_STAT_T(dupacks); COPY_STAT_T(segqlen); COPY_STAT(snd_numholes); COPY_STAT(snd_scale); COPY_STAT(rcv_scale); COPY_STAT_T(flags2); COPY_STAT_T(fbyte_in); COPY_STAT_T(fbyte_out); #undef COPY_STAT #undef COPY_STAT_T log_buf->tlb_flex1 = 0; log_buf->tlb_flex2 = 0; /* Copy stack-specific info. */ if (stackinfo != NULL) { memcpy(&log_buf->tlb_stackinfo, stackinfo, sizeof(log_buf->tlb_stackinfo)); log_buf->tlb_eventflags |= TLB_FLAG_STACKINFO; } /* The packet */ log_buf->tlb_len = len; if (th) { int optlen; log_buf->tlb_eventflags |= TLB_FLAG_HDR; log_buf->tlb_th = *th; if (th_hostorder) tcp_fields_to_net(&log_buf->tlb_th); optlen = (th->th_off << 2) - sizeof (struct tcphdr); if (optlen > 0) memcpy(log_buf->tlb_opts, th + 1, optlen); } /* Verbose information */ if (func != NULL) { log_buf->tlb_eventflags |= TLB_FLAG_VERBOSE; if (output_caller != NULL) strlcpy(log_verbose->tlv_snd_frm, output_caller, TCP_FUNC_LEN); else *log_verbose->tlv_snd_frm = 0; strlcpy(log_verbose->tlv_trace_func, func, TCP_FUNC_LEN); log_verbose->tlv_trace_line = line; } /* Insert the new log at the tail. */ STAILQ_INSERT_TAIL(&tp->t_logs, log_entry, tlm_queue); tcp_log_entry_refcnt_add(log_entry); return (log_buf); } /* * Change the logging state for a TCPCB. Returns 0 on success or an * error code on failure. */ int tcp_log_state_change(struct tcpcb *tp, int state) { struct tcp_log_mem *log_entry; INP_WLOCK_ASSERT(tp->t_inpcb); switch(state) { case TCP_LOG_STATE_CLEAR: while ((log_entry = STAILQ_FIRST(&tp->t_logs)) != NULL) tcp_log_remove_log_head(tp, log_entry); /* Fall through */ case TCP_LOG_STATE_OFF: tp->t_logstate = TCP_LOG_STATE_OFF; break; case TCP_LOG_STATE_TAIL: case TCP_LOG_STATE_HEAD: case TCP_LOG_STATE_CONTINUAL: case TCP_LOG_STATE_HEAD_AUTO: case TCP_LOG_STATE_TAIL_AUTO: tp->t_logstate = state; break; default: return (EINVAL); } if (tcp_disable_all_bb_logs) { /* We are prohibited from doing any logs */ tp->t_logstate = TCP_LOG_STATE_OFF; } tp->t_flags2 &= ~(TF2_LOG_AUTO); return (0); } /* If tcp_drain() is called, flush half the log entries. */ void tcp_log_drain(struct tcpcb *tp) { struct tcp_log_mem *log_entry, *next; int target, skip; INP_WLOCK_ASSERT(tp->t_inpcb); if ((target = tp->t_lognum / 2) == 0) return; /* * If we are logging the "head" packets, we want to discard * from the tail of the queue. Otherwise, we want to discard * from the head. */ if (tp->t_logstate == TCP_LOG_STATE_HEAD || tp->t_logstate == TCP_LOG_STATE_HEAD_AUTO) { skip = tp->t_lognum - target; STAILQ_FOREACH(log_entry, &tp->t_logs, tlm_queue) if (!--skip) break; KASSERT(log_entry != NULL, ("%s: skipped through all entries!", __func__)); if (log_entry == NULL) return; while ((next = STAILQ_NEXT(log_entry, tlm_queue)) != NULL) { STAILQ_REMOVE_AFTER(&tp->t_logs, log_entry, tlm_queue); tcp_log_entry_refcnt_rem(next); tcp_log_remove_log_cleanup(tp, next); #ifdef INVARIANTS target--; #endif } KASSERT(target == 0, ("%s: After removing from tail, target was %d", __func__, target)); } else if (tp->t_logstate == TCP_LOG_STATE_CONTINUAL) { (void)tcp_log_dump_tp_logbuf(tp, "auto-dumped from continual", M_NOWAIT, false); } else { while ((log_entry = STAILQ_FIRST(&tp->t_logs)) != NULL && target--) tcp_log_remove_log_head(tp, log_entry); KASSERT(target <= 0, ("%s: After removing from head, target was %d", __func__, target)); KASSERT(tp->t_lognum > 0, ("%s: After removing from head, tp->t_lognum was %d", __func__, target)); KASSERT(log_entry != NULL, ("%s: After removing from head, the tailq was empty", __func__)); } } static inline int tcp_log_copyout(struct sockopt *sopt, void *src, void *dst, size_t len) { if (sopt->sopt_td != NULL) return (copyout(src, dst, len)); bcopy(src, dst, len); return (0); } static int tcp_log_logs_to_buf(struct sockopt *sopt, struct tcp_log_stailq *log_tailqp, struct tcp_log_buffer **end, int count) { struct tcp_log_buffer *out_entry; struct tcp_log_mem *log_entry; size_t entrysize; int error; #ifdef INVARIANTS int orig_count = count; #endif /* Copy the data out. */ error = 0; out_entry = (struct tcp_log_buffer *) sopt->sopt_val; STAILQ_FOREACH(log_entry, log_tailqp, tlm_queue) { count--; KASSERT(count >= 0, ("%s:%d: Exceeded expected count (%d) processing list %p", __func__, __LINE__, orig_count, log_tailqp)); #ifdef TCPLOG_DEBUG_COUNTERS counter_u64_add(tcp_log_que_copyout, 1); #endif /* * Skip copying out the header if it isn't present. * Instead, copy out zeros (to ensure we don't leak info). * TODO: Make sure we truly do zero everything we don't * explicitly set. */ if (log_entry->tlm_buf.tlb_eventflags & TLB_FLAG_HDR) entrysize = sizeof(struct tcp_log_buffer); else entrysize = offsetof(struct tcp_log_buffer, tlb_th); error = tcp_log_copyout(sopt, &log_entry->tlm_buf, out_entry, entrysize); if (error) break; if (!(log_entry->tlm_buf.tlb_eventflags & TLB_FLAG_HDR)) { error = tcp_log_copyout(sopt, zerobuf, ((uint8_t *)out_entry) + entrysize, sizeof(struct tcp_log_buffer) - entrysize); } /* * Copy out the verbose bit, if needed. Either way, * increment the output pointer the correct amount. */ if (log_entry->tlm_buf.tlb_eventflags & TLB_FLAG_VERBOSE) { error = tcp_log_copyout(sopt, &log_entry->tlm_v, out_entry->tlb_verbose, sizeof(struct tcp_log_verbose)); if (error) break; out_entry = (struct tcp_log_buffer *) (((uint8_t *) (out_entry + 1)) + sizeof(struct tcp_log_verbose)); } else out_entry++; } *end = out_entry; KASSERT(error || count == 0, ("%s:%d: Less than expected count (%d) processing list %p" " (%d remain)", __func__, __LINE__, orig_count, log_tailqp, count)); return (error); } /* * Copy out the buffer. Note that we do incremental copying, so * sooptcopyout() won't work. However, the goal is to produce the same * end result as if we copied in the entire user buffer, updated it, * and then used sooptcopyout() to copy it out. * * NOTE: This should be called with a write lock on the PCB; however, * the function will drop it after it extracts the data from the TCPCB. */ int tcp_log_getlogbuf(struct sockopt *sopt, struct tcpcb *tp) { struct tcp_log_stailq log_tailq; struct tcp_log_mem *log_entry, *log_next; struct tcp_log_buffer *out_entry; struct inpcb *inp; size_t outsize, entrysize; int error, outnum; INP_WLOCK_ASSERT(tp->t_inpcb); inp = tp->t_inpcb; /* * Determine which log entries will fit in the buffer. As an * optimization, skip this if all the entries will clearly fit * in the buffer. (However, get an exact size if we are using * INVARIANTS.) */ #ifndef INVARIANTS if (sopt->sopt_valsize / (sizeof(struct tcp_log_buffer) + sizeof(struct tcp_log_verbose)) >= tp->t_lognum) { log_entry = STAILQ_LAST(&tp->t_logs, tcp_log_mem, tlm_queue); log_next = NULL; outsize = 0; outnum = tp->t_lognum; } else { #endif outsize = outnum = 0; log_entry = NULL; STAILQ_FOREACH(log_next, &tp->t_logs, tlm_queue) { entrysize = sizeof(struct tcp_log_buffer); if (log_next->tlm_buf.tlb_eventflags & TLB_FLAG_VERBOSE) entrysize += sizeof(struct tcp_log_verbose); if ((sopt->sopt_valsize - outsize) < entrysize) break; outsize += entrysize; outnum++; log_entry = log_next; } KASSERT(outsize <= sopt->sopt_valsize, ("%s: calculated output size (%zu) greater than available" "space (%zu)", __func__, outsize, sopt->sopt_valsize)); #ifndef INVARIANTS } #endif /* * Copy traditional sooptcopyout() behavior: if sopt->sopt_val * is NULL, silently skip the copy. However, in this case, we * will leave the list alone and return. Functionally, this * gives userspace a way to poll for an approximate buffer * size they will need to get the log entries. */ if (sopt->sopt_val == NULL) { INP_WUNLOCK(inp); if (outsize == 0) { outsize = outnum * (sizeof(struct tcp_log_buffer) + sizeof(struct tcp_log_verbose)); } if (sopt->sopt_valsize > outsize) sopt->sopt_valsize = outsize; return (0); } /* * Break apart the list. We'll save the ones we want to copy * out locally and remove them from the TCPCB list. We can * then drop the INPCB lock while we do the copyout. * * There are roughly three cases: * 1. There was nothing to copy out. That's easy: drop the * lock and return. * 2. We are copying out the entire list. Again, that's easy: * move the whole list. * 3. We are copying out a partial list. That's harder. We * need to update the list book-keeping entries. */ if (log_entry != NULL && log_next == NULL) { /* Move entire list. */ KASSERT(outnum == tp->t_lognum, ("%s:%d: outnum (%d) should match tp->t_lognum (%d)", __func__, __LINE__, outnum, tp->t_lognum)); log_tailq = tp->t_logs; tp->t_lognum = 0; STAILQ_INIT(&tp->t_logs); } else if (log_entry != NULL) { /* Move partial list. */ KASSERT(outnum < tp->t_lognum, ("%s:%d: outnum (%d) not less than tp->t_lognum (%d)", __func__, __LINE__, outnum, tp->t_lognum)); STAILQ_FIRST(&log_tailq) = STAILQ_FIRST(&tp->t_logs); STAILQ_FIRST(&tp->t_logs) = STAILQ_NEXT(log_entry, tlm_queue); KASSERT(STAILQ_NEXT(log_entry, tlm_queue) != NULL, ("%s:%d: tp->t_logs is unexpectedly shorter than expected" "(tp: %p, log_tailq: %p, outnum: %d, tp->t_lognum: %d)", __func__, __LINE__, tp, &log_tailq, outnum, tp->t_lognum)); STAILQ_NEXT(log_entry, tlm_queue) = NULL; log_tailq.stqh_last = &STAILQ_NEXT(log_entry, tlm_queue); tp->t_lognum -= outnum; } else STAILQ_INIT(&log_tailq); /* Drop the PCB lock. */ INP_WUNLOCK(inp); /* Copy the data out. */ error = tcp_log_logs_to_buf(sopt, &log_tailq, &out_entry, outnum); if (error) { /* Restore list */ INP_WLOCK(inp); - if ((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) == 0) { + if ((inp->inp_flags & INP_DROPPED) == 0) { tp = intotcpcb(inp); /* Merge the two lists. */ STAILQ_CONCAT(&log_tailq, &tp->t_logs); tp->t_logs = log_tailq; tp->t_lognum += outnum; } INP_WUNLOCK(inp); } else { /* Sanity check entries */ KASSERT(((caddr_t)out_entry - (caddr_t)sopt->sopt_val) == outsize, ("%s: Actual output size (%zu) != " "calculated output size (%zu)", __func__, (size_t)((caddr_t)out_entry - (caddr_t)sopt->sopt_val), outsize)); /* Free the entries we just copied out. */ STAILQ_FOREACH_SAFE(log_entry, &log_tailq, tlm_queue, log_next) { tcp_log_entry_refcnt_rem(log_entry); uma_zfree(tcp_log_zone, log_entry); } } sopt->sopt_valsize = (size_t)((caddr_t)out_entry - (caddr_t)sopt->sopt_val); return (error); } static void tcp_log_free_queue(struct tcp_log_dev_queue *param) { struct tcp_log_dev_log_queue *entry; KASSERT(param != NULL, ("%s: called with NULL param", __func__)); if (param == NULL) return; entry = (struct tcp_log_dev_log_queue *)param; /* Free the entries. */ tcp_log_free_entries(&entry->tldl_entries, &entry->tldl_count); /* Free the buffer, if it is allocated. */ if (entry->tldl_common.tldq_buf != NULL) free(entry->tldl_common.tldq_buf, M_TCPLOGDEV); /* Free the queue entry. */ free(entry, M_TCPLOGDEV); } static struct tcp_log_common_header * tcp_log_expandlogbuf(struct tcp_log_dev_queue *param) { struct tcp_log_dev_log_queue *entry; struct tcp_log_header *hdr; uint8_t *end; struct sockopt sopt; int error; entry = (struct tcp_log_dev_log_queue *)param; /* Take a worst-case guess at space needs. */ sopt.sopt_valsize = sizeof(struct tcp_log_header) + entry->tldl_count * (sizeof(struct tcp_log_buffer) + sizeof(struct tcp_log_verbose)); hdr = malloc(sopt.sopt_valsize, M_TCPLOGDEV, M_NOWAIT); if (hdr == NULL) { #ifdef TCPLOG_DEBUG_COUNTERS counter_u64_add(tcp_log_que_fail5, entry->tldl_count); #endif return (NULL); } sopt.sopt_val = hdr + 1; sopt.sopt_valsize -= sizeof(struct tcp_log_header); sopt.sopt_td = NULL; error = tcp_log_logs_to_buf(&sopt, &entry->tldl_entries, (struct tcp_log_buffer **)&end, entry->tldl_count); if (error) { free(hdr, M_TCPLOGDEV); return (NULL); } /* Free the entries. */ tcp_log_free_entries(&entry->tldl_entries, &entry->tldl_count); entry->tldl_count = 0; memset(hdr, 0, sizeof(struct tcp_log_header)); hdr->tlh_version = TCP_LOG_BUF_VER; hdr->tlh_type = TCP_LOG_DEV_TYPE_BBR; hdr->tlh_length = end - (uint8_t *)hdr; hdr->tlh_ie = entry->tldl_ie; hdr->tlh_af = entry->tldl_af; getboottime(&hdr->tlh_offset); strlcpy(hdr->tlh_id, entry->tldl_id, TCP_LOG_ID_LEN); strlcpy(hdr->tlh_tag, entry->tldl_tag, TCP_LOG_TAG_LEN); strlcpy(hdr->tlh_reason, entry->tldl_reason, TCP_LOG_REASON_LEN); return ((struct tcp_log_common_header *)hdr); } /* * Queue the tcpcb's log buffer for transmission via the log buffer facility. * * NOTE: This should be called with a write lock on the PCB. * * how should be M_WAITOK or M_NOWAIT. If M_WAITOK, the function will drop * and reacquire the INP lock if it needs to do so. * * If force is false, this will only dump auto-logged sessions if * tcp_log_auto_all is true or if there is a log ID defined for the session. */ int tcp_log_dump_tp_logbuf(struct tcpcb *tp, char *reason, int how, bool force) { struct tcp_log_dev_log_queue *entry; struct inpcb *inp; #ifdef TCPLOG_DEBUG_COUNTERS int num_entries; #endif inp = tp->t_inpcb; INP_WLOCK_ASSERT(inp); /* If there are no log entries, there is nothing to do. */ if (tp->t_lognum == 0) return (0); /* Check for a log ID. */ if (tp->t_lib == NULL && (tp->t_flags2 & TF2_LOG_AUTO) && !tcp_log_auto_all && !force) { struct tcp_log_mem *log_entry; /* * We needed a log ID and none was found. Free the log entries * and return success. Also, cancel further logging. If the * session doesn't have a log ID by now, we'll assume it isn't * going to get one. */ while ((log_entry = STAILQ_FIRST(&tp->t_logs)) != NULL) tcp_log_remove_log_head(tp, log_entry); KASSERT(tp->t_lognum == 0, ("%s: After freeing entries, tp->t_lognum=%d (expected 0)", __func__, tp->t_lognum)); tp->t_logstate = TCP_LOG_STATE_OFF; return (0); } /* * Allocate memory. If we must wait, we'll need to drop the locks * and reacquire them (and do all the related business that goes * along with that). */ entry = malloc(sizeof(struct tcp_log_dev_log_queue), M_TCPLOGDEV, M_NOWAIT); if (entry == NULL && (how & M_NOWAIT)) { #ifdef TCPLOG_DEBUG_COUNTERS counter_u64_add(tcp_log_que_fail3, 1); #endif return (ENOBUFS); } if (entry == NULL) { INP_WUNLOCK(inp); entry = malloc(sizeof(struct tcp_log_dev_log_queue), M_TCPLOGDEV, M_WAITOK); INP_WLOCK(inp); /* * Note that this check is slightly overly-restrictive in * that the TCB can survive either of these events. * However, there is currently not a good way to ensure * that is the case. So, if we hit this M_WAIT path, we * may end up dropping some entries. That seems like a * small price to pay for safety. */ - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { free(entry, M_TCPLOGDEV); #ifdef TCPLOG_DEBUG_COUNTERS counter_u64_add(tcp_log_que_fail2, 1); #endif return (ECONNRESET); } tp = intotcpcb(inp); if (tp->t_lognum == 0) { free(entry, M_TCPLOGDEV); return (0); } } /* Fill in the unique parts of the queue entry. */ if (tp->t_lib != NULL) { strlcpy(entry->tldl_id, tp->t_lib->tlb_id, TCP_LOG_ID_LEN); strlcpy(entry->tldl_tag, tp->t_lib->tlb_tag, TCP_LOG_TAG_LEN); } else { strlcpy(entry->tldl_id, "UNKNOWN", TCP_LOG_ID_LEN); strlcpy(entry->tldl_tag, "UNKNOWN", TCP_LOG_TAG_LEN); } if (reason != NULL) strlcpy(entry->tldl_reason, reason, TCP_LOG_REASON_LEN); else strlcpy(entry->tldl_reason, "UNKNOWN", TCP_LOG_ID_LEN); entry->tldl_ie = inp->inp_inc.inc_ie; if (inp->inp_inc.inc_flags & INC_ISIPV6) entry->tldl_af = AF_INET6; else entry->tldl_af = AF_INET; entry->tldl_entries = tp->t_logs; entry->tldl_count = tp->t_lognum; /* Fill in the common parts of the queue entry. */ entry->tldl_common.tldq_buf = NULL; entry->tldl_common.tldq_xform = tcp_log_expandlogbuf; entry->tldl_common.tldq_dtor = tcp_log_free_queue; /* Clear the log data from the TCPCB. */ #ifdef TCPLOG_DEBUG_COUNTERS num_entries = tp->t_lognum; #endif tp->t_lognum = 0; STAILQ_INIT(&tp->t_logs); /* Add the entry. If no one is listening, free the entry. */ if (tcp_log_dev_add_log((struct tcp_log_dev_queue *)entry)) { tcp_log_free_queue((struct tcp_log_dev_queue *)entry); #ifdef TCPLOG_DEBUG_COUNTERS counter_u64_add(tcp_log_que_fail1, num_entries); } else { counter_u64_add(tcp_log_queued, num_entries); #endif } return (0); } /* * Queue the log_id_node's log buffers for transmission via the log buffer * facility. * * NOTE: This should be called with the bucket locked and referenced. * * how should be M_WAITOK or M_NOWAIT. If M_WAITOK, the function will drop * and reacquire the bucket lock if it needs to do so. (The caller must * ensure that the tln is no longer on any lists so no one else will mess * with this while the lock is dropped!) */ static int tcp_log_dump_node_logbuf(struct tcp_log_id_node *tln, char *reason, int how) { struct tcp_log_dev_log_queue *entry; struct tcp_log_id_bucket *tlb; tlb = tln->tln_bucket; TCPID_BUCKET_LOCK_ASSERT(tlb); KASSERT(tlb->tlb_refcnt > 0, ("%s:%d: Called with unreferenced bucket (tln=%p, tlb=%p)", __func__, __LINE__, tln, tlb)); KASSERT(tln->tln_closed, ("%s:%d: Called for node with tln_closed==false (tln=%p)", __func__, __LINE__, tln)); /* If there are no log entries, there is nothing to do. */ if (tln->tln_count == 0) return (0); /* * Allocate memory. If we must wait, we'll need to drop the locks * and reacquire them (and do all the related business that goes * along with that). */ entry = malloc(sizeof(struct tcp_log_dev_log_queue), M_TCPLOGDEV, M_NOWAIT); if (entry == NULL && (how & M_NOWAIT)) return (ENOBUFS); if (entry == NULL) { TCPID_BUCKET_UNLOCK(tlb); entry = malloc(sizeof(struct tcp_log_dev_log_queue), M_TCPLOGDEV, M_WAITOK); TCPID_BUCKET_LOCK(tlb); } /* Fill in the common parts of the queue entry.. */ entry->tldl_common.tldq_buf = NULL; entry->tldl_common.tldq_xform = tcp_log_expandlogbuf; entry->tldl_common.tldq_dtor = tcp_log_free_queue; /* Fill in the unique parts of the queue entry. */ strlcpy(entry->tldl_id, tlb->tlb_id, TCP_LOG_ID_LEN); strlcpy(entry->tldl_tag, tlb->tlb_tag, TCP_LOG_TAG_LEN); if (reason != NULL) strlcpy(entry->tldl_reason, reason, TCP_LOG_REASON_LEN); else strlcpy(entry->tldl_reason, "UNKNOWN", TCP_LOG_ID_LEN); entry->tldl_ie = tln->tln_ie; entry->tldl_entries = tln->tln_entries; entry->tldl_count = tln->tln_count; entry->tldl_af = tln->tln_af; /* Add the entry. If no one is listening, free the entry. */ if (tcp_log_dev_add_log((struct tcp_log_dev_queue *)entry)) tcp_log_free_queue((struct tcp_log_dev_queue *)entry); return (0); } /* * Queue the log buffers for all sessions in a bucket for transmissions via * the log buffer facility. * * NOTE: This should be called with a locked bucket; however, the function * will drop the lock. */ #define LOCAL_SAVE 10 static void tcp_log_dumpbucketlogs(struct tcp_log_id_bucket *tlb, char *reason) { struct tcp_log_id_node local_entries[LOCAL_SAVE]; struct inpcb *inp; struct tcpcb *tp; struct tcp_log_id_node *cur_tln, *prev_tln, *tmp_tln; int i, num_local_entries, tree_locked; bool expireq_locked; TCPID_BUCKET_LOCK_ASSERT(tlb); /* * Take a reference on the bucket to keep it from disappearing until * we are done. */ TCPID_BUCKET_REF(tlb); /* * We'll try to create these without dropping locks. However, we * might very well need to drop locks to get memory. If that's the * case, we'll save up to 10 on the stack, and sacrifice the rest. * (Otherwise, we need to worry about finding our place again in a * potentially changed list. It just doesn't seem worth the trouble * to do that. */ expireq_locked = false; num_local_entries = 0; prev_tln = NULL; tree_locked = TREE_UNLOCKED; SLIST_FOREACH_SAFE(cur_tln, &tlb->tlb_head, tln_list, tmp_tln) { /* * If this isn't associated with a TCPCB, we can pull it off * the list now. We need to be careful that the expire timer * hasn't already taken ownership (tln_expiretime == SBT_MAX). * If so, we let the expire timer code free the data. */ if (cur_tln->tln_closed) { no_inp: /* * Get the expireq lock so we can get a consistent * read of tln_expiretime and so we can remove this * from the expireq. */ if (!expireq_locked) { TCPLOG_EXPIREQ_LOCK(); expireq_locked = true; } /* * We ignore entries with tln_expiretime == SBT_MAX. * The expire timer code already owns those. */ KASSERT(cur_tln->tln_expiretime > (sbintime_t) 0, ("%s:%d: node on the expire queue without positive " "expire time", __func__, __LINE__)); if (cur_tln->tln_expiretime == SBT_MAX) { prev_tln = cur_tln; continue; } /* Remove the entry from the expireq. */ STAILQ_REMOVE(&tcp_log_expireq_head, cur_tln, tcp_log_id_node, tln_expireq); /* Remove the entry from the bucket. */ if (prev_tln != NULL) SLIST_REMOVE_AFTER(prev_tln, tln_list); else SLIST_REMOVE_HEAD(&tlb->tlb_head, tln_list); /* * Drop the INP and bucket reference counts. Due to * lock-ordering rules, we need to drop the expire * queue lock. */ TCPLOG_EXPIREQ_UNLOCK(); expireq_locked = false; /* Drop the INP reference. */ INP_WLOCK(cur_tln->tln_inp); if (!in_pcbrele_wlocked(cur_tln->tln_inp)) INP_WUNLOCK(cur_tln->tln_inp); if (tcp_log_unref_bucket(tlb, &tree_locked, NULL)) { #ifdef INVARIANTS panic("%s: Bucket refcount unexpectedly 0.", __func__); #endif /* * Recover as best we can: free the entry we * own. */ tcp_log_free_entries(&cur_tln->tln_entries, &cur_tln->tln_count); uma_zfree(tcp_log_id_node_zone, cur_tln); goto done; } if (tcp_log_dump_node_logbuf(cur_tln, reason, M_NOWAIT)) { /* * If we have sapce, save the entries locally. * Otherwise, free them. */ if (num_local_entries < LOCAL_SAVE) { local_entries[num_local_entries] = *cur_tln; num_local_entries++; } else { tcp_log_free_entries( &cur_tln->tln_entries, &cur_tln->tln_count); } } /* No matter what, we are done with the node now. */ uma_zfree(tcp_log_id_node_zone, cur_tln); /* * Because we removed this entry from the list, prev_tln * (which tracks the previous entry still on the tlb * list) remains unchanged. */ continue; } /* * If we get to this point, the session data is still held in * the TCPCB. So, we need to pull the data out of that. * * We will need to drop the expireq lock so we can lock the INP. * We can then try to extract the data the "easy" way. If that * fails, we'll save the log entries for later. */ if (expireq_locked) { TCPLOG_EXPIREQ_UNLOCK(); expireq_locked = false; } /* Lock the INP and then re-check the state. */ inp = cur_tln->tln_inp; INP_WLOCK(inp); /* * If we caught this while it was transitioning, the data * might have moved from the TCPCB to the tln (signified by * setting tln_closed to true. If so, treat this like an * inactive connection. */ if (cur_tln->tln_closed) { /* * It looks like we may have caught this connection * while it was transitioning from active to inactive. * Treat this like an inactive connection. */ INP_WUNLOCK(inp); goto no_inp; } /* * Try to dump the data from the tp without dropping the lock. * If this fails, try to save off the data locally. */ tp = cur_tln->tln_tp; if (tcp_log_dump_tp_logbuf(tp, reason, M_NOWAIT, true) && num_local_entries < LOCAL_SAVE) { tcp_log_move_tp_to_node(tp, &local_entries[num_local_entries]); local_entries[num_local_entries].tln_closed = 1; KASSERT(local_entries[num_local_entries].tln_bucket == tlb, ("%s: %d: bucket mismatch for node %p", __func__, __LINE__, cur_tln)); num_local_entries++; } INP_WUNLOCK(inp); /* * We are goint to leave the current tln on the list. It will * become the previous tln. */ prev_tln = cur_tln; } /* Drop our locks, if any. */ KASSERT(tree_locked == TREE_UNLOCKED, ("%s: %d: tree unexpectedly locked", __func__, __LINE__)); switch (tree_locked) { case TREE_WLOCKED: TCPID_TREE_WUNLOCK(); tree_locked = TREE_UNLOCKED; break; case TREE_RLOCKED: TCPID_TREE_RUNLOCK(); tree_locked = TREE_UNLOCKED; break; } if (expireq_locked) { TCPLOG_EXPIREQ_UNLOCK(); expireq_locked = false; } /* * Try again for any saved entries. tcp_log_dump_node_logbuf() is * guaranteed to free the log entries within the node. And, since * the node itself is on our stack, we don't need to free it. */ for (i = 0; i < num_local_entries; i++) tcp_log_dump_node_logbuf(&local_entries[i], reason, M_WAITOK); /* Drop our reference. */ if (!tcp_log_unref_bucket(tlb, &tree_locked, NULL)) TCPID_BUCKET_UNLOCK(tlb); done: /* Drop our locks, if any. */ switch (tree_locked) { case TREE_WLOCKED: TCPID_TREE_WUNLOCK(); break; case TREE_RLOCKED: TCPID_TREE_RUNLOCK(); break; } if (expireq_locked) TCPLOG_EXPIREQ_UNLOCK(); } #undef LOCAL_SAVE /* * Queue the log buffers for all sessions in a bucket for transmissions via * the log buffer facility. * * NOTE: This should be called with a locked INP; however, the function * will drop the lock. */ void tcp_log_dump_tp_bucket_logbufs(struct tcpcb *tp, char *reason) { struct tcp_log_id_bucket *tlb; int tree_locked; /* Figure out our bucket and lock it. */ INP_WLOCK_ASSERT(tp->t_inpcb); tlb = tp->t_lib; if (tlb == NULL) { /* * No bucket; treat this like a request to dump a single * session's traces. */ (void)tcp_log_dump_tp_logbuf(tp, reason, M_WAITOK, true); INP_WUNLOCK(tp->t_inpcb); return; } TCPID_BUCKET_REF(tlb); INP_WUNLOCK(tp->t_inpcb); TCPID_BUCKET_LOCK(tlb); /* If we are the last reference, we have nothing more to do here. */ tree_locked = TREE_UNLOCKED; if (tcp_log_unref_bucket(tlb, &tree_locked, NULL)) { switch (tree_locked) { case TREE_WLOCKED: TCPID_TREE_WUNLOCK(); break; case TREE_RLOCKED: TCPID_TREE_RUNLOCK(); break; } return; } /* Turn this over to tcp_log_dumpbucketlogs() to finish the work. */ tcp_log_dumpbucketlogs(tlb, reason); } /* * Mark the end of a flow with the current stack. A stack can add * stack-specific info to this trace event by overriding this * function (see bbr_log_flowend() for example). */ void tcp_log_flowend(struct tcpcb *tp) { if (tp->t_logstate != TCP_LOG_STATE_OFF) { struct socket *so = tp->t_inpcb->inp_socket; TCP_LOG_EVENT(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_FLOWEND, 0, 0, NULL, false); } } diff --git a/sys/netinet/tcp_lro.c b/sys/netinet/tcp_lro.c index a4fc5580dfc7..9ec7736aef2f 100644 --- a/sys/netinet/tcp_lro.c +++ b/sys/netinet/tcp_lro.c @@ -1,2063 +1,2062 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2007, Myricom Inc. * Copyright (c) 2008, Intel Corporation. * Copyright (c) 2012 The FreeBSD Foundation * Copyright (c) 2016-2021 Mellanox Technologies. * All rights reserved. * * Portions of this software were developed by Bjoern Zeeb * under sponsorship from the FreeBSD Foundation. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include static MALLOC_DEFINE(M_LRO, "LRO", "LRO control structures"); #define TCP_LRO_TS_OPTION \ ntohl((TCPOPT_NOP << 24) | (TCPOPT_NOP << 16) | \ (TCPOPT_TIMESTAMP << 8) | TCPOLEN_TIMESTAMP) static void tcp_lro_rx_done(struct lro_ctrl *lc); static int tcp_lro_rx_common(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum, bool use_hash); #ifdef TCPHPTS static bool do_bpf_strip_and_compress(struct inpcb *, struct lro_ctrl *, struct lro_entry *, struct mbuf **, struct mbuf **, struct mbuf **, bool *, bool, bool, struct ifnet *, bool); #endif SYSCTL_NODE(_net_inet_tcp, OID_AUTO, lro, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "TCP LRO"); static long tcplro_stacks_wanting_mbufq; counter_u64_t tcp_inp_lro_direct_queue; counter_u64_t tcp_inp_lro_wokeup_queue; counter_u64_t tcp_inp_lro_compressed; counter_u64_t tcp_inp_lro_locks_taken; counter_u64_t tcp_extra_mbuf; counter_u64_t tcp_would_have_but; counter_u64_t tcp_comp_total; counter_u64_t tcp_uncomp_total; counter_u64_t tcp_bad_csums; static unsigned tcp_lro_entries = TCP_LRO_ENTRIES; SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, entries, CTLFLAG_RDTUN | CTLFLAG_MPSAFE, &tcp_lro_entries, 0, "default number of LRO entries"); static uint32_t tcp_lro_cpu_set_thresh = TCP_LRO_CPU_DECLARATION_THRESH; SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, lro_cpu_threshold, CTLFLAG_RDTUN | CTLFLAG_MPSAFE, &tcp_lro_cpu_set_thresh, 0, "Number of interrupts in a row on the same CPU that will make us declare an 'affinity' cpu?"); static uint32_t tcp_less_accurate_lro_ts = 0; SYSCTL_UINT(_net_inet_tcp_lro, OID_AUTO, lro_less_accurate, CTLFLAG_MPSAFE, &tcp_less_accurate_lro_ts, 0, "Do we trade off efficency by doing less timestamp operations for time accuracy?"); SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, fullqueue, CTLFLAG_RD, &tcp_inp_lro_direct_queue, "Number of lro's fully queued to transport"); SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, wokeup, CTLFLAG_RD, &tcp_inp_lro_wokeup_queue, "Number of lro's where we woke up transport via hpts"); SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, compressed, CTLFLAG_RD, &tcp_inp_lro_compressed, "Number of lro's compressed and sent to transport"); SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, lockcnt, CTLFLAG_RD, &tcp_inp_lro_locks_taken, "Number of lro's inp_wlocks taken"); SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, extra_mbuf, CTLFLAG_RD, &tcp_extra_mbuf, "Number of times we had an extra compressed ack dropped into the tp"); SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, would_have_but, CTLFLAG_RD, &tcp_would_have_but, "Number of times we would have had an extra compressed, but mget failed"); SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, with_m_ackcmp, CTLFLAG_RD, &tcp_comp_total, "Number of mbufs queued with M_ACKCMP flags set"); SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, without_m_ackcmp, CTLFLAG_RD, &tcp_uncomp_total, "Number of mbufs queued without M_ACKCMP"); SYSCTL_COUNTER_U64(_net_inet_tcp_lro, OID_AUTO, lro_badcsum, CTLFLAG_RD, &tcp_bad_csums, "Number of packets that the common code saw with bad csums"); void tcp_lro_reg_mbufq(void) { atomic_fetchadd_long(&tcplro_stacks_wanting_mbufq, 1); } void tcp_lro_dereg_mbufq(void) { atomic_fetchadd_long(&tcplro_stacks_wanting_mbufq, -1); } static __inline void tcp_lro_active_insert(struct lro_ctrl *lc, struct lro_head *bucket, struct lro_entry *le) { LIST_INSERT_HEAD(&lc->lro_active, le, next); LIST_INSERT_HEAD(bucket, le, hash_next); } static __inline void tcp_lro_active_remove(struct lro_entry *le) { LIST_REMOVE(le, next); /* active list */ LIST_REMOVE(le, hash_next); /* hash bucket */ } int tcp_lro_init(struct lro_ctrl *lc) { return (tcp_lro_init_args(lc, NULL, tcp_lro_entries, 0)); } int tcp_lro_init_args(struct lro_ctrl *lc, struct ifnet *ifp, unsigned lro_entries, unsigned lro_mbufs) { struct lro_entry *le; size_t size; unsigned i, elements; lc->lro_bad_csum = 0; lc->lro_queued = 0; lc->lro_flushed = 0; lc->lro_mbuf_count = 0; lc->lro_mbuf_max = lro_mbufs; lc->lro_cnt = lro_entries; lc->lro_ackcnt_lim = TCP_LRO_ACKCNT_MAX; lc->lro_length_lim = TCP_LRO_LENGTH_MAX; lc->ifp = ifp; LIST_INIT(&lc->lro_free); LIST_INIT(&lc->lro_active); /* create hash table to accelerate entry lookup */ if (lro_entries > lro_mbufs) elements = lro_entries; else elements = lro_mbufs; lc->lro_hash = phashinit_flags(elements, M_LRO, &lc->lro_hashsz, HASH_NOWAIT); if (lc->lro_hash == NULL) { memset(lc, 0, sizeof(*lc)); return (ENOMEM); } /* compute size to allocate */ size = (lro_mbufs * sizeof(struct lro_mbuf_sort)) + (lro_entries * sizeof(*le)); lc->lro_mbuf_data = (struct lro_mbuf_sort *) malloc(size, M_LRO, M_NOWAIT | M_ZERO); /* check for out of memory */ if (lc->lro_mbuf_data == NULL) { free(lc->lro_hash, M_LRO); memset(lc, 0, sizeof(*lc)); return (ENOMEM); } /* compute offset for LRO entries */ le = (struct lro_entry *) (lc->lro_mbuf_data + lro_mbufs); /* setup linked list */ for (i = 0; i != lro_entries; i++) LIST_INSERT_HEAD(&lc->lro_free, le + i, next); return (0); } struct vxlan_header { uint32_t vxlh_flags; uint32_t vxlh_vni; }; static inline void * tcp_lro_low_level_parser(void *ptr, struct lro_parser *parser, bool update_data, bool is_vxlan, int mlen) { const struct ether_vlan_header *eh; void *old; uint16_t eth_type; if (update_data) memset(parser, 0, sizeof(*parser)); old = ptr; if (is_vxlan) { const struct vxlan_header *vxh; vxh = ptr; ptr = (uint8_t *)ptr + sizeof(*vxh); if (update_data) { parser->data.vxlan_vni = vxh->vxlh_vni & htonl(0xffffff00); } } eh = ptr; if (__predict_false(eh->evl_encap_proto == htons(ETHERTYPE_VLAN))) { eth_type = eh->evl_proto; if (update_data) { /* strip priority and keep VLAN ID only */ parser->data.vlan_id = eh->evl_tag & htons(EVL_VLID_MASK); } /* advance to next header */ ptr = (uint8_t *)ptr + ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN; mlen -= (ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN); } else { eth_type = eh->evl_encap_proto; /* advance to next header */ mlen -= ETHER_HDR_LEN; ptr = (uint8_t *)ptr + ETHER_HDR_LEN; } if (__predict_false(mlen <= 0)) return (NULL); switch (eth_type) { #ifdef INET case htons(ETHERTYPE_IP): parser->ip4 = ptr; if (__predict_false(mlen < sizeof(struct ip))) return (NULL); /* Ensure there are no IPv4 options. */ if ((parser->ip4->ip_hl << 2) != sizeof (*parser->ip4)) break; /* .. and the packet is not fragmented. */ if (parser->ip4->ip_off & htons(IP_MF|IP_OFFMASK)) break; ptr = (uint8_t *)ptr + (parser->ip4->ip_hl << 2); mlen -= sizeof(struct ip); if (update_data) { parser->data.s_addr.v4 = parser->ip4->ip_src; parser->data.d_addr.v4 = parser->ip4->ip_dst; } switch (parser->ip4->ip_p) { case IPPROTO_UDP: if (__predict_false(mlen < sizeof(struct udphdr))) return (NULL); parser->udp = ptr; if (update_data) { parser->data.lro_type = LRO_TYPE_IPV4_UDP; parser->data.s_port = parser->udp->uh_sport; parser->data.d_port = parser->udp->uh_dport; } else { MPASS(parser->data.lro_type == LRO_TYPE_IPV4_UDP); } ptr = ((uint8_t *)ptr + sizeof(*parser->udp)); parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old; return (ptr); case IPPROTO_TCP: parser->tcp = ptr; if (__predict_false(mlen < sizeof(struct tcphdr))) return (NULL); if (update_data) { parser->data.lro_type = LRO_TYPE_IPV4_TCP; parser->data.s_port = parser->tcp->th_sport; parser->data.d_port = parser->tcp->th_dport; } else { MPASS(parser->data.lro_type == LRO_TYPE_IPV4_TCP); } if (__predict_false(mlen < (parser->tcp->th_off << 2))) return (NULL); ptr = (uint8_t *)ptr + (parser->tcp->th_off << 2); parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old; return (ptr); default: break; } break; #endif #ifdef INET6 case htons(ETHERTYPE_IPV6): parser->ip6 = ptr; if (__predict_false(mlen < sizeof(struct ip6_hdr))) return (NULL); ptr = (uint8_t *)ptr + sizeof(*parser->ip6); if (update_data) { parser->data.s_addr.v6 = parser->ip6->ip6_src; parser->data.d_addr.v6 = parser->ip6->ip6_dst; } mlen -= sizeof(struct ip6_hdr); switch (parser->ip6->ip6_nxt) { case IPPROTO_UDP: if (__predict_false(mlen < sizeof(struct udphdr))) return (NULL); parser->udp = ptr; if (update_data) { parser->data.lro_type = LRO_TYPE_IPV6_UDP; parser->data.s_port = parser->udp->uh_sport; parser->data.d_port = parser->udp->uh_dport; } else { MPASS(parser->data.lro_type == LRO_TYPE_IPV6_UDP); } ptr = (uint8_t *)ptr + sizeof(*parser->udp); parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old; return (ptr); case IPPROTO_TCP: if (__predict_false(mlen < sizeof(struct tcphdr))) return (NULL); parser->tcp = ptr; if (update_data) { parser->data.lro_type = LRO_TYPE_IPV6_TCP; parser->data.s_port = parser->tcp->th_sport; parser->data.d_port = parser->tcp->th_dport; } else { MPASS(parser->data.lro_type == LRO_TYPE_IPV6_TCP); } if (__predict_false(mlen < (parser->tcp->th_off << 2))) return (NULL); ptr = (uint8_t *)ptr + (parser->tcp->th_off << 2); parser->total_hdr_len = (uint8_t *)ptr - (uint8_t *)old; return (ptr); default: break; } break; #endif default: break; } /* Invalid packet - cannot parse */ return (NULL); } static const int vxlan_csum = CSUM_INNER_L3_CALC | CSUM_INNER_L3_VALID | CSUM_INNER_L4_CALC | CSUM_INNER_L4_VALID; static inline struct lro_parser * tcp_lro_parser(struct mbuf *m, struct lro_parser *po, struct lro_parser *pi, bool update_data) { void *data_ptr; /* Try to parse outer headers first. */ data_ptr = tcp_lro_low_level_parser(m->m_data, po, update_data, false, m->m_len); if (data_ptr == NULL || po->total_hdr_len > m->m_len) return (NULL); if (update_data) { /* Store VLAN ID, if any. */ if (__predict_false(m->m_flags & M_VLANTAG)) { po->data.vlan_id = htons(m->m_pkthdr.ether_vtag) & htons(EVL_VLID_MASK); } /* Store decrypted flag, if any. */ if (__predict_false((m->m_pkthdr.csum_flags & CSUM_TLS_MASK) == CSUM_TLS_DECRYPTED)) po->data.lro_flags |= LRO_FLAG_DECRYPTED; } switch (po->data.lro_type) { case LRO_TYPE_IPV4_UDP: case LRO_TYPE_IPV6_UDP: /* Check for VXLAN headers. */ if ((m->m_pkthdr.csum_flags & vxlan_csum) != vxlan_csum) break; /* Try to parse inner headers. */ data_ptr = tcp_lro_low_level_parser(data_ptr, pi, update_data, true, (m->m_len - ((caddr_t)data_ptr - m->m_data))); if (data_ptr == NULL || (pi->total_hdr_len + po->total_hdr_len) > m->m_len) break; /* Verify supported header types. */ switch (pi->data.lro_type) { case LRO_TYPE_IPV4_TCP: case LRO_TYPE_IPV6_TCP: return (pi); default: break; } break; case LRO_TYPE_IPV4_TCP: case LRO_TYPE_IPV6_TCP: if (update_data) memset(pi, 0, sizeof(*pi)); return (po); default: break; } return (NULL); } static inline int tcp_lro_trim_mbuf_chain(struct mbuf *m, const struct lro_parser *po) { int len; switch (po->data.lro_type) { #ifdef INET case LRO_TYPE_IPV4_TCP: len = ((uint8_t *)po->ip4 - (uint8_t *)m->m_data) + ntohs(po->ip4->ip_len); break; #endif #ifdef INET6 case LRO_TYPE_IPV6_TCP: len = ((uint8_t *)po->ip6 - (uint8_t *)m->m_data) + ntohs(po->ip6->ip6_plen) + sizeof(*po->ip6); break; #endif default: return (TCP_LRO_CANNOT); } /* * If the frame is padded beyond the end of the IP packet, * then trim the extra bytes off: */ if (__predict_true(m->m_pkthdr.len == len)) { return (0); } else if (m->m_pkthdr.len > len) { m_adj(m, len - m->m_pkthdr.len); return (0); } return (TCP_LRO_CANNOT); } static struct tcphdr * tcp_lro_get_th(struct mbuf *m) { return ((struct tcphdr *)((uint8_t *)m->m_data + m->m_pkthdr.lro_tcp_h_off)); } static void lro_free_mbuf_chain(struct mbuf *m) { struct mbuf *save; while (m) { save = m->m_nextpkt; m->m_nextpkt = NULL; m_freem(m); m = save; } } void tcp_lro_free(struct lro_ctrl *lc) { struct lro_entry *le; unsigned x; /* reset LRO free list */ LIST_INIT(&lc->lro_free); /* free active mbufs, if any */ while ((le = LIST_FIRST(&lc->lro_active)) != NULL) { tcp_lro_active_remove(le); lro_free_mbuf_chain(le->m_head); } /* free hash table */ free(lc->lro_hash, M_LRO); lc->lro_hash = NULL; lc->lro_hashsz = 0; /* free mbuf array, if any */ for (x = 0; x != lc->lro_mbuf_count; x++) m_freem(lc->lro_mbuf_data[x].mb); lc->lro_mbuf_count = 0; /* free allocated memory, if any */ free(lc->lro_mbuf_data, M_LRO); lc->lro_mbuf_data = NULL; } static uint16_t tcp_lro_rx_csum_tcphdr(const struct tcphdr *th) { const uint16_t *ptr; uint32_t csum; uint16_t len; csum = -th->th_sum; /* exclude checksum field */ len = th->th_off; ptr = (const uint16_t *)th; while (len--) { csum += *ptr; ptr++; csum += *ptr; ptr++; } while (csum > 0xffff) csum = (csum >> 16) + (csum & 0xffff); return (csum); } static uint16_t tcp_lro_rx_csum_data(const struct lro_parser *pa, uint16_t tcp_csum) { uint32_t c; uint16_t cs; c = tcp_csum; switch (pa->data.lro_type) { #ifdef INET6 case LRO_TYPE_IPV6_TCP: /* Compute full pseudo IPv6 header checksum. */ cs = in6_cksum_pseudo(pa->ip6, ntohs(pa->ip6->ip6_plen), pa->ip6->ip6_nxt, 0); break; #endif #ifdef INET case LRO_TYPE_IPV4_TCP: /* Compute full pseudo IPv4 header checsum. */ cs = in_addword(ntohs(pa->ip4->ip_len) - sizeof(*pa->ip4), IPPROTO_TCP); cs = in_pseudo(pa->ip4->ip_src.s_addr, pa->ip4->ip_dst.s_addr, htons(cs)); break; #endif default: cs = 0; /* Keep compiler happy. */ break; } /* Complement checksum. */ cs = ~cs; c += cs; /* Remove TCP header checksum. */ cs = ~tcp_lro_rx_csum_tcphdr(pa->tcp); c += cs; /* Compute checksum remainder. */ while (c > 0xffff) c = (c >> 16) + (c & 0xffff); return (c); } static void tcp_lro_rx_done(struct lro_ctrl *lc) { struct lro_entry *le; while ((le = LIST_FIRST(&lc->lro_active)) != NULL) { tcp_lro_active_remove(le); tcp_lro_flush(lc, le); } } static void tcp_lro_flush_active(struct lro_ctrl *lc) { struct lro_entry *le; /* * Walk through the list of le entries, and * any one that does have packets flush. This * is called because we have an inbound packet * (e.g. SYN) that has to have all others flushed * in front of it. Note we have to do the remove * because tcp_lro_flush() assumes that the entry * is being freed. This is ok it will just get * reallocated again like it was new. */ LIST_FOREACH(le, &lc->lro_active, next) { if (le->m_head != NULL) { tcp_lro_active_remove(le); tcp_lro_flush(lc, le); } } } void tcp_lro_flush_inactive(struct lro_ctrl *lc, const struct timeval *timeout) { struct lro_entry *le, *le_tmp; uint64_t now, tov; struct bintime bt; NET_EPOCH_ASSERT(); if (LIST_EMPTY(&lc->lro_active)) return; /* get timeout time and current time in ns */ binuptime(&bt); now = bintime2ns(&bt); tov = ((timeout->tv_sec * 1000000000) + (timeout->tv_usec * 1000)); LIST_FOREACH_SAFE(le, &lc->lro_active, next, le_tmp) { if (now >= (bintime2ns(&le->alloc_time) + tov)) { tcp_lro_active_remove(le); tcp_lro_flush(lc, le); } } } #ifdef INET static int tcp_lro_rx_ipv4(struct lro_ctrl *lc, struct mbuf *m, struct ip *ip4) { uint16_t csum; /* Legacy IP has a header checksum that needs to be correct. */ if (m->m_pkthdr.csum_flags & CSUM_IP_CHECKED) { if (__predict_false((m->m_pkthdr.csum_flags & CSUM_IP_VALID) == 0)) { lc->lro_bad_csum++; return (TCP_LRO_CANNOT); } } else { csum = in_cksum_hdr(ip4); if (__predict_false(csum != 0)) { lc->lro_bad_csum++; return (TCP_LRO_CANNOT); } } return (0); } #endif #ifdef TCPHPTS static void tcp_lro_log(struct tcpcb *tp, const struct lro_ctrl *lc, const struct lro_entry *le, const struct mbuf *m, int frm, int32_t tcp_data_len, uint32_t th_seq, uint32_t th_ack, uint16_t th_win) { if (tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv, btv; uint32_t cts; cts = tcp_get_usecs(&tv); memset(&log, 0, sizeof(union tcp_log_stackspecific)); log.u_bbr.flex8 = frm; log.u_bbr.flex1 = tcp_data_len; if (m) log.u_bbr.flex2 = m->m_pkthdr.len; else log.u_bbr.flex2 = 0; if (le->m_head) { log.u_bbr.flex3 = le->m_head->m_pkthdr.lro_nsegs; log.u_bbr.flex4 = le->m_head->m_pkthdr.lro_tcp_d_len; log.u_bbr.flex5 = le->m_head->m_pkthdr.len; log.u_bbr.delRate = le->m_head->m_flags; log.u_bbr.rttProp = le->m_head->m_pkthdr.rcv_tstmp; } log.u_bbr.inflight = th_seq; log.u_bbr.delivered = th_ack; log.u_bbr.timeStamp = cts; log.u_bbr.epoch = le->next_seq; log.u_bbr.lt_epoch = le->ack_seq; log.u_bbr.pacing_gain = th_win; log.u_bbr.cwnd_gain = le->window; log.u_bbr.lost = curcpu; log.u_bbr.cur_del_rate = (uintptr_t)m; log.u_bbr.bw_inuse = (uintptr_t)le->m_head; bintime2timeval(&lc->lro_last_queue_time, &btv); log.u_bbr.flex6 = tcp_tv_to_usectick(&btv); log.u_bbr.flex7 = le->compressed; log.u_bbr.pacing_gain = le->uncompressed; if (in_epoch(net_epoch_preempt)) log.u_bbr.inhpts = 1; else log.u_bbr.inhpts = 0; TCP_LOG_EVENTP(tp, NULL, &tp->t_inpcb->inp_socket->so_rcv, &tp->t_inpcb->inp_socket->so_snd, TCP_LOG_LRO, 0, 0, &log, false, &tv); } } #endif static inline void tcp_lro_assign_and_checksum_16(uint16_t *ptr, uint16_t value, uint16_t *psum) { uint32_t csum; csum = 0xffff - *ptr + value; while (csum > 0xffff) csum = (csum >> 16) + (csum & 0xffff); *ptr = value; *psum = csum; } static uint16_t tcp_lro_update_checksum(const struct lro_parser *pa, const struct lro_entry *le, uint16_t payload_len, uint16_t delta_sum) { uint32_t csum; uint16_t tlen; uint16_t temp[5] = {}; switch (pa->data.lro_type) { case LRO_TYPE_IPV4_TCP: /* Compute new IPv4 length. */ tlen = (pa->ip4->ip_hl << 2) + (pa->tcp->th_off << 2) + payload_len; tcp_lro_assign_and_checksum_16(&pa->ip4->ip_len, htons(tlen), &temp[0]); /* Subtract delta from current IPv4 checksum. */ csum = pa->ip4->ip_sum + 0xffff - temp[0]; while (csum > 0xffff) csum = (csum >> 16) + (csum & 0xffff); tcp_lro_assign_and_checksum_16(&pa->ip4->ip_sum, csum, &temp[1]); goto update_tcp_header; case LRO_TYPE_IPV6_TCP: /* Compute new IPv6 length. */ tlen = (pa->tcp->th_off << 2) + payload_len; tcp_lro_assign_and_checksum_16(&pa->ip6->ip6_plen, htons(tlen), &temp[0]); goto update_tcp_header; case LRO_TYPE_IPV4_UDP: /* Compute new IPv4 length. */ tlen = (pa->ip4->ip_hl << 2) + sizeof(*pa->udp) + payload_len; tcp_lro_assign_and_checksum_16(&pa->ip4->ip_len, htons(tlen), &temp[0]); /* Subtract delta from current IPv4 checksum. */ csum = pa->ip4->ip_sum + 0xffff - temp[0]; while (csum > 0xffff) csum = (csum >> 16) + (csum & 0xffff); tcp_lro_assign_and_checksum_16(&pa->ip4->ip_sum, csum, &temp[1]); goto update_udp_header; case LRO_TYPE_IPV6_UDP: /* Compute new IPv6 length. */ tlen = sizeof(*pa->udp) + payload_len; tcp_lro_assign_and_checksum_16(&pa->ip6->ip6_plen, htons(tlen), &temp[0]); goto update_udp_header; default: return (0); } update_tcp_header: /* Compute current TCP header checksum. */ temp[2] = tcp_lro_rx_csum_tcphdr(pa->tcp); /* Incorporate the latest ACK into the TCP header. */ pa->tcp->th_ack = le->ack_seq; pa->tcp->th_win = le->window; /* Incorporate latest timestamp into the TCP header. */ if (le->timestamp != 0) { uint32_t *ts_ptr; ts_ptr = (uint32_t *)(pa->tcp + 1); ts_ptr[1] = htonl(le->tsval); ts_ptr[2] = le->tsecr; } /* Compute new TCP header checksum. */ temp[3] = tcp_lro_rx_csum_tcphdr(pa->tcp); /* Compute new TCP checksum. */ csum = pa->tcp->th_sum + 0xffff - delta_sum + 0xffff - temp[0] + 0xffff - temp[3] + temp[2]; while (csum > 0xffff) csum = (csum >> 16) + (csum & 0xffff); /* Assign new TCP checksum. */ tcp_lro_assign_and_checksum_16(&pa->tcp->th_sum, csum, &temp[4]); /* Compute all modififications affecting next checksum. */ csum = temp[0] + temp[1] + 0xffff - temp[2] + temp[3] + temp[4] + delta_sum; while (csum > 0xffff) csum = (csum >> 16) + (csum & 0xffff); /* Return delta checksum to next stage, if any. */ return (csum); update_udp_header: tlen = sizeof(*pa->udp) + payload_len; /* Assign new UDP length and compute checksum delta. */ tcp_lro_assign_and_checksum_16(&pa->udp->uh_ulen, htons(tlen), &temp[2]); /* Check if there is a UDP checksum. */ if (__predict_false(pa->udp->uh_sum != 0)) { /* Compute new UDP checksum. */ csum = pa->udp->uh_sum + 0xffff - delta_sum + 0xffff - temp[0] + 0xffff - temp[2]; while (csum > 0xffff) csum = (csum >> 16) + (csum & 0xffff); /* Assign new UDP checksum. */ tcp_lro_assign_and_checksum_16(&pa->udp->uh_sum, csum, &temp[3]); } /* Compute all modififications affecting next checksum. */ csum = temp[0] + temp[1] + temp[2] + temp[3] + delta_sum; while (csum > 0xffff) csum = (csum >> 16) + (csum & 0xffff); /* Return delta checksum to next stage, if any. */ return (csum); } static void tcp_flush_out_entry(struct lro_ctrl *lc, struct lro_entry *le) { /* Check if we need to recompute any checksums. */ if (le->needs_merge) { uint16_t csum; switch (le->inner.data.lro_type) { case LRO_TYPE_IPV4_TCP: csum = tcp_lro_update_checksum(&le->inner, le, le->m_head->m_pkthdr.lro_tcp_d_len, le->m_head->m_pkthdr.lro_tcp_d_csum); csum = tcp_lro_update_checksum(&le->outer, NULL, le->m_head->m_pkthdr.lro_tcp_d_len + le->inner.total_hdr_len, csum); le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID | CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID; le->m_head->m_pkthdr.csum_data = 0xffff; if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED)) le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED; break; case LRO_TYPE_IPV6_TCP: csum = tcp_lro_update_checksum(&le->inner, le, le->m_head->m_pkthdr.lro_tcp_d_len, le->m_head->m_pkthdr.lro_tcp_d_csum); csum = tcp_lro_update_checksum(&le->outer, NULL, le->m_head->m_pkthdr.lro_tcp_d_len + le->inner.total_hdr_len, csum); le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID | CSUM_PSEUDO_HDR; le->m_head->m_pkthdr.csum_data = 0xffff; if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED)) le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED; break; case LRO_TYPE_NONE: switch (le->outer.data.lro_type) { case LRO_TYPE_IPV4_TCP: csum = tcp_lro_update_checksum(&le->outer, le, le->m_head->m_pkthdr.lro_tcp_d_len, le->m_head->m_pkthdr.lro_tcp_d_csum); le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID | CSUM_PSEUDO_HDR | CSUM_IP_CHECKED | CSUM_IP_VALID; le->m_head->m_pkthdr.csum_data = 0xffff; if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED)) le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED; break; case LRO_TYPE_IPV6_TCP: csum = tcp_lro_update_checksum(&le->outer, le, le->m_head->m_pkthdr.lro_tcp_d_len, le->m_head->m_pkthdr.lro_tcp_d_csum); le->m_head->m_pkthdr.csum_flags = CSUM_DATA_VALID | CSUM_PSEUDO_HDR; le->m_head->m_pkthdr.csum_data = 0xffff; if (__predict_false(le->outer.data.lro_flags & LRO_FLAG_DECRYPTED)) le->m_head->m_pkthdr.csum_flags |= CSUM_TLS_DECRYPTED; break; default: break; } break; default: break; } } /* * Break any chain, this is not set to NULL on the singleton * case m_nextpkt points to m_head. Other case set them * m_nextpkt to NULL in push_and_replace. */ le->m_head->m_nextpkt = NULL; lc->lro_queued += le->m_head->m_pkthdr.lro_nsegs; (*lc->ifp->if_input)(lc->ifp, le->m_head); } static void tcp_set_entry_to_mbuf(struct lro_ctrl *lc, struct lro_entry *le, struct mbuf *m, struct tcphdr *th) { uint32_t *ts_ptr; uint16_t tcp_data_len; uint16_t tcp_opt_len; ts_ptr = (uint32_t *)(th + 1); tcp_opt_len = (th->th_off << 2); tcp_opt_len -= sizeof(*th); /* Check if there is a timestamp option. */ if (tcp_opt_len == 0 || __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA || *ts_ptr != TCP_LRO_TS_OPTION)) { /* We failed to find the timestamp option. */ le->timestamp = 0; } else { le->timestamp = 1; le->tsval = ntohl(*(ts_ptr + 1)); le->tsecr = *(ts_ptr + 2); } tcp_data_len = m->m_pkthdr.lro_tcp_d_len; /* Pull out TCP sequence numbers and window size. */ le->next_seq = ntohl(th->th_seq) + tcp_data_len; le->ack_seq = th->th_ack; le->window = th->th_win; le->flags = tcp_get_flags(th); le->needs_merge = 0; /* Setup new data pointers. */ le->m_head = m; le->m_tail = m_last(m); } static void tcp_push_and_replace(struct lro_ctrl *lc, struct lro_entry *le, struct mbuf *m) { struct lro_parser *pa; /* * Push up the stack of the current entry * and replace it with "m". */ struct mbuf *msave; /* Grab off the next and save it */ msave = le->m_head->m_nextpkt; le->m_head->m_nextpkt = NULL; /* Now push out the old entry */ tcp_flush_out_entry(lc, le); /* Re-parse new header, should not fail. */ pa = tcp_lro_parser(m, &le->outer, &le->inner, false); KASSERT(pa != NULL, ("tcp_push_and_replace: LRO parser failed on m=%p\n", m)); /* * Now to replace the data properly in the entry * we have to reset the TCP header and * other fields. */ tcp_set_entry_to_mbuf(lc, le, m, pa->tcp); /* Restore the next list */ m->m_nextpkt = msave; } static void tcp_lro_mbuf_append_pkthdr(struct lro_entry *le, const struct mbuf *p) { struct mbuf *m; uint32_t csum; m = le->m_head; if (m->m_pkthdr.lro_nsegs == 1) { /* Compute relative checksum. */ csum = p->m_pkthdr.lro_tcp_d_csum; } else { /* Merge TCP data checksums. */ csum = (uint32_t)m->m_pkthdr.lro_tcp_d_csum + (uint32_t)p->m_pkthdr.lro_tcp_d_csum; while (csum > 0xffff) csum = (csum >> 16) + (csum & 0xffff); } /* Update various counters. */ m->m_pkthdr.len += p->m_pkthdr.lro_tcp_d_len; m->m_pkthdr.lro_tcp_d_csum = csum; m->m_pkthdr.lro_tcp_d_len += p->m_pkthdr.lro_tcp_d_len; m->m_pkthdr.lro_nsegs += p->m_pkthdr.lro_nsegs; le->needs_merge = 1; } static void tcp_lro_condense(struct lro_ctrl *lc, struct lro_entry *le) { /* * Walk through the mbuf chain we * have on tap and compress/condense * as required. */ uint32_t *ts_ptr; struct mbuf *m; struct tcphdr *th; uint32_t tcp_data_len_total; uint32_t tcp_data_seg_total; uint16_t tcp_data_len; uint16_t tcp_opt_len; /* * First we must check the lead (m_head) * we must make sure that it is *not* * something that should be sent up * right away (sack etc). */ again: m = le->m_head->m_nextpkt; if (m == NULL) { /* Just one left. */ return; } th = tcp_lro_get_th(m); tcp_opt_len = (th->th_off << 2); tcp_opt_len -= sizeof(*th); ts_ptr = (uint32_t *)(th + 1); if (tcp_opt_len != 0 && __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA || *ts_ptr != TCP_LRO_TS_OPTION)) { /* * Its not the timestamp. We can't * use this guy as the head. */ le->m_head->m_nextpkt = m->m_nextpkt; tcp_push_and_replace(lc, le, m); goto again; } if ((tcp_get_flags(th) & ~(TH_ACK | TH_PUSH)) != 0) { /* * Make sure that previously seen segments/ACKs are delivered * before this segment, e.g. FIN. */ le->m_head->m_nextpkt = m->m_nextpkt; tcp_push_and_replace(lc, le, m); goto again; } while((m = le->m_head->m_nextpkt) != NULL) { /* * condense m into le, first * pull m out of the list. */ le->m_head->m_nextpkt = m->m_nextpkt; m->m_nextpkt = NULL; /* Setup my data */ tcp_data_len = m->m_pkthdr.lro_tcp_d_len; th = tcp_lro_get_th(m); ts_ptr = (uint32_t *)(th + 1); tcp_opt_len = (th->th_off << 2); tcp_opt_len -= sizeof(*th); tcp_data_len_total = le->m_head->m_pkthdr.lro_tcp_d_len + tcp_data_len; tcp_data_seg_total = le->m_head->m_pkthdr.lro_nsegs + m->m_pkthdr.lro_nsegs; if (tcp_data_seg_total >= lc->lro_ackcnt_lim || tcp_data_len_total >= lc->lro_length_lim) { /* Flush now if appending will result in overflow. */ tcp_push_and_replace(lc, le, m); goto again; } if (tcp_opt_len != 0 && __predict_false(tcp_opt_len != TCPOLEN_TSTAMP_APPA || *ts_ptr != TCP_LRO_TS_OPTION)) { /* * Maybe a sack in the new one? We need to * start all over after flushing the * current le. We will go up to the beginning * and flush it (calling the replace again possibly * or just returning). */ tcp_push_and_replace(lc, le, m); goto again; } if ((tcp_get_flags(th) & ~(TH_ACK | TH_PUSH)) != 0) { tcp_push_and_replace(lc, le, m); goto again; } if (tcp_opt_len != 0) { uint32_t tsval = ntohl(*(ts_ptr + 1)); /* Make sure timestamp values are increasing. */ if (TSTMP_GT(le->tsval, tsval)) { tcp_push_and_replace(lc, le, m); goto again; } le->tsval = tsval; le->tsecr = *(ts_ptr + 2); } /* Try to append the new segment. */ if (__predict_false(ntohl(th->th_seq) != le->next_seq || ((tcp_get_flags(th) & TH_ACK) != (le->flags & TH_ACK)) || (tcp_data_len == 0 && le->ack_seq == th->th_ack && le->window == th->th_win))) { /* Out of order packet, non-ACK + ACK or dup ACK. */ tcp_push_and_replace(lc, le, m); goto again; } if (tcp_data_len != 0 || SEQ_GT(ntohl(th->th_ack), ntohl(le->ack_seq))) { le->next_seq += tcp_data_len; le->ack_seq = th->th_ack; le->window = th->th_win; le->needs_merge = 1; } else if (th->th_ack == le->ack_seq) { if (WIN_GT(th->th_win, le->window)) { le->window = th->th_win; le->needs_merge = 1; } } if (tcp_data_len == 0) { m_freem(m); continue; } /* Merge TCP data checksum and length to head mbuf. */ tcp_lro_mbuf_append_pkthdr(le, m); /* * Adjust the mbuf so that m_data points to the first byte of * the ULP payload. Adjust the mbuf to avoid complications and * append new segment to existing mbuf chain. */ m_adj(m, m->m_pkthdr.len - tcp_data_len); m_demote_pkthdr(m); le->m_tail->m_next = m; le->m_tail = m_last(m); } } #ifdef TCPHPTS static void tcp_queue_pkts(struct inpcb *inp, struct tcpcb *tp, struct lro_entry *le) { INP_WLOCK_ASSERT(inp); if (tp->t_in_pkt == NULL) { /* Nothing yet there */ tp->t_in_pkt = le->m_head; tp->t_tail_pkt = le->m_last_mbuf; } else { /* Already some there */ tp->t_tail_pkt->m_nextpkt = le->m_head; tp->t_tail_pkt = le->m_last_mbuf; } le->m_head = NULL; le->m_last_mbuf = NULL; } static struct mbuf * tcp_lro_get_last_if_ackcmp(struct lro_ctrl *lc, struct lro_entry *le, struct inpcb *inp, int32_t *new_m, bool can_append_old_cmp) { struct tcpcb *tp; struct mbuf *m; tp = intotcpcb(inp); if (__predict_false(tp == NULL)) return (NULL); /* Look at the last mbuf if any in queue */ if (can_append_old_cmp) { m = tp->t_tail_pkt; if (m != NULL && (m->m_flags & M_ACKCMP) != 0) { if (M_TRAILINGSPACE(m) >= sizeof(struct tcp_ackent)) { tcp_lro_log(tp, lc, le, NULL, 23, 0, 0, 0, 0); *new_m = 0; counter_u64_add(tcp_extra_mbuf, 1); return (m); } else { /* Mark we ran out of space */ inp->inp_flags2 |= INP_MBUF_L_ACKS; } } } /* Decide mbuf size. */ tcp_lro_log(tp, lc, le, NULL, 21, 0, 0, 0, 0); if (inp->inp_flags2 & INP_MBUF_L_ACKS) m = m_getcl(M_NOWAIT, MT_DATA, M_ACKCMP | M_PKTHDR); else m = m_gethdr(M_NOWAIT, MT_DATA); if (__predict_false(m == NULL)) { counter_u64_add(tcp_would_have_but, 1); return (NULL); } counter_u64_add(tcp_comp_total, 1); m->m_pkthdr.rcvif = lc->ifp; m->m_flags |= M_ACKCMP; *new_m = 1; return (m); } static struct inpcb * tcp_lro_lookup(struct ifnet *ifp, struct lro_parser *pa) { struct inpcb *inp; switch (pa->data.lro_type) { #ifdef INET6 case LRO_TYPE_IPV6_TCP: inp = in6_pcblookup(&V_tcbinfo, &pa->data.s_addr.v6, pa->data.s_port, &pa->data.d_addr.v6, pa->data.d_port, INPLOOKUP_WLOCKPCB, ifp); break; #endif #ifdef INET case LRO_TYPE_IPV4_TCP: inp = in_pcblookup(&V_tcbinfo, pa->data.s_addr.v4, pa->data.s_port, pa->data.d_addr.v4, pa->data.d_port, INPLOOKUP_WLOCKPCB, ifp); break; #endif default: inp = NULL; break; } return (inp); } static inline bool tcp_lro_ack_valid(struct mbuf *m, struct tcphdr *th, uint32_t **ppts, bool *other_opts) { /* * This function returns two bits of valuable information. * a) Is what is present capable of being ack-compressed, * we can ack-compress if there is no options or just * a timestamp option, and of course the th_flags must * be correct as well. * b) Our other options present such as SACK. This is * used to determine if we want to wakeup or not. */ bool ret = true; switch (th->th_off << 2) { case (sizeof(*th) + TCPOLEN_TSTAMP_APPA): *ppts = (uint32_t *)(th + 1); /* Check if we have only one timestamp option. */ if (**ppts == TCP_LRO_TS_OPTION) *other_opts = false; else { *other_opts = true; ret = false; } break; case (sizeof(*th)): /* No options. */ *ppts = NULL; *other_opts = false; break; default: *ppts = NULL; *other_opts = true; ret = false; break; } /* For ACKCMP we only accept ACK, PUSH, ECE and CWR. */ if ((tcp_get_flags(th) & ~(TH_ACK | TH_PUSH | TH_ECE | TH_CWR)) != 0) ret = false; /* If it has data on it we cannot compress it */ if (m->m_pkthdr.lro_tcp_d_len) ret = false; /* ACK flag must be set. */ if (!(tcp_get_flags(th) & TH_ACK)) ret = false; return (ret); } static int tcp_lro_flush_tcphpts(struct lro_ctrl *lc, struct lro_entry *le) { struct inpcb *inp; struct tcpcb *tp; struct mbuf **pp, *cmp, *mv_to; struct ifnet *lagg_ifp; bool bpf_req, lagg_bpf_req, should_wake, can_append_old_cmp; /* Check if packet doesn't belongs to our network interface. */ if ((tcplro_stacks_wanting_mbufq == 0) || (le->outer.data.vlan_id != 0) || (le->inner.data.lro_type != LRO_TYPE_NONE)) return (TCP_LRO_CANNOT); #ifdef INET6 /* * Be proactive about unspecified IPv6 address in source. As * we use all-zero to indicate unbounded/unconnected pcb, * unspecified IPv6 address can be used to confuse us. * * Note that packets with unspecified IPv6 destination is * already dropped in ip6_input. */ if (__predict_false(le->outer.data.lro_type == LRO_TYPE_IPV6_TCP && IN6_IS_ADDR_UNSPECIFIED(&le->outer.data.s_addr.v6))) return (TCP_LRO_CANNOT); if (__predict_false(le->inner.data.lro_type == LRO_TYPE_IPV6_TCP && IN6_IS_ADDR_UNSPECIFIED(&le->inner.data.s_addr.v6))) return (TCP_LRO_CANNOT); #endif /* Lookup inp, if any. */ inp = tcp_lro_lookup(lc->ifp, (le->inner.data.lro_type == LRO_TYPE_NONE) ? &le->outer : &le->inner); if (inp == NULL) return (TCP_LRO_CANNOT); counter_u64_add(tcp_inp_lro_locks_taken, 1); /* Get TCP control structure. */ tp = intotcpcb(inp); /* Check if the inp is dead, Jim. */ - if (tp == NULL || - (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT))) { + if (tp == NULL || (inp->inp_flags & INP_DROPPED)) { INP_WUNLOCK(inp); return (TCP_LRO_CANNOT); } if ((inp->inp_irq_cpu_set == 0) && (lc->lro_cpu_is_set == 1)) { inp->inp_irq_cpu = lc->lro_last_cpu; inp->inp_irq_cpu_set = 1; } /* Check if the transport doesn't support the needed optimizations. */ if ((inp->inp_flags2 & (INP_SUPPORTS_MBUFQ | INP_MBUF_ACKCMP)) == 0) { INP_WUNLOCK(inp); return (TCP_LRO_CANNOT); } if (inp->inp_flags2 & INP_MBUF_QUEUE_READY) should_wake = false; else should_wake = true; /* Check if packets should be tapped to BPF. */ bpf_req = bpf_peers_present(lc->ifp->if_bpf); lagg_bpf_req = false; lagg_ifp = NULL; if (lc->ifp->if_type == IFT_IEEE8023ADLAG || lc->ifp->if_type == IFT_INFINIBANDLAG) { struct lagg_port *lp = lc->ifp->if_lagg; struct lagg_softc *sc = lp->lp_softc; lagg_ifp = sc->sc_ifp; if (lagg_ifp != NULL) lagg_bpf_req = bpf_peers_present(lagg_ifp->if_bpf); } /* Strip and compress all the incoming packets. */ can_append_old_cmp = true; cmp = NULL; for (pp = &le->m_head; *pp != NULL; ) { mv_to = NULL; if (do_bpf_strip_and_compress(inp, lc, le, pp, &cmp, &mv_to, &should_wake, bpf_req, lagg_bpf_req, lagg_ifp, can_append_old_cmp) == false) { /* Advance to next mbuf. */ pp = &(*pp)->m_nextpkt; /* * Once we have appended we can't look in the pending * inbound packets for a compressed ack to append to. */ can_append_old_cmp = false; /* * Once we append we also need to stop adding to any * compressed ack we were remembering. A new cmp * ack will be required. */ cmp = NULL; tcp_lro_log(tp, lc, le, NULL, 25, 0, 0, 0, 0); } else if (mv_to != NULL) { /* We are asked to move pp up */ pp = &mv_to->m_nextpkt; tcp_lro_log(tp, lc, le, NULL, 24, 0, 0, 0, 0); } else tcp_lro_log(tp, lc, le, NULL, 26, 0, 0, 0, 0); } /* Update "m_last_mbuf", if any. */ if (pp == &le->m_head) le->m_last_mbuf = *pp; else le->m_last_mbuf = __containerof(pp, struct mbuf, m_nextpkt); /* Check if any data mbufs left. */ if (le->m_head != NULL) { counter_u64_add(tcp_inp_lro_direct_queue, 1); tcp_lro_log(tp, lc, le, NULL, 22, 1, inp->inp_flags2, 0, 1); tcp_queue_pkts(inp, tp, le); } if (should_wake) { /* Wakeup */ counter_u64_add(tcp_inp_lro_wokeup_queue, 1); if ((*tp->t_fb->tfb_do_queued_segments)(inp->inp_socket, tp, 0)) inp = NULL; } if (inp != NULL) INP_WUNLOCK(inp); return (0); /* Success. */ } #endif void tcp_lro_flush(struct lro_ctrl *lc, struct lro_entry *le) { /* Only optimise if there are multiple packets waiting. */ #ifdef TCPHPTS int error; #endif NET_EPOCH_ASSERT(); #ifdef TCPHPTS CURVNET_SET(lc->ifp->if_vnet); error = tcp_lro_flush_tcphpts(lc, le); CURVNET_RESTORE(); if (error != 0) { #endif tcp_lro_condense(lc, le); tcp_flush_out_entry(lc, le); #ifdef TCPHPTS } #endif lc->lro_flushed++; bzero(le, sizeof(*le)); LIST_INSERT_HEAD(&lc->lro_free, le, next); } #ifdef HAVE_INLINE_FLSLL #define tcp_lro_msb_64(x) (1ULL << (flsll(x) - 1)) #else static inline uint64_t tcp_lro_msb_64(uint64_t x) { x |= (x >> 1); x |= (x >> 2); x |= (x >> 4); x |= (x >> 8); x |= (x >> 16); x |= (x >> 32); return (x & ~(x >> 1)); } #endif /* * The tcp_lro_sort() routine is comparable to qsort(), except it has * a worst case complexity limit of O(MIN(N,64)*N), where N is the * number of elements to sort and 64 is the number of sequence bits * available. The algorithm is bit-slicing the 64-bit sequence number, * sorting one bit at a time from the most significant bit until the * least significant one, skipping the constant bits. This is * typically called a radix sort. */ static void tcp_lro_sort(struct lro_mbuf_sort *parray, uint32_t size) { struct lro_mbuf_sort temp; uint64_t ones; uint64_t zeros; uint32_t x; uint32_t y; repeat: /* for small arrays insertion sort is faster */ if (size <= 12) { for (x = 1; x < size; x++) { temp = parray[x]; for (y = x; y > 0 && temp.seq < parray[y - 1].seq; y--) parray[y] = parray[y - 1]; parray[y] = temp; } return; } /* compute sequence bits which are constant */ ones = 0; zeros = 0; for (x = 0; x != size; x++) { ones |= parray[x].seq; zeros |= ~parray[x].seq; } /* compute bits which are not constant into "ones" */ ones &= zeros; if (ones == 0) return; /* pick the most significant bit which is not constant */ ones = tcp_lro_msb_64(ones); /* * Move entries having cleared sequence bits to the beginning * of the array: */ for (x = y = 0; y != size; y++) { /* skip set bits */ if (parray[y].seq & ones) continue; /* swap entries */ temp = parray[x]; parray[x] = parray[y]; parray[y] = temp; x++; } KASSERT(x != 0 && x != size, ("Memory is corrupted\n")); /* sort zeros */ tcp_lro_sort(parray, x); /* sort ones */ parray += x; size -= x; goto repeat; } void tcp_lro_flush_all(struct lro_ctrl *lc) { uint64_t seq; uint64_t nseq; unsigned x; NET_EPOCH_ASSERT(); /* check if no mbufs to flush */ if (lc->lro_mbuf_count == 0) goto done; if (lc->lro_cpu_is_set == 0) { if (lc->lro_last_cpu == curcpu) { lc->lro_cnt_of_same_cpu++; /* Have we reached the threshold to declare a cpu? */ if (lc->lro_cnt_of_same_cpu > tcp_lro_cpu_set_thresh) lc->lro_cpu_is_set = 1; } else { lc->lro_last_cpu = curcpu; lc->lro_cnt_of_same_cpu = 0; } } CURVNET_SET(lc->ifp->if_vnet); /* get current time */ binuptime(&lc->lro_last_queue_time); /* sort all mbufs according to stream */ tcp_lro_sort(lc->lro_mbuf_data, lc->lro_mbuf_count); /* input data into LRO engine, stream by stream */ seq = 0; for (x = 0; x != lc->lro_mbuf_count; x++) { struct mbuf *mb; /* get mbuf */ mb = lc->lro_mbuf_data[x].mb; /* get sequence number, masking away the packet index */ nseq = lc->lro_mbuf_data[x].seq & (-1ULL << 24); /* check for new stream */ if (seq != nseq) { seq = nseq; /* flush active streams */ tcp_lro_rx_done(lc); } /* add packet to LRO engine */ if (tcp_lro_rx_common(lc, mb, 0, false) != 0) { /* Flush anything we have acummulated */ tcp_lro_flush_active(lc); /* input packet to network layer */ (*lc->ifp->if_input)(lc->ifp, mb); lc->lro_queued++; lc->lro_flushed++; } } CURVNET_RESTORE(); done: /* flush active streams */ tcp_lro_rx_done(lc); #ifdef TCPHPTS tcp_run_hpts(); #endif lc->lro_mbuf_count = 0; } #ifdef TCPHPTS static void build_ack_entry(struct tcp_ackent *ae, struct tcphdr *th, struct mbuf *m, uint32_t *ts_ptr, uint16_t iptos) { /* * Given a TCP ACK, summarize it down into the small TCP ACK * entry. */ ae->timestamp = m->m_pkthdr.rcv_tstmp; ae->flags = 0; if (m->m_flags & M_TSTMP_LRO) ae->flags |= TSTMP_LRO; else if (m->m_flags & M_TSTMP) ae->flags |= TSTMP_HDWR; ae->seq = ntohl(th->th_seq); ae->ack = ntohl(th->th_ack); ae->flags |= tcp_get_flags(th); if (ts_ptr != NULL) { ae->ts_value = ntohl(ts_ptr[1]); ae->ts_echo = ntohl(ts_ptr[2]); ae->flags |= HAS_TSTMP; } ae->win = ntohs(th->th_win); ae->codepoint = iptos; } /* * Do BPF tap for either ACK_CMP packets or MBUF QUEUE type packets * and strip all, but the IPv4/IPv6 header. */ static bool do_bpf_strip_and_compress(struct inpcb *inp, struct lro_ctrl *lc, struct lro_entry *le, struct mbuf **pp, struct mbuf **cmp, struct mbuf **mv_to, bool *should_wake, bool bpf_req, bool lagg_bpf_req, struct ifnet *lagg_ifp, bool can_append_old_cmp) { union { void *ptr; struct ip *ip4; struct ip6_hdr *ip6; } l3; struct mbuf *m; struct mbuf *nm; struct tcphdr *th; struct tcp_ackent *ack_ent; uint32_t *ts_ptr; int32_t n_mbuf; bool other_opts, can_compress; uint8_t lro_type; uint16_t iptos; int tcp_hdr_offset; int idx; /* Get current mbuf. */ m = *pp; /* Let the BPF see the packet */ if (__predict_false(bpf_req)) ETHER_BPF_MTAP(lc->ifp, m); if (__predict_false(lagg_bpf_req)) ETHER_BPF_MTAP(lagg_ifp, m); tcp_hdr_offset = m->m_pkthdr.lro_tcp_h_off; lro_type = le->inner.data.lro_type; switch (lro_type) { case LRO_TYPE_NONE: lro_type = le->outer.data.lro_type; switch (lro_type) { case LRO_TYPE_IPV4_TCP: tcp_hdr_offset -= sizeof(*le->outer.ip4); m->m_pkthdr.lro_etype = ETHERTYPE_IP; break; case LRO_TYPE_IPV6_TCP: tcp_hdr_offset -= sizeof(*le->outer.ip6); m->m_pkthdr.lro_etype = ETHERTYPE_IPV6; break; default: goto compressed; } break; case LRO_TYPE_IPV4_TCP: tcp_hdr_offset -= sizeof(*le->outer.ip4); m->m_pkthdr.lro_etype = ETHERTYPE_IP; break; case LRO_TYPE_IPV6_TCP: tcp_hdr_offset -= sizeof(*le->outer.ip6); m->m_pkthdr.lro_etype = ETHERTYPE_IPV6; break; default: goto compressed; } MPASS(tcp_hdr_offset >= 0); m_adj(m, tcp_hdr_offset); m->m_flags |= M_LRO_EHDRSTRP; m->m_flags &= ~M_ACKCMP; m->m_pkthdr.lro_tcp_h_off -= tcp_hdr_offset; th = tcp_lro_get_th(m); th->th_sum = 0; /* TCP checksum is valid. */ /* Check if ACK can be compressed */ can_compress = tcp_lro_ack_valid(m, th, &ts_ptr, &other_opts); /* Now lets look at the should wake states */ if ((other_opts == true) && ((inp->inp_flags2 & INP_DONT_SACK_QUEUE) == 0)) { /* * If there are other options (SACK?) and the * tcp endpoint has not expressly told us it does * not care about SACKS, then we should wake up. */ *should_wake = true; } /* Is the ack compressable? */ if (can_compress == false) goto done; /* Does the TCP endpoint support ACK compression? */ if ((inp->inp_flags2 & INP_MBUF_ACKCMP) == 0) goto done; /* Lets get the TOS/traffic class field */ l3.ptr = mtod(m, void *); switch (lro_type) { case LRO_TYPE_IPV4_TCP: iptos = l3.ip4->ip_tos; break; case LRO_TYPE_IPV6_TCP: iptos = IPV6_TRAFFIC_CLASS(l3.ip6); break; default: iptos = 0; /* Keep compiler happy. */ break; } /* Now lets get space if we don't have some already */ if (*cmp == NULL) { new_one: nm = tcp_lro_get_last_if_ackcmp(lc, le, inp, &n_mbuf, can_append_old_cmp); if (__predict_false(nm == NULL)) goto done; *cmp = nm; if (n_mbuf) { /* * Link in the new cmp ack to our in-order place, * first set our cmp ack's next to where we are. */ nm->m_nextpkt = m; (*pp) = nm; /* * Set it up so mv_to is advanced to our * compressed ack. This way the caller can * advance pp to the right place. */ *mv_to = nm; /* * Advance it here locally as well. */ pp = &nm->m_nextpkt; } } else { /* We have one already we are working on */ nm = *cmp; if (M_TRAILINGSPACE(nm) < sizeof(struct tcp_ackent)) { /* We ran out of space */ inp->inp_flags2 |= INP_MBUF_L_ACKS; goto new_one; } } MPASS(M_TRAILINGSPACE(nm) >= sizeof(struct tcp_ackent)); counter_u64_add(tcp_inp_lro_compressed, 1); le->compressed++; /* We can add in to the one on the tail */ ack_ent = mtod(nm, struct tcp_ackent *); idx = (nm->m_len / sizeof(struct tcp_ackent)); build_ack_entry(&ack_ent[idx], th, m, ts_ptr, iptos); /* Bump the size of both pkt-hdr and len */ nm->m_len += sizeof(struct tcp_ackent); nm->m_pkthdr.len += sizeof(struct tcp_ackent); compressed: /* Advance to next mbuf before freeing. */ *pp = m->m_nextpkt; m->m_nextpkt = NULL; m_freem(m); return (true); done: counter_u64_add(tcp_uncomp_total, 1); le->uncompressed++; return (false); } #endif static struct lro_head * tcp_lro_rx_get_bucket(struct lro_ctrl *lc, struct mbuf *m, struct lro_parser *parser) { u_long hash; if (M_HASHTYPE_ISHASH(m)) { hash = m->m_pkthdr.flowid; } else { for (unsigned i = hash = 0; i != LRO_RAW_ADDRESS_MAX; i++) hash += parser->data.raw[i]; } return (&lc->lro_hash[hash % lc->lro_hashsz]); } static int tcp_lro_rx_common(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum, bool use_hash) { struct lro_parser pi; /* inner address data */ struct lro_parser po; /* outer address data */ struct lro_parser *pa; /* current parser for TCP stream */ struct lro_entry *le; struct lro_head *bucket; struct tcphdr *th; int tcp_data_len; int tcp_opt_len; int error; uint16_t tcp_data_sum; #ifdef INET /* Quickly decide if packet cannot be LRO'ed */ if (__predict_false(V_ipforwarding != 0)) return (TCP_LRO_CANNOT); #endif #ifdef INET6 /* Quickly decide if packet cannot be LRO'ed */ if (__predict_false(V_ip6_forwarding != 0)) return (TCP_LRO_CANNOT); #endif if (((m->m_pkthdr.csum_flags & (CSUM_DATA_VALID | CSUM_PSEUDO_HDR)) != ((CSUM_DATA_VALID | CSUM_PSEUDO_HDR))) || (m->m_pkthdr.csum_data != 0xffff)) { /* * The checksum either did not have hardware offload * or it was a bad checksum. We can't LRO such * a packet. */ counter_u64_add(tcp_bad_csums, 1); return (TCP_LRO_CANNOT); } /* We expect a contiguous header [eh, ip, tcp]. */ pa = tcp_lro_parser(m, &po, &pi, true); if (__predict_false(pa == NULL)) return (TCP_LRO_NOT_SUPPORTED); /* We don't expect any padding. */ error = tcp_lro_trim_mbuf_chain(m, pa); if (__predict_false(error != 0)) return (error); #ifdef INET switch (pa->data.lro_type) { case LRO_TYPE_IPV4_TCP: error = tcp_lro_rx_ipv4(lc, m, pa->ip4); if (__predict_false(error != 0)) return (error); break; default: break; } #endif /* If no hardware or arrival stamp on the packet add timestamp */ if ((m->m_flags & (M_TSTMP_LRO | M_TSTMP)) == 0) { m->m_pkthdr.rcv_tstmp = bintime2ns(&lc->lro_last_queue_time); m->m_flags |= M_TSTMP_LRO; } /* Get pointer to TCP header. */ th = pa->tcp; /* Don't process SYN packets. */ if (__predict_false(tcp_get_flags(th) & TH_SYN)) return (TCP_LRO_CANNOT); /* Get total TCP header length and compute payload length. */ tcp_opt_len = (th->th_off << 2); tcp_data_len = m->m_pkthdr.len - ((uint8_t *)th - (uint8_t *)m->m_data) - tcp_opt_len; tcp_opt_len -= sizeof(*th); /* Don't process invalid TCP headers. */ if (__predict_false(tcp_opt_len < 0 || tcp_data_len < 0)) return (TCP_LRO_CANNOT); /* Compute TCP data only checksum. */ if (tcp_data_len == 0) tcp_data_sum = 0; /* no data, no checksum */ else if (__predict_false(csum != 0)) tcp_data_sum = tcp_lro_rx_csum_data(pa, ~csum); else tcp_data_sum = tcp_lro_rx_csum_data(pa, ~th->th_sum); /* Save TCP info in mbuf. */ m->m_nextpkt = NULL; m->m_pkthdr.rcvif = lc->ifp; m->m_pkthdr.lro_tcp_d_csum = tcp_data_sum; m->m_pkthdr.lro_tcp_d_len = tcp_data_len; m->m_pkthdr.lro_tcp_h_off = ((uint8_t *)th - (uint8_t *)m->m_data); m->m_pkthdr.lro_nsegs = 1; /* Get hash bucket. */ if (!use_hash) { bucket = &lc->lro_hash[0]; } else { bucket = tcp_lro_rx_get_bucket(lc, m, pa); } /* Try to find a matching previous segment. */ LIST_FOREACH(le, bucket, hash_next) { /* Compare addresses and ports. */ if (lro_address_compare(&po.data, &le->outer.data) == false || lro_address_compare(&pi.data, &le->inner.data) == false) continue; /* Check if no data and old ACK. */ if (tcp_data_len == 0 && SEQ_LT(ntohl(th->th_ack), ntohl(le->ack_seq))) { m_freem(m); return (0); } /* Mark "m" in the last spot. */ le->m_last_mbuf->m_nextpkt = m; /* Now set the tail to "m". */ le->m_last_mbuf = m; return (0); } /* Try to find an empty slot. */ if (LIST_EMPTY(&lc->lro_free)) return (TCP_LRO_NO_ENTRIES); /* Start a new segment chain. */ le = LIST_FIRST(&lc->lro_free); LIST_REMOVE(le, next); tcp_lro_active_insert(lc, bucket, le); /* Make sure the headers are set. */ le->inner = pi; le->outer = po; /* Store time this entry was allocated. */ le->alloc_time = lc->lro_last_queue_time; tcp_set_entry_to_mbuf(lc, le, m, th); /* Now set the tail to "m". */ le->m_last_mbuf = m; return (0); } int tcp_lro_rx(struct lro_ctrl *lc, struct mbuf *m, uint32_t csum) { int error; if (((m->m_pkthdr.csum_flags & (CSUM_DATA_VALID | CSUM_PSEUDO_HDR)) != ((CSUM_DATA_VALID | CSUM_PSEUDO_HDR))) || (m->m_pkthdr.csum_data != 0xffff)) { /* * The checksum either did not have hardware offload * or it was a bad checksum. We can't LRO such * a packet. */ counter_u64_add(tcp_bad_csums, 1); return (TCP_LRO_CANNOT); } /* get current time */ binuptime(&lc->lro_last_queue_time); CURVNET_SET(lc->ifp->if_vnet); error = tcp_lro_rx_common(lc, m, csum, true); if (__predict_false(error != 0)) { /* * Flush anything we have acummulated * ahead of this packet that can't * be LRO'd. This preserves order. */ tcp_lro_flush_active(lc); } CURVNET_RESTORE(); return (error); } void tcp_lro_queue_mbuf(struct lro_ctrl *lc, struct mbuf *mb) { NET_EPOCH_ASSERT(); /* sanity checks */ if (__predict_false(lc->ifp == NULL || lc->lro_mbuf_data == NULL || lc->lro_mbuf_max == 0)) { /* packet drop */ m_freem(mb); return; } /* check if packet is not LRO capable */ if (__predict_false((lc->ifp->if_capenable & IFCAP_LRO) == 0)) { /* input packet to network layer */ (*lc->ifp->if_input) (lc->ifp, mb); return; } /* If no hardware or arrival stamp on the packet add timestamp */ if ((tcplro_stacks_wanting_mbufq > 0) && (tcp_less_accurate_lro_ts == 0) && ((mb->m_flags & M_TSTMP) == 0)) { /* Add in an LRO time since no hardware */ binuptime(&lc->lro_last_queue_time); mb->m_pkthdr.rcv_tstmp = bintime2ns(&lc->lro_last_queue_time); mb->m_flags |= M_TSTMP_LRO; } /* create sequence number */ lc->lro_mbuf_data[lc->lro_mbuf_count].seq = (((uint64_t)M_HASHTYPE_GET(mb)) << 56) | (((uint64_t)mb->m_pkthdr.flowid) << 24) | ((uint64_t)lc->lro_mbuf_count); /* enter mbuf */ lc->lro_mbuf_data[lc->lro_mbuf_count].mb = mb; /* flush if array is full */ if (__predict_false(++lc->lro_mbuf_count == lc->lro_mbuf_max)) tcp_lro_flush_all(lc); } /* end */ diff --git a/sys/netinet/tcp_stacks/bbr.c b/sys/netinet/tcp_stacks/bbr.c index 31b5c2cc78dc..40d3a40dbe38 100644 --- a/sys/netinet/tcp_stacks/bbr.c +++ b/sys/netinet/tcp_stacks/bbr.c @@ -1,14887 +1,14887 @@ /*- * Copyright (c) 2016-2020 Netflix, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * */ /** * Author: Randall Stewart * This work is based on the ACM Queue paper * BBR - Congestion Based Congestion Control * and also numerous discussions with Neal, Yuchung and Van. */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include "opt_ipsec.h" #include "opt_tcpdebug.h" #include "opt_ratelimit.h" #include #include #include #include #include #ifdef TCP_HHOOK #include #endif #include #include #include #include #include #include #include #ifdef STATS #include #include #include /* Must come after qmath.h and tree.h */ #endif #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #define TCPSTATES /* for logging */ #include #include #include #include #include /* required for icmp_var.h */ #include /* for ICMP_BANDLIM */ #include #include #include #include #define TCPOUTFLAGS #include #include #include #include #include #include #include #include #include #include #include #ifdef TCPDEBUG #include #endif /* TCPDEBUG */ #ifdef TCP_OFFLOAD #include #endif #ifdef INET6 #include #endif #include #include #include #include #include #if defined(IPSEC) || defined(IPSEC_SUPPORT) #include #include #endif /* IPSEC */ #include #include #include #ifdef MAC #include #endif #include "sack_filter.h" #include "tcp_bbr.h" #include "rack_bbr_common.h" uma_zone_t bbr_zone; uma_zone_t bbr_pcb_zone; struct sysctl_ctx_list bbr_sysctl_ctx; struct sysctl_oid *bbr_sysctl_root; #define TCPT_RANGESET_NOSLOP(tv, value, tvmin, tvmax) do { \ (tv) = (value); \ if ((u_long)(tv) < (u_long)(tvmin)) \ (tv) = (tvmin); \ if ((u_long)(tv) > (u_long)(tvmax)) \ (tv) = (tvmax); \ } while(0) /*#define BBR_INVARIANT 1*/ /* * initial window */ static uint32_t bbr_def_init_win = 10; static int32_t bbr_persist_min = 250000; /* 250ms */ static int32_t bbr_persist_max = 1000000; /* 1 Second */ static int32_t bbr_cwnd_may_shrink = 0; static int32_t bbr_cwndtarget_rtt_touse = BBR_RTT_PROP; static int32_t bbr_num_pktepo_for_del_limit = BBR_NUM_RTTS_FOR_DEL_LIMIT; static int32_t bbr_hardware_pacing_limit = 8000; static int32_t bbr_quanta = 3; /* How much extra quanta do we get? */ static int32_t bbr_no_retran = 0; static int32_t bbr_error_base_paceout = 10000; /* usec to pace */ static int32_t bbr_max_net_error_cnt = 10; /* Should the following be dynamic too -- loss wise */ static int32_t bbr_rtt_gain_thresh = 0; /* Measurement controls */ static int32_t bbr_use_google_algo = 1; static int32_t bbr_ts_limiting = 1; static int32_t bbr_ts_can_raise = 0; static int32_t bbr_do_red = 600; static int32_t bbr_red_scale = 20000; static int32_t bbr_red_mul = 1; static int32_t bbr_red_div = 2; static int32_t bbr_red_growth_restrict = 1; static int32_t bbr_target_is_bbunit = 0; static int32_t bbr_drop_limit = 0; /* * How much gain do we need to see to * stay in startup? */ static int32_t bbr_marks_rxt_sack_passed = 0; static int32_t bbr_start_exit = 25; static int32_t bbr_low_start_exit = 25; /* When we are in reduced gain */ static int32_t bbr_startup_loss_thresh = 2000; /* 20.00% loss */ static int32_t bbr_hptsi_max_mul = 1; /* These two mul/div assure a min pacing */ static int32_t bbr_hptsi_max_div = 2; /* time, 0 means turned off. We need this * if we go back ever to where the pacer * has priority over timers. */ static int32_t bbr_policer_call_from_rack_to = 0; static int32_t bbr_policer_detection_enabled = 1; static int32_t bbr_min_measurements_req = 1; /* We need at least 2 * measurements before we are * "good" note that 2 == 1. * This is because we use a > * comparison. This means if * min_measure was 0, it takes * num-measures > min(0) and * you get 1 measurement and * you are good. Set to 1, you * have to have two * measurements (this is done * to prevent it from being ok * to have no measurements). */ static int32_t bbr_no_pacing_until = 4; static int32_t bbr_min_usec_delta = 20000; /* 20,000 usecs */ static int32_t bbr_min_peer_delta = 20; /* 20 units */ static int32_t bbr_delta_percent = 150; /* 15.0 % */ static int32_t bbr_target_cwnd_mult_limit = 8; /* * bbr_cwnd_min_val is the number of * segments we hold to in the RTT probe * state typically 4. */ static int32_t bbr_cwnd_min_val = BBR_PROBERTT_NUM_MSS; static int32_t bbr_cwnd_min_val_hs = BBR_HIGHSPEED_NUM_MSS; static int32_t bbr_gain_to_target = 1; static int32_t bbr_gain_gets_extra_too = 1; /* * bbr_high_gain is the 2/ln(2) value we need * to double the sending rate in startup. This * is used for both cwnd and hptsi gain's. */ static int32_t bbr_high_gain = BBR_UNIT * 2885 / 1000 + 1; static int32_t bbr_startup_lower = BBR_UNIT * 1500 / 1000 + 1; static int32_t bbr_use_lower_gain_in_startup = 1; /* thresholds for reduction on drain in sub-states/drain */ static int32_t bbr_drain_rtt = BBR_SRTT; static int32_t bbr_drain_floor = 88; static int32_t google_allow_early_out = 1; static int32_t google_consider_lost = 1; static int32_t bbr_drain_drop_mul = 4; static int32_t bbr_drain_drop_div = 5; static int32_t bbr_rand_ot = 50; static int32_t bbr_can_force_probertt = 0; static int32_t bbr_can_adjust_probertt = 1; static int32_t bbr_probertt_sets_rtt = 0; static int32_t bbr_can_use_ts_for_rtt = 1; static int32_t bbr_is_ratio = 0; static int32_t bbr_sub_drain_app_limit = 1; static int32_t bbr_prtt_slam_cwnd = 1; static int32_t bbr_sub_drain_slam_cwnd = 1; static int32_t bbr_slam_cwnd_in_main_drain = 1; static int32_t bbr_filter_len_sec = 6; /* How long does the rttProp filter * hold */ static uint32_t bbr_rtt_probe_limit = (USECS_IN_SECOND * 4); /* * bbr_drain_gain is the reverse of the high_gain * designed to drain back out the standing queue * that is formed in startup by causing a larger * hptsi gain and thus drainging the packets * in flight. */ static int32_t bbr_drain_gain = BBR_UNIT * 1000 / 2885; static int32_t bbr_rttprobe_gain = 192; /* * The cwnd_gain is the default cwnd gain applied when * calculating a target cwnd. Note that the cwnd is * a secondary factor in the way BBR works (see the * paper and think about it, it will take some time). * Basically the hptsi_gain spreads the packets out * so you never get more than BDP to the peer even * if the cwnd is high. In our implemenation that * means in non-recovery/retransmission scenarios * cwnd will never be reached by the flight-size. */ static int32_t bbr_cwnd_gain = BBR_UNIT * 2; static int32_t bbr_tlp_type_to_use = BBR_SRTT; static int32_t bbr_delack_time = 100000; /* 100ms in useconds */ static int32_t bbr_sack_not_required = 0; /* set to one to allow non-sack to use bbr */ static int32_t bbr_initial_bw_bps = 62500; /* 500kbps in bytes ps */ static int32_t bbr_ignore_data_after_close = 1; static int16_t bbr_hptsi_gain[] = { (BBR_UNIT *5 / 4), (BBR_UNIT * 3 / 4), BBR_UNIT, BBR_UNIT, BBR_UNIT, BBR_UNIT, BBR_UNIT, BBR_UNIT }; int32_t bbr_use_rack_resend_cheat = 1; int32_t bbr_sends_full_iwnd = 1; #define BBR_HPTSI_GAIN_MAX 8 /* * The BBR module incorporates a number of * TCP ideas that have been put out into the IETF * over the last few years: * - Yuchung Cheng's RACK TCP (for which its named) that * will stop us using the number of dup acks and instead * use time as the gage of when we retransmit. * - Reorder Detection of RFC4737 and the Tail-Loss probe draft * of Dukkipati et.al. * - Van Jacobson's et.al BBR. * * RACK depends on SACK, so if an endpoint arrives that * cannot do SACK the state machine below will shuttle the * connection back to using the "default" TCP stack that is * in FreeBSD. * * To implement BBR and RACK the original TCP stack was first decomposed * into a functional state machine with individual states * for each of the possible TCP connection states. The do_segment * functions role in life is to mandate the connection supports SACK * initially and then assure that the RACK state matches the conenction * state before calling the states do_segment function. Data processing * of inbound segments also now happens in the hpts_do_segment in general * with only one exception. This is so we can keep the connection on * a single CPU. * * Each state is simplified due to the fact that the original do_segment * has been decomposed and we *know* what state we are in (no * switches on the state) and all tests for SACK are gone. This * greatly simplifies what each state does. * * TCP output is also over-written with a new version since it * must maintain the new rack scoreboard and has had hptsi * integrated as a requirment. Still todo is to eliminate the * use of the callout_() system and use the hpts for all * timers as well. */ static uint32_t bbr_rtt_probe_time = 200000; /* 200ms in micro seconds */ static uint32_t bbr_rtt_probe_cwndtarg = 4; /* How many mss's outstanding */ static const int32_t bbr_min_req_free = 2; /* The min we must have on the * free list */ static int32_t bbr_tlp_thresh = 1; static int32_t bbr_reorder_thresh = 2; static int32_t bbr_reorder_fade = 60000000; /* 0 - never fade, def * 60,000,000 - 60 seconds */ static int32_t bbr_pkt_delay = 1000; static int32_t bbr_min_to = 1000; /* Number of usec's minimum timeout */ static int32_t bbr_incr_timers = 1; static int32_t bbr_tlp_min = 10000; /* 10ms in usecs */ static int32_t bbr_delayed_ack_time = 200000; /* 200ms in usecs */ static int32_t bbr_exit_startup_at_loss = 1; /* * bbr_lt_bw_ratio is 1/8th * bbr_lt_bw_diff is < 4 Kbit/sec */ static uint64_t bbr_lt_bw_diff = 4000 / 8; /* In bytes per second */ static uint64_t bbr_lt_bw_ratio = 8; /* For 1/8th */ static uint32_t bbr_lt_bw_max_rtts = 48; /* How many rtt's do we use * the lt_bw for */ static uint32_t bbr_lt_intvl_min_rtts = 4; /* Min num of RTT's to measure * lt_bw */ static int32_t bbr_lt_intvl_fp = 0; /* False positive epoch diff */ static int32_t bbr_lt_loss_thresh = 196; /* Lost vs delivered % */ static int32_t bbr_lt_fd_thresh = 100; /* false detection % */ static int32_t bbr_verbose_logging = 0; /* * Currently regular tcp has a rto_min of 30ms * the backoff goes 12 times so that ends up * being a total of 122.850 seconds before a * connection is killed. */ static int32_t bbr_rto_min_ms = 30; /* 30ms same as main freebsd */ static int32_t bbr_rto_max_sec = 4; /* 4 seconds */ /****************************************************/ /* DEFAULT TSO SIZING (cpu performance impacting) */ /****************************************************/ /* What amount is our formula using to get TSO size */ static int32_t bbr_hptsi_per_second = 1000; /* * For hptsi under bbr_cross_over connections what is delay * target 7ms (in usec) combined with a seg_max of 2 * gets us close to identical google behavior in * TSO size selection (possibly more 1MSS sends). */ static int32_t bbr_hptsi_segments_delay_tar = 7000; /* Does pacing delay include overhead's in its time calculations? */ static int32_t bbr_include_enet_oh = 0; static int32_t bbr_include_ip_oh = 1; static int32_t bbr_include_tcp_oh = 1; static int32_t bbr_google_discount = 10; /* Do we use (nf mode) pkt-epoch to drive us or rttProp? */ static int32_t bbr_state_is_pkt_epoch = 0; static int32_t bbr_state_drain_2_tar = 1; /* What is the max the 0 - bbr_cross_over MBPS TSO target * can reach using our delay target. Note that this * value becomes the floor for the cross over * algorithm. */ static int32_t bbr_hptsi_segments_max = 2; static int32_t bbr_hptsi_segments_floor = 1; static int32_t bbr_hptsi_utter_max = 0; /* What is the min the 0 - bbr_cross-over MBPS TSO target can be */ static int32_t bbr_hptsi_bytes_min = 1460; static int32_t bbr_all_get_min = 0; /* Cross over point from algo-a to algo-b */ static uint32_t bbr_cross_over = TWENTY_THREE_MBPS; /* Do we deal with our restart state? */ static int32_t bbr_uses_idle_restart = 0; static int32_t bbr_idle_restart_threshold = 100000; /* 100ms in useconds */ /* Do we allow hardware pacing? */ static int32_t bbr_allow_hdwr_pacing = 0; static int32_t bbr_hdwr_pace_adjust = 2; /* multipler when we calc the tso size */ static int32_t bbr_hdwr_pace_floor = 1; static int32_t bbr_hdwr_pacing_delay_cnt = 10; /****************************************************/ static int32_t bbr_resends_use_tso = 0; static int32_t bbr_tlp_max_resend = 2; static int32_t bbr_sack_block_limit = 128; #define BBR_MAX_STAT 19 counter_u64_t bbr_state_time[BBR_MAX_STAT]; counter_u64_t bbr_state_lost[BBR_MAX_STAT]; counter_u64_t bbr_state_resend[BBR_MAX_STAT]; counter_u64_t bbr_stat_arry[BBR_STAT_SIZE]; counter_u64_t bbr_opts_arry[BBR_OPTS_SIZE]; counter_u64_t bbr_out_size[TCP_MSS_ACCT_SIZE]; counter_u64_t bbr_flows_whdwr_pacing; counter_u64_t bbr_flows_nohdwr_pacing; counter_u64_t bbr_nohdwr_pacing_enobuf; counter_u64_t bbr_hdwr_pacing_enobuf; static inline uint64_t bbr_get_bw(struct tcp_bbr *bbr); /* * Static defintions we need for forward declarations. */ static uint32_t bbr_get_pacing_length(struct tcp_bbr *bbr, uint16_t gain, uint32_t useconds_time, uint64_t bw); static uint32_t bbr_get_a_state_target(struct tcp_bbr *bbr, uint32_t gain); static void bbr_set_state(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t win); static void bbr_set_probebw_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses); static void bbr_substate_change(struct tcp_bbr *bbr, uint32_t cts, int line, int dolog); static uint32_t bbr_get_target_cwnd(struct tcp_bbr *bbr, uint64_t bw, uint32_t gain); static void bbr_state_change(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, int32_t pkt_epoch, uint32_t losses); static uint32_t bbr_calc_thresh_rack(struct tcp_bbr *bbr, uint32_t srtt, uint32_t cts, struct bbr_sendmap *rsm); static uint32_t bbr_initial_cwnd(struct tcp_bbr *bbr, struct tcpcb *tp); static uint32_t bbr_calc_thresh_tlp(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t srtt, uint32_t cts); static void bbr_exit_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line); static void bbr_set_state_target(struct tcp_bbr *bbr, int line); static void bbr_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts, int32_t line); static void bbr_log_progress_event(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t tick, int event, int line); static void tcp_bbr_tso_size_check(struct tcp_bbr *bbr, uint32_t cts); static void bbr_setup_red_bw(struct tcp_bbr *bbr, uint32_t cts); static void bbr_log_rtt_shrinks(struct tcp_bbr *bbr, uint32_t cts, uint32_t applied, uint32_t rtt, uint32_t line, uint8_t is_start, uint16_t set); static struct bbr_sendmap * bbr_find_lowest_rsm(struct tcp_bbr *bbr); static __inline uint32_t bbr_get_rtt(struct tcp_bbr *bbr, int32_t rtt_type); static void bbr_log_to_start(struct tcp_bbr *bbr, uint32_t cts, uint32_t to, int32_t slot, uint8_t which); static void bbr_log_timer_var(struct tcp_bbr *bbr, int mode, uint32_t cts, uint32_t time_since_sent, uint32_t srtt, uint32_t thresh, uint32_t to); static void bbr_log_hpts_diag(struct tcp_bbr *bbr, uint32_t cts, struct hpts_diag *diag); static void bbr_log_type_bbrsnd(struct tcp_bbr *bbr, uint32_t len, uint32_t slot, uint32_t del_by, uint32_t cts, uint32_t sloton, uint32_t prev_delay); static void bbr_enter_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line); static void bbr_stop_all_timers(struct tcpcb *tp); static void bbr_exit_probe_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts); static void bbr_check_probe_rtt_limits(struct tcp_bbr *bbr, uint32_t cts); static void bbr_timer_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts); static void bbr_log_pacing_delay_calc(struct tcp_bbr *bbr, uint16_t gain, uint32_t len, uint32_t cts, uint32_t usecs, uint64_t bw, uint32_t override, int mod); static int bbr_ctloutput(struct inpcb *inp, struct sockopt *sopt); static inline uint8_t bbr_state_val(struct tcp_bbr *bbr) { return(bbr->rc_bbr_substate); } static inline uint32_t get_min_cwnd(struct tcp_bbr *bbr) { int mss; mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs); if (bbr_get_rtt(bbr, BBR_RTT_PROP) < BBR_HIGH_SPEED) return (bbr_cwnd_min_val_hs * mss); else return (bbr_cwnd_min_val * mss); } static uint32_t bbr_get_persists_timer_val(struct tcpcb *tp, struct tcp_bbr *bbr) { uint64_t srtt, var; uint64_t ret_val; bbr->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT; if (tp->t_srtt == 0) { srtt = (uint64_t)BBR_INITIAL_RTO; var = 0; } else { srtt = ((uint64_t)TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT); var = ((uint64_t)TICKS_2_USEC(tp->t_rttvar) >> TCP_RTT_SHIFT); } TCPT_RANGESET_NOSLOP(ret_val, ((srtt + var) * tcp_backoff[tp->t_rxtshift]), bbr_persist_min, bbr_persist_max); return ((uint32_t)ret_val); } static uint32_t bbr_timer_start(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) { /* * Start the FR timer, we do this based on getting the first one in * the rc_tmap. Note that if its NULL we must stop the timer. in all * events we need to stop the running timer (if its running) before * starting the new one. */ uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse; int32_t idx; int32_t is_tlp_timer = 0; struct bbr_sendmap *rsm; if (bbr->rc_all_timers_stopped) { /* All timers have been stopped none are to run */ return (0); } if (bbr->rc_in_persist) { /* We can't start any timer in persists */ return (bbr_get_persists_timer_val(tp, bbr)); } rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); if ((rsm == NULL) || ((tp->t_flags & TF_SACK_PERMIT) == 0) || (tp->t_state < TCPS_ESTABLISHED)) { /* Nothing on the send map */ activate_rxt: if (SEQ_LT(tp->snd_una, tp->snd_max) || sbavail(&(tp->t_inpcb->inp_socket->so_snd))) { uint64_t tov; time_since_sent = 0; rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); if (rsm) { idx = rsm->r_rtr_cnt - 1; if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], bbr->r_ctl.rc_tlp_rxt_last_time)) tstmp_touse = rsm->r_tim_lastsent[idx]; else tstmp_touse = bbr->r_ctl.rc_tlp_rxt_last_time; if (TSTMP_GT(tstmp_touse, cts)) time_since_sent = cts - tstmp_touse; } bbr->r_ctl.rc_hpts_flags |= PACE_TMR_RXT; if (tp->t_srtt == 0) tov = BBR_INITIAL_RTO; else tov = ((uint64_t)(TICKS_2_USEC(tp->t_srtt) + ((uint64_t)TICKS_2_USEC(tp->t_rttvar) * (uint64_t)4)) >> TCP_RTT_SHIFT); if (tp->t_rxtshift) tov *= tcp_backoff[tp->t_rxtshift]; if (tov > time_since_sent) tov -= time_since_sent; else tov = bbr->r_ctl.rc_min_to; TCPT_RANGESET_NOSLOP(to, tov, (bbr->r_ctl.rc_min_rto_ms * MS_IN_USEC), (bbr->rc_max_rto_sec * USECS_IN_SECOND)); bbr_log_timer_var(bbr, 2, cts, 0, srtt, 0, to); return (to); } return (0); } if (rsm->r_flags & BBR_ACKED) { rsm = bbr_find_lowest_rsm(bbr); if (rsm == NULL) { /* No lowest? */ goto activate_rxt; } } /* Convert from ms to usecs */ if (rsm->r_flags & BBR_SACK_PASSED) { if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) == 1) && (rsm->r_flags & BBR_HAS_FIN)) { /* * We don't start a bbr rack timer if all we have is * a FIN outstanding. */ goto activate_rxt; } srtt = bbr_get_rtt(bbr, BBR_RTT_RACK); thresh = bbr_calc_thresh_rack(bbr, srtt, cts, rsm); idx = rsm->r_rtr_cnt - 1; exp = rsm->r_tim_lastsent[idx] + thresh; if (SEQ_GEQ(exp, cts)) { to = exp - cts; if (to < bbr->r_ctl.rc_min_to) { to = bbr->r_ctl.rc_min_to; } } else { to = bbr->r_ctl.rc_min_to; } } else { /* Ok we need to do a TLP not RACK */ if (bbr->rc_tlp_in_progress != 0) { /* * The previous send was a TLP. */ goto activate_rxt; } rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_tmap, bbr_sendmap, r_tnext); if (rsm == NULL) { /* We found no rsm to TLP with. */ goto activate_rxt; } if (rsm->r_flags & BBR_HAS_FIN) { /* If its a FIN we don't do TLP */ rsm = NULL; goto activate_rxt; } time_since_sent = 0; idx = rsm->r_rtr_cnt - 1; if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], bbr->r_ctl.rc_tlp_rxt_last_time)) tstmp_touse = rsm->r_tim_lastsent[idx]; else tstmp_touse = bbr->r_ctl.rc_tlp_rxt_last_time; if (TSTMP_GT(tstmp_touse, cts)) time_since_sent = cts - tstmp_touse; is_tlp_timer = 1; srtt = bbr_get_rtt(bbr, bbr_tlp_type_to_use); thresh = bbr_calc_thresh_tlp(tp, bbr, rsm, srtt, cts); if (thresh > time_since_sent) to = thresh - time_since_sent; else to = bbr->r_ctl.rc_min_to; if (to > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) { /* * If the TLP time works out to larger than the max * RTO lets not do TLP.. just RTO. */ goto activate_rxt; } if ((bbr->rc_tlp_rtx_out == 1) && (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq)) { /* * Second retransmit of the same TLP * lets not. */ bbr->rc_tlp_rtx_out = 0; goto activate_rxt; } if (rsm->r_start != bbr->r_ctl.rc_last_tlp_seq) { /* * The tail is no longer the last one I did a probe * on */ bbr->r_ctl.rc_tlp_seg_send_cnt = 0; bbr->r_ctl.rc_last_tlp_seq = rsm->r_start; } } if (is_tlp_timer == 0) { BBR_STAT_INC(bbr_to_arm_rack); bbr->r_ctl.rc_hpts_flags |= PACE_TMR_RACK; } else { bbr_log_timer_var(bbr, 1, cts, time_since_sent, srtt, thresh, to); if (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend) { /* * We have exceeded how many times we can retran the * current TLP timer, switch to the RTO timer. */ goto activate_rxt; } else { BBR_STAT_INC(bbr_to_arm_tlp); bbr->r_ctl.rc_hpts_flags |= PACE_TMR_TLP; } } return (to); } static inline int32_t bbr_minseg(struct tcp_bbr *bbr) { return (bbr->r_ctl.rc_pace_min_segs - bbr->rc_last_options); } static void bbr_start_hpts_timer(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t cts, int32_t frm, int32_t slot, uint32_t tot_len) { struct inpcb *inp; struct hpts_diag diag; uint32_t delayed_ack = 0; uint32_t left = 0; uint32_t hpts_timeout; uint8_t stopped; int32_t delay_calc = 0; uint32_t prev_delay = 0; inp = tp->t_inpcb; if (tcp_in_hpts(inp)) { /* A previous call is already set up */ return; } if ((tp->t_state == TCPS_CLOSED) || (tp->t_state == TCPS_LISTEN)) { return; } stopped = bbr->rc_tmr_stopped; if (stopped && TSTMP_GT(bbr->r_ctl.rc_timer_exp, cts)) { left = bbr->r_ctl.rc_timer_exp - cts; } bbr->r_ctl.rc_hpts_flags = 0; bbr->r_ctl.rc_timer_exp = 0; prev_delay = bbr->r_ctl.rc_last_delay_val; if (bbr->r_ctl.rc_last_delay_val && (slot == 0)) { /* * If a previous pacer delay was in place we * are not coming from the output side (where * we calculate a delay, more likely a timer). */ slot = bbr->r_ctl.rc_last_delay_val; if (TSTMP_GT(cts, bbr->rc_pacer_started)) { /* Compensate for time passed */ delay_calc = cts - bbr->rc_pacer_started; if (delay_calc <= slot) slot -= delay_calc; } } /* Do we have early to make up for by pushing out the pacing time? */ if (bbr->r_agg_early_set) { bbr_log_pacing_delay_calc(bbr, 0, bbr->r_ctl.rc_agg_early, cts, slot, 0, bbr->r_agg_early_set, 2); slot += bbr->r_ctl.rc_agg_early; bbr->r_ctl.rc_agg_early = 0; bbr->r_agg_early_set = 0; } /* Are we running a total debt that needs to be compensated for? */ if (bbr->r_ctl.rc_hptsi_agg_delay) { if (slot > bbr->r_ctl.rc_hptsi_agg_delay) { /* We nuke the delay */ slot -= bbr->r_ctl.rc_hptsi_agg_delay; bbr->r_ctl.rc_hptsi_agg_delay = 0; } else { /* We nuke some of the delay, put in a minimal 100usecs */ bbr->r_ctl.rc_hptsi_agg_delay -= slot; bbr->r_ctl.rc_last_delay_val = slot = 100; } } bbr->r_ctl.rc_last_delay_val = slot; hpts_timeout = bbr_timer_start(tp, bbr, cts); if (tp->t_flags & TF_DELACK) { if (bbr->rc_in_persist == 0) { delayed_ack = bbr_delack_time; } else { /* * We are in persists and have * gotten a new data element. */ if (hpts_timeout > bbr_delack_time) { /* * Lets make the persists timer (which acks) * be the smaller of hpts_timeout and bbr_delack_time. */ hpts_timeout = bbr_delack_time; } } } if (delayed_ack && ((hpts_timeout == 0) || (delayed_ack < hpts_timeout))) { /* We need a Delayed ack timer */ bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK; hpts_timeout = delayed_ack; } if (slot) { /* Mark that we have a pacing timer up */ BBR_STAT_INC(bbr_paced_segments); bbr->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT; } /* * If no timers are going to run and we will fall off thfe hptsi * wheel, we resort to a keep-alive timer if its configured. */ if ((hpts_timeout == 0) && (slot == 0)) { if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && (tp->t_state <= TCPS_CLOSING)) { /* * Ok we have no timer (persists, rack, tlp, rxt or * del-ack), we don't have segments being paced. So * all that is left is the keepalive timer. */ if (TCPS_HAVEESTABLISHED(tp->t_state)) { hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp)); } else { hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp)); } bbr->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP; } } if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) == (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) { /* * RACK, TLP, persists and RXT timers all are restartable * based on actions input .. i.e we received a packet (ack * or sack) and that changes things (rw, or snd_una etc). * Thus we can restart them with a new value. For * keep-alive, delayed_ack we keep track of what was left * and restart the timer with a smaller value. */ if (left < hpts_timeout) hpts_timeout = left; } if (bbr->r_ctl.rc_incr_tmrs && slot && (bbr->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) { /* * If configured to do so, and the timer is either * the TLP or RXT timer, we need to increase the timeout * by the pacing time. Consider the bottleneck at my * machine as an example, we are sending something * to start a TLP on. The last packet won't be emitted * fully until the pacing time (the bottleneck will hold * the data in place). Once the packet is emitted that * is when we want to start waiting for the TLP. This * is most evident with hardware pacing (where the nic * is holding the packet(s) before emitting). But it * can also show up in the network so we do it for all * cases. Technically we would take off one packet from * this extra delay but this is easier and being more * conservative is probably better. */ hpts_timeout += slot; } if (hpts_timeout) { /* * Hack alert for now we can't time-out over 2147 seconds (a * bit more than 35min) */ if (hpts_timeout > 0x7ffffffe) hpts_timeout = 0x7ffffffe; bbr->r_ctl.rc_timer_exp = cts + hpts_timeout; } else bbr->r_ctl.rc_timer_exp = 0; if ((slot) && (bbr->rc_use_google || bbr->output_error_seen || (slot <= hpts_timeout)) ) { /* * Tell LRO that it can queue packets while * we pace. */ bbr->rc_inp->inp_flags2 |= INP_MBUF_QUEUE_READY; if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) && (bbr->rc_cwnd_limited == 0)) { /* * If we are not cwnd limited and we * are running a rack timer we put on * the do not disturbe even for sack. */ inp->inp_flags2 |= INP_DONT_SACK_QUEUE; } else inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE; bbr->rc_pacer_started = cts; (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(slot), __LINE__, &diag); bbr->rc_timer_first = 0; bbr->bbr_timer_src = frm; bbr_log_to_start(bbr, cts, hpts_timeout, slot, 1); bbr_log_hpts_diag(bbr, cts, &diag); } else if (hpts_timeout) { (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout), __LINE__, &diag); /* * We add the flag here as well if the slot is set, * since hpts will call in to clear the queue first before * calling the output routine (which does our timers). * We don't want to set the flag if its just a timer * else the arrival of data might (that causes us * to send more) might get delayed. Imagine being * on a keep-alive timer and a request comes in for * more data. */ if (slot) bbr->rc_pacer_started = cts; if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) && (bbr->rc_cwnd_limited == 0)) { /* * For a rack timer, don't wake us even * if a sack arrives as long as we are * not cwnd limited. */ bbr->rc_inp->inp_flags2 |= INP_MBUF_QUEUE_READY; inp->inp_flags2 |= INP_DONT_SACK_QUEUE; } else { /* All other timers wake us up */ bbr->rc_inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY; inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE; } bbr->bbr_timer_src = frm; bbr_log_to_start(bbr, cts, hpts_timeout, slot, 0); bbr_log_hpts_diag(bbr, cts, &diag); bbr->rc_timer_first = 1; } bbr->rc_tmr_stopped = 0; bbr_log_type_bbrsnd(bbr, tot_len, slot, delay_calc, cts, frm, prev_delay); } static void bbr_timer_audit(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, struct sockbuf *sb) { /* * We received an ack, and then did not call send or were bounced * out due to the hpts was running. Now a timer is up as well, is it * the right timer? */ struct inpcb *inp; struct bbr_sendmap *rsm; uint32_t hpts_timeout; int tmr_up; tmr_up = bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK; if (bbr->rc_in_persist && (tmr_up == PACE_TMR_PERSIT)) return; rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) && (tmr_up == PACE_TMR_RXT)) { /* Should be an RXT */ return; } inp = bbr->rc_inp; if (rsm == NULL) { /* Nothing outstanding? */ if (tp->t_flags & TF_DELACK) { if (tmr_up == PACE_TMR_DELACK) /* * We are supposed to have delayed ack up * and we do */ return; } else if (sbavail(&inp->inp_socket->so_snd) && (tmr_up == PACE_TMR_RXT)) { /* * if we hit enobufs then we would expect the * possibility of nothing outstanding and the RXT up * (and the hptsi timer). */ return; } else if (((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && (tp->t_state <= TCPS_CLOSING)) && (tmr_up == PACE_TMR_KEEP) && (tp->snd_max == tp->snd_una)) { /* We should have keep alive up and we do */ return; } } if (rsm && (rsm->r_flags & BBR_SACK_PASSED)) { if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) == 1) && (rsm->r_flags & BBR_HAS_FIN)) { /* needs to be a RXT */ if (tmr_up == PACE_TMR_RXT) return; else goto wrong_timer; } else if (tmr_up == PACE_TMR_RACK) return; else goto wrong_timer; } else if (rsm && (tmr_up == PACE_TMR_RACK)) { /* Rack timer has priority if we have data out */ return; } else if (SEQ_GT(tp->snd_max, tp->snd_una) && ((tmr_up == PACE_TMR_TLP) || (tmr_up == PACE_TMR_RXT))) { /* * Either a TLP or RXT is fine if no sack-passed is in place * and data is outstanding. */ return; } else if (tmr_up == PACE_TMR_DELACK) { /* * If the delayed ack was going to go off before the * rtx/tlp/rack timer were going to expire, then that would * be the timer in control. Note we don't check the time * here trusting the code is correct. */ return; } if (SEQ_GT(tp->snd_max, tp->snd_una) && ((tmr_up == PACE_TMR_RXT) || (tmr_up == PACE_TMR_TLP) || (tmr_up == PACE_TMR_RACK))) { /* * We have outstanding data and * we *do* have a RACK, TLP or RXT * timer running. We won't restart * anything here since thats probably ok we * will get called with some timer here shortly. */ return; } /* * Ok the timer originally started is not what we want now. We will * force the hpts to be stopped if any, and restart with the slot * set to what was in the saved slot. */ wrong_timer: if ((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) { if (tcp_in_hpts(inp)) tcp_hpts_remove(inp); bbr_timer_cancel(bbr, __LINE__, cts); bbr_start_hpts_timer(bbr, tp, cts, 1, bbr->r_ctl.rc_last_delay_val, 0); } else { /* * Output is hptsi so we just need to switch the type of * timer. We don't bother with keep-alive, since when we * jump through the output, it will start the keep-alive if * nothing is sent. * * We only need a delayed-ack added and or the hpts_timeout. */ hpts_timeout = bbr_timer_start(tp, bbr, cts); if (tp->t_flags & TF_DELACK) { if (hpts_timeout == 0) { hpts_timeout = bbr_delack_time; bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK; } else if (hpts_timeout > bbr_delack_time) { hpts_timeout = bbr_delack_time; bbr->r_ctl.rc_hpts_flags = PACE_TMR_DELACK; } } if (hpts_timeout) { if (hpts_timeout > 0x7ffffffe) hpts_timeout = 0x7ffffffe; bbr->r_ctl.rc_timer_exp = cts + hpts_timeout; } } } int32_t bbr_clear_lost = 0; /* * Considers the two time values now (cts) and earlier. * If cts is smaller than earlier, we could have * had a sequence wrap (our counter wraps every * 70 min or so) or it could be just clock skew * getting us two different time values. Clock skew * will show up within 10ms or so. So in such * a case (where cts is behind earlier time by * less than 10ms) we return 0. Otherwise we * return the true difference between them. */ static inline uint32_t bbr_calc_time(uint32_t cts, uint32_t earlier_time) { /* * Given two timestamps, the current time stamp cts, and some other * time-stamp taken in theory earlier return the difference. The * trick is here sometimes locking will get the other timestamp * after the cts. If this occurs we need to return 0. */ if (TSTMP_GEQ(cts, earlier_time)) return (cts - earlier_time); /* * cts is behind earlier_time if its less than 10ms consider it 0. * If its more than 10ms difference then we had a time wrap. Else * its just the normal locking foo. I wonder if we should not go to * 64bit TS and get rid of this issue. */ if (TSTMP_GEQ((cts + 10000), earlier_time)) return (0); /* * Ok the time must have wrapped. So we need to answer a large * amount of time, which the normal subtraction should do. */ return (cts - earlier_time); } static int sysctl_bbr_clear_lost(SYSCTL_HANDLER_ARGS) { uint32_t stat; int32_t error; error = SYSCTL_OUT(req, &bbr_clear_lost, sizeof(uint32_t)); if (error || req->newptr == NULL) return error; error = SYSCTL_IN(req, &stat, sizeof(uint32_t)); if (error) return (error); if (stat == 1) { #ifdef BBR_INVARIANTS printf("Clearing BBR lost counters\n"); #endif COUNTER_ARRAY_ZERO(bbr_state_lost, BBR_MAX_STAT); COUNTER_ARRAY_ZERO(bbr_state_time, BBR_MAX_STAT); COUNTER_ARRAY_ZERO(bbr_state_resend, BBR_MAX_STAT); } else if (stat == 2) { #ifdef BBR_INVARIANTS printf("Clearing BBR option counters\n"); #endif COUNTER_ARRAY_ZERO(bbr_opts_arry, BBR_OPTS_SIZE); } else if (stat == 3) { #ifdef BBR_INVARIANTS printf("Clearing BBR stats counters\n"); #endif COUNTER_ARRAY_ZERO(bbr_stat_arry, BBR_STAT_SIZE); } else if (stat == 4) { #ifdef BBR_INVARIANTS printf("Clearing BBR out-size counters\n"); #endif COUNTER_ARRAY_ZERO(bbr_out_size, TCP_MSS_ACCT_SIZE); } bbr_clear_lost = 0; return (0); } static void bbr_init_sysctls(void) { struct sysctl_oid *bbr_probertt; struct sysctl_oid *bbr_hptsi; struct sysctl_oid *bbr_measure; struct sysctl_oid *bbr_cwnd; struct sysctl_oid *bbr_timeout; struct sysctl_oid *bbr_states; struct sysctl_oid *bbr_startup; struct sysctl_oid *bbr_policer; /* Probe rtt controls */ bbr_probertt = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "probertt", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, ""); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_probertt), OID_AUTO, "gain", CTLFLAG_RW, &bbr_rttprobe_gain, 192, "What is the filter gain drop in probe_rtt (0=disable)?"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_probertt), OID_AUTO, "cwnd", CTLFLAG_RW, &bbr_rtt_probe_cwndtarg, 4, "How many mss's are outstanding during probe-rtt"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_probertt), OID_AUTO, "int", CTLFLAG_RW, &bbr_rtt_probe_limit, 4000000, "If RTT has not shrank in this many micro-seconds enter probe-rtt"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_probertt), OID_AUTO, "mintime", CTLFLAG_RW, &bbr_rtt_probe_time, 200000, "How many microseconds in probe-rtt"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_probertt), OID_AUTO, "filter_len_sec", CTLFLAG_RW, &bbr_filter_len_sec, 6, "How long in seconds does the rttProp filter run?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_probertt), OID_AUTO, "drain_rtt", CTLFLAG_RW, &bbr_drain_rtt, BBR_SRTT, "What is the drain rtt to use in probeRTT (rtt_prop=0, rtt_rack=1, rtt_pkt=2, rtt_srtt=3?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_probertt), OID_AUTO, "can_force", CTLFLAG_RW, &bbr_can_force_probertt, 0, "If we keep setting new low rtt's but delay going in probe-rtt can we force in??"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_probertt), OID_AUTO, "enter_sets_force", CTLFLAG_RW, &bbr_probertt_sets_rtt, 0, "In NF mode, do we imitate google_mode and set the rttProp on entry to probe-rtt?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_probertt), OID_AUTO, "can_adjust", CTLFLAG_RW, &bbr_can_adjust_probertt, 1, "Can we dynamically adjust the probe-rtt limits and times?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_probertt), OID_AUTO, "is_ratio", CTLFLAG_RW, &bbr_is_ratio, 0, "is the limit to filter a ratio?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_probertt), OID_AUTO, "use_cwnd", CTLFLAG_RW, &bbr_prtt_slam_cwnd, 0, "Should we set/recover cwnd?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_probertt), OID_AUTO, "can_use_ts", CTLFLAG_RW, &bbr_can_use_ts_for_rtt, 1, "Can we use the ms timestamp if available for retransmistted rtt calculations?"); /* Pacing controls */ bbr_hptsi = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "pacing", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, ""); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "hw_pacing", CTLFLAG_RW, &bbr_allow_hdwr_pacing, 1, "Do we allow hardware pacing?"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "hw_pacing_limit", CTLFLAG_RW, &bbr_hardware_pacing_limit, 4000, "Do we have a limited number of connections for pacing chelsio (0=no limit)?"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "hw_pacing_adj", CTLFLAG_RW, &bbr_hdwr_pace_adjust, 2, "Multiplier to calculated tso size?"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "hw_pacing_floor", CTLFLAG_RW, &bbr_hdwr_pace_floor, 1, "Do we invoke the hardware pacing floor?"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "hw_pacing_delay_cnt", CTLFLAG_RW, &bbr_hdwr_pacing_delay_cnt, 10, "How many packets must be sent after hdwr pacing is enabled"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "bw_cross", CTLFLAG_RW, &bbr_cross_over, 3000000, "What is the point where we cross over to linux like TSO size set"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "seg_deltarg", CTLFLAG_RW, &bbr_hptsi_segments_delay_tar, 7000, "What is the worse case delay target for hptsi < 48Mbp connections"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "enet_oh", CTLFLAG_RW, &bbr_include_enet_oh, 0, "Do we include the ethernet overhead in calculating pacing delay?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "ip_oh", CTLFLAG_RW, &bbr_include_ip_oh, 1, "Do we include the IP overhead in calculating pacing delay?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "tcp_oh", CTLFLAG_RW, &bbr_include_tcp_oh, 0, "Do we include the TCP overhead in calculating pacing delay?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "google_discount", CTLFLAG_RW, &bbr_google_discount, 10, "What is the default google discount percentage wise for pacing (11 = 1.1%%)?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "all_get_min", CTLFLAG_RW, &bbr_all_get_min, 0, "If you are less than a MSS do you just get the min?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "tso_min", CTLFLAG_RW, &bbr_hptsi_bytes_min, 1460, "For 0 -> 24Mbps what is floor number of segments for TSO"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "seg_tso_max", CTLFLAG_RW, &bbr_hptsi_segments_max, 6, "For 0 -> 24Mbps what is top number of segments for TSO"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "seg_floor", CTLFLAG_RW, &bbr_hptsi_segments_floor, 1, "Minimum TSO size we will fall too in segments"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "utter_max", CTLFLAG_RW, &bbr_hptsi_utter_max, 0, "The absolute maximum that any pacing (outside of hardware) can be"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "seg_divisor", CTLFLAG_RW, &bbr_hptsi_per_second, 100, "What is the divisor in our hptsi TSO calculation 512Mbps < X > 24Mbps "); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "srtt_mul", CTLFLAG_RW, &bbr_hptsi_max_mul, 1, "The multiplier for pace len max"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_hptsi), OID_AUTO, "srtt_div", CTLFLAG_RW, &bbr_hptsi_max_div, 2, "The divisor for pace len max"); /* Measurement controls */ bbr_measure = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "measure", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Measurement controls"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_measure), OID_AUTO, "min_i_bw", CTLFLAG_RW, &bbr_initial_bw_bps, 62500, "Minimum initial b/w in bytes per second"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_measure), OID_AUTO, "no_sack_needed", CTLFLAG_RW, &bbr_sack_not_required, 0, "Do we allow bbr to run on connections not supporting SACK?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_measure), OID_AUTO, "use_google", CTLFLAG_RW, &bbr_use_google_algo, 0, "Use has close to google V1.0 has possible?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_measure), OID_AUTO, "ts_limiting", CTLFLAG_RW, &bbr_ts_limiting, 1, "Do we attempt to use the peers timestamp to limit b/w caculations?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_measure), OID_AUTO, "ts_can_raise", CTLFLAG_RW, &bbr_ts_can_raise, 0, "Can we raise the b/w via timestamp b/w calculation?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_measure), OID_AUTO, "ts_delta", CTLFLAG_RW, &bbr_min_usec_delta, 20000, "How long in usec between ts of our sends in ts validation code?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_measure), OID_AUTO, "ts_peer_delta", CTLFLAG_RW, &bbr_min_peer_delta, 20, "What min numerical value should be between the peer deltas?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_measure), OID_AUTO, "ts_delta_percent", CTLFLAG_RW, &bbr_delta_percent, 150, "What percentage (150 = 15.0) do we allow variance for?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_measure), OID_AUTO, "min_measure_good_bw", CTLFLAG_RW, &bbr_min_measurements_req, 1, "What is the minimum measurement count we need before we switch to our b/w estimate"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_measure), OID_AUTO, "min_measure_before_pace", CTLFLAG_RW, &bbr_no_pacing_until, 4, "How many pkt-epoch's (0 is off) do we need before pacing is on?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_measure), OID_AUTO, "quanta", CTLFLAG_RW, &bbr_quanta, 2, "Extra quanta to add when calculating the target (ID section 4.2.3.2)."); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_measure), OID_AUTO, "noretran", CTLFLAG_RW, &bbr_no_retran, 0, "Should google mode not use retransmission measurements for the b/w estimation?"); /* State controls */ bbr_states = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "states", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "State controls"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "idle_restart", CTLFLAG_RW, &bbr_uses_idle_restart, 0, "Do we use a new special idle_restart state to ramp back up quickly?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "idle_restart_threshold", CTLFLAG_RW, &bbr_idle_restart_threshold, 100000, "How long must we be idle before we restart??"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "use_pkt_epoch", CTLFLAG_RW, &bbr_state_is_pkt_epoch, 0, "Do we use a pkt-epoch for substate if 0 rttProp?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "startup_rtt_gain", CTLFLAG_RW, &bbr_rtt_gain_thresh, 0, "What increase in RTT triggers us to stop ignoring no-loss and possibly exit startup?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "drain_floor", CTLFLAG_RW, &bbr_drain_floor, 88, "What is the lowest we can drain (pg) too?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "drain_2_target", CTLFLAG_RW, &bbr_state_drain_2_tar, 1, "Do we drain to target in drain substate?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "gain_2_target", CTLFLAG_RW, &bbr_gain_to_target, 1, "Does probe bw gain to target??"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "gain_extra_time", CTLFLAG_RW, &bbr_gain_gets_extra_too, 1, "Does probe bw gain get the extra time too?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "ld_div", CTLFLAG_RW, &bbr_drain_drop_div, 5, "Long drain drop divider?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "ld_mul", CTLFLAG_RW, &bbr_drain_drop_mul, 4, "Long drain drop multiplier?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "rand_ot_disc", CTLFLAG_RW, &bbr_rand_ot, 50, "Random discount of the ot?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "dr_filter_life", CTLFLAG_RW, &bbr_num_pktepo_for_del_limit, BBR_NUM_RTTS_FOR_DEL_LIMIT, "How many packet-epochs does the b/w delivery rate last?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "subdrain_applimited", CTLFLAG_RW, &bbr_sub_drain_app_limit, 0, "Does our sub-state drain invoke app limited if its long?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "use_cwnd_subdrain", CTLFLAG_RW, &bbr_sub_drain_slam_cwnd, 0, "Should we set/recover cwnd for sub-state drain?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "use_cwnd_maindrain", CTLFLAG_RW, &bbr_slam_cwnd_in_main_drain, 0, "Should we set/recover cwnd for main-state drain?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "google_gets_earlyout", CTLFLAG_RW, &google_allow_early_out, 1, "Should we allow google probe-bw/drain to exit early at flight target?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_states), OID_AUTO, "google_exit_loss", CTLFLAG_RW, &google_consider_lost, 1, "Should we have losses exit gain of probebw in google mode??"); /* Startup controls */ bbr_startup = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "startup", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Startup controls"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_startup), OID_AUTO, "cheat_iwnd", CTLFLAG_RW, &bbr_sends_full_iwnd, 1, "Do we not pace but burst out initial windows has our TSO size?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_startup), OID_AUTO, "loss_threshold", CTLFLAG_RW, &bbr_startup_loss_thresh, 2000, "In startup what is the loss threshold in a pe that will exit us from startup?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_startup), OID_AUTO, "use_lowerpg", CTLFLAG_RW, &bbr_use_lower_gain_in_startup, 1, "Should we use a lower hptsi gain if we see loss in startup?"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_startup), OID_AUTO, "gain", CTLFLAG_RW, &bbr_start_exit, 25, "What gain percent do we need to see to stay in startup??"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_startup), OID_AUTO, "low_gain", CTLFLAG_RW, &bbr_low_start_exit, 15, "What gain percent do we need to see to stay in the lower gain startup??"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_startup), OID_AUTO, "loss_exit", CTLFLAG_RW, &bbr_exit_startup_at_loss, 1, "Should we exit startup at loss in an epoch if we are not gaining?"); /* CWND controls */ bbr_cwnd = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "cwnd", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Cwnd controls"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "tar_rtt", CTLFLAG_RW, &bbr_cwndtarget_rtt_touse, 0, "Target cwnd rtt measurement to use (0=rtt_prop, 1=rtt_rack, 2=pkt_rtt, 3=srtt)?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "may_shrink", CTLFLAG_RW, &bbr_cwnd_may_shrink, 0, "Can the cwnd shrink if it would grow to more than the target?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "max_target_limit", CTLFLAG_RW, &bbr_target_cwnd_mult_limit, 8, "Do we limit the cwnd to some multiple of the cwnd target if cwnd can't shrink 0=no?"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "highspeed_min", CTLFLAG_RW, &bbr_cwnd_min_val_hs, BBR_HIGHSPEED_NUM_MSS, "What is the high-speed min cwnd (rttProp under 1ms)"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "lowspeed_min", CTLFLAG_RW, &bbr_cwnd_min_val, BBR_PROBERTT_NUM_MSS, "What is the min cwnd (rttProp > 1ms)"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "initwin", CTLFLAG_RW, &bbr_def_init_win, 10, "What is the BBR initial window, if 0 use tcp version"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "do_loss_red", CTLFLAG_RW, &bbr_do_red, 600, "Do we reduce the b/w at exit from recovery based on ratio of prop/srtt (800=80.0, 0=off)?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "red_scale", CTLFLAG_RW, &bbr_red_scale, 20000, "What RTT do we scale with?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "red_growslow", CTLFLAG_RW, &bbr_red_growth_restrict, 1, "Do we restrict cwnd growth for whats in flight?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "red_div", CTLFLAG_RW, &bbr_red_div, 2, "If we reduce whats the divisor?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "red_mul", CTLFLAG_RW, &bbr_red_mul, 1, "If we reduce whats the mulitiplier?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "target_is_unit", CTLFLAG_RW, &bbr_target_is_bbunit, 0, "Is the state target the pacing_gain or BBR_UNIT?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_cwnd), OID_AUTO, "drop_limit", CTLFLAG_RW, &bbr_drop_limit, 0, "Number of segments limit for drop (0=use min_cwnd w/flight)?"); /* Timeout controls */ bbr_timeout = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "timeout", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Time out controls"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_timeout), OID_AUTO, "delack", CTLFLAG_RW, &bbr_delack_time, 100000, "BBR's delayed ack time"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_timeout), OID_AUTO, "tlp_uses", CTLFLAG_RW, &bbr_tlp_type_to_use, 3, "RTT that TLP uses in its calculations, 0=rttProp, 1=Rack_rtt, 2=pkt_rtt and 3=srtt"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_timeout), OID_AUTO, "persmin", CTLFLAG_RW, &bbr_persist_min, 250000, "What is the minimum time in microseconds between persists"); SYSCTL_ADD_U32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_timeout), OID_AUTO, "persmax", CTLFLAG_RW, &bbr_persist_max, 1000000, "What is the largest delay in microseconds between persists"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_timeout), OID_AUTO, "tlp_minto", CTLFLAG_RW, &bbr_tlp_min, 10000, "TLP Min timeout in usecs"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_timeout), OID_AUTO, "tlp_dack_time", CTLFLAG_RW, &bbr_delayed_ack_time, 200000, "TLP delayed ack compensation value"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "minrto", CTLFLAG_RW, &bbr_rto_min_ms, 30, "Minimum RTO in ms"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_timeout), OID_AUTO, "maxrto", CTLFLAG_RW, &bbr_rto_max_sec, 4, "Maximum RTO in seconds -- should be at least as large as min_rto"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_timeout), OID_AUTO, "tlp_retry", CTLFLAG_RW, &bbr_tlp_max_resend, 2, "How many times does TLP retry a single segment or multiple with no ACK"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_timeout), OID_AUTO, "minto", CTLFLAG_RW, &bbr_min_to, 1000, "Minimum rack timeout in useconds"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_timeout), OID_AUTO, "pktdelay", CTLFLAG_RW, &bbr_pkt_delay, 1000, "Extra RACK time (in useconds) besides reordering thresh"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_timeout), OID_AUTO, "incr_tmrs", CTLFLAG_RW, &bbr_incr_timers, 1, "Increase the RXT/TLP timer by the pacing time used?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_timeout), OID_AUTO, "rxtmark_sackpassed", CTLFLAG_RW, &bbr_marks_rxt_sack_passed, 0, "Mark sack passed on all those not ack'd when a RXT hits?"); /* Policer controls */ bbr_policer = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "policer", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Policer controls"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_policer), OID_AUTO, "detect_enable", CTLFLAG_RW, &bbr_policer_detection_enabled, 1, "Is policer detection enabled??"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_policer), OID_AUTO, "min_pes", CTLFLAG_RW, &bbr_lt_intvl_min_rtts, 4, "Minimum number of PE's?"); SYSCTL_ADD_U64(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_policer), OID_AUTO, "bwdiff", CTLFLAG_RW, &bbr_lt_bw_diff, (4000/8), "Minimal bw diff?"); SYSCTL_ADD_U64(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_policer), OID_AUTO, "bwratio", CTLFLAG_RW, &bbr_lt_bw_ratio, 8, "Minimal bw diff?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_policer), OID_AUTO, "from_rack_rxt", CTLFLAG_RW, &bbr_policer_call_from_rack_to, 0, "Do we call the policer detection code from a rack-timeout?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_policer), OID_AUTO, "false_postive", CTLFLAG_RW, &bbr_lt_intvl_fp, 0, "What packet epoch do we do false-positive detection at (0=no)?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_policer), OID_AUTO, "loss_thresh", CTLFLAG_RW, &bbr_lt_loss_thresh, 196, "Loss threshold 196 = 19.6%?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_policer), OID_AUTO, "false_postive_thresh", CTLFLAG_RW, &bbr_lt_fd_thresh, 100, "What percentage is the false detection threshold (150=15.0)?"); /* All the rest */ SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "cheat_rxt", CTLFLAG_RW, &bbr_use_rack_resend_cheat, 0, "Do we burst 1ms between sends on retransmissions (like rack)?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "error_paceout", CTLFLAG_RW, &bbr_error_base_paceout, 10000, "When we hit an error what is the min to pace out in usec's?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "kill_paceout", CTLFLAG_RW, &bbr_max_net_error_cnt, 10, "When we hit this many errors in a row, kill the session?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "data_after_close", CTLFLAG_RW, &bbr_ignore_data_after_close, 1, "Do we hold off sending a RST until all pending data is ack'd"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "resend_use_tso", CTLFLAG_RW, &bbr_resends_use_tso, 0, "Can resends use TSO?"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "sblklimit", CTLFLAG_RW, &bbr_sack_block_limit, 128, "When do we start ignoring small sack blocks"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "bb_verbose", CTLFLAG_RW, &bbr_verbose_logging, 0, "Should BBR black box logging be verbose"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "reorder_thresh", CTLFLAG_RW, &bbr_reorder_thresh, 2, "What factor for rack will be added when seeing reordering (shift right)"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "reorder_fade", CTLFLAG_RW, &bbr_reorder_fade, 0, "Does reorder detection fade, if so how many ms (0 means never)"); SYSCTL_ADD_S32(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW, &bbr_tlp_thresh, 1, "what divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)"); /* Stats and counters */ /* The pacing counters for hdwr/software can't be in the array */ bbr_nohdwr_pacing_enobuf = counter_u64_alloc(M_WAITOK); bbr_hdwr_pacing_enobuf = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "enob_hdwr_pacing", CTLFLAG_RD, &bbr_hdwr_pacing_enobuf, "Total number of enobufs for hardware paced flows"); SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "enob_no_hdwr_pacing", CTLFLAG_RD, &bbr_nohdwr_pacing_enobuf, "Total number of enobufs for non-hardware paced flows"); bbr_flows_whdwr_pacing = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "hdwr_pacing", CTLFLAG_RD, &bbr_flows_whdwr_pacing, "Total number of hardware paced flows"); bbr_flows_nohdwr_pacing = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "software_pacing", CTLFLAG_RD, &bbr_flows_nohdwr_pacing, "Total number of software paced flows"); COUNTER_ARRAY_ALLOC(bbr_stat_arry, BBR_STAT_SIZE, M_WAITOK); SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "stats", CTLFLAG_RD, bbr_stat_arry, BBR_STAT_SIZE, "BBR Stats"); COUNTER_ARRAY_ALLOC(bbr_opts_arry, BBR_OPTS_SIZE, M_WAITOK); SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "opts", CTLFLAG_RD, bbr_opts_arry, BBR_OPTS_SIZE, "BBR Option Stats"); COUNTER_ARRAY_ALLOC(bbr_state_lost, BBR_MAX_STAT, M_WAITOK); SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "lost", CTLFLAG_RD, bbr_state_lost, BBR_MAX_STAT, "Stats of when losses occur"); COUNTER_ARRAY_ALLOC(bbr_state_resend, BBR_MAX_STAT, M_WAITOK); SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "stateresend", CTLFLAG_RD, bbr_state_resend, BBR_MAX_STAT, "Stats of what states resend"); COUNTER_ARRAY_ALLOC(bbr_state_time, BBR_MAX_STAT, M_WAITOK); SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "statetime", CTLFLAG_RD, bbr_state_time, BBR_MAX_STAT, "Stats of time spent in the states"); COUNTER_ARRAY_ALLOC(bbr_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK); SYSCTL_ADD_COUNTER_U64_ARRAY(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "outsize", CTLFLAG_RD, bbr_out_size, TCP_MSS_ACCT_SIZE, "Size of output calls"); SYSCTL_ADD_PROC(&bbr_sysctl_ctx, SYSCTL_CHILDREN(bbr_sysctl_root), OID_AUTO, "clrlost", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE, &bbr_clear_lost, 0, sysctl_bbr_clear_lost, "IU", "Clear lost counters"); } static void bbr_counter_destroy(void) { COUNTER_ARRAY_FREE(bbr_stat_arry, BBR_STAT_SIZE); COUNTER_ARRAY_FREE(bbr_opts_arry, BBR_OPTS_SIZE); COUNTER_ARRAY_FREE(bbr_out_size, TCP_MSS_ACCT_SIZE); COUNTER_ARRAY_FREE(bbr_state_lost, BBR_MAX_STAT); COUNTER_ARRAY_FREE(bbr_state_time, BBR_MAX_STAT); COUNTER_ARRAY_FREE(bbr_state_resend, BBR_MAX_STAT); counter_u64_free(bbr_nohdwr_pacing_enobuf); counter_u64_free(bbr_hdwr_pacing_enobuf); counter_u64_free(bbr_flows_whdwr_pacing); counter_u64_free(bbr_flows_nohdwr_pacing); } static __inline void bbr_fill_in_logging_data(struct tcp_bbr *bbr, struct tcp_log_bbr *l, uint32_t cts) { memset(l, 0, sizeof(union tcp_log_stackspecific)); l->cur_del_rate = bbr->r_ctl.rc_bbr_cur_del_rate; l->delRate = get_filter_value(&bbr->r_ctl.rc_delrate); l->rttProp = get_filter_value_small(&bbr->r_ctl.rc_rttprop); l->bw_inuse = bbr_get_bw(bbr); l->inflight = ctf_flight_size(bbr->rc_tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); l->applimited = bbr->r_ctl.r_app_limited_until; l->delivered = bbr->r_ctl.rc_delivered; l->timeStamp = cts; l->lost = bbr->r_ctl.rc_lost; l->bbr_state = bbr->rc_bbr_state; l->bbr_substate = bbr_state_val(bbr); l->epoch = bbr->r_ctl.rc_rtt_epoch; l->lt_epoch = bbr->r_ctl.rc_lt_epoch; l->pacing_gain = bbr->r_ctl.rc_bbr_hptsi_gain; l->cwnd_gain = bbr->r_ctl.rc_bbr_cwnd_gain; l->inhpts = tcp_in_hpts(bbr->rc_inp); l->use_lt_bw = bbr->rc_lt_use_bw; l->pkts_out = bbr->r_ctl.rc_flight_at_input; l->pkt_epoch = bbr->r_ctl.rc_pkt_epoch; } static void bbr_log_type_bw_reduce(struct tcp_bbr *bbr, int reason) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); log.u_bbr.flex1 = 0; log.u_bbr.flex2 = 0; log.u_bbr.flex5 = 0; log.u_bbr.flex3 = 0; log.u_bbr.flex4 = bbr->r_ctl.rc_pkt_epoch_loss_rate; log.u_bbr.flex7 = reason; log.u_bbr.flex6 = bbr->r_ctl.rc_bbr_enters_probertt; log.u_bbr.flex8 = 0; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_BW_RED_EV, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_type_rwnd_collapse(struct tcp_bbr *bbr, int seq, int mode, uint32_t count) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); log.u_bbr.flex1 = seq; log.u_bbr.flex2 = count; log.u_bbr.flex8 = mode; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_LOWGAIN, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_type_just_return(struct tcp_bbr *bbr, uint32_t cts, uint32_t tlen, uint8_t hpts_calling, uint8_t reason, uint32_t p_maxseg, int len) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = p_maxseg; log.u_bbr.flex2 = bbr->r_ctl.rc_hpts_flags; log.u_bbr.flex3 = bbr->r_ctl.rc_timer_exp; log.u_bbr.flex4 = reason; log.u_bbr.flex5 = bbr->rc_in_persist; log.u_bbr.flex6 = bbr->r_ctl.rc_last_delay_val; log.u_bbr.flex7 = p_maxseg; log.u_bbr.flex8 = bbr->rc_in_persist; log.u_bbr.pkts_out = 0; log.u_bbr.applimited = len; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_JUSTRET, 0, tlen, &log, false, &bbr->rc_tv); } } static void bbr_log_type_enter_rec(struct tcp_bbr *bbr, uint32_t seq) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); log.u_bbr.flex1 = seq; log.u_bbr.flex2 = bbr->r_ctl.rc_cwnd_on_ent; log.u_bbr.flex3 = bbr->r_ctl.rc_recovery_start; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_ENTREC, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_msgsize_fail(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t len, uint32_t maxseg, uint32_t mtu, int32_t csum_flags, int32_t tso, uint32_t cts) { if (tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = tso; log.u_bbr.flex2 = maxseg; log.u_bbr.flex3 = mtu; log.u_bbr.flex4 = csum_flags; TCP_LOG_EVENTP(tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_MSGSIZE, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_flowend(struct tcp_bbr *bbr) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct sockbuf *r, *s; struct timeval tv; if (bbr->rc_inp->inp_socket) { r = &bbr->rc_inp->inp_socket->so_rcv; s = &bbr->rc_inp->inp_socket->so_snd; } else { r = s = NULL; } bbr_fill_in_logging_data(bbr, &log.u_bbr, tcp_get_usecs(&tv)); TCP_LOG_EVENTP(bbr->rc_tp, NULL, r, s, TCP_LOG_FLOWEND, 0, 0, &log, false, &tv); } } static void bbr_log_pkt_epoch(struct tcp_bbr *bbr, uint32_t cts, uint32_t line, uint32_t lost, uint32_t del) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = lost; log.u_bbr.flex2 = del; log.u_bbr.flex3 = bbr->r_ctl.rc_bbr_lastbtlbw; log.u_bbr.flex4 = bbr->r_ctl.rc_pkt_epoch_rtt; log.u_bbr.flex5 = bbr->r_ctl.rc_bbr_last_startup_epoch; log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup; log.u_bbr.flex7 = line; log.u_bbr.flex8 = 0; log.u_bbr.inflight = bbr->r_ctl.r_measurement_count; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_PKT_EPOCH, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_time_epoch(struct tcp_bbr *bbr, uint32_t cts, uint32_t line, uint32_t epoch_time) { if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = bbr->r_ctl.rc_lost; log.u_bbr.flex2 = bbr->rc_inp->inp_socket->so_snd.sb_lowat; log.u_bbr.flex3 = bbr->rc_inp->inp_socket->so_snd.sb_hiwat; log.u_bbr.flex7 = line; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_TIME_EPOCH, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_set_of_state_target(struct tcp_bbr *bbr, uint32_t new_tar, int line, int meth) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); log.u_bbr.flex1 = bbr->r_ctl.rc_target_at_state; log.u_bbr.flex2 = new_tar; log.u_bbr.flex3 = line; log.u_bbr.flex4 = bbr->r_ctl.rc_pace_max_segs; log.u_bbr.flex5 = bbr_quanta; log.u_bbr.flex6 = bbr->r_ctl.rc_pace_min_segs; log.u_bbr.flex7 = bbr->rc_last_options; log.u_bbr.flex8 = meth; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_STATE_TARGET, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_type_statechange(struct tcp_bbr *bbr, uint32_t cts, int32_t line) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = line; log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks; log.u_bbr.flex3 = bbr->r_ctl.rc_probertt_int; if (bbr_state_is_pkt_epoch) log.u_bbr.flex4 = bbr_get_rtt(bbr, BBR_RTT_PKTRTT); else log.u_bbr.flex4 = bbr_get_rtt(bbr, BBR_RTT_PROP); log.u_bbr.flex5 = bbr->r_ctl.rc_bbr_last_startup_epoch; log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup; log.u_bbr.flex7 = (bbr->r_ctl.rc_target_at_state/1000); log.u_bbr.lt_epoch = bbr->r_ctl.rc_level_state_extra; log.u_bbr.pkts_out = bbr->r_ctl.rc_target_at_state; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_STATE, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_rtt_shrinks(struct tcp_bbr *bbr, uint32_t cts, uint32_t applied, uint32_t rtt, uint32_t line, uint8_t reas, uint16_t cond) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = line; log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks; log.u_bbr.flex3 = bbr->r_ctl.last_in_probertt; log.u_bbr.flex4 = applied; log.u_bbr.flex5 = rtt; log.u_bbr.flex6 = bbr->r_ctl.rc_target_at_state; log.u_bbr.flex7 = cond; log.u_bbr.flex8 = reas; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_RTT_SHRINKS, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_type_exit_rec(struct tcp_bbr *bbr) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); log.u_bbr.flex1 = bbr->r_ctl.rc_recovery_start; log.u_bbr.flex2 = bbr->r_ctl.rc_cwnd_on_ent; log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_EXITREC, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_type_cwndupd(struct tcp_bbr *bbr, uint32_t bytes_this_ack, uint32_t chg, uint32_t prev_acked, int32_t meth, uint32_t target, uint32_t th_ack, int32_t line) { if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); log.u_bbr.flex1 = line; log.u_bbr.flex2 = prev_acked; log.u_bbr.flex3 = bytes_this_ack; log.u_bbr.flex4 = chg; log.u_bbr.flex5 = th_ack; log.u_bbr.flex6 = target; log.u_bbr.flex8 = meth; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_CWND, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_rtt_sample(struct tcp_bbr *bbr, uint32_t rtt, uint32_t tsin) { /* * Log the rtt sample we are applying to the srtt algorithm in * useconds. */ if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); log.u_bbr.flex1 = rtt; log.u_bbr.flex2 = bbr->r_ctl.rc_bbr_state_time; log.u_bbr.flex3 = bbr->r_ctl.rc_ack_hdwr_delay; log.u_bbr.flex4 = bbr->rc_tp->ts_offset; log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; log.u_bbr.pkts_out = tcp_tv_to_mssectick(&bbr->rc_tv); log.u_bbr.flex6 = tsin; log.u_bbr.flex7 = 0; log.u_bbr.flex8 = bbr->rc_ack_was_delayed; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, TCP_LOG_RTT, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_type_pesist(struct tcp_bbr *bbr, uint32_t cts, uint32_t time_in, int32_t line, uint8_t enter_exit) { if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = time_in; log.u_bbr.flex2 = line; log.u_bbr.flex8 = enter_exit; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_PERSIST, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_ack_clear(struct tcp_bbr *bbr, uint32_t cts) { if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = bbr->rc_tp->ts_recent_age; log.u_bbr.flex2 = bbr->r_ctl.rc_rtt_shrinks; log.u_bbr.flex3 = bbr->r_ctl.rc_probertt_int; log.u_bbr.flex4 = bbr->r_ctl.rc_went_idle_time; log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_ACKCLEAR, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_ack_event(struct tcp_bbr *bbr, struct tcphdr *th, struct tcpopt *to, uint32_t tlen, uint16_t nsegs, uint32_t cts, int32_t nxt_pkt, struct mbuf *m) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = nsegs; log.u_bbr.flex2 = bbr->r_ctl.rc_lost_bytes; if (m) { struct timespec ts; log.u_bbr.flex3 = m->m_flags; if (m->m_flags & M_TSTMP) { mbuf_tstmp2timespec(m, &ts); tv.tv_sec = ts.tv_sec; tv.tv_usec = ts.tv_nsec / 1000; log.u_bbr.lt_epoch = tcp_tv_to_usectick(&tv); } else { log.u_bbr.lt_epoch = 0; } if (m->m_flags & M_TSTMP_LRO) { mbuf_tstmp2timeval(m, &tv); log.u_bbr.flex5 = tcp_tv_to_usectick(&tv); } else { /* No arrival timestamp */ log.u_bbr.flex5 = 0; } log.u_bbr.pkts_out = tcp_get_usecs(&tv); } else { log.u_bbr.flex3 = 0; log.u_bbr.flex5 = 0; log.u_bbr.flex6 = 0; log.u_bbr.pkts_out = 0; } log.u_bbr.flex4 = bbr->r_ctl.rc_target_at_state; log.u_bbr.flex7 = bbr->r_wanted_output; log.u_bbr.flex8 = bbr->rc_in_persist; TCP_LOG_EVENTP(bbr->rc_tp, th, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, TCP_LOG_IN, 0, tlen, &log, true, &bbr->rc_tv); } } static void bbr_log_doseg_done(struct tcp_bbr *bbr, uint32_t cts, int32_t nxt_pkt, int32_t did_out) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = did_out; log.u_bbr.flex2 = nxt_pkt; log.u_bbr.flex3 = bbr->r_ctl.rc_last_delay_val; log.u_bbr.flex4 = bbr->r_ctl.rc_hpts_flags; log.u_bbr.flex5 = bbr->r_ctl.rc_timer_exp; log.u_bbr.flex6 = bbr->r_ctl.rc_lost_bytes; log.u_bbr.flex7 = bbr->r_wanted_output; log.u_bbr.flex8 = bbr->rc_in_persist; log.u_bbr.pkts_out = bbr->r_ctl.highest_hdwr_delay; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_DOSEG_DONE, 0, 0, &log, true, &bbr->rc_tv); } } static void bbr_log_enobuf_jmp(struct tcp_bbr *bbr, uint32_t len, uint32_t cts, int32_t line, uint32_t o_len, uint32_t segcnt, uint32_t segsiz) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = line; log.u_bbr.flex2 = o_len; log.u_bbr.flex3 = segcnt; log.u_bbr.flex4 = segsiz; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_ENOBUF_JMP, ENOBUFS, len, &log, true, &bbr->rc_tv); } } static void bbr_log_to_processing(struct tcp_bbr *bbr, uint32_t cts, int32_t ret, int32_t timers, uint8_t hpts_calling) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = timers; log.u_bbr.flex2 = ret; log.u_bbr.flex3 = bbr->r_ctl.rc_timer_exp; log.u_bbr.flex4 = bbr->r_ctl.rc_hpts_flags; log.u_bbr.flex5 = cts; log.u_bbr.flex6 = bbr->r_ctl.rc_target_at_state; log.u_bbr.flex8 = hpts_calling; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_TO_PROCESS, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_to_event(struct tcp_bbr *bbr, uint32_t cts, int32_t to_num) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; uint64_t ar; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = bbr->bbr_timer_src; log.u_bbr.flex2 = 0; log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags; ar = (uint64_t)(bbr->r_ctl.rc_resend); ar >>= 32; ar &= 0x00000000ffffffff; log.u_bbr.flex4 = (uint32_t)ar; ar = (uint64_t)bbr->r_ctl.rc_resend; ar &= 0x00000000ffffffff; log.u_bbr.flex5 = (uint32_t)ar; log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); log.u_bbr.flex8 = to_num; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_RTO, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_startup_event(struct tcp_bbr *bbr, uint32_t cts, uint32_t flex1, uint32_t flex2, uint32_t flex3, uint8_t reason) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = flex1; log.u_bbr.flex2 = flex2; log.u_bbr.flex3 = flex3; log.u_bbr.flex4 = 0; log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; log.u_bbr.flex6 = bbr->r_ctl.rc_lost_at_startup; log.u_bbr.flex8 = reason; log.u_bbr.cur_del_rate = bbr->r_ctl.rc_bbr_lastbtlbw; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_REDUCE, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_hpts_diag(struct tcp_bbr *bbr, uint32_t cts, struct hpts_diag *diag) { if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = diag->p_nxt_slot; log.u_bbr.flex2 = diag->p_cur_slot; log.u_bbr.flex3 = diag->slot_req; log.u_bbr.flex4 = diag->inp_hptsslot; log.u_bbr.flex5 = diag->slot_remaining; log.u_bbr.flex6 = diag->need_new_to; log.u_bbr.flex7 = diag->p_hpts_active; log.u_bbr.flex8 = diag->p_on_min_sleep; /* Hijack other fields as needed */ log.u_bbr.epoch = diag->have_slept; log.u_bbr.lt_epoch = diag->yet_to_sleep; log.u_bbr.pkts_out = diag->co_ret; log.u_bbr.applimited = diag->hpts_sleep_time; log.u_bbr.delivered = diag->p_prev_slot; log.u_bbr.inflight = diag->p_runningslot; log.u_bbr.bw_inuse = diag->wheel_slot; log.u_bbr.rttProp = diag->wheel_cts; log.u_bbr.delRate = diag->maxslots; log.u_bbr.cur_del_rate = diag->p_curtick; log.u_bbr.cur_del_rate <<= 32; log.u_bbr.cur_del_rate |= diag->p_lasttick; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_HPTSDIAG, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_timer_var(struct tcp_bbr *bbr, int mode, uint32_t cts, uint32_t time_since_sent, uint32_t srtt, uint32_t thresh, uint32_t to) { if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = bbr->rc_tp->t_rttvar; log.u_bbr.flex2 = time_since_sent; log.u_bbr.flex3 = srtt; log.u_bbr.flex4 = thresh; log.u_bbr.flex5 = to; log.u_bbr.flex6 = bbr->rc_tp->t_srtt; log.u_bbr.flex8 = mode; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_TIMERPREP, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_pacing_delay_calc(struct tcp_bbr *bbr, uint16_t gain, uint32_t len, uint32_t cts, uint32_t usecs, uint64_t bw, uint32_t override, int mod) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = usecs; log.u_bbr.flex2 = len; log.u_bbr.flex3 = (uint32_t)((bw >> 32) & 0x00000000ffffffff); log.u_bbr.flex4 = (uint32_t)(bw & 0x00000000ffffffff); if (override) log.u_bbr.flex5 = (1 << 2); else log.u_bbr.flex5 = 0; log.u_bbr.flex6 = override; log.u_bbr.flex7 = gain; log.u_bbr.flex8 = mod; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_HPTSI_CALC, 0, len, &log, false, &bbr->rc_tv); } } static void bbr_log_to_start(struct tcp_bbr *bbr, uint32_t cts, uint32_t to, int32_t slot, uint8_t which) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = bbr->bbr_timer_src; log.u_bbr.flex2 = to; log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags; log.u_bbr.flex4 = slot; log.u_bbr.flex5 = bbr->rc_inp->inp_hptsslot; log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); log.u_bbr.pkts_out = bbr->rc_inp->inp_flags2; log.u_bbr.flex8 = which; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_TIMERSTAR, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_thresh_choice(struct tcp_bbr *bbr, uint32_t cts, uint32_t thresh, uint32_t lro, uint32_t srtt, struct bbr_sendmap *rsm, uint8_t frm) { if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = thresh; log.u_bbr.flex2 = lro; log.u_bbr.flex3 = bbr->r_ctl.rc_reorder_ts; log.u_bbr.flex4 = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; log.u_bbr.flex5 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); log.u_bbr.flex6 = srtt; log.u_bbr.flex7 = bbr->r_ctl.rc_reorder_shift; log.u_bbr.flex8 = frm; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_THRESH_CALC, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_to_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts, uint8_t hpts_removed) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = line; log.u_bbr.flex2 = bbr->bbr_timer_src; log.u_bbr.flex3 = bbr->r_ctl.rc_hpts_flags; log.u_bbr.flex4 = bbr->rc_in_persist; log.u_bbr.flex5 = bbr->r_ctl.rc_target_at_state; log.u_bbr.flex6 = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); log.u_bbr.flex8 = hpts_removed; log.u_bbr.pkts_out = bbr->rc_pacer_started; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_TIMERCANC, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_tstmp_validation(struct tcp_bbr *bbr, uint64_t peer_delta, uint64_t delta) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); log.u_bbr.flex1 = bbr->r_ctl.bbr_peer_tsratio; log.u_bbr.flex2 = (peer_delta >> 32); log.u_bbr.flex3 = (peer_delta & 0x00000000ffffffff); log.u_bbr.flex4 = (delta >> 32); log.u_bbr.flex5 = (delta & 0x00000000ffffffff); log.u_bbr.flex7 = bbr->rc_ts_clock_set; log.u_bbr.flex8 = bbr->rc_ts_cant_be_used; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_TSTMP_VAL, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_type_tsosize(struct tcp_bbr *bbr, uint32_t cts, uint32_t tsosz, uint32_t tls, uint32_t old_val, uint32_t maxseg, int hdwr) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = tsosz; log.u_bbr.flex2 = tls; log.u_bbr.flex3 = tcp_min_hptsi_time; log.u_bbr.flex4 = bbr->r_ctl.bbr_hptsi_bytes_min; log.u_bbr.flex5 = old_val; log.u_bbr.flex6 = maxseg; log.u_bbr.flex7 = bbr->rc_no_pacing; log.u_bbr.flex7 <<= 1; log.u_bbr.flex7 |= bbr->rc_past_init_win; if (hdwr) log.u_bbr.flex8 = 0x80 | bbr->rc_use_google; else log.u_bbr.flex8 = bbr->rc_use_google; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_BBRTSO, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_type_rsmclear(struct tcp_bbr *bbr, uint32_t cts, struct bbr_sendmap *rsm, uint32_t flags, uint32_t line) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = line; log.u_bbr.flex2 = rsm->r_start; log.u_bbr.flex3 = rsm->r_end; log.u_bbr.flex4 = rsm->r_delivered; log.u_bbr.flex5 = rsm->r_rtr_cnt; log.u_bbr.flex6 = rsm->r_dupack; log.u_bbr.flex7 = rsm->r_tim_lastsent[0]; log.u_bbr.flex8 = rsm->r_flags; /* Hijack the pkts_out fids */ log.u_bbr.applimited = flags; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_RSM_CLEARED, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_type_bbrupd(struct tcp_bbr *bbr, uint8_t flex8, uint32_t cts, uint32_t flex3, uint32_t flex2, uint32_t flex5, uint32_t flex6, uint32_t pkts_out, int flex7, uint32_t flex4, uint32_t flex1) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = flex1; log.u_bbr.flex2 = flex2; log.u_bbr.flex3 = flex3; log.u_bbr.flex4 = flex4; log.u_bbr.flex5 = flex5; log.u_bbr.flex6 = flex6; log.u_bbr.flex7 = flex7; /* Hijack the pkts_out fids */ log.u_bbr.pkts_out = pkts_out; log.u_bbr.flex8 = flex8; if (bbr->rc_ack_was_delayed) log.u_bbr.epoch = bbr->r_ctl.rc_ack_hdwr_delay; else log.u_bbr.epoch = 0; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_BBRUPD, 0, flex2, &log, false, &bbr->rc_tv); } } static void bbr_log_type_ltbw(struct tcp_bbr *bbr, uint32_t cts, int32_t reason, uint32_t newbw, uint32_t obw, uint32_t diff, uint32_t tim) { if (/*bbr_verbose_logging && */(bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = reason; log.u_bbr.flex2 = newbw; log.u_bbr.flex3 = obw; log.u_bbr.flex4 = diff; log.u_bbr.flex5 = bbr->r_ctl.rc_lt_lost; log.u_bbr.flex6 = bbr->r_ctl.rc_lt_del; log.u_bbr.flex7 = bbr->rc_lt_is_sampling; log.u_bbr.pkts_out = tim; log.u_bbr.bw_inuse = bbr->r_ctl.rc_lt_bw; if (bbr->rc_lt_use_bw == 0) log.u_bbr.epoch = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch; else log.u_bbr.epoch = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch_use; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_BWSAMP, 0, 0, &log, false, &bbr->rc_tv); } } static inline void bbr_log_progress_event(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t tick, int event, int line) { if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); log.u_bbr.flex1 = line; log.u_bbr.flex2 = tick; log.u_bbr.flex3 = tp->t_maxunacktime; log.u_bbr.flex4 = tp->t_acktime; log.u_bbr.flex8 = event; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_PROGRESS, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_type_log_hdwr_pacing(struct tcp_bbr *bbr, const struct ifnet *ifp, uint64_t rate, uint64_t hw_rate, int line, uint32_t cts, int error) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff); log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff); log.u_bbr.flex3 = (((uint64_t)ifp >> 32) & 0x00000000ffffffff); log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff); log.u_bbr.bw_inuse = rate; log.u_bbr.flex5 = line; log.u_bbr.flex6 = error; log.u_bbr.flex8 = bbr->skip_gain; log.u_bbr.flex8 <<= 1; log.u_bbr.flex8 |= bbr->gain_is_limited; log.u_bbr.flex8 <<= 1; log.u_bbr.flex8 |= bbr->bbr_hdrw_pacing; log.u_bbr.pkts_out = bbr->rc_tp->t_maxseg; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_HDWR_PACE, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_type_bbrsnd(struct tcp_bbr *bbr, uint32_t len, uint32_t slot, uint32_t del_by, uint32_t cts, uint32_t line, uint32_t prev_delay) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = slot; log.u_bbr.flex2 = del_by; log.u_bbr.flex3 = prev_delay; log.u_bbr.flex4 = line; log.u_bbr.flex5 = bbr->r_ctl.rc_last_delay_val; log.u_bbr.flex6 = bbr->r_ctl.rc_hptsi_agg_delay; log.u_bbr.flex7 = (0x0000ffff & bbr->r_ctl.rc_hpts_flags); log.u_bbr.flex8 = bbr->rc_in_persist; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_BBRSND, 0, len, &log, false, &bbr->rc_tv); } } static void bbr_log_type_bbrrttprop(struct tcp_bbr *bbr, uint32_t t, uint32_t end, uint32_t tsconv, uint32_t cts, int32_t match, uint32_t seq, uint8_t flags) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = bbr->r_ctl.rc_delivered; log.u_bbr.flex2 = 0; log.u_bbr.flex3 = bbr->r_ctl.rc_lowest_rtt; log.u_bbr.flex4 = end; log.u_bbr.flex5 = seq; log.u_bbr.flex6 = t; log.u_bbr.flex7 = match; log.u_bbr.flex8 = flags; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_BBRRTT, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_exit_gain(struct tcp_bbr *bbr, uint32_t cts, int32_t entry_method) { if (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = bbr->r_ctl.rc_target_at_state; log.u_bbr.flex2 = (bbr->rc_tp->t_maxseg - bbr->rc_last_options); log.u_bbr.flex3 = bbr->r_ctl.gain_epoch; log.u_bbr.flex4 = bbr->r_ctl.rc_pace_max_segs; log.u_bbr.flex5 = bbr->r_ctl.rc_pace_min_segs; log.u_bbr.flex6 = bbr->r_ctl.rc_bbr_state_atflight; log.u_bbr.flex7 = 0; log.u_bbr.flex8 = entry_method; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_EXIT_GAIN, 0, 0, &log, false, &bbr->rc_tv); } } static void bbr_log_settings_change(struct tcp_bbr *bbr, int settings_desired) { if (bbr_verbose_logging && (bbr->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, bbr->r_ctl.rc_rcvtime); /* R-HU */ log.u_bbr.flex1 = 0; log.u_bbr.flex2 = 0; log.u_bbr.flex3 = 0; log.u_bbr.flex4 = 0; log.u_bbr.flex7 = 0; log.u_bbr.flex8 = settings_desired; TCP_LOG_EVENTP(bbr->rc_tp, NULL, &bbr->rc_inp->inp_socket->so_rcv, &bbr->rc_inp->inp_socket->so_snd, BBR_LOG_SETTINGS_CHG, 0, 0, &log, false, &bbr->rc_tv); } } /* * Returns the bw from the our filter. */ static inline uint64_t bbr_get_full_bw(struct tcp_bbr *bbr) { uint64_t bw; bw = get_filter_value(&bbr->r_ctl.rc_delrate); return (bw); } static inline void bbr_set_pktepoch(struct tcp_bbr *bbr, uint32_t cts, int32_t line) { uint64_t calclr; uint32_t lost, del; if (bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_pktepoch) lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lost_at_pktepoch; else lost = 0; del = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_pkt_epoch_del; if (lost == 0) { calclr = 0; } else if (del) { calclr = lost; calclr *= (uint64_t)1000; calclr /= (uint64_t)del; } else { /* Nothing delivered? 100.0% loss */ calclr = 1000; } bbr->r_ctl.rc_pkt_epoch_loss_rate = (uint32_t)calclr; if (IN_RECOVERY(bbr->rc_tp->t_flags)) bbr->r_ctl.recovery_lr += (uint32_t)calclr; bbr->r_ctl.rc_pkt_epoch++; if (bbr->rc_no_pacing && (bbr->r_ctl.rc_pkt_epoch >= bbr->no_pacing_until)) { bbr->rc_no_pacing = 0; tcp_bbr_tso_size_check(bbr, cts); } bbr->r_ctl.rc_pkt_epoch_rtt = bbr_calc_time(cts, bbr->r_ctl.rc_pkt_epoch_time); bbr->r_ctl.rc_pkt_epoch_time = cts; /* What was our loss rate */ bbr_log_pkt_epoch(bbr, cts, line, lost, del); bbr->r_ctl.rc_pkt_epoch_del = bbr->r_ctl.rc_delivered; bbr->r_ctl.rc_lost_at_pktepoch = bbr->r_ctl.rc_lost; } static inline void bbr_set_epoch(struct tcp_bbr *bbr, uint32_t cts, int32_t line) { uint32_t epoch_time; /* Tick the RTT clock */ bbr->r_ctl.rc_rtt_epoch++; epoch_time = cts - bbr->r_ctl.rc_rcv_epoch_start; bbr_log_time_epoch(bbr, cts, line, epoch_time); bbr->r_ctl.rc_rcv_epoch_start = cts; } static inline void bbr_isit_a_pkt_epoch(struct tcp_bbr *bbr, uint32_t cts, struct bbr_sendmap *rsm, int32_t line, int32_t cum_acked) { if (SEQ_GEQ(rsm->r_delivered, bbr->r_ctl.rc_pkt_epoch_del)) { bbr->rc_is_pkt_epoch_now = 1; } } /* * Returns the bw from either the b/w filter * or from the lt_bw (if the connection is being * policed). */ static inline uint64_t __bbr_get_bw(struct tcp_bbr *bbr) { uint64_t bw, min_bw; uint64_t rtt; int gm_measure_cnt = 1; /* * For startup we make, like google, a * minimum b/w. This is generated from the * IW and the rttProp. We do fall back to srtt * if for some reason (initial handshake) we don't * have a rttProp. We, in the worst case, fall back * to the configured min_bw (rc_initial_hptsi_bw). */ if (bbr->rc_bbr_state == BBR_STATE_STARTUP) { /* Attempt first to use rttProp */ rtt = (uint64_t)get_filter_value_small(&bbr->r_ctl.rc_rttprop); if (rtt && (rtt < 0xffffffff)) { measure: min_bw = (uint64_t)(bbr_initial_cwnd(bbr, bbr->rc_tp)) * ((uint64_t)1000000); min_bw /= rtt; if (min_bw < bbr->r_ctl.rc_initial_hptsi_bw) { min_bw = bbr->r_ctl.rc_initial_hptsi_bw; } } else if (bbr->rc_tp->t_srtt != 0) { /* No rttProp, use srtt? */ rtt = bbr_get_rtt(bbr, BBR_SRTT); goto measure; } else { min_bw = bbr->r_ctl.rc_initial_hptsi_bw; } } else min_bw = 0; if ((bbr->rc_past_init_win == 0) && (bbr->r_ctl.rc_delivered > bbr_initial_cwnd(bbr, bbr->rc_tp))) bbr->rc_past_init_win = 1; if ((bbr->rc_use_google) && (bbr->r_ctl.r_measurement_count >= 1)) gm_measure_cnt = 0; if (gm_measure_cnt && ((bbr->r_ctl.r_measurement_count < bbr_min_measurements_req) || (bbr->rc_past_init_win == 0))) { /* For google we use our guess rate until we get 1 measurement */ use_initial_window: rtt = (uint64_t)get_filter_value_small(&bbr->r_ctl.rc_rttprop); if (rtt && (rtt < 0xffffffff)) { /* * We have an RTT measurement. Use that in * combination with our initial window to calculate * a b/w. */ bw = (uint64_t)(bbr_initial_cwnd(bbr, bbr->rc_tp)) * ((uint64_t)1000000); bw /= rtt; if (bw < bbr->r_ctl.rc_initial_hptsi_bw) { bw = bbr->r_ctl.rc_initial_hptsi_bw; } } else { /* Drop back to the 40 and punt to a default */ bw = bbr->r_ctl.rc_initial_hptsi_bw; } if (bw < 1) /* Probably should panic */ bw = 1; if (bw > min_bw) return (bw); else return (min_bw); } if (bbr->rc_lt_use_bw) bw = bbr->r_ctl.rc_lt_bw; else if (bbr->r_recovery_bw && (bbr->rc_use_google == 0)) bw = bbr->r_ctl.red_bw; else bw = get_filter_value(&bbr->r_ctl.rc_delrate); if (bbr->rc_tp->t_peakrate_thr && (bbr->rc_use_google == 0)) { /* * Enforce user set rate limit, keep in mind that * t_peakrate_thr is in B/s already */ bw = uqmin((uint64_t)bbr->rc_tp->t_peakrate_thr, bw); } if (bw == 0) { /* We should not be at 0, go to the initial window then */ goto use_initial_window; } if (bw < 1) /* Probably should panic */ bw = 1; if (bw < min_bw) bw = min_bw; return (bw); } static inline uint64_t bbr_get_bw(struct tcp_bbr *bbr) { uint64_t bw; bw = __bbr_get_bw(bbr); return (bw); } static inline void bbr_reset_lt_bw_interval(struct tcp_bbr *bbr, uint32_t cts) { bbr->r_ctl.rc_lt_epoch = bbr->r_ctl.rc_pkt_epoch; bbr->r_ctl.rc_lt_time = bbr->r_ctl.rc_del_time; bbr->r_ctl.rc_lt_del = bbr->r_ctl.rc_delivered; bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; } static inline void bbr_reset_lt_bw_sampling(struct tcp_bbr *bbr, uint32_t cts) { bbr->rc_lt_is_sampling = 0; bbr->rc_lt_use_bw = 0; bbr->r_ctl.rc_lt_bw = 0; bbr_reset_lt_bw_interval(bbr, cts); } static inline void bbr_lt_bw_samp_done(struct tcp_bbr *bbr, uint64_t bw, uint32_t cts, uint32_t timin) { uint64_t diff; /* Do we have a previous sample? */ if (bbr->r_ctl.rc_lt_bw) { /* Get the diff in bytes per second */ if (bbr->r_ctl.rc_lt_bw > bw) diff = bbr->r_ctl.rc_lt_bw - bw; else diff = bw - bbr->r_ctl.rc_lt_bw; if ((diff <= bbr_lt_bw_diff) || (diff <= (bbr->r_ctl.rc_lt_bw / bbr_lt_bw_ratio))) { /* Consider us policed */ uint32_t saved_bw; saved_bw = (uint32_t)bbr->r_ctl.rc_lt_bw; bbr->r_ctl.rc_lt_bw = (bw + bbr->r_ctl.rc_lt_bw) / 2; /* average of two */ bbr->rc_lt_use_bw = 1; bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; /* * Use pkt based epoch for measuring length of * policer up */ bbr->r_ctl.rc_lt_epoch_use = bbr->r_ctl.rc_pkt_epoch; /* * reason 4 is we need to start consider being * policed */ bbr_log_type_ltbw(bbr, cts, 4, (uint32_t)bw, saved_bw, (uint32_t)diff, timin); return; } } bbr->r_ctl.rc_lt_bw = bw; bbr_reset_lt_bw_interval(bbr, cts); bbr_log_type_ltbw(bbr, cts, 5, 0, (uint32_t)bw, 0, timin); } static void bbr_randomize_extra_state_time(struct tcp_bbr *bbr) { uint32_t ran, deduct; ran = arc4random_uniform(bbr_rand_ot); if (ran) { deduct = bbr->r_ctl.rc_level_state_extra / ran; bbr->r_ctl.rc_level_state_extra -= deduct; } } /* * Return randomly the starting state * to use in probebw. */ static uint8_t bbr_pick_probebw_substate(struct tcp_bbr *bbr, uint32_t cts) { uint32_t ran; uint8_t ret_val; /* Initialize the offset to 0 */ bbr->r_ctl.rc_exta_time_gd = 0; bbr->rc_hit_state_1 = 0; bbr->r_ctl.rc_level_state_extra = 0; ran = arc4random_uniform((BBR_SUBSTATE_COUNT-1)); /* * The math works funny here :) the return value is used to set the * substate and then the state change is called which increments by * one. So if we return 1 (DRAIN) we will increment to 2 (LEVEL1) when * we fully enter the state. Note that the (8 - 1 - ran) assures that * we return 1 - 7, so we dont return 0 and end up starting in * state 1 (DRAIN). */ ret_val = BBR_SUBSTATE_COUNT - 1 - ran; /* Set an epoch */ if ((cts - bbr->r_ctl.rc_rcv_epoch_start) >= bbr_get_rtt(bbr, BBR_RTT_PROP)) bbr_set_epoch(bbr, cts, __LINE__); bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; return (ret_val); } static void bbr_lt_bw_sampling(struct tcp_bbr *bbr, uint32_t cts, int32_t loss_detected) { uint32_t diff, d_time; uint64_t del_time, bw, lost, delivered; if (bbr->r_use_policer == 0) return; if (bbr->rc_lt_use_bw) { /* We are using lt bw do we stop yet? */ diff = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch_use; if (diff > bbr_lt_bw_max_rtts) { /* Reset it all */ reset_all: bbr_reset_lt_bw_sampling(bbr, cts); if (bbr->rc_filled_pipe) { bbr_set_epoch(bbr, cts, __LINE__); bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts); bbr_substate_change(bbr, cts, __LINE__, 0); bbr->rc_bbr_state = BBR_STATE_PROBE_BW; bbr_log_type_statechange(bbr, cts, __LINE__); } else { /* * This should not happen really * unless we remove the startup/drain * restrictions above. */ bbr->rc_bbr_state = BBR_STATE_STARTUP; bbr_set_epoch(bbr, cts, __LINE__); bbr->r_ctl.rc_bbr_state_time = cts; bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg; bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg; bbr_set_state_target(bbr, __LINE__); bbr_log_type_statechange(bbr, cts, __LINE__); } /* reason 0 is to stop using lt-bw */ bbr_log_type_ltbw(bbr, cts, 0, 0, 0, 0, 0); return; } if (bbr_lt_intvl_fp == 0) { /* Not doing false-positive detection */ return; } /* False positive detection */ if (diff == bbr_lt_intvl_fp) { /* At bbr_lt_intvl_fp we record the lost */ bbr->r_ctl.rc_lt_del = bbr->r_ctl.rc_delivered; bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; } else if (diff > (bbr_lt_intvl_min_rtts + bbr_lt_intvl_fp)) { /* Now is our loss rate still high? */ lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lt_lost; delivered = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_lt_del; if ((delivered == 0) || (((lost * 1000)/delivered) < bbr_lt_fd_thresh)) { /* No still below our threshold */ bbr_log_type_ltbw(bbr, cts, 7, lost, delivered, 0, 0); } else { /* Yikes its still high, it must be a false positive */ bbr_log_type_ltbw(bbr, cts, 8, lost, delivered, 0, 0); goto reset_all; } } return; } /* * Wait for the first loss before sampling, to let the policer * exhaust its tokens and estimate the steady-state rate allowed by * the policer. Starting samples earlier includes bursts that * over-estimate the bw. */ if (bbr->rc_lt_is_sampling == 0) { /* reason 1 is to begin doing the sampling */ if (loss_detected == 0) return; bbr_reset_lt_bw_interval(bbr, cts); bbr->rc_lt_is_sampling = 1; bbr_log_type_ltbw(bbr, cts, 1, 0, 0, 0, 0); return; } /* Now how long were we delivering long term last> */ if (TSTMP_GEQ(bbr->r_ctl.rc_del_time, bbr->r_ctl.rc_lt_time)) d_time = bbr->r_ctl.rc_del_time - bbr->r_ctl.rc_lt_time; else d_time = 0; /* To avoid underestimates, reset sampling if we run out of data. */ if (bbr->r_ctl.r_app_limited_until) { /* Can not measure in app-limited state */ bbr_reset_lt_bw_sampling(bbr, cts); /* reason 2 is to reset sampling due to app limits */ bbr_log_type_ltbw(bbr, cts, 2, 0, 0, 0, d_time); return; } diff = bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_lt_epoch; if (diff < bbr_lt_intvl_min_rtts) { /* * need more samples (we don't * start on a round like linux so * we need 1 more). */ /* 6 is not_enough time or no-loss */ bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time); return; } if (diff > (4 * bbr_lt_intvl_min_rtts)) { /* * For now if we wait too long, reset all sampling. We need * to do some research here, its possible that we should * base this on how much loss as occurred.. something like * if its under 10% (or some thresh) reset all otherwise * don't. Thats for phase II I guess. */ bbr_reset_lt_bw_sampling(bbr, cts); /* reason 3 is to reset sampling due too long of sampling */ bbr_log_type_ltbw(bbr, cts, 3, 0, 0, 0, d_time); return; } /* * End sampling interval when a packet is lost, so we estimate the * policer tokens were exhausted. Stopping the sampling before the * tokens are exhausted under-estimates the policed rate. */ if (loss_detected == 0) { /* 6 is not_enough time or no-loss */ bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time); return; } /* Calculate packets lost and delivered in sampling interval. */ lost = bbr->r_ctl.rc_lost - bbr->r_ctl.rc_lt_lost; delivered = bbr->r_ctl.rc_delivered - bbr->r_ctl.rc_lt_del; if ((delivered == 0) || (((lost * 1000)/delivered) < bbr_lt_loss_thresh)) { bbr_log_type_ltbw(bbr, cts, 6, lost, delivered, 0, d_time); return; } if (d_time < 1000) { /* Not enough time. wait */ /* 6 is not_enough time or no-loss */ bbr_log_type_ltbw(bbr, cts, 6, 0, 0, 0, d_time); return; } if (d_time >= (0xffffffff / USECS_IN_MSEC)) { /* Too long */ bbr_reset_lt_bw_sampling(bbr, cts); /* reason 3 is to reset sampling due too long of sampling */ bbr_log_type_ltbw(bbr, cts, 3, 0, 0, 0, d_time); return; } del_time = d_time; bw = delivered; bw *= (uint64_t)USECS_IN_SECOND; bw /= del_time; bbr_lt_bw_samp_done(bbr, bw, cts, d_time); } /* * Allocate a sendmap from our zone. */ static struct bbr_sendmap * bbr_alloc(struct tcp_bbr *bbr) { struct bbr_sendmap *rsm; BBR_STAT_INC(bbr_to_alloc); rsm = uma_zalloc(bbr_zone, (M_NOWAIT | M_ZERO)); if (rsm) { bbr->r_ctl.rc_num_maps_alloced++; return (rsm); } if (bbr->r_ctl.rc_free_cnt) { BBR_STAT_INC(bbr_to_alloc_emerg); rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free); TAILQ_REMOVE(&bbr->r_ctl.rc_free, rsm, r_next); bbr->r_ctl.rc_free_cnt--; return (rsm); } BBR_STAT_INC(bbr_to_alloc_failed); return (NULL); } static struct bbr_sendmap * bbr_alloc_full_limit(struct tcp_bbr *bbr) { if ((V_tcp_map_entries_limit > 0) && (bbr->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) { BBR_STAT_INC(bbr_alloc_limited); if (!bbr->alloc_limit_reported) { bbr->alloc_limit_reported = 1; BBR_STAT_INC(bbr_alloc_limited_conns); } return (NULL); } return (bbr_alloc(bbr)); } /* wrapper to allocate a sendmap entry, subject to a specific limit */ static struct bbr_sendmap * bbr_alloc_limit(struct tcp_bbr *bbr, uint8_t limit_type) { struct bbr_sendmap *rsm; if (limit_type) { /* currently there is only one limit type */ if (V_tcp_map_split_limit > 0 && bbr->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) { BBR_STAT_INC(bbr_split_limited); if (!bbr->alloc_limit_reported) { bbr->alloc_limit_reported = 1; BBR_STAT_INC(bbr_alloc_limited_conns); } return (NULL); } } /* allocate and mark in the limit type, if set */ rsm = bbr_alloc(bbr); if (rsm != NULL && limit_type) { rsm->r_limit_type = limit_type; bbr->r_ctl.rc_num_split_allocs++; } return (rsm); } static void bbr_free(struct tcp_bbr *bbr, struct bbr_sendmap *rsm) { if (rsm->r_limit_type) { /* currently there is only one limit type */ bbr->r_ctl.rc_num_split_allocs--; } if (rsm->r_is_smallmap) bbr->r_ctl.rc_num_small_maps_alloced--; if (bbr->r_ctl.rc_tlp_send == rsm) bbr->r_ctl.rc_tlp_send = NULL; if (bbr->r_ctl.rc_resend == rsm) { bbr->r_ctl.rc_resend = NULL; } if (bbr->r_ctl.rc_next == rsm) bbr->r_ctl.rc_next = NULL; if (bbr->r_ctl.rc_sacklast == rsm) bbr->r_ctl.rc_sacklast = NULL; if (bbr->r_ctl.rc_free_cnt < bbr_min_req_free) { memset(rsm, 0, sizeof(struct bbr_sendmap)); TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_free, rsm, r_next); rsm->r_limit_type = 0; bbr->r_ctl.rc_free_cnt++; return; } bbr->r_ctl.rc_num_maps_alloced--; uma_zfree(bbr_zone, rsm); } /* * Returns the BDP. */ static uint64_t bbr_get_bw_delay_prod(uint64_t rtt, uint64_t bw) { /* * Calculate the bytes in flight needed given the bw (in bytes per * second) and the specifyed rtt in useconds. We need to put out the * returned value per RTT to match that rate. Gain will normally * raise it up from there. * * This should not overflow as long as the bandwidth is below 1 * TByte per second (bw < 10**12 = 2**40) and the rtt is smaller * than 1000 seconds (rtt < 10**3 * 10**6 = 10**9 = 2**30). */ uint64_t usec_per_sec; usec_per_sec = USECS_IN_SECOND; return ((rtt * bw) / usec_per_sec); } /* * Return the initial cwnd. */ static uint32_t bbr_initial_cwnd(struct tcp_bbr *bbr, struct tcpcb *tp) { uint32_t i_cwnd; if (bbr->rc_init_win) { i_cwnd = bbr->rc_init_win * tp->t_maxseg; } else if (V_tcp_initcwnd_segments) i_cwnd = min((V_tcp_initcwnd_segments * tp->t_maxseg), max(2 * tp->t_maxseg, 14600)); else if (V_tcp_do_rfc3390) i_cwnd = min(4 * tp->t_maxseg, max(2 * tp->t_maxseg, 4380)); else { /* Per RFC5681 Section 3.1 */ if (tp->t_maxseg > 2190) i_cwnd = 2 * tp->t_maxseg; else if (tp->t_maxseg > 1095) i_cwnd = 3 * tp->t_maxseg; else i_cwnd = 4 * tp->t_maxseg; } return (i_cwnd); } /* * Given a specified gain, return the target * cwnd based on that gain. */ static uint32_t bbr_get_raw_target_cwnd(struct tcp_bbr *bbr, uint32_t gain, uint64_t bw) { uint64_t bdp, rtt; uint32_t cwnd; if ((get_filter_value_small(&bbr->r_ctl.rc_rttprop) == 0xffffffff) || (bbr_get_full_bw(bbr) == 0)) { /* No measurements yet */ return (bbr_initial_cwnd(bbr, bbr->rc_tp)); } /* * Get bytes per RTT needed (rttProp is normally in * bbr_cwndtarget_rtt_touse) */ rtt = bbr_get_rtt(bbr, bbr_cwndtarget_rtt_touse); /* Get the bdp from the two values */ bdp = bbr_get_bw_delay_prod(rtt, bw); /* Now apply the gain */ cwnd = (uint32_t)(((bdp * ((uint64_t)gain)) + (uint64_t)(BBR_UNIT - 1)) / ((uint64_t)BBR_UNIT)); return (cwnd); } static uint32_t bbr_get_target_cwnd(struct tcp_bbr *bbr, uint64_t bw, uint32_t gain) { uint32_t cwnd, mss; mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs); /* Get the base cwnd with gain rounded to a mss */ cwnd = roundup(bbr_get_raw_target_cwnd(bbr, bw, gain), mss); /* * Add in N (2 default since we do not have a * fq layer to trap packets in) quanta's per the I-D * section 4.2.3.2 quanta adjust. */ cwnd += (bbr_quanta * bbr->r_ctl.rc_pace_max_segs); if (bbr->rc_use_google) { if((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) && (bbr_state_val(bbr) == BBR_SUB_GAIN)) { /* * The linux implementation adds * an extra 2 x mss in gain cycle which * is documented no-where except in the code. * so we add more for Neal undocumented feature */ cwnd += 2 * mss; } if ((cwnd / mss) & 0x1) { /* Round up for odd num mss */ cwnd += mss; } } /* Are we below the min cwnd? */ if (cwnd < get_min_cwnd(bbr)) return (get_min_cwnd(bbr)); return (cwnd); } static uint16_t bbr_gain_adjust(struct tcp_bbr *bbr, uint16_t gain) { if (gain < 1) gain = 1; return (gain); } static uint32_t bbr_get_header_oh(struct tcp_bbr *bbr) { int seg_oh; seg_oh = 0; if (bbr->r_ctl.rc_inc_tcp_oh) { /* Do we include TCP overhead? */ seg_oh = (bbr->rc_last_options + sizeof(struct tcphdr)); } if (bbr->r_ctl.rc_inc_ip_oh) { /* Do we include IP overhead? */ #ifdef INET6 if (bbr->r_is_v6) { seg_oh += sizeof(struct ip6_hdr); } else #endif { #ifdef INET seg_oh += sizeof(struct ip); #endif } } if (bbr->r_ctl.rc_inc_enet_oh) { /* Do we include the ethernet overhead? */ seg_oh += sizeof(struct ether_header); } return(seg_oh); } static uint32_t bbr_get_pacing_length(struct tcp_bbr *bbr, uint16_t gain, uint32_t useconds_time, uint64_t bw) { uint64_t divor, res, tim; if (useconds_time == 0) return (0); gain = bbr_gain_adjust(bbr, gain); divor = (uint64_t)USECS_IN_SECOND * (uint64_t)BBR_UNIT; tim = useconds_time; res = (tim * bw * gain) / divor; if (res == 0) res = 1; return ((uint32_t)res); } /* * Given a gain and a length return the delay in useconds that * should be used to evenly space out packets * on the connection (based on the gain factor). */ static uint32_t bbr_get_pacing_delay(struct tcp_bbr *bbr, uint16_t gain, int32_t len, uint32_t cts, int nolog) { uint64_t bw, lentim, res; uint32_t usecs, srtt, over = 0; uint32_t seg_oh, num_segs, maxseg; if (len == 0) return (0); maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options; num_segs = (len + maxseg - 1) / maxseg; if (bbr->rc_use_google == 0) { seg_oh = bbr_get_header_oh(bbr); len += (num_segs * seg_oh); } gain = bbr_gain_adjust(bbr, gain); bw = bbr_get_bw(bbr); if (bbr->rc_use_google) { uint64_t cbw; /* * Reduce the b/w by the google discount * factor 10 = 1%. */ cbw = bw * (uint64_t)(1000 - bbr->r_ctl.bbr_google_discount); cbw /= (uint64_t)1000; /* We don't apply a discount if it results in 0 */ if (cbw > 0) bw = cbw; } lentim = ((uint64_t)len * (uint64_t)USECS_IN_SECOND * (uint64_t)BBR_UNIT); res = lentim / ((uint64_t)gain * bw); if (res == 0) res = 1; usecs = (uint32_t)res; srtt = bbr_get_rtt(bbr, BBR_SRTT); if (bbr_hptsi_max_mul && bbr_hptsi_max_div && (bbr->rc_use_google == 0) && (usecs > ((srtt * bbr_hptsi_max_mul) / bbr_hptsi_max_div))) { /* * We cannot let the delay be more than 1/2 the srtt time. * Otherwise we cannot pace out or send properly. */ over = usecs = (srtt * bbr_hptsi_max_mul) / bbr_hptsi_max_div; BBR_STAT_INC(bbr_hpts_min_time); } if (!nolog) bbr_log_pacing_delay_calc(bbr, gain, len, cts, usecs, bw, over, 1); return (usecs); } static void bbr_ack_received(struct tcpcb *tp, struct tcp_bbr *bbr, struct tcphdr *th, uint32_t bytes_this_ack, uint32_t sack_changed, uint32_t prev_acked, int32_t line, uint32_t losses) { INP_WLOCK_ASSERT(tp->t_inpcb); uint64_t bw; uint32_t cwnd, target_cwnd, saved_bytes, maxseg; int32_t meth; #ifdef STATS if ((tp->t_flags & TF_GPUTINPROG) && SEQ_GEQ(th->th_ack, tp->gput_ack)) { /* * Strech acks and compressed acks will cause this to * oscillate but we are doing it the same way as the main * stack so it will be compariable (though possibly not * ideal). */ int32_t cgput; int64_t gput, time_stamp; gput = (int64_t) (th->th_ack - tp->gput_seq) * 8; time_stamp = max(1, ((bbr->r_ctl.rc_rcvtime - tp->gput_ts) / 1000)); cgput = gput / time_stamp; stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT, cgput); if (tp->t_stats_gput_prev > 0) stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_GPUT_ND, ((gput - tp->t_stats_gput_prev) * 100) / tp->t_stats_gput_prev); tp->t_flags &= ~TF_GPUTINPROG; tp->t_stats_gput_prev = cgput; } #endif if ((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) && ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google)) { /* We don't change anything in probe-rtt */ return; } maxseg = tp->t_maxseg - bbr->rc_last_options; saved_bytes = bytes_this_ack; bytes_this_ack += sack_changed; if (bytes_this_ack > prev_acked) { bytes_this_ack -= prev_acked; /* * A byte ack'd gives us a full mss * to be like linux i.e. they count packets. */ if ((bytes_this_ack < maxseg) && bbr->rc_use_google) bytes_this_ack = maxseg; } else { /* Unlikely */ bytes_this_ack = 0; } cwnd = tp->snd_cwnd; bw = get_filter_value(&bbr->r_ctl.rc_delrate); if (bw) target_cwnd = bbr_get_target_cwnd(bbr, bw, (uint32_t)bbr->r_ctl.rc_bbr_cwnd_gain); else target_cwnd = bbr_initial_cwnd(bbr, bbr->rc_tp); if (IN_RECOVERY(tp->t_flags) && (bbr->bbr_prev_in_rec == 0)) { /* * We are entering recovery and * thus packet conservation. */ bbr->pkt_conservation = 1; bbr->r_ctl.rc_recovery_start = bbr->r_ctl.rc_rcvtime; cwnd = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + bytes_this_ack; } if (IN_RECOVERY(tp->t_flags)) { uint32_t flight; bbr->bbr_prev_in_rec = 1; if (cwnd > losses) { cwnd -= losses; if (cwnd < maxseg) cwnd = maxseg; } else cwnd = maxseg; flight = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); bbr_log_type_cwndupd(bbr, flight, 0, losses, 10, 0, 0, line); if (bbr->pkt_conservation) { uint32_t time_in; if (TSTMP_GEQ(bbr->r_ctl.rc_rcvtime, bbr->r_ctl.rc_recovery_start)) time_in = bbr->r_ctl.rc_rcvtime - bbr->r_ctl.rc_recovery_start; else time_in = 0; if (time_in >= bbr_get_rtt(bbr, BBR_RTT_PROP)) { /* Clear packet conservation after an rttProp */ bbr->pkt_conservation = 0; } else { if ((flight + bytes_this_ack) > cwnd) cwnd = flight + bytes_this_ack; if (cwnd < get_min_cwnd(bbr)) cwnd = get_min_cwnd(bbr); tp->snd_cwnd = cwnd; bbr_log_type_cwndupd(bbr, saved_bytes, sack_changed, prev_acked, 1, target_cwnd, th->th_ack, line); return; } } } else bbr->bbr_prev_in_rec = 0; if ((bbr->rc_use_google == 0) && bbr->r_ctl.restrict_growth) { bbr->r_ctl.restrict_growth--; if (bytes_this_ack > maxseg) bytes_this_ack = maxseg; } if (bbr->rc_filled_pipe) { /* * Here we have exited startup and filled the pipe. We will * thus allow the cwnd to shrink to the target. We hit here * mostly. */ uint32_t s_cwnd; meth = 2; s_cwnd = min((cwnd + bytes_this_ack), target_cwnd); if (s_cwnd > cwnd) cwnd = s_cwnd; else if (bbr_cwnd_may_shrink || bbr->rc_use_google || bbr->rc_no_pacing) cwnd = s_cwnd; } else { /* * Here we are still in startup, we increase cwnd by what * has been acked. */ if ((cwnd < target_cwnd) || (bbr->rc_past_init_win == 0)) { meth = 3; cwnd += bytes_this_ack; } else { /* * Method 4 means we are at target so no gain in * startup and past the initial window. */ meth = 4; } } tp->snd_cwnd = max(cwnd, get_min_cwnd(bbr)); bbr_log_type_cwndupd(bbr, saved_bytes, sack_changed, prev_acked, meth, target_cwnd, th->th_ack, line); } static void tcp_bbr_partialack(struct tcpcb *tp) { struct tcp_bbr *bbr; bbr = (struct tcp_bbr *)tp->t_fb_ptr; INP_WLOCK_ASSERT(tp->t_inpcb); if (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <= tp->snd_cwnd) { bbr->r_wanted_output = 1; } } static void bbr_post_recovery(struct tcpcb *tp) { struct tcp_bbr *bbr; uint32_t flight; INP_WLOCK_ASSERT(tp->t_inpcb); bbr = (struct tcp_bbr *)tp->t_fb_ptr; /* * Here we just exit recovery. */ EXIT_RECOVERY(tp->t_flags); /* Lock in our b/w reduction for the specified number of pkt-epochs */ bbr->r_recovery_bw = 0; tp->snd_recover = tp->snd_una; tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime); bbr->pkt_conservation = 0; if (bbr->rc_use_google == 0) { /* * For non-google mode lets * go ahead and make sure we clear * the recovery state so if we * bounce back in to recovery we * will do PC. */ bbr->bbr_prev_in_rec = 0; } bbr_log_type_exit_rec(bbr); if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) { tp->snd_cwnd = max(tp->snd_cwnd, bbr->r_ctl.rc_cwnd_on_ent); bbr_log_type_cwndupd(bbr, 0, 0, 0, 15, 0, 0, __LINE__); } else { /* For probe-rtt case lets fix up its saved_cwnd */ if (bbr->r_ctl.rc_saved_cwnd < bbr->r_ctl.rc_cwnd_on_ent) { bbr->r_ctl.rc_saved_cwnd = bbr->r_ctl.rc_cwnd_on_ent; bbr_log_type_cwndupd(bbr, 0, 0, 0, 16, 0, 0, __LINE__); } } flight = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); if ((bbr->rc_use_google == 0) && bbr_do_red) { uint64_t val, lr2use; uint32_t maxseg, newcwnd, acks_inflight, ratio, cwnd; uint32_t *cwnd_p; if (bbr_get_rtt(bbr, BBR_SRTT)) { val = ((uint64_t)bbr_get_rtt(bbr, BBR_RTT_PROP) * (uint64_t)1000); val /= bbr_get_rtt(bbr, BBR_SRTT); ratio = (uint32_t)val; } else ratio = 1000; bbr_log_type_cwndupd(bbr, bbr_red_mul, bbr_red_div, bbr->r_ctl.recovery_lr, 21, ratio, bbr->r_ctl.rc_red_cwnd_pe, __LINE__); if ((ratio < bbr_do_red) || (bbr_do_red == 0)) goto done; if (((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) && bbr_prtt_slam_cwnd) || (bbr_sub_drain_slam_cwnd && (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) && bbr->rc_hit_state_1 && (bbr_state_val(bbr) == BBR_SUB_DRAIN)) || ((bbr->rc_bbr_state == BBR_STATE_DRAIN) && bbr_slam_cwnd_in_main_drain)) { /* * Here we must poke at the saved cwnd * as well as the cwnd. */ cwnd = bbr->r_ctl.rc_saved_cwnd; cwnd_p = &bbr->r_ctl.rc_saved_cwnd; } else { cwnd = tp->snd_cwnd; cwnd_p = &tp->snd_cwnd; } maxseg = tp->t_maxseg - bbr->rc_last_options; /* Add the overall lr with the recovery lr */ if (bbr->r_ctl.rc_lost == 0) lr2use = 0; else if (bbr->r_ctl.rc_delivered == 0) lr2use = 1000; else { lr2use = bbr->r_ctl.rc_lost * 1000; lr2use /= bbr->r_ctl.rc_delivered; } lr2use += bbr->r_ctl.recovery_lr; acks_inflight = (flight / (maxseg * 2)); if (bbr_red_scale) { lr2use *= bbr_get_rtt(bbr, BBR_SRTT); lr2use /= bbr_red_scale; if ((bbr_red_growth_restrict) && ((bbr_get_rtt(bbr, BBR_SRTT)/bbr_red_scale) > 1)) bbr->r_ctl.restrict_growth += acks_inflight; } if (lr2use) { val = (uint64_t)cwnd * lr2use; val /= 1000; if (cwnd > val) newcwnd = roundup((cwnd - val), maxseg); else newcwnd = maxseg; } else { val = (uint64_t)cwnd * (uint64_t)bbr_red_mul; val /= (uint64_t)bbr_red_div; newcwnd = roundup((uint32_t)val, maxseg); } /* with standard delayed acks how many acks can I expect? */ if (bbr_drop_limit == 0) { /* * Anticpate how much we will * raise the cwnd based on the acks. */ if ((newcwnd + (acks_inflight * maxseg)) < get_min_cwnd(bbr)) { /* We do enforce the min (with the acks) */ newcwnd = (get_min_cwnd(bbr) - acks_inflight); } } else { /* * A strict drop limit of N is inplace */ if (newcwnd < (bbr_drop_limit * maxseg)) { newcwnd = bbr_drop_limit * maxseg; } } /* For the next N acks do we restrict the growth */ *cwnd_p = newcwnd; if (tp->snd_cwnd > newcwnd) tp->snd_cwnd = newcwnd; bbr_log_type_cwndupd(bbr, bbr_red_mul, bbr_red_div, val, 22, (uint32_t)lr2use, bbr_get_rtt(bbr, BBR_SRTT), __LINE__); bbr->r_ctl.rc_red_cwnd_pe = bbr->r_ctl.rc_pkt_epoch; } done: bbr->r_ctl.recovery_lr = 0; if (flight <= tp->snd_cwnd) { bbr->r_wanted_output = 1; } tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime); } static void bbr_setup_red_bw(struct tcp_bbr *bbr, uint32_t cts) { bbr->r_ctl.red_bw = get_filter_value(&bbr->r_ctl.rc_delrate); /* Limit the drop in b/w to 1/2 our current filter. */ if (bbr->r_ctl.red_bw > bbr->r_ctl.rc_bbr_cur_del_rate) bbr->r_ctl.red_bw = bbr->r_ctl.rc_bbr_cur_del_rate; if (bbr->r_ctl.red_bw < (get_filter_value(&bbr->r_ctl.rc_delrate) / 2)) bbr->r_ctl.red_bw = get_filter_value(&bbr->r_ctl.rc_delrate) / 2; tcp_bbr_tso_size_check(bbr, cts); } static void bbr_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type, struct bbr_sendmap *rsm) { struct tcp_bbr *bbr; INP_WLOCK_ASSERT(tp->t_inpcb); #ifdef STATS stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type); #endif bbr = (struct tcp_bbr *)tp->t_fb_ptr; switch (type) { case CC_NDUPACK: if (!IN_RECOVERY(tp->t_flags)) { tp->snd_recover = tp->snd_max; /* Start a new epoch */ bbr_set_pktepoch(bbr, bbr->r_ctl.rc_rcvtime, __LINE__); if (bbr->rc_lt_is_sampling || bbr->rc_lt_use_bw) { /* * Move forward the lt epoch * so it won't count the truncated * epoch. */ bbr->r_ctl.rc_lt_epoch++; } if (bbr->rc_bbr_state == BBR_STATE_STARTUP) { /* * Just like the policer detection code * if we are in startup we must push * forward the last startup epoch * to hide the truncated PE. */ bbr->r_ctl.rc_bbr_last_startup_epoch++; } bbr->r_ctl.rc_cwnd_on_ent = tp->snd_cwnd; ENTER_RECOVERY(tp->t_flags); bbr->rc_tlp_rtx_out = 0; bbr->r_ctl.recovery_lr = bbr->r_ctl.rc_pkt_epoch_loss_rate; tcp_bbr_tso_size_check(bbr, bbr->r_ctl.rc_rcvtime); if (tcp_in_hpts(bbr->rc_inp) && ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK) == 0)) { /* * When we enter recovery, we need to restart * any timers. This may mean we gain an agg * early, which will be made up for at the last * rxt out. */ bbr->rc_timer_first = 1; bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); } /* * Calculate a new cwnd based on to the current * delivery rate with no gain. We get the bdp * without gaining it up like we normally would and * we use the last cur_del_rate. */ if ((bbr->rc_use_google == 0) && (bbr->r_ctl.bbr_rttprobe_gain_val || (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT))) { tp->snd_cwnd = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + (tp->t_maxseg - bbr->rc_last_options); if (tp->snd_cwnd < get_min_cwnd(bbr)) { /* We always gate to min cwnd */ tp->snd_cwnd = get_min_cwnd(bbr); } bbr_log_type_cwndupd(bbr, 0, 0, 0, 14, 0, 0, __LINE__); } bbr_log_type_enter_rec(bbr, rsm->r_start); } break; case CC_RTO_ERR: KMOD_TCPSTAT_INC(tcps_sndrexmitbad); /* RTO was unnecessary, so reset everything. */ bbr_reset_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime); if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) { tp->snd_cwnd = tp->snd_cwnd_prev; tp->snd_ssthresh = tp->snd_ssthresh_prev; tp->snd_recover = tp->snd_recover_prev; tp->snd_cwnd = max(tp->snd_cwnd, bbr->r_ctl.rc_cwnd_on_ent); bbr_log_type_cwndupd(bbr, 0, 0, 0, 13, 0, 0, __LINE__); } tp->t_badrxtwin = 0; break; } } /* * Indicate whether this ack should be delayed. We can delay the ack if * following conditions are met: * - There is no delayed ack timer in progress. * - Our last ack wasn't a 0-sized window. We never want to delay * the ack that opens up a 0-sized window. * - LRO wasn't used for this segment. We make sure by checking that the * segment size is not larger than the MSS. * - Delayed acks are enabled or this is a half-synchronized T/TCP * connection. * - The data being acked is less than a full segment (a stretch ack * of more than a segment we should ack. * - nsegs is 1 (if its more than that we received more than 1 ack). */ #define DELAY_ACK(tp, bbr, nsegs) \ (((tp->t_flags & TF_RXWIN0SENT) == 0) && \ ((tp->t_flags & TF_DELACK) == 0) && \ ((bbr->bbr_segs_rcvd + nsegs) < tp->t_delayed_ack) && \ (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN))) /* * Return the lowest RSM in the map of * packets still in flight that is not acked. * This should normally find on the first one * since we remove packets from the send * map after they are marked ACKED. */ static struct bbr_sendmap * bbr_find_lowest_rsm(struct tcp_bbr *bbr) { struct bbr_sendmap *rsm; /* * Walk the time-order transmitted list looking for an rsm that is * not acked. This will be the one that was sent the longest time * ago that is still outstanding. */ TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_tmap, r_tnext) { if (rsm->r_flags & BBR_ACKED) { continue; } goto finish; } finish: return (rsm); } static struct bbr_sendmap * bbr_find_high_nonack(struct tcp_bbr *bbr, struct bbr_sendmap *rsm) { struct bbr_sendmap *prsm; /* * Walk the sequence order list backward until we hit and arrive at * the highest seq not acked. In theory when this is called it * should be the last segment (which it was not). */ prsm = rsm; TAILQ_FOREACH_REVERSE_FROM(prsm, &bbr->r_ctl.rc_map, bbr_head, r_next) { if (prsm->r_flags & (BBR_ACKED | BBR_HAS_FIN)) { continue; } return (prsm); } return (NULL); } /* * Returns to the caller the number of microseconds that * the packet can be outstanding before we think we * should have had an ack returned. */ static uint32_t bbr_calc_thresh_rack(struct tcp_bbr *bbr, uint32_t srtt, uint32_t cts, struct bbr_sendmap *rsm) { /* * lro is the flag we use to determine if we have seen reordering. * If it gets set we have seen reordering. The reorder logic either * works in one of two ways: * * If reorder-fade is configured, then we track the last time we saw * re-ordering occur. If we reach the point where enough time as * passed we no longer consider reordering has occuring. * * Or if reorder-face is 0, then once we see reordering we consider * the connection to alway be subject to reordering and just set lro * to 1. * * In the end if lro is non-zero we add the extra time for * reordering in. */ int32_t lro; uint32_t thresh, t_rxtcur; if (srtt == 0) srtt = 1; if (bbr->r_ctl.rc_reorder_ts) { if (bbr->r_ctl.rc_reorder_fade) { if (SEQ_GEQ(cts, bbr->r_ctl.rc_reorder_ts)) { lro = cts - bbr->r_ctl.rc_reorder_ts; if (lro == 0) { /* * No time as passed since the last * reorder, mark it as reordering. */ lro = 1; } } else { /* Negative time? */ lro = 0; } if (lro > bbr->r_ctl.rc_reorder_fade) { /* Turn off reordering seen too */ bbr->r_ctl.rc_reorder_ts = 0; lro = 0; } } else { /* Reodering does not fade */ lro = 1; } } else { lro = 0; } thresh = srtt + bbr->r_ctl.rc_pkt_delay; if (lro) { /* It must be set, if not you get 1/4 rtt */ if (bbr->r_ctl.rc_reorder_shift) thresh += (srtt >> bbr->r_ctl.rc_reorder_shift); else thresh += (srtt >> 2); } else { thresh += 1000; } /* We don't let the rack timeout be above a RTO */ if ((bbr->rc_tp)->t_srtt == 0) t_rxtcur = BBR_INITIAL_RTO; else t_rxtcur = TICKS_2_USEC(bbr->rc_tp->t_rxtcur); if (thresh > t_rxtcur) { thresh = t_rxtcur; } /* And we don't want it above the RTO max either */ if (thresh > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) { thresh = (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND); } bbr_log_thresh_choice(bbr, cts, thresh, lro, srtt, rsm, BBR_TO_FRM_RACK); return (thresh); } /* * Return to the caller the amount of time in mico-seconds * that should be used for the TLP timer from the last * send time of this packet. */ static uint32_t bbr_calc_thresh_tlp(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t srtt, uint32_t cts) { uint32_t thresh, len, maxseg, t_rxtcur; struct bbr_sendmap *prsm; if (srtt == 0) srtt = 1; if (bbr->rc_tlp_threshold) thresh = srtt + (srtt / bbr->rc_tlp_threshold); else thresh = (srtt * 2); maxseg = tp->t_maxseg - bbr->rc_last_options; /* Get the previous sent packet, if any */ len = rsm->r_end - rsm->r_start; /* 2.1 behavior */ prsm = TAILQ_PREV(rsm, bbr_head, r_tnext); if (prsm && (len <= maxseg)) { /* * Two packets outstanding, thresh should be (2*srtt) + * possible inter-packet delay (if any). */ uint32_t inter_gap = 0; int idx, nidx; idx = rsm->r_rtr_cnt - 1; nidx = prsm->r_rtr_cnt - 1; if (TSTMP_GEQ(rsm->r_tim_lastsent[nidx], prsm->r_tim_lastsent[idx])) { /* Yes it was sent later (or at the same time) */ inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx]; } thresh += inter_gap; } else if (len <= maxseg) { /* * Possibly compensate for delayed-ack. */ uint32_t alt_thresh; alt_thresh = srtt + (srtt / 2) + bbr_delayed_ack_time; if (alt_thresh > thresh) thresh = alt_thresh; } /* Not above the current RTO */ if (tp->t_srtt == 0) t_rxtcur = BBR_INITIAL_RTO; else t_rxtcur = TICKS_2_USEC(tp->t_rxtcur); bbr_log_thresh_choice(bbr, cts, thresh, t_rxtcur, srtt, rsm, BBR_TO_FRM_TLP); /* Not above an RTO */ if (thresh > t_rxtcur) { thresh = t_rxtcur; } /* Not above a RTO max */ if (thresh > (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND)) { thresh = (((uint32_t)bbr->rc_max_rto_sec) * USECS_IN_SECOND); } /* And now apply the user TLP min */ if (thresh < bbr_tlp_min) { thresh = bbr_tlp_min; } return (thresh); } /* * Return one of three RTTs to use (in microseconds). */ static __inline uint32_t bbr_get_rtt(struct tcp_bbr *bbr, int32_t rtt_type) { uint32_t f_rtt; uint32_t srtt; f_rtt = get_filter_value_small(&bbr->r_ctl.rc_rttprop); if (get_filter_value_small(&bbr->r_ctl.rc_rttprop) == 0xffffffff) { /* We have no rtt at all */ if (bbr->rc_tp->t_srtt == 0) f_rtt = BBR_INITIAL_RTO; else f_rtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT); /* * Since we don't know how good the rtt is apply a * delayed-ack min */ if (f_rtt < bbr_delayed_ack_time) { f_rtt = bbr_delayed_ack_time; } } /* Take the filter version or last measured pkt-rtt */ if (rtt_type == BBR_RTT_PROP) { srtt = f_rtt; } else if (rtt_type == BBR_RTT_PKTRTT) { if (bbr->r_ctl.rc_pkt_epoch_rtt) { srtt = bbr->r_ctl.rc_pkt_epoch_rtt; } else { /* No pkt rtt yet */ srtt = f_rtt; } } else if (rtt_type == BBR_RTT_RACK) { srtt = bbr->r_ctl.rc_last_rtt; /* We need to add in any internal delay for our timer */ if (bbr->rc_ack_was_delayed) srtt += bbr->r_ctl.rc_ack_hdwr_delay; } else if (rtt_type == BBR_SRTT) { srtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT); } else { /* TSNH */ srtt = f_rtt; #ifdef BBR_INVARIANTS panic("Unknown rtt request type %d", rtt_type); #endif } return (srtt); } static int bbr_is_lost(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t cts) { uint32_t thresh; thresh = bbr_calc_thresh_rack(bbr, bbr_get_rtt(bbr, BBR_RTT_RACK), cts, rsm); if ((cts - rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]) >= thresh) { /* It is lost (past time) */ return (1); } return (0); } /* * Return a sendmap if we need to retransmit something. */ static struct bbr_sendmap * bbr_check_recovery_mode(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) { /* * Check to see that we don't need to fall into recovery. We will * need to do so if our oldest transmit is past the time we should * have had an ack. */ struct bbr_sendmap *rsm; int32_t idx; if (TAILQ_EMPTY(&bbr->r_ctl.rc_map)) { /* Nothing outstanding that we know of */ return (NULL); } rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); if (rsm == NULL) { /* Nothing in the transmit map */ return (NULL); } if (tp->t_flags & TF_SENTFIN) { /* Fin restricted, don't find anything once a fin is sent */ return (NULL); } if (rsm->r_flags & BBR_ACKED) { /* * Ok the first one is acked (this really should not happen * since we remove the from the tmap once they are acked) */ rsm = bbr_find_lowest_rsm(bbr); if (rsm == NULL) return (NULL); } idx = rsm->r_rtr_cnt - 1; if (SEQ_LEQ(cts, rsm->r_tim_lastsent[idx])) { /* Send timestamp is the same or less? can't be ready */ return (NULL); } /* Get our RTT time */ if (bbr_is_lost(bbr, rsm, cts) && ((rsm->r_dupack >= DUP_ACK_THRESHOLD) || (rsm->r_flags & BBR_SACK_PASSED))) { if ((rsm->r_flags & BBR_MARKED_LOST) == 0) { rsm->r_flags |= BBR_MARKED_LOST; bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start; bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start; } bbr_cong_signal(tp, NULL, CC_NDUPACK, rsm); #ifdef BBR_INVARIANTS if ((rsm->r_end - rsm->r_start) == 0) panic("tp:%p bbr:%p rsm:%p length is 0?", tp, bbr, rsm); #endif return (rsm); } return (NULL); } /* * RACK Timer, here we simply do logging and house keeping. * the normal bbr_output_wtime() function will call the * appropriate thing to check if we need to do a RACK retransmit. * We return 1, saying don't proceed with bbr_output_wtime only * when all timers have been stopped (destroyed PCB?). */ static int bbr_timeout_rack(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) { /* * This timer simply provides an internal trigger to send out data. * The check_recovery_mode call will see if there are needed * retransmissions, if so we will enter fast-recovery. The output * call may or may not do the same thing depending on sysctl * settings. */ uint32_t lost; if (bbr->rc_all_timers_stopped) { return (1); } if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) { /* Its not time yet */ return (0); } BBR_STAT_INC(bbr_to_tot); lost = bbr->r_ctl.rc_lost; if (bbr->r_state && (bbr->r_state != tp->t_state)) bbr_set_state(tp, bbr, 0); bbr_log_to_event(bbr, cts, BBR_TO_FRM_RACK); if (bbr->r_ctl.rc_resend == NULL) { /* Lets do the check here */ bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts); } if (bbr_policer_call_from_rack_to) bbr_lt_bw_sampling(bbr, cts, (bbr->r_ctl.rc_lost > lost)); bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK; return (0); } static __inline void bbr_clone_rsm(struct tcp_bbr *bbr, struct bbr_sendmap *nrsm, struct bbr_sendmap *rsm, uint32_t start) { int idx; nrsm->r_start = start; nrsm->r_end = rsm->r_end; nrsm->r_rtr_cnt = rsm->r_rtr_cnt; nrsm-> r_rtt_not_allowed = rsm->r_rtt_not_allowed; nrsm->r_flags = rsm->r_flags; /* We don't transfer forward the SYN flag */ nrsm->r_flags &= ~BBR_HAS_SYN; /* We move forward the FIN flag, not that this should happen */ rsm->r_flags &= ~BBR_HAS_FIN; nrsm->r_dupack = rsm->r_dupack; nrsm->r_rtr_bytes = 0; nrsm->r_is_gain = rsm->r_is_gain; nrsm->r_is_drain = rsm->r_is_drain; nrsm->r_delivered = rsm->r_delivered; nrsm->r_ts_valid = rsm->r_ts_valid; nrsm->r_del_ack_ts = rsm->r_del_ack_ts; nrsm->r_del_time = rsm->r_del_time; nrsm->r_app_limited = rsm->r_app_limited; nrsm->r_first_sent_time = rsm->r_first_sent_time; nrsm->r_flight_at_send = rsm->r_flight_at_send; /* We split a piece the lower section looses any just_ret flag. */ nrsm->r_bbr_state = rsm->r_bbr_state; for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; } rsm->r_end = nrsm->r_start; idx = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs); idx /= 8; /* Check if we got too small */ if ((rsm->r_is_smallmap == 0) && ((rsm->r_end - rsm->r_start) <= idx)) { bbr->r_ctl.rc_num_small_maps_alloced++; rsm->r_is_smallmap = 1; } /* Check the new one as well */ if ((nrsm->r_end - nrsm->r_start) <= idx) { bbr->r_ctl.rc_num_small_maps_alloced++; nrsm->r_is_smallmap = 1; } } static int bbr_sack_mergable(struct bbr_sendmap *at, uint32_t start, uint32_t end) { /* * Given a sack block defined by * start and end, and a current position * at. Return 1 if either side of at * would show that the block is mergable * to that side. A block to be mergable * must have overlap with the start/end * and be in the SACK'd state. */ struct bbr_sendmap *l_rsm; struct bbr_sendmap *r_rsm; /* first get the either side blocks */ l_rsm = TAILQ_PREV(at, bbr_head, r_next); r_rsm = TAILQ_NEXT(at, r_next); if (l_rsm && (l_rsm->r_flags & BBR_ACKED)) { /* Potentially mergeable */ if ((l_rsm->r_end == start) || (SEQ_LT(start, l_rsm->r_end) && SEQ_GT(end, l_rsm->r_end))) { /* * map blk |------| * sack blk |------| * * map blk |------| * sack blk |------| */ return (1); } } if (r_rsm && (r_rsm->r_flags & BBR_ACKED)) { /* Potentially mergeable */ if ((r_rsm->r_start == end) || (SEQ_LT(start, r_rsm->r_start) && SEQ_GT(end, r_rsm->r_start))) { /* * map blk |---------| * sack blk |----| * * map blk |---------| * sack blk |-------| */ return (1); } } return (0); } static struct bbr_sendmap * bbr_merge_rsm(struct tcp_bbr *bbr, struct bbr_sendmap *l_rsm, struct bbr_sendmap *r_rsm) { /* * We are merging two ack'd RSM's, * the l_rsm is on the left (lower seq * values) and the r_rsm is on the right * (higher seq value). The simplest way * to merge these is to move the right * one into the left. I don't think there * is any reason we need to try to find * the oldest (or last oldest retransmitted). */ l_rsm->r_end = r_rsm->r_end; if (l_rsm->r_dupack < r_rsm->r_dupack) l_rsm->r_dupack = r_rsm->r_dupack; if (r_rsm->r_rtr_bytes) l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes; if (r_rsm->r_in_tmap) { /* This really should not happen */ TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, r_rsm, r_tnext); } if (r_rsm->r_app_limited) l_rsm->r_app_limited = r_rsm->r_app_limited; /* Now the flags */ if (r_rsm->r_flags & BBR_HAS_FIN) l_rsm->r_flags |= BBR_HAS_FIN; if (r_rsm->r_flags & BBR_TLP) l_rsm->r_flags |= BBR_TLP; if (r_rsm->r_flags & BBR_RWND_COLLAPSED) l_rsm->r_flags |= BBR_RWND_COLLAPSED; if (r_rsm->r_flags & BBR_MARKED_LOST) { /* This really should not happen */ bbr->r_ctl.rc_lost_bytes -= r_rsm->r_end - r_rsm->r_start; } TAILQ_REMOVE(&bbr->r_ctl.rc_map, r_rsm, r_next); if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) { /* Transfer the split limit to the map we free */ r_rsm->r_limit_type = l_rsm->r_limit_type; l_rsm->r_limit_type = 0; } bbr_free(bbr, r_rsm); return(l_rsm); } /* * TLP Timer, here we simply setup what segment we want to * have the TLP expire on, the normal bbr_output_wtime() will then * send it out. * * We return 1, saying don't proceed with bbr_output_wtime only * when all timers have been stopped (destroyed PCB?). */ static int bbr_timeout_tlp(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) { /* * Tail Loss Probe. */ struct bbr_sendmap *rsm = NULL; struct socket *so; uint32_t amm; uint32_t out, avail; uint32_t maxseg; int collapsed_win = 0; if (bbr->rc_all_timers_stopped) { return (1); } if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) { /* Its not time yet */ return (0); } if (ctf_progress_timeout_check(tp, true)) { bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); return (-ETIMEDOUT); /* tcp_drop() */ } /* Did we somehow get into persists? */ if (bbr->rc_in_persist) { return (0); } if (bbr->r_state && (bbr->r_state != tp->t_state)) bbr_set_state(tp, bbr, 0); BBR_STAT_INC(bbr_tlp_tot); maxseg = tp->t_maxseg - bbr->rc_last_options; /* * A TLP timer has expired. We have been idle for 2 rtts. So we now * need to figure out how to force a full MSS segment out. */ so = tp->t_inpcb->inp_socket; avail = sbavail(&so->so_snd); out = ctf_outstanding(tp); if (out > tp->snd_wnd) { /* special case, we need a retransmission */ collapsed_win = 1; goto need_retran; } if (avail > out) { /* New data is available */ amm = avail - out; if (amm > maxseg) { amm = maxseg; } else if ((amm < maxseg) && ((tp->t_flags & TF_NODELAY) == 0)) { /* not enough to fill a MTU and no-delay is off */ goto need_retran; } /* Set the send-new override */ if ((out + amm) <= tp->snd_wnd) { bbr->rc_tlp_new_data = 1; } else { goto need_retran; } bbr->r_ctl.rc_tlp_seg_send_cnt = 0; bbr->r_ctl.rc_last_tlp_seq = tp->snd_max; bbr->r_ctl.rc_tlp_send = NULL; /* cap any slots */ BBR_STAT_INC(bbr_tlp_newdata); goto send; } need_retran: /* * Ok we need to arrange the last un-acked segment to be re-sent, or * optionally the first un-acked segment. */ if (collapsed_win == 0) { rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next); if (rsm && (BBR_ACKED | BBR_HAS_FIN)) { rsm = bbr_find_high_nonack(bbr, rsm); } if (rsm == NULL) { goto restore; } } else { /* * We must find the last segment * that was acceptable by the client. */ TAILQ_FOREACH_REVERSE(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) { if ((rsm->r_flags & BBR_RWND_COLLAPSED) == 0) { /* Found one */ break; } } if (rsm == NULL) { /* None? if so send the first */ rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); if (rsm == NULL) goto restore; } } if ((rsm->r_end - rsm->r_start) > maxseg) { /* * We need to split this the last segment in two. */ struct bbr_sendmap *nrsm; nrsm = bbr_alloc_full_limit(bbr); if (nrsm == NULL) { /* * We can't get memory to split, we can either just * not split it. Or retransmit the whole piece, lets * do the large send (BTLP :-) ). */ goto go_for_it; } bbr_clone_rsm(bbr, nrsm, rsm, (rsm->r_end - maxseg)); TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } rsm->r_flags &= (~BBR_HAS_FIN); rsm = nrsm; } go_for_it: bbr->r_ctl.rc_tlp_send = rsm; bbr->rc_tlp_rtx_out = 1; if (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq) { bbr->r_ctl.rc_tlp_seg_send_cnt++; tp->t_rxtshift++; } else { bbr->r_ctl.rc_last_tlp_seq = rsm->r_start; bbr->r_ctl.rc_tlp_seg_send_cnt = 1; } send: if (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend) { /* * Can't [re]/transmit a segment we have retranmitted the * max times. We need the retransmit timer to take over. */ restore: bbr->rc_tlp_new_data = 0; bbr->r_ctl.rc_tlp_send = NULL; if (rsm) rsm->r_flags &= ~BBR_TLP; BBR_STAT_INC(bbr_tlp_retran_fail); return (0); } else if (rsm) { rsm->r_flags |= BBR_TLP; } if (rsm && (rsm->r_start == bbr->r_ctl.rc_last_tlp_seq) && (bbr->r_ctl.rc_tlp_seg_send_cnt > bbr_tlp_max_resend)) { /* * We have retransmitted to many times for TLP. Switch to * the regular RTO timer */ goto restore; } bbr_log_to_event(bbr, cts, BBR_TO_FRM_TLP); bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP; return (0); } /* * Delayed ack Timer, here we simply need to setup the * ACK_NOW flag and remove the DELACK flag. From there * the output routine will send the ack out. * * We only return 1, saying don't proceed, if all timers * are stopped (destroyed PCB?). */ static int bbr_timeout_delack(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) { if (bbr->rc_all_timers_stopped) { return (1); } bbr_log_to_event(bbr, cts, BBR_TO_FRM_DELACK); tp->t_flags &= ~TF_DELACK; tp->t_flags |= TF_ACKNOW; KMOD_TCPSTAT_INC(tcps_delack); bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK; return (0); } /* * Here we send a KEEP-ALIVE like probe to the * peer, we do not send data. * * We only return 1, saying don't proceed, if all timers * are stopped (destroyed PCB?). */ static int bbr_timeout_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) { struct tcptemp *t_template; int32_t retval = 1; if (bbr->rc_all_timers_stopped) { return (1); } if (bbr->rc_in_persist == 0) return (0); KASSERT(tp->t_inpcb != NULL, ("%s: tp %p tp->t_inpcb == NULL", __func__, tp)); /* * Persistence timer into zero window. Force a byte to be output, if * possible. */ bbr_log_to_event(bbr, cts, BBR_TO_FRM_PERSIST); bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT; KMOD_TCPSTAT_INC(tcps_persisttimeo); /* * Have we exceeded the user specified progress time? */ if (ctf_progress_timeout_check(tp, true)) { bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); return (-ETIMEDOUT); /* tcp_drop() */ } /* * Hack: if the peer is dead/unreachable, we do not time out if the * window is closed. After a full backoff, drop the connection if * the idle time (no responses to probes) reaches the maximum * backoff that we would use if retransmitting. */ if (tp->t_rxtshift == TCP_MAXRXTSHIFT && (ticks - tp->t_rcvtime >= tcp_maxpersistidle || ticks - tp->t_rcvtime >= TCP_REXMTVAL(tp) * tcp_totbackoff)) { KMOD_TCPSTAT_INC(tcps_persistdrop); tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX); return (-ETIMEDOUT); /* tcp_drop() */ } if ((sbavail(&bbr->rc_inp->inp_socket->so_snd) == 0) && tp->snd_una == tp->snd_max) { bbr_exit_persist(tp, bbr, cts, __LINE__); retval = 0; goto out; } /* * If the user has closed the socket then drop a persisting * connection after a much reduced timeout. */ if (tp->t_state > TCPS_CLOSE_WAIT && (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) { KMOD_TCPSTAT_INC(tcps_persistdrop); tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX); return (-ETIMEDOUT); /* tcp_drop() */ } t_template = tcpip_maketemplate(bbr->rc_inp); if (t_template) { tcp_respond(tp, t_template->tt_ipgen, &t_template->tt_t, (struct mbuf *)NULL, tp->rcv_nxt, tp->snd_una - 1, 0); /* This sends an ack */ if (tp->t_flags & TF_DELACK) tp->t_flags &= ~TF_DELACK; free(t_template, M_TEMP); } if (tp->t_rxtshift < TCP_MAXRXTSHIFT) tp->t_rxtshift++; bbr_start_hpts_timer(bbr, tp, cts, 3, 0, 0); out: return (retval); } /* * If a keepalive goes off, we had no other timers * happening. We always return 1 here since this * routine either drops the connection or sends * out a segment with respond. */ static int bbr_timeout_keepalive(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) { struct tcptemp *t_template; struct inpcb *inp; if (bbr->rc_all_timers_stopped) { return (1); } bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP; inp = tp->t_inpcb; bbr_log_to_event(bbr, cts, BBR_TO_FRM_KEEP); /* * Keep-alive timer went off; send something or drop connection if * idle for too long. */ KMOD_TCPSTAT_INC(tcps_keeptimeo); if (tp->t_state < TCPS_ESTABLISHED) goto dropit; if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && tp->t_state <= TCPS_CLOSING) { if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp)) goto dropit; /* * Send a packet designed to force a response if the peer is * up and reachable: either an ACK if the connection is * still alive, or an RST if the peer has closed the * connection due to timeout or reboot. Using sequence * number tp->snd_una-1 causes the transmitted zero-length * segment to lie outside the receive window; by the * protocol spec, this requires the correspondent TCP to * respond. */ KMOD_TCPSTAT_INC(tcps_keepprobe); t_template = tcpip_maketemplate(inp); if (t_template) { tcp_respond(tp, t_template->tt_ipgen, &t_template->tt_t, (struct mbuf *)NULL, tp->rcv_nxt, tp->snd_una - 1, 0); free(t_template, M_TEMP); } } bbr_start_hpts_timer(bbr, tp, cts, 4, 0, 0); return (1); dropit: KMOD_TCPSTAT_INC(tcps_keepdrops); tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX); return (-ETIMEDOUT); /* tcp_drop() */ } /* * Retransmit helper function, clear up all the ack * flags and take care of important book keeping. */ static void bbr_remxt_tmr(struct tcpcb *tp) { /* * The retransmit timer went off, all sack'd blocks must be * un-acked. */ struct bbr_sendmap *rsm, *trsm = NULL; struct tcp_bbr *bbr; uint32_t cts, lost; bbr = (struct tcp_bbr *)tp->t_fb_ptr; cts = tcp_get_usecs(&bbr->rc_tv); lost = bbr->r_ctl.rc_lost; if (bbr->r_state && (bbr->r_state != tp->t_state)) bbr_set_state(tp, bbr, 0); TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { if (rsm->r_flags & BBR_ACKED) { uint32_t old_flags; rsm->r_dupack = 0; if (rsm->r_in_tmap == 0) { /* We must re-add it back to the tlist */ if (trsm == NULL) { TAILQ_INSERT_HEAD(&bbr->r_ctl.rc_tmap, rsm, r_tnext); } else { TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, trsm, rsm, r_tnext); } rsm->r_in_tmap = 1; } old_flags = rsm->r_flags; rsm->r_flags |= BBR_RXT_CLEARED; rsm->r_flags &= ~(BBR_ACKED | BBR_SACK_PASSED | BBR_WAS_SACKPASS); bbr_log_type_rsmclear(bbr, cts, rsm, old_flags, __LINE__); } else { if ((tp->t_state < TCPS_ESTABLISHED) && (rsm->r_start == tp->snd_una)) { /* * Special case for TCP FO. Where * we sent more data beyond the snd_max. * We don't mark that as lost and stop here. */ break; } if ((rsm->r_flags & BBR_MARKED_LOST) == 0) { bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start; bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start; } if (bbr_marks_rxt_sack_passed) { /* * With this option, we will rack out * in 1ms increments the rest of the packets. */ rsm->r_flags |= BBR_SACK_PASSED | BBR_MARKED_LOST; rsm->r_flags &= ~BBR_WAS_SACKPASS; } else { /* * With this option we only mark them lost * and remove all sack'd markings. We will run * another RXT or a TLP. This will cause * us to eventually send more based on what * ack's come in. */ rsm->r_flags |= BBR_MARKED_LOST; rsm->r_flags &= ~BBR_WAS_SACKPASS; rsm->r_flags &= ~BBR_SACK_PASSED; } } trsm = rsm; } bbr->r_ctl.rc_resend = TAILQ_FIRST(&bbr->r_ctl.rc_map); /* Clear the count (we just un-acked them) */ bbr_log_to_event(bbr, cts, BBR_TO_FRM_TMR); bbr->rc_tlp_new_data = 0; bbr->r_ctl.rc_tlp_seg_send_cnt = 0; /* zap the behindness on a rxt */ bbr->r_ctl.rc_hptsi_agg_delay = 0; bbr->r_agg_early_set = 0; bbr->r_ctl.rc_agg_early = 0; bbr->rc_tlp_rtx_out = 0; bbr->r_ctl.rc_sacked = 0; bbr->r_ctl.rc_sacklast = NULL; bbr->r_timer_override = 1; bbr_lt_bw_sampling(bbr, cts, (bbr->r_ctl.rc_lost > lost)); } /* * Re-transmit timeout! If we drop the PCB we will return 1, otherwise * we will setup to retransmit the lowest seq number outstanding. */ static int bbr_timeout_rxt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) { int32_t rexmt; int32_t retval = 0; bool isipv6; bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT; if (bbr->rc_all_timers_stopped) { return (1); } if (TCPS_HAVEESTABLISHED(tp->t_state) && (tp->snd_una == tp->snd_max)) { /* Nothing outstanding .. nothing to do */ return (0); } /* * Retransmission timer went off. Message has not been acked within * retransmit interval. Back off to a longer retransmit interval * and retransmit one segment. */ if (ctf_progress_timeout_check(tp, true)) { bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); return (-ETIMEDOUT); /* tcp_drop() */ } bbr_remxt_tmr(tp); if ((bbr->r_ctl.rc_resend == NULL) || ((bbr->r_ctl.rc_resend->r_flags & BBR_RWND_COLLAPSED) == 0)) { /* * If the rwnd collapsed on * the one we are retransmitting * it does not count against the * rxt count. */ tp->t_rxtshift++; } if (tp->t_rxtshift > TCP_MAXRXTSHIFT) { tp->t_rxtshift = TCP_MAXRXTSHIFT; KMOD_TCPSTAT_INC(tcps_timeoutdrop); tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN); /* XXXGL: previously t_softerror was casted to uint16_t */ MPASS(tp->t_softerror >= 0); retval = tp->t_softerror ? -tp->t_softerror : -ETIMEDOUT; return (retval); /* tcp_drop() */ } if (tp->t_state == TCPS_SYN_SENT) { /* * If the SYN was retransmitted, indicate CWND to be limited * to 1 segment in cc_conn_init(). */ tp->snd_cwnd = 1; } else if (tp->t_rxtshift == 1) { /* * first retransmit; record ssthresh and cwnd so they can be * recovered if this turns out to be a "bad" retransmit. A * retransmit is considered "bad" if an ACK for this segment * is received within RTT/2 interval; the assumption here is * that the ACK was already in flight. See "On Estimating * End-to-End Network Path Properties" by Allman and Paxson * for more details. */ tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options; if (!IN_RECOVERY(tp->t_flags)) { tp->snd_cwnd_prev = tp->snd_cwnd; tp->snd_ssthresh_prev = tp->snd_ssthresh; tp->snd_recover_prev = tp->snd_recover; tp->t_badrxtwin = ticks + (tp->t_srtt >> (TCP_RTT_SHIFT + 1)); tp->t_flags |= TF_PREVVALID; } else { tp->t_flags &= ~TF_PREVVALID; } tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options; } else { tp->snd_cwnd = tp->t_maxseg - bbr->rc_last_options; tp->t_flags &= ~TF_PREVVALID; } KMOD_TCPSTAT_INC(tcps_rexmttimeo); if ((tp->t_state == TCPS_SYN_SENT) || (tp->t_state == TCPS_SYN_RECEIVED)) rexmt = USEC_2_TICKS(BBR_INITIAL_RTO) * tcp_backoff[tp->t_rxtshift]; else rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift]; TCPT_RANGESET(tp->t_rxtcur, rexmt, MSEC_2_TICKS(bbr->r_ctl.rc_min_rto_ms), MSEC_2_TICKS(((uint32_t)bbr->rc_max_rto_sec) * 1000)); /* * We enter the path for PLMTUD if connection is established or, if * connection is FIN_WAIT_1 status, reason for the last is that if * amount of data we send is very small, we could send it in couple * of packets and process straight to FIN. In that case we won't * catch ESTABLISHED state. */ #ifdef INET6 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? true : false; #else isipv6 = false; #endif if (((V_tcp_pmtud_blackhole_detect == 1) || (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) || (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) && ((tp->t_state == TCPS_ESTABLISHED) || (tp->t_state == TCPS_FIN_WAIT_1))) { /* * Idea here is that at each stage of mtu probe (usually, * 1448 -> 1188 -> 524) should be given 2 chances to recover * before further clamping down. 'tp->t_rxtshift % 2 == 0' * should take care of that. */ if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) == (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) && (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 && tp->t_rxtshift % 2 == 0)) { /* * Enter Path MTU Black-hole Detection mechanism: - * Disable Path MTU Discovery (IP "DF" bit). - * Reduce MTU to lower value than what we negotiated * with peer. */ if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) { /* * Record that we may have found a black * hole. */ tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE; /* Keep track of previous MSS. */ tp->t_pmtud_saved_maxseg = tp->t_maxseg; } /* * Reduce the MSS to blackhole value or to the * default in an attempt to retransmit. */ #ifdef INET6 isipv6 = bbr->r_is_v6; if (isipv6 && tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) { /* Use the sysctl tuneable blackhole MSS. */ tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss; KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated); } else if (isipv6) { /* Use the default MSS. */ tp->t_maxseg = V_tcp_v6mssdflt; /* * Disable Path MTU Discovery when we switch * to minmss. */ tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); } #endif #if defined(INET6) && defined(INET) else #endif #ifdef INET if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) { /* Use the sysctl tuneable blackhole MSS. */ tp->t_maxseg = V_tcp_pmtud_blackhole_mss; KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated); } else { /* Use the default MSS. */ tp->t_maxseg = V_tcp_mssdflt; /* * Disable Path MTU Discovery when we switch * to minmss. */ tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); } #endif } else { /* * If further retransmissions are still unsuccessful * with a lowered MTU, maybe this isn't a blackhole * and we restore the previous MSS and blackhole * detection flags. The limit '6' is determined by * giving each probe stage (1448, 1188, 524) 2 * chances to recover. */ if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) && (tp->t_rxtshift >= 6)) { tp->t_flags2 |= TF2_PLPMTU_PMTUD; tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE; tp->t_maxseg = tp->t_pmtud_saved_maxseg; KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed); } } } /* * Disable RFC1323 and SACK if we haven't got any response to our * third SYN to work-around some broken terminal servers (most of * which have hopefully been retired) that have bad VJ header * compression code which trashes TCP segments containing * unknown-to-them TCP options. */ if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) && (tp->t_rxtshift == 3)) tp->t_flags &= ~(TF_REQ_SCALE | TF_REQ_TSTMP | TF_SACK_PERMIT); /* * If we backed off this far, our srtt estimate is probably bogus. * Clobber it so we'll take the next rtt measurement as our srtt; * move the current srtt into rttvar to keep the current retransmit * times until then. */ if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) { #ifdef INET6 if (bbr->r_is_v6) in6_losing(tp->t_inpcb); else #endif in_losing(tp->t_inpcb); tp->t_rttvar += (tp->t_srtt >> TCP_RTT_SHIFT); tp->t_srtt = 0; } sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una); tp->snd_recover = tp->snd_max; tp->t_flags |= TF_ACKNOW; tp->t_rtttime = 0; return (retval); } static int bbr_process_timers(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, uint8_t hpts_calling) { int32_t ret = 0; int32_t timers = (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK); if (timers == 0) { return (0); } if (tp->t_state == TCPS_LISTEN) { /* no timers on listen sockets */ if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) return (0); return (1); } if (TSTMP_LT(cts, bbr->r_ctl.rc_timer_exp)) { uint32_t left; if (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) { ret = -1; bbr_log_to_processing(bbr, cts, ret, 0, hpts_calling); return (0); } if (hpts_calling == 0) { ret = -2; bbr_log_to_processing(bbr, cts, ret, 0, hpts_calling); return (0); } /* * Ok our timer went off early and we are not paced false * alarm, go back to sleep. */ left = bbr->r_ctl.rc_timer_exp - cts; ret = -3; bbr_log_to_processing(bbr, cts, ret, left, hpts_calling); tcp_hpts_insert(tp->t_inpcb, HPTS_USEC_TO_SLOTS(left)); return (1); } bbr->rc_tmr_stopped = 0; bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK; if (timers & PACE_TMR_DELACK) { ret = bbr_timeout_delack(tp, bbr, cts); } else if (timers & PACE_TMR_PERSIT) { ret = bbr_timeout_persist(tp, bbr, cts); } else if (timers & PACE_TMR_RACK) { bbr->r_ctl.rc_tlp_rxt_last_time = cts; ret = bbr_timeout_rack(tp, bbr, cts); } else if (timers & PACE_TMR_TLP) { bbr->r_ctl.rc_tlp_rxt_last_time = cts; ret = bbr_timeout_tlp(tp, bbr, cts); } else if (timers & PACE_TMR_RXT) { bbr->r_ctl.rc_tlp_rxt_last_time = cts; ret = bbr_timeout_rxt(tp, bbr, cts); } else if (timers & PACE_TMR_KEEP) { ret = bbr_timeout_keepalive(tp, bbr, cts); } bbr_log_to_processing(bbr, cts, ret, timers, hpts_calling); return (ret); } static void bbr_timer_cancel(struct tcp_bbr *bbr, int32_t line, uint32_t cts) { if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { uint8_t hpts_removed = 0; if (tcp_in_hpts(bbr->rc_inp) && (bbr->rc_timer_first == 1)) { /* * If we are canceling timer's when we have the * timer ahead of the output being paced. We also * must remove ourselves from the hpts. */ hpts_removed = 1; tcp_hpts_remove(bbr->rc_inp); if (bbr->r_ctl.rc_last_delay_val) { /* Update the last hptsi delay too */ uint32_t time_since_send; if (TSTMP_GT(cts, bbr->rc_pacer_started)) time_since_send = cts - bbr->rc_pacer_started; else time_since_send = 0; if (bbr->r_ctl.rc_last_delay_val > time_since_send) { /* Cut down our slot time */ bbr->r_ctl.rc_last_delay_val -= time_since_send; } else { bbr->r_ctl.rc_last_delay_val = 0; } bbr->rc_pacer_started = cts; } } bbr->rc_timer_first = 0; bbr_log_to_cancel(bbr, line, cts, hpts_removed); bbr->rc_tmr_stopped = bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK; bbr->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK); } } static void bbr_timer_stop(struct tcpcb *tp, uint32_t timer_type) { struct tcp_bbr *bbr; bbr = (struct tcp_bbr *)tp->t_fb_ptr; bbr->rc_all_timers_stopped = 1; return; } /* * stop all timers always returning 0. */ static int bbr_stopall(struct tcpcb *tp) { return (0); } static void bbr_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta) { return; } /* * return true if a bbr timer (rack or tlp) is active. */ static int bbr_timer_active(struct tcpcb *tp, uint32_t timer_type) { return (0); } static uint32_t bbr_get_earliest_send_outstanding(struct tcp_bbr *bbr, struct bbr_sendmap *u_rsm, uint32_t cts) { struct bbr_sendmap *rsm; rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); if ((rsm == NULL) || (u_rsm == rsm)) return (cts); return(rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]); } static void bbr_update_rsm(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t cts, uint32_t pacing_time) { int32_t idx; rsm->r_rtr_cnt++; rsm->r_dupack = 0; if (rsm->r_rtr_cnt > BBR_NUM_OF_RETRANS) { rsm->r_rtr_cnt = BBR_NUM_OF_RETRANS; rsm->r_flags |= BBR_OVERMAX; } if (rsm->r_flags & BBR_RWND_COLLAPSED) { /* Take off the collapsed flag at rxt */ rsm->r_flags &= ~BBR_RWND_COLLAPSED; } if (rsm->r_flags & BBR_MARKED_LOST) { /* We have retransmitted, its no longer lost */ rsm->r_flags &= ~BBR_MARKED_LOST; bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; } if (rsm->r_flags & BBR_RXT_CLEARED) { /* * We hit a RXT timer on it and * we cleared the "acked" flag. * We now have it going back into * flight, we can remove the cleared * flag and possibly do accounting on * this piece. */ rsm->r_flags &= ~BBR_RXT_CLEARED; } if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & BBR_TLP) == 0)) { bbr->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start); rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start); } idx = rsm->r_rtr_cnt - 1; rsm->r_tim_lastsent[idx] = cts; rsm->r_pacing_delay = pacing_time; rsm->r_delivered = bbr->r_ctl.rc_delivered; rsm->r_ts_valid = bbr->rc_ts_valid; if (bbr->rc_ts_valid) rsm->r_del_ack_ts = bbr->r_ctl.last_inbound_ts; if (bbr->r_ctl.r_app_limited_until) rsm->r_app_limited = 1; else rsm->r_app_limited = 0; if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) rsm->r_bbr_state = bbr_state_val(bbr); else rsm->r_bbr_state = 8; if (rsm->r_flags & BBR_ACKED) { /* Problably MTU discovery messing with us */ uint32_t old_flags; old_flags = rsm->r_flags; rsm->r_flags &= ~BBR_ACKED; bbr_log_type_rsmclear(bbr, cts, rsm, old_flags, __LINE__); bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); if (bbr->r_ctl.rc_sacked == 0) bbr->r_ctl.rc_sacklast = NULL; } if (rsm->r_in_tmap) { TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); } TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 1; if (rsm->r_flags & BBR_SACK_PASSED) { /* We have retransmitted due to the SACK pass */ rsm->r_flags &= ~BBR_SACK_PASSED; rsm->r_flags |= BBR_WAS_SACKPASS; } rsm->r_first_sent_time = bbr_get_earliest_send_outstanding(bbr, rsm, cts); rsm->r_flight_at_send = ctf_flight_size(bbr->rc_tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); bbr->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next); if (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT) { rsm->r_is_gain = 1; rsm->r_is_drain = 0; } else if (bbr->r_ctl.rc_bbr_hptsi_gain < BBR_UNIT) { rsm->r_is_drain = 1; rsm->r_is_gain = 0; } else { rsm->r_is_drain = 0; rsm->r_is_gain = 0; } rsm->r_del_time = bbr->r_ctl.rc_del_time; /* TEMP GOOGLE CODE */ } /* * Returns 0, or the sequence where we stopped * updating. We also update the lenp to be the amount * of data left. */ static uint32_t bbr_update_entry(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t cts, int32_t *lenp, uint32_t pacing_time) { /* * We (re-)transmitted starting at rsm->r_start for some length * (possibly less than r_end. */ struct bbr_sendmap *nrsm; uint32_t c_end; int32_t len; len = *lenp; c_end = rsm->r_start + len; if (SEQ_GEQ(c_end, rsm->r_end)) { /* * We retransmitted the whole piece or more than the whole * slopping into the next rsm. */ bbr_update_rsm(tp, bbr, rsm, cts, pacing_time); if (c_end == rsm->r_end) { *lenp = 0; return (0); } else { int32_t act_len; /* Hangs over the end return whats left */ act_len = rsm->r_end - rsm->r_start; *lenp = (len - act_len); return (rsm->r_end); } /* We don't get out of this block. */ } /* * Here we retransmitted less than the whole thing which means we * have to split this into what was transmitted and what was not. */ nrsm = bbr_alloc_full_limit(bbr); if (nrsm == NULL) { *lenp = 0; return (0); } /* * So here we are going to take the original rsm and make it what we * retransmitted. nrsm will be the tail portion we did not * retransmit. For example say the chunk was 1, 11 (10 bytes). And * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to * 1, 6 and the new piece will be 6, 11. */ bbr_clone_rsm(bbr, nrsm, rsm, c_end); TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); nrsm->r_dupack = 0; if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } rsm->r_flags &= (~BBR_HAS_FIN); bbr_update_rsm(tp, bbr, rsm, cts, pacing_time); *lenp = 0; return (0); } static uint64_t bbr_get_hardware_rate(struct tcp_bbr *bbr) { uint64_t bw; bw = bbr_get_bw(bbr); bw *= (uint64_t)bbr_hptsi_gain[BBR_SUB_GAIN]; bw /= (uint64_t)BBR_UNIT; return(bw); } static void bbr_setup_less_of_rate(struct tcp_bbr *bbr, uint32_t cts, uint64_t act_rate, uint64_t rate_wanted) { /* * We could not get a full gains worth * of rate. */ if (get_filter_value(&bbr->r_ctl.rc_delrate) >= act_rate) { /* we can't even get the real rate */ uint64_t red; bbr->skip_gain = 1; bbr->gain_is_limited = 0; red = get_filter_value(&bbr->r_ctl.rc_delrate) - act_rate; if (red) filter_reduce_by(&bbr->r_ctl.rc_delrate, red, cts); } else { /* We can use a lower gain */ bbr->skip_gain = 0; bbr->gain_is_limited = 1; } } static void bbr_update_hardware_pacing_rate(struct tcp_bbr *bbr, uint32_t cts) { const struct tcp_hwrate_limit_table *nrte; int error, rate = -1; if (bbr->r_ctl.crte == NULL) return; if ((bbr->rc_inp->inp_route.ro_nh == NULL) || (bbr->rc_inp->inp_route.ro_nh->nh_ifp == NULL)) { /* Lost our routes? */ /* Clear the way for a re-attempt */ bbr->bbr_attempt_hdwr_pace = 0; lost_rate: bbr->gain_is_limited = 0; bbr->skip_gain = 0; bbr->bbr_hdrw_pacing = 0; counter_u64_add(bbr_flows_whdwr_pacing, -1); counter_u64_add(bbr_flows_nohdwr_pacing, 1); tcp_bbr_tso_size_check(bbr, cts); return; } rate = bbr_get_hardware_rate(bbr); nrte = tcp_chg_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp, bbr->rc_inp->inp_route.ro_nh->nh_ifp, rate, (RS_PACING_GEQ|RS_PACING_SUB_OK), &error, NULL); if (nrte == NULL) { goto lost_rate; } if (nrte != bbr->r_ctl.crte) { bbr->r_ctl.crte = nrte; if (error == 0) { BBR_STAT_INC(bbr_hdwr_rl_mod_ok); if (bbr->r_ctl.crte->rate < rate) { /* We have a problem */ bbr_setup_less_of_rate(bbr, cts, bbr->r_ctl.crte->rate, rate); } else { /* We are good */ bbr->gain_is_limited = 0; bbr->skip_gain = 0; } } else { /* A failure should release the tag */ BBR_STAT_INC(bbr_hdwr_rl_mod_fail); bbr->gain_is_limited = 0; bbr->skip_gain = 0; bbr->bbr_hdrw_pacing = 0; } bbr_type_log_hdwr_pacing(bbr, bbr->r_ctl.crte->ptbl->rs_ifp, rate, ((bbr->r_ctl.crte == NULL) ? 0 : bbr->r_ctl.crte->rate), __LINE__, cts, error); } } static void bbr_adjust_for_hw_pacing(struct tcp_bbr *bbr, uint32_t cts) { /* * If we have hardware pacing support * we need to factor that in for our * TSO size. */ const struct tcp_hwrate_limit_table *rlp; uint32_t cur_delay, seg_sz, maxseg, new_tso, delta, hdwr_delay; if ((bbr->bbr_hdrw_pacing == 0) || (IN_RECOVERY(bbr->rc_tp->t_flags)) || (bbr->r_ctl.crte == NULL)) return; if (bbr->hw_pacing_set == 0) { /* Not yet by the hdwr pacing count delay */ return; } if (bbr_hdwr_pace_adjust == 0) { /* No adjustment */ return; } rlp = bbr->r_ctl.crte; if (bbr->rc_tp->t_maxseg > bbr->rc_last_options) maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options; else maxseg = BBR_MIN_SEG - bbr->rc_last_options; /* * So lets first get the * time we will take between * TSO sized sends currently without * hardware help. */ cur_delay = bbr_get_pacing_delay(bbr, BBR_UNIT, bbr->r_ctl.rc_pace_max_segs, cts, 1); hdwr_delay = bbr->r_ctl.rc_pace_max_segs / maxseg; hdwr_delay *= rlp->time_between; if (cur_delay > hdwr_delay) delta = cur_delay - hdwr_delay; else delta = 0; bbr_log_type_tsosize(bbr, cts, delta, cur_delay, hdwr_delay, (bbr->r_ctl.rc_pace_max_segs / maxseg), 1); if (delta && (delta < (max(rlp->time_between, bbr->r_ctl.bbr_hptsi_segments_delay_tar)))) { /* * Now lets divide by the pacing * time between each segment the * hardware sends rounding up and * derive a bytes from that. We multiply * that by bbr_hdwr_pace_adjust to get * more bang for our buck. * * The goal is to have the software pacer * waiting no more than an additional * pacing delay if we can (without the * compensation i.e. x bbr_hdwr_pace_adjust). */ seg_sz = max(((cur_delay + rlp->time_between)/rlp->time_between), (bbr->r_ctl.rc_pace_max_segs/maxseg)); seg_sz *= bbr_hdwr_pace_adjust; if (bbr_hdwr_pace_floor && (seg_sz < bbr->r_ctl.crte->ptbl->rs_min_seg)) { /* Currently hardware paces * out rs_min_seg segments at a time. * We need to make sure we always send at least * a full burst of bbr_hdwr_pace_floor down. */ seg_sz = bbr->r_ctl.crte->ptbl->rs_min_seg; } seg_sz *= maxseg; } else if (delta == 0) { /* * The highest pacing rate is * above our b/w gained. This means * we probably are going quite fast at * the hardware highest rate. Lets just multiply * the calculated TSO size by the * multiplier factor (its probably * 4 segments in the default config for * mlx). */ seg_sz = bbr->r_ctl.rc_pace_max_segs * bbr_hdwr_pace_adjust; if (bbr_hdwr_pace_floor && (seg_sz < bbr->r_ctl.crte->ptbl->rs_min_seg)) { /* Currently hardware paces * out rs_min_seg segments at a time. * We need to make sure we always send at least * a full burst of bbr_hdwr_pace_floor down. */ seg_sz = bbr->r_ctl.crte->ptbl->rs_min_seg; } } else { /* * The pacing time difference is so * big that the hardware will * pace out more rapidly then we * really want and then we * will have a long delay. Lets just keep * the same TSO size so its as if * we were not using hdwr pacing (we * just gain a bit of spacing from the * hardware if seg_sz > 1). */ seg_sz = bbr->r_ctl.rc_pace_max_segs; } if (seg_sz > bbr->r_ctl.rc_pace_max_segs) new_tso = seg_sz; else new_tso = bbr->r_ctl.rc_pace_max_segs; if (new_tso >= (PACE_MAX_IP_BYTES-maxseg)) new_tso = PACE_MAX_IP_BYTES - maxseg; if (new_tso != bbr->r_ctl.rc_pace_max_segs) { bbr_log_type_tsosize(bbr, cts, new_tso, 0, bbr->r_ctl.rc_pace_max_segs, maxseg, 0); bbr->r_ctl.rc_pace_max_segs = new_tso; } } static void tcp_bbr_tso_size_check(struct tcp_bbr *bbr, uint32_t cts) { uint64_t bw; uint32_t old_tso = 0, new_tso; uint32_t maxseg, bytes; uint32_t tls_seg=0; /* * Google/linux uses the following algorithm to determine * the TSO size based on the b/w of the link (from Neal Cardwell email 9/27/18): * * bytes = bw_in_bytes_per_second / 1000 * bytes = min(bytes, 64k) * tso_segs = bytes / MSS * if (bw < 1.2Mbs) * min_tso_segs = 1 * else * min_tso_segs = 2 * tso_segs = max(tso_segs, min_tso_segs) * * * Note apply a device specific limit (we apply this in the * tcp_m_copym). * Note that before the initial measurement is made google bursts out * a full iwnd just like new-reno/cubic. * * We do not use this algorithm. Instead we * use a two phased approach: * * if ( bw <= per-tcb-cross-over) * goal_tso = calculate how much with this bw we * can send in goal-time seconds. * if (goal_tso > mss) * seg = goal_tso / mss * tso = seg * mss * else * tso = mss * if (tso > per-tcb-max) * tso = per-tcb-max * else if ( bw > 512Mbps) * tso = max-tso (64k/mss) * else * goal_tso = bw / per-tcb-divsor * seg = (goal_tso + mss-1)/mss * tso = seg * mss * * if (tso < per-tcb-floor) * tso = per-tcb-floor * if (tso > per-tcb-utter_max) * tso = per-tcb-utter_max * * Note the default per-tcb-divisor is 1000 (same as google). * the goal cross over is 30Mbps however. To recreate googles * algorithm you need to set: * * cross-over = 23,168,000 bps * goal-time = 18000 * per-tcb-max = 2 * per-tcb-divisor = 1000 * per-tcb-floor = 1 * * This will get you "google bbr" behavior with respect to tso size. * * Note we do set anything TSO size until we are past the initial * window. Before that we gnerally use either a single MSS * or we use the full IW size (so we burst a IW at a time) */ if (bbr->rc_tp->t_maxseg > bbr->rc_last_options) { maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options; } else { maxseg = BBR_MIN_SEG - bbr->rc_last_options; } old_tso = bbr->r_ctl.rc_pace_max_segs; if (bbr->rc_past_init_win == 0) { /* * Not enough data has been acknowledged to make a * judgement. Set up the initial TSO based on if we * are sending a full IW at once or not. */ if (bbr->rc_use_google) bbr->r_ctl.rc_pace_max_segs = ((bbr->rc_tp->t_maxseg - bbr->rc_last_options) * 2); else if (bbr->bbr_init_win_cheat) bbr->r_ctl.rc_pace_max_segs = bbr_initial_cwnd(bbr, bbr->rc_tp); else bbr->r_ctl.rc_pace_max_segs = bbr->rc_tp->t_maxseg - bbr->rc_last_options; if (bbr->r_ctl.rc_pace_min_segs != bbr->rc_tp->t_maxseg) bbr->r_ctl.rc_pace_min_segs = bbr->rc_tp->t_maxseg; if (bbr->r_ctl.rc_pace_max_segs == 0) { bbr->r_ctl.rc_pace_max_segs = maxseg; } bbr_log_type_tsosize(bbr, cts, bbr->r_ctl.rc_pace_max_segs, tls_seg, old_tso, maxseg, 0); bbr_adjust_for_hw_pacing(bbr, cts); return; } /** * Now lets set the TSO goal based on our delivery rate in * bytes per second. Note we only do this if * we have acked at least the initial cwnd worth of data. */ bw = bbr_get_bw(bbr); if (IN_RECOVERY(bbr->rc_tp->t_flags) && (bbr->rc_use_google == 0)) { /* We clamp to one MSS in recovery */ new_tso = maxseg; } else if (bbr->rc_use_google) { int min_tso_segs; /* Google considers the gain too */ if (bbr->r_ctl.rc_bbr_hptsi_gain != BBR_UNIT) { bw *= bbr->r_ctl.rc_bbr_hptsi_gain; bw /= BBR_UNIT; } bytes = bw / 1024; if (bytes > (64 * 1024)) bytes = 64 * 1024; new_tso = bytes / maxseg; if (bw < ONE_POINT_TWO_MEG) min_tso_segs = 1; else min_tso_segs = 2; if (new_tso < min_tso_segs) new_tso = min_tso_segs; new_tso *= maxseg; } else if (bbr->rc_no_pacing) { new_tso = (PACE_MAX_IP_BYTES / maxseg) * maxseg; } else if (bw <= bbr->r_ctl.bbr_cross_over) { /* * Calculate the worse case b/w TSO if we are inserting no * more than a delay_target number of TSO's. */ uint32_t tso_len, min_tso; tso_len = bbr_get_pacing_length(bbr, BBR_UNIT, bbr->r_ctl.bbr_hptsi_segments_delay_tar, bw); if (tso_len > maxseg) { new_tso = tso_len / maxseg; if (new_tso > bbr->r_ctl.bbr_hptsi_segments_max) new_tso = bbr->r_ctl.bbr_hptsi_segments_max; new_tso *= maxseg; } else { /* * less than a full sized frame yikes.. long rtt or * low bw? */ min_tso = bbr_minseg(bbr); if ((tso_len > min_tso) && (bbr_all_get_min == 0)) new_tso = rounddown(tso_len, min_tso); else new_tso = min_tso; } } else if (bw > FIVETWELVE_MBPS) { /* * This guy is so fast b/w wise that we can TSO as large as * possible of segments that the NIC will allow. */ new_tso = rounddown(PACE_MAX_IP_BYTES, maxseg); } else { /* * This formula is based on attempting to send a segment or * more every bbr_hptsi_per_second. The default is 1000 * which means you are targeting what you can send every 1ms * based on the peers bw. * * If the number drops to say 500, then you are looking more * at 2ms and you will raise how much we send in a single * TSO thus saving CPU (less bbr_output_wtime() calls). The * trade off of course is you will send more at once and * thus tend to clump up the sends into larger "bursts" * building a queue. */ bw /= bbr->r_ctl.bbr_hptsi_per_second; new_tso = roundup(bw, (uint64_t)maxseg); /* * Gate the floor to match what our lower than 48Mbps * algorithm does. The ceiling (bbr_hptsi_segments_max) thus * becomes the floor for this calculation. */ if (new_tso < (bbr->r_ctl.bbr_hptsi_segments_max * maxseg)) new_tso = (bbr->r_ctl.bbr_hptsi_segments_max * maxseg); } if (bbr->r_ctl.bbr_hptsi_segments_floor && (new_tso < (maxseg * bbr->r_ctl.bbr_hptsi_segments_floor))) new_tso = maxseg * bbr->r_ctl.bbr_hptsi_segments_floor; if (new_tso > PACE_MAX_IP_BYTES) new_tso = rounddown(PACE_MAX_IP_BYTES, maxseg); /* Enforce an utter maximum. */ if (bbr->r_ctl.bbr_utter_max && (new_tso > (bbr->r_ctl.bbr_utter_max * maxseg))) { new_tso = bbr->r_ctl.bbr_utter_max * maxseg; } if (old_tso != new_tso) { /* Only log changes */ bbr_log_type_tsosize(bbr, cts, new_tso, tls_seg, old_tso, maxseg, 0); bbr->r_ctl.rc_pace_max_segs = new_tso; } /* We have hardware pacing! */ bbr_adjust_for_hw_pacing(bbr, cts); } static void bbr_log_output(struct tcp_bbr *bbr, struct tcpcb *tp, struct tcpopt *to, int32_t len, uint32_t seq_out, uint16_t th_flags, int32_t err, uint32_t cts, struct mbuf *mb, int32_t * abandon, struct bbr_sendmap *hintrsm, uint32_t delay_calc, struct sockbuf *sb) { struct bbr_sendmap *rsm, *nrsm; register uint32_t snd_max, snd_una; uint32_t pacing_time; /* * Add to the RACK log of packets in flight or retransmitted. If * there is a TS option we will use the TS echoed, if not we will * grab a TS. * * Retransmissions will increment the count and move the ts to its * proper place. Note that if options do not include TS's then we * won't be able to effectively use the ACK for an RTT on a retran. * * Notes about r_start and r_end. Lets consider a send starting at * sequence 1 for 10 bytes. In such an example the r_start would be * 1 (starting sequence) but the r_end would be r_start+len i.e. 11. * This means that r_end is actually the first sequence for the next * slot (11). * */ INP_WLOCK_ASSERT(tp->t_inpcb); if (err) { /* * We don't log errors -- we could but snd_max does not * advance in this case either. */ return; } if (th_flags & TH_RST) { /* * We don't log resets and we return immediately from * sending */ *abandon = 1; return; } snd_una = tp->snd_una; if (th_flags & (TH_SYN | TH_FIN) && (hintrsm == NULL)) { /* * The call to bbr_log_output is made before bumping * snd_max. This means we can record one extra byte on a SYN * or FIN if seq_out is adding more on and a FIN is present * (and we are not resending). */ if ((th_flags & TH_SYN) && (tp->iss == seq_out)) len++; if (th_flags & TH_FIN) len++; } if (SEQ_LEQ((seq_out + len), snd_una)) { /* Are sending an old segment to induce an ack (keep-alive)? */ return; } if (SEQ_LT(seq_out, snd_una)) { /* huh? should we panic? */ uint32_t end; end = seq_out + len; seq_out = snd_una; len = end - seq_out; } snd_max = tp->snd_max; if (len == 0) { /* We don't log zero window probes */ return; } pacing_time = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, len, cts, 1); /* First question is it a retransmission? */ if (seq_out == snd_max) { again: rsm = bbr_alloc(bbr); if (rsm == NULL) { return; } rsm->r_flags = 0; if (th_flags & TH_SYN) rsm->r_flags |= BBR_HAS_SYN; if (th_flags & TH_FIN) rsm->r_flags |= BBR_HAS_FIN; rsm->r_tim_lastsent[0] = cts; rsm->r_rtr_cnt = 1; rsm->r_rtr_bytes = 0; rsm->r_start = seq_out; rsm->r_end = rsm->r_start + len; rsm->r_dupack = 0; rsm->r_delivered = bbr->r_ctl.rc_delivered; rsm->r_pacing_delay = pacing_time; rsm->r_ts_valid = bbr->rc_ts_valid; if (bbr->rc_ts_valid) rsm->r_del_ack_ts = bbr->r_ctl.last_inbound_ts; rsm->r_del_time = bbr->r_ctl.rc_del_time; if (bbr->r_ctl.r_app_limited_until) rsm->r_app_limited = 1; else rsm->r_app_limited = 0; rsm->r_first_sent_time = bbr_get_earliest_send_outstanding(bbr, rsm, cts); rsm->r_flight_at_send = ctf_flight_size(bbr->rc_tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); /* * Here we must also add in this rsm since snd_max * is updated after we return from a new send. */ rsm->r_flight_at_send += len; TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_map, rsm, r_next); TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 1; if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) rsm->r_bbr_state = bbr_state_val(bbr); else rsm->r_bbr_state = 8; if (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT) { rsm->r_is_gain = 1; rsm->r_is_drain = 0; } else if (bbr->r_ctl.rc_bbr_hptsi_gain < BBR_UNIT) { rsm->r_is_drain = 1; rsm->r_is_gain = 0; } else { rsm->r_is_drain = 0; rsm->r_is_gain = 0; } return; } /* * If we reach here its a retransmission and we need to find it. */ more: if (hintrsm && (hintrsm->r_start == seq_out)) { rsm = hintrsm; hintrsm = NULL; } else if (bbr->r_ctl.rc_next) { /* We have a hint from a previous run */ rsm = bbr->r_ctl.rc_next; } else { /* No hints sorry */ rsm = NULL; } if ((rsm) && (rsm->r_start == seq_out)) { /* * We used rc_next or hintrsm to retransmit, hopefully the * likely case. */ seq_out = bbr_update_entry(tp, bbr, rsm, cts, &len, pacing_time); if (len == 0) { return; } else { goto more; } } /* Ok it was not the last pointer go through it the hard way. */ TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { if (rsm->r_start == seq_out) { seq_out = bbr_update_entry(tp, bbr, rsm, cts, &len, pacing_time); bbr->r_ctl.rc_next = TAILQ_NEXT(rsm, r_next); if (len == 0) { return; } else { continue; } } if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) { /* Transmitted within this piece */ /* * Ok we must split off the front and then let the * update do the rest */ nrsm = bbr_alloc_full_limit(bbr); if (nrsm == NULL) { bbr_update_rsm(tp, bbr, rsm, cts, pacing_time); return; } /* * copy rsm to nrsm and then trim the front of rsm * to not include this part. */ bbr_clone_rsm(bbr, nrsm, rsm, seq_out); TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } rsm->r_flags &= (~BBR_HAS_FIN); seq_out = bbr_update_entry(tp, bbr, nrsm, cts, &len, pacing_time); if (len == 0) { return; } } } /* * Hmm not found in map did they retransmit both old and on into the * new? */ if (seq_out == tp->snd_max) { goto again; } else if (SEQ_LT(seq_out, tp->snd_max)) { #ifdef BBR_INVARIANTS printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n", seq_out, len, tp->snd_una, tp->snd_max); printf("Starting Dump of all rack entries\n"); TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { printf("rsm:%p start:%u end:%u\n", rsm, rsm->r_start, rsm->r_end); } printf("Dump complete\n"); panic("seq_out not found rack:%p tp:%p", bbr, tp); #endif } else { #ifdef BBR_INVARIANTS /* * Hmm beyond sndmax? (only if we are using the new rtt-pack * flag) */ panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p", seq_out, len, tp->snd_max, tp); #endif } } static void bbr_collapse_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, int32_t rtt) { /* * Collapse timeout back the cum-ack moved. */ tp->t_rxtshift = 0; tp->t_softerror = 0; } static void tcp_bbr_xmit_timer(struct tcp_bbr *bbr, uint32_t rtt_usecs, uint32_t rsm_send_time, uint32_t r_start, uint32_t tsin) { bbr->rtt_valid = 1; bbr->r_ctl.cur_rtt = rtt_usecs; bbr->r_ctl.ts_in = tsin; if (rsm_send_time) bbr->r_ctl.cur_rtt_send_time = rsm_send_time; } static void bbr_make_timestamp_determination(struct tcp_bbr *bbr) { /** * We have in our bbr control: * 1) The timestamp we started observing cum-acks (bbr->r_ctl.bbr_ts_check_tstmp). * 2) Our timestamp indicating when we sent that packet (bbr->r_ctl.rsm->bbr_ts_check_our_cts). * 3) The current timestamp that just came in (bbr->r_ctl.last_inbound_ts) * 4) The time that the packet that generated that ack was sent (bbr->r_ctl.cur_rtt_send_time) * * Now we can calculate the time between the sends by doing: * * delta = bbr->r_ctl.cur_rtt_send_time - bbr->r_ctl.bbr_ts_check_our_cts * * And the peer's time between receiving them by doing: * * peer_delta = bbr->r_ctl.last_inbound_ts - bbr->r_ctl.bbr_ts_check_tstmp * * We want to figure out if the timestamp values are in msec, 10msec or usec. * We also may find that we can't use the timestamps if say we see * that the peer_delta indicates that though we may have taken 10ms to * pace out the data, it only saw 1ms between the two packets. This would * indicate that somewhere on the path is a batching entity that is giving * out time-slices of the actual b/w. This would mean we could not use * reliably the peers timestamps. * * We expect delta > peer_delta initially. Until we figure out the * timestamp difference which we will store in bbr->r_ctl.bbr_peer_tsratio. * If we place 1000 there then its a ms vs our usec. If we place 10000 there * then its 10ms vs our usec. If the peer is running a usec clock we would * put a 1 there. If the value is faster then ours, we will disable the * use of timestamps (though we could revist this later if we find it to be not * just an isolated one or two flows)). * * To detect the batching middle boxes we will come up with our compensation and * if with it in place, we find the peer is drastically off (by some margin) in * the smaller direction, then we will assume the worst case and disable use of timestamps. * */ uint64_t delta, peer_delta, delta_up; delta = bbr->r_ctl.cur_rtt_send_time - bbr->r_ctl.bbr_ts_check_our_cts; if (delta < bbr_min_usec_delta) { /* * Have not seen a min amount of time * between our send times so we can * make a determination of the timestamp * yet. */ return; } peer_delta = bbr->r_ctl.last_inbound_ts - bbr->r_ctl.bbr_ts_check_tstmp; if (peer_delta < bbr_min_peer_delta) { /* * We may have enough in the form of * our delta but the peers number * has not changed that much. It could * be its clock ratio is such that * we need more data (10ms tick) or * there may be other compression scenarios * going on. In any event we need the * spread to be larger. */ return; } /* Ok lets first see which way our delta is going */ if (peer_delta > delta) { /* Very unlikely, the peer without * compensation shows that it saw * the two sends arrive further apart * then we saw then in micro-seconds. */ if (peer_delta < (delta + ((delta * (uint64_t)1000)/ (uint64_t)bbr_delta_percent))) { /* well it looks like the peer is a micro-second clock. */ bbr->rc_ts_clock_set = 1; bbr->r_ctl.bbr_peer_tsratio = 1; } else { bbr->rc_ts_cant_be_used = 1; bbr->rc_ts_clock_set = 1; } return; } /* Ok we know that the peer_delta is smaller than our send distance */ bbr->rc_ts_clock_set = 1; /* First question is it within the percentage that they are using usec time? */ delta_up = (peer_delta * 1000) / (uint64_t)bbr_delta_percent; if ((peer_delta + delta_up) >= delta) { /* Its a usec clock */ bbr->r_ctl.bbr_peer_tsratio = 1; bbr_log_tstmp_validation(bbr, peer_delta, delta); return; } /* Ok if not usec, what about 10usec (though unlikely)? */ delta_up = (peer_delta * 1000 * 10) / (uint64_t)bbr_delta_percent; if (((peer_delta * 10) + delta_up) >= delta) { bbr->r_ctl.bbr_peer_tsratio = 10; bbr_log_tstmp_validation(bbr, peer_delta, delta); return; } /* And what about 100usec (though again unlikely)? */ delta_up = (peer_delta * 1000 * 100) / (uint64_t)bbr_delta_percent; if (((peer_delta * 100) + delta_up) >= delta) { bbr->r_ctl.bbr_peer_tsratio = 100; bbr_log_tstmp_validation(bbr, peer_delta, delta); return; } /* And how about 1 msec (the most likely one)? */ delta_up = (peer_delta * 1000 * 1000) / (uint64_t)bbr_delta_percent; if (((peer_delta * 1000) + delta_up) >= delta) { bbr->r_ctl.bbr_peer_tsratio = 1000; bbr_log_tstmp_validation(bbr, peer_delta, delta); return; } /* Ok if not msec could it be 10 msec? */ delta_up = (peer_delta * 1000 * 10000) / (uint64_t)bbr_delta_percent; if (((peer_delta * 10000) + delta_up) >= delta) { bbr->r_ctl.bbr_peer_tsratio = 10000; return; } /* If we fall down here the clock tick so slowly we can't use it */ bbr->rc_ts_cant_be_used = 1; bbr->r_ctl.bbr_peer_tsratio = 0; bbr_log_tstmp_validation(bbr, peer_delta, delta); } /* * Collect new round-trip time estimate * and update averages and current timeout. */ static void tcp_bbr_xmit_timer_commit(struct tcp_bbr *bbr, struct tcpcb *tp, uint32_t cts) { int32_t delta; uint32_t rtt, tsin; int32_t rtt_ticks; if (bbr->rtt_valid == 0) /* No valid sample */ return; rtt = bbr->r_ctl.cur_rtt; tsin = bbr->r_ctl.ts_in; if (bbr->rc_prtt_set_ts) { /* * We are to force feed the rttProp filter due * to an entry into PROBE_RTT. This assures * that the times are sync'd between when we * go into PROBE_RTT and the filter expiration. * * Google does not use a true filter, so they do * this implicitly since they only keep one value * and when they enter probe-rtt they update the * value to the newest rtt. */ uint32_t rtt_prop; bbr->rc_prtt_set_ts = 0; rtt_prop = get_filter_value_small(&bbr->r_ctl.rc_rttprop); if (rtt > rtt_prop) filter_increase_by_small(&bbr->r_ctl.rc_rttprop, (rtt - rtt_prop), cts); else apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts); } if (bbr->rc_ack_was_delayed) rtt += bbr->r_ctl.rc_ack_hdwr_delay; if (rtt < bbr->r_ctl.rc_lowest_rtt) bbr->r_ctl.rc_lowest_rtt = rtt; bbr_log_rtt_sample(bbr, rtt, tsin); if (bbr->r_init_rtt) { /* * The initial rtt is not-trusted, nuke it and lets get * our first valid measurement in. */ bbr->r_init_rtt = 0; tp->t_srtt = 0; } if ((bbr->rc_ts_clock_set == 0) && bbr->rc_ts_valid) { /* * So we have not yet figured out * what the peers TSTMP value is * in (most likely ms). We need a * series of cum-ack's to determine * this reliably. */ if (bbr->rc_ack_is_cumack) { if (bbr->rc_ts_data_set) { /* Lets attempt to determine the timestamp granularity. */ bbr_make_timestamp_determination(bbr); } else { bbr->rc_ts_data_set = 1; bbr->r_ctl.bbr_ts_check_tstmp = bbr->r_ctl.last_inbound_ts; bbr->r_ctl.bbr_ts_check_our_cts = bbr->r_ctl.cur_rtt_send_time; } } else { /* * We have to have consecutive acks * reset any "filled" state to none. */ bbr->rc_ts_data_set = 0; } } /* Round it up */ rtt_ticks = USEC_2_TICKS((rtt + (USECS_IN_MSEC - 1))); if (rtt_ticks == 0) rtt_ticks = 1; if (tp->t_srtt != 0) { /* * srtt is stored as fixed point with 5 bits after the * binary point (i.e., scaled by 8). The following magic is * equivalent to the smoothing algorithm in rfc793 with an * alpha of .875 (srtt = rtt/8 + srtt*7/8 in fixed point). * Adjust rtt to origin 0. */ delta = ((rtt_ticks - 1) << TCP_DELTA_SHIFT) - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT)); tp->t_srtt += delta; if (tp->t_srtt <= 0) tp->t_srtt = 1; /* * We accumulate a smoothed rtt variance (actually, a * smoothed mean difference), then set the retransmit timer * to smoothed rtt + 4 times the smoothed variance. rttvar * is stored as fixed point with 4 bits after the binary * point (scaled by 16). The following is equivalent to * rfc793 smoothing with an alpha of .75 (rttvar = * rttvar*3/4 + |delta| / 4). This replaces rfc793's * wired-in beta. */ if (delta < 0) delta = -delta; delta -= tp->t_rttvar >> (TCP_RTTVAR_SHIFT - TCP_DELTA_SHIFT); tp->t_rttvar += delta; if (tp->t_rttvar <= 0) tp->t_rttvar = 1; if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar) tp->t_rttbest = tp->t_srtt + tp->t_rttvar; } else { /* * No rtt measurement yet - use the unsmoothed rtt. Set the * variance to half the rtt (so our first retransmit happens * at 3*rtt). */ tp->t_srtt = rtt_ticks << TCP_RTT_SHIFT; tp->t_rttvar = rtt_ticks << (TCP_RTTVAR_SHIFT - 1); tp->t_rttbest = tp->t_srtt + tp->t_rttvar; } KMOD_TCPSTAT_INC(tcps_rttupdated); tp->t_rttupdated++; #ifdef STATS stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt_ticks)); #endif /* * the retransmit should happen at rtt + 4 * rttvar. Because of the * way we do the smoothing, srtt and rttvar will each average +1/2 * tick of bias. When we compute the retransmit timer, we want 1/2 * tick of rounding and 1 extra tick because of +-1/2 tick * uncertainty in the firing of the timer. The bias will give us * exactly the 1.5 tick we need. But, because the bias is * statistical, we have to test that we don't drop below the minimum * feasible timer (which is 2 ticks). */ TCPT_RANGESET(tp->t_rxtcur, TCP_REXMTVAL(tp), max(MSEC_2_TICKS(bbr->r_ctl.rc_min_rto_ms), rtt_ticks + 2), MSEC_2_TICKS(((uint32_t)bbr->rc_max_rto_sec) * 1000)); /* * We received an ack for a packet that wasn't retransmitted; it is * probably safe to discard any error indications we've received * recently. This isn't quite right, but close enough for now (a * route might have failed after we sent a segment, and the return * path might not be symmetrical). */ tp->t_softerror = 0; rtt = (TICKS_2_USEC(bbr->rc_tp->t_srtt) >> TCP_RTT_SHIFT); if (bbr->r_ctl.bbr_smallest_srtt_this_state > rtt) bbr->r_ctl.bbr_smallest_srtt_this_state = rtt; } static void bbr_set_reduced_rtt(struct tcp_bbr *bbr, uint32_t cts, uint32_t line) { bbr->r_ctl.rc_rtt_shrinks = cts; if (bbr_can_force_probertt && (TSTMP_GT(cts, bbr->r_ctl.last_in_probertt)) && ((cts - bbr->r_ctl.last_in_probertt) > bbr->r_ctl.rc_probertt_int)) { /* * We should enter probe-rtt its been too long * since we have been there. */ bbr_enter_probe_rtt(bbr, cts, __LINE__); } else bbr_check_probe_rtt_limits(bbr, cts); } static void tcp_bbr_commit_bw(struct tcp_bbr *bbr, uint32_t cts) { uint64_t orig_bw; if (bbr->r_ctl.rc_bbr_cur_del_rate == 0) { /* We never apply a zero measurement */ bbr_log_type_bbrupd(bbr, 20, cts, 0, 0, 0, 0, 0, 0, 0, 0); return; } if (bbr->r_ctl.r_measurement_count < 0xffffffff) bbr->r_ctl.r_measurement_count++; orig_bw = get_filter_value(&bbr->r_ctl.rc_delrate); apply_filter_max(&bbr->r_ctl.rc_delrate, bbr->r_ctl.rc_bbr_cur_del_rate, bbr->r_ctl.rc_pkt_epoch); bbr_log_type_bbrupd(bbr, 21, cts, (uint32_t)orig_bw, (uint32_t)get_filter_value(&bbr->r_ctl.rc_delrate), 0, 0, 0, 0, 0, 0); if (orig_bw && (orig_bw != get_filter_value(&bbr->r_ctl.rc_delrate))) { if (bbr->bbr_hdrw_pacing) { /* * Apply a new rate to the hardware * possibly. */ bbr_update_hardware_pacing_rate(bbr, cts); } bbr_set_state_target(bbr, __LINE__); tcp_bbr_tso_size_check(bbr, cts); if (bbr->r_recovery_bw) { bbr_setup_red_bw(bbr, cts); bbr_log_type_bw_reduce(bbr, BBR_RED_BW_USELRBW); } } else if ((orig_bw == 0) && get_filter_value(&bbr->r_ctl.rc_delrate)) tcp_bbr_tso_size_check(bbr, cts); } static void bbr_nf_measurement(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts) { if (bbr->rc_in_persist == 0) { /* We log only when not in persist */ /* Translate to a Bytes Per Second */ uint64_t tim, bw, ts_diff, ts_bw; uint32_t delivered; if (TSTMP_GT(bbr->r_ctl.rc_del_time, rsm->r_del_time)) tim = (uint64_t)(bbr->r_ctl.rc_del_time - rsm->r_del_time); else tim = 1; /* * Now that we have processed the tim (skipping the sample * or possibly updating the time, go ahead and * calculate the cdr. */ delivered = (bbr->r_ctl.rc_delivered - rsm->r_delivered); bw = (uint64_t)delivered; bw *= (uint64_t)USECS_IN_SECOND; bw /= tim; if (bw == 0) { /* We must have a calculatable amount */ return; } /* * If we are using this b/w shove it in now so we * can see in the trace viewer if it gets over-ridden. */ if (rsm->r_ts_valid && bbr->rc_ts_valid && bbr->rc_ts_clock_set && (bbr->rc_ts_cant_be_used == 0) && bbr->rc_use_ts_limit) { ts_diff = max((bbr->r_ctl.last_inbound_ts - rsm->r_del_ack_ts), 1); ts_diff *= bbr->r_ctl.bbr_peer_tsratio; if ((delivered == 0) || (rtt < 1000)) { /* Can't use the ts */ bbr_log_type_bbrupd(bbr, 61, cts, ts_diff, bbr->r_ctl.last_inbound_ts, rsm->r_del_ack_ts, 0, 0, 0, 0, delivered); } else { ts_bw = (uint64_t)delivered; ts_bw *= (uint64_t)USECS_IN_SECOND; ts_bw /= ts_diff; bbr_log_type_bbrupd(bbr, 62, cts, (ts_bw >> 32), (ts_bw & 0xffffffff), 0, 0, 0, 0, ts_diff, delivered); if ((bbr->ts_can_raise) && (ts_bw > bw)) { bbr_log_type_bbrupd(bbr, 8, cts, delivered, ts_diff, (bw >> 32), (bw & 0x00000000ffffffff), 0, 0, 0, 0); bw = ts_bw; } else if (ts_bw && (ts_bw < bw)) { bbr_log_type_bbrupd(bbr, 7, cts, delivered, ts_diff, (bw >> 32), (bw & 0x00000000ffffffff), 0, 0, 0, 0); bw = ts_bw; } } } if (rsm->r_first_sent_time && TSTMP_GT(rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)],rsm->r_first_sent_time)) { uint64_t sbw, sti; /* * We use what was in flight at the time of our * send and the size of this send to figure * out what we have been sending at (amount). * For the time we take from the time of * the send of the first send outstanding * until this send plus this sends pacing * time. This gives us a good calculation * as to the rate we have been sending at. */ sbw = (uint64_t)(rsm->r_flight_at_send); sbw *= (uint64_t)USECS_IN_SECOND; sti = rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)] - rsm->r_first_sent_time; sti += rsm->r_pacing_delay; sbw /= sti; if (sbw < bw) { bbr_log_type_bbrupd(bbr, 6, cts, delivered, (uint32_t)sti, (bw >> 32), (uint32_t)bw, rsm->r_first_sent_time, 0, (sbw >> 32), (uint32_t)sbw); bw = sbw; } } /* Use the google algorithm for b/w measurements */ bbr->r_ctl.rc_bbr_cur_del_rate = bw; if ((rsm->r_app_limited == 0) || (bw > get_filter_value(&bbr->r_ctl.rc_delrate))) { tcp_bbr_commit_bw(bbr, cts); bbr_log_type_bbrupd(bbr, 10, cts, (uint32_t)tim, delivered, 0, 0, 0, 0, bbr->r_ctl.rc_del_time, rsm->r_del_time); } } } static void bbr_google_measurement(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts) { if (bbr->rc_in_persist == 0) { /* We log only when not in persist */ /* Translate to a Bytes Per Second */ uint64_t tim, bw; uint32_t delivered; int no_apply = 0; if (TSTMP_GT(bbr->r_ctl.rc_del_time, rsm->r_del_time)) tim = (uint64_t)(bbr->r_ctl.rc_del_time - rsm->r_del_time); else tim = 1; /* * Now that we have processed the tim (skipping the sample * or possibly updating the time, go ahead and * calculate the cdr. */ delivered = (bbr->r_ctl.rc_delivered - rsm->r_delivered); bw = (uint64_t)delivered; bw *= (uint64_t)USECS_IN_SECOND; bw /= tim; if (tim < bbr->r_ctl.rc_lowest_rtt) { bbr_log_type_bbrupd(bbr, 99, cts, (uint32_t)tim, delivered, tim, bbr->r_ctl.rc_lowest_rtt, 0, 0, 0, 0); no_apply = 1; } /* * If we are using this b/w shove it in now so we * can see in the trace viewer if it gets over-ridden. */ bbr->r_ctl.rc_bbr_cur_del_rate = bw; /* Gate by the sending rate */ if (rsm->r_first_sent_time && TSTMP_GT(rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)],rsm->r_first_sent_time)) { uint64_t sbw, sti; /* * We use what was in flight at the time of our * send and the size of this send to figure * out what we have been sending at (amount). * For the time we take from the time of * the send of the first send outstanding * until this send plus this sends pacing * time. This gives us a good calculation * as to the rate we have been sending at. */ sbw = (uint64_t)(rsm->r_flight_at_send); sbw *= (uint64_t)USECS_IN_SECOND; sti = rsm->r_tim_lastsent[(rsm->r_rtr_cnt -1)] - rsm->r_first_sent_time; sti += rsm->r_pacing_delay; sbw /= sti; if (sbw < bw) { bbr_log_type_bbrupd(bbr, 6, cts, delivered, (uint32_t)sti, (bw >> 32), (uint32_t)bw, rsm->r_first_sent_time, 0, (sbw >> 32), (uint32_t)sbw); bw = sbw; } if ((sti > tim) && (sti < bbr->r_ctl.rc_lowest_rtt)) { bbr_log_type_bbrupd(bbr, 99, cts, (uint32_t)tim, delivered, (uint32_t)sti, bbr->r_ctl.rc_lowest_rtt, 0, 0, 0, 0); no_apply = 1; } else no_apply = 0; } bbr->r_ctl.rc_bbr_cur_del_rate = bw; if ((no_apply == 0) && ((rsm->r_app_limited == 0) || (bw > get_filter_value(&bbr->r_ctl.rc_delrate)))) { tcp_bbr_commit_bw(bbr, cts); bbr_log_type_bbrupd(bbr, 10, cts, (uint32_t)tim, delivered, 0, 0, 0, 0, bbr->r_ctl.rc_del_time, rsm->r_del_time); } } } static void bbr_update_bbr_info(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t rtt, uint32_t cts, uint32_t tsin, uint32_t uts, int32_t match, uint32_t rsm_send_time, int32_t ack_type, struct tcpopt *to) { uint64_t old_rttprop; /* Update our delivery time and amount */ bbr->r_ctl.rc_delivered += (rsm->r_end - rsm->r_start); bbr->r_ctl.rc_del_time = cts; if (rtt == 0) { /* * 0 means its a retransmit, for now we don't use these for * the rest of BBR. */ return; } if ((bbr->rc_use_google == 0) && (match != BBR_RTT_BY_EXACTMATCH) && (match != BBR_RTT_BY_TIMESTAMP)){ /* * We get a lot of rtt updates, lets not pay attention to * any that are not an exact match. That way we don't have * to worry about timestamps and the whole nonsense of * unsure if its a retransmission etc (if we ever had the * timestamp fixed to always have the last thing sent this * would not be a issue). */ return; } if ((bbr_no_retran && bbr->rc_use_google) && (match != BBR_RTT_BY_EXACTMATCH) && (match != BBR_RTT_BY_TIMESTAMP)){ /* * We only do measurements in google mode * with bbr_no_retran on for sure things. */ return; } /* Only update srtt if we know by exact match */ tcp_bbr_xmit_timer(bbr, rtt, rsm_send_time, rsm->r_start, tsin); if (ack_type == BBR_CUM_ACKED) bbr->rc_ack_is_cumack = 1; else bbr->rc_ack_is_cumack = 0; old_rttprop = bbr_get_rtt(bbr, BBR_RTT_PROP); /* * Note the following code differs to the original * BBR spec. It calls for <= not <. However after a * long discussion in email with Neal, he acknowledged * that it should be < than so that we will have flows * going into probe-rtt (we were seeing cases where that * did not happen and caused ugly things to occur). We * have added this agreed upon fix to our code base. */ if (rtt < old_rttprop) { /* Update when we last saw a rtt drop */ bbr_log_rtt_shrinks(bbr, cts, 0, rtt, __LINE__, BBR_RTTS_NEWRTT, 0); bbr_set_reduced_rtt(bbr, cts, __LINE__); } bbr_log_type_bbrrttprop(bbr, rtt, (rsm ? rsm->r_end : 0), uts, cts, match, rsm->r_start, rsm->r_flags); apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts); if (old_rttprop != bbr_get_rtt(bbr, BBR_RTT_PROP)) { /* * The RTT-prop moved, reset the target (may be a * nop for some states). */ bbr_set_state_target(bbr, __LINE__); if (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_NEW_TARGET, 0); else if (old_rttprop < bbr_get_rtt(bbr, BBR_RTT_PROP)) /* It went up */ bbr_check_probe_rtt_limits(bbr, cts); } if ((bbr->rc_use_google == 0) && (match == BBR_RTT_BY_TIMESTAMP)) { /* * We don't do b/w update with * these since they are not really * reliable. */ return; } if (bbr->r_ctl.r_app_limited_until && (bbr->r_ctl.rc_delivered >= bbr->r_ctl.r_app_limited_until)) { /* We are no longer app-limited */ bbr->r_ctl.r_app_limited_until = 0; } if (bbr->rc_use_google) { bbr_google_measurement(bbr, rsm, rtt, cts); } else { bbr_nf_measurement(bbr, rsm, rtt, cts); } } /* * Convert a timestamp that the main stack * uses (milliseconds) into one that bbr uses * (microseconds). Return that converted timestamp. */ static uint32_t bbr_ts_convert(uint32_t cts) { uint32_t sec, msec; sec = cts / MS_IN_USEC; msec = cts - (MS_IN_USEC * sec); return ((sec * USECS_IN_SECOND) + (msec * MS_IN_USEC)); } /* * Return 0 if we did not update the RTT time, return * 1 if we did. */ static int bbr_update_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, uint32_t th_ack) { int32_t i; uint32_t t, uts = 0; if ((rsm->r_flags & BBR_ACKED) || (rsm->r_flags & BBR_WAS_RENEGED) || (rsm->r_flags & BBR_RXT_CLEARED)) { /* Already done */ return (0); } if (rsm->r_rtt_not_allowed) { /* Not allowed */ return (0); } if (rsm->r_rtr_cnt == 1) { /* * Only one transmit. Hopefully the normal case. */ if (TSTMP_GT(cts, rsm->r_tim_lastsent[0])) t = cts - rsm->r_tim_lastsent[0]; else t = 1; if ((int)t <= 0) t = 1; bbr->r_ctl.rc_last_rtt = t; bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, 0, BBR_RTT_BY_EXACTMATCH, rsm->r_tim_lastsent[0], ack_type, to); return (1); } /* Convert to usecs */ if ((bbr_can_use_ts_for_rtt == 1) && (bbr->rc_use_google == 1) && (ack_type == BBR_CUM_ACKED) && (to->to_flags & TOF_TS) && (to->to_tsecr != 0)) { t = tcp_tv_to_mssectick(&bbr->rc_tv) - to->to_tsecr; if (t < 1) t = 1; t *= MS_IN_USEC; bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, 0, BBR_RTT_BY_TIMESTAMP, rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)], ack_type, to); return (1); } uts = bbr_ts_convert(to->to_tsecr); if ((to->to_flags & TOF_TS) && (to->to_tsecr != 0) && (ack_type == BBR_CUM_ACKED) && ((rsm->r_flags & BBR_OVERMAX) == 0)) { /* * Now which timestamp does it match? In this block the ACK * may be coming from a previous transmission. */ uint32_t fudge; fudge = BBR_TIMER_FUDGE; for (i = 0; i < rsm->r_rtr_cnt; i++) { if ((SEQ_GEQ(uts, (rsm->r_tim_lastsent[i] - fudge))) && (SEQ_LEQ(uts, (rsm->r_tim_lastsent[i] + fudge)))) { if (TSTMP_GT(cts, rsm->r_tim_lastsent[i])) t = cts - rsm->r_tim_lastsent[i]; else t = 1; if ((int)t <= 0) t = 1; bbr->r_ctl.rc_last_rtt = t; bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_TSMATCHING, rsm->r_tim_lastsent[i], ack_type, to); if ((i + 1) < rsm->r_rtr_cnt) { /* Likely */ return (0); } else if (rsm->r_flags & BBR_TLP) { bbr->rc_tlp_rtx_out = 0; } return (1); } } /* Fall through if we can't find a matching timestamp */ } /* * Ok its a SACK block that we retransmitted. or a windows * machine without timestamps. We can tell nothing from the * time-stamp since its not there or the time the peer last * recieved a segment that moved forward its cum-ack point. * * Lets look at the last retransmit and see what we can tell * (with BBR for space we only keep 2 note we have to keep * at least 2 so the map can not be condensed more). */ i = rsm->r_rtr_cnt - 1; if (TSTMP_GT(cts, rsm->r_tim_lastsent[i])) t = cts - rsm->r_tim_lastsent[i]; else goto not_sure; if (t < bbr->r_ctl.rc_lowest_rtt) { /* * We retransmitted and the ack came back in less * than the smallest rtt we have observed in the * windowed rtt. We most likey did an improper * retransmit as outlined in 4.2 Step 3 point 2 in * the rack-draft. * * Use the prior transmission to update all the * information as long as there is only one prior * transmission. */ if ((rsm->r_flags & BBR_OVERMAX) == 0) { #ifdef BBR_INVARIANTS if (rsm->r_rtr_cnt == 1) panic("rsm:%p bbr:%p rsm has overmax and only 1 retranmit flags:%x?", rsm, bbr, rsm->r_flags); #endif i = rsm->r_rtr_cnt - 2; if (TSTMP_GT(cts, rsm->r_tim_lastsent[i])) t = cts - rsm->r_tim_lastsent[i]; else t = 1; bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_EARLIER_RET, rsm->r_tim_lastsent[i], ack_type, to); return (0); } else { /* * Too many prior transmissions, just * updated BBR delivered */ not_sure: bbr_update_bbr_info(bbr, rsm, 0, cts, to->to_tsecr, uts, BBR_RTT_BY_SOME_RETRAN, 0, ack_type, to); } } else { /* * We retransmitted it and the retransmit did the * job. */ if (rsm->r_flags & BBR_TLP) bbr->rc_tlp_rtx_out = 0; if ((rsm->r_flags & BBR_OVERMAX) == 0) bbr_update_bbr_info(bbr, rsm, t, cts, to->to_tsecr, uts, BBR_RTT_BY_THIS_RETRAN, 0, ack_type, to); else bbr_update_bbr_info(bbr, rsm, 0, cts, to->to_tsecr, uts, BBR_RTT_BY_SOME_RETRAN, 0, ack_type, to); return (1); } return (0); } /* * Mark the SACK_PASSED flag on all entries prior to rsm send wise. */ static void bbr_log_sack_passed(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm) { struct bbr_sendmap *nrsm; nrsm = rsm; TAILQ_FOREACH_REVERSE_FROM(nrsm, &bbr->r_ctl.rc_tmap, bbr_head, r_tnext) { if (nrsm == rsm) { /* Skip orginal segment he is acked */ continue; } if (nrsm->r_flags & BBR_ACKED) { /* Skip ack'd segments */ continue; } if (nrsm->r_flags & BBR_SACK_PASSED) { /* * We found one that is already marked * passed, we have been here before and * so all others below this are marked. */ break; } BBR_STAT_INC(bbr_sack_passed); nrsm->r_flags |= BBR_SACK_PASSED; if (((nrsm->r_flags & BBR_MARKED_LOST) == 0) && bbr_is_lost(bbr, nrsm, bbr->r_ctl.rc_rcvtime)) { bbr->r_ctl.rc_lost += nrsm->r_end - nrsm->r_start; bbr->r_ctl.rc_lost_bytes += nrsm->r_end - nrsm->r_start; nrsm->r_flags |= BBR_MARKED_LOST; } nrsm->r_flags &= ~BBR_WAS_SACKPASS; } } /* * Returns the number of bytes that were * newly ack'd by sack blocks. */ static uint32_t bbr_proc_sack_blk(struct tcpcb *tp, struct tcp_bbr *bbr, struct sackblk *sack, struct tcpopt *to, struct bbr_sendmap **prsm, uint32_t cts) { int32_t times = 0; uint32_t start, end, changed = 0; struct bbr_sendmap *rsm, *nrsm; int32_t used_ref = 1; uint8_t went_back = 0, went_fwd = 0; start = sack->start; end = sack->end; rsm = *prsm; if (rsm == NULL) used_ref = 0; /* Do we locate the block behind where we last were? */ if (rsm && SEQ_LT(start, rsm->r_start)) { went_back = 1; TAILQ_FOREACH_REVERSE_FROM(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) { if (SEQ_GEQ(start, rsm->r_start) && SEQ_LT(start, rsm->r_end)) { goto do_rest_ofb; } } } start_at_beginning: went_fwd = 1; /* * Ok lets locate the block where this guy is fwd from rsm (if its * set) */ TAILQ_FOREACH_FROM(rsm, &bbr->r_ctl.rc_map, r_next) { if (SEQ_GEQ(start, rsm->r_start) && SEQ_LT(start, rsm->r_end)) { break; } } do_rest_ofb: if (rsm == NULL) { /* * This happens when we get duplicate sack blocks with the * same end. For example SACK 4: 100 SACK 3: 100 The sort * will not change there location so we would just start at * the end of the first one and get lost. */ if (tp->t_flags & TF_SENTFIN) { /* * Check to see if we have not logged the FIN that * went out. */ nrsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next); if (nrsm && (nrsm->r_end + 1) == tp->snd_max) { /* * Ok we did not get the FIN logged. */ nrsm->r_end++; rsm = nrsm; goto do_rest_ofb; } } if (times == 1) { #ifdef BBR_INVARIANTS panic("tp:%p bbr:%p sack:%p to:%p prsm:%p", tp, bbr, sack, to, prsm); #else goto out; #endif } times++; BBR_STAT_INC(bbr_sack_proc_restart); rsm = NULL; goto start_at_beginning; } /* Ok we have an ACK for some piece of rsm */ if (rsm->r_start != start) { /* * Need to split this in two pieces the before and after. */ if (bbr_sack_mergable(rsm, start, end)) nrsm = bbr_alloc_full_limit(bbr); else nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT); if (nrsm == NULL) { /* We could not allocate ignore the sack */ struct sackblk blk; blk.start = start; blk.end = end; sack_filter_reject(&bbr->r_ctl.bbr_sf, &blk); goto out; } bbr_clone_rsm(bbr, nrsm, rsm, start); TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } rsm->r_flags &= (~BBR_HAS_FIN); rsm = nrsm; } if (SEQ_GEQ(end, rsm->r_end)) { /* * The end of this block is either beyond this guy or right * at this guy. */ if ((rsm->r_flags & BBR_ACKED) == 0) { bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_SACKED, 0); changed += (rsm->r_end - rsm->r_start); bbr->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start); bbr_log_sack_passed(tp, bbr, rsm); if (rsm->r_flags & BBR_MARKED_LOST) { bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; } /* Is Reordering occuring? */ if (rsm->r_flags & BBR_SACK_PASSED) { BBR_STAT_INC(bbr_reorder_seen); bbr->r_ctl.rc_reorder_ts = cts; if (rsm->r_flags & BBR_MARKED_LOST) { bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) /* LT sampling also needs adjustment */ bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; } } rsm->r_flags |= BBR_ACKED; rsm->r_flags &= ~(BBR_TLP|BBR_WAS_RENEGED|BBR_RXT_CLEARED|BBR_MARKED_LOST); if (rsm->r_in_tmap) { TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 0; } } bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_SACKED); if (end == rsm->r_end) { /* This block only - done */ goto out; } /* There is more not coverend by this rsm move on */ start = rsm->r_end; nrsm = TAILQ_NEXT(rsm, r_next); rsm = nrsm; times = 0; goto do_rest_ofb; } if (rsm->r_flags & BBR_ACKED) { /* Been here done that */ goto out; } /* Ok we need to split off this one at the tail */ if (bbr_sack_mergable(rsm, start, end)) nrsm = bbr_alloc_full_limit(bbr); else nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT); if (nrsm == NULL) { /* failed XXXrrs what can we do but loose the sack info? */ struct sackblk blk; blk.start = start; blk.end = end; sack_filter_reject(&bbr->r_ctl.bbr_sf, &blk); goto out; } /* Clone it */ bbr_clone_rsm(bbr, nrsm, rsm, end); /* The sack block does not cover this guy fully */ rsm->r_flags &= (~BBR_HAS_FIN); TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } nrsm->r_dupack = 0; bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_SACKED, 0); bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_SACKED); changed += (rsm->r_end - rsm->r_start); bbr->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start); bbr_log_sack_passed(tp, bbr, rsm); /* Is Reordering occuring? */ if (rsm->r_flags & BBR_MARKED_LOST) { bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; } if (rsm->r_flags & BBR_SACK_PASSED) { BBR_STAT_INC(bbr_reorder_seen); bbr->r_ctl.rc_reorder_ts = cts; if (rsm->r_flags & BBR_MARKED_LOST) { bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) /* LT sampling also needs adjustment */ bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; } } rsm->r_flags &= ~(BBR_TLP|BBR_WAS_RENEGED|BBR_RXT_CLEARED|BBR_MARKED_LOST); rsm->r_flags |= BBR_ACKED; if (rsm->r_in_tmap) { TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 0; } out: if (rsm && (rsm->r_flags & BBR_ACKED)) { /* * Now can we merge this newly acked * block with either the previous or * next block? */ nrsm = TAILQ_NEXT(rsm, r_next); if (nrsm && (nrsm->r_flags & BBR_ACKED)) { /* yep this and next can be merged */ rsm = bbr_merge_rsm(bbr, rsm, nrsm); } /* Now what about the previous? */ nrsm = TAILQ_PREV(rsm, bbr_head, r_next); if (nrsm && (nrsm->r_flags & BBR_ACKED)) { /* yep the previous and this can be merged */ rsm = bbr_merge_rsm(bbr, nrsm, rsm); } } if (used_ref == 0) { BBR_STAT_INC(bbr_sack_proc_all); } else { BBR_STAT_INC(bbr_sack_proc_short); } if (went_fwd && went_back) { BBR_STAT_INC(bbr_sack_search_both); } else if (went_fwd) { BBR_STAT_INC(bbr_sack_search_fwd); } else if (went_back) { BBR_STAT_INC(bbr_sack_search_back); } /* Save off where the next seq is */ if (rsm) bbr->r_ctl.rc_sacklast = TAILQ_NEXT(rsm, r_next); else bbr->r_ctl.rc_sacklast = NULL; *prsm = rsm; return (changed); } static void inline bbr_peer_reneges(struct tcp_bbr *bbr, struct bbr_sendmap *rsm, tcp_seq th_ack) { struct bbr_sendmap *tmap; BBR_STAT_INC(bbr_reneges_seen); tmap = NULL; while (rsm && (rsm->r_flags & BBR_ACKED)) { /* Its no longer sacked, mark it so */ uint32_t oflags; bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); #ifdef BBR_INVARIANTS if (rsm->r_in_tmap) { panic("bbr:%p rsm:%p flags:0x%x in tmap?", bbr, rsm, rsm->r_flags); } #endif oflags = rsm->r_flags; if (rsm->r_flags & BBR_MARKED_LOST) { bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) /* LT sampling also needs adjustment */ bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; } rsm->r_flags &= ~(BBR_ACKED | BBR_SACK_PASSED | BBR_WAS_SACKPASS | BBR_MARKED_LOST); rsm->r_flags |= BBR_WAS_RENEGED; rsm->r_flags |= BBR_RXT_CLEARED; bbr_log_type_rsmclear(bbr, bbr->r_ctl.rc_rcvtime, rsm, oflags, __LINE__); /* Rebuild it into our tmap */ if (tmap == NULL) { TAILQ_INSERT_HEAD(&bbr->r_ctl.rc_tmap, rsm, r_tnext); tmap = rsm; } else { TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, tmap, rsm, r_tnext); tmap = rsm; } tmap->r_in_tmap = 1; /* * XXXrrs Delivered? Should we do anything here? * * Of course we don't on a rxt timeout so maybe its ok that * we don't? * * For now lets not. */ rsm = TAILQ_NEXT(rsm, r_next); } /* * Now lets possibly clear the sack filter so we start recognizing * sacks that cover this area. */ sack_filter_clear(&bbr->r_ctl.bbr_sf, th_ack); } static void bbr_log_syn(struct tcpcb *tp, struct tcpopt *to) { struct tcp_bbr *bbr; struct bbr_sendmap *rsm; uint32_t cts; bbr = (struct tcp_bbr *)tp->t_fb_ptr; cts = bbr->r_ctl.rc_rcvtime; rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); if (rsm && (rsm->r_flags & BBR_HAS_SYN)) { if ((rsm->r_end - rsm->r_start) <= 1) { /* Log out the SYN completely */ bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; rsm->r_rtr_bytes = 0; TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next); if (rsm->r_in_tmap) { TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 0; } if (bbr->r_ctl.rc_next == rsm) { /* scoot along the marker */ bbr->r_ctl.rc_next = TAILQ_FIRST(&bbr->r_ctl.rc_map); } if (to != NULL) bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_CUM_ACKED, 0); bbr_free(bbr, rsm); } else { /* There is more (Fast open)? strip out SYN. */ rsm->r_flags &= ~BBR_HAS_SYN; rsm->r_start++; } } } /* * Returns the number of bytes that were * acknowledged by SACK blocks. */ static uint32_t bbr_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, uint32_t *prev_acked) { uint32_t changed, last_seq, entered_recovery = 0; struct tcp_bbr *bbr; struct bbr_sendmap *rsm; struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1]; register uint32_t th_ack; int32_t i, j, k, new_sb, num_sack_blks = 0; uint32_t cts, acked, ack_point, sack_changed = 0; uint32_t p_maxseg, maxseg, p_acked = 0; INP_WLOCK_ASSERT(tp->t_inpcb); if (tcp_get_flags(th) & TH_RST) { /* We don't log resets */ return (0); } bbr = (struct tcp_bbr *)tp->t_fb_ptr; cts = bbr->r_ctl.rc_rcvtime; rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); changed = 0; maxseg = tp->t_maxseg - bbr->rc_last_options; p_maxseg = min(bbr->r_ctl.rc_pace_max_segs, maxseg); th_ack = th->th_ack; if (SEQ_GT(th_ack, tp->snd_una)) { acked = th_ack - tp->snd_una; bbr_log_progress_event(bbr, tp, ticks, PROGRESS_UPDATE, __LINE__); bbr->rc_tp->t_acktime = ticks; } else acked = 0; if (SEQ_LEQ(th_ack, tp->snd_una)) { /* Only sent here for sack processing */ goto proc_sack; } if (rsm && SEQ_GT(th_ack, rsm->r_start)) { changed = th_ack - rsm->r_start; } else if ((rsm == NULL) && ((th_ack - 1) == tp->iss)) { /* * For the SYN incoming case we will not have called * tcp_output for the sending of the SYN, so there will be * no map. All other cases should probably be a panic. */ if ((to->to_flags & TOF_TS) && (to->to_tsecr != 0)) { /* * We have a timestamp that can be used to generate * an initial RTT. */ uint32_t ts, now, rtt; ts = bbr_ts_convert(to->to_tsecr); now = bbr_ts_convert(tcp_tv_to_mssectick(&bbr->rc_tv)); rtt = now - ts; if (rtt < 1) rtt = 1; bbr_log_type_bbrrttprop(bbr, rtt, tp->iss, 0, cts, BBR_RTT_BY_TIMESTAMP, tp->iss, 0); apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts); changed = 1; bbr->r_wanted_output = 1; goto out; } goto proc_sack; } else if (rsm == NULL) { goto out; } if (changed) { /* * The ACK point is advancing to th_ack, we must drop off * the packets in the rack log and calculate any eligble * RTT's. */ bbr->r_wanted_output = 1; more: if (rsm == NULL) { if (tp->t_flags & TF_SENTFIN) { /* if we send a FIN we will not hav a map */ goto proc_sack; } #ifdef BBR_INVARIANTS panic("No rack map tp:%p for th:%p state:%d bbr:%p snd_una:%u snd_max:%u chg:%d\n", tp, th, tp->t_state, bbr, tp->snd_una, tp->snd_max, changed); #endif goto proc_sack; } } if (SEQ_LT(th_ack, rsm->r_start)) { /* Huh map is missing this */ #ifdef BBR_INVARIANTS printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d bbr:%p\n", rsm->r_start, th_ack, tp->t_state, bbr->r_state, bbr); panic("th-ack is bad bbr:%p tp:%p", bbr, tp); #endif goto proc_sack; } else if (th_ack == rsm->r_start) { /* None here to ack */ goto proc_sack; } /* * Clear the dup ack counter, it will * either be freed or if there is some * remaining we need to start it at zero. */ rsm->r_dupack = 0; /* Now do we consume the whole thing? */ if (SEQ_GEQ(th_ack, rsm->r_end)) { /* Its all consumed. */ uint32_t left; if (rsm->r_flags & BBR_ACKED) { /* * It was acked on the scoreboard -- remove it from * total */ p_acked += (rsm->r_end - rsm->r_start); bbr->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); if (bbr->r_ctl.rc_sacked == 0) bbr->r_ctl.rc_sacklast = NULL; } else { bbr_update_rtt(tp, bbr, rsm, to, cts, BBR_CUM_ACKED, th_ack); if (rsm->r_flags & BBR_MARKED_LOST) { bbr->r_ctl.rc_lost_bytes -= rsm->r_end - rsm->r_start; } if (rsm->r_flags & BBR_SACK_PASSED) { /* * There are acked segments ACKED on the * scoreboard further up. We are seeing * reordering. */ BBR_STAT_INC(bbr_reorder_seen); bbr->r_ctl.rc_reorder_ts = cts; if (rsm->r_flags & BBR_MARKED_LOST) { bbr->r_ctl.rc_lost -= rsm->r_end - rsm->r_start; if (SEQ_GT(bbr->r_ctl.rc_lt_lost, bbr->r_ctl.rc_lost)) /* LT sampling also needs adjustment */ bbr->r_ctl.rc_lt_lost = bbr->r_ctl.rc_lost; } } rsm->r_flags &= ~BBR_MARKED_LOST; } bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; rsm->r_rtr_bytes = 0; TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next); if (rsm->r_in_tmap) { TAILQ_REMOVE(&bbr->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 0; } if (bbr->r_ctl.rc_next == rsm) { /* scoot along the marker */ bbr->r_ctl.rc_next = TAILQ_FIRST(&bbr->r_ctl.rc_map); } bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_CUM_ACKED); /* Adjust the packet counts */ left = th_ack - rsm->r_end; /* Free back to zone */ bbr_free(bbr, rsm); if (left) { rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); goto more; } goto proc_sack; } if (rsm->r_flags & BBR_ACKED) { /* * It was acked on the scoreboard -- remove it from total * for the part being cum-acked. */ p_acked += (rsm->r_end - rsm->r_start); bbr->r_ctl.rc_sacked -= (th_ack - rsm->r_start); if (bbr->r_ctl.rc_sacked == 0) bbr->r_ctl.rc_sacklast = NULL; } else { /* * It was acked up to th_ack point for the first time */ struct bbr_sendmap lrsm; memcpy(&lrsm, rsm, sizeof(struct bbr_sendmap)); lrsm.r_end = th_ack; bbr_update_rtt(tp, bbr, &lrsm, to, cts, BBR_CUM_ACKED, th_ack); } if ((rsm->r_flags & BBR_MARKED_LOST) && ((rsm->r_flags & BBR_ACKED) == 0)) { /* * It was marked lost and partly ack'd now * for the first time. We lower the rc_lost_bytes * and still leave it MARKED. */ bbr->r_ctl.rc_lost_bytes -= th_ack - rsm->r_start; } bbr_isit_a_pkt_epoch(bbr, cts, rsm, __LINE__, BBR_CUM_ACKED); bbr->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; rsm->r_rtr_bytes = 0; /* adjust packet count */ rsm->r_start = th_ack; proc_sack: /* Check for reneging */ rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); if (rsm && (rsm->r_flags & BBR_ACKED) && (th_ack == rsm->r_start)) { /* * The peer has moved snd_una up to the edge of this send, * i.e. one that it had previously acked. The only way that * can be true if the peer threw away data (space issues) * that it had previously sacked (else it would have given * us snd_una up to (rsm->r_end). We need to undo the acked * markings here. * * Note we have to look to make sure th_ack is our * rsm->r_start in case we get an old ack where th_ack is * behind snd_una. */ bbr_peer_reneges(bbr, rsm, th->th_ack); } if ((to->to_flags & TOF_SACK) == 0) { /* We are done nothing left to log */ goto out; } rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_map, bbr_sendmap, r_next); if (rsm) { last_seq = rsm->r_end; } else { last_seq = tp->snd_max; } /* Sack block processing */ if (SEQ_GT(th_ack, tp->snd_una)) ack_point = th_ack; else ack_point = tp->snd_una; for (i = 0; i < to->to_nsacks; i++) { bcopy((to->to_sacks + i * TCPOLEN_SACK), &sack, sizeof(sack)); sack.start = ntohl(sack.start); sack.end = ntohl(sack.end); if (SEQ_GT(sack.end, sack.start) && SEQ_GT(sack.start, ack_point) && SEQ_LT(sack.start, tp->snd_max) && SEQ_GT(sack.end, ack_point) && SEQ_LEQ(sack.end, tp->snd_max)) { if ((bbr->r_ctl.rc_num_small_maps_alloced > bbr_sack_block_limit) && (SEQ_LT(sack.end, last_seq)) && ((sack.end - sack.start) < (p_maxseg / 8))) { /* * Not the last piece and its smaller than * 1/8th of a p_maxseg. We ignore this. */ BBR_STAT_INC(bbr_runt_sacks); continue; } sack_blocks[num_sack_blks] = sack; num_sack_blks++; } else if (SEQ_LEQ(sack.start, th_ack) && SEQ_LEQ(sack.end, th_ack)) { /* * Its a D-SACK block. */ tcp_record_dsack(tp, sack.start, sack.end, 0); } } if (num_sack_blks == 0) goto out; /* * Sort the SACK blocks so we can update the rack scoreboard with * just one pass. */ new_sb = sack_filter_blks(&bbr->r_ctl.bbr_sf, sack_blocks, num_sack_blks, th->th_ack); ctf_log_sack_filter(bbr->rc_tp, new_sb, sack_blocks); BBR_STAT_ADD(bbr_sack_blocks, num_sack_blks); BBR_STAT_ADD(bbr_sack_blocks_skip, (num_sack_blks - new_sb)); num_sack_blks = new_sb; if (num_sack_blks < 2) { goto do_sack_work; } /* Sort the sacks */ for (i = 0; i < num_sack_blks; i++) { for (j = i + 1; j < num_sack_blks; j++) { if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) { sack = sack_blocks[i]; sack_blocks[i] = sack_blocks[j]; sack_blocks[j] = sack; } } } /* * Now are any of the sack block ends the same (yes some * implememtations send these)? */ again: if (num_sack_blks > 1) { for (i = 0; i < num_sack_blks; i++) { for (j = i + 1; j < num_sack_blks; j++) { if (sack_blocks[i].end == sack_blocks[j].end) { /* * Ok these two have the same end we * want the smallest end and then * throw away the larger and start * again. */ if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) { /* * The second block covers * more area use that */ sack_blocks[i].start = sack_blocks[j].start; } /* * Now collapse out the dup-sack and * lower the count */ for (k = (j + 1); k < num_sack_blks; k++) { sack_blocks[j].start = sack_blocks[k].start; sack_blocks[j].end = sack_blocks[k].end; j++; } num_sack_blks--; goto again; } } } } do_sack_work: rsm = bbr->r_ctl.rc_sacklast; for (i = 0; i < num_sack_blks; i++) { acked = bbr_proc_sack_blk(tp, bbr, &sack_blocks[i], to, &rsm, cts); if (acked) { bbr->r_wanted_output = 1; changed += acked; sack_changed += acked; } } out: *prev_acked = p_acked; if ((sack_changed) && (!IN_RECOVERY(tp->t_flags))) { /* * Ok we have a high probability that we need to go in to * recovery since we have data sack'd */ struct bbr_sendmap *rsm; rsm = bbr_check_recovery_mode(tp, bbr, cts); if (rsm) { /* Enter recovery */ entered_recovery = 1; bbr->r_wanted_output = 1; /* * When we enter recovery we need to assure we send * one packet. */ if (bbr->r_ctl.rc_resend == NULL) { bbr->r_ctl.rc_resend = rsm; } } } if (IN_RECOVERY(tp->t_flags) && (entered_recovery == 0)) { /* * See if we need to rack-retransmit anything if so set it * up as the thing to resend assuming something else is not * already in that position. */ if (bbr->r_ctl.rc_resend == NULL) { bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts); } } /* * We return the amount that changed via sack, this is used by the * ack-received code to augment what was changed between th_ack <-> * snd_una. */ return (sack_changed); } static void bbr_strike_dupack(struct tcp_bbr *bbr) { struct bbr_sendmap *rsm; rsm = TAILQ_FIRST(&bbr->r_ctl.rc_tmap); if (rsm && (rsm->r_dupack < 0xff)) { rsm->r_dupack++; if (rsm->r_dupack >= DUP_ACK_THRESHOLD) bbr->r_wanted_output = 1; } } /* * Return value of 1, we do not need to call bbr_process_data(). * return value of 0, bbr_process_data can be called. * For ret_val if its 0 the TCB is locked and valid, if its non-zero * its unlocked and probably unsafe to touch the TCB. */ static int bbr_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val) { int32_t ourfinisacked = 0; int32_t acked_amount; uint16_t nsegs; int32_t acked; uint32_t lost, sack_changed = 0; struct mbuf *mfree; struct tcp_bbr *bbr; uint32_t prev_acked = 0; bbr = (struct tcp_bbr *)tp->t_fb_ptr; lost = bbr->r_ctl.rc_lost; nsegs = max(1, m->m_pkthdr.lro_nsegs); if (SEQ_GT(th->th_ack, tp->snd_max)) { ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val); bbr->r_wanted_output = 1; return (1); } if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) { /* Process the ack */ if (bbr->rc_in_persist) tp->t_rxtshift = 0; if ((th->th_ack == tp->snd_una) && (tiwin == tp->snd_wnd)) bbr_strike_dupack(bbr); sack_changed = bbr_log_ack(tp, to, th, &prev_acked); } bbr_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime, (bbr->r_ctl.rc_lost > lost)); if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { /* * Old ack, behind the last one rcv'd or a duplicate ack * with SACK info. */ if (th->th_ack == tp->snd_una) { bbr_ack_received(tp, bbr, th, 0, sack_changed, prev_acked, __LINE__, 0); if (bbr->r_state == TCPS_SYN_SENT) { /* * Special case on where we sent SYN. When * the SYN-ACK is processed in syn_sent * state it bumps the snd_una. This causes * us to hit here even though we did ack 1 * byte. * * Go through the nothing left case so we * send data. */ goto nothing_left; } } return (0); } /* * If we reach this point, ACK is not a duplicate, i.e., it ACKs * something we sent. */ if (tp->t_flags & TF_NEEDSYN) { /* * T/TCP: Connection was half-synchronized, and our SYN has * been ACK'd (so connection is now fully synchronized). Go * to non-starred state, increment snd_una for ACK of SYN, * and check if we can do window scaling. */ tp->t_flags &= ~TF_NEEDSYN; tp->snd_una++; /* Do window scaling? */ if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == (TF_RCVD_SCALE | TF_REQ_SCALE)) { tp->rcv_scale = tp->request_r_scale; /* Send window already scaled. */ } } INP_WLOCK_ASSERT(tp->t_inpcb); acked = BYTES_THIS_ACK(tp, th); KMOD_TCPSTAT_ADD(tcps_rcvackpack, (int)nsegs); KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked); /* * If we just performed our first retransmit, and the ACK arrives * within our recovery window, then it was a mistake to do the * retransmit in the first place. Recover our original cwnd and * ssthresh, and proceed to transmit where we left off. */ if (tp->t_flags & TF_PREVVALID) { tp->t_flags &= ~TF_PREVVALID; if (tp->t_rxtshift == 1 && (int)(ticks - tp->t_badrxtwin) < 0) bbr_cong_signal(tp, th, CC_RTO_ERR, NULL); } SOCKBUF_LOCK(&so->so_snd); acked_amount = min(acked, (int)sbavail(&so->so_snd)); tp->snd_wnd -= acked_amount; mfree = sbcut_locked(&so->so_snd, acked_amount); /* NB: sowwakeup_locked() does an implicit unlock. */ sowwakeup_locked(so); m_freem(mfree); if (SEQ_GT(th->th_ack, tp->snd_una)) { bbr_collapse_rtt(tp, bbr, TCP_REXMTVAL(tp)); } tp->snd_una = th->th_ack; bbr_ack_received(tp, bbr, th, acked, sack_changed, prev_acked, __LINE__, (bbr->r_ctl.rc_lost - lost)); if (IN_RECOVERY(tp->t_flags)) { if (SEQ_LT(th->th_ack, tp->snd_recover) && (SEQ_LT(th->th_ack, tp->snd_max))) { tcp_bbr_partialack(tp); } else { bbr_post_recovery(tp); } } if (SEQ_GT(tp->snd_una, tp->snd_recover)) { tp->snd_recover = tp->snd_una; } if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { tp->snd_nxt = tp->snd_max; } if (tp->snd_una == tp->snd_max) { /* Nothing left outstanding */ nothing_left: bbr_log_progress_event(bbr, tp, ticks, PROGRESS_CLEAR, __LINE__); if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0) bbr->rc_tp->t_acktime = 0; if ((sbused(&so->so_snd) == 0) && (tp->t_flags & TF_SENTFIN)) { ourfinisacked = 1; } bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); if (bbr->rc_in_persist == 0) { bbr->r_ctl.rc_went_idle_time = bbr->r_ctl.rc_rcvtime; } sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una); bbr_log_ack_clear(bbr, bbr->r_ctl.rc_rcvtime); /* * We invalidate the last ack here since we * don't want to transfer forward the time * for our sum's calculations. */ if ((tp->t_state >= TCPS_FIN_WAIT_1) && (sbavail(&so->so_snd) == 0) && (tp->t_flags2 & TF2_DROP_AF_DATA)) { /* * The socket was gone and the peer sent data, time * to reset him. */ *ret_val = 1; tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE); /* tcp_close will kill the inp pre-log the Reset */ tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); tp = tcp_close(tp); ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen); BBR_STAT_INC(bbr_dropped_af_data); return (1); } /* Set need output so persist might get set */ bbr->r_wanted_output = 1; } if (ofia) *ofia = ourfinisacked; return (0); } static void bbr_enter_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line) { if (bbr->rc_in_persist == 0) { bbr_timer_cancel(bbr, __LINE__, cts); bbr->r_ctl.rc_last_delay_val = 0; tp->t_rxtshift = 0; bbr->rc_in_persist = 1; bbr->r_ctl.rc_went_idle_time = cts; /* We should be capped when rw went to 0 but just in case */ bbr_log_type_pesist(bbr, cts, 0, line, 1); /* Time freezes for the state, so do the accounting now */ if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { uint32_t time_in; time_in = cts - bbr->r_ctl.rc_bbr_state_time; if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) { int32_t idx; idx = bbr_state_val(bbr); counter_u64_add(bbr_state_time[(idx + 5)], time_in); } else { counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); } } bbr->r_ctl.rc_bbr_state_time = cts; } } static void bbr_restart_after_idle(struct tcp_bbr *bbr, uint32_t cts, uint32_t idle_time) { /* * Note that if idle time does not exceed our * threshold, we do nothing continuing the state * transitions we were last walking through. */ if (idle_time >= bbr_idle_restart_threshold) { if (bbr->rc_use_idle_restart) { bbr->rc_bbr_state = BBR_STATE_IDLE_EXIT; /* * Set our target using BBR_UNIT, so * we increase at a dramatic rate but * we stop when we get the pipe * full again for our current b/w estimate. */ bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT; bbr_set_state_target(bbr, __LINE__); /* Now setup our gains to ramp up */ bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg; bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg; bbr_log_type_statechange(bbr, cts, __LINE__); } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) { bbr_substate_change(bbr, cts, __LINE__, 1); } } } static void bbr_exit_persist(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts, int32_t line) { uint32_t idle_time; if (bbr->rc_in_persist == 0) return; idle_time = bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time); bbr->rc_in_persist = 0; bbr->rc_hit_state_1 = 0; bbr->r_ctl.rc_del_time = cts; /* * We invalidate the last ack here since we * don't want to transfer forward the time * for our sum's calculations. */ if (tcp_in_hpts(bbr->rc_inp)) { tcp_hpts_remove(bbr->rc_inp); bbr->rc_timer_first = 0; bbr->r_ctl.rc_hpts_flags = 0; bbr->r_ctl.rc_last_delay_val = 0; bbr->r_ctl.rc_hptsi_agg_delay = 0; bbr->r_agg_early_set = 0; bbr->r_ctl.rc_agg_early = 0; } bbr_log_type_pesist(bbr, cts, idle_time, line, 0); if (idle_time >= bbr_rtt_probe_time) { /* * This qualifies as a RTT_PROBE session since we drop the * data outstanding to nothing and waited more than * bbr_rtt_probe_time. */ bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_PERSIST, 0); bbr->r_ctl.last_in_probertt = bbr->r_ctl.rc_rtt_shrinks = cts; } tp->t_rxtshift = 0; /* * If in probeBW and we have persisted more than an RTT lets do * special handling. */ /* Force a time based epoch */ bbr_set_epoch(bbr, cts, __LINE__); /* * Setup the lost so we don't count anything against the guy * we have been stuck with during persists. */ bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; /* Time un-freezes for the state */ bbr->r_ctl.rc_bbr_state_time = cts; if ((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) || (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT)) { /* * If we are going back to probe-bw * or probe_rtt, we may need to possibly * do a fast restart. */ bbr_restart_after_idle(bbr, cts, idle_time); } } static void bbr_collapsed_window(struct tcp_bbr *bbr) { /* * Now we must walk the * send map and divide the * ones left stranded. These * guys can't cause us to abort * the connection and are really * "unsent". However if a buggy * client actually did keep some * of the data i.e. collapsed the win * and refused to ack and then opened * the win and acked that data. We would * get into an ack war, the simplier * method then of just pretending we * did not send those segments something * won't work. */ struct bbr_sendmap *rsm, *nrsm; tcp_seq max_seq; uint32_t maxseg; int can_split = 0; int fnd = 0; maxseg = bbr->rc_tp->t_maxseg - bbr->rc_last_options; max_seq = bbr->rc_tp->snd_una + bbr->rc_tp->snd_wnd; bbr_log_type_rwnd_collapse(bbr, max_seq, 1, 0); TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { /* Find the first seq past or at maxseq */ if (rsm->r_flags & BBR_RWND_COLLAPSED) rsm->r_flags &= ~BBR_RWND_COLLAPSED; if (SEQ_GEQ(max_seq, rsm->r_start) && SEQ_GEQ(rsm->r_end, max_seq)) { fnd = 1; break; } } bbr->rc_has_collapsed = 0; if (!fnd) { /* Nothing to do strange */ return; } /* * Now can we split? * * We don't want to split if splitting * would generate too many small segments * less we let an attacker fragment our * send_map and leave us out of memory. */ if ((max_seq != rsm->r_start) && (max_seq != rsm->r_end)){ /* can we split? */ int res1, res2; res1 = max_seq - rsm->r_start; res2 = rsm->r_end - max_seq; if ((res1 >= (maxseg/8)) && (res2 >= (maxseg/8))) { /* No small pieces here */ can_split = 1; } else if (bbr->r_ctl.rc_num_small_maps_alloced < bbr_sack_block_limit) { /* We are under the limit */ can_split = 1; } } /* Ok do we need to split this rsm? */ if (max_seq == rsm->r_start) { /* It's this guy no split required */ nrsm = rsm; } else if (max_seq == rsm->r_end) { /* It's the next one no split required. */ nrsm = TAILQ_NEXT(rsm, r_next); if (nrsm == NULL) { /* Huh? */ return; } } else if (can_split && SEQ_LT(max_seq, rsm->r_end)) { /* yep we need to split it */ nrsm = bbr_alloc_limit(bbr, BBR_LIMIT_TYPE_SPLIT); if (nrsm == NULL) { /* failed XXXrrs what can we do mark the whole? */ nrsm = rsm; goto no_split; } /* Clone it */ bbr_log_type_rwnd_collapse(bbr, max_seq, 3, 0); bbr_clone_rsm(bbr, nrsm, rsm, max_seq); TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_map, rsm, nrsm, r_next); if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&bbr->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } } else { /* * Split not allowed just start here just * use this guy. */ nrsm = rsm; } no_split: BBR_STAT_INC(bbr_collapsed_win); /* reuse fnd as a count */ fnd = 0; TAILQ_FOREACH_FROM(nrsm, &bbr->r_ctl.rc_map, r_next) { nrsm->r_flags |= BBR_RWND_COLLAPSED; fnd++; bbr->rc_has_collapsed = 1; } bbr_log_type_rwnd_collapse(bbr, max_seq, 4, fnd); } static void bbr_un_collapse_window(struct tcp_bbr *bbr) { struct bbr_sendmap *rsm; int cleared = 0; TAILQ_FOREACH_REVERSE(rsm, &bbr->r_ctl.rc_map, bbr_head, r_next) { if (rsm->r_flags & BBR_RWND_COLLAPSED) { /* Clear the flag */ rsm->r_flags &= ~BBR_RWND_COLLAPSED; cleared++; } else break; } bbr_log_type_rwnd_collapse(bbr, (bbr->rc_tp->snd_una + bbr->rc_tp->snd_wnd), 0, cleared); bbr->rc_has_collapsed = 0; } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCB is still * locked. */ static int bbr_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) { /* * Update window information. Don't look at window if no ACK: TAC's * send garbage on first SYN. */ uint16_t nsegs; int32_t tfo_syn; struct tcp_bbr *bbr; bbr = (struct tcp_bbr *)tp->t_fb_ptr; INP_WLOCK_ASSERT(tp->t_inpcb); nsegs = max(1, m->m_pkthdr.lro_nsegs); if ((thflags & TH_ACK) && (SEQ_LT(tp->snd_wl1, th->th_seq) || (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) || (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) { /* keep track of pure window updates */ if (tlen == 0 && tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd) KMOD_TCPSTAT_INC(tcps_rcvwinupd); tp->snd_wnd = tiwin; tp->snd_wl1 = th->th_seq; tp->snd_wl2 = th->th_ack; if (tp->snd_wnd > tp->max_sndwnd) tp->max_sndwnd = tp->snd_wnd; bbr->r_wanted_output = 1; } else if (thflags & TH_ACK) { if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) { tp->snd_wnd = tiwin; tp->snd_wl1 = th->th_seq; tp->snd_wl2 = th->th_ack; } } if (tp->snd_wnd < ctf_outstanding(tp)) /* The peer collapsed its window on us */ bbr_collapsed_window(bbr); else if (bbr->rc_has_collapsed) bbr_un_collapse_window(bbr); /* Was persist timer active and now we have window space? */ if ((bbr->rc_in_persist != 0) && (tp->snd_wnd >= min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr)))) { /* * Make the rate persist at end of persist mode if idle long * enough */ bbr_exit_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); /* Make sure we output to start the timer */ bbr->r_wanted_output = 1; } /* Do we need to enter persist? */ if ((bbr->rc_in_persist == 0) && (tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) && TCPS_HAVEESTABLISHED(tp->t_state) && (tp->snd_max == tp->snd_una) && sbavail(&tp->t_inpcb->inp_socket->so_snd) && (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) { /* No send window.. we must enter persist */ bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); } if (tp->t_flags2 & TF2_DROP_AF_DATA) { m_freem(m); return (0); } /* * We don't support urgent data but * drag along the up just to make sure * if there is a stack switch no one * is surprised. */ tp->rcv_up = tp->rcv_nxt; INP_WLOCK_ASSERT(tp->t_inpcb); /* * Process the segment text, merging it into the TCP sequencing * queue, and arranging for acknowledgment of receipt if necessary. * This process logically involves adjusting tp->rcv_wnd as data is * presented to the user (this happens in tcp_usrreq.c, case * PRU_RCVD). If a FIN has already been received on this connection * then we just ignore the text. */ tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) && IS_FASTOPEN(tp->t_flags)); if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) && TCPS_HAVERCVDFIN(tp->t_state) == 0) { tcp_seq save_start = th->th_seq; tcp_seq save_rnxt = tp->rcv_nxt; int save_tlen = tlen; m_adj(m, drop_hdrlen); /* delayed header drop */ /* * Insert segment which includes th into TCP reassembly * queue with control block tp. Set thflags to whether * reassembly now includes a segment with FIN. This handles * the common case inline (segment is the next to be * received on an established connection, and the queue is * empty), avoiding linkage into and removal from the queue * and repetition of various conversions. Set DELACK for * segments received in order, but ack immediately when * segments are out of order (so fast retransmit can work). */ if (th->th_seq == tp->rcv_nxt && SEGQ_EMPTY(tp) && (TCPS_HAVEESTABLISHED(tp->t_state) || tfo_syn)) { #ifdef NETFLIX_SB_LIMITS u_int mcnt, appended; if (so->so_rcv.sb_shlim) { mcnt = m_memcnt(m); appended = 0; if (counter_fo_get(so->so_rcv.sb_shlim, mcnt, CFO_NOSLEEP, NULL) == false) { counter_u64_add(tcp_sb_shlim_fails, 1); m_freem(m); return (0); } } #endif if (DELAY_ACK(tp, bbr, nsegs) || tfo_syn) { bbr->bbr_segs_rcvd += max(1, nsegs); tp->t_flags |= TF_DELACK; bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); } else { bbr->r_wanted_output = 1; tp->t_flags |= TF_ACKNOW; } tp->rcv_nxt += tlen; if (tlen && ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) && (tp->t_fbyte_in == 0)) { tp->t_fbyte_in = ticks; if (tp->t_fbyte_in == 0) tp->t_fbyte_in = 1; if (tp->t_fbyte_out && tp->t_fbyte_in) tp->t_flags2 |= TF2_FBYTES_COMPLETE; } thflags = tcp_get_flags(th) & TH_FIN; KMOD_TCPSTAT_ADD(tcps_rcvpack, (int)nsegs); KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen); SOCKBUF_LOCK(&so->so_rcv); if (so->so_rcv.sb_state & SBS_CANTRCVMORE) m_freem(m); else #ifdef NETFLIX_SB_LIMITS appended = #endif sbappendstream_locked(&so->so_rcv, m, 0); /* NB: sorwakeup_locked() does an implicit unlock. */ sorwakeup_locked(so); #ifdef NETFLIX_SB_LIMITS if (so->so_rcv.sb_shlim && appended != mcnt) counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended); #endif } else { /* * XXX: Due to the header drop above "th" is * theoretically invalid by now. Fortunately * m_adj() doesn't actually frees any mbufs when * trimming from the head. */ tcp_seq temp = save_start; thflags = tcp_reass(tp, th, &temp, &tlen, m); tp->t_flags |= TF_ACKNOW; if (tp->t_flags & TF_WAKESOR) { tp->t_flags &= ~TF_WAKESOR; /* NB: sorwakeup_locked() does an implicit unlock. */ sorwakeup_locked(so); } } if ((tp->t_flags & TF_SACK_PERMIT) && (save_tlen > 0) && TCPS_HAVEESTABLISHED(tp->t_state)) { if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) { /* * DSACK actually handled in the fastpath * above. */ tcp_update_sack_list(tp, save_start, save_start + save_tlen); } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) { if ((tp->rcv_numsacks >= 1) && (tp->sackblks[0].end == save_start)) { /* * Partial overlap, recorded at todrop * above. */ tcp_update_sack_list(tp, tp->sackblks[0].start, tp->sackblks[0].end); } else { tcp_update_dsack_list(tp, save_start, save_start + save_tlen); } } else if (tlen >= save_tlen) { /* Update of sackblks. */ tcp_update_dsack_list(tp, save_start, save_start + save_tlen); } else if (tlen > 0) { tcp_update_dsack_list(tp, save_start, save_start + tlen); } } } else { m_freem(m); thflags &= ~TH_FIN; } /* * If FIN is received ACK the FIN and let the user know that the * connection is closing. */ if (thflags & TH_FIN) { if (TCPS_HAVERCVDFIN(tp->t_state) == 0) { /* The socket upcall is handled by socantrcvmore. */ socantrcvmore(so); /* * If connection is half-synchronized (ie NEEDSYN * flag on) then delay ACK, so it may be piggybacked * when SYN is sent. Otherwise, since we received a * FIN then no more input can be expected, send ACK * now. */ if (tp->t_flags & TF_NEEDSYN) { tp->t_flags |= TF_DELACK; bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); } else { tp->t_flags |= TF_ACKNOW; } tp->rcv_nxt++; } switch (tp->t_state) { /* * In SYN_RECEIVED and ESTABLISHED STATES enter the * CLOSE_WAIT state. */ case TCPS_SYN_RECEIVED: tp->t_starttime = ticks; /* FALLTHROUGH */ case TCPS_ESTABLISHED: tcp_state_change(tp, TCPS_CLOSE_WAIT); break; /* * If still in FIN_WAIT_1 STATE FIN has not been * acked so enter the CLOSING state. */ case TCPS_FIN_WAIT_1: tcp_state_change(tp, TCPS_CLOSING); break; /* * In FIN_WAIT_2 state enter the TIME_WAIT state, * starting the time-wait timer, turning off the * other standard timers. */ case TCPS_FIN_WAIT_2: bbr->rc_timer_first = 1; bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); INP_WLOCK_ASSERT(tp->t_inpcb); tcp_twstart(tp); return (1); } } /* * Return any desired output. */ if ((tp->t_flags & TF_ACKNOW) || (sbavail(&so->so_snd) > ctf_outstanding(tp))) { bbr->r_wanted_output = 1; } INP_WLOCK_ASSERT(tp->t_inpcb); return (0); } /* * Here nothing is really faster, its just that we * have broken out the fast-data path also just like * the fast-ack. Return 1 if we processed the packet * return 0 if you need to take the "slow-path". */ static int bbr_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t nxt_pkt) { uint16_t nsegs; int32_t newsize = 0; /* automatic sockbuf scaling */ struct tcp_bbr *bbr; #ifdef NETFLIX_SB_LIMITS u_int mcnt, appended; #endif #ifdef TCPDEBUG /* * The size of tcp_saveipgen must be the size of the max ip header, * now IPv6. */ u_char tcp_saveipgen[IP6_HDR_LEN]; struct tcphdr tcp_savetcp; short ostate = 0; #endif /* On the hpts and we would have called output */ bbr = (struct tcp_bbr *)tp->t_fb_ptr; /* * If last ACK falls within this segment's sequence numbers, record * the timestamp. NOTE that the test is modified according to the * latest proposal of the tcplw@cray.com list (Braden 1993/04/26). */ if (bbr->r_ctl.rc_resend != NULL) { return (0); } if (tiwin && tiwin != tp->snd_wnd) { return (0); } if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) { return (0); } if (__predict_false((to->to_flags & TOF_TS) && (TSTMP_LT(to->to_tsval, tp->ts_recent)))) { return (0); } if (__predict_false((th->th_ack != tp->snd_una))) { return (0); } if (__predict_false(tlen > sbspace(&so->so_rcv))) { return (0); } if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); tp->ts_recent = to->to_tsval; } /* * This is a pure, in-sequence data packet with nothing on the * reassembly queue and we have enough buffer space to take it. */ nsegs = max(1, m->m_pkthdr.lro_nsegs); #ifdef NETFLIX_SB_LIMITS if (so->so_rcv.sb_shlim) { mcnt = m_memcnt(m); appended = 0; if (counter_fo_get(so->so_rcv.sb_shlim, mcnt, CFO_NOSLEEP, NULL) == false) { counter_u64_add(tcp_sb_shlim_fails, 1); m_freem(m); return (1); } } #endif /* Clean receiver SACK report if present */ if (tp->rcv_numsacks) tcp_clean_sackreport(tp); KMOD_TCPSTAT_INC(tcps_preddat); tp->rcv_nxt += tlen; if (tlen && ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) && (tp->t_fbyte_in == 0)) { tp->t_fbyte_in = ticks; if (tp->t_fbyte_in == 0) tp->t_fbyte_in = 1; if (tp->t_fbyte_out && tp->t_fbyte_in) tp->t_flags2 |= TF2_FBYTES_COMPLETE; } /* * Pull snd_wl1 up to prevent seq wrap relative to th_seq. */ tp->snd_wl1 = th->th_seq; /* * Pull rcv_up up to prevent seq wrap relative to rcv_nxt. */ tp->rcv_up = tp->rcv_nxt; KMOD_TCPSTAT_ADD(tcps_rcvpack, (int)nsegs); KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen); #ifdef TCPDEBUG if (so->so_options & SO_DEBUG) tcp_trace(TA_INPUT, ostate, tp, (void *)tcp_saveipgen, &tcp_savetcp, 0); #endif newsize = tcp_autorcvbuf(m, th, so, tp, tlen); /* Add data to socket buffer. */ SOCKBUF_LOCK(&so->so_rcv); if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { m_freem(m); } else { /* * Set new socket buffer size. Give up when limit is * reached. */ if (newsize) if (!sbreserve_locked(so, SO_RCV, newsize, NULL)) so->so_rcv.sb_flags &= ~SB_AUTOSIZE; m_adj(m, drop_hdrlen); /* delayed header drop */ #ifdef NETFLIX_SB_LIMITS appended = #endif sbappendstream_locked(&so->so_rcv, m, 0); ctf_calc_rwin(so, tp); } /* NB: sorwakeup_locked() does an implicit unlock. */ sorwakeup_locked(so); #ifdef NETFLIX_SB_LIMITS if (so->so_rcv.sb_shlim && mcnt != appended) counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended); #endif if (DELAY_ACK(tp, bbr, nsegs)) { bbr->bbr_segs_rcvd += max(1, nsegs); tp->t_flags |= TF_DELACK; bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); } else { bbr->r_wanted_output = 1; tp->t_flags |= TF_ACKNOW; } return (1); } /* * This subfunction is used to try to highly optimize the * fast path. We again allow window updates that are * in sequence to remain in the fast-path. We also add * in the __predict's to attempt to help the compiler. * Note that if we return a 0, then we can *not* process * it and the caller should push the packet into the * slow-path. If we return 1, then all is well and * the packet is fully processed. */ static int bbr_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos) { int32_t acked; uint16_t nsegs; uint32_t sack_changed; #ifdef TCPDEBUG /* * The size of tcp_saveipgen must be the size of the max ip header, * now IPv6. */ u_char tcp_saveipgen[IP6_HDR_LEN]; struct tcphdr tcp_savetcp; short ostate = 0; #endif uint32_t prev_acked = 0; struct tcp_bbr *bbr; if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { /* Old ack, behind (or duplicate to) the last one rcv'd */ return (0); } if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) { /* Above what we have sent? */ return (0); } if (__predict_false(tiwin == 0)) { /* zero window */ return (0); } if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) { /* We need a SYN or a FIN, unlikely.. */ return (0); } if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) { /* Timestamp is behind .. old ack with seq wrap? */ return (0); } if (__predict_false(IN_RECOVERY(tp->t_flags))) { /* Still recovering */ return (0); } bbr = (struct tcp_bbr *)tp->t_fb_ptr; if (__predict_false(bbr->r_ctl.rc_resend != NULL)) { /* We are retransmitting */ return (0); } if (__predict_false(bbr->rc_in_persist != 0)) { /* In persist mode */ return (0); } if (bbr->r_ctl.rc_sacked) { /* We have sack holes on our scoreboard */ return (0); } /* Ok if we reach here, we can process a fast-ack */ nsegs = max(1, m->m_pkthdr.lro_nsegs); sack_changed = bbr_log_ack(tp, to, th, &prev_acked); /* * We never detect loss in fast ack [we can't * have a sack and can't be in recovery so * we always pass 0 (nothing detected)]. */ bbr_lt_bw_sampling(bbr, bbr->r_ctl.rc_rcvtime, 0); /* Did the window get updated? */ if (tiwin != tp->snd_wnd) { tp->snd_wnd = tiwin; tp->snd_wl1 = th->th_seq; if (tp->snd_wnd > tp->max_sndwnd) tp->max_sndwnd = tp->snd_wnd; } /* Do we need to exit persists? */ if ((bbr->rc_in_persist != 0) && (tp->snd_wnd >= min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr)))) { bbr_exit_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); bbr->r_wanted_output = 1; } /* Do we need to enter persists? */ if ((bbr->rc_in_persist == 0) && (tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) && TCPS_HAVEESTABLISHED(tp->t_state) && (tp->snd_max == tp->snd_una) && sbavail(&tp->t_inpcb->inp_socket->so_snd) && (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) { /* No send window.. we must enter persist */ bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); } /* * If last ACK falls within this segment's sequence numbers, record * the timestamp. NOTE that the test is modified according to the * latest proposal of the tcplw@cray.com list (Braden 1993/04/26). */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { tp->ts_recent_age = bbr->r_ctl.rc_rcvtime; tp->ts_recent = to->to_tsval; } /* * This is a pure ack for outstanding data. */ KMOD_TCPSTAT_INC(tcps_predack); /* * "bad retransmit" recovery. */ if (tp->t_flags & TF_PREVVALID) { tp->t_flags &= ~TF_PREVVALID; if (tp->t_rxtshift == 1 && (int)(ticks - tp->t_badrxtwin) < 0) bbr_cong_signal(tp, th, CC_RTO_ERR, NULL); } /* * Recalculate the transmit timer / rtt. * * Some boxes send broken timestamp replies during the SYN+ACK * phase, ignore timestamps of 0 or we could calculate a huge RTT * and blow up the retransmit timer. */ acked = BYTES_THIS_ACK(tp, th); #ifdef TCP_HHOOK /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */ hhook_run_tcp_est_in(tp, th, to); #endif KMOD_TCPSTAT_ADD(tcps_rcvackpack, (int)nsegs); KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked); sbdrop(&so->so_snd, acked); if (SEQ_GT(th->th_ack, tp->snd_una)) bbr_collapse_rtt(tp, bbr, TCP_REXMTVAL(tp)); tp->snd_una = th->th_ack; if (tp->snd_wnd < ctf_outstanding(tp)) /* The peer collapsed its window on us */ bbr_collapsed_window(bbr); else if (bbr->rc_has_collapsed) bbr_un_collapse_window(bbr); if (SEQ_GT(tp->snd_una, tp->snd_recover)) { tp->snd_recover = tp->snd_una; } bbr_ack_received(tp, bbr, th, acked, sack_changed, prev_acked, __LINE__, 0); /* * Pull snd_wl2 up to prevent seq wrap relative to th_ack. */ tp->snd_wl2 = th->th_ack; m_freem(m); /* * If all outstanding data are acked, stop retransmit timer, * otherwise restart timer using current (possibly backed-off) * value. If process is waiting for space, wakeup/selwakeup/signal. * If data are ready to send, let tcp_output decide between more * output or persist. */ #ifdef TCPDEBUG if (so->so_options & SO_DEBUG) tcp_trace(TA_INPUT, ostate, tp, (void *)tcp_saveipgen, &tcp_savetcp, 0); #endif /* Wake up the socket if we have room to write more */ sowwakeup(so); if (tp->snd_una == tp->snd_max) { /* Nothing left outstanding */ bbr_log_progress_event(bbr, tp, ticks, PROGRESS_CLEAR, __LINE__); if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0) bbr->rc_tp->t_acktime = 0; bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); if (bbr->rc_in_persist == 0) { bbr->r_ctl.rc_went_idle_time = bbr->r_ctl.rc_rcvtime; } sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una); bbr_log_ack_clear(bbr, bbr->r_ctl.rc_rcvtime); /* * We invalidate the last ack here since we * don't want to transfer forward the time * for our sum's calculations. */ bbr->r_wanted_output = 1; } if (sbavail(&so->so_snd)) { bbr->r_wanted_output = 1; } return (1); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCB is still * locked. */ static int bbr_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t todrop; int32_t ourfinisacked = 0; struct tcp_bbr *bbr; int32_t ret_val = 0; bbr = (struct tcp_bbr *)tp->t_fb_ptr; ctf_calc_rwin(so, tp); /* * If the state is SYN_SENT: if seg contains an ACK, but not for our * SYN, drop the input. if seg contains a RST, then drop the * connection. if seg does not contain SYN, then drop it. Otherwise * this is an acceptable SYN segment initialize tp->rcv_nxt and * tp->irs if seg contains ack then advance tp->snd_una. BRR does * not support ECN so we will not say we are capable. if SYN has * been acked change to ESTABLISHED else SYN_RCVD state arrange for * segment to be acked (eventually) continue processing rest of * data/controls, beginning with URG */ if ((thflags & TH_ACK) && (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) { tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) { TCP_PROBE5(connect__refused, NULL, tp, mtod(m, const char *), tp, th); tp = tcp_drop(tp, ECONNREFUSED); ctf_do_drop(m, tp); return (1); } if (thflags & TH_RST) { ctf_do_drop(m, tp); return (1); } if (!(thflags & TH_SYN)) { ctf_do_drop(m, tp); return (1); } tp->irs = th->th_seq; tcp_rcvseqinit(tp); if (thflags & TH_ACK) { int tfo_partial = 0; KMOD_TCPSTAT_INC(tcps_connects); soisconnected(so); #ifdef MAC mac_socketpeer_set_from_mbuf(m, so); #endif /* Do window scaling on this connection? */ if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == (TF_RCVD_SCALE | TF_REQ_SCALE)) { tp->rcv_scale = tp->request_r_scale; } tp->rcv_adv += min(tp->rcv_wnd, TCP_MAXWIN << tp->rcv_scale); /* * If not all the data that was sent in the TFO SYN * has been acked, resend the remainder right away. */ if (IS_FASTOPEN(tp->t_flags) && (tp->snd_una != tp->snd_max)) { tp->snd_nxt = th->th_ack; tfo_partial = 1; } /* * If there's data, delay ACK; if there's also a FIN ACKNOW * will be turned on later. */ if (DELAY_ACK(tp, bbr, 1) && tlen != 0 && !tfo_partial) { bbr->bbr_segs_rcvd += 1; tp->t_flags |= TF_DELACK; bbr_timer_cancel(bbr, __LINE__, bbr->r_ctl.rc_rcvtime); } else { bbr->r_wanted_output = 1; tp->t_flags |= TF_ACKNOW; } if (SEQ_GT(th->th_ack, tp->iss)) { /* * The SYN is acked * handle it specially. */ bbr_log_syn(tp, to); } if (SEQ_GT(th->th_ack, tp->snd_una)) { /* * We advance snd_una for the * fast open case. If th_ack is * acknowledging data beyond * snd_una we can't just call * ack-processing since the * data stream in our send-map * will start at snd_una + 1 (one * beyond the SYN). If its just * equal we don't need to do that * and there is no send_map. */ tp->snd_una++; } /* * Received in SYN_SENT[*] state. Transitions: * SYN_SENT --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1 */ tp->t_starttime = ticks; if (tp->t_flags & TF_NEEDFIN) { tcp_state_change(tp, TCPS_FIN_WAIT_1); tp->t_flags &= ~TF_NEEDFIN; thflags &= ~TH_SYN; } else { tcp_state_change(tp, TCPS_ESTABLISHED); TCP_PROBE5(connect__established, NULL, tp, mtod(m, const char *), tp, th); cc_conn_init(tp); } } else { /* * Received initial SYN in SYN-SENT[*] state => simultaneous * open. If segment contains CC option and there is a * cached CC, apply TAO test. If it succeeds, connection is * * half-synchronized. Otherwise, do 3-way handshake: * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If * there was no CC option, clear cached CC value. */ tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN | TF_SONOTCONN); tcp_state_change(tp, TCPS_SYN_RECEIVED); } INP_WLOCK_ASSERT(tp->t_inpcb); /* * Advance th->th_seq to correspond to first data byte. If data, * trim to stay within window, dropping FIN if necessary. */ th->th_seq++; if (tlen > tp->rcv_wnd) { todrop = tlen - tp->rcv_wnd; m_adj(m, -todrop); tlen = tp->rcv_wnd; thflags &= ~TH_FIN; KMOD_TCPSTAT_INC(tcps_rcvpackafterwin); KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop); } tp->snd_wl1 = th->th_seq - 1; tp->rcv_up = th->th_seq; /* * Client side of transaction: already sent SYN and data. If the * remote host used T/TCP to validate the SYN, our data will be * ACK'd; if so, enter normal data segment processing in the middle * of step 5, ack processing. Otherwise, goto step 6. */ if (thflags & TH_ACK) { if ((to->to_flags & TOF_TS) != 0) { uint32_t t, rtt; t = tcp_tv_to_mssectick(&bbr->rc_tv); if (TSTMP_GEQ(t, to->to_tsecr)) { rtt = t - to->to_tsecr; if (rtt == 0) { rtt = 1; } rtt *= MS_IN_USEC; tcp_bbr_xmit_timer(bbr, rtt, 0, 0, 0); apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, bbr->r_ctl.rc_rcvtime); } } if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) return (ret_val); /* We may have changed to FIN_WAIT_1 above */ if (tp->t_state == TCPS_FIN_WAIT_1) { /* * In FIN_WAIT_1 STATE in addition to the processing * for the ESTABLISHED state if our FIN is now * acknowledged then enter FIN_WAIT_2. */ if (ourfinisacked) { /* * If we can't receive any more data, then * closing user can proceed. Starting the * timer is contrary to the specification, * but if we don't get a FIN we'll hang * forever. * * XXXjl: we should release the tp also, and * use a compressed state. */ if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { soisdisconnected(so); tcp_timer_activate(tp, TT_2MSL, (tcp_fast_finwait2_recycle ? tcp_finwait2_timeout : TP_MAXIDLE(tp))); } tcp_state_change(tp, TCPS_FIN_WAIT_2); } } } return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCB is still * locked. */ static int bbr_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t ourfinisacked = 0; int32_t ret_val; struct tcp_bbr *bbr; bbr = (struct tcp_bbr *)tp->t_fb_ptr; ctf_calc_rwin(so, tp); if ((thflags & TH_ACK) && (SEQ_LEQ(th->th_ack, tp->snd_una) || SEQ_GT(th->th_ack, tp->snd_max))) { tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } if (IS_FASTOPEN(tp->t_flags)) { /* * When a TFO connection is in SYN_RECEIVED, the only valid * packets are the initial SYN, a retransmit/copy of the * initial SYN (possibly with a subset of the original * data), a valid ACK, a FIN, or a RST. */ if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) { tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } else if (thflags & TH_SYN) { /* non-initial SYN is ignored */ if ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_RXT) || (bbr->r_ctl.rc_hpts_flags & PACE_TMR_TLP) || (bbr->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) { ctf_do_drop(m, NULL); return (0); } } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) { ctf_do_drop(m, NULL); return (0); } } if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (ctf_process_rst(m, th, so, tp)); /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } /* * In the SYN-RECEIVED state, validate that the packet belongs to * this connection before trimming the data to fit the receive * window. Check the sequence number versus IRS since we know the * sequence numbers haven't wrapped. This is a partial fix for the * "LAND" DoS attack. */ if (SEQ_LT(th->th_seq, tp->irs)) { tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } INP_WLOCK_ASSERT(tp->t_inpcb); if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { return (ret_val); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); tp->ts_recent = to->to_tsval; } tp->snd_wnd = tiwin; /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (IS_FASTOPEN(tp->t_flags)) { cc_conn_init(tp); } return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } KMOD_TCPSTAT_INC(tcps_connects); if (tp->t_flags & TF_SONOTCONN) { tp->t_flags &= ~TF_SONOTCONN; soisconnected(so); } /* Do window scaling? */ if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == (TF_RCVD_SCALE | TF_REQ_SCALE)) { tp->rcv_scale = tp->request_r_scale; } /* * ok for the first time in lets see if we can use the ts to figure * out what the initial RTT was. */ if ((to->to_flags & TOF_TS) != 0) { uint32_t t, rtt; t = tcp_tv_to_mssectick(&bbr->rc_tv); if (TSTMP_GEQ(t, to->to_tsecr)) { rtt = t - to->to_tsecr; if (rtt == 0) { rtt = 1; } rtt *= MS_IN_USEC; tcp_bbr_xmit_timer(bbr, rtt, 0, 0, 0); apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, bbr->r_ctl.rc_rcvtime); } } /* Drop off any SYN in the send map (probably not there) */ if (thflags & TH_ACK) bbr_log_syn(tp, to); if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) { tcp_fastopen_decrement_counter(tp->t_tfo_pending); tp->t_tfo_pending = NULL; } /* * Make transitions: SYN-RECEIVED -> ESTABLISHED SYN-RECEIVED* -> * FIN-WAIT-1 */ tp->t_starttime = ticks; if (tp->t_flags & TF_NEEDFIN) { tcp_state_change(tp, TCPS_FIN_WAIT_1); tp->t_flags &= ~TF_NEEDFIN; } else { tcp_state_change(tp, TCPS_ESTABLISHED); TCP_PROBE5(accept__established, NULL, tp, mtod(m, const char *), tp, th); /* * TFO connections call cc_conn_init() during SYN * processing. Calling it again here for such connections * is not harmless as it would undo the snd_cwnd reduction * that occurs when a TFO SYN|ACK is retransmitted. */ if (!IS_FASTOPEN(tp->t_flags)) cc_conn_init(tp); } /* * Account for the ACK of our SYN prior to * regular ACK processing below, except for * simultaneous SYN, which is handled later. */ if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN)) tp->snd_una++; /* * If segment contains data or ACK, will call tcp_reass() later; if * not, do so now to pass queued data to user. */ if (tlen == 0 && (thflags & TH_FIN) == 0) { (void)tcp_reass(tp, (struct tcphdr *)0, NULL, 0, (struct mbuf *)0); if (tp->t_flags & TF_WAKESOR) { tp->t_flags &= ~TF_WAKESOR; /* NB: sorwakeup_locked() does an implicit unlock. */ sorwakeup_locked(so); } } tp->snd_wl1 = th->th_seq - 1; if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { return (ret_val); } if (tp->t_state == TCPS_FIN_WAIT_1) { /* We could have went to FIN_WAIT_1 (or EST) above */ /* * In FIN_WAIT_1 STATE in addition to the processing for the * ESTABLISHED state if our FIN is now acknowledged then * enter FIN_WAIT_2. */ if (ourfinisacked) { /* * If we can't receive any more data, then closing * user can proceed. Starting the timer is contrary * to the specification, but if we don't get a FIN * we'll hang forever. * * XXXjl: we should release the tp also, and use a * compressed state. */ if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { soisdisconnected(so); tcp_timer_activate(tp, TT_2MSL, (tcp_fast_finwait2_recycle ? tcp_finwait2_timeout : TP_MAXIDLE(tp))); } tcp_state_change(tp, TCPS_FIN_WAIT_2); } } return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCB is still * locked. */ static int bbr_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { struct tcp_bbr *bbr; int32_t ret_val; /* * Header prediction: check for the two common cases of a * uni-directional data xfer. If the packet has no control flags, * is in-sequence, the window didn't change and we're not * retransmitting, it's a candidate. If the length is zero and the * ack moved forward, we're the sender side of the xfer. Just free * the data acked & wake any higher level process that was blocked * waiting for space. If the length is non-zero and the ack didn't * move, we're the receiver side. If we're getting packets in-order * (the reassembly queue is empty), add the data toc The socket * buffer and note that we need a delayed ack. Make sure that the * hidden state-flags are also off. Since we check for * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN. */ bbr = (struct tcp_bbr *)tp->t_fb_ptr; if (bbr->r_ctl.rc_delivered < (4 * tp->t_maxseg)) { /* * If we have delived under 4 segments increase the initial * window if raised by the peer. We use this to determine * dynamic and static rwnd's at the end of a connection. */ bbr->r_ctl.rc_init_rwnd = max(tiwin, tp->snd_wnd); } if (__predict_true(((to->to_flags & TOF_SACK) == 0)) && __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_URG | TH_ACK)) == TH_ACK) && __predict_true(SEGQ_EMPTY(tp)) && __predict_true(th->th_seq == tp->rcv_nxt)) { if (tlen == 0) { if (bbr_fastack(m, th, so, tp, to, drop_hdrlen, tlen, tiwin, nxt_pkt, iptos)) { return (0); } } else { if (bbr_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen, tiwin, nxt_pkt)) { return (0); } } } ctf_calc_rwin(so, tp); if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (ctf_process_rst(m, th, so, tp)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } INP_WLOCK_ASSERT(tp->t_inpcb); if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { return (ret_val); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); bbr->r_wanted_output = 1; return (ret_val); } else { ctf_do_drop(m, NULL); return (0); } } /* * Ack processing. */ if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) { return (ret_val); } if (sbavail(&so->so_snd)) { if (ctf_progress_timeout_check(tp, true)) { bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } } /* State changes only happen in bbr_process_data() */ return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCB is still * locked. */ static int bbr_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { struct tcp_bbr *bbr; int32_t ret_val; bbr = (struct tcp_bbr *)tp->t_fb_ptr; ctf_calc_rwin(so, tp); if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (ctf_process_rst(m, th, so, tp)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } INP_WLOCK_ASSERT(tp->t_inpcb); if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { return (ret_val); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); bbr->r_wanted_output = 1; return (ret_val); } else { ctf_do_drop(m, NULL); return (0); } } /* * Ack processing. */ if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) { return (ret_val); } if (sbavail(&so->so_snd)) { if (ctf_progress_timeout_check(tp, true)) { bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } } return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } static int bbr_check_data_after_close(struct mbuf *m, struct tcp_bbr *bbr, struct tcpcb *tp, int32_t * tlen, struct tcphdr *th, struct socket *so) { if (bbr->rc_allow_data_af_clo == 0) { close_now: tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE); /* tcp_close will kill the inp pre-log the Reset */ tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); tp = tcp_close(tp); KMOD_TCPSTAT_INC(tcps_rcvafterclose); ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen)); return (1); } if (sbavail(&so->so_snd) == 0) goto close_now; /* Ok we allow data that is ignored and a followup reset */ tp->rcv_nxt = th->th_seq + *tlen; tp->t_flags2 |= TF2_DROP_AF_DATA; bbr->r_wanted_output = 1; *tlen = 0; return (0); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCB is still * locked. */ static int bbr_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t ourfinisacked = 0; int32_t ret_val; struct tcp_bbr *bbr; bbr = (struct tcp_bbr *)tp->t_fb_ptr; ctf_calc_rwin(so, tp); if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (ctf_process_rst(m, th, so, tp)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } INP_WLOCK_ASSERT(tp->t_inpcb); if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { return (ret_val); } /* * If new data are received on a connection after the user processes * are gone, then RST the other end. * We call a new function now so we might continue and setup * to reset at all data being ack'd. */ if ((tp->t_flags & TF_CLOSED) && tlen && bbr_check_data_after_close(m, bbr, tp, &tlen, th, so)) return (1); /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); bbr->r_wanted_output = 1; return (ret_val); } else { ctf_do_drop(m, NULL); return (0); } } /* * Ack processing. */ if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { return (ret_val); } if (ourfinisacked) { /* * If we can't receive any more data, then closing user can * proceed. Starting the timer is contrary to the * specification, but if we don't get a FIN we'll hang * forever. * * XXXjl: we should release the tp also, and use a * compressed state. */ if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { soisdisconnected(so); tcp_timer_activate(tp, TT_2MSL, (tcp_fast_finwait2_recycle ? tcp_finwait2_timeout : TP_MAXIDLE(tp))); } tcp_state_change(tp, TCPS_FIN_WAIT_2); } if (sbavail(&so->so_snd)) { if (ctf_progress_timeout_check(tp, true)) { bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } } return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCB is still * locked. */ static int bbr_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t ourfinisacked = 0; int32_t ret_val; struct tcp_bbr *bbr; bbr = (struct tcp_bbr *)tp->t_fb_ptr; ctf_calc_rwin(so, tp); if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (ctf_process_rst(m, th, so, tp)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } INP_WLOCK_ASSERT(tp->t_inpcb); if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { return (ret_val); } /* * If new data are received on a connection after the user processes * are gone, then RST the other end. * We call a new function now so we might continue and setup * to reset at all data being ack'd. */ if ((tp->t_flags & TF_CLOSED) && tlen && bbr_check_data_after_close(m, bbr, tp, &tlen, th, so)) return (1); /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); bbr->r_wanted_output = 1; return (ret_val); } else { ctf_do_drop(m, NULL); return (0); } } /* * Ack processing. */ if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { return (ret_val); } if (ourfinisacked) { tcp_twstart(tp); m_freem(m); return (1); } if (sbavail(&so->so_snd)) { if (ctf_progress_timeout_check(tp, true)) { bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } } return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCB is still * locked. */ static int bbr_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t ourfinisacked = 0; int32_t ret_val; struct tcp_bbr *bbr; bbr = (struct tcp_bbr *)tp->t_fb_ptr; ctf_calc_rwin(so, tp); if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (ctf_process_rst(m, th, so, tp)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } INP_WLOCK_ASSERT(tp->t_inpcb); if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { return (ret_val); } /* * If new data are received on a connection after the user processes * are gone, then RST the other end. * We call a new function now so we might continue and setup * to reset at all data being ack'd. */ if ((tp->t_flags & TF_CLOSED) && tlen && bbr_check_data_after_close(m, bbr, tp, &tlen, th, so)) return (1); /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); bbr->r_wanted_output = 1; return (ret_val); } else { ctf_do_drop(m, NULL); return (0); } } /* * case TCPS_LAST_ACK: Ack processing. */ if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { return (ret_val); } if (ourfinisacked) { tp = tcp_close(tp); ctf_do_drop(m, tp); return (1); } if (sbavail(&so->so_snd)) { if (ctf_progress_timeout_check(tp, true)) { bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } } return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCB is still * locked. */ static int bbr_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t ourfinisacked = 0; int32_t ret_val; struct tcp_bbr *bbr; bbr = (struct tcp_bbr *)tp->t_fb_ptr; ctf_calc_rwin(so, tp); /* Reset receive buffer auto scaling when not in bulk receive mode. */ if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (ctf_process_rst(m, th, so, tp)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, &ret_val); return (ret_val); } INP_WLOCK_ASSERT(tp->t_inpcb); /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } INP_WLOCK_ASSERT(tp->t_inpcb); if (ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val)) { return (ret_val); } /* * If new data are received on a connection after the user processes * are gone, then we may RST the other end depending on the outcome * of bbr_check_data_after_close. * We call a new function now so we might continue and setup * to reset at all data being ack'd. */ if ((tp->t_flags & TF_CLOSED) && tlen && bbr_check_data_after_close(m, bbr, tp, &tlen, th, so)) return (1); /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ INP_WLOCK_ASSERT(tp->t_inpcb); if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); bbr->r_wanted_output = 1; return (ret_val); } else { ctf_do_drop(m, NULL); return (0); } } /* * Ack processing. */ INP_WLOCK_ASSERT(tp->t_inpcb); if (bbr_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { return (ret_val); } if (sbavail(&so->so_snd)) { if (ctf_progress_timeout_check(tp, true)) { bbr_log_progress_event(bbr, tp, tick, PROGRESS_DROP, __LINE__); ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } } INP_WLOCK_ASSERT(tp->t_inpcb); return (bbr_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } static void bbr_stop_all_timers(struct tcpcb *tp) { struct tcp_bbr *bbr; /* * Assure no timers are running. */ if (tcp_timer_active(tp, TT_PERSIST)) { /* We enter in persists, set the flag appropriately */ bbr = (struct tcp_bbr *)tp->t_fb_ptr; bbr->rc_in_persist = 1; } tcp_timer_suspend(tp, TT_PERSIST); tcp_timer_suspend(tp, TT_REXMT); tcp_timer_suspend(tp, TT_KEEP); tcp_timer_suspend(tp, TT_DELACK); } static void bbr_google_mode_on(struct tcp_bbr *bbr) { bbr->rc_use_google = 1; bbr->rc_no_pacing = 0; bbr->r_ctl.bbr_google_discount = bbr_google_discount; bbr->r_use_policer = bbr_policer_detection_enabled; bbr->r_ctl.rc_probertt_int = (USECS_IN_SECOND * 10); bbr->bbr_use_rack_cheat = 0; bbr->r_ctl.rc_incr_tmrs = 0; bbr->r_ctl.rc_inc_tcp_oh = 0; bbr->r_ctl.rc_inc_ip_oh = 0; bbr->r_ctl.rc_inc_enet_oh = 0; reset_time(&bbr->r_ctl.rc_delrate, BBR_NUM_RTTS_FOR_GOOG_DEL_LIMIT); reset_time_small(&bbr->r_ctl.rc_rttprop, (11 * USECS_IN_SECOND)); tcp_bbr_tso_size_check(bbr, tcp_get_usecs(&bbr->rc_tv)); } static void bbr_google_mode_off(struct tcp_bbr *bbr) { bbr->rc_use_google = 0; bbr->r_ctl.bbr_google_discount = 0; bbr->no_pacing_until = bbr_no_pacing_until; bbr->r_use_policer = 0; if (bbr->no_pacing_until) bbr->rc_no_pacing = 1; else bbr->rc_no_pacing = 0; if (bbr_use_rack_resend_cheat) bbr->bbr_use_rack_cheat = 1; else bbr->bbr_use_rack_cheat = 0; if (bbr_incr_timers) bbr->r_ctl.rc_incr_tmrs = 1; else bbr->r_ctl.rc_incr_tmrs = 0; if (bbr_include_tcp_oh) bbr->r_ctl.rc_inc_tcp_oh = 1; else bbr->r_ctl.rc_inc_tcp_oh = 0; if (bbr_include_ip_oh) bbr->r_ctl.rc_inc_ip_oh = 1; else bbr->r_ctl.rc_inc_ip_oh = 0; if (bbr_include_enet_oh) bbr->r_ctl.rc_inc_enet_oh = 1; else bbr->r_ctl.rc_inc_enet_oh = 0; bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit; reset_time(&bbr->r_ctl.rc_delrate, bbr_num_pktepo_for_del_limit); reset_time_small(&bbr->r_ctl.rc_rttprop, (bbr_filter_len_sec * USECS_IN_SECOND)); tcp_bbr_tso_size_check(bbr, tcp_get_usecs(&bbr->rc_tv)); } /* * Return 0 on success, non-zero on failure * which indicates the error (usually no memory). */ static int bbr_init(struct tcpcb *tp) { struct tcp_bbr *bbr = NULL; struct inpcb *inp; uint32_t cts; tp->t_fb_ptr = uma_zalloc(bbr_pcb_zone, (M_NOWAIT | M_ZERO)); if (tp->t_fb_ptr == NULL) { /* * We need to allocate memory but cant. The INP and INP_INFO * locks and they are recursive (happens during setup. So a * scheme to drop the locks fails :( * */ return (ENOMEM); } bbr = (struct tcp_bbr *)tp->t_fb_ptr; bbr->rtt_valid = 0; inp = tp->t_inpcb; inp->inp_flags2 |= INP_CANNOT_DO_ECN; inp->inp_flags2 |= INP_SUPPORTS_MBUFQ; TAILQ_INIT(&bbr->r_ctl.rc_map); TAILQ_INIT(&bbr->r_ctl.rc_free); TAILQ_INIT(&bbr->r_ctl.rc_tmap); bbr->rc_tp = tp; if (tp->t_inpcb) { bbr->rc_inp = tp->t_inpcb; } cts = tcp_get_usecs(&bbr->rc_tv); tp->t_acktime = 0; bbr->rc_allow_data_af_clo = bbr_ignore_data_after_close; bbr->r_ctl.rc_reorder_fade = bbr_reorder_fade; bbr->rc_tlp_threshold = bbr_tlp_thresh; bbr->r_ctl.rc_reorder_shift = bbr_reorder_thresh; bbr->r_ctl.rc_pkt_delay = bbr_pkt_delay; bbr->r_ctl.rc_min_to = bbr_min_to; bbr->rc_bbr_state = BBR_STATE_STARTUP; bbr->r_ctl.bbr_lost_at_state = 0; bbr->r_ctl.rc_lost_at_startup = 0; bbr->rc_all_timers_stopped = 0; bbr->r_ctl.rc_bbr_lastbtlbw = 0; bbr->r_ctl.rc_pkt_epoch_del = 0; bbr->r_ctl.rc_pkt_epoch = 0; bbr->r_ctl.rc_lowest_rtt = 0xffffffff; bbr->r_ctl.rc_bbr_hptsi_gain = bbr_high_gain; bbr->r_ctl.rc_bbr_cwnd_gain = bbr_high_gain; bbr->r_ctl.rc_went_idle_time = cts; bbr->rc_pacer_started = cts; bbr->r_ctl.rc_pkt_epoch_time = cts; bbr->r_ctl.rc_rcvtime = cts; bbr->r_ctl.rc_bbr_state_time = cts; bbr->r_ctl.rc_del_time = cts; bbr->r_ctl.rc_tlp_rxt_last_time = cts; bbr->r_ctl.last_in_probertt = cts; bbr->skip_gain = 0; bbr->gain_is_limited = 0; bbr->no_pacing_until = bbr_no_pacing_until; if (bbr->no_pacing_until) bbr->rc_no_pacing = 1; if (bbr_use_google_algo) { bbr->rc_no_pacing = 0; bbr->rc_use_google = 1; bbr->r_ctl.bbr_google_discount = bbr_google_discount; bbr->r_use_policer = bbr_policer_detection_enabled; } else { bbr->rc_use_google = 0; bbr->r_ctl.bbr_google_discount = 0; bbr->r_use_policer = 0; } if (bbr_ts_limiting) bbr->rc_use_ts_limit = 1; else bbr->rc_use_ts_limit = 0; if (bbr_ts_can_raise) bbr->ts_can_raise = 1; else bbr->ts_can_raise = 0; if (V_tcp_delack_enabled == 1) tp->t_delayed_ack = 2; else if (V_tcp_delack_enabled == 0) tp->t_delayed_ack = 0; else if (V_tcp_delack_enabled < 100) tp->t_delayed_ack = V_tcp_delack_enabled; else tp->t_delayed_ack = 2; if (bbr->rc_use_google == 0) bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit; else bbr->r_ctl.rc_probertt_int = (USECS_IN_SECOND * 10); bbr->r_ctl.rc_min_rto_ms = bbr_rto_min_ms; bbr->rc_max_rto_sec = bbr_rto_max_sec; bbr->rc_init_win = bbr_def_init_win; if (tp->t_flags & TF_REQ_TSTMP) bbr->rc_last_options = TCP_TS_OVERHEAD; bbr->r_ctl.rc_pace_max_segs = tp->t_maxseg - bbr->rc_last_options; bbr->r_ctl.rc_high_rwnd = tp->snd_wnd; bbr->r_init_rtt = 1; counter_u64_add(bbr_flows_nohdwr_pacing, 1); if (bbr_allow_hdwr_pacing) bbr->bbr_hdw_pace_ena = 1; else bbr->bbr_hdw_pace_ena = 0; if (bbr_sends_full_iwnd) bbr->bbr_init_win_cheat = 1; else bbr->bbr_init_win_cheat = 0; bbr->r_ctl.bbr_utter_max = bbr_hptsi_utter_max; bbr->r_ctl.rc_drain_pg = bbr_drain_gain; bbr->r_ctl.rc_startup_pg = bbr_high_gain; bbr->rc_loss_exit = bbr_exit_startup_at_loss; bbr->r_ctl.bbr_rttprobe_gain_val = bbr_rttprobe_gain; bbr->r_ctl.bbr_hptsi_per_second = bbr_hptsi_per_second; bbr->r_ctl.bbr_hptsi_segments_delay_tar = bbr_hptsi_segments_delay_tar; bbr->r_ctl.bbr_hptsi_segments_max = bbr_hptsi_segments_max; bbr->r_ctl.bbr_hptsi_segments_floor = bbr_hptsi_segments_floor; bbr->r_ctl.bbr_hptsi_bytes_min = bbr_hptsi_bytes_min; bbr->r_ctl.bbr_cross_over = bbr_cross_over; bbr->r_ctl.rc_rtt_shrinks = cts; if (bbr->rc_use_google) { setup_time_filter(&bbr->r_ctl.rc_delrate, FILTER_TYPE_MAX, BBR_NUM_RTTS_FOR_GOOG_DEL_LIMIT); setup_time_filter_small(&bbr->r_ctl.rc_rttprop, FILTER_TYPE_MIN, (11 * USECS_IN_SECOND)); } else { setup_time_filter(&bbr->r_ctl.rc_delrate, FILTER_TYPE_MAX, bbr_num_pktepo_for_del_limit); setup_time_filter_small(&bbr->r_ctl.rc_rttprop, FILTER_TYPE_MIN, (bbr_filter_len_sec * USECS_IN_SECOND)); } bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_INIT, 0); if (bbr_uses_idle_restart) bbr->rc_use_idle_restart = 1; else bbr->rc_use_idle_restart = 0; bbr->r_ctl.rc_bbr_cur_del_rate = 0; bbr->r_ctl.rc_initial_hptsi_bw = bbr_initial_bw_bps; if (bbr_resends_use_tso) bbr->rc_resends_use_tso = 1; #ifdef NETFLIX_PEAKRATE tp->t_peakrate_thr = tp->t_maxpeakrate; #endif if (tp->snd_una != tp->snd_max) { /* Create a send map for the current outstanding data */ struct bbr_sendmap *rsm; rsm = bbr_alloc(bbr); if (rsm == NULL) { uma_zfree(bbr_pcb_zone, tp->t_fb_ptr); tp->t_fb_ptr = NULL; return (ENOMEM); } rsm->r_rtt_not_allowed = 1; rsm->r_tim_lastsent[0] = cts; rsm->r_rtr_cnt = 1; rsm->r_rtr_bytes = 0; rsm->r_start = tp->snd_una; rsm->r_end = tp->snd_max; rsm->r_dupack = 0; rsm->r_delivered = bbr->r_ctl.rc_delivered; rsm->r_ts_valid = 0; rsm->r_del_ack_ts = tp->ts_recent; rsm->r_del_time = cts; if (bbr->r_ctl.r_app_limited_until) rsm->r_app_limited = 1; else rsm->r_app_limited = 0; TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_map, rsm, r_next); TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 1; if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) rsm->r_bbr_state = bbr_state_val(bbr); else rsm->r_bbr_state = 8; } if (bbr_use_rack_resend_cheat && (bbr->rc_use_google == 0)) bbr->bbr_use_rack_cheat = 1; if (bbr_incr_timers && (bbr->rc_use_google == 0)) bbr->r_ctl.rc_incr_tmrs = 1; if (bbr_include_tcp_oh && (bbr->rc_use_google == 0)) bbr->r_ctl.rc_inc_tcp_oh = 1; if (bbr_include_ip_oh && (bbr->rc_use_google == 0)) bbr->r_ctl.rc_inc_ip_oh = 1; if (bbr_include_enet_oh && (bbr->rc_use_google == 0)) bbr->r_ctl.rc_inc_enet_oh = 1; bbr_log_type_statechange(bbr, cts, __LINE__); if (TCPS_HAVEESTABLISHED(tp->t_state) && (tp->t_srtt)) { uint32_t rtt; rtt = (TICKS_2_USEC(tp->t_srtt) >> TCP_RTT_SHIFT); apply_filter_min_small(&bbr->r_ctl.rc_rttprop, rtt, cts); } /* announce the settings and state */ bbr_log_settings_change(bbr, BBR_RECOVERY_LOWRTT); tcp_bbr_tso_size_check(bbr, cts); /* * Now call the generic function to start a timer. This will place * the TCB on the hptsi wheel if a timer is needed with appropriate * flags. */ bbr_stop_all_timers(tp); bbr_start_hpts_timer(bbr, tp, cts, 5, 0, 0); return (0); } /* * Return 0 if we can accept the connection. Return * non-zero if we can't handle the connection. A EAGAIN * means you need to wait until the connection is up. * a EADDRNOTAVAIL means we can never handle the connection * (no SACK). */ static int bbr_handoff_ok(struct tcpcb *tp) { if ((tp->t_state == TCPS_CLOSED) || (tp->t_state == TCPS_LISTEN)) { /* Sure no problem though it may not stick */ return (0); } if ((tp->t_state == TCPS_SYN_SENT) || (tp->t_state == TCPS_SYN_RECEIVED)) { /* * We really don't know you have to get to ESTAB or beyond * to tell. */ return (EAGAIN); } if (tp->t_flags & TF_SENTFIN) return (EINVAL); if ((tp->t_flags & TF_SACK_PERMIT) || bbr_sack_not_required) { return (0); } /* * If we reach here we don't do SACK on this connection so we can * never do rack. */ return (EINVAL); } static void bbr_fini(struct tcpcb *tp, int32_t tcb_is_purged) { if (tp->t_fb_ptr) { uint32_t calc; struct tcp_bbr *bbr; struct bbr_sendmap *rsm; bbr = (struct tcp_bbr *)tp->t_fb_ptr; if (bbr->r_ctl.crte) tcp_rel_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp); bbr_log_flowend(bbr); bbr->rc_tp = NULL; if (tp->t_inpcb) { /* Backout any flags2 we applied */ tp->t_inpcb->inp_flags2 &= ~INP_CANNOT_DO_ECN; tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ; tp->t_inpcb->inp_flags2 &= ~INP_MBUF_QUEUE_READY; } if (bbr->bbr_hdrw_pacing) counter_u64_add(bbr_flows_whdwr_pacing, -1); else counter_u64_add(bbr_flows_nohdwr_pacing, -1); if (bbr->r_ctl.crte != NULL) { tcp_rel_pacing_rate(bbr->r_ctl.crte, tp); bbr->r_ctl.crte = NULL; } rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); while (rsm) { TAILQ_REMOVE(&bbr->r_ctl.rc_map, rsm, r_next); uma_zfree(bbr_zone, rsm); rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); } rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free); while (rsm) { TAILQ_REMOVE(&bbr->r_ctl.rc_free, rsm, r_next); uma_zfree(bbr_zone, rsm); rsm = TAILQ_FIRST(&bbr->r_ctl.rc_free); } calc = bbr->r_ctl.rc_high_rwnd - bbr->r_ctl.rc_init_rwnd; if (calc > (bbr->r_ctl.rc_init_rwnd / 10)) BBR_STAT_INC(bbr_dynamic_rwnd); else BBR_STAT_INC(bbr_static_rwnd); bbr->r_ctl.rc_free_cnt = 0; uma_zfree(bbr_pcb_zone, tp->t_fb_ptr); tp->t_fb_ptr = NULL; } /* Make sure snd_nxt is correctly set */ tp->snd_nxt = tp->snd_max; } static void bbr_set_state(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t win) { switch (tp->t_state) { case TCPS_SYN_SENT: bbr->r_state = TCPS_SYN_SENT; bbr->r_substate = bbr_do_syn_sent; break; case TCPS_SYN_RECEIVED: bbr->r_state = TCPS_SYN_RECEIVED; bbr->r_substate = bbr_do_syn_recv; break; case TCPS_ESTABLISHED: bbr->r_ctl.rc_init_rwnd = max(win, bbr->rc_tp->snd_wnd); bbr->r_state = TCPS_ESTABLISHED; bbr->r_substate = bbr_do_established; break; case TCPS_CLOSE_WAIT: bbr->r_state = TCPS_CLOSE_WAIT; bbr->r_substate = bbr_do_close_wait; break; case TCPS_FIN_WAIT_1: bbr->r_state = TCPS_FIN_WAIT_1; bbr->r_substate = bbr_do_fin_wait_1; break; case TCPS_CLOSING: bbr->r_state = TCPS_CLOSING; bbr->r_substate = bbr_do_closing; break; case TCPS_LAST_ACK: bbr->r_state = TCPS_LAST_ACK; bbr->r_substate = bbr_do_lastack; break; case TCPS_FIN_WAIT_2: bbr->r_state = TCPS_FIN_WAIT_2; bbr->r_substate = bbr_do_fin_wait_2; break; case TCPS_LISTEN: case TCPS_CLOSED: case TCPS_TIME_WAIT: default: break; }; } static void bbr_substate_change(struct tcp_bbr *bbr, uint32_t cts, int32_t line, int dolog) { /* * Now what state are we going into now? Is there adjustments * needed? */ int32_t old_state; old_state = bbr_state_val(bbr); if (bbr_state_val(bbr) == BBR_SUB_LEVEL1) { /* Save the lowest srtt we saw in our end of the sub-state */ bbr->rc_hit_state_1 = 0; if (bbr->r_ctl.bbr_smallest_srtt_this_state != 0xffffffff) bbr->r_ctl.bbr_smallest_srtt_state2 = bbr->r_ctl.bbr_smallest_srtt_this_state; } bbr->rc_bbr_substate++; if (bbr->rc_bbr_substate >= BBR_SUBSTATE_COUNT) { /* Cycle back to first state-> gain */ bbr->rc_bbr_substate = 0; } if (bbr_state_val(bbr) == BBR_SUB_GAIN) { /* * We enter the gain(5/4) cycle (possibly less if * shallow buffer detection is enabled) */ if (bbr->skip_gain) { /* * Hardware pacing has set our rate to * the max and limited our b/w just * do level i.e. no gain. */ bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_LEVEL1]; } else if (bbr->gain_is_limited && bbr->bbr_hdrw_pacing && bbr->r_ctl.crte) { /* * We can't gain above the hardware pacing * rate which is less than our rate + the gain * calculate the gain needed to reach the hardware * pacing rate.. */ uint64_t bw, rate, gain_calc; bw = bbr_get_bw(bbr); rate = bbr->r_ctl.crte->rate; if ((rate > bw) && (((bw * (uint64_t)bbr_hptsi_gain[BBR_SUB_GAIN]) / (uint64_t)BBR_UNIT) > rate)) { gain_calc = (rate * BBR_UNIT) / bw; if (gain_calc < BBR_UNIT) gain_calc = BBR_UNIT; bbr->r_ctl.rc_bbr_hptsi_gain = (uint16_t)gain_calc; } else { bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_GAIN]; } } else bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_GAIN]; if ((bbr->rc_use_google == 0) && (bbr_gain_to_target == 0)) { bbr->r_ctl.rc_bbr_state_atflight = cts; } else bbr->r_ctl.rc_bbr_state_atflight = 0; } else if (bbr_state_val(bbr) == BBR_SUB_DRAIN) { bbr->rc_hit_state_1 = 1; bbr->r_ctl.rc_exta_time_gd = 0; bbr->r_ctl.flightsize_at_drain = ctf_flight_size(bbr->rc_tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); if (bbr_state_drain_2_tar) { bbr->r_ctl.rc_bbr_state_atflight = 0; } else bbr->r_ctl.rc_bbr_state_atflight = cts; bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[BBR_SUB_DRAIN]; } else { /* All other cycles hit here 2-7 */ if ((old_state == BBR_SUB_DRAIN) && bbr->rc_hit_state_1) { if (bbr_sub_drain_slam_cwnd && (bbr->rc_use_google == 0) && (bbr->rc_tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) { bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd; bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); } if ((cts - bbr->r_ctl.rc_bbr_state_time) > bbr_get_rtt(bbr, BBR_RTT_PROP)) bbr->r_ctl.rc_exta_time_gd += ((cts - bbr->r_ctl.rc_bbr_state_time) - bbr_get_rtt(bbr, BBR_RTT_PROP)); else bbr->r_ctl.rc_exta_time_gd = 0; if (bbr->r_ctl.rc_exta_time_gd) { bbr->r_ctl.rc_level_state_extra = bbr->r_ctl.rc_exta_time_gd; /* Now chop up the time for each state (div by 7) */ bbr->r_ctl.rc_level_state_extra /= 7; if (bbr_rand_ot && bbr->r_ctl.rc_level_state_extra) { /* Add a randomization */ bbr_randomize_extra_state_time(bbr); } } } bbr->r_ctl.rc_bbr_state_atflight = max(1, cts); bbr->r_ctl.rc_bbr_hptsi_gain = bbr_hptsi_gain[bbr_state_val(bbr)]; } if (bbr->rc_use_google) { bbr->r_ctl.rc_bbr_state_atflight = max(1, cts); } bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; bbr->r_ctl.rc_bbr_cwnd_gain = bbr_cwnd_gain; if (dolog) bbr_log_type_statechange(bbr, cts, line); if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { uint32_t time_in; time_in = cts - bbr->r_ctl.rc_bbr_state_time; if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) { counter_u64_add(bbr_state_time[(old_state + 5)], time_in); } else { counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); } } bbr->r_ctl.bbr_smallest_srtt_this_state = 0xffffffff; bbr_set_state_target(bbr, __LINE__); if (bbr_sub_drain_slam_cwnd && (bbr->rc_use_google == 0) && (bbr_state_val(bbr) == BBR_SUB_DRAIN)) { /* Slam down the cwnd */ bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd; bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; if (bbr_sub_drain_app_limit) { /* Go app limited if we are on a long drain */ bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.rc_delivered + ctf_flight_size(bbr->rc_tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes))); } bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); } if (bbr->rc_lt_use_bw) { /* In policed mode we clamp pacing_gain to BBR_UNIT */ bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; } /* Google changes TSO size every cycle */ if (bbr->rc_use_google) tcp_bbr_tso_size_check(bbr, cts); bbr->r_ctl.gain_epoch = cts; bbr->r_ctl.rc_bbr_state_time = cts; bbr->r_ctl.substate_pe = bbr->r_ctl.rc_pkt_epoch; } static void bbr_set_probebw_google_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses) { if ((bbr_state_val(bbr) == BBR_SUB_DRAIN) && (google_allow_early_out == 1) && (bbr->r_ctl.rc_flight_at_input <= bbr->r_ctl.rc_target_at_state)) { /* We have reached out target flight size possibly early */ goto change_state; } if (TSTMP_LT(cts, bbr->r_ctl.rc_bbr_state_time)) { return; } if ((cts - bbr->r_ctl.rc_bbr_state_time) < bbr_get_rtt(bbr, BBR_RTT_PROP)) { /* * Must be a rttProp movement forward before * we can change states. */ return; } if (bbr_state_val(bbr) == BBR_SUB_GAIN) { /* * The needed time has passed but for * the gain cycle extra rules apply: * 1) If we have seen loss, we exit * 2) If we have not reached the target * we stay in GAIN (gain-to-target). */ if (google_consider_lost && losses) goto change_state; if (bbr->r_ctl.rc_target_at_state > bbr->r_ctl.rc_flight_at_input) { return; } } change_state: /* For gain we must reach our target, all others last 1 rttProp */ bbr_substate_change(bbr, cts, __LINE__, 1); } static void bbr_set_probebw_gains(struct tcp_bbr *bbr, uint32_t cts, uint32_t losses) { uint32_t flight, bbr_cur_cycle_time; if (bbr->rc_use_google) { bbr_set_probebw_google_gains(bbr, cts, losses); return; } if (cts == 0) { /* * Never alow cts to be 0 we * do this so we can judge if * we have set a timestamp. */ cts = 1; } if (bbr_state_is_pkt_epoch) bbr_cur_cycle_time = bbr_get_rtt(bbr, BBR_RTT_PKTRTT); else bbr_cur_cycle_time = bbr_get_rtt(bbr, BBR_RTT_PROP); if (bbr->r_ctl.rc_bbr_state_atflight == 0) { if (bbr_state_val(bbr) == BBR_SUB_DRAIN) { flight = ctf_flight_size(bbr->rc_tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); if (bbr_sub_drain_slam_cwnd && bbr->rc_hit_state_1) { /* Keep it slam down */ if (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state) { bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); } if (bbr_sub_drain_app_limit) { /* Go app limited if we are on a long drain */ bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.rc_delivered + flight); } } if (TSTMP_GT(cts, bbr->r_ctl.gain_epoch) && (((cts - bbr->r_ctl.gain_epoch) > bbr_get_rtt(bbr, BBR_RTT_PROP)) || (flight >= bbr->r_ctl.flightsize_at_drain))) { /* * Still here after the same time as * the gain. We need to drain harder * for the next srtt. Reduce by a set amount * the gain drop is capped at DRAIN states * value (88). */ bbr->r_ctl.flightsize_at_drain = flight; if (bbr_drain_drop_mul && bbr_drain_drop_div && (bbr_drain_drop_mul < bbr_drain_drop_div)) { /* Use your specific drop value (def 4/5 = 20%) */ bbr->r_ctl.rc_bbr_hptsi_gain *= bbr_drain_drop_mul; bbr->r_ctl.rc_bbr_hptsi_gain /= bbr_drain_drop_div; } else { /* You get drop of 20% */ bbr->r_ctl.rc_bbr_hptsi_gain *= 4; bbr->r_ctl.rc_bbr_hptsi_gain /= 5; } if (bbr->r_ctl.rc_bbr_hptsi_gain <= bbr_drain_floor) { /* Reduce our gain again to the bottom */ bbr->r_ctl.rc_bbr_hptsi_gain = max(bbr_drain_floor, 1); } bbr_log_exit_gain(bbr, cts, 4); /* * Extend out so we wait another * epoch before dropping again. */ bbr->r_ctl.gain_epoch = cts; } if (flight <= bbr->r_ctl.rc_target_at_state) { if (bbr_sub_drain_slam_cwnd && (bbr->rc_use_google == 0) && (bbr->rc_tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) { bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd; bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); } bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1); bbr_log_exit_gain(bbr, cts, 3); } } else { /* Its a gain */ if (bbr->r_ctl.rc_lost > bbr->r_ctl.bbr_lost_at_state) { bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1); goto change_state; } if ((ctf_outstanding(bbr->rc_tp) >= bbr->r_ctl.rc_target_at_state) || ((ctf_outstanding(bbr->rc_tp) + bbr->rc_tp->t_maxseg - 1) >= bbr->rc_tp->snd_wnd)) { bbr->r_ctl.rc_bbr_state_atflight = max(cts, 1); bbr_log_exit_gain(bbr, cts, 2); } } /** * We fall through and return always one of two things has * occurred. * 1) We are still not at target * * 2) We reached the target and set rc_bbr_state_atflight * which means we no longer hit this block * next time we are called. */ return; } change_state: if (TSTMP_LT(cts, bbr->r_ctl.rc_bbr_state_time)) return; if ((cts - bbr->r_ctl.rc_bbr_state_time) < bbr_cur_cycle_time) { /* Less than a full time-period has passed */ return; } if (bbr->r_ctl.rc_level_state_extra && (bbr_state_val(bbr) > BBR_SUB_DRAIN) && ((cts - bbr->r_ctl.rc_bbr_state_time) < (bbr_cur_cycle_time + bbr->r_ctl.rc_level_state_extra))) { /* Less than a full time-period + extra has passed */ return; } if (bbr_gain_gets_extra_too && bbr->r_ctl.rc_level_state_extra && (bbr_state_val(bbr) == BBR_SUB_GAIN) && ((cts - bbr->r_ctl.rc_bbr_state_time) < (bbr_cur_cycle_time + bbr->r_ctl.rc_level_state_extra))) { /* Less than a full time-period + extra has passed */ return; } bbr_substate_change(bbr, cts, __LINE__, 1); } static uint32_t bbr_get_a_state_target(struct tcp_bbr *bbr, uint32_t gain) { uint32_t mss, tar; if (bbr->rc_use_google) { /* Google just uses the cwnd target */ tar = bbr_get_target_cwnd(bbr, bbr_get_bw(bbr), gain); } else { mss = min((bbr->rc_tp->t_maxseg - bbr->rc_last_options), bbr->r_ctl.rc_pace_max_segs); /* Get the base cwnd with gain rounded to a mss */ tar = roundup(bbr_get_raw_target_cwnd(bbr, bbr_get_bw(bbr), gain), mss); /* Make sure it is within our min */ if (tar < get_min_cwnd(bbr)) return (get_min_cwnd(bbr)); } return (tar); } static void bbr_set_state_target(struct tcp_bbr *bbr, int line) { uint32_t tar, meth; if ((bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) && ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google)) { /* Special case using old probe-rtt method */ tar = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options); meth = 1; } else { /* Non-probe-rtt case and reduced probe-rtt */ if ((bbr->rc_bbr_state == BBR_STATE_PROBE_BW) && (bbr->r_ctl.rc_bbr_hptsi_gain > BBR_UNIT)) { /* For gain cycle we use the hptsi gain */ tar = bbr_get_a_state_target(bbr, bbr->r_ctl.rc_bbr_hptsi_gain); meth = 2; } else if ((bbr_target_is_bbunit) || bbr->rc_use_google) { /* * If configured, or for google all other states * get BBR_UNIT. */ tar = bbr_get_a_state_target(bbr, BBR_UNIT); meth = 3; } else { /* * Or we set a target based on the pacing gain * for non-google mode and default (non-configured). * Note we don't set a target goal below drain (192). */ if (bbr->r_ctl.rc_bbr_hptsi_gain < bbr_hptsi_gain[BBR_SUB_DRAIN]) { tar = bbr_get_a_state_target(bbr, bbr_hptsi_gain[BBR_SUB_DRAIN]); meth = 4; } else { tar = bbr_get_a_state_target(bbr, bbr->r_ctl.rc_bbr_hptsi_gain); meth = 5; } } } bbr_log_set_of_state_target(bbr, tar, line, meth); bbr->r_ctl.rc_target_at_state = tar; } static void bbr_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts, int32_t line) { /* Change to probe_rtt */ uint32_t time_in; bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; bbr->r_ctl.flightsize_at_drain = ctf_flight_size(bbr->rc_tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); bbr->r_ctl.r_app_limited_until = (bbr->r_ctl.flightsize_at_drain + bbr->r_ctl.rc_delivered); /* Setup so we force feed the filter */ if (bbr->rc_use_google || bbr_probertt_sets_rtt) bbr->rc_prtt_set_ts = 1; if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { time_in = cts - bbr->r_ctl.rc_bbr_state_time; counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); } bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_ENTERPROBE, 0); bbr->r_ctl.rc_rtt_shrinks = cts; bbr->r_ctl.last_in_probertt = cts; bbr->r_ctl.rc_probertt_srttchktim = cts; bbr->r_ctl.rc_bbr_state_time = cts; bbr->rc_bbr_state = BBR_STATE_PROBE_RTT; /* We need to force the filter to update */ if ((bbr_sub_drain_slam_cwnd) && bbr->rc_hit_state_1 && (bbr->rc_use_google == 0) && (bbr_state_val(bbr) == BBR_SUB_DRAIN)) { if (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_saved_cwnd) bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd; } else bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd; /* Update the lost */ bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; if ((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google){ /* Set to the non-configurable default of 4 (PROBE_RTT_MIN) */ bbr->rc_tp->snd_cwnd = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options); bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT; bbr_log_set_of_state_target(bbr, bbr->rc_tp->snd_cwnd, __LINE__, 6); bbr->r_ctl.rc_target_at_state = bbr->rc_tp->snd_cwnd; } else { /* * We bring it down slowly by using a hptsi gain that is * probably 75%. This will slowly float down our outstanding * without tampering with the cwnd. */ bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.bbr_rttprobe_gain_val; bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT; bbr_set_state_target(bbr, __LINE__); if (bbr_prtt_slam_cwnd && (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) { bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); } } if (ctf_flight_size(bbr->rc_tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <= bbr->r_ctl.rc_target_at_state) { /* We are at target */ bbr->r_ctl.rc_bbr_enters_probertt = cts; } else { /* We need to come down to reach target before our time begins */ bbr->r_ctl.rc_bbr_enters_probertt = 0; } bbr->r_ctl.rc_pe_of_prtt = bbr->r_ctl.rc_pkt_epoch; BBR_STAT_INC(bbr_enter_probertt); bbr_log_exit_gain(bbr, cts, 0); bbr_log_type_statechange(bbr, cts, line); } static void bbr_check_probe_rtt_limits(struct tcp_bbr *bbr, uint32_t cts) { /* * Sanity check on probe-rtt intervals. * In crazy situations where we are competing * against new-reno flows with huge buffers * our rtt-prop interval could come to dominate * things if we can't get through a full set * of cycles, we need to adjust it. */ if (bbr_can_adjust_probertt && (bbr->rc_use_google == 0)) { uint16_t val = 0; uint32_t cur_rttp, fval, newval, baseval; /* Are we to small and go into probe-rtt to often? */ baseval = (bbr_get_rtt(bbr, BBR_RTT_PROP) * (BBR_SUBSTATE_COUNT + 1)); cur_rttp = roundup(baseval, USECS_IN_SECOND); fval = bbr_filter_len_sec * USECS_IN_SECOND; if (bbr_is_ratio == 0) { if (fval > bbr_rtt_probe_limit) newval = cur_rttp + (fval - bbr_rtt_probe_limit); else newval = cur_rttp; } else { int mul; mul = fval / bbr_rtt_probe_limit; newval = cur_rttp * mul; } if (cur_rttp > bbr->r_ctl.rc_probertt_int) { bbr->r_ctl.rc_probertt_int = cur_rttp; reset_time_small(&bbr->r_ctl.rc_rttprop, newval); val = 1; } else { /* * No adjustments were made * do we need to shrink it? */ if (bbr->r_ctl.rc_probertt_int > bbr_rtt_probe_limit) { if (cur_rttp <= bbr_rtt_probe_limit) { /* * Things have calmed down lets * shrink all the way to default */ bbr->r_ctl.rc_probertt_int = bbr_rtt_probe_limit; reset_time_small(&bbr->r_ctl.rc_rttprop, (bbr_filter_len_sec * USECS_IN_SECOND)); cur_rttp = bbr_rtt_probe_limit; newval = (bbr_filter_len_sec * USECS_IN_SECOND); val = 2; } else { /* * Well does some adjustment make sense? */ if (cur_rttp < bbr->r_ctl.rc_probertt_int) { /* We can reduce interval time some */ bbr->r_ctl.rc_probertt_int = cur_rttp; reset_time_small(&bbr->r_ctl.rc_rttprop, newval); val = 3; } } } } if (val) bbr_log_rtt_shrinks(bbr, cts, cur_rttp, newval, __LINE__, BBR_RTTS_RESETS_VALUES, val); } } static void bbr_exit_probe_rtt(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t cts) { /* Exit probe-rtt */ if (tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd) { tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd; bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); } bbr_log_exit_gain(bbr, cts, 1); bbr->rc_hit_state_1 = 0; bbr->r_ctl.rc_rtt_shrinks = cts; bbr->r_ctl.last_in_probertt = cts; bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_RTTPROBE, 0); bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; bbr->r_ctl.r_app_limited_until = (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + bbr->r_ctl.rc_delivered); if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { uint32_t time_in; time_in = cts - bbr->r_ctl.rc_bbr_state_time; counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); } if (bbr->rc_filled_pipe) { /* Switch to probe_bw */ bbr->rc_bbr_state = BBR_STATE_PROBE_BW; bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts); bbr->r_ctl.rc_bbr_cwnd_gain = bbr_cwnd_gain; bbr_substate_change(bbr, cts, __LINE__, 0); bbr_log_type_statechange(bbr, cts, __LINE__); } else { /* Back to startup */ bbr->rc_bbr_state = BBR_STATE_STARTUP; bbr->r_ctl.rc_bbr_state_time = cts; /* * We don't want to give a complete free 3 * measurements until we exit, so we use * the number of pe's we were in probe-rtt * to add to the startup_epoch. That way * we will still retain the old state. */ bbr->r_ctl.rc_bbr_last_startup_epoch += (bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_pe_of_prtt); bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; /* Make sure to use the lower pg when shifting back in */ if (bbr->r_ctl.rc_lost && bbr_use_lower_gain_in_startup && (bbr->rc_use_google == 0)) bbr->r_ctl.rc_bbr_hptsi_gain = bbr_startup_lower; else bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_startup_pg; bbr->r_ctl.rc_bbr_cwnd_gain = bbr->r_ctl.rc_startup_pg; /* Probably not needed but set it anyway */ bbr_set_state_target(bbr, __LINE__); bbr_log_type_statechange(bbr, cts, __LINE__); bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 0); } bbr_check_probe_rtt_limits(bbr, cts); } static int32_t inline bbr_should_enter_probe_rtt(struct tcp_bbr *bbr, uint32_t cts) { if ((bbr->rc_past_init_win == 1) && (bbr->rc_in_persist == 0) && (bbr_calc_time(cts, bbr->r_ctl.rc_rtt_shrinks) >= bbr->r_ctl.rc_probertt_int)) { return (1); } if (bbr_can_force_probertt && (bbr->rc_in_persist == 0) && (TSTMP_GT(cts, bbr->r_ctl.last_in_probertt)) && ((cts - bbr->r_ctl.last_in_probertt) > bbr->r_ctl.rc_probertt_int)) { return (1); } return (0); } static int32_t bbr_google_startup(struct tcp_bbr *bbr, uint32_t cts, int32_t pkt_epoch) { uint64_t btlbw, gain; if (pkt_epoch == 0) { /* * Need to be on a pkt-epoch to continue. */ return (0); } btlbw = bbr_get_full_bw(bbr); gain = ((bbr->r_ctl.rc_bbr_lastbtlbw * (uint64_t)bbr_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw; if (btlbw >= gain) { bbr->r_ctl.rc_bbr_last_startup_epoch = bbr->r_ctl.rc_pkt_epoch; bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 3); bbr->r_ctl.rc_bbr_lastbtlbw = btlbw; } if ((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS) return (1); bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 8); return(0); } static int32_t inline bbr_state_startup(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, int32_t pkt_epoch) { /* Have we gained 25% in the last 3 packet based epoch's? */ uint64_t btlbw, gain; int do_exit; int delta, rtt_gain; if ((bbr->rc_tp->snd_una == bbr->rc_tp->snd_max) && (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) { /* * This qualifies as a RTT_PROBE session since we drop the * data outstanding to nothing and waited more than * bbr_rtt_probe_time. */ bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0); bbr_set_reduced_rtt(bbr, cts, __LINE__); } if (bbr_should_enter_probe_rtt(bbr, cts)) { bbr_enter_probe_rtt(bbr, cts, __LINE__); return (0); } if (bbr->rc_use_google) return (bbr_google_startup(bbr, cts, pkt_epoch)); if ((bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_startup) && (bbr_use_lower_gain_in_startup)) { /* Drop to a lower gain 1.5 x since we saw loss */ bbr->r_ctl.rc_bbr_hptsi_gain = bbr_startup_lower; } if (pkt_epoch == 0) { /* * Need to be on a pkt-epoch to continue. */ return (0); } if (bbr_rtt_gain_thresh) { /* * Do we allow a flow to stay * in startup with no loss and no * gain in rtt over a set threshold? */ if (bbr->r_ctl.rc_pkt_epoch_rtt && bbr->r_ctl.startup_last_srtt && (bbr->r_ctl.rc_pkt_epoch_rtt > bbr->r_ctl.startup_last_srtt)) { delta = bbr->r_ctl.rc_pkt_epoch_rtt - bbr->r_ctl.startup_last_srtt; rtt_gain = (delta * 100) / bbr->r_ctl.startup_last_srtt; } else rtt_gain = 0; if ((bbr->r_ctl.startup_last_srtt == 0) || (bbr->r_ctl.rc_pkt_epoch_rtt < bbr->r_ctl.startup_last_srtt)) /* First time or new lower value */ bbr->r_ctl.startup_last_srtt = bbr->r_ctl.rc_pkt_epoch_rtt; if ((bbr->r_ctl.rc_lost == 0) && (rtt_gain < bbr_rtt_gain_thresh)) { /* * No loss, and we are under * our gain threhold for * increasing RTT. */ if (bbr->r_ctl.rc_bbr_last_startup_epoch < bbr->r_ctl.rc_pkt_epoch) bbr->r_ctl.rc_bbr_last_startup_epoch++; bbr_log_startup_event(bbr, cts, rtt_gain, delta, bbr->r_ctl.startup_last_srtt, 10); return (0); } } if ((bbr->r_ctl.r_measurement_count == bbr->r_ctl.last_startup_measure) && (bbr->r_ctl.rc_lost_at_startup == bbr->r_ctl.rc_lost) && (!IN_RECOVERY(bbr->rc_tp->t_flags))) { /* * We only assess if we have a new measurement when * we have no loss and are not in recovery. * Drag up by one our last_startup epoch so we will hold * the number of non-gain we have already accumulated. */ if (bbr->r_ctl.rc_bbr_last_startup_epoch < bbr->r_ctl.rc_pkt_epoch) bbr->r_ctl.rc_bbr_last_startup_epoch++; bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 9); return (0); } /* Case where we reduced the lost (bad retransmit) */ if (bbr->r_ctl.rc_lost_at_startup > bbr->r_ctl.rc_lost) bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; bbr->r_ctl.last_startup_measure = bbr->r_ctl.r_measurement_count; btlbw = bbr_get_full_bw(bbr); if (bbr->r_ctl.rc_bbr_hptsi_gain == bbr_startup_lower) gain = ((bbr->r_ctl.rc_bbr_lastbtlbw * (uint64_t)bbr_low_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw; else gain = ((bbr->r_ctl.rc_bbr_lastbtlbw * (uint64_t)bbr_start_exit) / (uint64_t)100) + bbr->r_ctl.rc_bbr_lastbtlbw; do_exit = 0; if (btlbw > bbr->r_ctl.rc_bbr_lastbtlbw) bbr->r_ctl.rc_bbr_lastbtlbw = btlbw; if (btlbw >= gain) { bbr->r_ctl.rc_bbr_last_startup_epoch = bbr->r_ctl.rc_pkt_epoch; /* Update the lost so we won't exit in next set of tests */ bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 3); } if ((bbr->rc_loss_exit && (bbr->r_ctl.rc_lost > bbr->r_ctl.rc_lost_at_startup) && (bbr->r_ctl.rc_pkt_epoch_loss_rate > bbr_startup_loss_thresh)) && ((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS)) { /* * If we had no gain, we had loss and that loss was above * our threshould, the rwnd is not constrained, and we have * had at least 3 packet epochs exit. Note that this is * switched off by sysctl. Google does not do this by the * way. */ if ((ctf_flight_size(bbr->rc_tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + (2 * max(bbr->r_ctl.rc_pace_max_segs, bbr->rc_tp->t_maxseg))) <= bbr->rc_tp->snd_wnd) { do_exit = 1; bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 4); } else { /* Just record an updated loss value */ bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 5); } } else bbr->r_ctl.rc_lost_at_startup = bbr->r_ctl.rc_lost; if (((bbr->r_ctl.rc_pkt_epoch - bbr->r_ctl.rc_bbr_last_startup_epoch) >= BBR_STARTUP_EPOCHS) || do_exit) { /* Return 1 to exit the startup state. */ return (1); } /* Stay in startup */ bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 8); return (0); } static void bbr_state_change(struct tcp_bbr *bbr, uint32_t cts, int32_t epoch, int32_t pkt_epoch, uint32_t losses) { /* * A tick occurred in the rtt epoch do we need to do anything? */ #ifdef BBR_INVARIANTS if ((bbr->rc_bbr_state != BBR_STATE_STARTUP) && (bbr->rc_bbr_state != BBR_STATE_DRAIN) && (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) && (bbr->rc_bbr_state != BBR_STATE_IDLE_EXIT) && (bbr->rc_bbr_state != BBR_STATE_PROBE_BW)) { /* Debug code? */ panic("Unknown BBR state %d?\n", bbr->rc_bbr_state); } #endif if (bbr->rc_bbr_state == BBR_STATE_STARTUP) { /* Do we exit the startup state? */ if (bbr_state_startup(bbr, cts, epoch, pkt_epoch)) { uint32_t time_in; bbr_log_startup_event(bbr, cts, bbr->r_ctl.rc_bbr_last_startup_epoch, bbr->r_ctl.rc_lost_at_startup, bbr_start_exit, 6); bbr->rc_filled_pipe = 1; bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { time_in = cts - bbr->r_ctl.rc_bbr_state_time; counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); } else time_in = 0; if (bbr->rc_no_pacing) bbr->rc_no_pacing = 0; bbr->r_ctl.rc_bbr_state_time = cts; bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.rc_drain_pg; bbr->rc_bbr_state = BBR_STATE_DRAIN; bbr_set_state_target(bbr, __LINE__); if ((bbr->rc_use_google == 0) && bbr_slam_cwnd_in_main_drain) { /* Here we don't have to worry about probe-rtt */ bbr->r_ctl.rc_saved_cwnd = bbr->rc_tp->snd_cwnd; bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); } bbr->r_ctl.rc_bbr_cwnd_gain = bbr_high_gain; bbr_log_type_statechange(bbr, cts, __LINE__); if (ctf_flight_size(bbr->rc_tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) <= bbr->r_ctl.rc_target_at_state) { /* * Switch to probe_bw if we are already * there */ bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts); bbr_substate_change(bbr, cts, __LINE__, 0); bbr->rc_bbr_state = BBR_STATE_PROBE_BW; bbr_log_type_statechange(bbr, cts, __LINE__); } } } else if (bbr->rc_bbr_state == BBR_STATE_IDLE_EXIT) { uint32_t inflight; struct tcpcb *tp; tp = bbr->rc_tp; inflight = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); if (inflight >= bbr->r_ctl.rc_target_at_state) { /* We have reached a flight of the cwnd target */ bbr->rc_bbr_state = BBR_STATE_PROBE_BW; bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; bbr->r_ctl.rc_bbr_cwnd_gain = BBR_UNIT; bbr_set_state_target(bbr, __LINE__); /* * Rig it so we don't do anything crazy and * start fresh with a new randomization. */ bbr->r_ctl.bbr_smallest_srtt_this_state = 0xffffffff; bbr->rc_bbr_substate = BBR_SUB_LEVEL6; bbr_substate_change(bbr, cts, __LINE__, 1); } } else if (bbr->rc_bbr_state == BBR_STATE_DRAIN) { /* Has in-flight reached the bdp (or less)? */ uint32_t inflight; struct tcpcb *tp; tp = bbr->rc_tp; inflight = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); if ((bbr->rc_use_google == 0) && bbr_slam_cwnd_in_main_drain && (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) { /* * Here we don't have to worry about probe-rtt * re-slam it, but keep it slammed down. */ bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); } if (inflight <= bbr->r_ctl.rc_target_at_state) { /* We have drained */ bbr->rc_bbr_state = BBR_STATE_PROBE_BW; bbr->r_ctl.bbr_lost_at_state = bbr->r_ctl.rc_lost; if (SEQ_GT(cts, bbr->r_ctl.rc_bbr_state_time)) { uint32_t time_in; time_in = cts - bbr->r_ctl.rc_bbr_state_time; counter_u64_add(bbr_state_time[bbr->rc_bbr_state], time_in); } if ((bbr->rc_use_google == 0) && bbr_slam_cwnd_in_main_drain && (tp->snd_cwnd < bbr->r_ctl.rc_saved_cwnd)) { /* Restore the cwnd */ tp->snd_cwnd = bbr->r_ctl.rc_saved_cwnd; bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); } /* Setup probe-rtt has being done now RRS-HERE */ bbr->r_ctl.rc_rtt_shrinks = cts; bbr->r_ctl.last_in_probertt = cts; bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_LEAVE_DRAIN, 0); /* Randomly pick a sub-state */ bbr->rc_bbr_substate = bbr_pick_probebw_substate(bbr, cts); bbr_substate_change(bbr, cts, __LINE__, 0); bbr_log_type_statechange(bbr, cts, __LINE__); } } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_RTT) { uint32_t flight; flight = ctf_flight_size(bbr->rc_tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); bbr->r_ctl.r_app_limited_until = (flight + bbr->r_ctl.rc_delivered); if (((bbr->r_ctl.bbr_rttprobe_gain_val == 0) || bbr->rc_use_google) && (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) { /* * We must keep cwnd at the desired MSS. */ bbr->rc_tp->snd_cwnd = bbr_rtt_probe_cwndtarg * (bbr->rc_tp->t_maxseg - bbr->rc_last_options); bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); } else if ((bbr_prtt_slam_cwnd) && (bbr->rc_tp->snd_cwnd > bbr->r_ctl.rc_target_at_state)) { /* Re-slam it */ bbr->rc_tp->snd_cwnd = bbr->r_ctl.rc_target_at_state; bbr_log_type_cwndupd(bbr, 0, 0, 0, 12, 0, 0, __LINE__); } if (bbr->r_ctl.rc_bbr_enters_probertt == 0) { /* Has outstanding reached our target? */ if (flight <= bbr->r_ctl.rc_target_at_state) { bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_REACHTAR, 0); bbr->r_ctl.rc_bbr_enters_probertt = cts; /* If time is exactly 0, be 1usec off */ if (bbr->r_ctl.rc_bbr_enters_probertt == 0) bbr->r_ctl.rc_bbr_enters_probertt = 1; if (bbr->rc_use_google == 0) { /* * Restore any lowering that as occurred to * reach here */ if (bbr->r_ctl.bbr_rttprobe_gain_val) bbr->r_ctl.rc_bbr_hptsi_gain = bbr->r_ctl.bbr_rttprobe_gain_val; else bbr->r_ctl.rc_bbr_hptsi_gain = BBR_UNIT; } } if ((bbr->r_ctl.rc_bbr_enters_probertt == 0) && (bbr->rc_use_google == 0) && bbr->r_ctl.bbr_rttprobe_gain_val && (((cts - bbr->r_ctl.rc_probertt_srttchktim) > bbr_get_rtt(bbr, bbr_drain_rtt)) || (flight >= bbr->r_ctl.flightsize_at_drain))) { /* * We have doddled with our current hptsi * gain an srtt and have still not made it * to target, or we have increased our flight. * Lets reduce the gain by xx% * flooring the reduce at DRAIN (based on * mul/div) */ int red; bbr->r_ctl.flightsize_at_drain = flight; bbr->r_ctl.rc_probertt_srttchktim = cts; red = max((bbr->r_ctl.bbr_rttprobe_gain_val / 10), 1); if ((bbr->r_ctl.rc_bbr_hptsi_gain - red) > max(bbr_drain_floor, 1)) { /* Reduce our gain again */ bbr->r_ctl.rc_bbr_hptsi_gain -= red; bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_SHRINK_PG, 0); } else if (bbr->r_ctl.rc_bbr_hptsi_gain > max(bbr_drain_floor, 1)) { /* one more chance before we give up */ bbr->r_ctl.rc_bbr_hptsi_gain = max(bbr_drain_floor, 1); bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_SHRINK_PG_FINAL, 0); } else { /* At the very bottom */ bbr->r_ctl.rc_bbr_hptsi_gain = max((bbr_drain_floor-1), 1); } } } if (bbr->r_ctl.rc_bbr_enters_probertt && (TSTMP_GT(cts, bbr->r_ctl.rc_bbr_enters_probertt)) && ((cts - bbr->r_ctl.rc_bbr_enters_probertt) >= bbr_rtt_probe_time)) { /* Time to exit probe RTT normally */ bbr_exit_probe_rtt(bbr->rc_tp, bbr, cts); } } else if (bbr->rc_bbr_state == BBR_STATE_PROBE_BW) { if ((bbr->rc_tp->snd_una == bbr->rc_tp->snd_max) && (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) { /* * This qualifies as a RTT_PROBE session since we * drop the data outstanding to nothing and waited * more than bbr_rtt_probe_time. */ bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0); bbr_set_reduced_rtt(bbr, cts, __LINE__); } if (bbr_should_enter_probe_rtt(bbr, cts)) { bbr_enter_probe_rtt(bbr, cts, __LINE__); } else { bbr_set_probebw_gains(bbr, cts, losses); } } } static void bbr_check_bbr_for_state(struct tcp_bbr *bbr, uint32_t cts, int32_t line, uint32_t losses) { int32_t epoch = 0; if ((cts - bbr->r_ctl.rc_rcv_epoch_start) >= bbr_get_rtt(bbr, BBR_RTT_PROP)) { bbr_set_epoch(bbr, cts, line); /* At each epoch doe lt bw sampling */ epoch = 1; } bbr_state_change(bbr, cts, epoch, bbr->rc_is_pkt_epoch_now, losses); } static int bbr_do_segment_nounlock(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos, int32_t nxt_pkt, struct timeval *tv) { int32_t thflags, retval; uint32_t cts, lcts; uint32_t tiwin; struct tcpopt to; struct tcp_bbr *bbr; struct bbr_sendmap *rsm; struct timeval ltv; int32_t did_out = 0; uint16_t nsegs; int32_t prev_state; uint32_t lost; nsegs = max(1, m->m_pkthdr.lro_nsegs); bbr = (struct tcp_bbr *)tp->t_fb_ptr; /* add in our stats */ kern_prefetch(bbr, &prev_state); prev_state = 0; thflags = tcp_get_flags(th); /* * If this is either a state-changing packet or current state isn't * established, we require a write lock on tcbinfo. Otherwise, we * allow the tcbinfo to be in either alocked or unlocked, as the * caller may have unnecessarily acquired a write lock due to a * race. */ INP_WLOCK_ASSERT(tp->t_inpcb); KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN", __func__)); tp->t_rcvtime = ticks; /* * Unscale the window into a 32-bit value. For the SYN_SENT state * the scale is zero. */ tiwin = th->th_win << tp->snd_scale; #ifdef STATS stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin); #endif if (m->m_flags & M_TSTMP) { /* Prefer the hardware timestamp if present */ struct timespec ts; mbuf_tstmp2timespec(m, &ts); bbr->rc_tv.tv_sec = ts.tv_sec; bbr->rc_tv.tv_usec = ts.tv_nsec / 1000; bbr->r_ctl.rc_rcvtime = cts = tcp_tv_to_usectick(&bbr->rc_tv); } else if (m->m_flags & M_TSTMP_LRO) { /* Next the arrival timestamp */ struct timespec ts; mbuf_tstmp2timespec(m, &ts); bbr->rc_tv.tv_sec = ts.tv_sec; bbr->rc_tv.tv_usec = ts.tv_nsec / 1000; bbr->r_ctl.rc_rcvtime = cts = tcp_tv_to_usectick(&bbr->rc_tv); } else { /* * Ok just get the current time. */ bbr->r_ctl.rc_rcvtime = lcts = cts = tcp_get_usecs(&bbr->rc_tv); } /* * Parse options on any incoming segment. */ tcp_dooptions(&to, (u_char *)(th + 1), (th->th_off << 2) - sizeof(struct tcphdr), (thflags & TH_SYN) ? TO_SYN : 0); /* * If timestamps were negotiated during SYN/ACK and a * segment without a timestamp is received, silently drop * the segment, unless it is a RST segment or missing timestamps are * tolerated. * See section 3.2 of RFC 7323. */ if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) && ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) { retval = 0; m_freem(m); goto done_with_input; } /* * If echoed timestamp is later than the current time, fall back to * non RFC1323 RTT calculation. Normalize timestamp if syncookies * were used when this connection was established. */ if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) { to.to_tsecr -= tp->ts_offset; if (TSTMP_GT(to.to_tsecr, tcp_tv_to_mssectick(&bbr->rc_tv))) to.to_tsecr = 0; } /* * If its the first time in we need to take care of options and * verify we can do SACK for rack! */ if (bbr->r_state == 0) { /* * Process options only when we get SYN/ACK back. The SYN * case for incoming connections is handled in tcp_syncache. * According to RFC1323 the window field in a SYN (i.e., a * or ) segment itself is never scaled. XXX * this is traditional behavior, may need to be cleaned up. */ if (bbr->rc_inp == NULL) { bbr->rc_inp = tp->t_inpcb; } /* * We need to init rc_inp here since its not init'd when * bbr_init is called */ if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) { if ((to.to_flags & TOF_SCALE) && (tp->t_flags & TF_REQ_SCALE)) { tp->t_flags |= TF_RCVD_SCALE; tp->snd_scale = to.to_wscale; } else tp->t_flags &= ~TF_REQ_SCALE; /* * Initial send window. It will be updated with the * next incoming segment to the scaled value. */ tp->snd_wnd = th->th_win; if ((to.to_flags & TOF_TS) && (tp->t_flags & TF_REQ_TSTMP)) { tp->t_flags |= TF_RCVD_TSTMP; tp->ts_recent = to.to_tsval; tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); } else tp->t_flags &= ~TF_REQ_TSTMP; if (to.to_flags & TOF_MSS) tcp_mss(tp, to.to_mss); if ((tp->t_flags & TF_SACK_PERMIT) && (to.to_flags & TOF_SACKPERM) == 0) tp->t_flags &= ~TF_SACK_PERMIT; if (IS_FASTOPEN(tp->t_flags)) { if (to.to_flags & TOF_FASTOPEN) { uint16_t mss; if (to.to_flags & TOF_MSS) mss = to.to_mss; else if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0) mss = TCP6_MSS; else mss = TCP_MSS; tcp_fastopen_update_cache(tp, mss, to.to_tfo_len, to.to_tfo_cookie); } else tcp_fastopen_disable_path(tp); } } /* * At this point we are at the initial call. Here we decide * if we are doing RACK or not. We do this by seeing if * TF_SACK_PERMIT is set, if not rack is *not* possible and * we switch to the default code. */ if ((tp->t_flags & TF_SACK_PERMIT) == 0) { /* Bail */ tcp_switch_back_to_default(tp); (*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen, tlen, iptos); return (1); } /* Set the flag */ bbr->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0; tcp_set_hpts(tp->t_inpcb); sack_filter_clear(&bbr->r_ctl.bbr_sf, th->th_ack); } if (thflags & TH_ACK) { /* Track ack types */ if (to.to_flags & TOF_SACK) BBR_STAT_INC(bbr_acks_with_sacks); else BBR_STAT_INC(bbr_plain_acks); } /* * This is the one exception case where we set the rack state * always. All other times (timers etc) we must have a rack-state * set (so we assure we have done the checks above for SACK). */ if (thflags & TH_FIN) tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN); if (bbr->r_state != tp->t_state) bbr_set_state(tp, bbr, tiwin); if (SEQ_GT(th->th_ack, tp->snd_una) && (rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map)) != NULL) kern_prefetch(rsm, &prev_state); prev_state = bbr->r_state; bbr->rc_ack_was_delayed = 0; lost = bbr->r_ctl.rc_lost; bbr->rc_is_pkt_epoch_now = 0; if (m->m_flags & (M_TSTMP|M_TSTMP_LRO)) { /* Get the real time into lcts and figure the real delay */ lcts = tcp_get_usecs(<v); if (TSTMP_GT(lcts, cts)) { bbr->r_ctl.rc_ack_hdwr_delay = lcts - cts; bbr->rc_ack_was_delayed = 1; if (TSTMP_GT(bbr->r_ctl.rc_ack_hdwr_delay, bbr->r_ctl.highest_hdwr_delay)) bbr->r_ctl.highest_hdwr_delay = bbr->r_ctl.rc_ack_hdwr_delay; } else { bbr->r_ctl.rc_ack_hdwr_delay = 0; bbr->rc_ack_was_delayed = 0; } } else { bbr->r_ctl.rc_ack_hdwr_delay = 0; bbr->rc_ack_was_delayed = 0; } bbr_log_ack_event(bbr, th, &to, tlen, nsegs, cts, nxt_pkt, m); if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) { retval = 0; m_freem(m); goto done_with_input; } /* * If a segment with the ACK-bit set arrives in the SYN-SENT state * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9. */ if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) && (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) { tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } if (tiwin > bbr->r_ctl.rc_high_rwnd) bbr->r_ctl.rc_high_rwnd = tiwin; #ifdef BBR_INVARIANTS if ((tp->t_inpcb->inp_flags & INP_DROPPED) || (tp->t_inpcb->inp_flags2 & INP_FREED)) { panic("tp:%p bbr:%p given a dropped inp:%p", tp, bbr, tp->t_inpcb); } #endif bbr->r_ctl.rc_flight_at_input = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); bbr->rtt_valid = 0; if (to.to_flags & TOF_TS) { bbr->rc_ts_valid = 1; bbr->r_ctl.last_inbound_ts = to.to_tsval; } else { bbr->rc_ts_valid = 0; bbr->r_ctl.last_inbound_ts = 0; } retval = (*bbr->r_substate) (m, th, so, tp, &to, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt, iptos); #ifdef BBR_INVARIANTS if ((retval == 0) && (tp->t_inpcb == NULL)) { panic("retval:%d tp:%p t_inpcb:NULL state:%d", retval, tp, prev_state); } #endif if (nxt_pkt == 0) BBR_STAT_INC(bbr_rlock_left_ret0); else BBR_STAT_INC(bbr_rlock_left_ret1); if (retval == 0) { /* * If retval is 1 the tcb is unlocked and most likely the tp * is gone. */ INP_WLOCK_ASSERT(tp->t_inpcb); tcp_bbr_xmit_timer_commit(bbr, tp, cts); if (bbr->rc_is_pkt_epoch_now) bbr_set_pktepoch(bbr, cts, __LINE__); bbr_check_bbr_for_state(bbr, cts, __LINE__, (bbr->r_ctl.rc_lost - lost)); if (nxt_pkt == 0) { if (bbr->r_wanted_output != 0) { bbr->rc_output_starts_timer = 0; did_out = 1; if (tcp_output(tp) < 0) return (1); } else bbr_start_hpts_timer(bbr, tp, cts, 6, 0, 0); } if ((nxt_pkt == 0) && ((bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) && (SEQ_GT(tp->snd_max, tp->snd_una) || (tp->t_flags & TF_DELACK) || ((V_tcp_always_keepalive || bbr->rc_inp->inp_socket->so_options & SO_KEEPALIVE) && (tp->t_state <= TCPS_CLOSING)))) { /* * We could not send (probably in the hpts but * stopped the timer)? */ if ((tp->snd_max == tp->snd_una) && ((tp->t_flags & TF_DELACK) == 0) && (tcp_in_hpts(bbr->rc_inp)) && (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { /* * keep alive not needed if we are hptsi * output yet */ ; } else { if (tcp_in_hpts(bbr->rc_inp)) { tcp_hpts_remove(bbr->rc_inp); if ((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) && (TSTMP_GT(lcts, bbr->rc_pacer_started))) { uint32_t del; del = lcts - bbr->rc_pacer_started; if (bbr->r_ctl.rc_last_delay_val > del) { BBR_STAT_INC(bbr_force_timer_start); bbr->r_ctl.rc_last_delay_val -= del; bbr->rc_pacer_started = lcts; } else { /* We are late */ bbr->r_ctl.rc_last_delay_val = 0; BBR_STAT_INC(bbr_force_output); if (tcp_output(tp) < 0) return (1); } } } bbr_start_hpts_timer(bbr, tp, cts, 8, bbr->r_ctl.rc_last_delay_val, 0); } } else if ((bbr->rc_output_starts_timer == 0) && (nxt_pkt == 0)) { /* Do we have the correct timer running? */ bbr_timer_audit(tp, bbr, lcts, &so->so_snd); } /* Do we have a new state */ if (bbr->r_state != tp->t_state) bbr_set_state(tp, bbr, tiwin); done_with_input: bbr_log_doseg_done(bbr, cts, nxt_pkt, did_out); if (did_out) bbr->r_wanted_output = 0; #ifdef BBR_INVARIANTS if (tp->t_inpcb == NULL) { panic("OP:%d retval:%d tp:%p t_inpcb:NULL state:%d", did_out, retval, tp, prev_state); } #endif } return (retval); } static void bbr_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos) { struct timeval tv; int retval; /* First lets see if we have old packets */ if (tp->t_in_pkt) { if (ctf_do_queued_segments(so, tp, 1)) { m_freem(m); return; } } if (m->m_flags & M_TSTMP_LRO) { mbuf_tstmp2timeval(m, &tv); } else { /* Should not be should we kassert instead? */ tcp_get_usecs(&tv); } retval = bbr_do_segment_nounlock(m, th, so, tp, drop_hdrlen, tlen, iptos, 0, &tv); if (retval == 0) { INP_WUNLOCK(tp->t_inpcb); } } /* * Return how much data can be sent without violating the * cwnd or rwnd. */ static inline uint32_t bbr_what_can_we_send(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t sendwin, uint32_t avail, int32_t sb_offset, uint32_t cts) { uint32_t len; if (ctf_outstanding(tp) >= tp->snd_wnd) { /* We never want to go over our peers rcv-window */ len = 0; } else { uint32_t flight; flight = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)); if (flight >= sendwin) { /* * We have in flight what we are allowed by cwnd (if * it was rwnd blocking it would have hit above out * >= tp->snd_wnd). */ return (0); } len = sendwin - flight; if ((len + ctf_outstanding(tp)) > tp->snd_wnd) { /* We would send too much (beyond the rwnd) */ len = tp->snd_wnd - ctf_outstanding(tp); } if ((len + sb_offset) > avail) { /* * We don't have that much in the SB, how much is * there? */ len = avail - sb_offset; } } return (len); } static inline void bbr_do_error_accounting(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, int32_t len, int32_t error) { #ifdef NETFLIX_STATS KMOD_TCPSTAT_INC(tcps_sndpack_error); KMOD_TCPSTAT_ADD(tcps_sndbyte_error, len); #endif } static inline void bbr_do_send_accounting(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, int32_t len, int32_t error) { if (error) { bbr_do_error_accounting(tp, bbr, rsm, len, error); return; } if (rsm) { if (rsm->r_flags & BBR_TLP) { /* * TLP should not count in retran count, but in its * own bin */ #ifdef NETFLIX_STATS KMOD_TCPSTAT_INC(tcps_tlpresends); KMOD_TCPSTAT_ADD(tcps_tlpresend_bytes, len); #endif } else { /* Retransmit */ tp->t_sndrexmitpack++; KMOD_TCPSTAT_INC(tcps_sndrexmitpack); KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len); #ifdef STATS stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB, len); #endif } /* * Logs in 0 - 8, 8 is all non probe_bw states 0-7 is * sub-state */ counter_u64_add(bbr_state_lost[rsm->r_bbr_state], len); if (bbr->rc_bbr_state != BBR_STATE_PROBE_BW) { /* Non probe_bw log in 1, 2, or 4. */ counter_u64_add(bbr_state_resend[bbr->rc_bbr_state], len); } else { /* * Log our probe state 3, and log also 5-13 to show * us the recovery sub-state for the send. This * means that 3 == (5+6+7+8+9+10+11+12+13) */ counter_u64_add(bbr_state_resend[BBR_STATE_PROBE_BW], len); counter_u64_add(bbr_state_resend[(bbr_state_val(bbr) + 5)], len); } /* Place in both 16's the totals of retransmitted */ counter_u64_add(bbr_state_lost[16], len); counter_u64_add(bbr_state_resend[16], len); /* Place in 17's the total sent */ counter_u64_add(bbr_state_resend[17], len); counter_u64_add(bbr_state_lost[17], len); } else { /* New sends */ KMOD_TCPSTAT_INC(tcps_sndpack); KMOD_TCPSTAT_ADD(tcps_sndbyte, len); /* Place in 17's the total sent */ counter_u64_add(bbr_state_resend[17], len); counter_u64_add(bbr_state_lost[17], len); #ifdef STATS stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB, len); #endif } } static void bbr_cwnd_limiting(struct tcpcb *tp, struct tcp_bbr *bbr, uint32_t in_level) { if (bbr->rc_filled_pipe && bbr_target_cwnd_mult_limit && (bbr->rc_use_google == 0)) { /* * Limit the cwnd to not be above N x the target plus whats * is outstanding. The target is based on the current b/w * estimate. */ uint32_t target; target = bbr_get_target_cwnd(bbr, bbr_get_bw(bbr), BBR_UNIT); target += ctf_outstanding(tp); target *= bbr_target_cwnd_mult_limit; if (tp->snd_cwnd > target) tp->snd_cwnd = target; bbr_log_type_cwndupd(bbr, 0, 0, 0, 10, 0, 0, __LINE__); } } static int bbr_window_update_needed(struct tcpcb *tp, struct socket *so, uint32_t recwin, int32_t maxseg) { /* * "adv" is the amount we could increase the window, taking into * account that we are limited by TCP_MAXWIN << tp->rcv_scale. */ int32_t adv; int32_t oldwin; adv = recwin; if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { oldwin = (tp->rcv_adv - tp->rcv_nxt); if (adv > oldwin) adv -= oldwin; else { /* We can't increase the window */ adv = 0; } } else oldwin = 0; /* * If the new window size ends up being the same as or less * than the old size when it is scaled, then don't force * a window update. */ if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale) return (0); if (adv >= (2 * maxseg) && (adv >= (so->so_rcv.sb_hiwat / 4) || recwin <= (so->so_rcv.sb_hiwat / 8) || so->so_rcv.sb_hiwat <= 8 * maxseg)) { return (1); } if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) return (1); return (0); } /* * Return 0 on success and a errno on failure to send. * Note that a 0 return may not mean we sent anything * if the TCB was on the hpts. A non-zero return * does indicate the error we got from ip[6]_output. */ static int bbr_output_wtime(struct tcpcb *tp, const struct timeval *tv) { struct socket *so; int32_t len; uint32_t cts; uint32_t recwin, sendwin; int32_t sb_offset; int32_t flags, abandon, error = 0; struct tcp_log_buffer *lgb = NULL; struct mbuf *m; struct mbuf *mb; uint32_t if_hw_tsomaxsegcount = 0; uint32_t if_hw_tsomaxsegsize = 0; uint32_t if_hw_tsomax = 0; struct ip *ip = NULL; #ifdef TCPDEBUG struct ipovly *ipov = NULL; #endif struct tcp_bbr *bbr; struct tcphdr *th; struct udphdr *udp = NULL; u_char opt[TCP_MAXOLEN]; unsigned ipoptlen, optlen, hdrlen; unsigned ulen; uint32_t bbr_seq; uint32_t delay_calc=0; uint8_t doing_tlp = 0; uint8_t local_options; #ifdef BBR_INVARIANTS uint8_t doing_retran_from = 0; uint8_t picked_up_retran = 0; #endif uint8_t wanted_cookie = 0; uint8_t more_to_rxt=0; int32_t prefetch_so_done = 0; int32_t prefetch_rsm = 0; uint32_t tot_len = 0; uint32_t maxseg, pace_max_segs, p_maxseg; int32_t csum_flags = 0; int32_t hw_tls; #if defined(IPSEC) || defined(IPSEC_SUPPORT) unsigned ipsec_optlen = 0; #endif volatile int32_t sack_rxmit; struct bbr_sendmap *rsm = NULL; int32_t tso, mtu; struct tcpopt to; int32_t slot = 0; struct inpcb *inp; struct sockbuf *sb; uint32_t hpts_calling; #ifdef INET6 struct ip6_hdr *ip6 = NULL; int32_t isipv6; #endif uint8_t app_limited = BBR_JR_SENT_DATA; uint8_t filled_all = 0; bbr = (struct tcp_bbr *)tp->t_fb_ptr; /* We take a cache hit here */ memcpy(&bbr->rc_tv, tv, sizeof(struct timeval)); cts = tcp_tv_to_usectick(&bbr->rc_tv); inp = bbr->rc_inp; so = inp->inp_socket; sb = &so->so_snd; if (sb->sb_flags & SB_TLS_IFNET) hw_tls = 1; else hw_tls = 0; kern_prefetch(sb, &maxseg); maxseg = tp->t_maxseg - bbr->rc_last_options; if (bbr_minseg(bbr) < maxseg) { tcp_bbr_tso_size_check(bbr, cts); } /* Remove any flags that indicate we are pacing on the inp */ pace_max_segs = bbr->r_ctl.rc_pace_max_segs; p_maxseg = min(maxseg, pace_max_segs); INP_WLOCK_ASSERT(inp); #ifdef TCP_OFFLOAD if (tp->t_flags & TF_TOE) return (tcp_offload_output(tp)); #endif #ifdef INET6 if (bbr->r_state) { /* Use the cache line loaded if possible */ isipv6 = bbr->r_is_v6; } else { isipv6 = (inp->inp_vflag & INP_IPV6) != 0; } #endif if (((bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) && tcp_in_hpts(inp)) { /* * We are on the hpts for some timer but not hptsi output. * Possibly remove from the hpts so we can send/recv etc. */ if ((tp->t_flags & TF_ACKNOW) == 0) { /* * No immediate demand right now to send an ack, but * the user may have read, making room for new data * (a window update). If so we may want to cancel * whatever timer is running (KEEP/DEL-ACK?) and * continue to send out a window update. Or we may * have gotten more data into the socket buffer to * send. */ recwin = lmin(lmax(sbspace(&so->so_rcv), 0), (long)TCP_MAXWIN << tp->rcv_scale); if ((bbr_window_update_needed(tp, so, recwin, maxseg) == 0) && ((tcp_outflags[tp->t_state] & TH_RST) == 0) && ((sbavail(sb) + ((tcp_outflags[tp->t_state] & TH_FIN) ? 1 : 0)) <= (tp->snd_max - tp->snd_una))) { /* * Nothing new to send and no window update * is needed to send. Lets just return and * let the timer-run off. */ return (0); } } tcp_hpts_remove(inp); bbr_timer_cancel(bbr, __LINE__, cts); } if (bbr->r_ctl.rc_last_delay_val) { /* Calculate a rough delay for early escape to sending */ if (SEQ_GT(cts, bbr->rc_pacer_started)) delay_calc = cts - bbr->rc_pacer_started; if (delay_calc >= bbr->r_ctl.rc_last_delay_val) delay_calc -= bbr->r_ctl.rc_last_delay_val; else delay_calc = 0; } /* Mark that we have called bbr_output(). */ if ((bbr->r_timer_override) || (tp->t_state < TCPS_ESTABLISHED)) { /* Timeouts or early states are exempt */ if (tcp_in_hpts(inp)) tcp_hpts_remove(inp); } else if (tcp_in_hpts(inp)) { if ((bbr->r_ctl.rc_last_delay_val) && (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) && delay_calc) { /* * We were being paced for output and the delay has * already exceeded when we were supposed to be * called, lets go ahead and pull out of the hpts * and call output. */ counter_u64_add(bbr_out_size[TCP_MSS_ACCT_LATE], 1); bbr->r_ctl.rc_last_delay_val = 0; tcp_hpts_remove(inp); } else if (tp->t_state == TCPS_CLOSED) { bbr->r_ctl.rc_last_delay_val = 0; tcp_hpts_remove(inp); } else { /* * On the hpts, you shall not pass! even if ACKNOW * is on, we will when the hpts fires, unless of * course we are overdue. */ counter_u64_add(bbr_out_size[TCP_MSS_ACCT_INPACE], 1); return (0); } } bbr->rc_cwnd_limited = 0; if (bbr->r_ctl.rc_last_delay_val) { /* recalculate the real delay and deal with over/under */ if (SEQ_GT(cts, bbr->rc_pacer_started)) delay_calc = cts - bbr->rc_pacer_started; else delay_calc = 0; if (delay_calc >= bbr->r_ctl.rc_last_delay_val) /* Setup the delay which will be added in */ delay_calc -= bbr->r_ctl.rc_last_delay_val; else { /* * We are early setup to adjust * our slot time. */ uint64_t merged_val; bbr->r_ctl.rc_agg_early += (bbr->r_ctl.rc_last_delay_val - delay_calc); bbr->r_agg_early_set = 1; if (bbr->r_ctl.rc_hptsi_agg_delay) { if (bbr->r_ctl.rc_hptsi_agg_delay >= bbr->r_ctl.rc_agg_early) { /* Nope our previous late cancels out the early */ bbr->r_ctl.rc_hptsi_agg_delay -= bbr->r_ctl.rc_agg_early; bbr->r_agg_early_set = 0; bbr->r_ctl.rc_agg_early = 0; } else { bbr->r_ctl.rc_agg_early -= bbr->r_ctl.rc_hptsi_agg_delay; bbr->r_ctl.rc_hptsi_agg_delay = 0; } } merged_val = bbr->rc_pacer_started; merged_val <<= 32; merged_val |= bbr->r_ctl.rc_last_delay_val; bbr_log_pacing_delay_calc(bbr, inp->inp_hpts_calls, bbr->r_ctl.rc_agg_early, cts, delay_calc, merged_val, bbr->r_agg_early_set, 3); bbr->r_ctl.rc_last_delay_val = 0; BBR_STAT_INC(bbr_early); delay_calc = 0; } } else { /* We were not delayed due to hptsi */ if (bbr->r_agg_early_set) bbr->r_ctl.rc_agg_early = 0; bbr->r_agg_early_set = 0; delay_calc = 0; } if (delay_calc) { /* * We had a hptsi delay which means we are falling behind on * sending at the expected rate. Calculate an extra amount * of data we can send, if any, to put us back on track. */ if ((bbr->r_ctl.rc_hptsi_agg_delay + delay_calc) < bbr->r_ctl.rc_hptsi_agg_delay) bbr->r_ctl.rc_hptsi_agg_delay = 0xffffffff; else bbr->r_ctl.rc_hptsi_agg_delay += delay_calc; } sendwin = min(tp->snd_wnd, tp->snd_cwnd); if ((tp->snd_una == tp->snd_max) && (bbr->rc_bbr_state != BBR_STATE_IDLE_EXIT) && (sbavail(sb))) { /* * Ok we have been idle with nothing outstanding * we possibly need to start fresh with either a new * suite of states or a fast-ramp up. */ bbr_restart_after_idle(bbr, cts, bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time)); } /* * Now was there a hptsi delay where we are behind? We only count * being behind if: a) We are not in recovery. b) There was a delay. * c) We had room to send something. * */ hpts_calling = inp->inp_hpts_calls; inp->inp_hpts_calls = 0; if (bbr->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { int retval; retval = bbr_process_timers(tp, bbr, cts, hpts_calling); if (retval != 0) { counter_u64_add(bbr_out_size[TCP_MSS_ACCT_ATIMER], 1); /* * If timers want tcp_drop(), then pass error out, * otherwise suppress it. */ return (retval < 0 ? retval : 0); } } bbr->rc_inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY; if (hpts_calling && (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { bbr->r_ctl.rc_last_delay_val = 0; } bbr->r_timer_override = 0; bbr->r_wanted_output = 0; /* * For TFO connections in SYN_RECEIVED, only allow the initial * SYN|ACK and those sent by the retransmit timer. */ if (IS_FASTOPEN(tp->t_flags) && ((tp->t_state == TCPS_SYN_RECEIVED) || (tp->t_state == TCPS_SYN_SENT)) && SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */ (tp->t_rxtshift == 0)) { /* not a retransmit */ len = 0; goto just_return_nolock; } /* * Before sending anything check for a state update. For hpts * calling without input this is important. If its input calling * then this was already done. */ if (bbr->rc_use_google == 0) bbr_check_bbr_for_state(bbr, cts, __LINE__, 0); again: /* * If we've recently taken a timeout, snd_max will be greater than * snd_max. BBR in general does not pay much attention to snd_nxt * for historic reasons the persist timer still uses it. This means * we have to look at it. All retransmissions that are not persits * use the rsm that needs to be sent so snd_nxt is ignored. At the * end of this routine we pull snd_nxt always up to snd_max. */ doing_tlp = 0; #ifdef BBR_INVARIANTS doing_retran_from = picked_up_retran = 0; #endif error = 0; tso = 0; slot = 0; mtu = 0; sendwin = min(tp->snd_wnd, tp->snd_cwnd); sb_offset = tp->snd_max - tp->snd_una; flags = tcp_outflags[tp->t_state]; sack_rxmit = 0; len = 0; rsm = NULL; if (flags & TH_RST) { SOCKBUF_LOCK(sb); goto send; } recheck_resend: while (bbr->r_ctl.rc_free_cnt < bbr_min_req_free) { /* We need to always have one in reserve */ rsm = bbr_alloc(bbr); if (rsm == NULL) { error = ENOMEM; /* Lie to get on the hpts */ tot_len = tp->t_maxseg; if (hpts_calling) /* Retry in a ms */ slot = 1001; goto just_return_nolock; } TAILQ_INSERT_TAIL(&bbr->r_ctl.rc_free, rsm, r_next); bbr->r_ctl.rc_free_cnt++; rsm = NULL; } /* What do we send, a resend? */ if (bbr->r_ctl.rc_resend == NULL) { /* Check for rack timeout */ bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts); if (bbr->r_ctl.rc_resend) { #ifdef BBR_INVARIANTS picked_up_retran = 1; #endif bbr_cong_signal(tp, NULL, CC_NDUPACK, bbr->r_ctl.rc_resend); } } if (bbr->r_ctl.rc_resend) { rsm = bbr->r_ctl.rc_resend; #ifdef BBR_INVARIANTS doing_retran_from = 1; #endif /* Remove any TLP flags its a RACK or T-O */ rsm->r_flags &= ~BBR_TLP; bbr->r_ctl.rc_resend = NULL; if (SEQ_LT(rsm->r_start, tp->snd_una)) { #ifdef BBR_INVARIANTS panic("Huh, tp:%p bbr:%p rsm:%p start:%u < snd_una:%u\n", tp, bbr, rsm, rsm->r_start, tp->snd_una); goto recheck_resend; #else /* TSNH */ rsm = NULL; goto recheck_resend; #endif } if (rsm->r_flags & BBR_HAS_SYN) { /* Only retransmit a SYN by itself */ len = 0; if ((flags & TH_SYN) == 0) { /* Huh something is wrong */ rsm->r_start++; if (rsm->r_start == rsm->r_end) { /* Clean it up, somehow we missed the ack? */ bbr_log_syn(tp, NULL); } else { /* TFO with data? */ rsm->r_flags &= ~BBR_HAS_SYN; len = rsm->r_end - rsm->r_start; } } else { /* Retransmitting SYN */ rsm = NULL; SOCKBUF_LOCK(sb); goto send; } } else len = rsm->r_end - rsm->r_start; if ((bbr->rc_resends_use_tso == 0) && (len > maxseg)) { len = maxseg; more_to_rxt = 1; } sb_offset = rsm->r_start - tp->snd_una; if (len > 0) { sack_rxmit = 1; KMOD_TCPSTAT_INC(tcps_sack_rexmits); KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes, min(len, maxseg)); } else { /* I dont think this can happen */ rsm = NULL; goto recheck_resend; } BBR_STAT_INC(bbr_resends_set); } else if (bbr->r_ctl.rc_tlp_send) { /* * Tail loss probe */ doing_tlp = 1; rsm = bbr->r_ctl.rc_tlp_send; bbr->r_ctl.rc_tlp_send = NULL; sack_rxmit = 1; len = rsm->r_end - rsm->r_start; if ((bbr->rc_resends_use_tso == 0) && (len > maxseg)) len = maxseg; if (SEQ_GT(tp->snd_una, rsm->r_start)) { #ifdef BBR_INVARIANTS panic("tp:%p bbc:%p snd_una:%u rsm:%p r_start:%u", tp, bbr, tp->snd_una, rsm, rsm->r_start); #else /* TSNH */ rsm = NULL; goto recheck_resend; #endif } sb_offset = rsm->r_start - tp->snd_una; BBR_STAT_INC(bbr_tlp_set); } /* * Enforce a connection sendmap count limit if set * as long as we are not retransmiting. */ if ((rsm == NULL) && (V_tcp_map_entries_limit > 0) && (bbr->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) { BBR_STAT_INC(bbr_alloc_limited); if (!bbr->alloc_limit_reported) { bbr->alloc_limit_reported = 1; BBR_STAT_INC(bbr_alloc_limited_conns); } goto just_return_nolock; } #ifdef BBR_INVARIANTS if (rsm && SEQ_LT(rsm->r_start, tp->snd_una)) { panic("tp:%p bbr:%p rsm:%p sb_offset:%u len:%u", tp, bbr, rsm, sb_offset, len); } #endif /* * Get standard flags, and add SYN or FIN if requested by 'hidden' * state flags. */ if (tp->t_flags & TF_NEEDFIN && (rsm == NULL)) flags |= TH_FIN; if (tp->t_flags & TF_NEEDSYN) flags |= TH_SYN; if (rsm && (rsm->r_flags & BBR_HAS_FIN)) { /* we are retransmitting the fin */ len--; if (len) { /* * When retransmitting data do *not* include the * FIN. This could happen from a TLP probe if we * allowed data with a FIN. */ flags &= ~TH_FIN; } } else if (rsm) { if (flags & TH_FIN) flags &= ~TH_FIN; } if ((sack_rxmit == 0) && (prefetch_rsm == 0)) { void *end_rsm; end_rsm = TAILQ_LAST_FAST(&bbr->r_ctl.rc_tmap, bbr_sendmap, r_tnext); if (end_rsm) kern_prefetch(end_rsm, &prefetch_rsm); prefetch_rsm = 1; } SOCKBUF_LOCK(sb); /* * If snd_nxt == snd_max and we have transmitted a FIN, the * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a * negative length. This can also occur when TCP opens up its * congestion window while receiving additional duplicate acks after * fast-retransmit because TCP will reset snd_nxt to snd_max after * the fast-retransmit. * * In the normal retransmit-FIN-only case, however, snd_nxt will be * set to snd_una, the sb_offset will be 0, and the length may wind * up 0. * * If sack_rxmit is true we are retransmitting from the scoreboard * in which case len is already set. */ if (sack_rxmit == 0) { uint32_t avail; avail = sbavail(sb); if (SEQ_GT(tp->snd_max, tp->snd_una)) sb_offset = tp->snd_max - tp->snd_una; else sb_offset = 0; if (bbr->rc_tlp_new_data) { /* TLP is forcing out new data */ uint32_t tlplen; doing_tlp = 1; tlplen = maxseg; if (tlplen > (uint32_t)(avail - sb_offset)) { tlplen = (uint32_t)(avail - sb_offset); } if (tlplen > tp->snd_wnd) { len = tp->snd_wnd; } else { len = tlplen; } bbr->rc_tlp_new_data = 0; } else { len = bbr_what_can_we_send(tp, bbr, sendwin, avail, sb_offset, cts); if ((len < p_maxseg) && (bbr->rc_in_persist == 0) && (ctf_outstanding(tp) >= (2 * p_maxseg)) && ((avail - sb_offset) >= p_maxseg)) { /* * We are not completing whats in the socket * buffer (i.e. there is at least a segment * waiting to send) and we have 2 or more * segments outstanding. There is no sense * of sending a little piece. Lets defer and * and wait until we can send a whole * segment. */ len = 0; } if (bbr->rc_in_persist) { /* * We are in persists, figure out if * a retransmit is available (maybe the previous * persists we sent) or if we have to send new * data. */ rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); if (rsm) { len = rsm->r_end - rsm->r_start; if (rsm->r_flags & BBR_HAS_FIN) len--; if ((bbr->rc_resends_use_tso == 0) && (len > maxseg)) len = maxseg; if (len > 1) BBR_STAT_INC(bbr_persist_reneg); /* * XXXrrs we could force the len to * 1 byte here to cause the chunk to * split apart.. but that would then * mean we always retransmit it as * one byte even after the window * opens. */ sack_rxmit = 1; sb_offset = rsm->r_start - tp->snd_una; } else { /* * First time through in persists or peer * acked our one byte. Though we do have * to have something in the sb. */ len = 1; sb_offset = 0; if (avail == 0) len = 0; } } } } if (prefetch_so_done == 0) { kern_prefetch(so, &prefetch_so_done); prefetch_so_done = 1; } /* * Lop off SYN bit if it has already been sent. However, if this is * SYN-SENT state and if segment contains data and if we don't know * that foreign host supports TAO, suppress sending segment. */ if ((flags & TH_SYN) && (rsm == NULL) && SEQ_GT(tp->snd_max, tp->snd_una)) { if (tp->t_state != TCPS_SYN_RECEIVED) flags &= ~TH_SYN; /* * When sending additional segments following a TFO SYN|ACK, * do not include the SYN bit. */ if (IS_FASTOPEN(tp->t_flags) && (tp->t_state == TCPS_SYN_RECEIVED)) flags &= ~TH_SYN; sb_offset--, len++; if (sbavail(sb) == 0) len = 0; } else if ((flags & TH_SYN) && rsm) { /* * Subtract one from the len for the SYN being * retransmitted. */ len--; } /* * Be careful not to send data and/or FIN on SYN segments. This * measure is needed to prevent interoperability problems with not * fully conformant TCP implementations. */ if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { len = 0; flags &= ~TH_FIN; } /* * On TFO sockets, ensure no data is sent in the following cases: * * - When retransmitting SYN|ACK on a passively-created socket * - When retransmitting SYN on an actively created socket * - When sending a zero-length cookie (cookie request) on an * actively created socket * - When the socket is in the CLOSED state (RST is being sent) */ if (IS_FASTOPEN(tp->t_flags) && (((flags & TH_SYN) && (tp->t_rxtshift > 0)) || ((tp->t_state == TCPS_SYN_SENT) && (tp->t_tfo_client_cookie_len == 0)) || (flags & TH_RST))) { len = 0; sack_rxmit = 0; rsm = NULL; } /* Without fast-open there should never be data sent on a SYN */ if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) len = 0; if (len <= 0) { /* * If FIN has been sent but not acked, but we haven't been * called to retransmit, len will be < 0. Otherwise, window * shrank after we sent into it. If window shrank to 0, * cancel pending retransmit, pull snd_nxt back to (closed) * window, and set the persist timer if it isn't already * going. If the window didn't close completely, just wait * for an ACK. * * We also do a general check here to ensure that we will * set the persist timer when we have data to send, but a * 0-byte window. This makes sure the persist timer is set * even if the packet hits one of the "goto send" lines * below. */ len = 0; if ((tp->snd_wnd == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) && (tp->snd_una == tp->snd_max) && (sb_offset < (int)sbavail(sb))) { /* * Not enough room in the rwnd to send * a paced segment out. */ bbr_enter_persist(tp, bbr, cts, __LINE__); } } else if ((rsm == NULL) && (doing_tlp == 0) && (len < bbr->r_ctl.rc_pace_max_segs)) { /* * We are not sending a full segment for * some reason. Should we not send anything (think * sws or persists)? */ if ((tp->snd_wnd < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) && (TCPS_HAVEESTABLISHED(tp->t_state)) && (len < (int)(sbavail(sb) - sb_offset))) { /* * Here the rwnd is less than * the pacing size, this is not a retransmit, * we are established and * the send is not the last in the socket buffer * lets not send, and possibly enter persists. */ len = 0; if (tp->snd_max == tp->snd_una) bbr_enter_persist(tp, bbr, cts, __LINE__); } else if ((tp->snd_cwnd >= bbr->r_ctl.rc_pace_max_segs) && (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) > (2 * maxseg)) && (len < (int)(sbavail(sb) - sb_offset)) && (len < bbr_minseg(bbr))) { /* * Here we are not retransmitting, and * the cwnd is not so small that we could * not send at least a min size (rxt timer * not having gone off), We have 2 segments or * more already in flight, its not the tail end * of the socket buffer and the cwnd is blocking * us from sending out minimum pacing segment size. * Lets not send anything. */ bbr->rc_cwnd_limited = 1; len = 0; } else if (((tp->snd_wnd - ctf_outstanding(tp)) < min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) && (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) > (2 * maxseg)) && (len < (int)(sbavail(sb) - sb_offset)) && (TCPS_HAVEESTABLISHED(tp->t_state))) { /* * Here we have a send window but we have * filled it up and we can't send another pacing segment. * We also have in flight more than 2 segments * and we are not completing the sb i.e. we allow * the last bytes of the sb to go out even if * its not a full pacing segment. */ len = 0; } } /* len will be >= 0 after this point. */ KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); tcp_sndbuf_autoscale(tp, so, sendwin); /* * */ if (bbr->rc_in_persist && len && (rsm == NULL) && (len < min((bbr->r_ctl.rc_high_rwnd/2), bbr->r_ctl.rc_pace_max_segs))) { /* * We are in persist, not doing a retransmit and don't have enough space * yet to send a full TSO. So is it at the end of the sb * if so we need to send else nuke to 0 and don't send. */ int sbleft; if (sbavail(sb) > sb_offset) sbleft = sbavail(sb) - sb_offset; else sbleft = 0; if (sbleft >= min((bbr->r_ctl.rc_high_rwnd/2), bbr->r_ctl.rc_pace_max_segs)) { /* not at end of sb lets not send */ len = 0; } } /* * Decide if we can use TCP Segmentation Offloading (if supported by * hardware). * * TSO may only be used if we are in a pure bulk sending state. The * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP * options prevent using TSO. With TSO the TCP header is the same * (except for the sequence number) for all generated packets. This * makes it impossible to transmit any options which vary per * generated segment or packet. * * IPv4 handling has a clear separation of ip options and ip header * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() * does the right thing below to provide length of just ip options * and thus checking for ipoptlen is enough to decide if ip options * are present. */ #ifdef INET6 if (isipv6) ipoptlen = ip6_optlen(inp); else #endif if (inp->inp_options) ipoptlen = inp->inp_options->m_len - offsetof(struct ipoption, ipopt_list); else ipoptlen = 0; #if defined(IPSEC) || defined(IPSEC_SUPPORT) /* * Pre-calculate here as we save another lookup into the darknesses * of IPsec that way and can actually decide if TSO is ok. */ #ifdef INET6 if (isipv6 && IPSEC_ENABLED(ipv6)) ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp); #ifdef INET else #endif #endif /* INET6 */ #ifdef INET if (IPSEC_ENABLED(ipv4)) ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp); #endif /* INET */ #endif /* IPSEC */ #if defined(IPSEC) || defined(IPSEC_SUPPORT) ipoptlen += ipsec_optlen; #endif if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && (len > maxseg) && (tp->t_port == 0) && ((tp->t_flags & TF_SIGNATURE) == 0) && tp->rcv_numsacks == 0 && ipoptlen == 0) tso = 1; recwin = lmin(lmax(sbspace(&so->so_rcv), 0), (long)TCP_MAXWIN << tp->rcv_scale); /* * Sender silly window avoidance. We transmit under the following * conditions when len is non-zero: * * - We have a full segment (or more with TSO) - This is the last * buffer in a write()/send() and we are either idle or running * NODELAY - we've timed out (e.g. persist timer) - we have more * then 1/2 the maximum send window's worth of data (receiver may be * limited the window size) - we need to retransmit */ if (rsm) goto send; if (len) { if (sack_rxmit) goto send; if (len >= p_maxseg) goto send; /* * NOTE! on localhost connections an 'ack' from the remote * end may occur synchronously with the output and cause us * to flush a buffer queued with moretocome. XXX * */ if (((tp->t_flags & TF_MORETOCOME) == 0) && /* normal case */ ((tp->t_flags & TF_NODELAY) || ((uint32_t)len + (uint32_t)sb_offset) >= sbavail(&so->so_snd)) && (tp->t_flags & TF_NOPUSH) == 0) { goto send; } if ((tp->snd_una == tp->snd_max) && len) { /* Nothing outstanding */ goto send; } if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) { goto send; } } /* * Sending of standalone window updates. * * Window updates are important when we close our window due to a * full socket buffer and are opening it again after the application * reads data from it. Once the window has opened again and the * remote end starts to send again the ACK clock takes over and * provides the most current window information. * * We must avoid the silly window syndrome whereas every read from * the receive buffer, no matter how small, causes a window update * to be sent. We also should avoid sending a flurry of window * updates when the socket buffer had queued a lot of data and the * application is doing small reads. * * Prevent a flurry of pointless window updates by only sending an * update when we can increase the advertized window by more than * 1/4th of the socket buffer capacity. When the buffer is getting * full or is very small be more aggressive and send an update * whenever we can increase by two mss sized segments. In all other * situations the ACK's to new incoming data will carry further * window increases. * * Don't send an independent window update if a delayed ACK is * pending (it will get piggy-backed on it) or the remote side * already has done a half-close and won't send more data. Skip * this if the connection is in T/TCP half-open state. */ if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && !(tp->t_flags & TF_DELACK) && !TCPS_HAVERCVDFIN(tp->t_state)) { /* Check to see if we should do a window update */ if (bbr_window_update_needed(tp, so, recwin, maxseg)) goto send; } /* * Send if we owe the peer an ACK, RST, SYN. ACKNOW * is also a catch-all for the retransmit timer timeout case. */ if (tp->t_flags & TF_ACKNOW) { goto send; } if (flags & TH_RST) { /* Always send a RST if one is due */ goto send; } if ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0) { goto send; } /* * If our state indicates that FIN should be sent and we have not * yet done so, then we need to send. */ if (flags & TH_FIN && ((tp->t_flags & TF_SENTFIN) == 0)) { goto send; } /* * No reason to send a segment, just return. */ just_return: SOCKBUF_UNLOCK(sb); just_return_nolock: if (tot_len) slot = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, tot_len, cts, 0); if (bbr->rc_no_pacing) slot = 0; if (tot_len == 0) { if ((ctf_outstanding(tp) + min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr))) >= tp->snd_wnd) { BBR_STAT_INC(bbr_rwnd_limited); app_limited = BBR_JR_RWND_LIMITED; bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp)); if ((bbr->rc_in_persist == 0) && TCPS_HAVEESTABLISHED(tp->t_state) && (tp->snd_max == tp->snd_una) && sbavail(&tp->t_inpcb->inp_socket->so_snd)) { /* No send window.. we must enter persist */ bbr_enter_persist(tp, bbr, bbr->r_ctl.rc_rcvtime, __LINE__); } } else if (ctf_outstanding(tp) >= sbavail(sb)) { BBR_STAT_INC(bbr_app_limited); app_limited = BBR_JR_APP_LIMITED; bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp)); } else if ((ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + p_maxseg) >= tp->snd_cwnd) { BBR_STAT_INC(bbr_cwnd_limited); app_limited = BBR_JR_CWND_LIMITED; bbr_cwnd_limiting(tp, bbr, ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes))); bbr->rc_cwnd_limited = 1; } else { BBR_STAT_INC(bbr_app_limited); app_limited = BBR_JR_APP_LIMITED; bbr_cwnd_limiting(tp, bbr, ctf_outstanding(tp)); } bbr->r_ctl.rc_hptsi_agg_delay = 0; bbr->r_agg_early_set = 0; bbr->r_ctl.rc_agg_early = 0; bbr->r_ctl.rc_last_delay_val = 0; } else if (bbr->rc_use_google == 0) bbr_check_bbr_for_state(bbr, cts, __LINE__, 0); /* Are we app limited? */ if ((app_limited == BBR_JR_APP_LIMITED) || (app_limited == BBR_JR_RWND_LIMITED)) { /** * We are application limited. */ bbr->r_ctl.r_app_limited_until = (ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + bbr->r_ctl.rc_delivered); } if (tot_len == 0) counter_u64_add(bbr_out_size[TCP_MSS_ACCT_JUSTRET], 1); /* Dont update the time if we did not send */ bbr->r_ctl.rc_last_delay_val = 0; bbr->rc_output_starts_timer = 1; bbr_start_hpts_timer(bbr, tp, cts, 9, slot, tot_len); bbr_log_type_just_return(bbr, cts, tot_len, hpts_calling, app_limited, p_maxseg, len); if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { /* Make sure snd_nxt is drug up */ tp->snd_nxt = tp->snd_max; } return (error); send: if (doing_tlp == 0) { /* * Data not a TLP, and its not the rxt firing. If it is the * rxt firing, we want to leave the tlp_in_progress flag on * so we don't send another TLP. It has to be a rack timer * or normal send (response to acked data) to clear the tlp * in progress flag. */ bbr->rc_tlp_in_progress = 0; bbr->rc_tlp_rtx_out = 0; } else { /* * Its a TLP. */ bbr->rc_tlp_in_progress = 1; } bbr_timer_cancel(bbr, __LINE__, cts); if (rsm == NULL) { if (sbused(sb) > 0) { /* * This is sub-optimal. We only send a stand alone * FIN on its own segment. */ if (flags & TH_FIN) { flags &= ~TH_FIN; if ((len == 0) && ((tp->t_flags & TF_ACKNOW) == 0)) { /* Lets not send this */ slot = 0; goto just_return; } } } } else { /* * We do *not* send a FIN on a retransmit if it has data. * The if clause here where len > 1 should never come true. */ if ((len > 0) && (((rsm->r_flags & BBR_HAS_FIN) == 0) && (flags & TH_FIN))) { flags &= ~TH_FIN; len--; } } SOCKBUF_LOCK_ASSERT(sb); if (len > 0) { if ((tp->snd_una == tp->snd_max) && (bbr_calc_time(cts, bbr->r_ctl.rc_went_idle_time) >= bbr_rtt_probe_time)) { /* * This qualifies as a RTT_PROBE session since we * drop the data outstanding to nothing and waited * more than bbr_rtt_probe_time. */ bbr_log_rtt_shrinks(bbr, cts, 0, 0, __LINE__, BBR_RTTS_WASIDLE, 0); bbr_set_reduced_rtt(bbr, cts, __LINE__); } if (len >= maxseg) tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT; else tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT; } /* * Before ESTABLISHED, force sending of initial options unless TCP * set not to do any options. NOTE: we assume that the IP/TCP header * plus TCP options always fit in a single mbuf, leaving room for a * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr) * + optlen <= MCLBYTES */ optlen = 0; #ifdef INET6 if (isipv6) hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); else #endif hdrlen = sizeof(struct tcpiphdr); /* * Compute options for segment. We only have to care about SYN and * established connection segments. Options for SYN-ACK segments * are handled in TCP syncache. */ to.to_flags = 0; local_options = 0; if ((tp->t_flags & TF_NOOPT) == 0) { /* Maximum segment size. */ if (flags & TH_SYN) { to.to_mss = tcp_mssopt(&inp->inp_inc); if (tp->t_port) to.to_mss -= V_tcp_udp_tunneling_overhead; to.to_flags |= TOF_MSS; /* * On SYN or SYN|ACK transmits on TFO connections, * only include the TFO option if it is not a * retransmit, as the presence of the TFO option may * have caused the original SYN or SYN|ACK to have * been dropped by a middlebox. */ if (IS_FASTOPEN(tp->t_flags) && (tp->t_rxtshift == 0)) { if (tp->t_state == TCPS_SYN_RECEIVED) { to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; to.to_tfo_cookie = (u_int8_t *)&tp->t_tfo_cookie.server; to.to_flags |= TOF_FASTOPEN; wanted_cookie = 1; } else if (tp->t_state == TCPS_SYN_SENT) { to.to_tfo_len = tp->t_tfo_client_cookie_len; to.to_tfo_cookie = tp->t_tfo_cookie.client; to.to_flags |= TOF_FASTOPEN; wanted_cookie = 1; } } } /* Window scaling. */ if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { to.to_wscale = tp->request_r_scale; to.to_flags |= TOF_SCALE; } /* Timestamps. */ if ((tp->t_flags & TF_RCVD_TSTMP) || ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { to.to_tsval = tcp_tv_to_mssectick(&bbr->rc_tv) + tp->ts_offset; to.to_tsecr = tp->ts_recent; to.to_flags |= TOF_TS; local_options += TCPOLEN_TIMESTAMP + 2; } /* Set receive buffer autosizing timestamp. */ if (tp->rfbuf_ts == 0 && (so->so_rcv.sb_flags & SB_AUTOSIZE)) tp->rfbuf_ts = tcp_tv_to_mssectick(&bbr->rc_tv); /* Selective ACK's. */ if (flags & TH_SYN) to.to_flags |= TOF_SACKPERM; else if (TCPS_HAVEESTABLISHED(tp->t_state) && tp->rcv_numsacks > 0) { to.to_flags |= TOF_SACK; to.to_nsacks = tp->rcv_numsacks; to.to_sacks = (u_char *)tp->sackblks; } #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) /* TCP-MD5 (RFC2385). */ if (tp->t_flags & TF_SIGNATURE) to.to_flags |= TOF_SIGNATURE; #endif /* TCP_SIGNATURE */ /* Processing the options. */ hdrlen += (optlen = tcp_addoptions(&to, opt)); /* * If we wanted a TFO option to be added, but it was unable * to fit, ensure no data is sent. */ if (IS_FASTOPEN(tp->t_flags) && wanted_cookie && !(to.to_flags & TOF_FASTOPEN)) len = 0; } if (tp->t_port) { if (V_tcp_udp_tunneling_port == 0) { /* The port was removed?? */ SOCKBUF_UNLOCK(&so->so_snd); return (EHOSTUNREACH); } hdrlen += sizeof(struct udphdr); } #ifdef INET6 if (isipv6) ipoptlen = ip6_optlen(tp->t_inpcb); else #endif if (tp->t_inpcb->inp_options) ipoptlen = tp->t_inpcb->inp_options->m_len - offsetof(struct ipoption, ipopt_list); else ipoptlen = 0; ipoptlen = 0; #if defined(IPSEC) || defined(IPSEC_SUPPORT) ipoptlen += ipsec_optlen; #endif if (bbr->rc_last_options != local_options) { /* * Cache the options length this generally does not change * on a connection. We use this to calculate TSO. */ bbr->rc_last_options = local_options; } maxseg = tp->t_maxseg - (ipoptlen + optlen); p_maxseg = min(maxseg, pace_max_segs); /* * Adjust data length if insertion of options will bump the packet * length beyond the t_maxseg length. Clear the FIN bit because we * cut off the tail of the segment. */ if (len > maxseg) { if (len != 0 && (flags & TH_FIN)) { flags &= ~TH_FIN; } if (tso) { uint32_t moff; int32_t max_len; /* extract TSO information */ if_hw_tsomax = tp->t_tsomax; if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; KASSERT(ipoptlen == 0, ("%s: TSO can't do IP options", __func__)); /* * Check if we should limit by maximum payload * length: */ if (if_hw_tsomax != 0) { /* compute maximum TSO length */ max_len = (if_hw_tsomax - hdrlen - max_linkhdr); if (max_len <= 0) { len = 0; } else if (len > max_len) { len = max_len; } } /* * Prevent the last segment from being fractional * unless the send sockbuf can be emptied: */ if ((sb_offset + len) < sbavail(sb)) { moff = len % (uint32_t)maxseg; if (moff != 0) { len -= moff; } } /* * In case there are too many small fragments don't * use TSO: */ if (len <= maxseg) { len = maxseg; tso = 0; } } else { /* Not doing TSO */ if (optlen + ipoptlen >= tp->t_maxseg) { /* * Since we don't have enough space to put * the IP header chain and the TCP header in * one packet as required by RFC 7112, don't * send it. Also ensure that at least one * byte of the payload can be put into the * TCP segment. */ SOCKBUF_UNLOCK(&so->so_snd); error = EMSGSIZE; sack_rxmit = 0; goto out; } len = maxseg; } } else { /* Not doing TSO */ if_hw_tsomaxsegcount = 0; tso = 0; } KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET, ("%s: len > IP_MAXPACKET", __func__)); #ifdef DIAGNOSTIC #ifdef INET6 if (max_linkhdr + hdrlen > MCLBYTES) #else if (max_linkhdr + hdrlen > MHLEN) #endif panic("tcphdr too big"); #endif /* * This KASSERT is here to catch edge cases at a well defined place. * Before, those had triggered (random) panic conditions further * down. */ #ifdef BBR_INVARIANTS if (sack_rxmit) { if (SEQ_LT(rsm->r_start, tp->snd_una)) { panic("RSM:%p TP:%p bbr:%p start:%u is < snd_una:%u", rsm, tp, bbr, rsm->r_start, tp->snd_una); } } #endif KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); if ((len == 0) && (flags & TH_FIN) && (sbused(sb))) { /* * We have outstanding data, don't send a fin by itself!. */ slot = 0; goto just_return; } /* * Grab a header mbuf, attaching a copy of data to be transmitted, * and initialize the header from the template for sends on this * connection. */ if (len) { uint32_t moff; /* * We place a limit on sending with hptsi. */ if ((rsm == NULL) && len > pace_max_segs) len = pace_max_segs; if (len <= maxseg) tso = 0; #ifdef INET6 if (MHLEN < hdrlen + max_linkhdr) m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); else #endif m = m_gethdr(M_NOWAIT, MT_DATA); if (m == NULL) { BBR_STAT_INC(bbr_failed_mbuf_aloc); bbr_log_enobuf_jmp(bbr, len, cts, __LINE__, len, 0, 0); SOCKBUF_UNLOCK(sb); error = ENOBUFS; sack_rxmit = 0; goto out; } m->m_data += max_linkhdr; m->m_len = hdrlen; /* * Start the m_copy functions from the closest mbuf to the * sb_offset in the socket buffer chain. */ if ((sb_offset > sbavail(sb)) || ((len + sb_offset) > sbavail(sb))) { #ifdef BBR_INVARIANTS if ((len + sb_offset) > (sbavail(sb) + ((flags & (TH_FIN | TH_SYN)) ? 1 : 0))) panic("tp:%p bbr:%p len:%u sb_offset:%u sbavail:%u rsm:%p %u:%u:%u", tp, bbr, len, sb_offset, sbavail(sb), rsm, doing_retran_from, picked_up_retran, doing_tlp); #endif /* * In this messed up situation we have two choices, * a) pretend the send worked, and just start timers * and what not (not good since that may lead us * back here a lot). b) Send the lowest segment * in the map. c) Drop the connection. Lets do * which if it continues to happen will lead to * via timeouts. */ BBR_STAT_INC(bbr_offset_recovery); rsm = TAILQ_FIRST(&bbr->r_ctl.rc_map); sb_offset = 0; if (rsm == NULL) { sack_rxmit = 0; len = sbavail(sb); } else { sack_rxmit = 1; if (rsm->r_start != tp->snd_una) { /* * Things are really messed up, * is the only thing to do. */ BBR_STAT_INC(bbr_offset_drop); SOCKBUF_UNLOCK(sb); (void)m_free(m); return (-EFAULT); /* tcp_drop() */ } len = rsm->r_end - rsm->r_start; } if (len > sbavail(sb)) len = sbavail(sb); if (len > maxseg) len = maxseg; } mb = sbsndptr_noadv(sb, sb_offset, &moff); if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) { m_copydata(mb, moff, (int)len, mtod(m, caddr_t)+hdrlen); if (rsm == NULL) sbsndptr_adv(sb, mb, len); m->m_len += len; } else { struct sockbuf *msb; if (rsm) msb = NULL; else msb = sb; #ifdef BBR_INVARIANTS if ((len + moff) > (sbavail(sb) + ((flags & (TH_FIN | TH_SYN)) ? 1 : 0))) { if (rsm) { panic("tp:%p bbr:%p len:%u moff:%u sbavail:%u rsm:%p snd_una:%u rsm_start:%u flg:%x %u:%u:%u sr:%d ", tp, bbr, len, moff, sbavail(sb), rsm, tp->snd_una, rsm->r_flags, rsm->r_start, doing_retran_from, picked_up_retran, doing_tlp, sack_rxmit); } else { panic("tp:%p bbr:%p len:%u moff:%u sbavail:%u sb_offset:%u snd_una:%u", tp, bbr, len, moff, sbavail(sb), sb_offset, tp->snd_una); } } #endif m->m_next = tcp_m_copym( mb, moff, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb, ((rsm == NULL) ? hw_tls : 0) #ifdef NETFLIX_COPY_ARGS , &filled_all #endif ); if (len <= maxseg) { /* * Must have ran out of mbufs for the copy * shorten it to no longer need tso. Lets * not put on sendalot since we are low on * mbufs. */ tso = 0; } if (m->m_next == NULL) { SOCKBUF_UNLOCK(sb); (void)m_free(m); error = ENOBUFS; sack_rxmit = 0; goto out; } } #ifdef BBR_INVARIANTS if (tso && len < maxseg) { panic("tp:%p tso on, but len:%d < maxseg:%d", tp, len, maxseg); } if (tso && if_hw_tsomaxsegcount) { int32_t seg_cnt = 0; struct mbuf *foo; foo = m; while (foo) { seg_cnt++; foo = foo->m_next; } if (seg_cnt > if_hw_tsomaxsegcount) { panic("seg_cnt:%d > max:%d", seg_cnt, if_hw_tsomaxsegcount); } } #endif /* * If we're sending everything we've got, set PUSH. (This * will keep happy those implementations which only give * data to the user when a buffer fills or a PUSH comes in.) */ if (sb_offset + len == sbused(sb) && sbused(sb) && !(flags & TH_SYN)) { flags |= TH_PUSH; } SOCKBUF_UNLOCK(sb); } else { SOCKBUF_UNLOCK(sb); if (tp->t_flags & TF_ACKNOW) KMOD_TCPSTAT_INC(tcps_sndacks); else if (flags & (TH_SYN | TH_FIN | TH_RST)) KMOD_TCPSTAT_INC(tcps_sndctrl); else KMOD_TCPSTAT_INC(tcps_sndwinup); m = m_gethdr(M_NOWAIT, MT_DATA); if (m == NULL) { BBR_STAT_INC(bbr_failed_mbuf_aloc); bbr_log_enobuf_jmp(bbr, len, cts, __LINE__, len, 0, 0); error = ENOBUFS; /* Fudge the send time since we could not send */ sack_rxmit = 0; goto out; } #ifdef INET6 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && MHLEN >= hdrlen) { M_ALIGN(m, hdrlen); } else #endif m->m_data += max_linkhdr; m->m_len = hdrlen; } SOCKBUF_UNLOCK_ASSERT(sb); m->m_pkthdr.rcvif = (struct ifnet *)0; #ifdef MAC mac_inpcb_create_mbuf(inp, m); #endif #ifdef INET6 if (isipv6) { ip6 = mtod(m, struct ip6_hdr *); if (tp->t_port) { udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr)); udp->uh_sport = htons(V_tcp_udp_tunneling_port); udp->uh_dport = tp->t_port; ulen = hdrlen + len - sizeof(struct ip6_hdr); udp->uh_ulen = htons(ulen); th = (struct tcphdr *)(udp + 1); } else { th = (struct tcphdr *)(ip6 + 1); } tcpip_fillheaders(inp, tp->t_port, ip6, th); } else #endif /* INET6 */ { ip = mtod(m, struct ip *); #ifdef TCPDEBUG ipov = (struct ipovly *)ip; #endif if (tp->t_port) { udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip)); udp->uh_sport = htons(V_tcp_udp_tunneling_port); udp->uh_dport = tp->t_port; ulen = hdrlen + len - sizeof(struct ip); udp->uh_ulen = htons(ulen); th = (struct tcphdr *)(udp + 1); } else { th = (struct tcphdr *)(ip + 1); } tcpip_fillheaders(inp, tp->t_port, ip, th); } /* * If we are doing retransmissions, then snd_nxt will not reflect * the first unsent octet. For ACK only packets, we do not want the * sequence number of the retransmitted packet, we want the sequence * number of the next unsent octet. So, if there is no data (and no * SYN or FIN), use snd_max instead of snd_nxt when filling in * ti_seq. But if we are in persist state, snd_max might reflect * one byte beyond the right edge of the window, so use snd_nxt in * that case, since we know we aren't doing a retransmission. * (retransmit and persist are mutually exclusive...) */ if (sack_rxmit == 0) { if (len && ((flags & (TH_FIN | TH_SYN | TH_RST)) == 0)) { /* New data (including new persists) */ th->th_seq = htonl(tp->snd_max); bbr_seq = tp->snd_max; } else if (flags & TH_SYN) { /* Syn's always send from iss */ th->th_seq = htonl(tp->iss); bbr_seq = tp->iss; } else if (flags & TH_FIN) { if (flags & TH_FIN && tp->t_flags & TF_SENTFIN) { /* * If we sent the fin already its 1 minus * snd_max */ th->th_seq = (htonl(tp->snd_max - 1)); bbr_seq = (tp->snd_max - 1); } else { /* First time FIN use snd_max */ th->th_seq = htonl(tp->snd_max); bbr_seq = tp->snd_max; } } else { /* * len == 0 and not persist we use snd_max, sending * an ack unless we have sent the fin then its 1 * minus. */ /* * XXXRRS Question if we are in persists and we have * nothing outstanding to send and we have not sent * a FIN, we will send an ACK. In such a case it * might be better to send (tp->snd_una - 1) which * would force the peer to ack. */ if (tp->t_flags & TF_SENTFIN) { th->th_seq = htonl(tp->snd_max - 1); bbr_seq = (tp->snd_max - 1); } else { th->th_seq = htonl(tp->snd_max); bbr_seq = tp->snd_max; } } } else { /* All retransmits use the rsm to guide the send */ th->th_seq = htonl(rsm->r_start); bbr_seq = rsm->r_start; } th->th_ack = htonl(tp->rcv_nxt); if (optlen) { bcopy(opt, th + 1, optlen); th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; } tcp_set_flags(th, flags); /* * Calculate receive window. Don't shrink window, but avoid silly * window syndrome. */ if ((flags & TH_RST) || ((recwin < (so->so_rcv.sb_hiwat / 4) && recwin < maxseg))) recwin = 0; if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) && recwin < (tp->rcv_adv - tp->rcv_nxt)) recwin = (tp->rcv_adv - tp->rcv_nxt); if (recwin > TCP_MAXWIN << tp->rcv_scale) recwin = TCP_MAXWIN << tp->rcv_scale; /* * According to RFC1323 the window field in a SYN (i.e., a or * ) segment itself is never scaled. The case is * handled in syncache. */ if (flags & TH_SYN) th->th_win = htons((u_short) (min(sbspace(&so->so_rcv), TCP_MAXWIN))); else { /* Avoid shrinking window with window scaling. */ recwin = roundup2(recwin, 1 << tp->rcv_scale); th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); } /* * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0 * window. This may cause the remote transmitter to stall. This * flag tells soreceive() to disable delayed acknowledgements when * draining the buffer. This can occur if the receiver is * attempting to read more data than can be buffered prior to * transmitting on the connection. */ if (th->th_win == 0) { tp->t_sndzerowin++; tp->t_flags |= TF_RXWIN0SENT; } else tp->t_flags &= ~TF_RXWIN0SENT; /* * We don't support urgent data, but drag along * the pointer in case of a stack switch. */ tp->snd_up = tp->snd_una; #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) if (to.to_flags & TOF_SIGNATURE) { /* * Calculate MD5 signature and put it into the place * determined before. NOTE: since TCP options buffer doesn't * point into mbuf's data, calculate offset and use it. */ if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th, (u_char *)(th + 1) + (to.to_signature - opt)) != 0) { /* * Do not send segment if the calculation of MD5 * digest has failed. */ goto out; } } #endif /* * Put TCP length in extended header, and then checksum extended * header and data. */ m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ #ifdef INET6 if (isipv6) { /* * ip6_plen is not need to be filled now, and will be filled * in ip6_output. */ if (tp->t_port) { m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0); th->th_sum = htons(0); UDPSTAT_INC(udps_opackets); } else { csum_flags = m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP, 0); } } #endif #if defined(INET6) && defined(INET) else #endif #ifdef INET { if (tp->t_port) { m->m_pkthdr.csum_flags = CSUM_UDP; m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); udp->uh_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); th->th_sum = htons(0); UDPSTAT_INC(udps_opackets); } else { csum_flags = m->m_pkthdr.csum_flags = CSUM_TCP; m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen)); } /* IP version must be set here for ipv4/ipv6 checking later */ KASSERT(ip->ip_v == IPVERSION, ("%s: IP version incorrect: %d", __func__, ip->ip_v)); } #endif /* * Enable TSO and specify the size of the segments. The TCP pseudo * header checksum is always provided. XXX: Fixme: This is currently * not the case for IPv6. */ if (tso) { KASSERT(len > maxseg, ("%s: len:%d <= tso_segsz:%d", __func__, len, maxseg)); m->m_pkthdr.csum_flags |= CSUM_TSO; csum_flags |= CSUM_TSO; m->m_pkthdr.tso_segsz = maxseg; } KASSERT(len + hdrlen == m_length(m, NULL), ("%s: mbuf chain different than expected: %d + %u != %u", __func__, len, hdrlen, m_length(m, NULL))); #ifdef TCP_HHOOK /* Run HHOOK_TC_ESTABLISHED_OUT helper hooks. */ hhook_run_tcp_est_out(tp, th, &to, len, tso); #endif #ifdef TCPDEBUG /* * Trace. */ if (so->so_options & SO_DEBUG) { u_short save = 0; #ifdef INET6 if (!isipv6) #endif { save = ipov->ih_len; ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + * (th->th_off << 2) */ ); } tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0); #ifdef INET6 if (!isipv6) #endif ipov->ih_len = save; } #endif /* TCPDEBUG */ /* Log to the black box */ if (tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); /* Record info on type of transmission */ log.u_bbr.flex1 = bbr->r_ctl.rc_hptsi_agg_delay; log.u_bbr.flex2 = (bbr->r_recovery_bw << 3); log.u_bbr.flex3 = maxseg; log.u_bbr.flex4 = delay_calc; /* Encode filled_all into the upper flex5 bit */ log.u_bbr.flex5 = bbr->rc_past_init_win; log.u_bbr.flex5 <<= 1; log.u_bbr.flex5 |= bbr->rc_no_pacing; log.u_bbr.flex5 <<= 29; if (filled_all) log.u_bbr.flex5 |= 0x80000000; log.u_bbr.flex5 |= tp->t_maxseg; log.u_bbr.flex6 = bbr->r_ctl.rc_pace_max_segs; log.u_bbr.flex7 = (bbr->rc_bbr_state << 8) | bbr_state_val(bbr); /* lets poke in the low and the high here for debugging */ log.u_bbr.pkts_out = bbr->rc_tp->t_maxseg; if (rsm || sack_rxmit) { if (doing_tlp) log.u_bbr.flex8 = 2; else log.u_bbr.flex8 = 1; } else { log.u_bbr.flex8 = 0; } lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK, len, &log, false, NULL, NULL, 0, tv); } else { lgb = NULL; } /* * Fill in IP length and desired time to live and send to IP level. * There should be a better way to handle ttl and tos; we could keep * them in the template, but need a way to checksum without them. */ /* * m->m_pkthdr.len should have been set before cksum calcuration, * because in6_cksum() need it. */ #ifdef INET6 if (isipv6) { /* * we separately set hoplimit for every segment, since the * user might want to change the value via setsockopt. Also, * desired default hop limit might be changed via Neighbor * Discovery. */ ip6->ip6_hlim = in6_selecthlim(inp, NULL); /* * Set the packet size here for the benefit of DTrace * probes. ip6_output() will set it properly; it's supposed * to include the option header lengths as well. */ ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); if (V_path_mtu_discovery && maxseg > V_tcp_minmss) tp->t_flags2 |= TF2_PLPMTU_PMTUD; else tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; if (tp->t_state == TCPS_SYN_SENT) TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th); TCP_PROBE5(send, NULL, tp, ip6, tp, th); /* TODO: IPv6 IP6TOS_ECT bit on */ error = ip6_output(m, inp->in6p_outputopts, &inp->inp_route6, ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), NULL, NULL, inp); if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL) mtu = inp->inp_route6.ro_nh->nh_mtu; } #endif /* INET6 */ #if defined(INET) && defined(INET6) else #endif #ifdef INET { ip->ip_len = htons(m->m_pkthdr.len); #ifdef INET6 if (isipv6) ip->ip_ttl = in6_selecthlim(inp, NULL); #endif /* INET6 */ /* * If we do path MTU discovery, then we set DF on every * packet. This might not be the best thing to do according * to RFC3390 Section 2. However the tcp hostcache migitates * the problem so it affects only the first tcp connection * with a host. * * NB: Don't set DF on small MTU/MSS to have a safe * fallback. */ if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) { tp->t_flags2 |= TF2_PLPMTU_PMTUD; if (tp->t_port == 0 || len < V_tcp_minmss) { ip->ip_off |= htons(IP_DF); } } else { tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; } if (tp->t_state == TCPS_SYN_SENT) TCP_PROBE5(connect__request, NULL, tp, ip, tp, th); TCP_PROBE5(send, NULL, tp, ip, tp, th); error = ip_output(m, inp->inp_options, &inp->inp_route, ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0, inp); if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL) mtu = inp->inp_route.ro_nh->nh_mtu; } #endif /* INET */ out: if (lgb) { lgb->tlb_errno = error; lgb = NULL; } /* * In transmit state, time the transmission and arrange for the * retransmit. In persist state, just set snd_max. */ if (error == 0) { tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls); if (TCPS_HAVEESTABLISHED(tp->t_state) && (tp->t_flags & TF_SACK_PERMIT) && tp->rcv_numsacks > 0) tcp_clean_dsack_blocks(tp); /* We sent an ack clear the bbr_segs_rcvd count */ bbr->output_error_seen = 0; bbr->oerror_cnt = 0; bbr->bbr_segs_rcvd = 0; if (len == 0) counter_u64_add(bbr_out_size[TCP_MSS_ACCT_SNDACK], 1); /* Do accounting for new sends */ if ((len > 0) && (rsm == NULL)) { int idx; if (tp->snd_una == tp->snd_max) { /* * Special case to match google, when * nothing is in flight the delivered * time does get updated to the current * time (see tcp_rate_bsd.c). */ bbr->r_ctl.rc_del_time = cts; } if (len >= maxseg) { idx = (len / maxseg) + 3; if (idx >= TCP_MSS_ACCT_ATIMER) counter_u64_add(bbr_out_size[(TCP_MSS_ACCT_ATIMER - 1)], 1); else counter_u64_add(bbr_out_size[idx], 1); } else { /* smaller than a MSS */ idx = len / (bbr_hptsi_bytes_min - bbr->rc_last_options); if (idx >= TCP_MSS_SMALL_MAX_SIZE_DIV) idx = (TCP_MSS_SMALL_MAX_SIZE_DIV - 1); counter_u64_add(bbr_out_size[(idx + TCP_MSS_SMALL_SIZE_OFF)], 1); } } } abandon = 0; /* * We must do the send accounting before we log the output, * otherwise the state of the rsm could change and we account to the * wrong bucket. */ if (len > 0) { bbr_do_send_accounting(tp, bbr, rsm, len, error); if (error == 0) { if (tp->snd_una == tp->snd_max) bbr->r_ctl.rc_tlp_rxt_last_time = cts; } } bbr_log_output(bbr, tp, &to, len, bbr_seq, (uint8_t) flags, error, cts, mb, &abandon, rsm, 0, sb); if (abandon) { /* * If bbr_log_output destroys the TCB or sees a TH_RST being * sent we should hit this condition. */ return (0); } if (bbr->rc_in_persist == 0) { /* * Advance snd_nxt over sequence space of this segment. */ if (error) /* We don't log or do anything with errors */ goto skip_upd; if (tp->snd_una == tp->snd_max && (len || (flags & (TH_SYN | TH_FIN)))) { /* * Update the time we just added data since none was * outstanding. */ bbr_log_progress_event(bbr, tp, ticks, PROGRESS_START, __LINE__); bbr->rc_tp->t_acktime = ticks; } if (flags & (TH_SYN | TH_FIN) && (rsm == NULL)) { if (flags & TH_SYN) { /* * Smack the snd_max to iss + 1 * if its a FO we will add len below. */ tp->snd_max = tp->iss + 1; } if ((flags & TH_FIN) && ((tp->t_flags & TF_SENTFIN) == 0)) { tp->snd_max++; tp->t_flags |= TF_SENTFIN; } } if (sack_rxmit == 0) tp->snd_max += len; skip_upd: if ((error == 0) && len) tot_len += len; } else { /* Persists case */ int32_t xlen = len; if (error) goto nomore; if (flags & TH_SYN) ++xlen; if ((flags & TH_FIN) && ((tp->t_flags & TF_SENTFIN) == 0)) { ++xlen; tp->t_flags |= TF_SENTFIN; } if (xlen && (tp->snd_una == tp->snd_max)) { /* * Update the time we just added data since none was * outstanding. */ bbr_log_progress_event(bbr, tp, ticks, PROGRESS_START, __LINE__); bbr->rc_tp->t_acktime = ticks; } if (sack_rxmit == 0) tp->snd_max += xlen; tot_len += (len + optlen + ipoptlen); } nomore: if (error) { /* * Failures do not advance the seq counter above. For the * case of ENOBUFS we will fall out and become ack-clocked. * capping the cwnd at the current flight. * Everything else will just have to retransmit with the timer * (no pacer). */ SOCKBUF_UNLOCK_ASSERT(sb); BBR_STAT_INC(bbr_saw_oerr); /* Clear all delay/early tracks */ bbr->r_ctl.rc_hptsi_agg_delay = 0; bbr->r_ctl.rc_agg_early = 0; bbr->r_agg_early_set = 0; bbr->output_error_seen = 1; if (bbr->oerror_cnt < 0xf) bbr->oerror_cnt++; if (bbr_max_net_error_cnt && (bbr->oerror_cnt >= bbr_max_net_error_cnt)) { /* drop the session */ return (-ENETDOWN); } switch (error) { case ENOBUFS: /* * Make this guy have to get ack's to send * more but lets make sure we don't * slam him below a T-O (1MSS). */ if (bbr->rc_bbr_state != BBR_STATE_PROBE_RTT) { tp->snd_cwnd = ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) - maxseg; if (tp->snd_cwnd < maxseg) tp->snd_cwnd = maxseg; } slot = (bbr_error_base_paceout + 1) << bbr->oerror_cnt; BBR_STAT_INC(bbr_saw_enobuf); if (bbr->bbr_hdrw_pacing) counter_u64_add(bbr_hdwr_pacing_enobuf, 1); else counter_u64_add(bbr_nohdwr_pacing_enobuf, 1); /* * Here even in the enobuf's case we want to do our * state update. The reason being we may have been * called by the input function. If so we have had * things change. */ error = 0; goto enobufs; case EMSGSIZE: /* * For some reason the interface we used initially * to send segments changed to another or lowered * its MTU. If TSO was active we either got an * interface without TSO capabilits or TSO was * turned off. If we obtained mtu from ip_output() * then update it and try again. */ /* Turn on tracing (or try to) */ { int old_maxseg; old_maxseg = tp->t_maxseg; BBR_STAT_INC(bbr_saw_emsgsiz); bbr_log_msgsize_fail(bbr, tp, len, maxseg, mtu, csum_flags, tso, cts); if (mtu != 0) tcp_mss_update(tp, -1, mtu, NULL, NULL); if (old_maxseg <= tp->t_maxseg) { /* Huh it did not shrink? */ tp->t_maxseg = old_maxseg - 40; bbr_log_msgsize_fail(bbr, tp, len, maxseg, mtu, 0, tso, cts); } /* * Nuke all other things that can interfere * with slot */ if ((tot_len + len) && (len >= tp->t_maxseg)) { slot = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, (tot_len + len), cts, 0); if (slot < bbr_error_base_paceout) slot = (bbr_error_base_paceout + 2) << bbr->oerror_cnt; } else slot = (bbr_error_base_paceout + 2) << bbr->oerror_cnt; bbr->rc_output_starts_timer = 1; bbr_start_hpts_timer(bbr, tp, cts, 10, slot, tot_len); return (error); } case EPERM: tp->t_softerror = error; /* Fall through */ case EHOSTDOWN: case EHOSTUNREACH: case ENETDOWN: case ENETUNREACH: if (TCPS_HAVERCVDSYN(tp->t_state)) { tp->t_softerror = error; } /* FALLTHROUGH */ default: slot = (bbr_error_base_paceout + 3) << bbr->oerror_cnt; bbr->rc_output_starts_timer = 1; bbr_start_hpts_timer(bbr, tp, cts, 11, slot, 0); return (error); } #ifdef STATS } else if (((tp->t_flags & TF_GPUTINPROG) == 0) && len && (rsm == NULL) && (bbr->rc_in_persist == 0)) { tp->gput_seq = bbr_seq; tp->gput_ack = bbr_seq + min(sbavail(&so->so_snd) - sb_offset, sendwin); tp->gput_ts = cts; tp->t_flags |= TF_GPUTINPROG; #endif } KMOD_TCPSTAT_INC(tcps_sndtotal); if ((bbr->bbr_hdw_pace_ena) && (bbr->bbr_attempt_hdwr_pace == 0) && (bbr->rc_past_init_win) && (bbr->rc_bbr_state != BBR_STATE_STARTUP) && (get_filter_value(&bbr->r_ctl.rc_delrate)) && (inp->inp_route.ro_nh && inp->inp_route.ro_nh->nh_ifp)) { /* * We are past the initial window and * have at least one measurement so we * could use hardware pacing if its available. * We have an interface and we have not attempted * to setup hardware pacing, lets try to now. */ uint64_t rate_wanted; int err = 0; rate_wanted = bbr_get_hardware_rate(bbr); bbr->bbr_attempt_hdwr_pace = 1; bbr->r_ctl.crte = tcp_set_pacing_rate(bbr->rc_tp, inp->inp_route.ro_nh->nh_ifp, rate_wanted, (RS_PACING_GEQ|RS_PACING_SUB_OK), &err, NULL); if (bbr->r_ctl.crte) { bbr_type_log_hdwr_pacing(bbr, bbr->r_ctl.crte->ptbl->rs_ifp, rate_wanted, bbr->r_ctl.crte->rate, __LINE__, cts, err); BBR_STAT_INC(bbr_hdwr_rl_add_ok); counter_u64_add(bbr_flows_nohdwr_pacing, -1); counter_u64_add(bbr_flows_whdwr_pacing, 1); bbr->bbr_hdrw_pacing = 1; /* Now what is our gain status? */ if (bbr->r_ctl.crte->rate < rate_wanted) { /* We have a problem */ bbr_setup_less_of_rate(bbr, cts, bbr->r_ctl.crte->rate, rate_wanted); } else { /* We are good */ bbr->gain_is_limited = 0; bbr->skip_gain = 0; } tcp_bbr_tso_size_check(bbr, cts); } else { bbr_type_log_hdwr_pacing(bbr, inp->inp_route.ro_nh->nh_ifp, rate_wanted, 0, __LINE__, cts, err); BBR_STAT_INC(bbr_hdwr_rl_add_fail); } } if (bbr->bbr_hdrw_pacing) { /* * Worry about cases where the route * changes or something happened that we * lost our hardware pacing possibly during * the last ip_output call. */ if (inp->inp_snd_tag == NULL) { /* A change during ip output disabled hw pacing? */ bbr->bbr_hdrw_pacing = 0; } else if ((inp->inp_route.ro_nh == NULL) || (inp->inp_route.ro_nh->nh_ifp != inp->inp_snd_tag->ifp)) { /* * We had an interface or route change, * detach from the current hdwr pacing * and setup to re-attempt next go * round. */ bbr->bbr_hdrw_pacing = 0; bbr->bbr_attempt_hdwr_pace = 0; tcp_rel_pacing_rate(bbr->r_ctl.crte, bbr->rc_tp); tcp_bbr_tso_size_check(bbr, cts); } } /* * Data sent (as far as we can tell). If this advertises a larger * window than any other segment, then remember the size of the * advertised window. Any pending ACK has now been sent. */ if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) tp->rcv_adv = tp->rcv_nxt + recwin; tp->last_ack_sent = tp->rcv_nxt; if ((error == 0) && (bbr->r_ctl.rc_pace_max_segs > tp->t_maxseg) && (doing_tlp == 0) && (tso == 0) && (len > 0) && ((flags & TH_RST) == 0) && ((flags & TH_SYN) == 0) && (IN_RECOVERY(tp->t_flags) == 0) && (bbr->rc_in_persist == 0) && (tot_len < bbr->r_ctl.rc_pace_max_segs)) { /* * For non-tso we need to goto again until we have sent out * enough data to match what we are hptsi out every hptsi * interval. */ if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { /* Make sure snd_nxt is drug up */ tp->snd_nxt = tp->snd_max; } if (rsm != NULL) { rsm = NULL; goto skip_again; } rsm = NULL; sack_rxmit = 0; tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); goto again; } skip_again: if ((error == 0) && (flags & TH_FIN)) tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN); if ((error == 0) && (flags & TH_RST)) tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); if (((flags & (TH_RST | TH_SYN | TH_FIN)) == 0) && tot_len) { /* * Calculate/Re-Calculate the hptsi slot in usecs based on * what we have sent so far */ slot = bbr_get_pacing_delay(bbr, bbr->r_ctl.rc_bbr_hptsi_gain, tot_len, cts, 0); if (bbr->rc_no_pacing) slot = 0; } tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); enobufs: if (bbr->rc_use_google == 0) bbr_check_bbr_for_state(bbr, cts, __LINE__, 0); bbr_cwnd_limiting(tp, bbr, ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes))); bbr->rc_output_starts_timer = 1; if (bbr->bbr_use_rack_cheat && (more_to_rxt || ((bbr->r_ctl.rc_resend = bbr_check_recovery_mode(tp, bbr, cts)) != NULL))) { /* Rack cheats and shotguns out all rxt's 1ms apart */ if (slot > 1000) slot = 1000; } if (bbr->bbr_hdrw_pacing && (bbr->hw_pacing_set == 0)) { /* * We don't change the tso size until some number of sends * to give the hardware commands time to get down * to the interface. */ bbr->r_ctl.bbr_hdwr_cnt_noset_snt++; if (bbr->r_ctl.bbr_hdwr_cnt_noset_snt >= bbr_hdwr_pacing_delay_cnt) { bbr->hw_pacing_set = 1; tcp_bbr_tso_size_check(bbr, cts); } } bbr_start_hpts_timer(bbr, tp, cts, 12, slot, tot_len); if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { /* Make sure snd_nxt is drug up */ tp->snd_nxt = tp->snd_max; } return (error); } /* * See bbr_output_wtime() for return values. */ static int bbr_output(struct tcpcb *tp) { int32_t ret; struct timeval tv; NET_EPOCH_ASSERT(); INP_WLOCK_ASSERT(tp->t_inpcb); (void)tcp_get_usecs(&tv); ret = bbr_output_wtime(tp, &tv); return (ret); } static void bbr_mtu_chg(struct tcpcb *tp) { struct tcp_bbr *bbr; struct bbr_sendmap *rsm, *frsm = NULL; uint32_t maxseg; /* * The MTU has changed. a) Clear the sack filter. b) Mark everything * over the current size as SACK_PASS so a retransmit will occur. */ bbr = (struct tcp_bbr *)tp->t_fb_ptr; maxseg = tp->t_maxseg - bbr->rc_last_options; sack_filter_clear(&bbr->r_ctl.bbr_sf, tp->snd_una); TAILQ_FOREACH(rsm, &bbr->r_ctl.rc_map, r_next) { /* Don't mess with ones acked (by sack?) */ if (rsm->r_flags & BBR_ACKED) continue; if ((rsm->r_end - rsm->r_start) > maxseg) { /* * We mark sack-passed on all the previous large * sends we did. This will force them to retransmit. */ rsm->r_flags |= BBR_SACK_PASSED; if (((rsm->r_flags & BBR_MARKED_LOST) == 0) && bbr_is_lost(bbr, rsm, bbr->r_ctl.rc_rcvtime)) { bbr->r_ctl.rc_lost_bytes += rsm->r_end - rsm->r_start; bbr->r_ctl.rc_lost += rsm->r_end - rsm->r_start; rsm->r_flags |= BBR_MARKED_LOST; } if (frsm == NULL) frsm = rsm; } } if (frsm) { bbr->r_ctl.rc_resend = frsm; } } static int bbr_pru_options(struct tcpcb *tp, int flags) { if (flags & PRUS_OOB) return (EOPNOTSUPP); return (0); } struct tcp_function_block __tcp_bbr = { .tfb_tcp_block_name = __XSTRING(STACKNAME), .tfb_tcp_output = bbr_output, .tfb_do_queued_segments = ctf_do_queued_segments, .tfb_do_segment_nounlock = bbr_do_segment_nounlock, .tfb_tcp_do_segment = bbr_do_segment, .tfb_tcp_ctloutput = bbr_ctloutput, .tfb_tcp_fb_init = bbr_init, .tfb_tcp_fb_fini = bbr_fini, .tfb_tcp_timer_stop_all = bbr_stopall, .tfb_tcp_timer_activate = bbr_timer_activate, .tfb_tcp_timer_active = bbr_timer_active, .tfb_tcp_timer_stop = bbr_timer_stop, .tfb_tcp_rexmit_tmr = bbr_remxt_tmr, .tfb_tcp_handoff_ok = bbr_handoff_ok, .tfb_tcp_mtu_chg = bbr_mtu_chg, .tfb_pru_options = bbr_pru_options, .tfb_flags = TCP_FUNC_OUTPUT_CANDROP, }; /* * bbr_ctloutput() must drop the inpcb lock before performing copyin on * socket option arguments. When it re-acquires the lock after the copy, it * has to revalidate that the connection is still valid for the socket * option. */ static int bbr_set_sockopt(struct inpcb *inp, struct sockopt *sopt) { struct epoch_tracker et; struct tcpcb *tp; struct tcp_bbr *bbr; int32_t error = 0, optval; switch (sopt->sopt_level) { case IPPROTO_IPV6: case IPPROTO_IP: return (tcp_default_ctloutput(inp, sopt)); } switch (sopt->sopt_name) { case TCP_RACK_PACE_MAX_SEG: case TCP_RACK_MIN_TO: case TCP_RACK_REORD_THRESH: case TCP_RACK_REORD_FADE: case TCP_RACK_TLP_THRESH: case TCP_RACK_PKT_DELAY: case TCP_BBR_ALGORITHM: case TCP_BBR_TSLIMITS: case TCP_BBR_IWINTSO: case TCP_BBR_RECFORCE: case TCP_BBR_STARTUP_PG: case TCP_BBR_DRAIN_PG: case TCP_BBR_RWND_IS_APP: case TCP_BBR_PROBE_RTT_INT: case TCP_BBR_PROBE_RTT_GAIN: case TCP_BBR_PROBE_RTT_LEN: case TCP_BBR_STARTUP_LOSS_EXIT: case TCP_BBR_USEDEL_RATE: case TCP_BBR_MIN_RTO: case TCP_BBR_MAX_RTO: case TCP_BBR_PACE_PER_SEC: case TCP_DELACK: case TCP_BBR_PACE_DEL_TAR: case TCP_BBR_SEND_IWND_IN_TSO: case TCP_BBR_EXTRA_STATE: case TCP_BBR_UTTER_MAX_TSO: case TCP_BBR_MIN_TOPACEOUT: case TCP_BBR_FLOOR_MIN_TSO: case TCP_BBR_TSTMP_RAISES: case TCP_BBR_POLICER_DETECT: case TCP_BBR_USE_RACK_CHEAT: case TCP_DATA_AFTER_CLOSE: case TCP_BBR_HDWR_PACE: case TCP_BBR_PACE_SEG_MAX: case TCP_BBR_PACE_SEG_MIN: case TCP_BBR_PACE_CROSS: case TCP_BBR_PACE_OH: #ifdef NETFLIX_PEAKRATE case TCP_MAXPEAKRATE: #endif case TCP_BBR_TMR_PACE_OH: case TCP_BBR_RACK_RTT_USE: case TCP_BBR_RETRAN_WTSO: break; default: return (tcp_default_ctloutput(inp, sopt)); break; } INP_WUNLOCK(inp); error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval)); if (error) return (error); INP_WLOCK(inp); - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { INP_WUNLOCK(inp); return (ECONNRESET); } tp = intotcpcb(inp); if (tp->t_fb != &__tcp_bbr) { INP_WUNLOCK(inp); return (ENOPROTOOPT); } bbr = (struct tcp_bbr *)tp->t_fb_ptr; switch (sopt->sopt_name) { case TCP_BBR_PACE_PER_SEC: BBR_OPTS_INC(tcp_bbr_pace_per_sec); bbr->r_ctl.bbr_hptsi_per_second = optval; break; case TCP_BBR_PACE_DEL_TAR: BBR_OPTS_INC(tcp_bbr_pace_del_tar); bbr->r_ctl.bbr_hptsi_segments_delay_tar = optval; break; case TCP_BBR_PACE_SEG_MAX: BBR_OPTS_INC(tcp_bbr_pace_seg_max); bbr->r_ctl.bbr_hptsi_segments_max = optval; break; case TCP_BBR_PACE_SEG_MIN: BBR_OPTS_INC(tcp_bbr_pace_seg_min); bbr->r_ctl.bbr_hptsi_bytes_min = optval; break; case TCP_BBR_PACE_CROSS: BBR_OPTS_INC(tcp_bbr_pace_cross); bbr->r_ctl.bbr_cross_over = optval; break; case TCP_BBR_ALGORITHM: BBR_OPTS_INC(tcp_bbr_algorithm); if (optval && (bbr->rc_use_google == 0)) { /* Turn on the google mode */ bbr_google_mode_on(bbr); if ((optval > 3) && (optval < 500)) { /* * Must be at least greater than .3% * and must be less than 50.0%. */ bbr->r_ctl.bbr_google_discount = optval; } } else if ((optval == 0) && (bbr->rc_use_google == 1)) { /* Turn off the google mode */ bbr_google_mode_off(bbr); } break; case TCP_BBR_TSLIMITS: BBR_OPTS_INC(tcp_bbr_tslimits); if (optval == 1) bbr->rc_use_ts_limit = 1; else if (optval == 0) bbr->rc_use_ts_limit = 0; else error = EINVAL; break; case TCP_BBR_IWINTSO: BBR_OPTS_INC(tcp_bbr_iwintso); if ((optval >= 0) && (optval < 128)) { uint32_t twin; bbr->rc_init_win = optval; twin = bbr_initial_cwnd(bbr, tp); if ((bbr->rc_past_init_win == 0) && (twin > tp->snd_cwnd)) tp->snd_cwnd = twin; else error = EBUSY; } else error = EINVAL; break; case TCP_BBR_STARTUP_PG: BBR_OPTS_INC(tcp_bbr_startup_pg); if ((optval > 0) && (optval < BBR_MAX_GAIN_VALUE)) { bbr->r_ctl.rc_startup_pg = optval; if (bbr->rc_bbr_state == BBR_STATE_STARTUP) { bbr->r_ctl.rc_bbr_hptsi_gain = optval; } } else error = EINVAL; break; case TCP_BBR_DRAIN_PG: BBR_OPTS_INC(tcp_bbr_drain_pg); if ((optval > 0) && (optval < BBR_MAX_GAIN_VALUE)) bbr->r_ctl.rc_drain_pg = optval; else error = EINVAL; break; case TCP_BBR_PROBE_RTT_LEN: BBR_OPTS_INC(tcp_bbr_probertt_len); if (optval <= 1) reset_time_small(&bbr->r_ctl.rc_rttprop, (optval * USECS_IN_SECOND)); else error = EINVAL; break; case TCP_BBR_PROBE_RTT_GAIN: BBR_OPTS_INC(tcp_bbr_probertt_gain); if (optval <= BBR_UNIT) bbr->r_ctl.bbr_rttprobe_gain_val = optval; else error = EINVAL; break; case TCP_BBR_PROBE_RTT_INT: BBR_OPTS_INC(tcp_bbr_probe_rtt_int); if (optval > 1000) bbr->r_ctl.rc_probertt_int = optval; else error = EINVAL; break; case TCP_BBR_MIN_TOPACEOUT: BBR_OPTS_INC(tcp_bbr_topaceout); if (optval == 0) { bbr->no_pacing_until = 0; bbr->rc_no_pacing = 0; } else if (optval <= 0x00ff) { bbr->no_pacing_until = optval; if ((bbr->r_ctl.rc_pkt_epoch < bbr->no_pacing_until) && (bbr->rc_bbr_state == BBR_STATE_STARTUP)){ /* Turn on no pacing */ bbr->rc_no_pacing = 1; } } else error = EINVAL; break; case TCP_BBR_STARTUP_LOSS_EXIT: BBR_OPTS_INC(tcp_bbr_startup_loss_exit); bbr->rc_loss_exit = optval; break; case TCP_BBR_USEDEL_RATE: error = EINVAL; break; case TCP_BBR_MIN_RTO: BBR_OPTS_INC(tcp_bbr_min_rto); bbr->r_ctl.rc_min_rto_ms = optval; break; case TCP_BBR_MAX_RTO: BBR_OPTS_INC(tcp_bbr_max_rto); bbr->rc_max_rto_sec = optval; break; case TCP_RACK_MIN_TO: /* Minimum time between rack t-o's in ms */ BBR_OPTS_INC(tcp_rack_min_to); bbr->r_ctl.rc_min_to = optval; break; case TCP_RACK_REORD_THRESH: /* RACK reorder threshold (shift amount) */ BBR_OPTS_INC(tcp_rack_reord_thresh); if ((optval > 0) && (optval < 31)) bbr->r_ctl.rc_reorder_shift = optval; else error = EINVAL; break; case TCP_RACK_REORD_FADE: /* Does reordering fade after ms time */ BBR_OPTS_INC(tcp_rack_reord_fade); bbr->r_ctl.rc_reorder_fade = optval; break; case TCP_RACK_TLP_THRESH: /* RACK TLP theshold i.e. srtt+(srtt/N) */ BBR_OPTS_INC(tcp_rack_tlp_thresh); if (optval) bbr->rc_tlp_threshold = optval; else error = EINVAL; break; case TCP_BBR_USE_RACK_CHEAT: BBR_OPTS_INC(tcp_use_rackcheat); if (bbr->rc_use_google) { error = EINVAL; break; } BBR_OPTS_INC(tcp_rack_cheat); if (optval) bbr->bbr_use_rack_cheat = 1; else bbr->bbr_use_rack_cheat = 0; break; case TCP_BBR_FLOOR_MIN_TSO: BBR_OPTS_INC(tcp_utter_max_tso); if ((optval >= 0) && (optval < 40)) bbr->r_ctl.bbr_hptsi_segments_floor = optval; else error = EINVAL; break; case TCP_BBR_UTTER_MAX_TSO: BBR_OPTS_INC(tcp_utter_max_tso); if ((optval >= 0) && (optval < 0xffff)) bbr->r_ctl.bbr_utter_max = optval; else error = EINVAL; break; case TCP_BBR_EXTRA_STATE: BBR_OPTS_INC(tcp_extra_state); if (optval) bbr->rc_use_idle_restart = 1; else bbr->rc_use_idle_restart = 0; break; case TCP_BBR_SEND_IWND_IN_TSO: BBR_OPTS_INC(tcp_iwnd_tso); if (optval) { bbr->bbr_init_win_cheat = 1; if (bbr->rc_past_init_win == 0) { uint32_t cts; cts = tcp_get_usecs(&bbr->rc_tv); tcp_bbr_tso_size_check(bbr, cts); } } else bbr->bbr_init_win_cheat = 0; break; case TCP_BBR_HDWR_PACE: BBR_OPTS_INC(tcp_hdwr_pacing); if (optval){ bbr->bbr_hdw_pace_ena = 1; bbr->bbr_attempt_hdwr_pace = 0; } else { bbr->bbr_hdw_pace_ena = 0; #ifdef RATELIMIT if (bbr->r_ctl.crte != NULL) { tcp_rel_pacing_rate(bbr->r_ctl.crte, tp); bbr->r_ctl.crte = NULL; } #endif } break; case TCP_DELACK: BBR_OPTS_INC(tcp_delack); if (optval < 100) { if (optval == 0) /* off */ tp->t_delayed_ack = 0; else if (optval == 1) /* on which is 2 */ tp->t_delayed_ack = 2; else /* higher than 2 and less than 100 */ tp->t_delayed_ack = optval; if (tp->t_flags & TF_DELACK) { tp->t_flags &= ~TF_DELACK; tp->t_flags |= TF_ACKNOW; NET_EPOCH_ENTER(et); bbr_output(tp); NET_EPOCH_EXIT(et); } } else error = EINVAL; break; case TCP_RACK_PKT_DELAY: /* RACK added ms i.e. rack-rtt + reord + N */ BBR_OPTS_INC(tcp_rack_pkt_delay); bbr->r_ctl.rc_pkt_delay = optval; break; #ifdef NETFLIX_PEAKRATE case TCP_MAXPEAKRATE: BBR_OPTS_INC(tcp_maxpeak); error = tcp_set_maxpeakrate(tp, optval); if (!error) tp->t_peakrate_thr = tp->t_maxpeakrate; break; #endif case TCP_BBR_RETRAN_WTSO: BBR_OPTS_INC(tcp_retran_wtso); if (optval) bbr->rc_resends_use_tso = 1; else bbr->rc_resends_use_tso = 0; break; case TCP_DATA_AFTER_CLOSE: BBR_OPTS_INC(tcp_data_ac); if (optval) bbr->rc_allow_data_af_clo = 1; else bbr->rc_allow_data_af_clo = 0; break; case TCP_BBR_POLICER_DETECT: BBR_OPTS_INC(tcp_policer_det); if (bbr->rc_use_google == 0) error = EINVAL; else if (optval) bbr->r_use_policer = 1; else bbr->r_use_policer = 0; break; case TCP_BBR_TSTMP_RAISES: BBR_OPTS_INC(tcp_ts_raises); if (optval) bbr->ts_can_raise = 1; else bbr->ts_can_raise = 0; break; case TCP_BBR_TMR_PACE_OH: BBR_OPTS_INC(tcp_pacing_oh_tmr); if (bbr->rc_use_google) { error = EINVAL; } else { if (optval) bbr->r_ctl.rc_incr_tmrs = 1; else bbr->r_ctl.rc_incr_tmrs = 0; } break; case TCP_BBR_PACE_OH: BBR_OPTS_INC(tcp_pacing_oh); if (bbr->rc_use_google) { error = EINVAL; } else { if (optval > (BBR_INCL_TCP_OH| BBR_INCL_IP_OH| BBR_INCL_ENET_OH)) { error = EINVAL; break; } if (optval & BBR_INCL_TCP_OH) bbr->r_ctl.rc_inc_tcp_oh = 1; else bbr->r_ctl.rc_inc_tcp_oh = 0; if (optval & BBR_INCL_IP_OH) bbr->r_ctl.rc_inc_ip_oh = 1; else bbr->r_ctl.rc_inc_ip_oh = 0; if (optval & BBR_INCL_ENET_OH) bbr->r_ctl.rc_inc_enet_oh = 1; else bbr->r_ctl.rc_inc_enet_oh = 0; } break; default: return (tcp_default_ctloutput(inp, sopt)); break; } #ifdef NETFLIX_STATS tcp_log_socket_option(tp, sopt->sopt_name, optval, error); #endif INP_WUNLOCK(inp); return (error); } /* * return 0 on success, error-num on failure */ static int bbr_get_sockopt(struct inpcb *inp, struct sockopt *sopt) { struct tcpcb *tp; struct tcp_bbr *bbr; int32_t error, optval; tp = intotcpcb(inp); bbr = (struct tcp_bbr *)tp->t_fb_ptr; if (bbr == NULL) { INP_WUNLOCK(inp); return (EINVAL); } /* * Because all our options are either boolean or an int, we can just * pull everything into optval and then unlock and copy. If we ever * add a option that is not a int, then this will have quite an * impact to this routine. */ switch (sopt->sopt_name) { case TCP_BBR_PACE_PER_SEC: optval = bbr->r_ctl.bbr_hptsi_per_second; break; case TCP_BBR_PACE_DEL_TAR: optval = bbr->r_ctl.bbr_hptsi_segments_delay_tar; break; case TCP_BBR_PACE_SEG_MAX: optval = bbr->r_ctl.bbr_hptsi_segments_max; break; case TCP_BBR_MIN_TOPACEOUT: optval = bbr->no_pacing_until; break; case TCP_BBR_PACE_SEG_MIN: optval = bbr->r_ctl.bbr_hptsi_bytes_min; break; case TCP_BBR_PACE_CROSS: optval = bbr->r_ctl.bbr_cross_over; break; case TCP_BBR_ALGORITHM: optval = bbr->rc_use_google; break; case TCP_BBR_TSLIMITS: optval = bbr->rc_use_ts_limit; break; case TCP_BBR_IWINTSO: optval = bbr->rc_init_win; break; case TCP_BBR_STARTUP_PG: optval = bbr->r_ctl.rc_startup_pg; break; case TCP_BBR_DRAIN_PG: optval = bbr->r_ctl.rc_drain_pg; break; case TCP_BBR_PROBE_RTT_INT: optval = bbr->r_ctl.rc_probertt_int; break; case TCP_BBR_PROBE_RTT_LEN: optval = (bbr->r_ctl.rc_rttprop.cur_time_limit / USECS_IN_SECOND); break; case TCP_BBR_PROBE_RTT_GAIN: optval = bbr->r_ctl.bbr_rttprobe_gain_val; break; case TCP_BBR_STARTUP_LOSS_EXIT: optval = bbr->rc_loss_exit; break; case TCP_BBR_USEDEL_RATE: error = EINVAL; break; case TCP_BBR_MIN_RTO: optval = bbr->r_ctl.rc_min_rto_ms; break; case TCP_BBR_MAX_RTO: optval = bbr->rc_max_rto_sec; break; case TCP_RACK_PACE_MAX_SEG: /* Max segments in a pace */ optval = bbr->r_ctl.rc_pace_max_segs; break; case TCP_RACK_MIN_TO: /* Minimum time between rack t-o's in ms */ optval = bbr->r_ctl.rc_min_to; break; case TCP_RACK_REORD_THRESH: /* RACK reorder threshold (shift amount) */ optval = bbr->r_ctl.rc_reorder_shift; break; case TCP_RACK_REORD_FADE: /* Does reordering fade after ms time */ optval = bbr->r_ctl.rc_reorder_fade; break; case TCP_BBR_USE_RACK_CHEAT: /* Do we use the rack cheat for rxt */ optval = bbr->bbr_use_rack_cheat; break; case TCP_BBR_FLOOR_MIN_TSO: optval = bbr->r_ctl.bbr_hptsi_segments_floor; break; case TCP_BBR_UTTER_MAX_TSO: optval = bbr->r_ctl.bbr_utter_max; break; case TCP_BBR_SEND_IWND_IN_TSO: /* Do we send TSO size segments initially */ optval = bbr->bbr_init_win_cheat; break; case TCP_BBR_EXTRA_STATE: optval = bbr->rc_use_idle_restart; break; case TCP_RACK_TLP_THRESH: /* RACK TLP theshold i.e. srtt+(srtt/N) */ optval = bbr->rc_tlp_threshold; break; case TCP_RACK_PKT_DELAY: /* RACK added ms i.e. rack-rtt + reord + N */ optval = bbr->r_ctl.rc_pkt_delay; break; case TCP_BBR_RETRAN_WTSO: optval = bbr->rc_resends_use_tso; break; case TCP_DATA_AFTER_CLOSE: optval = bbr->rc_allow_data_af_clo; break; case TCP_DELACK: optval = tp->t_delayed_ack; break; case TCP_BBR_HDWR_PACE: optval = bbr->bbr_hdw_pace_ena; break; case TCP_BBR_POLICER_DETECT: optval = bbr->r_use_policer; break; case TCP_BBR_TSTMP_RAISES: optval = bbr->ts_can_raise; break; case TCP_BBR_TMR_PACE_OH: optval = bbr->r_ctl.rc_incr_tmrs; break; case TCP_BBR_PACE_OH: optval = 0; if (bbr->r_ctl.rc_inc_tcp_oh) optval |= BBR_INCL_TCP_OH; if (bbr->r_ctl.rc_inc_ip_oh) optval |= BBR_INCL_IP_OH; if (bbr->r_ctl.rc_inc_enet_oh) optval |= BBR_INCL_ENET_OH; break; default: return (tcp_default_ctloutput(inp, sopt)); break; } INP_WUNLOCK(inp); error = sooptcopyout(sopt, &optval, sizeof optval); return (error); } /* * return 0 on success, error-num on failure */ static int bbr_ctloutput(struct inpcb *inp, struct sockopt *sopt) { if (sopt->sopt_dir == SOPT_SET) { return (bbr_set_sockopt(inp, sopt)); } else if (sopt->sopt_dir == SOPT_GET) { return (bbr_get_sockopt(inp, sopt)); } else { panic("%s: sopt_dir $%d", __func__, sopt->sopt_dir); } } static const char *bbr_stack_names[] = { __XSTRING(STACKNAME), #ifdef STACKALIAS __XSTRING(STACKALIAS), #endif }; static bool bbr_mod_inited = false; static int tcp_addbbr(module_t mod, int32_t type, void *data) { int32_t err = 0; int num_stacks; switch (type) { case MOD_LOAD: printf("Attempting to load " __XSTRING(MODNAME) "\n"); bbr_zone = uma_zcreate(__XSTRING(MODNAME) "_map", sizeof(struct bbr_sendmap), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); bbr_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb", sizeof(struct tcp_bbr), NULL, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0); sysctl_ctx_init(&bbr_sysctl_ctx); bbr_sysctl_root = SYSCTL_ADD_NODE(&bbr_sysctl_ctx, SYSCTL_STATIC_CHILDREN(_net_inet_tcp), OID_AUTO, #ifdef STACKALIAS __XSTRING(STACKALIAS), #else __XSTRING(STACKNAME), #endif CTLFLAG_RW | CTLFLAG_MPSAFE, 0, ""); if (bbr_sysctl_root == NULL) { printf("Failed to add sysctl node\n"); err = EFAULT; goto free_uma; } bbr_init_sysctls(); num_stacks = nitems(bbr_stack_names); err = register_tcp_functions_as_names(&__tcp_bbr, M_WAITOK, bbr_stack_names, &num_stacks); if (err) { printf("Failed to register %s stack name for " "%s module\n", bbr_stack_names[num_stacks], __XSTRING(MODNAME)); sysctl_ctx_free(&bbr_sysctl_ctx); free_uma: uma_zdestroy(bbr_zone); uma_zdestroy(bbr_pcb_zone); bbr_counter_destroy(); printf("Failed to register " __XSTRING(MODNAME) " module err:%d\n", err); return (err); } tcp_lro_reg_mbufq(); bbr_mod_inited = true; printf(__XSTRING(MODNAME) " is now available\n"); break; case MOD_QUIESCE: err = deregister_tcp_functions(&__tcp_bbr, true, false); break; case MOD_UNLOAD: err = deregister_tcp_functions(&__tcp_bbr, false, true); if (err == EBUSY) break; if (bbr_mod_inited) { uma_zdestroy(bbr_zone); uma_zdestroy(bbr_pcb_zone); sysctl_ctx_free(&bbr_sysctl_ctx); bbr_counter_destroy(); printf(__XSTRING(MODNAME) " is now no longer available\n"); bbr_mod_inited = false; } tcp_lro_dereg_mbufq(); err = 0; break; default: return (EOPNOTSUPP); } return (err); } static moduledata_t tcp_bbr = { .name = __XSTRING(MODNAME), .evhand = tcp_addbbr, .priv = 0 }; MODULE_VERSION(MODNAME, 1); DECLARE_MODULE(MODNAME, tcp_bbr, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY); MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1); diff --git a/sys/netinet/tcp_stacks/rack.c b/sys/netinet/tcp_stacks/rack.c index 740ec73a17df..74503bc8a1b2 100644 --- a/sys/netinet/tcp_stacks/rack.c +++ b/sys/netinet/tcp_stacks/rack.c @@ -1,20955 +1,20955 @@ /*- * Copyright (c) 2016-2020 Netflix, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include "opt_ipsec.h" #include "opt_tcpdebug.h" #include "opt_ratelimit.h" #include "opt_kern_tls.h" #include #include #include #include #ifdef TCP_HHOOK #include #endif #include #include #include #include #include #include /* for proc0 declaration */ #include #include #include #include #ifdef STATS #include #include #include /* Must come after qmath.h and tree.h */ #else #include #endif #include #include #include #include #include #include #include #ifdef TCP_ACCOUNTING #include #include #endif #include #include #include #include #define TCPSTATES /* for logging */ #include #include #include #include #include /* required for icmp_var.h */ #include /* for ICMP_BANDLIM */ #include #include #include #include #include #define TCPOUTFLAGS #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef NETFLIX_SHARED_CWND #include #endif #ifdef TCPDEBUG #include #endif /* TCPDEBUG */ #ifdef TCP_OFFLOAD #include #endif #ifdef INET6 #include #endif #include #include #if defined(IPSEC) || defined(IPSEC_SUPPORT) #include #include #endif /* IPSEC */ #include #include #include #ifdef MAC #include #endif #include "sack_filter.h" #include "tcp_rack.h" #include "rack_bbr_common.h" uma_zone_t rack_zone; uma_zone_t rack_pcb_zone; #ifndef TICKS2SBT #define TICKS2SBT(__t) (tick_sbt * ((sbintime_t)(__t))) #endif VNET_DECLARE(uint32_t, newreno_beta); VNET_DECLARE(uint32_t, newreno_beta_ecn); #define V_newreno_beta VNET(newreno_beta) #define V_newreno_beta_ecn VNET(newreno_beta_ecn) MALLOC_DEFINE(M_TCPFSB, "tcp_fsb", "TCP fast send block"); MALLOC_DEFINE(M_TCPDO, "tcp_do", "TCP deferred options"); struct sysctl_ctx_list rack_sysctl_ctx; struct sysctl_oid *rack_sysctl_root; #define CUM_ACKED 1 #define SACKED 2 /* * The RACK module incorporates a number of * TCP ideas that have been put out into the IETF * over the last few years: * - Matt Mathis's Rate Halving which slowly drops * the congestion window so that the ack clock can * be maintained during a recovery. * - Yuchung Cheng's RACK TCP (for which its named) that * will stop us using the number of dup acks and instead * use time as the gage of when we retransmit. * - Reorder Detection of RFC4737 and the Tail-Loss probe draft * of Dukkipati et.al. * RACK depends on SACK, so if an endpoint arrives that * cannot do SACK the state machine below will shuttle the * connection back to using the "default" TCP stack that is * in FreeBSD. * * To implement RACK the original TCP stack was first decomposed * into a functional state machine with individual states * for each of the possible TCP connection states. The do_segment * functions role in life is to mandate the connection supports SACK * initially and then assure that the RACK state matches the conenction * state before calling the states do_segment function. Each * state is simplified due to the fact that the original do_segment * has been decomposed and we *know* what state we are in (no * switches on the state) and all tests for SACK are gone. This * greatly simplifies what each state does. * * TCP output is also over-written with a new version since it * must maintain the new rack scoreboard. * */ static int32_t rack_tlp_thresh = 1; static int32_t rack_tlp_limit = 2; /* No more than 2 TLPs w-out new data */ static int32_t rack_tlp_use_greater = 1; static int32_t rack_reorder_thresh = 2; static int32_t rack_reorder_fade = 60000000; /* 0 - never fade, def 60,000,000 * - 60 seconds */ static uint8_t rack_req_measurements = 1; /* Attack threshold detections */ static uint32_t rack_highest_sack_thresh_seen = 0; static uint32_t rack_highest_move_thresh_seen = 0; static int32_t rack_enable_hw_pacing = 0; /* Due to CCSP keep it off by default */ static int32_t rack_hw_pace_extra_slots = 2; /* 2 extra MSS time betweens */ static int32_t rack_hw_rate_caps = 1; /* 1; */ static int32_t rack_hw_rate_min = 0; /* 1500000;*/ static int32_t rack_hw_rate_to_low = 0; /* 1200000; */ static int32_t rack_hw_up_only = 1; static int32_t rack_stats_gets_ms_rtt = 1; static int32_t rack_prr_addbackmax = 2; static int32_t rack_do_hystart = 0; static int32_t rack_apply_rtt_with_reduced_conf = 0; static int32_t rack_pkt_delay = 1000; static int32_t rack_send_a_lot_in_prr = 1; static int32_t rack_min_to = 1000; /* Number of microsecond min timeout */ static int32_t rack_verbose_logging = 0; static int32_t rack_ignore_data_after_close = 1; static int32_t rack_enable_shared_cwnd = 1; static int32_t rack_use_cmp_acks = 1; static int32_t rack_use_fsb = 1; static int32_t rack_use_rfo = 1; static int32_t rack_use_rsm_rfo = 1; static int32_t rack_max_abc_post_recovery = 2; static int32_t rack_client_low_buf = 0; static int32_t rack_dsack_std_based = 0x3; /* bit field bit 1 sets rc_rack_tmr_std_based and bit 2 sets rc_rack_use_dsack */ #ifdef TCP_ACCOUNTING static int32_t rack_tcp_accounting = 0; #endif static int32_t rack_limits_scwnd = 1; static int32_t rack_enable_mqueue_for_nonpaced = 0; static int32_t rack_disable_prr = 0; static int32_t use_rack_rr = 1; static int32_t rack_non_rxt_use_cr = 0; /* does a non-rxt in recovery use the configured rate (ss/ca)? */ static int32_t rack_persist_min = 250000; /* 250usec */ static int32_t rack_persist_max = 2000000; /* 2 Second in usec's */ static int32_t rack_sack_not_required = 1; /* set to one to allow non-sack to use rack */ static int32_t rack_default_init_window = 0; /* Use system default */ static int32_t rack_limit_time_with_srtt = 0; static int32_t rack_autosndbuf_inc = 20; /* In percentage form */ static int32_t rack_enobuf_hw_boost_mult = 2; /* How many times the hw rate we boost slot using time_between */ static int32_t rack_enobuf_hw_max = 12000; /* 12 ms in usecs */ static int32_t rack_enobuf_hw_min = 10000; /* 10 ms in usecs */ static int32_t rack_hw_rwnd_factor = 2; /* How many max_segs the rwnd must be before we hold off sending */ /* * Currently regular tcp has a rto_min of 30ms * the backoff goes 12 times so that ends up * being a total of 122.850 seconds before a * connection is killed. */ static uint32_t rack_def_data_window = 20; static uint32_t rack_goal_bdp = 2; static uint32_t rack_min_srtts = 1; static uint32_t rack_min_measure_usec = 0; static int32_t rack_tlp_min = 10000; /* 10ms */ static int32_t rack_rto_min = 30000; /* 30,000 usec same as main freebsd */ static int32_t rack_rto_max = 4000000; /* 4 seconds in usec's */ static const int32_t rack_free_cache = 2; static int32_t rack_hptsi_segments = 40; static int32_t rack_rate_sample_method = USE_RTT_LOW; static int32_t rack_pace_every_seg = 0; static int32_t rack_delayed_ack_time = 40000; /* 40ms in usecs */ static int32_t rack_slot_reduction = 4; static int32_t rack_wma_divisor = 8; /* For WMA calculation */ static int32_t rack_cwnd_block_ends_measure = 0; static int32_t rack_rwnd_block_ends_measure = 0; static int32_t rack_def_profile = 0; static int32_t rack_lower_cwnd_at_tlp = 0; static int32_t rack_limited_retran = 0; static int32_t rack_always_send_oldest = 0; static int32_t rack_tlp_threshold_use = TLP_USE_TWO_ONE; static uint16_t rack_per_of_gp_ss = 250; /* 250 % slow-start */ static uint16_t rack_per_of_gp_ca = 200; /* 200 % congestion-avoidance */ static uint16_t rack_per_of_gp_rec = 200; /* 200 % of bw */ /* Probertt */ static uint16_t rack_per_of_gp_probertt = 60; /* 60% of bw */ static uint16_t rack_per_of_gp_lowthresh = 40; /* 40% is bottom */ static uint16_t rack_per_of_gp_probertt_reduce = 10; /* 10% reduction */ static uint16_t rack_atexit_prtt_hbp = 130; /* Clamp to 130% on exit prtt if highly buffered path */ static uint16_t rack_atexit_prtt = 130; /* Clamp to 100% on exit prtt if non highly buffered path */ static uint32_t rack_max_drain_wait = 2; /* How man gp srtt's before we give up draining */ static uint32_t rack_must_drain = 1; /* How many GP srtt's we *must* wait */ static uint32_t rack_probertt_use_min_rtt_entry = 1; /* Use the min to calculate the goal else gp_srtt */ static uint32_t rack_probertt_use_min_rtt_exit = 0; static uint32_t rack_probe_rtt_sets_cwnd = 0; static uint32_t rack_probe_rtt_safety_val = 2000000; /* No more than 2 sec in probe-rtt */ static uint32_t rack_time_between_probertt = 9600000; /* 9.6 sec in usecs */ static uint32_t rack_probertt_gpsrtt_cnt_mul = 0; /* How many srtt periods does probe-rtt last top fraction */ static uint32_t rack_probertt_gpsrtt_cnt_div = 0; /* How many srtt periods does probe-rtt last bottom fraction */ static uint32_t rack_min_probertt_hold = 40000; /* Equal to delayed ack time */ static uint32_t rack_probertt_filter_life = 10000000; static uint32_t rack_probertt_lower_within = 10; static uint32_t rack_min_rtt_movement = 250000; /* Must move at least 250ms (in microseconds) to count as a lowering */ static int32_t rack_pace_one_seg = 0; /* Shall we pace for less than 1.4Meg 1MSS at a time */ static int32_t rack_probertt_clear_is = 1; static int32_t rack_max_drain_hbp = 1; /* Extra drain times gpsrtt for highly buffered paths */ static int32_t rack_hbp_thresh = 3; /* what is the divisor max_rtt/min_rtt to decided a hbp */ /* Part of pacing */ static int32_t rack_max_per_above = 30; /* When we go to increment stop if above 100+this% */ /* Timely information */ /* Combine these two gives the range of 'no change' to bw */ /* ie the up/down provide the upper and lower bound */ static int32_t rack_gp_per_bw_mul_up = 2; /* 2% */ static int32_t rack_gp_per_bw_mul_down = 4; /* 4% */ static int32_t rack_gp_rtt_maxmul = 3; /* 3 x maxmin */ static int32_t rack_gp_rtt_minmul = 1; /* minrtt + (minrtt/mindiv) is lower rtt */ static int32_t rack_gp_rtt_mindiv = 4; /* minrtt + (minrtt * minmul/mindiv) is lower rtt */ static int32_t rack_gp_decrease_per = 20; /* 20% decrease in multiplier */ static int32_t rack_gp_increase_per = 2; /* 2% increase in multiplier */ static int32_t rack_per_lower_bound = 50; /* Don't allow to drop below this multiplier */ static int32_t rack_per_upper_bound_ss = 0; /* Don't allow SS to grow above this */ static int32_t rack_per_upper_bound_ca = 0; /* Don't allow CA to grow above this */ static int32_t rack_do_dyn_mul = 0; /* Are the rack gp multipliers dynamic */ static int32_t rack_gp_no_rec_chg = 1; /* Prohibit recovery from reducing it's multiplier */ static int32_t rack_timely_dec_clear = 6; /* Do we clear decrement count at a value (6)? */ static int32_t rack_timely_max_push_rise = 3; /* One round of pushing */ static int32_t rack_timely_max_push_drop = 3; /* Three round of pushing */ static int32_t rack_timely_min_segs = 4; /* 4 segment minimum */ static int32_t rack_use_max_for_nobackoff = 0; static int32_t rack_timely_int_timely_only = 0; /* do interim timely's only use the timely algo (no b/w changes)? */ static int32_t rack_timely_no_stopping = 0; static int32_t rack_down_raise_thresh = 100; static int32_t rack_req_segs = 1; static uint64_t rack_bw_rate_cap = 0; static uint32_t rack_trace_point_config = 0; static uint32_t rack_trace_point_bb_mode = 4; static int32_t rack_trace_point_count = 0; /* Weird delayed ack mode */ static int32_t rack_use_imac_dack = 0; /* Rack specific counters */ counter_u64_t rack_saw_enobuf; counter_u64_t rack_saw_enobuf_hw; counter_u64_t rack_saw_enetunreach; counter_u64_t rack_persists_sends; counter_u64_t rack_persists_acks; counter_u64_t rack_persists_loss; counter_u64_t rack_persists_lost_ends; #ifdef INVARIANTS counter_u64_t rack_adjust_map_bw; #endif /* Tail loss probe counters */ counter_u64_t rack_tlp_tot; counter_u64_t rack_tlp_newdata; counter_u64_t rack_tlp_retran; counter_u64_t rack_tlp_retran_bytes; counter_u64_t rack_to_tot; counter_u64_t rack_hot_alloc; counter_u64_t rack_to_alloc; counter_u64_t rack_to_alloc_hard; counter_u64_t rack_to_alloc_emerg; counter_u64_t rack_to_alloc_limited; counter_u64_t rack_alloc_limited_conns; counter_u64_t rack_split_limited; counter_u64_t rack_multi_single_eq; counter_u64_t rack_proc_non_comp_ack; counter_u64_t rack_fto_send; counter_u64_t rack_fto_rsm_send; counter_u64_t rack_nfto_resend; counter_u64_t rack_non_fto_send; counter_u64_t rack_extended_rfo; counter_u64_t rack_sack_proc_all; counter_u64_t rack_sack_proc_short; counter_u64_t rack_sack_proc_restart; counter_u64_t rack_sack_attacks_detected; counter_u64_t rack_sack_attacks_reversed; counter_u64_t rack_sack_used_next_merge; counter_u64_t rack_sack_splits; counter_u64_t rack_sack_used_prev_merge; counter_u64_t rack_sack_skipped_acked; counter_u64_t rack_ack_total; counter_u64_t rack_express_sack; counter_u64_t rack_sack_total; counter_u64_t rack_move_none; counter_u64_t rack_move_some; counter_u64_t rack_input_idle_reduces; counter_u64_t rack_collapsed_win; counter_u64_t rack_collapsed_win_seen; counter_u64_t rack_collapsed_win_rxt; counter_u64_t rack_collapsed_win_rxt_bytes; counter_u64_t rack_try_scwnd; counter_u64_t rack_hw_pace_init_fail; counter_u64_t rack_hw_pace_lost; counter_u64_t rack_out_size[TCP_MSS_ACCT_SIZE]; counter_u64_t rack_opts_arry[RACK_OPTS_SIZE]; #define RACK_REXMTVAL(tp) max(rack_rto_min, ((tp)->t_srtt + ((tp)->t_rttvar << 2))) #define RACK_TCPT_RANGESET(tv, value, tvmin, tvmax, slop) do { \ (tv) = (value) + slop; \ if ((u_long)(tv) < (u_long)(tvmin)) \ (tv) = (tvmin); \ if ((u_long)(tv) > (u_long)(tvmax)) \ (tv) = (tvmax); \ } while (0) static void rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick, int event, int line); static int rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t * ret_val); static int rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt); static void rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, uint32_t th_ack, uint16_t nsegs, uint16_t type, int32_t recovery); static struct rack_sendmap *rack_alloc(struct tcp_rack *rack); static struct rack_sendmap *rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type); static struct rack_sendmap * rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused); static void rack_cong_signal(struct tcpcb *tp, uint32_t type, uint32_t ack, int ); static void rack_counter_destroy(void); static int rack_ctloutput(struct inpcb *inp, struct sockopt *sopt); static int32_t rack_ctor(void *mem, int32_t size, void *arg, int32_t how); static void rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override); static void rack_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos); static void rack_dtor(void *mem, int32_t size, void *arg); static void rack_log_alt_to_to_cancel(struct tcp_rack *rack, uint32_t flex1, uint32_t flex2, uint32_t flex3, uint32_t flex4, uint32_t flex5, uint32_t flex6, uint16_t flex7, uint8_t mod); static void rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint64_t bw_est, uint64_t bw, uint64_t len_time, int method, int line, struct rack_sendmap *rsm, uint8_t quality); static struct rack_sendmap * rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm); static struct rack_sendmap *rack_find_lowest_rsm(struct tcp_rack *rack); static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm); static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged); static int rack_get_sockopt(struct inpcb *inp, struct sockopt *sopt); static void rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack, int line, uint8_t quality); static uint32_t rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss); static int32_t rack_handoff_ok(struct tcpcb *tp); static int32_t rack_init(struct tcpcb *tp); static void rack_init_sysctls(void); static void rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, int entered_rec, int dup_ack_struck); static void rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len, uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t ts, struct rack_sendmap *hintrsm, uint16_t add_flags, struct mbuf *s_mb, uint32_t s_moff, int hw_tls); static void rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm); static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm); static int32_t rack_output(struct tcpcb *tp); static uint32_t rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts, int *moved_two); static void rack_post_recovery(struct tcpcb *tp, uint32_t th_seq); static void rack_remxt_tmr(struct tcpcb *tp); static int rack_set_sockopt(struct inpcb *inp, struct sockopt *sopt); static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack); static int32_t rack_stopall(struct tcpcb *tp); static void rack_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta); static int32_t rack_timer_active(struct tcpcb *tp, uint32_t timer_type); static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line); static void rack_timer_stop(struct tcpcb *tp, uint32_t timer_type); static uint32_t rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm, uint64_t ts, int32_t * lenp, uint16_t add_flag); static void rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag); static int rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack); static int32_t tcp_addrack(module_t mod, int32_t type, void *data); static int rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos); static int rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos); static int rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos); static int rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos); static int rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos); static int rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos); static int rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos); static int rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos); static int rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos); struct rack_sendmap * tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused); static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_tim, int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt); static void tcp_rack_partialack(struct tcpcb *tp); static int rack_set_profile(struct tcp_rack *rack, int prof); static void rack_apply_deferred_options(struct tcp_rack *rack); int32_t rack_clear_counter=0; static inline void rack_trace_point(struct tcp_rack *rack, int num) { if (((rack_trace_point_config == num) || (rack_trace_point_config = 0xffffffff)) && (rack_trace_point_bb_mode != 0) && (rack_trace_point_count > 0) && (rack->rc_tp->t_logstate == 0)) { int res; res = atomic_fetchadd_int(&rack_trace_point_count, -1); if (res > 0) { rack->rc_tp->t_logstate = rack_trace_point_bb_mode; } else { /* Loss a race assure its zero now */ rack_trace_point_count = 0; } } } static void rack_set_cc_pacing(struct tcp_rack *rack) { struct sockopt sopt; struct cc_newreno_opts opt; struct newreno old, *ptr; struct tcpcb *tp; int error; if (rack->rc_pacing_cc_set) return; tp = rack->rc_tp; if (tp->cc_algo == NULL) { /* Tcb is leaving */ return; } rack->rc_pacing_cc_set = 1; if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) { /* Not new-reno we can't play games with beta! */ goto out; } ptr = ((struct newreno *)tp->ccv->cc_data); if (CC_ALGO(tp)->ctl_output == NULL) { /* Huh, why does new_reno no longer have a set function? */ goto out; } if (ptr == NULL) { /* Just the default values */ old.beta = V_newreno_beta_ecn; old.beta_ecn = V_newreno_beta_ecn; old.newreno_flags = 0; } else { old.beta = ptr->beta; old.beta_ecn = ptr->beta_ecn; old.newreno_flags = ptr->newreno_flags; } sopt.sopt_valsize = sizeof(struct cc_newreno_opts); sopt.sopt_dir = SOPT_SET; opt.name = CC_NEWRENO_BETA; opt.val = rack->r_ctl.rc_saved_beta.beta; error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt); if (error) { goto out; } /* * Hack alert we need to set in our newreno_flags * so that Abe behavior is also applied. */ ((struct newreno *)tp->ccv->cc_data)->newreno_flags |= CC_NEWRENO_BETA_ECN_ENABLED; opt.name = CC_NEWRENO_BETA_ECN; opt.val = rack->r_ctl.rc_saved_beta.beta_ecn; error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt); if (error) { goto out; } /* Save off the original values for restoral */ memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno)); out: if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; struct timeval tv; ptr = ((struct newreno *)tp->ccv->cc_data); memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.timeStamp = tcp_get_usecs(&tv); if (ptr) { log.u_bbr.flex1 = ptr->beta; log.u_bbr.flex2 = ptr->beta_ecn; log.u_bbr.flex3 = ptr->newreno_flags; } log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta; log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn; log.u_bbr.flex6 = rack->r_ctl.rc_saved_beta.newreno_flags; log.u_bbr.flex7 = rack->gp_ready; log.u_bbr.flex7 <<= 1; log.u_bbr.flex7 |= rack->use_fixed_rate; log.u_bbr.flex7 <<= 1; log.u_bbr.flex7 |= rack->rc_pacing_cc_set; log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.flex8 = 3; tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, error, 0, &log, false, NULL, NULL, 0, &tv); } } static void rack_undo_cc_pacing(struct tcp_rack *rack) { struct newreno old, *ptr; struct tcpcb *tp; if (rack->rc_pacing_cc_set == 0) return; tp = rack->rc_tp; rack->rc_pacing_cc_set = 0; if (tp->cc_algo == NULL) /* Tcb is leaving */ return; if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) { /* Not new-reno nothing to do! */ return; } ptr = ((struct newreno *)tp->ccv->cc_data); if (ptr == NULL) { /* * This happens at rack_fini() if the * cc module gets freed on us. In that * case we loose our "new" settings but * thats ok, since the tcb is going away anyway. */ return; } /* Grab out our set values */ memcpy(&old, ptr, sizeof(struct newreno)); /* Copy back in the original values */ memcpy(ptr, &rack->r_ctl.rc_saved_beta, sizeof(struct newreno)); /* Now save back the values we had set in (for when pacing is restored) */ memcpy(&rack->r_ctl.rc_saved_beta, &old, sizeof(struct newreno)); if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; struct timeval tv; ptr = ((struct newreno *)tp->ccv->cc_data); memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.flex1 = ptr->beta; log.u_bbr.flex2 = ptr->beta_ecn; log.u_bbr.flex3 = ptr->newreno_flags; log.u_bbr.flex4 = rack->r_ctl.rc_saved_beta.beta; log.u_bbr.flex5 = rack->r_ctl.rc_saved_beta.beta_ecn; log.u_bbr.flex6 = rack->r_ctl.rc_saved_beta.newreno_flags; log.u_bbr.flex7 = rack->gp_ready; log.u_bbr.flex7 <<= 1; log.u_bbr.flex7 |= rack->use_fixed_rate; log.u_bbr.flex7 <<= 1; log.u_bbr.flex7 |= rack->rc_pacing_cc_set; log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.flex8 = 4; tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0, 0, &log, false, NULL, NULL, 0, &tv); } } #ifdef NETFLIX_PEAKRATE static inline void rack_update_peakrate_thr(struct tcpcb *tp) { /* Keep in mind that t_maxpeakrate is in B/s. */ uint64_t peak; peak = uqmax((tp->t_maxseg * 2), (((uint64_t)tp->t_maxpeakrate * (uint64_t)(tp->t_srtt)) / (uint64_t)HPTS_USEC_IN_SEC)); tp->t_peakrate_thr = (uint32_t)uqmin(peak, UINT32_MAX); } #endif static int sysctl_rack_clear(SYSCTL_HANDLER_ARGS) { uint32_t stat; int32_t error; error = SYSCTL_OUT(req, &rack_clear_counter, sizeof(uint32_t)); if (error || req->newptr == NULL) return error; error = SYSCTL_IN(req, &stat, sizeof(uint32_t)); if (error) return (error); if (stat == 1) { #ifdef INVARIANTS printf("Clearing RACK counters\n"); #endif counter_u64_zero(rack_tlp_tot); counter_u64_zero(rack_tlp_newdata); counter_u64_zero(rack_tlp_retran); counter_u64_zero(rack_tlp_retran_bytes); counter_u64_zero(rack_to_tot); counter_u64_zero(rack_saw_enobuf); counter_u64_zero(rack_saw_enobuf_hw); counter_u64_zero(rack_saw_enetunreach); counter_u64_zero(rack_persists_sends); counter_u64_zero(rack_persists_acks); counter_u64_zero(rack_persists_loss); counter_u64_zero(rack_persists_lost_ends); #ifdef INVARIANTS counter_u64_zero(rack_adjust_map_bw); #endif counter_u64_zero(rack_to_alloc_hard); counter_u64_zero(rack_to_alloc_emerg); counter_u64_zero(rack_sack_proc_all); counter_u64_zero(rack_fto_send); counter_u64_zero(rack_fto_rsm_send); counter_u64_zero(rack_extended_rfo); counter_u64_zero(rack_hw_pace_init_fail); counter_u64_zero(rack_hw_pace_lost); counter_u64_zero(rack_non_fto_send); counter_u64_zero(rack_nfto_resend); counter_u64_zero(rack_sack_proc_short); counter_u64_zero(rack_sack_proc_restart); counter_u64_zero(rack_to_alloc); counter_u64_zero(rack_to_alloc_limited); counter_u64_zero(rack_alloc_limited_conns); counter_u64_zero(rack_split_limited); counter_u64_zero(rack_multi_single_eq); counter_u64_zero(rack_proc_non_comp_ack); counter_u64_zero(rack_sack_attacks_detected); counter_u64_zero(rack_sack_attacks_reversed); counter_u64_zero(rack_sack_used_next_merge); counter_u64_zero(rack_sack_used_prev_merge); counter_u64_zero(rack_sack_splits); counter_u64_zero(rack_sack_skipped_acked); counter_u64_zero(rack_ack_total); counter_u64_zero(rack_express_sack); counter_u64_zero(rack_sack_total); counter_u64_zero(rack_move_none); counter_u64_zero(rack_move_some); counter_u64_zero(rack_try_scwnd); counter_u64_zero(rack_collapsed_win); counter_u64_zero(rack_collapsed_win_rxt); counter_u64_zero(rack_collapsed_win_seen); counter_u64_zero(rack_collapsed_win_rxt_bytes); } rack_clear_counter = 0; return (0); } static void rack_init_sysctls(void) { struct sysctl_oid *rack_counters; struct sysctl_oid *rack_attack; struct sysctl_oid *rack_pacing; struct sysctl_oid *rack_timely; struct sysctl_oid *rack_timers; struct sysctl_oid *rack_tlp; struct sysctl_oid *rack_misc; struct sysctl_oid *rack_features; struct sysctl_oid *rack_measure; struct sysctl_oid *rack_probertt; struct sysctl_oid *rack_hw_pacing; struct sysctl_oid *rack_tracepoint; rack_attack = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "sack_attack", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Rack Sack Attack Counters and Controls"); rack_counters = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "stats", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Rack Counters"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "rate_sample_method", CTLFLAG_RW, &rack_rate_sample_method , USE_RTT_LOW, "What method should we use for rate sampling 0=high, 1=low "); /* Probe rtt related controls */ rack_probertt = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "probertt", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "ProbeRTT related Controls"); SYSCTL_ADD_U16(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "exit_per_hpb", CTLFLAG_RW, &rack_atexit_prtt_hbp, 130, "What percentage above goodput do we clamp CA/SS to at exit on high-BDP path 110%"); SYSCTL_ADD_U16(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "exit_per_nonhpb", CTLFLAG_RW, &rack_atexit_prtt, 130, "What percentage above goodput do we clamp CA/SS to at exit on a non high-BDP path 100%"); SYSCTL_ADD_U16(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "gp_per_mul", CTLFLAG_RW, &rack_per_of_gp_probertt, 60, "What percentage of goodput do we pace at in probertt"); SYSCTL_ADD_U16(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "gp_per_reduce", CTLFLAG_RW, &rack_per_of_gp_probertt_reduce, 10, "What percentage of goodput do we reduce every gp_srtt"); SYSCTL_ADD_U16(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "gp_per_low", CTLFLAG_RW, &rack_per_of_gp_lowthresh, 40, "What percentage of goodput do we allow the multiplier to fall to"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "time_between", CTLFLAG_RW, & rack_time_between_probertt, 96000000, "How many useconds between the lowest rtt falling must past before we enter probertt"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "safety", CTLFLAG_RW, &rack_probe_rtt_safety_val, 2000000, "If not zero, provides a maximum usecond that you can stay in probertt (2sec = 2000000)"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "sets_cwnd", CTLFLAG_RW, &rack_probe_rtt_sets_cwnd, 0, "Do we set the cwnd too (if always_lower is on)"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "maxdrainsrtts", CTLFLAG_RW, &rack_max_drain_wait, 2, "Maximum number of gp_srtt's to hold in drain waiting for flight to reach goal"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "mustdrainsrtts", CTLFLAG_RW, &rack_must_drain, 1, "We must drain this many gp_srtt's waiting for flight to reach goal"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "goal_use_min_entry", CTLFLAG_RW, &rack_probertt_use_min_rtt_entry, 1, "Should we use the min-rtt to calculate the goal rtt (else gp_srtt) at entry"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "goal_use_min_exit", CTLFLAG_RW, &rack_probertt_use_min_rtt_exit, 0, "How to set cwnd at exit, 0 - dynamic, 1 - use min-rtt, 2 - use curgprtt, 3 - entry gp-rtt"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "length_div", CTLFLAG_RW, &rack_probertt_gpsrtt_cnt_div, 0, "How many recent goodput srtt periods plus hold tim does probertt last (bottom of fraction)"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "length_mul", CTLFLAG_RW, &rack_probertt_gpsrtt_cnt_mul, 0, "How many recent goodput srtt periods plus hold tim does probertt last (top of fraction)"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "holdtim_at_target", CTLFLAG_RW, &rack_min_probertt_hold, 200000, "What is the minimum time we hold probertt at target"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "filter_life", CTLFLAG_RW, &rack_probertt_filter_life, 10000000, "What is the time for the filters life in useconds"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "lower_within", CTLFLAG_RW, &rack_probertt_lower_within, 10, "If the rtt goes lower within this percentage of the time, go into probe-rtt"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "must_move", CTLFLAG_RW, &rack_min_rtt_movement, 250, "How much is the minimum movement in rtt to count as a drop for probertt purposes"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "clear_is_cnts", CTLFLAG_RW, &rack_probertt_clear_is, 1, "Do we clear I/S counts on exiting probe-rtt"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "hbp_extra_drain", CTLFLAG_RW, &rack_max_drain_hbp, 1, "How many extra drain gpsrtt's do we get in highly buffered paths"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_probertt), OID_AUTO, "hbp_threshold", CTLFLAG_RW, &rack_hbp_thresh, 3, "We are highly buffered if min_rtt_seen / max_rtt_seen > this-threshold"); rack_tracepoint = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "tp", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Rack tracepoint facility"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tracepoint), OID_AUTO, "number", CTLFLAG_RW, &rack_trace_point_config, 0, "What is the trace point number to activate (0=none, 0xffffffff = all)?"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tracepoint), OID_AUTO, "bbmode", CTLFLAG_RW, &rack_trace_point_bb_mode, 4, "What is BB logging mode that is activated?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tracepoint), OID_AUTO, "count", CTLFLAG_RW, &rack_trace_point_count, 0, "How many connections will have BB logging turned on that hit the tracepoint?"); /* Pacing related sysctls */ rack_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "pacing", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Pacing related Controls"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_pacing), OID_AUTO, "max_pace_over", CTLFLAG_RW, &rack_max_per_above, 30, "What is the maximum allowable percentage that we can pace above (so 30 = 130% of our goal)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_pacing), OID_AUTO, "pace_to_one", CTLFLAG_RW, &rack_pace_one_seg, 0, "Do we allow low b/w pacing of 1MSS instead of two"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_pacing), OID_AUTO, "limit_wsrtt", CTLFLAG_RW, &rack_limit_time_with_srtt, 0, "Do we limit pacing time based on srtt"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_pacing), OID_AUTO, "init_win", CTLFLAG_RW, &rack_default_init_window, 0, "Do we have a rack initial window 0 = system default"); SYSCTL_ADD_U16(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_pacing), OID_AUTO, "gp_per_ss", CTLFLAG_RW, &rack_per_of_gp_ss, 250, "If non zero, what percentage of goodput to pace at in slow start"); SYSCTL_ADD_U16(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_pacing), OID_AUTO, "gp_per_ca", CTLFLAG_RW, &rack_per_of_gp_ca, 150, "If non zero, what percentage of goodput to pace at in congestion avoidance"); SYSCTL_ADD_U16(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_pacing), OID_AUTO, "gp_per_rec", CTLFLAG_RW, &rack_per_of_gp_rec, 200, "If non zero, what percentage of goodput to pace at in recovery"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_pacing), OID_AUTO, "pace_max_seg", CTLFLAG_RW, &rack_hptsi_segments, 40, "What size is the max for TSO segments in pacing and burst mitigation"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_pacing), OID_AUTO, "burst_reduces", CTLFLAG_RW, &rack_slot_reduction, 4, "When doing only burst mitigation what is the reduce divisor"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "use_pacing", CTLFLAG_RW, &rack_pace_every_seg, 0, "If set we use pacing, if clear we use only the original burst mitigation"); SYSCTL_ADD_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_pacing), OID_AUTO, "rate_cap", CTLFLAG_RW, &rack_bw_rate_cap, 0, "If set we apply this value to the absolute rate cap used by pacing"); SYSCTL_ADD_U8(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "req_measure_cnt", CTLFLAG_RW, &rack_req_measurements, 1, "If doing dynamic pacing, how many measurements must be in before we start pacing?"); /* Hardware pacing */ rack_hw_pacing = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "hdwr_pacing", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Pacing related Controls"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_hw_pacing), OID_AUTO, "rwnd_factor", CTLFLAG_RW, &rack_hw_rwnd_factor, 2, "How many times does snd_wnd need to be bigger than pace_max_seg so we will hold off and get more acks?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_hw_pacing), OID_AUTO, "pace_enobuf_mult", CTLFLAG_RW, &rack_enobuf_hw_boost_mult, 2, "By how many time_betweens should we boost the pacing time if we see a ENOBUFS?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_hw_pacing), OID_AUTO, "pace_enobuf_max", CTLFLAG_RW, &rack_enobuf_hw_max, 2, "What is the max boost the pacing time if we see a ENOBUFS?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_hw_pacing), OID_AUTO, "pace_enobuf_min", CTLFLAG_RW, &rack_enobuf_hw_min, 2, "What is the min boost the pacing time if we see a ENOBUFS?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_hw_pacing), OID_AUTO, "enable", CTLFLAG_RW, &rack_enable_hw_pacing, 0, "Should RACK attempt to use hw pacing?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_hw_pacing), OID_AUTO, "rate_cap", CTLFLAG_RW, &rack_hw_rate_caps, 1, "Does the highest hardware pacing rate cap the rate we will send at??"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_hw_pacing), OID_AUTO, "rate_min", CTLFLAG_RW, &rack_hw_rate_min, 0, "Do we need a minimum estimate of this many bytes per second in order to engage hw pacing?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_hw_pacing), OID_AUTO, "rate_to_low", CTLFLAG_RW, &rack_hw_rate_to_low, 0, "If we fall below this rate, dis-engage hw pacing?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_hw_pacing), OID_AUTO, "up_only", CTLFLAG_RW, &rack_hw_up_only, 1, "Do we allow hw pacing to lower the rate selected?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_hw_pacing), OID_AUTO, "extra_mss_precise", CTLFLAG_RW, &rack_hw_pace_extra_slots, 2, "If the rates between software and hardware match precisely how many extra time_betweens do we get?"); rack_timely = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "timely", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Rack Timely RTT Controls"); /* Timely based GP dynmics */ SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "upper", CTLFLAG_RW, &rack_gp_per_bw_mul_up, 2, "Rack timely upper range for equal b/w (in percentage)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "lower", CTLFLAG_RW, &rack_gp_per_bw_mul_down, 4, "Rack timely lower range for equal b/w (in percentage)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "rtt_max_mul", CTLFLAG_RW, &rack_gp_rtt_maxmul, 3, "Rack timely multiplier of lowest rtt for rtt_max"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "rtt_min_div", CTLFLAG_RW, &rack_gp_rtt_mindiv, 4, "Rack timely divisor used for rtt + (rtt * mul/divisor) for check for lower rtt"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "rtt_min_mul", CTLFLAG_RW, &rack_gp_rtt_minmul, 1, "Rack timely multiplier used for rtt + (rtt * mul/divisor) for check for lower rtt"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "decrease", CTLFLAG_RW, &rack_gp_decrease_per, 20, "Rack timely decrease percentage of our GP multiplication factor"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "increase", CTLFLAG_RW, &rack_gp_increase_per, 2, "Rack timely increase perentage of our GP multiplication factor"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "lowerbound", CTLFLAG_RW, &rack_per_lower_bound, 50, "Rack timely lowest percentage we allow GP multiplier to fall to"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "upperboundss", CTLFLAG_RW, &rack_per_upper_bound_ss, 0, "Rack timely highest percentage we allow GP multiplier in SS to raise to (0 is no upperbound)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "upperboundca", CTLFLAG_RW, &rack_per_upper_bound_ca, 0, "Rack timely highest percentage we allow GP multiplier to CA raise to (0 is no upperbound)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "dynamicgp", CTLFLAG_RW, &rack_do_dyn_mul, 0, "Rack timely do we enable dynmaic timely goodput by default"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "no_rec_red", CTLFLAG_RW, &rack_gp_no_rec_chg, 1, "Rack timely do we prohibit the recovery multiplier from being lowered"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "red_clear_cnt", CTLFLAG_RW, &rack_timely_dec_clear, 6, "Rack timely what threshold do we count to before another boost during b/w decent"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "max_push_rise", CTLFLAG_RW, &rack_timely_max_push_rise, 3, "Rack timely how many times do we push up with b/w increase"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "max_push_drop", CTLFLAG_RW, &rack_timely_max_push_drop, 3, "Rack timely how many times do we push back on b/w decent"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "min_segs", CTLFLAG_RW, &rack_timely_min_segs, 4, "Rack timely when setting the cwnd what is the min num segments"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "noback_max", CTLFLAG_RW, &rack_use_max_for_nobackoff, 0, "Rack timely when deciding if to backoff on a loss, do we use under max rtt else min"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "interim_timely_only", CTLFLAG_RW, &rack_timely_int_timely_only, 0, "Rack timely when doing interim timely's do we only do timely (no b/w consideration)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "nonstop", CTLFLAG_RW, &rack_timely_no_stopping, 0, "Rack timely don't stop increase"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "dec_raise_thresh", CTLFLAG_RW, &rack_down_raise_thresh, 100, "If the CA or SS is below this threshold raise on the first 3 b/w lowers (0=always)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timely), OID_AUTO, "bottom_drag_segs", CTLFLAG_RW, &rack_req_segs, 1, "Bottom dragging if not these many segments outstanding and room"); /* TLP and Rack related parameters */ rack_tlp = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "tlp", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "TLP and Rack related Controls"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "use_rrr", CTLFLAG_RW, &use_rack_rr, 1, "Do we use Rack Rapid Recovery"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "post_rec_labc", CTLFLAG_RW, &rack_max_abc_post_recovery, 2, "Since we do early recovery, do we override the l_abc to a value, if so what?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "nonrxt_use_cr", CTLFLAG_RW, &rack_non_rxt_use_cr, 0, "Do we use ss/ca rate if in recovery we are transmitting a new data chunk"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "tlpmethod", CTLFLAG_RW, &rack_tlp_threshold_use, TLP_USE_TWO_ONE, "What method do we do for TLP time calc 0=no-de-ack-comp, 1=ID, 2=2.1, 3=2.2"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "limit", CTLFLAG_RW, &rack_tlp_limit, 2, "How many TLP's can be sent without sending new data"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "use_greater", CTLFLAG_RW, &rack_tlp_use_greater, 1, "Should we use the rack_rtt time if its greater than srtt"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "tlpminto", CTLFLAG_RW, &rack_tlp_min, 10000, "TLP minimum timeout per the specification (in microseconds)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "send_oldest", CTLFLAG_RW, &rack_always_send_oldest, 0, "Should we always send the oldest TLP and RACK-TLP"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "rack_tlimit", CTLFLAG_RW, &rack_limited_retran, 0, "How many times can a rack timeout drive out sends"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "tlp_cwnd_flag", CTLFLAG_RW, &rack_lower_cwnd_at_tlp, 0, "When a TLP completes a retran should we enter recovery"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "reorder_thresh", CTLFLAG_RW, &rack_reorder_thresh, 2, "What factor for rack will be added when seeing reordering (shift right)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "rtt_tlp_thresh", CTLFLAG_RW, &rack_tlp_thresh, 1, "What divisor for TLP rtt/retran will be added (1=rtt, 2=1/2 rtt etc)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "reorder_fade", CTLFLAG_RW, &rack_reorder_fade, 60000000, "Does reorder detection fade, if so how many microseconds (0 means never)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_tlp), OID_AUTO, "pktdelay", CTLFLAG_RW, &rack_pkt_delay, 1000, "Extra RACK time (in microseconds) besides reordering thresh"); /* Timer related controls */ rack_timers = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "timers", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Timer related controls"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timers), OID_AUTO, "persmin", CTLFLAG_RW, &rack_persist_min, 250000, "What is the minimum time in microseconds between persists"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timers), OID_AUTO, "persmax", CTLFLAG_RW, &rack_persist_max, 2000000, "What is the largest delay in microseconds between persists"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timers), OID_AUTO, "delayed_ack", CTLFLAG_RW, &rack_delayed_ack_time, 40000, "Delayed ack time (40ms in microseconds)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timers), OID_AUTO, "minrto", CTLFLAG_RW, &rack_rto_min, 30000, "Minimum RTO in microseconds -- set with caution below 1000 due to TLP"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timers), OID_AUTO, "maxrto", CTLFLAG_RW, &rack_rto_max, 4000000, "Maximum RTO in microseconds -- should be at least as large as min_rto"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_timers), OID_AUTO, "minto", CTLFLAG_RW, &rack_min_to, 1000, "Minimum rack timeout in microseconds"); /* Measure controls */ rack_measure = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "measure", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Measure related controls"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_measure), OID_AUTO, "wma_divisor", CTLFLAG_RW, &rack_wma_divisor, 8, "When doing b/w calculation what is the divisor for the WMA"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_measure), OID_AUTO, "end_cwnd", CTLFLAG_RW, &rack_cwnd_block_ends_measure, 0, "Does a cwnd just-return end the measurement window (app limited)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_measure), OID_AUTO, "end_rwnd", CTLFLAG_RW, &rack_rwnd_block_ends_measure, 0, "Does an rwnd just-return end the measurement window (app limited -- not persists)"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_measure), OID_AUTO, "min_target", CTLFLAG_RW, &rack_def_data_window, 20, "What is the minimum target window (in mss) for a GP measurements"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_measure), OID_AUTO, "goal_bdp", CTLFLAG_RW, &rack_goal_bdp, 2, "What is the goal BDP to measure"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_measure), OID_AUTO, "min_srtts", CTLFLAG_RW, &rack_min_srtts, 1, "What is the goal BDP to measure"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_measure), OID_AUTO, "min_measure_tim", CTLFLAG_RW, &rack_min_measure_usec, 0, "What is the Minimum time time for a measurement if 0, this is off"); /* Features */ rack_features = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "features", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Feature controls"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_features), OID_AUTO, "cmpack", CTLFLAG_RW, &rack_use_cmp_acks, 1, "Should RACK have LRO send compressed acks"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_features), OID_AUTO, "fsb", CTLFLAG_RW, &rack_use_fsb, 1, "Should RACK use the fast send block?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_features), OID_AUTO, "rfo", CTLFLAG_RW, &rack_use_rfo, 1, "Should RACK use rack_fast_output()?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_features), OID_AUTO, "rsmrfo", CTLFLAG_RW, &rack_use_rsm_rfo, 1, "Should RACK use rack_fast_rsm_output()?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_features), OID_AUTO, "non_paced_lro_queue", CTLFLAG_RW, &rack_enable_mqueue_for_nonpaced, 0, "Should RACK use mbuf queuing for non-paced connections"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_features), OID_AUTO, "hystartplusplus", CTLFLAG_RW, &rack_do_hystart, 0, "Should RACK enable HyStart++ on connections?"); /* Misc rack controls */ rack_misc = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "misc", CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Misc related controls"); #ifdef TCP_ACCOUNTING SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "tcp_acct", CTLFLAG_RW, &rack_tcp_accounting, 0, "Should we turn on TCP accounting for all rack sessions?"); #endif SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "apply_rtt_with_low_conf", CTLFLAG_RW, &rack_apply_rtt_with_reduced_conf, 0, "When a persist or keep-alive probe is not answered do we calculate rtt on subsequent answers?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "rack_dsack_ctl", CTLFLAG_RW, &rack_dsack_std_based, 3, "How do we process dsack with respect to rack timers, bit field, 3 is standards based?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "prr_addback_max", CTLFLAG_RW, &rack_prr_addbackmax, 2, "What is the maximum number of MSS we allow to be added back if prr can't send all its data?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "stats_gets_ms", CTLFLAG_RW, &rack_stats_gets_ms_rtt, 1, "What do we feed the stats framework (1 = ms_rtt, 0 = us_rtt, 2 = ms_rtt from hdwr, > 2 usec rtt from hdwr)?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "clientlowbuf", CTLFLAG_RW, &rack_client_low_buf, 0, "Client low buffer level (below this we are more aggressive in DGP exiting recovery (0 = off)?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "defprofile", CTLFLAG_RW, &rack_def_profile, 0, "Should RACK use a default profile (0=no, num == profile num)?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "shared_cwnd", CTLFLAG_RW, &rack_enable_shared_cwnd, 1, "Should RACK try to use the shared cwnd on connections where allowed"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "limits_on_scwnd", CTLFLAG_RW, &rack_limits_scwnd, 1, "Should RACK place low end time limits on the shared cwnd feature"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "iMac_dack", CTLFLAG_RW, &rack_use_imac_dack, 0, "Should RACK try to emulate iMac delayed ack"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "no_prr", CTLFLAG_RW, &rack_disable_prr, 0, "Should RACK not use prr and only pace (must have pacing on)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "bb_verbose", CTLFLAG_RW, &rack_verbose_logging, 0, "Should RACK black box logging be verbose"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "data_after_close", CTLFLAG_RW, &rack_ignore_data_after_close, 1, "Do we hold off sending a RST until all pending data is ack'd"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "no_sack_needed", CTLFLAG_RW, &rack_sack_not_required, 1, "Do we allow rack to run on connections not supporting SACK"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "prr_sendalot", CTLFLAG_RW, &rack_send_a_lot_in_prr, 1, "Send a lot in prr"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_misc), OID_AUTO, "autoscale", CTLFLAG_RW, &rack_autosndbuf_inc, 20, "What percentage should rack scale up its snd buffer by?"); /* Sack Attacker detection stuff */ SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "detect_highsackratio", CTLFLAG_RW, &rack_highest_sack_thresh_seen, 0, "Highest sack to ack ratio seen"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "detect_highmoveratio", CTLFLAG_RW, &rack_highest_move_thresh_seen, 0, "Highest move to non-move ratio seen"); rack_ack_total = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "acktotal", CTLFLAG_RD, &rack_ack_total, "Total number of Ack's"); rack_express_sack = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "exp_sacktotal", CTLFLAG_RD, &rack_express_sack, "Total expresss number of Sack's"); rack_sack_total = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "sacktotal", CTLFLAG_RD, &rack_sack_total, "Total number of SACKs"); rack_move_none = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "move_none", CTLFLAG_RD, &rack_move_none, "Total number of SACK index reuse of positions under threshold"); rack_move_some = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "move_some", CTLFLAG_RD, &rack_move_some, "Total number of SACK index reuse of positions over threshold"); rack_sack_attacks_detected = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "attacks", CTLFLAG_RD, &rack_sack_attacks_detected, "Total number of SACK attackers that had sack disabled"); rack_sack_attacks_reversed = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "reversed", CTLFLAG_RD, &rack_sack_attacks_reversed, "Total number of SACK attackers that were later determined false positive"); rack_sack_used_next_merge = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "nextmerge", CTLFLAG_RD, &rack_sack_used_next_merge, "Total number of times we used the next merge"); rack_sack_used_prev_merge = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "prevmerge", CTLFLAG_RD, &rack_sack_used_prev_merge, "Total number of times we used the prev merge"); /* Counters */ rack_fto_send = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "fto_send", CTLFLAG_RD, &rack_fto_send, "Total number of rack_fast_output sends"); rack_fto_rsm_send = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "fto_rsm_send", CTLFLAG_RD, &rack_fto_rsm_send, "Total number of rack_fast_rsm_output sends"); rack_nfto_resend = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "nfto_resend", CTLFLAG_RD, &rack_nfto_resend, "Total number of rack_output retransmissions"); rack_non_fto_send = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "nfto_send", CTLFLAG_RD, &rack_non_fto_send, "Total number of rack_output first sends"); rack_extended_rfo = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "rfo_extended", CTLFLAG_RD, &rack_extended_rfo, "Total number of times we extended rfo"); rack_hw_pace_init_fail = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "hwpace_init_fail", CTLFLAG_RD, &rack_hw_pace_init_fail, "Total number of times we failed to initialize hw pacing"); rack_hw_pace_lost = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "hwpace_lost", CTLFLAG_RD, &rack_hw_pace_lost, "Total number of times we failed to initialize hw pacing"); rack_tlp_tot = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "tlp_to_total", CTLFLAG_RD, &rack_tlp_tot, "Total number of tail loss probe expirations"); rack_tlp_newdata = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "tlp_new", CTLFLAG_RD, &rack_tlp_newdata, "Total number of tail loss probe sending new data"); rack_tlp_retran = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "tlp_retran", CTLFLAG_RD, &rack_tlp_retran, "Total number of tail loss probe sending retransmitted data"); rack_tlp_retran_bytes = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "tlp_retran_bytes", CTLFLAG_RD, &rack_tlp_retran_bytes, "Total bytes of tail loss probe sending retransmitted data"); rack_to_tot = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "rack_to_tot", CTLFLAG_RD, &rack_to_tot, "Total number of times the rack to expired"); rack_saw_enobuf = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "saw_enobufs", CTLFLAG_RD, &rack_saw_enobuf, "Total number of times a sends returned enobuf for non-hdwr paced connections"); rack_saw_enobuf_hw = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "saw_enobufs_hw", CTLFLAG_RD, &rack_saw_enobuf_hw, "Total number of times a send returned enobuf for hdwr paced connections"); rack_saw_enetunreach = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "saw_enetunreach", CTLFLAG_RD, &rack_saw_enetunreach, "Total number of times a send received a enetunreachable"); rack_hot_alloc = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "alloc_hot", CTLFLAG_RD, &rack_hot_alloc, "Total allocations from the top of our list"); rack_to_alloc = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "allocs", CTLFLAG_RD, &rack_to_alloc, "Total allocations of tracking structures"); rack_to_alloc_hard = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "allochard", CTLFLAG_RD, &rack_to_alloc_hard, "Total allocations done with sleeping the hard way"); rack_to_alloc_emerg = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "allocemerg", CTLFLAG_RD, &rack_to_alloc_emerg, "Total allocations done from emergency cache"); rack_to_alloc_limited = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "alloc_limited", CTLFLAG_RD, &rack_to_alloc_limited, "Total allocations dropped due to limit"); rack_alloc_limited_conns = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "alloc_limited_conns", CTLFLAG_RD, &rack_alloc_limited_conns, "Connections with allocations dropped due to limit"); rack_split_limited = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "split_limited", CTLFLAG_RD, &rack_split_limited, "Split allocations dropped due to limit"); rack_persists_sends = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "persist_sends", CTLFLAG_RD, &rack_persists_sends, "Number of times we sent a persist probe"); rack_persists_acks = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "persist_acks", CTLFLAG_RD, &rack_persists_acks, "Number of times a persist probe was acked"); rack_persists_loss = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "persist_loss", CTLFLAG_RD, &rack_persists_loss, "Number of times we detected a lost persist probe (no ack)"); rack_persists_lost_ends = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "persist_loss_ends", CTLFLAG_RD, &rack_persists_lost_ends, "Number of lost persist probe (no ack) that the run ended with a PERSIST abort"); #ifdef INVARIANTS rack_adjust_map_bw = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "map_adjust_req", CTLFLAG_RD, &rack_adjust_map_bw, "Number of times we hit the case where the sb went up and down on a sendmap entry"); #endif rack_multi_single_eq = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "cmp_ack_equiv", CTLFLAG_RD, &rack_multi_single_eq, "Number of compressed acks total represented"); rack_proc_non_comp_ack = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "cmp_ack_not", CTLFLAG_RD, &rack_proc_non_comp_ack, "Number of non compresseds acks that we processed"); rack_sack_proc_all = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "sack_long", CTLFLAG_RD, &rack_sack_proc_all, "Total times we had to walk whole list for sack processing"); rack_sack_proc_restart = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "sack_restart", CTLFLAG_RD, &rack_sack_proc_restart, "Total times we had to walk whole list due to a restart"); rack_sack_proc_short = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "sack_short", CTLFLAG_RD, &rack_sack_proc_short, "Total times we took shortcut for sack processing"); rack_sack_skipped_acked = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "skipacked", CTLFLAG_RD, &rack_sack_skipped_acked, "Total number of times we skipped previously sacked"); rack_sack_splits = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_attack), OID_AUTO, "ofsplit", CTLFLAG_RD, &rack_sack_splits, "Total number of times we did the old fashion tree split"); rack_input_idle_reduces = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "idle_reduce_oninput", CTLFLAG_RD, &rack_input_idle_reduces, "Total number of idle reductions on input"); rack_collapsed_win_seen = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "collapsed_win_seen", CTLFLAG_RD, &rack_collapsed_win_seen, "Total number of collapsed window events seen (where our window shrinks)"); rack_collapsed_win = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "collapsed_win", CTLFLAG_RD, &rack_collapsed_win, "Total number of collapsed window events where we mark packets"); rack_collapsed_win_rxt = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "collapsed_win_rxt", CTLFLAG_RD, &rack_collapsed_win_rxt, "Total number of packets that were retransmitted"); rack_collapsed_win_rxt_bytes = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "collapsed_win_bytes", CTLFLAG_RD, &rack_collapsed_win_rxt_bytes, "Total number of bytes that were retransmitted"); rack_try_scwnd = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "tried_scwnd", CTLFLAG_RD, &rack_try_scwnd, "Total number of scwnd attempts"); COUNTER_ARRAY_ALLOC(rack_out_size, TCP_MSS_ACCT_SIZE, M_WAITOK); SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "outsize", CTLFLAG_RD, rack_out_size, TCP_MSS_ACCT_SIZE, "MSS send sizes"); COUNTER_ARRAY_ALLOC(rack_opts_arry, RACK_OPTS_SIZE, M_WAITOK); SYSCTL_ADD_COUNTER_U64_ARRAY(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "opts", CTLFLAG_RD, rack_opts_arry, RACK_OPTS_SIZE, "RACK Option Stats"); SYSCTL_ADD_PROC(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "clear", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE, &rack_clear_counter, 0, sysctl_rack_clear, "IU", "Clear counters"); } static __inline int rb_map_cmp(struct rack_sendmap *b, struct rack_sendmap *a) { if (SEQ_GEQ(b->r_start, a->r_start) && SEQ_LT(b->r_start, a->r_end)) { /* * The entry b is within the * block a. i.e.: * a -- |-------------| * b -- |----| * * b -- |------| * * b -- |-----------| */ return (0); } else if (SEQ_GEQ(b->r_start, a->r_end)) { /* * b falls as either the next * sequence block after a so a * is said to be smaller than b. * i.e: * a -- |------| * b -- |--------| * or * b -- |-----| */ return (1); } /* * Whats left is where a is * larger than b. i.e: * a -- |-------| * b -- |---| * or even possibly * b -- |--------------| */ return (-1); } RB_PROTOTYPE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp); RB_GENERATE(rack_rb_tree_head, rack_sendmap, r_next, rb_map_cmp); static uint32_t rc_init_window(struct tcp_rack *rack) { uint32_t win; if (rack->rc_init_win == 0) { /* * Nothing set by the user, use the system stack * default. */ return (tcp_compute_initwnd(tcp_maxseg(rack->rc_tp))); } win = ctf_fixed_maxseg(rack->rc_tp) * rack->rc_init_win; return (win); } static uint64_t rack_get_fixed_pacing_bw(struct tcp_rack *rack) { if (IN_FASTRECOVERY(rack->rc_tp->t_flags)) return (rack->r_ctl.rc_fixed_pacing_rate_rec); else if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) return (rack->r_ctl.rc_fixed_pacing_rate_ss); else return (rack->r_ctl.rc_fixed_pacing_rate_ca); } static uint64_t rack_get_bw(struct tcp_rack *rack) { if (rack->use_fixed_rate) { /* Return the fixed pacing rate */ return (rack_get_fixed_pacing_bw(rack)); } if (rack->r_ctl.gp_bw == 0) { /* * We have yet no b/w measurement, * if we have a user set initial bw * return it. If we don't have that and * we have an srtt, use the tcp IW (10) to * calculate a fictional b/w over the SRTT * which is more or less a guess. Note * we don't use our IW from rack on purpose * so if we have like IW=30, we are not * calculating a "huge" b/w. */ uint64_t bw, srtt; if (rack->r_ctl.init_rate) return (rack->r_ctl.init_rate); /* Has the user set a max peak rate? */ #ifdef NETFLIX_PEAKRATE if (rack->rc_tp->t_maxpeakrate) return (rack->rc_tp->t_maxpeakrate); #endif /* Ok lets come up with the IW guess, if we have a srtt */ if (rack->rc_tp->t_srtt == 0) { /* * Go with old pacing method * i.e. burst mitigation only. */ return (0); } /* Ok lets get the initial TCP win (not racks) */ bw = tcp_compute_initwnd(tcp_maxseg(rack->rc_tp)); srtt = (uint64_t)rack->rc_tp->t_srtt; bw *= (uint64_t)USECS_IN_SECOND; bw /= srtt; if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap)) bw = rack->r_ctl.bw_rate_cap; return (bw); } else { uint64_t bw; if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) { /* Averaging is done, we can return the value */ bw = rack->r_ctl.gp_bw; } else { /* Still doing initial average must calculate */ bw = rack->r_ctl.gp_bw / rack->r_ctl.num_measurements; } #ifdef NETFLIX_PEAKRATE if ((rack->rc_tp->t_maxpeakrate) && (bw > rack->rc_tp->t_maxpeakrate)) { /* The user has set a peak rate to pace at * don't allow us to pace faster than that. */ return (rack->rc_tp->t_maxpeakrate); } #endif if (rack->r_ctl.bw_rate_cap && (bw > rack->r_ctl.bw_rate_cap)) bw = rack->r_ctl.bw_rate_cap; return (bw); } } static uint16_t rack_get_output_gain(struct tcp_rack *rack, struct rack_sendmap *rsm) { if (rack->use_fixed_rate) { return (100); } else if (rack->in_probe_rtt && (rsm == NULL)) return (rack->r_ctl.rack_per_of_gp_probertt); else if ((IN_FASTRECOVERY(rack->rc_tp->t_flags) && rack->r_ctl.rack_per_of_gp_rec)) { if (rsm) { /* a retransmission always use the recovery rate */ return (rack->r_ctl.rack_per_of_gp_rec); } else if (rack->rack_rec_nonrxt_use_cr) { /* Directed to use the configured rate */ goto configured_rate; } else if (rack->rack_no_prr && (rack->r_ctl.rack_per_of_gp_rec > 100)) { /* No PRR, lets just use the b/w estimate only */ return (100); } else { /* * Here we may have a non-retransmit but we * have no overrides, so just use the recovery * rate (prr is in effect). */ return (rack->r_ctl.rack_per_of_gp_rec); } } configured_rate: /* For the configured rate we look at our cwnd vs the ssthresh */ if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) return (rack->r_ctl.rack_per_of_gp_ss); else return (rack->r_ctl.rack_per_of_gp_ca); } static void rack_log_dsack_event(struct tcp_rack *rack, uint8_t mod, uint32_t flex4, uint32_t flex5, uint32_t flex6) { /* * Types of logs (mod value) * 1 = dsack_persists reduced by 1 via T-O or fast recovery exit. * 2 = a dsack round begins, persist is reset to 16. * 3 = a dsack round ends * 4 = Dsack option increases rack rtt flex5 is the srtt input, flex6 is thresh * 5 = Socket option set changing the control flags rc_rack_tmr_std_based, rc_rack_use_dsack * 6 = Final rack rtt, flex4 is srtt and flex6 is final limited thresh. */ if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log, 0, sizeof(log)); log.u_bbr.flex1 = rack->rc_rack_tmr_std_based; log.u_bbr.flex1 <<= 1; log.u_bbr.flex1 |= rack->rc_rack_use_dsack; log.u_bbr.flex1 <<= 1; log.u_bbr.flex1 |= rack->rc_dsack_round_seen; log.u_bbr.flex2 = rack->r_ctl.dsack_round_end; log.u_bbr.flex3 = rack->r_ctl.num_dsack; log.u_bbr.flex4 = flex4; log.u_bbr.flex5 = flex5; log.u_bbr.flex6 = flex6; log.u_bbr.flex7 = rack->r_ctl.dsack_persist; log.u_bbr.flex8 = mod; log.u_bbr.timeStamp = tcp_get_usecs(&tv); TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, RACK_DSACK_HANDLING, 0, 0, &log, false, &tv); } } static void rack_log_hdwr_pacing(struct tcp_rack *rack, uint64_t rate, uint64_t hw_rate, int line, int error, uint16_t mod) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; const struct ifnet *ifp; memset(&log, 0, sizeof(log)); log.u_bbr.flex1 = ((hw_rate >> 32) & 0x00000000ffffffff); log.u_bbr.flex2 = (hw_rate & 0x00000000ffffffff); if (rack->r_ctl.crte) { ifp = rack->r_ctl.crte->ptbl->rs_ifp; } else if (rack->rc_inp->inp_route.ro_nh && rack->rc_inp->inp_route.ro_nh->nh_ifp) { ifp = rack->rc_inp->inp_route.ro_nh->nh_ifp; } else ifp = NULL; if (ifp) { log.u_bbr.flex3 = (((uint64_t)ifp >> 32) & 0x00000000ffffffff); log.u_bbr.flex4 = ((uint64_t)ifp & 0x00000000ffffffff); } log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.bw_inuse = rate; log.u_bbr.flex5 = line; log.u_bbr.flex6 = error; log.u_bbr.flex7 = mod; log.u_bbr.applimited = rack->r_ctl.rc_pace_max_segs; log.u_bbr.flex8 = rack->use_fixed_rate; log.u_bbr.flex8 <<= 1; log.u_bbr.flex8 |= rack->rack_hdrw_pacing; log.u_bbr.pkts_out = rack->rc_tp->t_maxseg; log.u_bbr.delRate = rack->r_ctl.crte_prev_rate; if (rack->r_ctl.crte) log.u_bbr.cur_del_rate = rack->r_ctl.crte->rate; else log.u_bbr.cur_del_rate = 0; log.u_bbr.rttProp = rack->r_ctl.last_hw_bw_req; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_HDWR_PACE, 0, 0, &log, false, &tv); } } static uint64_t rack_get_output_bw(struct tcp_rack *rack, uint64_t bw, struct rack_sendmap *rsm, int *capped) { /* * We allow rack_per_of_gp_xx to dictate our bw rate we want. */ uint64_t bw_est, high_rate; uint64_t gain; gain = (uint64_t)rack_get_output_gain(rack, rsm); bw_est = bw * gain; bw_est /= (uint64_t)100; /* Never fall below the minimum (def 64kbps) */ if (bw_est < RACK_MIN_BW) bw_est = RACK_MIN_BW; if (rack->r_rack_hw_rate_caps) { /* Rate caps are in place */ if (rack->r_ctl.crte != NULL) { /* We have a hdwr rate already */ high_rate = tcp_hw_highest_rate(rack->r_ctl.crte); if (bw_est >= high_rate) { /* We are capping bw at the highest rate table entry */ rack_log_hdwr_pacing(rack, bw_est, high_rate, __LINE__, 0, 3); bw_est = high_rate; if (capped) *capped = 1; } } else if ((rack->rack_hdrw_pacing == 0) && (rack->rack_hdw_pace_ena) && (rack->rack_attempt_hdwr_pace == 0) && (rack->rc_inp->inp_route.ro_nh != NULL) && (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) { /* * Special case, we have not yet attempted hardware * pacing, and yet we may, when we do, find out if we are * above the highest rate. We need to know the maxbw for the interface * in question (if it supports ratelimiting). We get back * a 0, if the interface is not found in the RL lists. */ high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp); if (high_rate) { /* Yep, we have a rate is it above this rate? */ if (bw_est > high_rate) { bw_est = high_rate; if (capped) *capped = 1; } } } } return (bw_est); } static void rack_log_retran_reason(struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t tsused, uint32_t thresh, int mod) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; if ((mod != 1) && (rack_verbose_logging == 0)) { /* * We get 3 values currently for mod * 1 - We are retransmitting and this tells the reason. * 2 - We are clearing a dup-ack count. * 3 - We are incrementing a dup-ack count. * * The clear/increment are only logged * if you have BBverbose on. */ return; } memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.flex1 = tsused; log.u_bbr.flex2 = thresh; log.u_bbr.flex3 = rsm->r_flags; log.u_bbr.flex4 = rsm->r_dupack; log.u_bbr.flex5 = rsm->r_start; log.u_bbr.flex6 = rsm->r_end; log.u_bbr.flex8 = mod; log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp); log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto; log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto; log.u_bbr.pacing_gain = rack->r_must_retran; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_SETTINGS_CHG, 0, 0, &log, false, &tv); } } static void rack_log_to_start(struct tcp_rack *rack, uint32_t cts, uint32_t to, int32_t slot, uint8_t which) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.flex1 = rack->rc_tp->t_srtt; log.u_bbr.flex2 = to; log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags; log.u_bbr.flex4 = slot; log.u_bbr.flex5 = rack->rc_inp->inp_hptsslot; log.u_bbr.flex6 = rack->rc_tp->t_rxtcur; log.u_bbr.flex7 = rack->rc_in_persist; log.u_bbr.flex8 = which; if (rack->rack_no_prr) log.u_bbr.pkts_out = 0; else log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp); log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto; log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto; log.u_bbr.pacing_gain = rack->r_must_retran; log.u_bbr.cwnd_gain = rack->rc_has_collapsed; log.u_bbr.lt_epoch = rack->rc_tp->t_rxtshift; log.u_bbr.lost = rack_rto_min; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_TIMERSTAR, 0, 0, &log, false, &tv); } } static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num, struct rack_sendmap *rsm) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp); log.u_bbr.flex8 = to_num; log.u_bbr.flex1 = rack->r_ctl.rc_rack_min_rtt; log.u_bbr.flex2 = rack->rc_rack_rtt; if (rsm == NULL) log.u_bbr.flex3 = 0; else log.u_bbr.flex3 = rsm->r_end - rsm->r_start; if (rack->rack_no_prr) log.u_bbr.flex5 = 0; else log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto; log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto; log.u_bbr.pacing_gain = rack->r_must_retran; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_RTO, 0, 0, &log, false, &tv); } } static void rack_log_map_chg(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *prev, struct rack_sendmap *rsm, struct rack_sendmap *next, int flag, uint32_t th_ack, int line) { if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.flex8 = flag; log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp); log.u_bbr.cur_del_rate = (uint64_t)prev; log.u_bbr.delRate = (uint64_t)rsm; log.u_bbr.rttProp = (uint64_t)next; log.u_bbr.flex7 = 0; if (prev) { log.u_bbr.flex1 = prev->r_start; log.u_bbr.flex2 = prev->r_end; log.u_bbr.flex7 |= 0x4; } if (rsm) { log.u_bbr.flex3 = rsm->r_start; log.u_bbr.flex4 = rsm->r_end; log.u_bbr.flex7 |= 0x2; } if (next) { log.u_bbr.flex5 = next->r_start; log.u_bbr.flex6 = next->r_end; log.u_bbr.flex7 |= 0x1; } log.u_bbr.applimited = line; log.u_bbr.pkts_out = th_ack; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); if (rack->rack_no_prr) log.u_bbr.lost = 0; else log.u_bbr.lost = rack->r_ctl.rc_prr_sndcnt; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, TCP_LOG_MAPCHG, 0, 0, &log, false, &tv); } } static void rack_log_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, uint32_t t, uint32_t len, struct rack_sendmap *rsm, int conf) { if (tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp); log.u_bbr.flex1 = t; log.u_bbr.flex2 = len; log.u_bbr.flex3 = rack->r_ctl.rc_rack_min_rtt; log.u_bbr.flex4 = rack->r_ctl.rack_rs.rs_rtt_lowest; log.u_bbr.flex5 = rack->r_ctl.rack_rs.rs_rtt_highest; log.u_bbr.flex6 = rack->r_ctl.rack_rs.rs_us_rtrcnt; log.u_bbr.flex7 = conf; log.u_bbr.rttProp = (uint64_t)rack->r_ctl.rack_rs.rs_rtt_tot; log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.delivered = rack->r_ctl.rack_rs.rs_us_rtrcnt; log.u_bbr.pkts_out = rack->r_ctl.rack_rs.rs_flags; log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); if (rsm) { log.u_bbr.pkt_epoch = rsm->r_start; log.u_bbr.lost = rsm->r_end; log.u_bbr.cwnd_gain = rsm->r_rtr_cnt; /* We loose any upper of the 24 bits */ log.u_bbr.pacing_gain = (uint16_t)rsm->r_flags; } else { /* Its a SYN */ log.u_bbr.pkt_epoch = rack->rc_tp->iss; log.u_bbr.lost = 0; log.u_bbr.cwnd_gain = 0; log.u_bbr.pacing_gain = 0; } /* Write out general bits of interest rrs here */ log.u_bbr.use_lt_bw = rack->rc_highly_buffered; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->forced_ack; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->rc_gp_dyn_mul; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->in_probe_rtt; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->app_limited_needs_set; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->rc_gp_filled; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->rc_dragged_bottom; log.u_bbr.applimited = rack->r_ctl.rc_target_probertt_flight; log.u_bbr.epoch = rack->r_ctl.rc_time_probertt_starts; log.u_bbr.lt_epoch = rack->r_ctl.rc_time_probertt_entered; log.u_bbr.cur_del_rate = rack->r_ctl.rc_lower_rtt_us_cts; log.u_bbr.delRate = rack->r_ctl.rc_gp_srtt; log.u_bbr.bw_inuse = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time); log.u_bbr.bw_inuse <<= 32; if (rsm) log.u_bbr.bw_inuse |= ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]); TCP_LOG_EVENTP(tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_BBRRTT, 0, 0, &log, false, &tv); } } static void rack_log_rtt_sample(struct tcp_rack *rack, uint32_t rtt) { /* * Log the rtt sample we are * applying to the srtt algorithm in * useconds. */ if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; /* Convert our ms to a microsecond */ memset(&log, 0, sizeof(log)); log.u_bbr.flex1 = rtt; log.u_bbr.flex2 = rack->r_ctl.ack_count; log.u_bbr.flex3 = rack->r_ctl.sack_count; log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move; log.u_bbr.flex5 = rack->r_ctl.sack_moved_extra; log.u_bbr.flex6 = rack->rc_tp->t_rxtcur; log.u_bbr.flex7 = 1; log.u_bbr.flex8 = rack->sack_attack_disable; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto; log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto; log.u_bbr.pacing_gain = rack->r_must_retran; /* * We capture in delRate the upper 32 bits as * the confidence level we had declared, and the * lower 32 bits as the actual RTT using the arrival * timestamp. */ log.u_bbr.delRate = rack->r_ctl.rack_rs.confidence; log.u_bbr.delRate <<= 32; log.u_bbr.delRate |= rack->r_ctl.rack_rs.rs_us_rtt; /* Lets capture all the things that make up t_rtxcur */ log.u_bbr.applimited = rack_rto_min; log.u_bbr.epoch = rack_rto_max; log.u_bbr.lt_epoch = rack->r_ctl.timer_slop; log.u_bbr.lost = rack_rto_min; log.u_bbr.pkt_epoch = TICKS_2_USEC(tcp_rexmit_slop); log.u_bbr.rttProp = RACK_REXMTVAL(rack->rc_tp); log.u_bbr.bw_inuse = rack->r_ctl.act_rcv_time.tv_sec; log.u_bbr.bw_inuse *= HPTS_USEC_IN_SEC; log.u_bbr.bw_inuse += rack->r_ctl.act_rcv_time.tv_usec; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, TCP_LOG_RTT, 0, 0, &log, false, &tv); } } static void rack_log_rtt_sample_calc(struct tcp_rack *rack, uint32_t rtt, uint32_t send_time, uint32_t ack_time, int where) { if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; struct timeval tv; /* Convert our ms to a microsecond */ memset(&log, 0, sizeof(log)); log.u_bbr.flex1 = rtt; log.u_bbr.flex2 = send_time; log.u_bbr.flex3 = ack_time; log.u_bbr.flex4 = where; log.u_bbr.flex7 = 2; log.u_bbr.timeStamp = tcp_get_usecs(&tv); TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, TCP_LOG_RTT, 0, 0, &log, false, &tv); } } static inline void rack_log_progress_event(struct tcp_rack *rack, struct tcpcb *tp, uint32_t tick, int event, int line) { if (rack_verbose_logging && (tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp); log.u_bbr.flex1 = line; log.u_bbr.flex2 = tick; log.u_bbr.flex3 = tp->t_maxunacktime; log.u_bbr.flex4 = tp->t_acktime; log.u_bbr.flex8 = event; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto; log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto; log.u_bbr.pacing_gain = rack->r_must_retran; TCP_LOG_EVENTP(tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_PROGRESS, 0, 0, &log, false, &tv); } } static void rack_log_type_bbrsnd(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint32_t cts, struct timeval *tv) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp); log.u_bbr.flex1 = slot; if (rack->rack_no_prr) log.u_bbr.flex2 = 0; else log.u_bbr.flex2 = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.flex7 = (0x0000ffff & rack->r_ctl.rc_hpts_flags); log.u_bbr.flex8 = rack->rc_in_persist; log.u_bbr.timeStamp = cts; log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto; log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto; log.u_bbr.pacing_gain = rack->r_must_retran; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_BBRSND, 0, 0, &log, false, tv); } } static void rack_log_doseg_done(struct tcp_rack *rack, uint32_t cts, int32_t nxt_pkt, int32_t did_out, int way_out, int nsegs) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log, 0, sizeof(log)); log.u_bbr.flex1 = did_out; log.u_bbr.flex2 = nxt_pkt; log.u_bbr.flex3 = way_out; log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags; if (rack->rack_no_prr) log.u_bbr.flex5 = 0; else log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.flex6 = nsegs; log.u_bbr.applimited = rack->r_ctl.rc_pace_min_segs; log.u_bbr.flex7 = rack->rc_ack_can_sendout_data; /* Do we have ack-can-send set */ log.u_bbr.flex7 <<= 1; log.u_bbr.flex7 |= rack->r_fast_output; /* is fast output primed */ log.u_bbr.flex7 <<= 1; log.u_bbr.flex7 |= rack->r_wanted_output; /* Do we want output */ log.u_bbr.flex8 = rack->rc_in_persist; log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp); log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.use_lt_bw = rack->r_ent_rec_ns; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->r_might_revert; log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto; log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto; log.u_bbr.pacing_gain = rack->r_must_retran; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_DOSEG_DONE, 0, 0, &log, false, &tv); } } static void rack_log_type_pacing_sizes(struct tcpcb *tp, struct tcp_rack *rack, uint32_t arg1, uint32_t arg2, uint32_t arg3, uint8_t frm) { if (tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log, 0, sizeof(log)); log.u_bbr.flex1 = rack->r_ctl.rc_pace_min_segs; log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs; log.u_bbr.flex4 = arg1; log.u_bbr.flex5 = arg2; log.u_bbr.flex6 = arg3; log.u_bbr.flex8 = frm; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto; log.u_bbr.applimited = rack->r_ctl.rc_sacked; log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto; log.u_bbr.pacing_gain = rack->r_must_retran; TCP_LOG_EVENTP(tp, NULL, &tp->t_inpcb->inp_socket->so_rcv, &tp->t_inpcb->inp_socket->so_snd, TCP_HDWR_PACE_SIZE, 0, 0, &log, false, &tv); } } static void rack_log_type_just_return(struct tcp_rack *rack, uint32_t cts, uint32_t tlen, uint32_t slot, uint8_t hpts_calling, int reason, uint32_t cwnd_to_use) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp); log.u_bbr.flex1 = slot; log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags; log.u_bbr.flex4 = reason; if (rack->rack_no_prr) log.u_bbr.flex5 = 0; else log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.flex7 = hpts_calling; log.u_bbr.flex8 = rack->rc_in_persist; log.u_bbr.lt_epoch = cwnd_to_use; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto; log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto; log.u_bbr.pacing_gain = rack->r_must_retran; log.u_bbr.cwnd_gain = rack->rc_has_collapsed; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_JUSTRET, 0, tlen, &log, false, &tv); } } static void rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, int line, uint32_t us_cts, struct timeval *tv, uint32_t flags_on_entry) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp); log.u_bbr.flex1 = line; log.u_bbr.flex2 = rack->r_ctl.rc_last_output_to; log.u_bbr.flex3 = flags_on_entry; log.u_bbr.flex4 = us_cts; if (rack->rack_no_prr) log.u_bbr.flex5 = 0; else log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.flex6 = rack->rc_tp->t_rxtcur; log.u_bbr.flex7 = hpts_removed; log.u_bbr.flex8 = 1; log.u_bbr.applimited = rack->r_ctl.rc_hpts_flags; log.u_bbr.timeStamp = us_cts; log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto; log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto; log.u_bbr.pacing_gain = rack->r_must_retran; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_TIMERCANC, 0, 0, &log, false, tv); } } static void rack_log_alt_to_to_cancel(struct tcp_rack *rack, uint32_t flex1, uint32_t flex2, uint32_t flex3, uint32_t flex4, uint32_t flex5, uint32_t flex6, uint16_t flex7, uint8_t mod) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; if (mod == 1) { /* No you can't use 1, its for the real to cancel */ return; } memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.flex1 = flex1; log.u_bbr.flex2 = flex2; log.u_bbr.flex3 = flex3; log.u_bbr.flex4 = flex4; log.u_bbr.flex5 = flex5; log.u_bbr.flex6 = flex6; log.u_bbr.flex7 = flex7; log.u_bbr.flex8 = mod; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_TIMERCANC, 0, 0, &log, false, &tv); } } static void rack_log_to_processing(struct tcp_rack *rack, uint32_t cts, int32_t ret, int32_t timers) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.flex1 = timers; log.u_bbr.flex2 = ret; log.u_bbr.flex3 = rack->r_ctl.rc_timer_exp; log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags; log.u_bbr.flex5 = cts; if (rack->rack_no_prr) log.u_bbr.flex6 = 0; else log.u_bbr.flex6 = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.pkts_out = rack->r_ctl.rc_out_at_rto; log.u_bbr.delivered = rack->r_ctl.rc_snd_max_at_rto; log.u_bbr.pacing_gain = rack->r_must_retran; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_TO_PROCESS, 0, 0, &log, false, &tv); } } static void rack_log_to_prr(struct tcp_rack *rack, int frm, int orig_cwnd, int line) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.flex1 = rack->r_ctl.rc_prr_out; log.u_bbr.flex2 = rack->r_ctl.rc_prr_recovery_fs; if (rack->rack_no_prr) log.u_bbr.flex3 = 0; else log.u_bbr.flex3 = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.flex4 = rack->r_ctl.rc_prr_delivered; log.u_bbr.flex5 = rack->r_ctl.rc_sacked; log.u_bbr.flex6 = rack->r_ctl.rc_holes_rxt; log.u_bbr.flex7 = line; log.u_bbr.flex8 = frm; log.u_bbr.pkts_out = orig_cwnd; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.use_lt_bw = rack->r_ent_rec_ns; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->r_might_revert; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_BBRUPD, 0, 0, &log, false, &tv); } } #ifdef NETFLIX_EXP_DETECTION static void rack_log_sad(struct tcp_rack *rack, int event) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.flex1 = rack->r_ctl.sack_count; log.u_bbr.flex2 = rack->r_ctl.ack_count; log.u_bbr.flex3 = rack->r_ctl.sack_moved_extra; log.u_bbr.flex4 = rack->r_ctl.sack_noextra_move; log.u_bbr.flex5 = rack->r_ctl.rc_num_maps_alloced; log.u_bbr.flex6 = tcp_sack_to_ack_thresh; log.u_bbr.pkts_out = tcp_sack_to_move_thresh; log.u_bbr.lt_epoch = (tcp_force_detection << 8); log.u_bbr.lt_epoch |= rack->do_detection; log.u_bbr.applimited = tcp_map_minimum; log.u_bbr.flex7 = rack->sack_attack_disable; log.u_bbr.flex8 = event; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.delivered = tcp_sad_decay_val; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, TCP_SAD_DETECTION, 0, 0, &log, false, &tv); } } #endif static void rack_counter_destroy(void) { counter_u64_free(rack_fto_send); counter_u64_free(rack_fto_rsm_send); counter_u64_free(rack_nfto_resend); counter_u64_free(rack_hw_pace_init_fail); counter_u64_free(rack_hw_pace_lost); counter_u64_free(rack_non_fto_send); counter_u64_free(rack_extended_rfo); counter_u64_free(rack_ack_total); counter_u64_free(rack_express_sack); counter_u64_free(rack_sack_total); counter_u64_free(rack_move_none); counter_u64_free(rack_move_some); counter_u64_free(rack_sack_attacks_detected); counter_u64_free(rack_sack_attacks_reversed); counter_u64_free(rack_sack_used_next_merge); counter_u64_free(rack_sack_used_prev_merge); counter_u64_free(rack_tlp_tot); counter_u64_free(rack_tlp_newdata); counter_u64_free(rack_tlp_retran); counter_u64_free(rack_tlp_retran_bytes); counter_u64_free(rack_to_tot); counter_u64_free(rack_saw_enobuf); counter_u64_free(rack_saw_enobuf_hw); counter_u64_free(rack_saw_enetunreach); counter_u64_free(rack_hot_alloc); counter_u64_free(rack_to_alloc); counter_u64_free(rack_to_alloc_hard); counter_u64_free(rack_to_alloc_emerg); counter_u64_free(rack_to_alloc_limited); counter_u64_free(rack_alloc_limited_conns); counter_u64_free(rack_split_limited); counter_u64_free(rack_multi_single_eq); counter_u64_free(rack_proc_non_comp_ack); counter_u64_free(rack_sack_proc_all); counter_u64_free(rack_sack_proc_restart); counter_u64_free(rack_sack_proc_short); counter_u64_free(rack_sack_skipped_acked); counter_u64_free(rack_sack_splits); counter_u64_free(rack_input_idle_reduces); counter_u64_free(rack_collapsed_win); counter_u64_free(rack_collapsed_win_rxt); counter_u64_free(rack_collapsed_win_rxt_bytes); counter_u64_free(rack_collapsed_win_seen); counter_u64_free(rack_try_scwnd); counter_u64_free(rack_persists_sends); counter_u64_free(rack_persists_acks); counter_u64_free(rack_persists_loss); counter_u64_free(rack_persists_lost_ends); #ifdef INVARIANTS counter_u64_free(rack_adjust_map_bw); #endif COUNTER_ARRAY_FREE(rack_out_size, TCP_MSS_ACCT_SIZE); COUNTER_ARRAY_FREE(rack_opts_arry, RACK_OPTS_SIZE); } static struct rack_sendmap * rack_alloc(struct tcp_rack *rack) { struct rack_sendmap *rsm; /* * First get the top of the list it in * theory is the "hottest" rsm we have, * possibly just freed by ack processing. */ if (rack->rc_free_cnt > rack_free_cache) { rsm = TAILQ_FIRST(&rack->r_ctl.rc_free); TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext); counter_u64_add(rack_hot_alloc, 1); rack->rc_free_cnt--; return (rsm); } /* * Once we get under our free cache we probably * no longer have a "hot" one available. Lets * get one from UMA. */ rsm = uma_zalloc(rack_zone, M_NOWAIT); if (rsm) { rack->r_ctl.rc_num_maps_alloced++; counter_u64_add(rack_to_alloc, 1); return (rsm); } /* * Dig in to our aux rsm's (the last two) since * UMA failed to get us one. */ if (rack->rc_free_cnt) { counter_u64_add(rack_to_alloc_emerg, 1); rsm = TAILQ_FIRST(&rack->r_ctl.rc_free); TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext); rack->rc_free_cnt--; return (rsm); } return (NULL); } static struct rack_sendmap * rack_alloc_full_limit(struct tcp_rack *rack) { if ((V_tcp_map_entries_limit > 0) && (rack->do_detection == 0) && (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) { counter_u64_add(rack_to_alloc_limited, 1); if (!rack->alloc_limit_reported) { rack->alloc_limit_reported = 1; counter_u64_add(rack_alloc_limited_conns, 1); } return (NULL); } return (rack_alloc(rack)); } /* wrapper to allocate a sendmap entry, subject to a specific limit */ static struct rack_sendmap * rack_alloc_limit(struct tcp_rack *rack, uint8_t limit_type) { struct rack_sendmap *rsm; if (limit_type) { /* currently there is only one limit type */ if (V_tcp_map_split_limit > 0 && (rack->do_detection == 0) && rack->r_ctl.rc_num_split_allocs >= V_tcp_map_split_limit) { counter_u64_add(rack_split_limited, 1); if (!rack->alloc_limit_reported) { rack->alloc_limit_reported = 1; counter_u64_add(rack_alloc_limited_conns, 1); } return (NULL); } } /* allocate and mark in the limit type, if set */ rsm = rack_alloc(rack); if (rsm != NULL && limit_type) { rsm->r_limit_type = limit_type; rack->r_ctl.rc_num_split_allocs++; } return (rsm); } static void rack_free(struct tcp_rack *rack, struct rack_sendmap *rsm) { if (rsm->r_flags & RACK_APP_LIMITED) { if (rack->r_ctl.rc_app_limited_cnt > 0) { rack->r_ctl.rc_app_limited_cnt--; } } if (rsm->r_limit_type) { /* currently there is only one limit type */ rack->r_ctl.rc_num_split_allocs--; } if (rsm == rack->r_ctl.rc_first_appl) { if (rack->r_ctl.rc_app_limited_cnt == 0) rack->r_ctl.rc_first_appl = NULL; else { /* Follow the next one out */ struct rack_sendmap fe; fe.r_start = rsm->r_nseq_appl; rack->r_ctl.rc_first_appl = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe); } } if (rsm == rack->r_ctl.rc_resend) rack->r_ctl.rc_resend = NULL; if (rsm == rack->r_ctl.rc_end_appl) rack->r_ctl.rc_end_appl = NULL; if (rack->r_ctl.rc_tlpsend == rsm) rack->r_ctl.rc_tlpsend = NULL; if (rack->r_ctl.rc_sacklast == rsm) rack->r_ctl.rc_sacklast = NULL; memset(rsm, 0, sizeof(struct rack_sendmap)); TAILQ_INSERT_HEAD(&rack->r_ctl.rc_free, rsm, r_tnext); rack->rc_free_cnt++; } static void rack_free_trim(struct tcp_rack *rack) { struct rack_sendmap *rsm; /* * Free up all the tail entries until * we get our list down to the limit. */ while (rack->rc_free_cnt > rack_free_cache) { rsm = TAILQ_LAST(&rack->r_ctl.rc_free, rack_head); TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext); rack->rc_free_cnt--; uma_zfree(rack_zone, rsm); } } static uint32_t rack_get_measure_window(struct tcpcb *tp, struct tcp_rack *rack) { uint64_t srtt, bw, len, tim; uint32_t segsiz, def_len, minl; segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs); def_len = rack_def_data_window * segsiz; if (rack->rc_gp_filled == 0) { /* * We have no measurement (IW is in flight?) so * we can only guess using our data_window sysctl * value (usually 20MSS). */ return (def_len); } /* * Now we have a number of factors to consider. * * 1) We have a desired BDP which is usually * at least 2. * 2) We have a minimum number of rtt's usually 1 SRTT * but we allow it too to be more. * 3) We want to make sure a measurement last N useconds (if * we have set rack_min_measure_usec. * * We handle the first concern here by trying to create a data * window of max(rack_def_data_window, DesiredBDP). The * second concern we handle in not letting the measurement * window end normally until at least the required SRTT's * have gone by which is done further below in * rack_enough_for_measurement(). Finally the third concern * we also handle here by calculating how long that time * would take at the current BW and then return the * max of our first calculation and that length. Note * that if rack_min_measure_usec is 0, we don't deal * with concern 3. Also for both Concern 1 and 3 an * application limited period could end the measurement * earlier. * * So lets calculate the BDP with the "known" b/w using * the SRTT has our rtt and then multiply it by the * goal. */ bw = rack_get_bw(rack); srtt = (uint64_t)tp->t_srtt; len = bw * srtt; len /= (uint64_t)HPTS_USEC_IN_SEC; len *= max(1, rack_goal_bdp); /* Now we need to round up to the nearest MSS */ len = roundup(len, segsiz); if (rack_min_measure_usec) { /* Now calculate our min length for this b/w */ tim = rack_min_measure_usec; minl = (tim * bw) / (uint64_t)HPTS_USEC_IN_SEC; if (minl == 0) minl = 1; minl = roundup(minl, segsiz); if (len < minl) len = minl; } /* * Now if we have a very small window we want * to attempt to get the window that is * as small as possible. This happens on * low b/w connections and we don't want to * span huge numbers of rtt's between measurements. * * We basically include 2 over our "MIN window" so * that the measurement can be shortened (possibly) by * an ack'ed packet. */ if (len < def_len) return (max((uint32_t)len, ((MIN_GP_WIN+2) * segsiz))); else return (max((uint32_t)len, def_len)); } static int rack_enough_for_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack, uint8_t *quality) { uint32_t tim, srtts, segsiz; /* * Has enough time passed for the GP measurement to be valid? */ if ((tp->snd_max == tp->snd_una) || (th_ack == tp->snd_max)){ /* All is acked */ *quality = RACK_QUALITY_ALLACKED; return (1); } if (SEQ_LT(th_ack, tp->gput_seq)) { /* Not enough bytes yet */ return (0); } segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs); if (SEQ_LT(th_ack, tp->gput_ack) && ((th_ack - tp->gput_seq) < max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) { /* Not enough bytes yet */ return (0); } if (rack->r_ctl.rc_first_appl && (SEQ_GEQ(th_ack, rack->r_ctl.rc_first_appl->r_end))) { /* * We are up to the app limited send point * we have to measure irrespective of the time.. */ *quality = RACK_QUALITY_APPLIMITED; return (1); } /* Now what about time? */ srtts = (rack->r_ctl.rc_gp_srtt * rack_min_srtts); tim = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - tp->gput_ts; if (tim >= srtts) { *quality = RACK_QUALITY_HIGH; return (1); } /* Nope not even a full SRTT has passed */ return (0); } static void rack_log_timely(struct tcp_rack *rack, uint32_t logged, uint64_t cur_bw, uint64_t low_bnd, uint64_t up_bnd, int line, uint8_t method) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log, 0, sizeof(log)); log.u_bbr.flex1 = logged; log.u_bbr.flex2 = rack->rc_gp_timely_inc_cnt; log.u_bbr.flex2 <<= 4; log.u_bbr.flex2 |= rack->rc_gp_timely_dec_cnt; log.u_bbr.flex2 <<= 4; log.u_bbr.flex2 |= rack->rc_gp_incr; log.u_bbr.flex2 <<= 4; log.u_bbr.flex2 |= rack->rc_gp_bwred; log.u_bbr.flex3 = rack->rc_gp_incr; log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss; log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ca; log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_rec; log.u_bbr.flex7 = rack->rc_gp_bwred; log.u_bbr.flex8 = method; log.u_bbr.cur_del_rate = cur_bw; log.u_bbr.delRate = low_bnd; log.u_bbr.bw_inuse = up_bnd; log.u_bbr.rttProp = rack_get_bw(rack); log.u_bbr.pkt_epoch = line; log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt; log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt; log.u_bbr.cwnd_gain = rack->rc_dragged_bottom; log.u_bbr.cwnd_gain <<= 1; log.u_bbr.cwnd_gain |= rack->rc_gp_saw_rec; log.u_bbr.cwnd_gain <<= 1; log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss; log.u_bbr.cwnd_gain <<= 1; log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca; log.u_bbr.lost = rack->r_ctl.rc_loss_count; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, TCP_TIMELY_WORK, 0, 0, &log, false, &tv); } } static int rack_bw_can_be_raised(struct tcp_rack *rack, uint64_t cur_bw, uint64_t last_bw_est, uint16_t mult) { /* * Before we increase we need to know if * the estimate just made was less than * our pacing goal (i.e. (cur_bw * mult) > last_bw_est) * * If we already are pacing at a fast enough * rate to push us faster there is no sense of * increasing. * * We first caculate our actual pacing rate (ss or ca multiplier * times our cur_bw). * * Then we take the last measured rate and multipy by our * maximum pacing overage to give us a max allowable rate. * * If our act_rate is smaller than our max_allowable rate * then we should increase. Else we should hold steady. * */ uint64_t act_rate, max_allow_rate; if (rack_timely_no_stopping) return (1); if ((cur_bw == 0) || (last_bw_est == 0)) { /* * Initial startup case or * everything is acked case. */ rack_log_timely(rack, mult, cur_bw, 0, 0, __LINE__, 9); return (1); } if (mult <= 100) { /* * We can always pace at or slightly above our rate. */ rack_log_timely(rack, mult, cur_bw, 0, 0, __LINE__, 9); return (1); } act_rate = cur_bw * (uint64_t)mult; act_rate /= 100; max_allow_rate = last_bw_est * ((uint64_t)rack_max_per_above + (uint64_t)100); max_allow_rate /= 100; if (act_rate < max_allow_rate) { /* * Here the rate we are actually pacing at * is smaller than 10% above our last measurement. * This means we are pacing below what we would * like to try to achieve (plus some wiggle room). */ rack_log_timely(rack, mult, cur_bw, act_rate, max_allow_rate, __LINE__, 9); return (1); } else { /* * Here we are already pacing at least rack_max_per_above(10%) * what we are getting back. This indicates most likely * that we are being limited (cwnd/rwnd/app) and can't * get any more b/w. There is no sense of trying to * raise up the pacing rate its not speeding us up * and we already are pacing faster than we are getting. */ rack_log_timely(rack, mult, cur_bw, act_rate, max_allow_rate, __LINE__, 8); return (0); } } static void rack_validate_multipliers_at_or_above100(struct tcp_rack *rack) { /* * When we drag bottom, we want to assure * that no multiplier is below 1.0, if so * we want to restore it to at least that. */ if (rack->r_ctl.rack_per_of_gp_rec < 100) { /* This is unlikely we usually do not touch recovery */ rack->r_ctl.rack_per_of_gp_rec = 100; } if (rack->r_ctl.rack_per_of_gp_ca < 100) { rack->r_ctl.rack_per_of_gp_ca = 100; } if (rack->r_ctl.rack_per_of_gp_ss < 100) { rack->r_ctl.rack_per_of_gp_ss = 100; } } static void rack_validate_multipliers_at_or_below_100(struct tcp_rack *rack) { if (rack->r_ctl.rack_per_of_gp_ca > 100) { rack->r_ctl.rack_per_of_gp_ca = 100; } if (rack->r_ctl.rack_per_of_gp_ss > 100) { rack->r_ctl.rack_per_of_gp_ss = 100; } } static void rack_increase_bw_mul(struct tcp_rack *rack, int timely_says, uint64_t cur_bw, uint64_t last_bw_est, int override) { int32_t calc, logged, plus; logged = 0; if (override) { /* * override is passed when we are * loosing b/w and making one last * gasp at trying to not loose out * to a new-reno flow. */ goto extra_boost; } /* In classic timely we boost by 5x if we have 5 increases in a row, lets not */ if (rack->rc_gp_incr && ((rack->rc_gp_timely_inc_cnt + 1) >= RACK_TIMELY_CNT_BOOST)) { /* * Reset and get 5 strokes more before the boost. Note * that the count is 0 based so we have to add one. */ extra_boost: plus = (uint32_t)rack_gp_increase_per * RACK_TIMELY_CNT_BOOST; rack->rc_gp_timely_inc_cnt = 0; } else plus = (uint32_t)rack_gp_increase_per; /* Must be at least 1% increase for true timely increases */ if ((plus < 1) && ((rack->r_ctl.rc_rtt_diff <= 0) || (timely_says <= 0))) plus = 1; if (rack->rc_gp_saw_rec && (rack->rc_gp_no_rec_chg == 0) && rack_bw_can_be_raised(rack, cur_bw, last_bw_est, rack->r_ctl.rack_per_of_gp_rec)) { /* We have been in recovery ding it too */ calc = rack->r_ctl.rack_per_of_gp_rec + plus; if (calc > 0xffff) calc = 0xffff; logged |= 1; rack->r_ctl.rack_per_of_gp_rec = (uint16_t)calc; if (rack_per_upper_bound_ss && (rack->rc_dragged_bottom == 0) && (rack->r_ctl.rack_per_of_gp_rec > rack_per_upper_bound_ss)) rack->r_ctl.rack_per_of_gp_rec = rack_per_upper_bound_ss; } if (rack->rc_gp_saw_ca && (rack->rc_gp_saw_ss == 0) && rack_bw_can_be_raised(rack, cur_bw, last_bw_est, rack->r_ctl.rack_per_of_gp_ca)) { /* In CA */ calc = rack->r_ctl.rack_per_of_gp_ca + plus; if (calc > 0xffff) calc = 0xffff; logged |= 2; rack->r_ctl.rack_per_of_gp_ca = (uint16_t)calc; if (rack_per_upper_bound_ca && (rack->rc_dragged_bottom == 0) && (rack->r_ctl.rack_per_of_gp_ca > rack_per_upper_bound_ca)) rack->r_ctl.rack_per_of_gp_ca = rack_per_upper_bound_ca; } if (rack->rc_gp_saw_ss && rack_bw_can_be_raised(rack, cur_bw, last_bw_est, rack->r_ctl.rack_per_of_gp_ss)) { /* In SS */ calc = rack->r_ctl.rack_per_of_gp_ss + plus; if (calc > 0xffff) calc = 0xffff; rack->r_ctl.rack_per_of_gp_ss = (uint16_t)calc; if (rack_per_upper_bound_ss && (rack->rc_dragged_bottom == 0) && (rack->r_ctl.rack_per_of_gp_ss > rack_per_upper_bound_ss)) rack->r_ctl.rack_per_of_gp_ss = rack_per_upper_bound_ss; logged |= 4; } if (logged && (rack->rc_gp_incr == 0)){ /* Go into increment mode */ rack->rc_gp_incr = 1; rack->rc_gp_timely_inc_cnt = 0; } if (rack->rc_gp_incr && logged && (rack->rc_gp_timely_inc_cnt < RACK_TIMELY_CNT_BOOST)) { rack->rc_gp_timely_inc_cnt++; } rack_log_timely(rack, logged, plus, 0, 0, __LINE__, 1); } static uint32_t rack_get_decrease(struct tcp_rack *rack, uint32_t curper, int32_t rtt_diff) { /* * norm_grad = rtt_diff / minrtt; * new_per = curper * (1 - B * norm_grad) * * B = rack_gp_decrease_per (default 10%) * rtt_dif = input var current rtt-diff * curper = input var current percentage * minrtt = from rack filter * */ uint64_t perf; perf = (((uint64_t)curper * ((uint64_t)1000000 - ((uint64_t)rack_gp_decrease_per * (uint64_t)10000 * (((uint64_t)rtt_diff * (uint64_t)1000000)/ (uint64_t)get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)))/ (uint64_t)1000000)) / (uint64_t)1000000); if (perf > curper) { /* TSNH */ perf = curper - 1; } return ((uint32_t)perf); } static uint32_t rack_decrease_highrtt(struct tcp_rack *rack, uint32_t curper, uint32_t rtt) { /* * highrttthresh * result = curper * (1 - (B * ( 1 - ------ )) * gp_srtt * * B = rack_gp_decrease_per (default 10%) * highrttthresh = filter_min * rack_gp_rtt_maxmul */ uint64_t perf; uint32_t highrttthresh; highrttthresh = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul; perf = (((uint64_t)curper * ((uint64_t)1000000 - ((uint64_t)rack_gp_decrease_per * ((uint64_t)1000000 - ((uint64_t)highrttthresh * (uint64_t)1000000) / (uint64_t)rtt)) / 100)) /(uint64_t)1000000); return (perf); } static void rack_decrease_bw_mul(struct tcp_rack *rack, int timely_says, uint32_t rtt, int32_t rtt_diff) { uint64_t logvar, logvar2, logvar3; uint32_t logged, new_per, ss_red, ca_red, rec_red, alt, val; if (rack->rc_gp_incr) { /* Turn off increment counting */ rack->rc_gp_incr = 0; rack->rc_gp_timely_inc_cnt = 0; } ss_red = ca_red = rec_red = 0; logged = 0; /* Calculate the reduction value */ if (rtt_diff < 0) { rtt_diff *= -1; } /* Must be at least 1% reduction */ if (rack->rc_gp_saw_rec && (rack->rc_gp_no_rec_chg == 0)) { /* We have been in recovery ding it too */ if (timely_says == 2) { new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_rec, rtt); alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff); if (alt < new_per) val = alt; else val = new_per; } else val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff); if (rack->r_ctl.rack_per_of_gp_rec > val) { rec_red = (rack->r_ctl.rack_per_of_gp_rec - val); rack->r_ctl.rack_per_of_gp_rec = (uint16_t)val; } else { rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound; rec_red = 0; } if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_rec) rack->r_ctl.rack_per_of_gp_rec = rack_per_lower_bound; logged |= 1; } if (rack->rc_gp_saw_ss) { /* Sent in SS */ if (timely_says == 2) { new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ss, rtt); alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff); if (alt < new_per) val = alt; else val = new_per; } else val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ss, rtt_diff); if (rack->r_ctl.rack_per_of_gp_ss > new_per) { ss_red = rack->r_ctl.rack_per_of_gp_ss - val; rack->r_ctl.rack_per_of_gp_ss = (uint16_t)val; } else { ss_red = new_per; rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound; logvar = new_per; logvar <<= 32; logvar |= alt; logvar2 = (uint32_t)rtt; logvar2 <<= 32; logvar2 |= (uint32_t)rtt_diff; logvar3 = rack_gp_rtt_maxmul; logvar3 <<= 32; logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt); rack_log_timely(rack, timely_says, logvar2, logvar3, logvar, __LINE__, 10); } if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ss) rack->r_ctl.rack_per_of_gp_ss = rack_per_lower_bound; logged |= 4; } else if (rack->rc_gp_saw_ca) { /* Sent in CA */ if (timely_says == 2) { new_per = rack_decrease_highrtt(rack, rack->r_ctl.rack_per_of_gp_ca, rtt); alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_rec, rtt_diff); if (alt < new_per) val = alt; else val = new_per; } else val = new_per = alt = rack_get_decrease(rack, rack->r_ctl.rack_per_of_gp_ca, rtt_diff); if (rack->r_ctl.rack_per_of_gp_ca > val) { ca_red = rack->r_ctl.rack_per_of_gp_ca - val; rack->r_ctl.rack_per_of_gp_ca = (uint16_t)val; } else { rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound; ca_red = 0; logvar = new_per; logvar <<= 32; logvar |= alt; logvar2 = (uint32_t)rtt; logvar2 <<= 32; logvar2 |= (uint32_t)rtt_diff; logvar3 = rack_gp_rtt_maxmul; logvar3 <<= 32; logvar3 |= get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt); rack_log_timely(rack, timely_says, logvar2, logvar3, logvar, __LINE__, 10); } if (rack_per_lower_bound > rack->r_ctl.rack_per_of_gp_ca) rack->r_ctl.rack_per_of_gp_ca = rack_per_lower_bound; logged |= 2; } if (rack->rc_gp_timely_dec_cnt < 0x7) { rack->rc_gp_timely_dec_cnt++; if (rack_timely_dec_clear && (rack->rc_gp_timely_dec_cnt == rack_timely_dec_clear)) rack->rc_gp_timely_dec_cnt = 0; } logvar = ss_red; logvar <<= 32; logvar |= ca_red; rack_log_timely(rack, logged, rec_red, rack_per_lower_bound, logvar, __LINE__, 2); } static void rack_log_rtt_shrinks(struct tcp_rack *rack, uint32_t us_cts, uint32_t rtt, uint32_t line, uint8_t reas) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.flex1 = line; log.u_bbr.flex2 = rack->r_ctl.rc_time_probertt_starts; log.u_bbr.flex3 = rack->r_ctl.rc_lower_rtt_us_cts; log.u_bbr.flex4 = rack->r_ctl.rack_per_of_gp_ss; log.u_bbr.flex5 = rtt; log.u_bbr.flex6 = rack->rc_highly_buffered; log.u_bbr.flex6 <<= 1; log.u_bbr.flex6 |= rack->forced_ack; log.u_bbr.flex6 <<= 1; log.u_bbr.flex6 |= rack->rc_gp_dyn_mul; log.u_bbr.flex6 <<= 1; log.u_bbr.flex6 |= rack->in_probe_rtt; log.u_bbr.flex6 <<= 1; log.u_bbr.flex6 |= rack->measure_saw_probe_rtt; log.u_bbr.flex7 = rack->r_ctl.rack_per_of_gp_probertt; log.u_bbr.pacing_gain = rack->r_ctl.rack_per_of_gp_ca; log.u_bbr.cwnd_gain = rack->r_ctl.rack_per_of_gp_rec; log.u_bbr.flex8 = reas; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.delRate = rack_get_bw(rack); log.u_bbr.cur_del_rate = rack->r_ctl.rc_highest_us_rtt; log.u_bbr.cur_del_rate <<= 32; log.u_bbr.cur_del_rate |= rack->r_ctl.rc_lowest_us_rtt; log.u_bbr.applimited = rack->r_ctl.rc_time_probertt_entered; log.u_bbr.pkts_out = rack->r_ctl.rc_rtt_diff; log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.epoch = rack->r_ctl.rc_gp_srtt; log.u_bbr.lt_epoch = rack->r_ctl.rc_prev_gp_srtt; log.u_bbr.pkt_epoch = rack->r_ctl.rc_lower_rtt_us_cts; log.u_bbr.delivered = rack->r_ctl.rc_target_probertt_flight; log.u_bbr.lost = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt); log.u_bbr.rttProp = us_cts; log.u_bbr.rttProp <<= 32; log.u_bbr.rttProp |= rack->r_ctl.rc_entry_gp_rtt; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_RTT_SHRINKS, 0, 0, &log, false, &rack->r_ctl.act_rcv_time); } } static void rack_set_prtt_target(struct tcp_rack *rack, uint32_t segsiz, uint32_t rtt) { uint64_t bwdp; bwdp = rack_get_bw(rack); bwdp *= (uint64_t)rtt; bwdp /= (uint64_t)HPTS_USEC_IN_SEC; rack->r_ctl.rc_target_probertt_flight = roundup((uint32_t)bwdp, segsiz); if (rack->r_ctl.rc_target_probertt_flight < (segsiz * rack_timely_min_segs)) { /* * A window protocol must be able to have 4 packets * outstanding as the floor in order to function * (especially considering delayed ack :D). */ rack->r_ctl.rc_target_probertt_flight = (segsiz * rack_timely_min_segs); } } static void rack_enter_probertt(struct tcp_rack *rack, uint32_t us_cts) { /** * ProbeRTT is a bit different in rack_pacing than in * BBR. It is like BBR in that it uses the lowering of * the RTT as a signal that we saw something new and * counts from there for how long between. But it is * different in that its quite simple. It does not * play with the cwnd and wait until we get down * to N segments outstanding and hold that for * 200ms. Instead it just sets the pacing reduction * rate to a set percentage (70 by default) and hold * that for a number of recent GP Srtt's. */ uint32_t segsiz; if (rack->rc_gp_dyn_mul == 0) return; if (rack->rc_tp->snd_max == rack->rc_tp->snd_una) { /* We are idle */ return; } if ((rack->rc_tp->t_flags & TF_GPUTINPROG) && SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) { /* * Stop the goodput now, the idea here is * that future measurements with in_probe_rtt * won't register if they are not greater so * we want to get what info (if any) is available * now. */ rack_do_goodput_measurement(rack->rc_tp, rack, rack->rc_tp->snd_una, __LINE__, RACK_QUALITY_PROBERTT); } rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt; rack->r_ctl.rc_time_probertt_entered = us_cts; segsiz = min(ctf_fixed_maxseg(rack->rc_tp), rack->r_ctl.rc_pace_min_segs); rack->in_probe_rtt = 1; rack->measure_saw_probe_rtt = 1; rack->r_ctl.rc_lower_rtt_us_cts = us_cts; rack->r_ctl.rc_time_probertt_starts = 0; rack->r_ctl.rc_entry_gp_rtt = rack->r_ctl.rc_gp_srtt; if (rack_probertt_use_min_rtt_entry) rack_set_prtt_target(rack, segsiz, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)); else rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_gp_srtt); rack_log_rtt_shrinks(rack, us_cts, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), __LINE__, RACK_RTTS_ENTERPROBE); } static void rack_exit_probertt(struct tcp_rack *rack, uint32_t us_cts) { struct rack_sendmap *rsm; uint32_t segsiz; segsiz = min(ctf_fixed_maxseg(rack->rc_tp), rack->r_ctl.rc_pace_min_segs); rack->in_probe_rtt = 0; if ((rack->rc_tp->t_flags & TF_GPUTINPROG) && SEQ_GT(rack->rc_tp->snd_una, rack->rc_tp->gput_seq)) { /* * Stop the goodput now, the idea here is * that future measurements with in_probe_rtt * won't register if they are not greater so * we want to get what info (if any) is available * now. */ rack_do_goodput_measurement(rack->rc_tp, rack, rack->rc_tp->snd_una, __LINE__, RACK_QUALITY_PROBERTT); } else if (rack->rc_tp->t_flags & TF_GPUTINPROG) { /* * We don't have enough data to make a measurement. * So lets just stop and start here after exiting * probe-rtt. We probably are not interested in * the results anyway. */ rack->rc_tp->t_flags &= ~TF_GPUTINPROG; } /* * Measurements through the current snd_max are going * to be limited by the slower pacing rate. * * We need to mark these as app-limited so we * don't collapse the b/w. */ rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree); if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) { if (rack->r_ctl.rc_app_limited_cnt == 0) rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm; else { /* * Go out to the end app limited and mark * this new one as next and move the end_appl up * to this guy. */ if (rack->r_ctl.rc_end_appl) rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start; rack->r_ctl.rc_end_appl = rsm; } rsm->r_flags |= RACK_APP_LIMITED; rack->r_ctl.rc_app_limited_cnt++; } /* * Now, we need to examine our pacing rate multipliers. * If its under 100%, we need to kick it back up to * 100%. We also don't let it be over our "max" above * the actual rate i.e. 100% + rack_clamp_atexit_prtt. * Note setting clamp_atexit_prtt to 0 has the effect * of setting CA/SS to 100% always at exit (which is * the default behavior). */ if (rack_probertt_clear_is) { rack->rc_gp_incr = 0; rack->rc_gp_bwred = 0; rack->rc_gp_timely_inc_cnt = 0; rack->rc_gp_timely_dec_cnt = 0; } /* Do we do any clamping at exit? */ if (rack->rc_highly_buffered && rack_atexit_prtt_hbp) { rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt_hbp; rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt_hbp; } if ((rack->rc_highly_buffered == 0) && rack_atexit_prtt) { rack->r_ctl.rack_per_of_gp_ca = rack_atexit_prtt; rack->r_ctl.rack_per_of_gp_ss = rack_atexit_prtt; } /* * Lets set rtt_diff to 0, so that we will get a "boost" * after exiting. */ rack->r_ctl.rc_rtt_diff = 0; /* Clear all flags so we start fresh */ rack->rc_tp->t_bytes_acked = 0; rack->rc_tp->ccv->flags &= ~CCF_ABC_SENTAWND; /* * If configured to, set the cwnd and ssthresh to * our targets. */ if (rack_probe_rtt_sets_cwnd) { uint64_t ebdp; uint32_t setto; /* Set ssthresh so we get into CA once we hit our target */ if (rack_probertt_use_min_rtt_exit == 1) { /* Set to min rtt */ rack_set_prtt_target(rack, segsiz, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt)); } else if (rack_probertt_use_min_rtt_exit == 2) { /* Set to current gp rtt */ rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_gp_srtt); } else if (rack_probertt_use_min_rtt_exit == 3) { /* Set to entry gp rtt */ rack_set_prtt_target(rack, segsiz, rack->r_ctl.rc_entry_gp_rtt); } else { uint64_t sum; uint32_t setval; sum = rack->r_ctl.rc_entry_gp_rtt; sum *= 10; sum /= (uint64_t)(max(1, rack->r_ctl.rc_gp_srtt)); if (sum >= 20) { /* * A highly buffered path needs * cwnd space for timely to work. * Lets set things up as if * we are heading back here again. */ setval = rack->r_ctl.rc_entry_gp_rtt; } else if (sum >= 15) { /* * Lets take the smaller of the * two since we are just somewhat * buffered. */ setval = rack->r_ctl.rc_gp_srtt; if (setval > rack->r_ctl.rc_entry_gp_rtt) setval = rack->r_ctl.rc_entry_gp_rtt; } else { /* * Here we are not highly buffered * and should pick the min we can to * keep from causing loss. */ setval = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt); } rack_set_prtt_target(rack, segsiz, setval); } if (rack_probe_rtt_sets_cwnd > 1) { /* There is a percentage here to boost */ ebdp = rack->r_ctl.rc_target_probertt_flight; ebdp *= rack_probe_rtt_sets_cwnd; ebdp /= 100; setto = rack->r_ctl.rc_target_probertt_flight + ebdp; } else setto = rack->r_ctl.rc_target_probertt_flight; rack->rc_tp->snd_cwnd = roundup(setto, segsiz); if (rack->rc_tp->snd_cwnd < (segsiz * rack_timely_min_segs)) { /* Enforce a min */ rack->rc_tp->snd_cwnd = segsiz * rack_timely_min_segs; } /* If we set in the cwnd also set the ssthresh point so we are in CA */ rack->rc_tp->snd_ssthresh = (rack->rc_tp->snd_cwnd - 1); } rack_log_rtt_shrinks(rack, us_cts, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), __LINE__, RACK_RTTS_EXITPROBE); /* Clear times last so log has all the info */ rack->r_ctl.rc_probertt_sndmax_atexit = rack->rc_tp->snd_max; rack->r_ctl.rc_time_probertt_entered = us_cts; rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts = us_cts; rack->r_ctl.rc_time_of_last_probertt = us_cts; } static void rack_check_probe_rtt(struct tcp_rack *rack, uint32_t us_cts) { /* Check in on probe-rtt */ if (rack->rc_gp_filled == 0) { /* We do not do p-rtt unless we have gp measurements */ return; } if (rack->in_probe_rtt) { uint64_t no_overflow; uint32_t endtime, must_stay; if (rack->r_ctl.rc_went_idle_time && ((us_cts - rack->r_ctl.rc_went_idle_time) > rack_min_probertt_hold)) { /* * We went idle during prtt, just exit now. */ rack_exit_probertt(rack, us_cts); } else if (rack_probe_rtt_safety_val && TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered) && ((us_cts - rack->r_ctl.rc_time_probertt_entered) > rack_probe_rtt_safety_val)) { /* * Probe RTT safety value triggered! */ rack_log_rtt_shrinks(rack, us_cts, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), __LINE__, RACK_RTTS_SAFETY); rack_exit_probertt(rack, us_cts); } /* Calculate the max we will wait */ endtime = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_max_drain_wait); if (rack->rc_highly_buffered) endtime += (rack->r_ctl.rc_gp_srtt * rack_max_drain_hbp); /* Calculate the min we must wait */ must_stay = rack->r_ctl.rc_time_probertt_entered + (rack->r_ctl.rc_gp_srtt * rack_must_drain); if ((ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.rc_target_probertt_flight) && TSTMP_LT(us_cts, endtime)) { uint32_t calc; /* Do we lower more? */ no_exit: if (TSTMP_GT(us_cts, rack->r_ctl.rc_time_probertt_entered)) calc = us_cts - rack->r_ctl.rc_time_probertt_entered; else calc = 0; calc /= max(rack->r_ctl.rc_gp_srtt, 1); if (calc) { /* Maybe */ calc *= rack_per_of_gp_probertt_reduce; rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt - calc; /* Limit it too */ if (rack->r_ctl.rack_per_of_gp_probertt < rack_per_of_gp_lowthresh) rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_lowthresh; } /* We must reach target or the time set */ return; } if (rack->r_ctl.rc_time_probertt_starts == 0) { if ((TSTMP_LT(us_cts, must_stay) && rack->rc_highly_buffered) || (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.rc_target_probertt_flight)) { /* We are not past the must_stay time */ goto no_exit; } rack_log_rtt_shrinks(rack, us_cts, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), __LINE__, RACK_RTTS_REACHTARGET); rack->r_ctl.rc_time_probertt_starts = us_cts; if (rack->r_ctl.rc_time_probertt_starts == 0) rack->r_ctl.rc_time_probertt_starts = 1; /* Restore back to our rate we want to pace at in prtt */ rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt; } /* * Setup our end time, some number of gp_srtts plus 200ms. */ no_overflow = ((uint64_t)rack->r_ctl.rc_gp_srtt * (uint64_t)rack_probertt_gpsrtt_cnt_mul); if (rack_probertt_gpsrtt_cnt_div) endtime = (uint32_t)(no_overflow / (uint64_t)rack_probertt_gpsrtt_cnt_div); else endtime = 0; endtime += rack_min_probertt_hold; endtime += rack->r_ctl.rc_time_probertt_starts; if (TSTMP_GEQ(us_cts, endtime)) { /* yes, exit probertt */ rack_exit_probertt(rack, us_cts); } } else if ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= rack_time_between_probertt) { /* Go into probertt, its been too long since we went lower */ rack_enter_probertt(rack, us_cts); } } static void rack_update_multiplier(struct tcp_rack *rack, int32_t timely_says, uint64_t last_bw_est, uint32_t rtt, int32_t rtt_diff) { uint64_t cur_bw, up_bnd, low_bnd, subfr; uint32_t losses; if ((rack->rc_gp_dyn_mul == 0) || (rack->use_fixed_rate) || (rack->in_probe_rtt) || (rack->rc_always_pace == 0)) { /* No dynamic GP multiplier in play */ return; } losses = rack->r_ctl.rc_loss_count - rack->r_ctl.rc_loss_at_start; cur_bw = rack_get_bw(rack); /* Calculate our up and down range */ up_bnd = rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_up; up_bnd /= 100; up_bnd += rack->r_ctl.last_gp_comp_bw; subfr = (uint64_t)rack->r_ctl.last_gp_comp_bw * (uint64_t)rack_gp_per_bw_mul_down; subfr /= 100; low_bnd = rack->r_ctl.last_gp_comp_bw - subfr; if ((timely_says == 2) && (rack->r_ctl.rc_no_push_at_mrtt)) { /* * This is the case where our RTT is above * the max target and we have been configured * to just do timely no bonus up stuff in that case. * * There are two configurations, set to 1, and we * just do timely if we are over our max. If its * set above 1 then we slam the multipliers down * to 100 and then decrement per timely. */ rack_log_timely(rack, timely_says, cur_bw, low_bnd, up_bnd, __LINE__, 3); if (rack->r_ctl.rc_no_push_at_mrtt > 1) rack_validate_multipliers_at_or_below_100(rack); rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff); } else if ((last_bw_est < low_bnd) && !losses) { /* * We are decreasing this is a bit complicated this * means we are loosing ground. This could be * because another flow entered and we are competing * for b/w with it. This will push the RTT up which * makes timely unusable unless we want to get shoved * into a corner and just be backed off (the age * old problem with delay based CC). * * On the other hand if it was a route change we * would like to stay somewhat contained and not * blow out the buffers. */ rack_log_timely(rack, timely_says, cur_bw, low_bnd, up_bnd, __LINE__, 3); rack->r_ctl.last_gp_comp_bw = cur_bw; if (rack->rc_gp_bwred == 0) { /* Go into reduction counting */ rack->rc_gp_bwred = 1; rack->rc_gp_timely_dec_cnt = 0; } if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) || (timely_says == 0)) { /* * Push another time with a faster pacing * to try to gain back (we include override to * get a full raise factor). */ if ((rack->rc_gp_saw_ca && rack->r_ctl.rack_per_of_gp_ca <= rack_down_raise_thresh) || (rack->rc_gp_saw_ss && rack->r_ctl.rack_per_of_gp_ss <= rack_down_raise_thresh) || (timely_says == 0) || (rack_down_raise_thresh == 0)) { /* * Do an override up in b/w if we were * below the threshold or if the threshold * is zero we always do the raise. */ rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 1); } else { /* Log it stays the same */ rack_log_timely(rack, 0, last_bw_est, low_bnd, 0, __LINE__, 11); } rack->rc_gp_timely_dec_cnt++; /* We are not incrementing really no-count */ rack->rc_gp_incr = 0; rack->rc_gp_timely_inc_cnt = 0; } else { /* * Lets just use the RTT * information and give up * pushing. */ goto use_timely; } } else if ((timely_says != 2) && !losses && (last_bw_est > up_bnd)) { /* * We are increasing b/w lets keep going, updating * our b/w and ignoring any timely input, unless * of course we are at our max raise (if there is one). */ rack_log_timely(rack, timely_says, cur_bw, low_bnd, up_bnd, __LINE__, 3); rack->r_ctl.last_gp_comp_bw = cur_bw; if (rack->rc_gp_saw_ss && rack_per_upper_bound_ss && (rack->r_ctl.rack_per_of_gp_ss == rack_per_upper_bound_ss)) { /* * In cases where we can't go higher * we should just use timely. */ goto use_timely; } if (rack->rc_gp_saw_ca && rack_per_upper_bound_ca && (rack->r_ctl.rack_per_of_gp_ca == rack_per_upper_bound_ca)) { /* * In cases where we can't go higher * we should just use timely. */ goto use_timely; } rack->rc_gp_bwred = 0; rack->rc_gp_timely_dec_cnt = 0; /* You get a set number of pushes if timely is trying to reduce */ if ((rack->rc_gp_incr < rack_timely_max_push_rise) || (timely_says == 0)) { rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0); } else { /* Log it stays the same */ rack_log_timely(rack, 0, last_bw_est, up_bnd, 0, __LINE__, 12); } return; } else { /* * We are staying between the lower and upper range bounds * so use timely to decide. */ rack_log_timely(rack, timely_says, cur_bw, low_bnd, up_bnd, __LINE__, 3); use_timely: if (timely_says) { rack->rc_gp_incr = 0; rack->rc_gp_timely_inc_cnt = 0; if ((rack->rc_gp_timely_dec_cnt < rack_timely_max_push_drop) && !losses && (last_bw_est < low_bnd)) { /* We are loosing ground */ rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0); rack->rc_gp_timely_dec_cnt++; /* We are not incrementing really no-count */ rack->rc_gp_incr = 0; rack->rc_gp_timely_inc_cnt = 0; } else rack_decrease_bw_mul(rack, timely_says, rtt, rtt_diff); } else { rack->rc_gp_bwred = 0; rack->rc_gp_timely_dec_cnt = 0; rack_increase_bw_mul(rack, timely_says, cur_bw, last_bw_est, 0); } } } static int32_t rack_make_timely_judgement(struct tcp_rack *rack, uint32_t rtt, int32_t rtt_diff, uint32_t prev_rtt) { int32_t timely_says; uint64_t log_mult, log_rtt_a_diff; log_rtt_a_diff = rtt; log_rtt_a_diff <<= 32; log_rtt_a_diff |= (uint32_t)rtt_diff; if (rtt >= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul)) { /* Reduce the b/w multiplier */ timely_says = 2; log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_maxmul; log_mult <<= 32; log_mult |= prev_rtt; rack_log_timely(rack, timely_says, log_mult, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 4); } else if (rtt <= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) + ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) / max(rack_gp_rtt_mindiv , 1)))) { /* Increase the b/w multiplier */ log_mult = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) + ((get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack_gp_rtt_minmul) / max(rack_gp_rtt_mindiv , 1)); log_mult <<= 32; log_mult |= prev_rtt; timely_says = 0; rack_log_timely(rack, timely_says, log_mult , get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 5); } else { /* * Use a gradient to find it the timely gradient * is: * grad = rc_rtt_diff / min_rtt; * * anything below or equal to 0 will be * a increase indication. Anything above * zero is a decrease. Note we take care * of the actual gradient calculation * in the reduction (its not needed for * increase). */ log_mult = prev_rtt; if (rtt_diff <= 0) { /* * Rttdiff is less than zero, increase the * b/w multiplier (its 0 or negative) */ timely_says = 0; rack_log_timely(rack, timely_says, log_mult, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 6); } else { /* Reduce the b/w multiplier */ timely_says = 1; rack_log_timely(rack, timely_says, log_mult, get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt), log_rtt_a_diff, __LINE__, 7); } } return (timely_says); } static void rack_do_goodput_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq th_ack, int line, uint8_t quality) { uint64_t tim, bytes_ps, ltim, stim, utim; uint32_t segsiz, bytes, reqbytes, us_cts; int32_t gput, new_rtt_diff, timely_says; uint64_t resid_bw, subpart = 0, addpart = 0, srtt; int did_add = 0; us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time); segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs); if (TSTMP_GEQ(us_cts, tp->gput_ts)) tim = us_cts - tp->gput_ts; else tim = 0; if (rack->r_ctl.rc_gp_cumack_ts > rack->r_ctl.rc_gp_output_ts) stim = rack->r_ctl.rc_gp_cumack_ts - rack->r_ctl.rc_gp_output_ts; else stim = 0; /* * Use the larger of the send time or ack time. This prevents us * from being influenced by ack artifacts to come up with too * high of measurement. Note that since we are spanning over many more * bytes in most of our measurements hopefully that is less likely to * occur. */ if (tim > stim) utim = max(tim, 1); else utim = max(stim, 1); /* Lets get a msec time ltim too for the old stuff */ ltim = max(1, (utim / HPTS_USEC_IN_MSEC)); gput = (((uint64_t) (th_ack - tp->gput_seq)) << 3) / ltim; reqbytes = min(rc_init_window(rack), (MIN_GP_WIN * segsiz)); if ((tim == 0) && (stim == 0)) { /* * Invalid measurement time, maybe * all on one ack/one send? */ bytes = 0; bytes_ps = 0; rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes, 0, 0, 0, 10, __LINE__, NULL, quality); goto skip_measurement; } if (rack->r_ctl.rc_gp_lowrtt == 0xffffffff) { /* We never made a us_rtt measurement? */ bytes = 0; bytes_ps = 0; rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes, 0, 0, 0, 10, __LINE__, NULL, quality); goto skip_measurement; } /* * Calculate the maximum possible b/w this connection * could have. We base our calculation on the lowest * rtt we have seen during the measurement and the * largest rwnd the client has given us in that time. This * forms a BDP that is the maximum that we could ever * get to the client. Anything larger is not valid. * * I originally had code here that rejected measurements * where the time was less than 1/2 the latest us_rtt. * But after thinking on that I realized its wrong since * say you had a 150Mbps or even 1Gbps link, and you * were a long way away.. example I am in Europe (100ms rtt) * talking to my 1Gbps link in S.C. Now measuring say 150,000 * bytes my time would be 1.2ms, and yet my rtt would say * the measurement was invalid the time was < 50ms. The * same thing is true for 150Mb (8ms of time). * * A better way I realized is to look at what the maximum * the connection could possibly do. This is gated on * the lowest RTT we have seen and the highest rwnd. * We should in theory never exceed that, if we are * then something on the path is storing up packets * and then feeding them all at once to our endpoint * messing up our measurement. */ rack->r_ctl.last_max_bw = rack->r_ctl.rc_gp_high_rwnd; rack->r_ctl.last_max_bw *= HPTS_USEC_IN_SEC; rack->r_ctl.last_max_bw /= rack->r_ctl.rc_gp_lowrtt; if (SEQ_LT(th_ack, tp->gput_seq)) { /* No measurement can be made */ bytes = 0; bytes_ps = 0; rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes, 0, 0, 0, 10, __LINE__, NULL, quality); goto skip_measurement; } else bytes = (th_ack - tp->gput_seq); bytes_ps = (uint64_t)bytes; /* * Don't measure a b/w for pacing unless we have gotten at least * an initial windows worth of data in this measurement interval. * * Small numbers of bytes get badly influenced by delayed ack and * other artifacts. Note we take the initial window or our * defined minimum GP (defaulting to 10 which hopefully is the * IW). */ if (rack->rc_gp_filled == 0) { /* * The initial estimate is special. We * have blasted out an IW worth of packets * without a real valid ack ts results. We * then setup the app_limited_needs_set flag, * this should get the first ack in (probably 2 * MSS worth) to be recorded as the timestamp. * We thus allow a smaller number of bytes i.e. * IW - 2MSS. */ reqbytes -= (2 * segsiz); /* Also lets fill previous for our first measurement to be neutral */ rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt; } if ((bytes_ps < reqbytes) || rack->app_limited_needs_set) { rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes, rack->r_ctl.rc_app_limited_cnt, 0, 0, 10, __LINE__, NULL, quality); goto skip_measurement; } /* * We now need to calculate the Timely like status so * we can update (possibly) the b/w multipliers. */ new_rtt_diff = (int32_t)rack->r_ctl.rc_gp_srtt - (int32_t)rack->r_ctl.rc_prev_gp_srtt; if (rack->rc_gp_filled == 0) { /* No previous reading */ rack->r_ctl.rc_rtt_diff = new_rtt_diff; } else { if (rack->measure_saw_probe_rtt == 0) { /* * We don't want a probertt to be counted * since it will be negative incorrectly. We * expect to be reducing the RTT when we * pace at a slower rate. */ rack->r_ctl.rc_rtt_diff -= (rack->r_ctl.rc_rtt_diff / 8); rack->r_ctl.rc_rtt_diff += (new_rtt_diff / 8); } } timely_says = rack_make_timely_judgement(rack, rack->r_ctl.rc_gp_srtt, rack->r_ctl.rc_rtt_diff, rack->r_ctl.rc_prev_gp_srtt ); bytes_ps *= HPTS_USEC_IN_SEC; bytes_ps /= utim; if (bytes_ps > rack->r_ctl.last_max_bw) { /* * Something is on path playing * since this b/w is not possible based * on our BDP (highest rwnd and lowest rtt * we saw in the measurement window). * * Another option here would be to * instead skip the measurement. */ rack_log_pacing_delay_calc(rack, bytes, reqbytes, bytes_ps, rack->r_ctl.last_max_bw, 0, 11, __LINE__, NULL, quality); bytes_ps = rack->r_ctl.last_max_bw; } /* We store gp for b/w in bytes per second */ if (rack->rc_gp_filled == 0) { /* Initial measurement */ if (bytes_ps) { rack->r_ctl.gp_bw = bytes_ps; rack->rc_gp_filled = 1; rack->r_ctl.num_measurements = 1; rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL); } else { rack_log_pacing_delay_calc(rack, bytes_ps, reqbytes, rack->r_ctl.rc_app_limited_cnt, 0, 0, 10, __LINE__, NULL, quality); } if (tcp_in_hpts(rack->rc_inp) && (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { /* * Ok we can't trust the pacer in this case * where we transition from un-paced to paced. * Or for that matter when the burst mitigation * was making a wild guess and got it wrong. * Stop the pacer and clear up all the aggregate * delays etc. */ tcp_hpts_remove(rack->rc_inp); rack->r_ctl.rc_hpts_flags = 0; rack->r_ctl.rc_last_output_to = 0; } did_add = 2; } else if (rack->r_ctl.num_measurements < RACK_REQ_AVG) { /* Still a small number run an average */ rack->r_ctl.gp_bw += bytes_ps; addpart = rack->r_ctl.num_measurements; rack->r_ctl.num_measurements++; if (rack->r_ctl.num_measurements >= RACK_REQ_AVG) { /* We have collected enough to move forward */ rack->r_ctl.gp_bw /= (uint64_t)rack->r_ctl.num_measurements; } did_add = 3; } else { /* * We want to take 1/wma of the goodput and add in to 7/8th * of the old value weighted by the srtt. So if your measurement * period is say 2 SRTT's long you would get 1/4 as the * value, if it was like 1/2 SRTT then you would get 1/16th. * * But we must be careful not to take too much i.e. if the * srtt is say 20ms and the measurement is taken over * 400ms our weight would be 400/20 i.e. 20. On the * other hand if we get a measurement over 1ms with a * 10ms rtt we only want to take a much smaller portion. */ if (rack->r_ctl.num_measurements < 0xff) { rack->r_ctl.num_measurements++; } srtt = (uint64_t)tp->t_srtt; if (srtt == 0) { /* * Strange why did t_srtt go back to zero? */ if (rack->r_ctl.rc_rack_min_rtt) srtt = rack->r_ctl.rc_rack_min_rtt; else srtt = HPTS_USEC_IN_MSEC; } /* * XXXrrs: Note for reviewers, in playing with * dynamic pacing I discovered this GP calculation * as done originally leads to some undesired results. * Basically you can get longer measurements contributing * too much to the WMA. Thus I changed it if you are doing * dynamic adjustments to only do the aportioned adjustment * if we have a very small (time wise) measurement. Longer * measurements just get there weight (defaulting to 1/8) * add to the WMA. We may want to think about changing * this to always do that for both sides i.e. dynamic * and non-dynamic... but considering lots of folks * were playing with this I did not want to change the * calculation per.se. without your thoughts.. Lawerence? * Peter?? */ if (rack->rc_gp_dyn_mul == 0) { subpart = rack->r_ctl.gp_bw * utim; subpart /= (srtt * 8); if (subpart < (rack->r_ctl.gp_bw / 2)) { /* * The b/w update takes no more * away then 1/2 our running total * so factor it in. */ addpart = bytes_ps * utim; addpart /= (srtt * 8); } else { /* * Don't allow a single measurement * to account for more than 1/2 of the * WMA. This could happen on a retransmission * where utim becomes huge compared to * srtt (multiple retransmissions when using * the sending rate which factors in all the * transmissions from the first one). */ subpart = rack->r_ctl.gp_bw / 2; addpart = bytes_ps / 2; } resid_bw = rack->r_ctl.gp_bw - subpart; rack->r_ctl.gp_bw = resid_bw + addpart; did_add = 1; } else { if ((utim / srtt) <= 1) { /* * The b/w update was over a small period * of time. The idea here is to prevent a small * measurement time period from counting * too much. So we scale it based on the * time so it attributes less than 1/rack_wma_divisor * of its measurement. */ subpart = rack->r_ctl.gp_bw * utim; subpart /= (srtt * rack_wma_divisor); addpart = bytes_ps * utim; addpart /= (srtt * rack_wma_divisor); } else { /* * The scaled measurement was long * enough so lets just add in the * portion of the measurement i.e. 1/rack_wma_divisor */ subpart = rack->r_ctl.gp_bw / rack_wma_divisor; addpart = bytes_ps / rack_wma_divisor; } if ((rack->measure_saw_probe_rtt == 0) || (bytes_ps > rack->r_ctl.gp_bw)) { /* * For probe-rtt we only add it in * if its larger, all others we just * add in. */ did_add = 1; resid_bw = rack->r_ctl.gp_bw - subpart; rack->r_ctl.gp_bw = resid_bw + addpart; } } } if ((rack->gp_ready == 0) && (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) { /* We have enough measurements now */ rack->gp_ready = 1; rack_set_cc_pacing(rack); if (rack->defer_options) rack_apply_deferred_options(rack); } rack_log_pacing_delay_calc(rack, subpart, addpart, bytes_ps, stim, rack_get_bw(rack), 22, did_add, NULL, quality); /* We do not update any multipliers if we are in or have seen a probe-rtt */ if ((rack->measure_saw_probe_rtt == 0) && rack->rc_gp_rtt_set) rack_update_multiplier(rack, timely_says, bytes_ps, rack->r_ctl.rc_gp_srtt, rack->r_ctl.rc_rtt_diff); rack_log_pacing_delay_calc(rack, bytes, tim, bytes_ps, stim, rack_get_bw(rack), 3, line, NULL, quality); /* reset the gp srtt and setup the new prev */ rack->r_ctl.rc_prev_gp_srtt = rack->r_ctl.rc_gp_srtt; /* Record the lost count for the next measurement */ rack->r_ctl.rc_loss_at_start = rack->r_ctl.rc_loss_count; /* * We restart our diffs based on the gpsrtt in the * measurement window. */ rack->rc_gp_rtt_set = 0; rack->rc_gp_saw_rec = 0; rack->rc_gp_saw_ca = 0; rack->rc_gp_saw_ss = 0; rack->rc_dragged_bottom = 0; skip_measurement: #ifdef STATS stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_GPUT, gput); /* * XXXLAS: This is a temporary hack, and should be * chained off VOI_TCP_GPUT when stats(9) grows an * API to deal with chained VOIs. */ if (tp->t_stats_gput_prev > 0) stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_GPUT_ND, ((gput - tp->t_stats_gput_prev) * 100) / tp->t_stats_gput_prev); #endif tp->t_flags &= ~TF_GPUTINPROG; tp->t_stats_gput_prev = gput; /* * Now are we app limited now and there is space from where we * were to where we want to go? * * We don't do the other case i.e. non-applimited here since * the next send will trigger us picking up the missing data. */ if (rack->r_ctl.rc_first_appl && TCPS_HAVEESTABLISHED(tp->t_state) && rack->r_ctl.rc_app_limited_cnt && (SEQ_GT(rack->r_ctl.rc_first_appl->r_start, th_ack)) && ((rack->r_ctl.rc_first_appl->r_end - th_ack) > max(rc_init_window(rack), (MIN_GP_WIN * segsiz)))) { /* * Yep there is enough outstanding to make a measurement here. */ struct rack_sendmap *rsm, fe; rack->r_ctl.rc_gp_lowrtt = 0xffffffff; rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd; tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time); rack->app_limited_needs_set = 0; tp->gput_seq = th_ack; if (rack->in_probe_rtt) rack->measure_saw_probe_rtt = 1; else if ((rack->measure_saw_probe_rtt) && (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit))) rack->measure_saw_probe_rtt = 0; if ((rack->r_ctl.rc_first_appl->r_end - th_ack) >= rack_get_measure_window(tp, rack)) { /* There is a full window to gain info from */ tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack); } else { /* We can only measure up to the applimited point */ tp->gput_ack = tp->gput_seq + (rack->r_ctl.rc_first_appl->r_end - th_ack); if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) { /* * We don't have enough to make a measurement. */ tp->t_flags &= ~TF_GPUTINPROG; rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq, 0, 0, 0, 6, __LINE__, NULL, quality); return; } } if (tp->t_state >= TCPS_FIN_WAIT_1) { /* * We will get no more data into the SB * this means we need to have the data available * before we start a measurement. */ if (sbavail(&tp->t_inpcb->inp_socket->so_snd) < (tp->gput_ack - tp->gput_seq)) { /* Nope not enough data. */ return; } } tp->t_flags |= TF_GPUTINPROG; /* * Now we need to find the timestamp of the send at tp->gput_seq * for the send based measurement. */ fe.r_start = tp->gput_seq; rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe); if (rsm) { /* Ok send-based limit is set */ if (SEQ_LT(rsm->r_start, tp->gput_seq)) { /* * Move back to include the earlier part * so our ack time lines up right (this may * make an overlapping measurement but thats * ok). */ tp->gput_seq = rsm->r_start; } if (rsm->r_flags & RACK_ACKED) tp->gput_ts = (uint32_t)rsm->r_ack_arrival; else rack->app_limited_needs_set = 1; rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]; } else { /* * If we don't find the rsm due to some * send-limit set the current time, which * basically disables the send-limit. */ struct timeval tv; microuptime(&tv); rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv); } rack_log_pacing_delay_calc(rack, tp->gput_seq, tp->gput_ack, (uint64_t)rsm, tp->gput_ts, rack->r_ctl.rc_app_limited_cnt, 9, __LINE__, NULL, quality); } } /* * CC wrapper hook functions */ static void rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, uint32_t th_ack, uint16_t nsegs, uint16_t type, int32_t recovery) { uint32_t prior_cwnd, acked; struct tcp_log_buffer *lgb = NULL; uint8_t labc_to_use, quality; INP_WLOCK_ASSERT(tp->t_inpcb); tp->ccv->nsegs = nsegs; acked = tp->ccv->bytes_this_ack = (th_ack - tp->snd_una); if ((recovery) && (rack->r_ctl.rc_early_recovery_segs)) { uint32_t max; max = rack->r_ctl.rc_early_recovery_segs * ctf_fixed_maxseg(tp); if (tp->ccv->bytes_this_ack > max) { tp->ccv->bytes_this_ack = max; } } #ifdef STATS stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF, ((int32_t)rack->r_ctl.cwnd_to_use) - tp->snd_wnd); #endif quality = RACK_QUALITY_NONE; if ((tp->t_flags & TF_GPUTINPROG) && rack_enough_for_measurement(tp, rack, th_ack, &quality)) { /* Measure the Goodput */ rack_do_goodput_measurement(tp, rack, th_ack, __LINE__, quality); #ifdef NETFLIX_PEAKRATE if ((type == CC_ACK) && (tp->t_maxpeakrate)) { /* * We update t_peakrate_thr. This gives us roughly * one update per round trip time. Note * it will only be used if pace_always is off i.e * we don't do this for paced flows. */ rack_update_peakrate_thr(tp); } #endif } /* Which way our we limited, if not cwnd limited no advance in CA */ if (tp->snd_cwnd <= tp->snd_wnd) tp->ccv->flags |= CCF_CWND_LIMITED; else tp->ccv->flags &= ~CCF_CWND_LIMITED; if (tp->snd_cwnd > tp->snd_ssthresh) { tp->t_bytes_acked += min(tp->ccv->bytes_this_ack, nsegs * V_tcp_abc_l_var * ctf_fixed_maxseg(tp)); /* For the setting of a window past use the actual scwnd we are using */ if (tp->t_bytes_acked >= rack->r_ctl.cwnd_to_use) { tp->t_bytes_acked -= rack->r_ctl.cwnd_to_use; tp->ccv->flags |= CCF_ABC_SENTAWND; } } else { tp->ccv->flags &= ~CCF_ABC_SENTAWND; tp->t_bytes_acked = 0; } prior_cwnd = tp->snd_cwnd; if ((recovery == 0) || (rack_max_abc_post_recovery == 0) || rack->r_use_labc_for_rec || (rack_client_low_buf && (rack->client_bufferlvl < rack_client_low_buf))) labc_to_use = rack->rc_labc; else labc_to_use = rack_max_abc_post_recovery; if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.flex1 = th_ack; log.u_bbr.flex2 = tp->ccv->flags; log.u_bbr.flex3 = tp->ccv->bytes_this_ack; log.u_bbr.flex4 = tp->ccv->nsegs; log.u_bbr.flex5 = labc_to_use; log.u_bbr.flex6 = prior_cwnd; log.u_bbr.flex7 = V_tcp_do_newsack; log.u_bbr.flex8 = 1; lgb = tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0, 0, &log, false, NULL, NULL, 0, &tv); } if (CC_ALGO(tp)->ack_received != NULL) { /* XXXLAS: Find a way to live without this */ tp->ccv->curack = th_ack; tp->ccv->labc = labc_to_use; tp->ccv->flags |= CCF_USE_LOCAL_ABC; CC_ALGO(tp)->ack_received(tp->ccv, type); } if (lgb) { lgb->tlb_stackinfo.u_bbr.flex6 = tp->snd_cwnd; } if (rack->r_must_retran) { if (SEQ_GEQ(th_ack, rack->r_ctl.rc_snd_max_at_rto)) { /* * We now are beyond the rxt point so lets disable * the flag. */ rack->r_ctl.rc_out_at_rto = 0; rack->r_must_retran = 0; } else if ((prior_cwnd + ctf_fixed_maxseg(tp)) <= tp->snd_cwnd) { /* * Only decrement the rc_out_at_rto if the cwnd advances * at least a whole segment. Otherwise next time the peer * acks, we won't be able to send this generaly happens * when we are in Congestion Avoidance. */ if (acked <= rack->r_ctl.rc_out_at_rto){ rack->r_ctl.rc_out_at_rto -= acked; } else { rack->r_ctl.rc_out_at_rto = 0; } } } #ifdef STATS stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, rack->r_ctl.cwnd_to_use); #endif if (rack->r_ctl.rc_rack_largest_cwnd < rack->r_ctl.cwnd_to_use) { rack->r_ctl.rc_rack_largest_cwnd = rack->r_ctl.cwnd_to_use; } #ifdef NETFLIX_PEAKRATE /* we enforce max peak rate if it is set and we are not pacing */ if ((rack->rc_always_pace == 0) && tp->t_peakrate_thr && (tp->snd_cwnd > tp->t_peakrate_thr)) { tp->snd_cwnd = tp->t_peakrate_thr; } #endif } static void tcp_rack_partialack(struct tcpcb *tp) { struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; INP_WLOCK_ASSERT(tp->t_inpcb); /* * If we are doing PRR and have enough * room to send we are pacing and prr * is disabled we will want to see if we * can send data (by setting r_wanted_output to * true). */ if ((rack->r_ctl.rc_prr_sndcnt > 0) || rack->rack_no_prr) rack->r_wanted_output = 1; } static void rack_post_recovery(struct tcpcb *tp, uint32_t th_ack) { struct tcp_rack *rack; uint32_t orig_cwnd; orig_cwnd = tp->snd_cwnd; INP_WLOCK_ASSERT(tp->t_inpcb); rack = (struct tcp_rack *)tp->t_fb_ptr; /* only alert CC if we alerted when we entered */ if (CC_ALGO(tp)->post_recovery != NULL) { tp->ccv->curack = th_ack; CC_ALGO(tp)->post_recovery(tp->ccv); if (tp->snd_cwnd < tp->snd_ssthresh) { /* * Rack has burst control and pacing * so lets not set this any lower than * snd_ssthresh per RFC-6582 (option 2). */ tp->snd_cwnd = tp->snd_ssthresh; } } if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.flex1 = th_ack; log.u_bbr.flex2 = tp->ccv->flags; log.u_bbr.flex3 = tp->ccv->bytes_this_ack; log.u_bbr.flex4 = tp->ccv->nsegs; log.u_bbr.flex5 = V_tcp_abc_l_var; log.u_bbr.flex6 = orig_cwnd; log.u_bbr.flex7 = V_tcp_do_newsack; log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.flex8 = 2; tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0, 0, &log, false, NULL, NULL, 0, &tv); } if ((rack->rack_no_prr == 0) && (rack->no_prr_addback == 0) && (rack->r_ctl.rc_prr_sndcnt > 0)) { /* * Suck the next prr cnt back into cwnd, but * only do that if we are not application limited. */ if (ctf_outstanding(tp) <= sbavail(&(tp->t_inpcb->inp_socket->so_snd))) { /* * We are allowed to add back to the cwnd the amount we did * not get out if: * a) no_prr_addback is off. * b) we are not app limited * c) we are doing prr * * d) it is bounded by rack_prr_addbackmax (if addback is 0, then none). */ tp->snd_cwnd += min((ctf_fixed_maxseg(tp) * rack_prr_addbackmax), rack->r_ctl.rc_prr_sndcnt); } rack->r_ctl.rc_prr_sndcnt = 0; rack_log_to_prr(rack, 1, 0, __LINE__); } rack_log_to_prr(rack, 14, orig_cwnd, __LINE__); tp->snd_recover = tp->snd_una; if (rack->r_ctl.dsack_persist) { rack->r_ctl.dsack_persist--; if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) { rack->r_ctl.num_dsack = 0; } rack_log_dsack_event(rack, 1, __LINE__, 0, 0); } EXIT_RECOVERY(tp->t_flags); } static void rack_cong_signal(struct tcpcb *tp, uint32_t type, uint32_t ack, int line) { struct tcp_rack *rack; uint32_t ssthresh_enter, cwnd_enter, in_rec_at_entry, orig_cwnd; INP_WLOCK_ASSERT(tp->t_inpcb); #ifdef STATS stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type); #endif if (IN_RECOVERY(tp->t_flags) == 0) { in_rec_at_entry = 0; ssthresh_enter = tp->snd_ssthresh; cwnd_enter = tp->snd_cwnd; } else in_rec_at_entry = 1; rack = (struct tcp_rack *)tp->t_fb_ptr; switch (type) { case CC_NDUPACK: tp->t_flags &= ~TF_WASFRECOVERY; tp->t_flags &= ~TF_WASCRECOVERY; if (!IN_FASTRECOVERY(tp->t_flags)) { rack->r_ctl.rc_prr_delivered = 0; rack->r_ctl.rc_prr_out = 0; if (rack->rack_no_prr == 0) { rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp); rack_log_to_prr(rack, 2, in_rec_at_entry, line); } rack->r_ctl.rc_prr_recovery_fs = tp->snd_max - tp->snd_una; tp->snd_recover = tp->snd_max; if (tp->t_flags2 & TF2_ECN_PERMIT) tp->t_flags2 |= TF2_ECN_SND_CWR; } break; case CC_ECN: if (!IN_CONGRECOVERY(tp->t_flags) || /* * Allow ECN reaction on ACK to CWR, if * that data segment was also CE marked. */ SEQ_GEQ(ack, tp->snd_recover)) { EXIT_CONGRECOVERY(tp->t_flags); KMOD_TCPSTAT_INC(tcps_ecn_rcwnd); tp->snd_recover = tp->snd_max + 1; if (tp->t_flags2 & TF2_ECN_PERMIT) tp->t_flags2 |= TF2_ECN_SND_CWR; } break; case CC_RTO: tp->t_dupacks = 0; tp->t_bytes_acked = 0; EXIT_RECOVERY(tp->t_flags); tp->snd_ssthresh = max(2, min(tp->snd_wnd, rack->r_ctl.cwnd_to_use) / 2 / ctf_fixed_maxseg(tp)) * ctf_fixed_maxseg(tp); orig_cwnd = tp->snd_cwnd; tp->snd_cwnd = ctf_fixed_maxseg(tp); rack_log_to_prr(rack, 16, orig_cwnd, line); if (tp->t_flags2 & TF2_ECN_PERMIT) tp->t_flags2 |= TF2_ECN_SND_CWR; break; case CC_RTO_ERR: KMOD_TCPSTAT_INC(tcps_sndrexmitbad); /* RTO was unnecessary, so reset everything. */ tp->snd_cwnd = tp->snd_cwnd_prev; tp->snd_ssthresh = tp->snd_ssthresh_prev; tp->snd_recover = tp->snd_recover_prev; if (tp->t_flags & TF_WASFRECOVERY) { ENTER_FASTRECOVERY(tp->t_flags); tp->t_flags &= ~TF_WASFRECOVERY; } if (tp->t_flags & TF_WASCRECOVERY) { ENTER_CONGRECOVERY(tp->t_flags); tp->t_flags &= ~TF_WASCRECOVERY; } tp->snd_nxt = tp->snd_max; tp->t_badrxtwin = 0; break; } if ((CC_ALGO(tp)->cong_signal != NULL) && (type != CC_RTO)){ tp->ccv->curack = ack; CC_ALGO(tp)->cong_signal(tp->ccv, type); } if ((in_rec_at_entry == 0) && IN_RECOVERY(tp->t_flags)) { rack_log_to_prr(rack, 15, cwnd_enter, line); rack->r_ctl.dsack_byte_cnt = 0; rack->r_ctl.retran_during_recovery = 0; rack->r_ctl.rc_cwnd_at_erec = cwnd_enter; rack->r_ctl.rc_ssthresh_at_erec = ssthresh_enter; rack->r_ent_rec_ns = 1; } } static inline void rack_cc_after_idle(struct tcp_rack *rack, struct tcpcb *tp) { uint32_t i_cwnd; INP_WLOCK_ASSERT(tp->t_inpcb); #ifdef NETFLIX_STATS KMOD_TCPSTAT_INC(tcps_idle_restarts); if (tp->t_state == TCPS_ESTABLISHED) KMOD_TCPSTAT_INC(tcps_idle_estrestarts); #endif if (CC_ALGO(tp)->after_idle != NULL) CC_ALGO(tp)->after_idle(tp->ccv); if (tp->snd_cwnd == 1) i_cwnd = tp->t_maxseg; /* SYN(-ACK) lost */ else i_cwnd = rc_init_window(rack); /* * Being idle is no different than the initial window. If the cc * clamps it down below the initial window raise it to the initial * window. */ if (tp->snd_cwnd < i_cwnd) { tp->snd_cwnd = i_cwnd; } } /* * Indicate whether this ack should be delayed. We can delay the ack if * following conditions are met: * - There is no delayed ack timer in progress. * - Our last ack wasn't a 0-sized window. We never want to delay * the ack that opens up a 0-sized window. * - LRO wasn't used for this segment. We make sure by checking that the * segment size is not larger than the MSS. * - Delayed acks are enabled or this is a half-synchronized T/TCP * connection. */ #define DELAY_ACK(tp, tlen) \ (((tp->t_flags & TF_RXWIN0SENT) == 0) && \ ((tp->t_flags & TF_DELACK) == 0) && \ (tlen <= tp->t_maxseg) && \ (tp->t_delayed_ack || (tp->t_flags & TF_NEEDSYN))) static struct rack_sendmap * rack_find_lowest_rsm(struct tcp_rack *rack) { struct rack_sendmap *rsm; /* * Walk the time-order transmitted list looking for an rsm that is * not acked. This will be the one that was sent the longest time * ago that is still outstanding. */ TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) { if (rsm->r_flags & RACK_ACKED) { continue; } goto finish; } finish: return (rsm); } static struct rack_sendmap * rack_find_high_nonack(struct tcp_rack *rack, struct rack_sendmap *rsm) { struct rack_sendmap *prsm; /* * Walk the sequence order list backward until we hit and arrive at * the highest seq not acked. In theory when this is called it * should be the last segment (which it was not). */ prsm = rsm; RB_FOREACH_REVERSE_FROM(prsm, rack_rb_tree_head, rsm) { if (prsm->r_flags & (RACK_ACKED | RACK_HAS_FIN)) { continue; } return (prsm); } return (NULL); } static uint32_t rack_calc_thresh_rack(struct tcp_rack *rack, uint32_t srtt, uint32_t cts) { int32_t lro; uint32_t thresh; /* * lro is the flag we use to determine if we have seen reordering. * If it gets set we have seen reordering. The reorder logic either * works in one of two ways: * * If reorder-fade is configured, then we track the last time we saw * re-ordering occur. If we reach the point where enough time as * passed we no longer consider reordering has occuring. * * Or if reorder-face is 0, then once we see reordering we consider * the connection to alway be subject to reordering and just set lro * to 1. * * In the end if lro is non-zero we add the extra time for * reordering in. */ if (srtt == 0) srtt = 1; if (rack->r_ctl.rc_reorder_ts) { if (rack->r_ctl.rc_reorder_fade) { if (SEQ_GEQ(cts, rack->r_ctl.rc_reorder_ts)) { lro = cts - rack->r_ctl.rc_reorder_ts; if (lro == 0) { /* * No time as passed since the last * reorder, mark it as reordering. */ lro = 1; } } else { /* Negative time? */ lro = 0; } if (lro > rack->r_ctl.rc_reorder_fade) { /* Turn off reordering seen too */ rack->r_ctl.rc_reorder_ts = 0; lro = 0; } } else { /* Reodering does not fade */ lro = 1; } } else { lro = 0; } if (rack->rc_rack_tmr_std_based == 0) { thresh = srtt + rack->r_ctl.rc_pkt_delay; } else { /* Standards based pkt-delay is 1/4 srtt */ thresh = srtt + (srtt >> 2); } if (lro && (rack->rc_rack_tmr_std_based == 0)) { /* It must be set, if not you get 1/4 rtt */ if (rack->r_ctl.rc_reorder_shift) thresh += (srtt >> rack->r_ctl.rc_reorder_shift); else thresh += (srtt >> 2); } if (rack->rc_rack_use_dsack && lro && (rack->r_ctl.num_dsack > 0)) { /* * We only increase the reordering window if we * have seen reordering we have a DSACK count. */ thresh += rack->r_ctl.num_dsack * (srtt >> 2); rack_log_dsack_event(rack, 4, __LINE__, srtt, thresh); } /* SRTT * 2 is the ceiling */ if (thresh > (srtt * 2)) { thresh = srtt * 2; } /* And we don't want it above the RTO max either */ if (thresh > rack_rto_max) { thresh = rack_rto_max; } rack_log_dsack_event(rack, 6, __LINE__, srtt, thresh); return (thresh); } static uint32_t rack_calc_thresh_tlp(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t srtt) { struct rack_sendmap *prsm; uint32_t thresh, len; int segsiz; if (srtt == 0) srtt = 1; if (rack->r_ctl.rc_tlp_threshold) thresh = srtt + (srtt / rack->r_ctl.rc_tlp_threshold); else thresh = (srtt * 2); /* Get the previous sent packet, if any */ segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs); len = rsm->r_end - rsm->r_start; if (rack->rack_tlp_threshold_use == TLP_USE_ID) { /* Exactly like the ID */ if (((tp->snd_max - tp->snd_una) - rack->r_ctl.rc_sacked + rack->r_ctl.rc_holes_rxt) <= segsiz) { uint32_t alt_thresh; /* * Compensate for delayed-ack with the d-ack time. */ alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time; if (alt_thresh > thresh) thresh = alt_thresh; } } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_ONE) { /* 2.1 behavior */ prsm = TAILQ_PREV(rsm, rack_head, r_tnext); if (prsm && (len <= segsiz)) { /* * Two packets outstanding, thresh should be (2*srtt) + * possible inter-packet delay (if any). */ uint32_t inter_gap = 0; int idx, nidx; idx = rsm->r_rtr_cnt - 1; nidx = prsm->r_rtr_cnt - 1; if (rsm->r_tim_lastsent[nidx] >= prsm->r_tim_lastsent[idx]) { /* Yes it was sent later (or at the same time) */ inter_gap = rsm->r_tim_lastsent[idx] - prsm->r_tim_lastsent[nidx]; } thresh += inter_gap; } else if (len <= segsiz) { /* * Possibly compensate for delayed-ack. */ uint32_t alt_thresh; alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time; if (alt_thresh > thresh) thresh = alt_thresh; } } else if (rack->rack_tlp_threshold_use == TLP_USE_TWO_TWO) { /* 2.2 behavior */ if (len <= segsiz) { uint32_t alt_thresh; /* * Compensate for delayed-ack with the d-ack time. */ alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time; if (alt_thresh > thresh) thresh = alt_thresh; } } /* Not above an RTO */ if (thresh > tp->t_rxtcur) { thresh = tp->t_rxtcur; } /* Not above a RTO max */ if (thresh > rack_rto_max) { thresh = rack_rto_max; } /* Apply user supplied min TLP */ if (thresh < rack_tlp_min) { thresh = rack_tlp_min; } return (thresh); } static uint32_t rack_grab_rtt(struct tcpcb *tp, struct tcp_rack *rack) { /* * We want the rack_rtt which is the * last rtt we measured. However if that * does not exist we fallback to the srtt (which * we probably will never do) and then as a last * resort we use RACK_INITIAL_RTO if no srtt is * yet set. */ if (rack->rc_rack_rtt) return (rack->rc_rack_rtt); else if (tp->t_srtt == 0) return (RACK_INITIAL_RTO); return (tp->t_srtt); } static struct rack_sendmap * rack_check_recovery_mode(struct tcpcb *tp, uint32_t tsused) { /* * Check to see that we don't need to fall into recovery. We will * need to do so if our oldest transmit is past the time we should * have had an ack. */ struct tcp_rack *rack; struct rack_sendmap *rsm; int32_t idx; uint32_t srtt, thresh; rack = (struct tcp_rack *)tp->t_fb_ptr; if (RB_EMPTY(&rack->r_ctl.rc_mtree)) { return (NULL); } rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); if (rsm == NULL) return (NULL); if (rsm->r_flags & RACK_ACKED) { rsm = rack_find_lowest_rsm(rack); if (rsm == NULL) return (NULL); } idx = rsm->r_rtr_cnt - 1; srtt = rack_grab_rtt(tp, rack); thresh = rack_calc_thresh_rack(rack, srtt, tsused); if (TSTMP_LT(tsused, ((uint32_t)rsm->r_tim_lastsent[idx]))) { return (NULL); } if ((tsused - ((uint32_t)rsm->r_tim_lastsent[idx])) < thresh) { return (NULL); } /* Ok if we reach here we are over-due and this guy can be sent */ rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__); return (rsm); } static uint32_t rack_get_persists_timer_val(struct tcpcb *tp, struct tcp_rack *rack) { int32_t t; int32_t tt; uint32_t ret_val; t = (tp->t_srtt + (tp->t_rttvar << 2)); RACK_TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], rack_persist_min, rack_persist_max, rack->r_ctl.timer_slop); rack->r_ctl.rc_hpts_flags |= PACE_TMR_PERSIT; ret_val = (uint32_t)tt; return (ret_val); } static uint32_t rack_timer_start(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int sup_rack) { /* * Start the FR timer, we do this based on getting the first one in * the rc_tmap. Note that if its NULL we must stop the timer. in all * events we need to stop the running timer (if its running) before * starting the new one. */ uint32_t thresh, exp, to, srtt, time_since_sent, tstmp_touse; uint32_t srtt_cur; int32_t idx; int32_t is_tlp_timer = 0; struct rack_sendmap *rsm; if (rack->t_timers_stopped) { /* All timers have been stopped none are to run */ return (0); } if (rack->rc_in_persist) { /* We can't start any timer in persists */ return (rack_get_persists_timer_val(tp, rack)); } rack->rc_on_min_to = 0; if ((tp->t_state < TCPS_ESTABLISHED) || ((tp->t_flags & TF_SACK_PERMIT) == 0)) { goto activate_rxt; } rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); if ((rsm == NULL) || sup_rack) { /* Nothing on the send map or no rack */ activate_rxt: time_since_sent = 0; rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); if (rsm) { /* * Should we discount the RTX timer any? * * We want to discount it the smallest amount. * If a timer (Rack/TLP or RXT) has gone off more * recently thats the discount we want to use (now - timer time). * If the retransmit of the oldest packet was more recent then * we want to use that (now - oldest-packet-last_transmit_time). * */ idx = rsm->r_rtr_cnt - 1; if (TSTMP_GEQ(rack->r_ctl.rc_tlp_rxt_last_time, ((uint32_t)rsm->r_tim_lastsent[idx]))) tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time; else tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx]; if (TSTMP_GT(cts, tstmp_touse)) time_since_sent = cts - tstmp_touse; } if (SEQ_LT(tp->snd_una, tp->snd_max) || sbavail(&(tp->t_inpcb->inp_socket->so_snd))) { rack->r_ctl.rc_hpts_flags |= PACE_TMR_RXT; to = tp->t_rxtcur; if (to > time_since_sent) to -= time_since_sent; else to = rack->r_ctl.rc_min_to; if (to == 0) to = 1; /* Special case for KEEPINIT */ if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) && (TP_KEEPINIT(tp) != 0) && rsm) { /* * We have to put a ceiling on the rxt timer * of the keep-init timeout. */ uint32_t max_time, red; max_time = TICKS_2_USEC(TP_KEEPINIT(tp)); if (TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) { red = (cts - (uint32_t)rsm->r_tim_lastsent[0]); if (red < max_time) max_time -= red; else max_time = 1; } /* Reduce timeout to the keep value if needed */ if (max_time < to) to = max_time; } return (to); } return (0); } if (rsm->r_flags & RACK_ACKED) { rsm = rack_find_lowest_rsm(rack); if (rsm == NULL) { /* No lowest? */ goto activate_rxt; } } if (rack->sack_attack_disable) { /* * We don't want to do * any TLP's if you are an attacker. * Though if you are doing what * is expected you may still have * SACK-PASSED marks. */ goto activate_rxt; } /* Convert from ms to usecs */ if ((rsm->r_flags & RACK_SACK_PASSED) || (rsm->r_flags & RACK_RWND_COLLAPSED) || (rsm->r_dupack >= DUP_ACK_THRESHOLD)) { if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) == 1) && (rsm->r_flags & RACK_HAS_FIN)) { /* * We don't start a rack timer if all we have is a * FIN outstanding. */ goto activate_rxt; } if ((rack->use_rack_rr == 0) && (IN_FASTRECOVERY(tp->t_flags)) && (rack->rack_no_prr == 0) && (rack->r_ctl.rc_prr_sndcnt < ctf_fixed_maxseg(tp))) { /* * We are not cheating, in recovery and * not enough ack's to yet get our next * retransmission out. * * Note that classified attackers do not * get to use the rack-cheat. */ goto activate_tlp; } srtt = rack_grab_rtt(tp, rack); thresh = rack_calc_thresh_rack(rack, srtt, cts); idx = rsm->r_rtr_cnt - 1; exp = ((uint32_t)rsm->r_tim_lastsent[idx]) + thresh; if (SEQ_GEQ(exp, cts)) { to = exp - cts; if (to < rack->r_ctl.rc_min_to) { to = rack->r_ctl.rc_min_to; if (rack->r_rr_config == 3) rack->rc_on_min_to = 1; } } else { to = rack->r_ctl.rc_min_to; if (rack->r_rr_config == 3) rack->rc_on_min_to = 1; } } else { /* Ok we need to do a TLP not RACK */ activate_tlp: if ((rack->rc_tlp_in_progress != 0) && (rack->r_ctl.rc_tlp_cnt_out >= rack_tlp_limit)) { /* * The previous send was a TLP and we have sent * N TLP's without sending new data. */ goto activate_rxt; } rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext); if (rsm == NULL) { /* We found no rsm to TLP with. */ goto activate_rxt; } if (rsm->r_flags & RACK_HAS_FIN) { /* If its a FIN we dont do TLP */ rsm = NULL; goto activate_rxt; } idx = rsm->r_rtr_cnt - 1; time_since_sent = 0; if (TSTMP_GEQ(((uint32_t)rsm->r_tim_lastsent[idx]), rack->r_ctl.rc_tlp_rxt_last_time)) tstmp_touse = (uint32_t)rsm->r_tim_lastsent[idx]; else tstmp_touse = (uint32_t)rack->r_ctl.rc_tlp_rxt_last_time; if (TSTMP_GT(cts, tstmp_touse)) time_since_sent = cts - tstmp_touse; is_tlp_timer = 1; if (tp->t_srtt) { if ((rack->rc_srtt_measure_made == 0) && (tp->t_srtt == 1)) { /* * If another stack as run and set srtt to 1, * then the srtt was 0, so lets use the initial. */ srtt = RACK_INITIAL_RTO; } else { srtt_cur = tp->t_srtt; srtt = srtt_cur; } } else srtt = RACK_INITIAL_RTO; /* * If the SRTT is not keeping up and the * rack RTT has spiked we want to use * the last RTT not the smoothed one. */ if (rack_tlp_use_greater && tp->t_srtt && (srtt < rack_grab_rtt(tp, rack))) { srtt = rack_grab_rtt(tp, rack); } thresh = rack_calc_thresh_tlp(tp, rack, rsm, srtt); if (thresh > time_since_sent) { to = thresh - time_since_sent; } else { to = rack->r_ctl.rc_min_to; rack_log_alt_to_to_cancel(rack, thresh, /* flex1 */ time_since_sent, /* flex2 */ tstmp_touse, /* flex3 */ rack->r_ctl.rc_tlp_rxt_last_time, /* flex4 */ (uint32_t)rsm->r_tim_lastsent[idx], srtt, idx, 99); } if (to < rack_tlp_min) { to = rack_tlp_min; } if (to > TICKS_2_USEC(TCPTV_REXMTMAX)) { /* * If the TLP time works out to larger than the max * RTO lets not do TLP.. just RTO. */ goto activate_rxt; } } if (is_tlp_timer == 0) { rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK; } else { rack->r_ctl.rc_hpts_flags |= PACE_TMR_TLP; } if (to == 0) to = 1; return (to); } static void rack_enter_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) { if (rack->rc_in_persist == 0) { if (tp->t_flags & TF_GPUTINPROG) { /* * Stop the goodput now, the calling of the * measurement function clears the flag. */ rack_do_goodput_measurement(tp, rack, tp->snd_una, __LINE__, RACK_QUALITY_PERSIST); } #ifdef NETFLIX_SHARED_CWND if (rack->r_ctl.rc_scw) { tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index); rack->rack_scwnd_is_idle = 1; } #endif rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL); if (rack->r_ctl.rc_went_idle_time == 0) rack->r_ctl.rc_went_idle_time = 1; rack_timer_cancel(tp, rack, cts, __LINE__); rack->r_ctl.persist_lost_ends = 0; rack->probe_not_answered = 0; rack->forced_ack = 0; tp->t_rxtshift = 0; RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp), rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop); rack->rc_in_persist = 1; } } static void rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) { if (tcp_in_hpts(rack->rc_inp)) { tcp_hpts_remove(rack->rc_inp); rack->r_ctl.rc_hpts_flags = 0; } #ifdef NETFLIX_SHARED_CWND if (rack->r_ctl.rc_scw) { tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index); rack->rack_scwnd_is_idle = 0; } #endif if (rack->rc_gp_dyn_mul && (rack->use_fixed_rate == 0) && (rack->rc_always_pace)) { /* * Do we count this as if a probe-rtt just * finished? */ uint32_t time_idle, idle_min; time_idle = tcp_get_usecs(NULL) - rack->r_ctl.rc_went_idle_time; idle_min = rack_min_probertt_hold; if (rack_probertt_gpsrtt_cnt_div) { uint64_t extra; extra = (uint64_t)rack->r_ctl.rc_gp_srtt * (uint64_t)rack_probertt_gpsrtt_cnt_mul; extra /= (uint64_t)rack_probertt_gpsrtt_cnt_div; idle_min += (uint32_t)extra; } if (time_idle >= idle_min) { /* Yes, we count it as a probe-rtt. */ uint32_t us_cts; us_cts = tcp_get_usecs(NULL); if (rack->in_probe_rtt == 0) { rack->r_ctl.rc_lower_rtt_us_cts = us_cts; rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts; rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts; rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts; } else { rack_exit_probertt(rack, us_cts); } } } rack->rc_in_persist = 0; rack->r_ctl.rc_went_idle_time = 0; tp->t_rxtshift = 0; RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp), rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop); rack->r_ctl.rc_agg_delayed = 0; rack->r_early = 0; rack->r_late = 0; rack->r_ctl.rc_agg_early = 0; } static void rack_log_hpts_diag(struct tcp_rack *rack, uint32_t cts, struct hpts_diag *diag, struct timeval *tv) { if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.flex1 = diag->p_nxt_slot; log.u_bbr.flex2 = diag->p_cur_slot; log.u_bbr.flex3 = diag->slot_req; log.u_bbr.flex4 = diag->inp_hptsslot; log.u_bbr.flex5 = diag->slot_remaining; log.u_bbr.flex6 = diag->need_new_to; log.u_bbr.flex7 = diag->p_hpts_active; log.u_bbr.flex8 = diag->p_on_min_sleep; /* Hijack other fields as needed */ log.u_bbr.epoch = diag->have_slept; log.u_bbr.lt_epoch = diag->yet_to_sleep; log.u_bbr.pkts_out = diag->co_ret; log.u_bbr.applimited = diag->hpts_sleep_time; log.u_bbr.delivered = diag->p_prev_slot; log.u_bbr.inflight = diag->p_runningslot; log.u_bbr.bw_inuse = diag->wheel_slot; log.u_bbr.rttProp = diag->wheel_cts; log.u_bbr.timeStamp = cts; log.u_bbr.delRate = diag->maxslots; log.u_bbr.cur_del_rate = diag->p_curtick; log.u_bbr.cur_del_rate <<= 32; log.u_bbr.cur_del_rate |= diag->p_lasttick; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_HPTSDIAG, 0, 0, &log, false, tv); } } static void rack_log_wakeup(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb, uint32_t len, int type) { if (rack_verbose_logging && rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.flex1 = sb->sb_flags; log.u_bbr.flex2 = len; log.u_bbr.flex3 = sb->sb_state; log.u_bbr.flex8 = type; log.u_bbr.timeStamp = tcp_get_usecs(&tv); TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, TCP_LOG_SB_WAKE, 0, len, &log, false, &tv); } } static void rack_start_hpts_timer(struct tcp_rack *rack, struct tcpcb *tp, uint32_t cts, int32_t slot, uint32_t tot_len_this_send, int sup_rack) { struct hpts_diag diag; struct inpcb *inp; struct timeval tv; uint32_t delayed_ack = 0; uint32_t hpts_timeout; uint32_t entry_slot = slot; uint8_t stopped; uint32_t left = 0; uint32_t us_cts; inp = tp->t_inpcb; if ((tp->t_state == TCPS_CLOSED) || (tp->t_state == TCPS_LISTEN)) { return; } if (tcp_in_hpts(inp)) { /* Already on the pacer */ return; } stopped = rack->rc_tmr_stopped; if (stopped && TSTMP_GT(rack->r_ctl.rc_timer_exp, cts)) { left = rack->r_ctl.rc_timer_exp - cts; } rack->r_ctl.rc_timer_exp = 0; rack->r_ctl.rc_hpts_flags = 0; us_cts = tcp_get_usecs(&tv); /* Now early/late accounting */ rack_log_pacing_delay_calc(rack, entry_slot, slot, 0, 0, 0, 26, __LINE__, NULL, 0); if (rack->r_early && (rack->rc_ack_can_sendout_data == 0)) { /* * We have a early carry over set, * we can always add more time so we * can always make this compensation. * * Note if ack's are allowed to wake us do not * penalize the next timer for being awoke * by an ack aka the rc_agg_early (non-paced mode). */ slot += rack->r_ctl.rc_agg_early; rack->r_early = 0; rack->r_ctl.rc_agg_early = 0; } if (rack->r_late) { /* * This is harder, we can * compensate some but it * really depends on what * the current pacing time is. */ if (rack->r_ctl.rc_agg_delayed >= slot) { /* * We can't compensate for it all. * And we have to have some time * on the clock. We always have a min * 10 slots (10 x 10 i.e. 100 usecs). */ if (slot <= HPTS_TICKS_PER_SLOT) { /* We gain delay */ rack->r_ctl.rc_agg_delayed += (HPTS_TICKS_PER_SLOT - slot); slot = HPTS_TICKS_PER_SLOT; } else { /* We take off some */ rack->r_ctl.rc_agg_delayed -= (slot - HPTS_TICKS_PER_SLOT); slot = HPTS_TICKS_PER_SLOT; } } else { slot -= rack->r_ctl.rc_agg_delayed; rack->r_ctl.rc_agg_delayed = 0; /* Make sure we have 100 useconds at minimum */ if (slot < HPTS_TICKS_PER_SLOT) { rack->r_ctl.rc_agg_delayed = HPTS_TICKS_PER_SLOT - slot; slot = HPTS_TICKS_PER_SLOT; } if (rack->r_ctl.rc_agg_delayed == 0) rack->r_late = 0; } } if (slot) { /* We are pacing too */ rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT; } hpts_timeout = rack_timer_start(tp, rack, cts, sup_rack); #ifdef NETFLIX_EXP_DETECTION if (rack->sack_attack_disable && (slot < tcp_sad_pacing_interval)) { /* * We have a potential attacker on * the line. We have possibly some * (or now) pacing time set. We want to * slow down the processing of sacks by some * amount (if it is an attacker). Set the default * slot for attackers in place (unless the orginal * interval is longer). Its stored in * micro-seconds, so lets convert to msecs. */ slot = tcp_sad_pacing_interval; } #endif if (tp->t_flags & TF_DELACK) { delayed_ack = TICKS_2_USEC(tcp_delacktime); rack->r_ctl.rc_hpts_flags |= PACE_TMR_DELACK; } if (delayed_ack && ((hpts_timeout == 0) || (delayed_ack < hpts_timeout))) hpts_timeout = delayed_ack; else rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK; /* * If no timers are going to run and we will fall off the hptsi * wheel, we resort to a keep-alive timer if its configured. */ if ((hpts_timeout == 0) && (slot == 0)) { if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && (tp->t_state <= TCPS_CLOSING)) { /* * Ok we have no timer (persists, rack, tlp, rxt or * del-ack), we don't have segments being paced. So * all that is left is the keepalive timer. */ if (TCPS_HAVEESTABLISHED(tp->t_state)) { /* Get the established keep-alive time */ hpts_timeout = TICKS_2_USEC(TP_KEEPIDLE(tp)); } else { /* * Get the initial setup keep-alive time, * note that this is probably not going to * happen, since rack will be running a rxt timer * if a SYN of some sort is outstanding. It is * actually handled in rack_timeout_rxt(). */ hpts_timeout = TICKS_2_USEC(TP_KEEPINIT(tp)); } rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP; if (rack->in_probe_rtt) { /* * We want to instead not wake up a long time from * now but to wake up about the time we would * exit probe-rtt and initiate a keep-alive ack. * This will get us out of probe-rtt and update * our min-rtt. */ hpts_timeout = rack_min_probertt_hold; } } } if (left && (stopped & (PACE_TMR_KEEP | PACE_TMR_DELACK)) == (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK)) { /* * RACK, TLP, persists and RXT timers all are restartable * based on actions input .. i.e we received a packet (ack * or sack) and that changes things (rw, or snd_una etc). * Thus we can restart them with a new value. For * keep-alive, delayed_ack we keep track of what was left * and restart the timer with a smaller value. */ if (left < hpts_timeout) hpts_timeout = left; } if (hpts_timeout) { /* * Hack alert for now we can't time-out over 2,147,483 * seconds (a bit more than 596 hours), which is probably ok * :). */ if (hpts_timeout > 0x7ffffffe) hpts_timeout = 0x7ffffffe; rack->r_ctl.rc_timer_exp = cts + hpts_timeout; } rack_log_pacing_delay_calc(rack, entry_slot, slot, hpts_timeout, 0, 0, 27, __LINE__, NULL, 0); if ((rack->gp_ready == 0) && (rack->use_fixed_rate == 0) && (hpts_timeout < slot) && (rack->r_ctl.rc_hpts_flags & (PACE_TMR_TLP|PACE_TMR_RXT))) { /* * We have no good estimate yet for the * old clunky burst mitigation or the * real pacing. And the tlp or rxt is smaller * than the pacing calculation. Lets not * pace that long since we know the calculation * so far is not accurate. */ slot = hpts_timeout; } /** * Turn off all the flags for queuing by default. The * flags have important meanings to what happens when * LRO interacts with the transport. Most likely (by default now) * mbuf_queueing and ack compression are on. So the transport * has a couple of flags that control what happens (if those * are not on then these flags won't have any effect since it * won't go through the queuing LRO path). * * INP_MBUF_QUEUE_READY - This flags says that I am busy * pacing output, so don't disturb. But * it also means LRO can wake me if there * is a SACK arrival. * * INP_DONT_SACK_QUEUE - This flag is used in conjunction * with the above flag (QUEUE_READY) and * when present it says don't even wake me * if a SACK arrives. * * The idea behind these flags is that if we are pacing we * set the MBUF_QUEUE_READY and only get woken up if * a SACK arrives (which could change things) or if * our pacing timer expires. If, however, we have a rack * timer running, then we don't even want a sack to wake * us since the rack timer has to expire before we can send. * * Other cases should usually have none of the flags set * so LRO can call into us. */ inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY); if (slot) { rack->r_ctl.rc_last_output_to = us_cts + slot; /* * A pacing timer (slot) is being set, in * such a case we cannot send (we are blocked by * the timer). So lets tell LRO that it should not * wake us unless there is a SACK. Note this only * will be effective if mbuf queueing is on or * compressed acks are being processed. */ inp->inp_flags2 |= INP_MBUF_QUEUE_READY; /* * But wait if we have a Rack timer running * even a SACK should not disturb us (with * the exception of r_rr_config 3). */ if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) && (rack->r_rr_config != 3)) inp->inp_flags2 |= INP_DONT_SACK_QUEUE; if (rack->rc_ack_can_sendout_data) { /* * Ahh but wait, this is that special case * where the pacing timer can be disturbed * backout the changes (used for non-paced * burst limiting). */ inp->inp_flags2 &= ~(INP_DONT_SACK_QUEUE|INP_MBUF_QUEUE_READY); } if ((rack->use_rack_rr) && (rack->r_rr_config < 2) && ((hpts_timeout) && (hpts_timeout < slot))) { /* * Arrange for the hpts to kick back in after the * t-o if the t-o does not cause a send. */ (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout), __LINE__, &diag); rack_log_hpts_diag(rack, us_cts, &diag, &tv); rack_log_to_start(rack, cts, hpts_timeout, slot, 0); } else { (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(slot), __LINE__, &diag); rack_log_hpts_diag(rack, us_cts, &diag, &tv); rack_log_to_start(rack, cts, hpts_timeout, slot, 1); } } else if (hpts_timeout) { /* * With respect to inp_flags2 here, lets let any new acks wake * us up here. Since we are not pacing (no pacing timer), output * can happen so we should let it. If its a Rack timer, then any inbound * packet probably won't change the sending (we will be blocked) * but it may change the prr stats so letting it in (the set defaults * at the start of this block) are good enough. */ (void)tcp_hpts_insert_diag(tp->t_inpcb, HPTS_USEC_TO_SLOTS(hpts_timeout), __LINE__, &diag); rack_log_hpts_diag(rack, us_cts, &diag, &tv); rack_log_to_start(rack, cts, hpts_timeout, slot, 0); } else { /* No timer starting */ #ifdef INVARIANTS if (SEQ_GT(tp->snd_max, tp->snd_una)) { panic("tp:%p rack:%p tlts:%d cts:%u slot:%u pto:%u -- no timer started?", tp, rack, tot_len_this_send, cts, slot, hpts_timeout); } #endif } rack->rc_tmr_stopped = 0; if (slot) rack_log_type_bbrsnd(rack, tot_len_this_send, slot, us_cts, &tv); } /* * RACK Timer, here we simply do logging and house keeping. * the normal rack_output() function will call the * appropriate thing to check if we need to do a RACK retransmit. * We return 1, saying don't proceed with rack_output only * when all timers have been stopped (destroyed PCB?). */ static int rack_timeout_rack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) { /* * This timer simply provides an internal trigger to send out data. * The check_recovery_mode call will see if there are needed * retransmissions, if so we will enter fast-recovery. The output * call may or may not do the same thing depending on sysctl * settings. */ struct rack_sendmap *rsm; if (tp->t_timers->tt_flags & TT_STOPPED) { return (1); } counter_u64_add(rack_to_tot, 1); if (rack->r_state && (rack->r_state != tp->t_state)) rack_set_state(tp, rack); rack->rc_on_min_to = 0; rsm = rack_check_recovery_mode(tp, cts); rack_log_to_event(rack, RACK_TO_FRM_RACK, rsm); if (rsm) { rack->r_ctl.rc_resend = rsm; rack->r_timer_override = 1; if (rack->use_rack_rr) { /* * Don't accumulate extra pacing delay * we are allowing the rack timer to * over-ride pacing i.e. rrr takes precedence * if the pacing interval is longer than the rrr * time (in other words we get the min pacing * time versus rrr pacing time). */ rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT; } } rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK; if (rsm == NULL) { /* restart a timer and return 1 */ rack_start_hpts_timer(rack, tp, cts, 0, 0, 0); return (1); } return (0); } static void rack_adjust_orig_mlen(struct rack_sendmap *rsm) { if (rsm->m->m_len > rsm->orig_m_len) { /* * Mbuf grew, caused by sbcompress, our offset does * not change. */ rsm->orig_m_len = rsm->m->m_len; } else if (rsm->m->m_len < rsm->orig_m_len) { /* * Mbuf shrank, trimmed off the top by an ack, our * offset changes. */ rsm->soff -= (rsm->orig_m_len - rsm->m->m_len); rsm->orig_m_len = rsm->m->m_len; } } static void rack_setup_offset_for_rsm(struct rack_sendmap *src_rsm, struct rack_sendmap *rsm) { struct mbuf *m; uint32_t soff; if (src_rsm->m && (src_rsm->orig_m_len != src_rsm->m->m_len)) { /* Fix up the orig_m_len and possibly the mbuf offset */ rack_adjust_orig_mlen(src_rsm); } m = src_rsm->m; soff = src_rsm->soff + (src_rsm->r_end - src_rsm->r_start); while (soff >= m->m_len) { /* Move out past this mbuf */ soff -= m->m_len; m = m->m_next; KASSERT((m != NULL), ("rsm:%p nrsm:%p hit at soff:%u null m", src_rsm, rsm, soff)); } rsm->m = m; rsm->soff = soff; rsm->orig_m_len = m->m_len; } static __inline void rack_clone_rsm(struct tcp_rack *rack, struct rack_sendmap *nrsm, struct rack_sendmap *rsm, uint32_t start) { int idx; nrsm->r_start = start; nrsm->r_end = rsm->r_end; nrsm->r_rtr_cnt = rsm->r_rtr_cnt; nrsm->r_flags = rsm->r_flags; nrsm->r_dupack = rsm->r_dupack; nrsm->r_no_rtt_allowed = rsm->r_no_rtt_allowed; nrsm->r_rtr_bytes = 0; nrsm->r_fas = rsm->r_fas; rsm->r_end = nrsm->r_start; nrsm->r_just_ret = rsm->r_just_ret; for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; } /* Now if we have SYN flag we keep it on the left edge */ if (nrsm->r_flags & RACK_HAS_SYN) nrsm->r_flags &= ~RACK_HAS_SYN; /* Now if we have a FIN flag we keep it on the right edge */ if (rsm->r_flags & RACK_HAS_FIN) rsm->r_flags &= ~RACK_HAS_FIN; /* Push bit must go to the right edge as well */ if (rsm->r_flags & RACK_HAD_PUSH) rsm->r_flags &= ~RACK_HAD_PUSH; /* Clone over the state of the hw_tls flag */ nrsm->r_hw_tls = rsm->r_hw_tls; /* * Now we need to find nrsm's new location in the mbuf chain * we basically calculate a new offset, which is soff + * how much is left in original rsm. Then we walk out the mbuf * chain to find the righ position, it may be the same mbuf * or maybe not. */ KASSERT(((rsm->m != NULL) || (rsm->r_flags & (RACK_HAS_SYN|RACK_HAS_FIN))), ("rsm:%p nrsm:%p rack:%p -- rsm->m is NULL?", rsm, nrsm, rack)); if (rsm->m) rack_setup_offset_for_rsm(rsm, nrsm); } static struct rack_sendmap * rack_merge_rsm(struct tcp_rack *rack, struct rack_sendmap *l_rsm, struct rack_sendmap *r_rsm) { /* * We are merging two ack'd RSM's, * the l_rsm is on the left (lower seq * values) and the r_rsm is on the right * (higher seq value). The simplest way * to merge these is to move the right * one into the left. I don't think there * is any reason we need to try to find * the oldest (or last oldest retransmitted). */ #ifdef INVARIANTS struct rack_sendmap *rm; #endif rack_log_map_chg(rack->rc_tp, rack, NULL, l_rsm, r_rsm, MAP_MERGE, r_rsm->r_end, __LINE__); l_rsm->r_end = r_rsm->r_end; if (l_rsm->r_dupack < r_rsm->r_dupack) l_rsm->r_dupack = r_rsm->r_dupack; if (r_rsm->r_rtr_bytes) l_rsm->r_rtr_bytes += r_rsm->r_rtr_bytes; if (r_rsm->r_in_tmap) { /* This really should not happen */ TAILQ_REMOVE(&rack->r_ctl.rc_tmap, r_rsm, r_tnext); r_rsm->r_in_tmap = 0; } /* Now the flags */ if (r_rsm->r_flags & RACK_HAS_FIN) l_rsm->r_flags |= RACK_HAS_FIN; if (r_rsm->r_flags & RACK_TLP) l_rsm->r_flags |= RACK_TLP; if (r_rsm->r_flags & RACK_RWND_COLLAPSED) l_rsm->r_flags |= RACK_RWND_COLLAPSED; if ((r_rsm->r_flags & RACK_APP_LIMITED) && ((l_rsm->r_flags & RACK_APP_LIMITED) == 0)) { /* * If both are app-limited then let the * free lower the count. If right is app * limited and left is not, transfer. */ l_rsm->r_flags |= RACK_APP_LIMITED; r_rsm->r_flags &= ~RACK_APP_LIMITED; if (r_rsm == rack->r_ctl.rc_first_appl) rack->r_ctl.rc_first_appl = l_rsm; } #ifndef INVARIANTS (void)RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, r_rsm); #else rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, r_rsm); if (rm != r_rsm) { panic("removing head in rack:%p rsm:%p rm:%p", rack, r_rsm, rm); } #endif if ((r_rsm->r_limit_type == 0) && (l_rsm->r_limit_type != 0)) { /* Transfer the split limit to the map we free */ r_rsm->r_limit_type = l_rsm->r_limit_type; l_rsm->r_limit_type = 0; } rack_free(rack, r_rsm); return (l_rsm); } /* * TLP Timer, here we simply setup what segment we want to * have the TLP expire on, the normal rack_output() will then * send it out. * * We return 1, saying don't proceed with rack_output only * when all timers have been stopped (destroyed PCB?). */ static int rack_timeout_tlp(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t *doing_tlp) { /* * Tail Loss Probe. */ struct rack_sendmap *rsm = NULL; #ifdef INVARIANTS struct rack_sendmap *insret; #endif struct socket *so; uint32_t amm; uint32_t out, avail; int collapsed_win = 0; if (tp->t_timers->tt_flags & TT_STOPPED) { return (1); } if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) { /* Its not time yet */ return (0); } if (ctf_progress_timeout_check(tp, true)) { rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__); return (-ETIMEDOUT); /* tcp_drop() */ } /* * A TLP timer has expired. We have been idle for 2 rtts. So we now * need to figure out how to force a full MSS segment out. */ rack_log_to_event(rack, RACK_TO_FRM_TLP, NULL); rack->r_ctl.retran_during_recovery = 0; rack->r_ctl.dsack_byte_cnt = 0; counter_u64_add(rack_tlp_tot, 1); if (rack->r_state && (rack->r_state != tp->t_state)) rack_set_state(tp, rack); so = tp->t_inpcb->inp_socket; avail = sbavail(&so->so_snd); out = tp->snd_max - tp->snd_una; if ((out > tp->snd_wnd) || rack->rc_has_collapsed) { /* special case, we need a retransmission */ collapsed_win = 1; goto need_retran; } if (rack->r_ctl.dsack_persist && (rack->r_ctl.rc_tlp_cnt_out >= 1)) { rack->r_ctl.dsack_persist--; if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) { rack->r_ctl.num_dsack = 0; } rack_log_dsack_event(rack, 1, __LINE__, 0, 0); } if ((tp->t_flags & TF_GPUTINPROG) && (rack->r_ctl.rc_tlp_cnt_out == 1)) { /* * If this is the second in a row * TLP and we are doing a measurement * its time to abandon the measurement. * Something is likely broken on * the clients network and measuring a * broken network does us no good. */ tp->t_flags &= ~TF_GPUTINPROG; rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/, rack->r_ctl.rc_gp_srtt /*flex1*/, tp->gput_seq, 0, 0, 18, __LINE__, NULL, 0); } /* * Check our send oldest always settings, and if * there is an oldest to send jump to the need_retran. */ if (rack_always_send_oldest && (TAILQ_EMPTY(&rack->r_ctl.rc_tmap) == 0)) goto need_retran; if (avail > out) { /* New data is available */ amm = avail - out; if (amm > ctf_fixed_maxseg(tp)) { amm = ctf_fixed_maxseg(tp); if ((amm + out) > tp->snd_wnd) { /* We are rwnd limited */ goto need_retran; } } else if (amm < ctf_fixed_maxseg(tp)) { /* not enough to fill a MTU */ goto need_retran; } if (IN_FASTRECOVERY(tp->t_flags)) { /* Unlikely */ if (rack->rack_no_prr == 0) { if (out + amm <= tp->snd_wnd) { rack->r_ctl.rc_prr_sndcnt = amm; rack->r_ctl.rc_tlp_new_data = amm; rack_log_to_prr(rack, 4, 0, __LINE__); } } else goto need_retran; } else { /* Set the send-new override */ if (out + amm <= tp->snd_wnd) rack->r_ctl.rc_tlp_new_data = amm; else goto need_retran; } rack->r_ctl.rc_tlpsend = NULL; counter_u64_add(rack_tlp_newdata, 1); goto send; } need_retran: /* * Ok we need to arrange the last un-acked segment to be re-sent, or * optionally the first un-acked segment. */ if (collapsed_win == 0) { if (rack_always_send_oldest) rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); else { rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree); if (rsm && (rsm->r_flags & (RACK_ACKED | RACK_HAS_FIN))) { rsm = rack_find_high_nonack(rack, rsm); } } if (rsm == NULL) { #ifdef TCP_BLACKBOX tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true); #endif goto out; } } else { /* * We must find the last segment * that was acceptable by the client. */ RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) { if ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0) { /* Found one */ break; } } if (rsm == NULL) { /* None? if so send the first */ rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree); if (rsm == NULL) { #ifdef TCP_BLACKBOX tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true); #endif goto out; } } } if ((rsm->r_end - rsm->r_start) > ctf_fixed_maxseg(tp)) { /* * We need to split this the last segment in two. */ struct rack_sendmap *nrsm; nrsm = rack_alloc_full_limit(rack); if (nrsm == NULL) { /* * No memory to split, we will just exit and punt * off to the RXT timer. */ goto out; } rack_clone_rsm(rack, nrsm, rsm, (rsm->r_end - ctf_fixed_maxseg(tp))); rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__); #ifndef INVARIANTS (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); #else insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); if (insret != NULL) { panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p", nrsm, insret, rack, rsm); } #endif if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } rsm = nrsm; } rack->r_ctl.rc_tlpsend = rsm; send: /* Make sure output path knows we are doing a TLP */ *doing_tlp = 1; rack->r_timer_override = 1; rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP; return (0); out: rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP; return (0); } /* * Delayed ack Timer, here we simply need to setup the * ACK_NOW flag and remove the DELACK flag. From there * the output routine will send the ack out. * * We only return 1, saying don't proceed, if all timers * are stopped (destroyed PCB?). */ static int rack_timeout_delack(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) { if (tp->t_timers->tt_flags & TT_STOPPED) { return (1); } rack_log_to_event(rack, RACK_TO_FRM_DELACK, NULL); tp->t_flags &= ~TF_DELACK; tp->t_flags |= TF_ACKNOW; KMOD_TCPSTAT_INC(tcps_delack); rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK; return (0); } /* * Persists timer, here we simply send the * same thing as a keepalive will. * the one byte send. * * We only return 1, saying don't proceed, if all timers * are stopped (destroyed PCB?). */ static int rack_timeout_persist(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) { struct tcptemp *t_template; #ifdef INVARIANTS struct inpcb *inp = tp->t_inpcb; #endif int32_t retval = 1; if (tp->t_timers->tt_flags & TT_STOPPED) { return (1); } if (rack->rc_in_persist == 0) return (0); if (ctf_progress_timeout_check(tp, false)) { tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX); rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__); counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends); return (-ETIMEDOUT); /* tcp_drop() */ } KASSERT(inp != NULL, ("%s: tp %p tp->t_inpcb == NULL", __func__, tp)); /* * Persistence timer into zero window. Force a byte to be output, if * possible. */ KMOD_TCPSTAT_INC(tcps_persisttimeo); /* * Hack: if the peer is dead/unreachable, we do not time out if the * window is closed. After a full backoff, drop the connection if * the idle time (no responses to probes) reaches the maximum * backoff that we would use if retransmitting. */ if (tp->t_rxtshift == TCP_MAXRXTSHIFT && (ticks - tp->t_rcvtime >= tcp_maxpersistidle || TICKS_2_USEC(ticks - tp->t_rcvtime) >= RACK_REXMTVAL(tp) * tcp_totbackoff)) { KMOD_TCPSTAT_INC(tcps_persistdrop); tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX); counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends); retval = -ETIMEDOUT; /* tcp_drop() */ goto out; } if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) && tp->snd_una == tp->snd_max) rack_exit_persist(tp, rack, cts); rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_PERSIT; /* * If the user has closed the socket then drop a persisting * connection after a much reduced timeout. */ if (tp->t_state > TCPS_CLOSE_WAIT && (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) { KMOD_TCPSTAT_INC(tcps_persistdrop); tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX); counter_u64_add(rack_persists_lost_ends, rack->r_ctl.persist_lost_ends); retval = -ETIMEDOUT; /* tcp_drop() */ goto out; } t_template = tcpip_maketemplate(rack->rc_inp); if (t_template) { /* only set it if we were answered */ if (rack->forced_ack == 0) { rack->forced_ack = 1; rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL); } else { rack->probe_not_answered = 1; counter_u64_add(rack_persists_loss, 1); rack->r_ctl.persist_lost_ends++; } counter_u64_add(rack_persists_sends, 1); tcp_respond(tp, t_template->tt_ipgen, &t_template->tt_t, (struct mbuf *)NULL, tp->rcv_nxt, tp->snd_una - 1, 0); /* This sends an ack */ if (tp->t_flags & TF_DELACK) tp->t_flags &= ~TF_DELACK; free(t_template, M_TEMP); } if (tp->t_rxtshift < TCP_MAXRXTSHIFT) tp->t_rxtshift++; out: rack_log_to_event(rack, RACK_TO_FRM_PERSIST, NULL); rack_start_hpts_timer(rack, tp, cts, 0, 0, 0); return (retval); } /* * If a keepalive goes off, we had no other timers * happening. We always return 1 here since this * routine either drops the connection or sends * out a segment with respond. */ static int rack_timeout_keepalive(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) { struct tcptemp *t_template; struct inpcb *inp; if (tp->t_timers->tt_flags & TT_STOPPED) { return (1); } rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_KEEP; inp = tp->t_inpcb; rack_log_to_event(rack, RACK_TO_FRM_KEEP, NULL); /* * Keep-alive timer went off; send something or drop connection if * idle for too long. */ KMOD_TCPSTAT_INC(tcps_keeptimeo); if (tp->t_state < TCPS_ESTABLISHED) goto dropit; if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && tp->t_state <= TCPS_CLOSING) { if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp)) goto dropit; /* * Send a packet designed to force a response if the peer is * up and reachable: either an ACK if the connection is * still alive, or an RST if the peer has closed the * connection due to timeout or reboot. Using sequence * number tp->snd_una-1 causes the transmitted zero-length * segment to lie outside the receive window; by the * protocol spec, this requires the correspondent TCP to * respond. */ KMOD_TCPSTAT_INC(tcps_keepprobe); t_template = tcpip_maketemplate(inp); if (t_template) { if (rack->forced_ack == 0) { rack->forced_ack = 1; rack->r_ctl.forced_ack_ts = tcp_get_usecs(NULL); } else { rack->probe_not_answered = 1; } tcp_respond(tp, t_template->tt_ipgen, &t_template->tt_t, (struct mbuf *)NULL, tp->rcv_nxt, tp->snd_una - 1, 0); free(t_template, M_TEMP); } } rack_start_hpts_timer(rack, tp, cts, 0, 0, 0); return (1); dropit: KMOD_TCPSTAT_INC(tcps_keepdrops); tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX); return (-ETIMEDOUT); /* tcp_drop() */ } /* * Retransmit helper function, clear up all the ack * flags and take care of important book keeping. */ static void rack_remxt_tmr(struct tcpcb *tp) { /* * The retransmit timer went off, all sack'd blocks must be * un-acked. */ struct rack_sendmap *rsm, *trsm = NULL; struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; rack_timer_cancel(tp, rack, tcp_get_usecs(NULL), __LINE__); rack_log_to_event(rack, RACK_TO_FRM_TMR, NULL); if (rack->r_state && (rack->r_state != tp->t_state)) rack_set_state(tp, rack); /* * Ideally we would like to be able to * mark SACK-PASS on anything not acked here. * * However, if we do that we would burst out * all that data 1ms apart. This would be unwise, * so for now we will just let the normal rxt timer * and tlp timer take care of it. * * Also we really need to stick them back in sequence * order. This way we send in the proper order and any * sacks that come floating in will "re-ack" the data. * To do this we zap the tmap with an INIT and then * walk through and place every rsm in the RB tree * back in its seq ordered place. */ TAILQ_INIT(&rack->r_ctl.rc_tmap); RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) { rsm->r_dupack = 0; rack_log_retran_reason(rack, rsm, __LINE__, 0, 2); /* We must re-add it back to the tlist */ if (trsm == NULL) { TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext); } else { TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, trsm, rsm, r_tnext); } rsm->r_in_tmap = 1; trsm = rsm; if (rsm->r_flags & RACK_ACKED) rsm->r_flags |= RACK_WAS_ACKED; rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS | RACK_RWND_COLLAPSED); rsm->r_flags |= RACK_MUST_RXT; } /* Clear the count (we just un-acked them) */ rack->r_ctl.rc_last_timeout_snduna = tp->snd_una; rack->r_ctl.rc_sacked = 0; rack->r_ctl.rc_sacklast = NULL; rack->r_ctl.rc_agg_delayed = 0; rack->r_early = 0; rack->r_ctl.rc_agg_early = 0; rack->r_late = 0; /* Clear the tlp rtx mark */ rack->r_ctl.rc_resend = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree); if (rack->r_ctl.rc_resend != NULL) rack->r_ctl.rc_resend->r_flags |= RACK_TO_REXT; rack->r_ctl.rc_prr_sndcnt = 0; rack_log_to_prr(rack, 6, 0, __LINE__); rack->r_timer_override = 1; if ((((tp->t_flags & TF_SACK_PERMIT) == 0) #ifdef NETFLIX_EXP_DETECTION || (rack->sack_attack_disable != 0) #endif ) && ((tp->t_flags & TF_SENTFIN) == 0)) { /* * For non-sack customers new data * needs to go out as retransmits until * we retransmit up to snd_max. */ rack->r_must_retran = 1; rack->r_ctl.rc_out_at_rto = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); } rack->r_ctl.rc_snd_max_at_rto = tp->snd_max; } static void rack_convert_rtts(struct tcpcb *tp) { if (tp->t_srtt > 1) { uint32_t val, frac; val = tp->t_srtt >> TCP_RTT_SHIFT; frac = tp->t_srtt & 0x1f; tp->t_srtt = TICKS_2_USEC(val); /* * frac is the fractional part of the srtt (if any) * but its in ticks and every bit represents * 1/32nd of a hz. */ if (frac) { if (hz == 1000) { frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE); } else { frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE)); } tp->t_srtt += frac; } } if (tp->t_rttvar) { uint32_t val, frac; val = tp->t_rttvar >> TCP_RTTVAR_SHIFT; frac = tp->t_rttvar & 0x1f; tp->t_rttvar = TICKS_2_USEC(val); /* * frac is the fractional part of the srtt (if any) * but its in ticks and every bit represents * 1/32nd of a hz. */ if (frac) { if (hz == 1000) { frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_MSEC) / (uint64_t)TCP_RTT_SCALE); } else { frac = (((uint64_t)frac * (uint64_t)HPTS_USEC_IN_SEC) / ((uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE)); } tp->t_rttvar += frac; } } tp->t_rxtcur = RACK_REXMTVAL(tp); if (TCPS_HAVEESTABLISHED(tp->t_state)) { tp->t_rxtcur += TICKS_2_USEC(tcp_rexmit_slop); } if (tp->t_rxtcur > rack_rto_max) { tp->t_rxtcur = rack_rto_max; } } static void rack_cc_conn_init(struct tcpcb *tp) { struct tcp_rack *rack; uint32_t srtt; rack = (struct tcp_rack *)tp->t_fb_ptr; srtt = tp->t_srtt; cc_conn_init(tp); /* * Now convert to rack's internal format, * if required. */ if ((srtt == 0) && (tp->t_srtt != 0)) rack_convert_rtts(tp); /* * We want a chance to stay in slowstart as * we create a connection. TCP spec says that * initially ssthresh is infinite. For our * purposes that is the snd_wnd. */ if (tp->snd_ssthresh < tp->snd_wnd) { tp->snd_ssthresh = tp->snd_wnd; } /* * We also want to assure a IW worth of * data can get inflight. */ if (rc_init_window(rack) < tp->snd_cwnd) tp->snd_cwnd = rc_init_window(rack); } /* * Re-transmit timeout! If we drop the PCB we will return 1, otherwise * we will setup to retransmit the lowest seq number outstanding. */ static int rack_timeout_rxt(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts) { int32_t rexmt; int32_t retval = 0; bool isipv6; if (tp->t_timers->tt_flags & TT_STOPPED) { return (1); } if ((tp->t_flags & TF_GPUTINPROG) && (tp->t_rxtshift)) { /* * We have had a second timeout * measurements on successive rxt's are not profitable. * It is unlikely to be of any use (the network is * broken or the client went away). */ tp->t_flags &= ~TF_GPUTINPROG; rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/, rack->r_ctl.rc_gp_srtt /*flex1*/, tp->gput_seq, 0, 0, 18, __LINE__, NULL, 0); } if (ctf_progress_timeout_check(tp, false)) { tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN); rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__); return (-ETIMEDOUT); /* tcp_drop() */ } rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT; rack->r_ctl.retran_during_recovery = 0; rack->rc_ack_required = 1; rack->r_ctl.dsack_byte_cnt = 0; if (IN_FASTRECOVERY(tp->t_flags)) tp->t_flags |= TF_WASFRECOVERY; else tp->t_flags &= ~TF_WASFRECOVERY; if (IN_CONGRECOVERY(tp->t_flags)) tp->t_flags |= TF_WASCRECOVERY; else tp->t_flags &= ~TF_WASCRECOVERY; if (TCPS_HAVEESTABLISHED(tp->t_state) && (tp->snd_una == tp->snd_max)) { /* Nothing outstanding .. nothing to do */ return (0); } if (rack->r_ctl.dsack_persist) { rack->r_ctl.dsack_persist--; if (rack->r_ctl.num_dsack && (rack->r_ctl.dsack_persist == 0)) { rack->r_ctl.num_dsack = 0; } rack_log_dsack_event(rack, 1, __LINE__, 0, 0); } /* * Rack can only run one timer at a time, so we cannot * run a KEEPINIT (gating SYN sending) and a retransmit * timer for the SYN. So if we are in a front state and * have a KEEPINIT timer we need to check the first transmit * against now to see if we have exceeded the KEEPINIT time * (if one is set). */ if ((TCPS_HAVEESTABLISHED(tp->t_state) == 0) && (TP_KEEPINIT(tp) != 0)) { struct rack_sendmap *rsm; rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree); if (rsm) { /* Ok we have something outstanding to test keepinit with */ if ((TSTMP_GT(cts, (uint32_t)rsm->r_tim_lastsent[0])) && ((cts - (uint32_t)rsm->r_tim_lastsent[0]) >= TICKS_2_USEC(TP_KEEPINIT(tp)))) { /* We have exceeded the KEEPINIT time */ tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX); goto drop_it; } } } /* * Retransmission timer went off. Message has not been acked within * retransmit interval. Back off to a longer retransmit interval * and retransmit one segment. */ rack_remxt_tmr(tp); if ((rack->r_ctl.rc_resend == NULL) || ((rack->r_ctl.rc_resend->r_flags & RACK_RWND_COLLAPSED) == 0)) { /* * If the rwnd collapsed on * the one we are retransmitting * it does not count against the * rxt count. */ tp->t_rxtshift++; } if (tp->t_rxtshift > TCP_MAXRXTSHIFT) { tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN); drop_it: tp->t_rxtshift = TCP_MAXRXTSHIFT; KMOD_TCPSTAT_INC(tcps_timeoutdrop); /* XXXGL: previously t_softerror was casted to uint16_t */ MPASS(tp->t_softerror >= 0); retval = tp->t_softerror ? -tp->t_softerror : -ETIMEDOUT; goto out; /* tcp_drop() */ } if (tp->t_state == TCPS_SYN_SENT) { /* * If the SYN was retransmitted, indicate CWND to be limited * to 1 segment in cc_conn_init(). */ tp->snd_cwnd = 1; } else if (tp->t_rxtshift == 1) { /* * first retransmit; record ssthresh and cwnd so they can be * recovered if this turns out to be a "bad" retransmit. A * retransmit is considered "bad" if an ACK for this segment * is received within RTT/2 interval; the assumption here is * that the ACK was already in flight. See "On Estimating * End-to-End Network Path Properties" by Allman and Paxson * for more details. */ tp->snd_cwnd_prev = tp->snd_cwnd; tp->snd_ssthresh_prev = tp->snd_ssthresh; tp->snd_recover_prev = tp->snd_recover; tp->t_badrxtwin = ticks + (USEC_2_TICKS(tp->t_srtt)/2); tp->t_flags |= TF_PREVVALID; } else if ((tp->t_flags & TF_RCVD_TSTMP) == 0) tp->t_flags &= ~TF_PREVVALID; KMOD_TCPSTAT_INC(tcps_rexmttimeo); if ((tp->t_state == TCPS_SYN_SENT) || (tp->t_state == TCPS_SYN_RECEIVED)) rexmt = RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift]; else rexmt = max(rack_rto_min, (tp->t_srtt + (tp->t_rttvar << 2))) * tcp_backoff[tp->t_rxtshift]; RACK_TCPT_RANGESET(tp->t_rxtcur, rexmt, max(rack_rto_min, rexmt), rack_rto_max, rack->r_ctl.timer_slop); /* * We enter the path for PLMTUD if connection is established or, if * connection is FIN_WAIT_1 status, reason for the last is that if * amount of data we send is very small, we could send it in couple * of packets and process straight to FIN. In that case we won't * catch ESTABLISHED state. */ #ifdef INET6 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? true : false; #else isipv6 = false; #endif if (((V_tcp_pmtud_blackhole_detect == 1) || (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) || (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) && ((tp->t_state == TCPS_ESTABLISHED) || (tp->t_state == TCPS_FIN_WAIT_1))) { /* * Idea here is that at each stage of mtu probe (usually, * 1448 -> 1188 -> 524) should be given 2 chances to recover * before further clamping down. 'tp->t_rxtshift % 2 == 0' * should take care of that. */ if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) == (TF2_PLPMTU_PMTUD | TF2_PLPMTU_MAXSEGSNT)) && (tp->t_rxtshift >= 2 && tp->t_rxtshift < 6 && tp->t_rxtshift % 2 == 0)) { /* * Enter Path MTU Black-hole Detection mechanism: - * Disable Path MTU Discovery (IP "DF" bit). - * Reduce MTU to lower value than what we negotiated * with peer. */ if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) { /* Record that we may have found a black hole. */ tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE; /* Keep track of previous MSS. */ tp->t_pmtud_saved_maxseg = tp->t_maxseg; } /* * Reduce the MSS to blackhole value or to the * default in an attempt to retransmit. */ #ifdef INET6 if (isipv6 && tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) { /* Use the sysctl tuneable blackhole MSS. */ tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss; KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated); } else if (isipv6) { /* Use the default MSS. */ tp->t_maxseg = V_tcp_v6mssdflt; /* * Disable Path MTU Discovery when we switch * to minmss. */ tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); } #endif #if defined(INET6) && defined(INET) else #endif #ifdef INET if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) { /* Use the sysctl tuneable blackhole MSS. */ tp->t_maxseg = V_tcp_pmtud_blackhole_mss; KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated); } else { /* Use the default MSS. */ tp->t_maxseg = V_tcp_mssdflt; /* * Disable Path MTU Discovery when we switch * to minmss. */ tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); } #endif } else { /* * If further retransmissions are still unsuccessful * with a lowered MTU, maybe this isn't a blackhole * and we restore the previous MSS and blackhole * detection flags. The limit '6' is determined by * giving each probe stage (1448, 1188, 524) 2 * chances to recover. */ if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) && (tp->t_rxtshift >= 6)) { tp->t_flags2 |= TF2_PLPMTU_PMTUD; tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE; tp->t_maxseg = tp->t_pmtud_saved_maxseg; KMOD_TCPSTAT_INC(tcps_pmtud_blackhole_failed); } } } /* * Disable RFC1323 and SACK if we haven't got any response to * our third SYN to work-around some broken terminal servers * (most of which have hopefully been retired) that have bad VJ * header compression code which trashes TCP segments containing * unknown-to-them TCP options. */ if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) && (tp->t_rxtshift == 3)) tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT); /* * If we backed off this far, our srtt estimate is probably bogus. * Clobber it so we'll take the next rtt measurement as our srtt; * move the current srtt into rttvar to keep the current retransmit * times until then. */ if (tp->t_rxtshift > TCP_MAXRXTSHIFT / 4) { #ifdef INET6 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0) in6_losing(tp->t_inpcb); else #endif in_losing(tp->t_inpcb); tp->t_rttvar += tp->t_srtt; tp->t_srtt = 0; } sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una); tp->snd_recover = tp->snd_max; tp->t_flags |= TF_ACKNOW; tp->t_rtttime = 0; rack_cong_signal(tp, CC_RTO, tp->snd_una, __LINE__); out: return (retval); } static int rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling, uint8_t *doing_tlp) { int32_t ret = 0; int32_t timers = (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK); if ((tp->t_state >= TCPS_FIN_WAIT_1) && (tp->t_flags & TF_GPUTINPROG)) { /* * We have a goodput in progress * and we have entered a late state. * Do we have enough data in the sb * to handle the GPUT request? */ uint32_t bytes; bytes = tp->gput_ack - tp->gput_seq; if (SEQ_GT(tp->gput_seq, tp->snd_una)) bytes += tp->gput_seq - tp->snd_una; if (bytes > sbavail(&tp->t_inpcb->inp_socket->so_snd)) { /* * There are not enough bytes in the socket * buffer that have been sent to cover this * measurement. Cancel it. */ rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/, rack->r_ctl.rc_gp_srtt /*flex1*/, tp->gput_seq, 0, 0, 18, __LINE__, NULL, 0); tp->t_flags &= ~TF_GPUTINPROG; } } if (timers == 0) { return (0); } if (tp->t_state == TCPS_LISTEN) { /* no timers on listen sockets */ if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) return (0); return (1); } if ((timers & PACE_TMR_RACK) && rack->rc_on_min_to) { /* * For the rack timer when we * are on a min-timeout (which means rrr_conf = 3) * we don't want to check the timer. It may * be going off for a pace and thats ok we * want to send the retransmit (if its ready). * * If its on a normal rack timer (non-min) then * we will check if its expired. */ goto skip_time_check; } if (TSTMP_LT(cts, rack->r_ctl.rc_timer_exp)) { uint32_t left; if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) { ret = -1; rack_log_to_processing(rack, cts, ret, 0); return (0); } if (hpts_calling == 0) { /* * A user send or queued mbuf (sack) has called us? We * return 0 and let the pacing guards * deal with it if they should or * should not cause a send. */ ret = -2; rack_log_to_processing(rack, cts, ret, 0); return (0); } /* * Ok our timer went off early and we are not paced false * alarm, go back to sleep. */ ret = -3; left = rack->r_ctl.rc_timer_exp - cts; tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(left)); rack_log_to_processing(rack, cts, ret, left); return (1); } skip_time_check: rack->rc_tmr_stopped = 0; rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_MASK; if (timers & PACE_TMR_DELACK) { ret = rack_timeout_delack(tp, rack, cts); } else if (timers & PACE_TMR_RACK) { rack->r_ctl.rc_tlp_rxt_last_time = cts; rack->r_fast_output = 0; ret = rack_timeout_rack(tp, rack, cts); } else if (timers & PACE_TMR_TLP) { rack->r_ctl.rc_tlp_rxt_last_time = cts; ret = rack_timeout_tlp(tp, rack, cts, doing_tlp); } else if (timers & PACE_TMR_RXT) { rack->r_ctl.rc_tlp_rxt_last_time = cts; rack->r_fast_output = 0; ret = rack_timeout_rxt(tp, rack, cts); } else if (timers & PACE_TMR_PERSIT) { ret = rack_timeout_persist(tp, rack, cts); } else if (timers & PACE_TMR_KEEP) { ret = rack_timeout_keepalive(tp, rack, cts); } rack_log_to_processing(rack, cts, ret, timers); return (ret); } static void rack_timer_cancel(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, int line) { struct timeval tv; uint32_t us_cts, flags_on_entry; uint8_t hpts_removed = 0; flags_on_entry = rack->r_ctl.rc_hpts_flags; us_cts = tcp_get_usecs(&tv); if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) && ((TSTMP_GEQ(us_cts, rack->r_ctl.rc_last_output_to)) || ((tp->snd_max - tp->snd_una) == 0))) { tcp_hpts_remove(rack->rc_inp); hpts_removed = 1; /* If we were not delayed cancel out the flag. */ if ((tp->snd_max - tp->snd_una) == 0) rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT; rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry); } if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK; if (tcp_in_hpts(rack->rc_inp) && ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)) { /* * Canceling timer's when we have no output being * paced. We also must remove ourselves from the * hpts. */ tcp_hpts_remove(rack->rc_inp); hpts_removed = 1; } rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK); } if (hpts_removed == 0) rack_log_to_cancel(rack, hpts_removed, line, us_cts, &tv, flags_on_entry); } static void rack_timer_stop(struct tcpcb *tp, uint32_t timer_type) { return; } static int rack_stopall(struct tcpcb *tp) { struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; rack->t_timers_stopped = 1; return (0); } static void rack_timer_activate(struct tcpcb *tp, uint32_t timer_type, uint32_t delta) { return; } static int rack_timer_active(struct tcpcb *tp, uint32_t timer_type) { return (0); } static void rack_stop_all_timers(struct tcpcb *tp) { struct tcp_rack *rack; /* * Assure no timers are running. */ if (tcp_timer_active(tp, TT_PERSIST)) { /* We enter in persists, set the flag appropriately */ rack = (struct tcp_rack *)tp->t_fb_ptr; rack->rc_in_persist = 1; } tcp_timer_suspend(tp, TT_PERSIST); tcp_timer_suspend(tp, TT_REXMT); tcp_timer_suspend(tp, TT_KEEP); tcp_timer_suspend(tp, TT_DELACK); } static void rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm, uint64_t ts, uint16_t add_flag) { int32_t idx; rsm->r_rtr_cnt++; rack_log_retran_reason(rack, rsm, __LINE__, 0, 2); rsm->r_dupack = 0; if (rsm->r_rtr_cnt > RACK_NUM_OF_RETRANS) { rsm->r_rtr_cnt = RACK_NUM_OF_RETRANS; rsm->r_flags |= RACK_OVERMAX; } if ((rsm->r_rtr_cnt > 1) && ((rsm->r_flags & RACK_TLP) == 0)) { rack->r_ctl.rc_holes_rxt += (rsm->r_end - rsm->r_start); rsm->r_rtr_bytes += (rsm->r_end - rsm->r_start); } idx = rsm->r_rtr_cnt - 1; rsm->r_tim_lastsent[idx] = ts; /* * Here we don't add in the len of send, since its already * in snduna <->snd_max. */ rsm->r_fas = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); if (rsm->r_flags & RACK_ACKED) { /* Problably MTU discovery messing with us */ rsm->r_flags &= ~RACK_ACKED; rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); } if (rsm->r_in_tmap) { TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 0; } TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 1; /* Take off the must retransmit flag, if its on */ if (rsm->r_flags & RACK_MUST_RXT) { if (rack->r_must_retran) rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start); if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) { /* * We have retransmitted all we need. Clear * any must retransmit flags. */ rack->r_must_retran = 0; rack->r_ctl.rc_out_at_rto = 0; } rsm->r_flags &= ~RACK_MUST_RXT; } if (rsm->r_flags & RACK_SACK_PASSED) { /* We have retransmitted due to the SACK pass */ rsm->r_flags &= ~RACK_SACK_PASSED; rsm->r_flags |= RACK_WAS_SACKPASS; } } static uint32_t rack_update_entry(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm, uint64_t ts, int32_t *lenp, uint16_t add_flag) { /* * We (re-)transmitted starting at rsm->r_start for some length * (possibly less than r_end. */ struct rack_sendmap *nrsm; #ifdef INVARIANTS struct rack_sendmap *insret; #endif uint32_t c_end; int32_t len; len = *lenp; c_end = rsm->r_start + len; if (SEQ_GEQ(c_end, rsm->r_end)) { /* * We retransmitted the whole piece or more than the whole * slopping into the next rsm. */ rack_update_rsm(tp, rack, rsm, ts, add_flag); if (c_end == rsm->r_end) { *lenp = 0; return (0); } else { int32_t act_len; /* Hangs over the end return whats left */ act_len = rsm->r_end - rsm->r_start; *lenp = (len - act_len); return (rsm->r_end); } /* We don't get out of this block. */ } /* * Here we retransmitted less than the whole thing which means we * have to split this into what was transmitted and what was not. */ nrsm = rack_alloc_full_limit(rack); if (nrsm == NULL) { /* * We can't get memory, so lets not proceed. */ *lenp = 0; return (0); } /* * So here we are going to take the original rsm and make it what we * retransmitted. nrsm will be the tail portion we did not * retransmit. For example say the chunk was 1, 11 (10 bytes). And * we retransmitted 5 bytes i.e. 1, 5. The original piece shrinks to * 1, 6 and the new piece will be 6, 11. */ rack_clone_rsm(rack, nrsm, rsm, c_end); nrsm->r_dupack = 0; rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2); #ifndef INVARIANTS (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); #else insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); if (insret != NULL) { panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p", nrsm, insret, rack, rsm); } #endif if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } rsm->r_flags &= (~RACK_HAS_FIN); rack_update_rsm(tp, rack, rsm, ts, add_flag); /* Log a split of rsm into rsm and nrsm */ rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__); *lenp = 0; return (0); } static void rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len, uint32_t seq_out, uint16_t th_flags, int32_t err, uint64_t cts, struct rack_sendmap *hintrsm, uint16_t add_flag, struct mbuf *s_mb, uint32_t s_moff, int hw_tls) { struct tcp_rack *rack; struct rack_sendmap *rsm, *nrsm, fe; #ifdef INVARIANTS struct rack_sendmap *insret; #endif register uint32_t snd_max, snd_una; /* * Add to the RACK log of packets in flight or retransmitted. If * there is a TS option we will use the TS echoed, if not we will * grab a TS. * * Retransmissions will increment the count and move the ts to its * proper place. Note that if options do not include TS's then we * won't be able to effectively use the ACK for an RTT on a retran. * * Notes about r_start and r_end. Lets consider a send starting at * sequence 1 for 10 bytes. In such an example the r_start would be * 1 (starting sequence) but the r_end would be r_start+len i.e. 11. * This means that r_end is actually the first sequence for the next * slot (11). * */ /* * If err is set what do we do XXXrrs? should we not add the thing? * -- i.e. return if err != 0 or should we pretend we sent it? -- * i.e. proceed with add ** do this for now. */ INP_WLOCK_ASSERT(tp->t_inpcb); if (err) /* * We don't log errors -- we could but snd_max does not * advance in this case either. */ return; if (th_flags & TH_RST) { /* * We don't log resets and we return immediately from * sending */ return; } rack = (struct tcp_rack *)tp->t_fb_ptr; snd_una = tp->snd_una; snd_max = tp->snd_max; if (th_flags & (TH_SYN | TH_FIN)) { /* * The call to rack_log_output is made before bumping * snd_max. This means we can record one extra byte on a SYN * or FIN if seq_out is adding more on and a FIN is present * (and we are not resending). */ if ((th_flags & TH_SYN) && (seq_out == tp->iss)) len++; if (th_flags & TH_FIN) len++; if (SEQ_LT(snd_max, tp->snd_nxt)) { /* * The add/update as not been done for the FIN/SYN * yet. */ snd_max = tp->snd_nxt; } } if (SEQ_LEQ((seq_out + len), snd_una)) { /* Are sending an old segment to induce an ack (keep-alive)? */ return; } if (SEQ_LT(seq_out, snd_una)) { /* huh? should we panic? */ uint32_t end; end = seq_out + len; seq_out = snd_una; if (SEQ_GEQ(end, seq_out)) len = end - seq_out; else len = 0; } if (len == 0) { /* We don't log zero window probes */ return; } if (IN_FASTRECOVERY(tp->t_flags)) { rack->r_ctl.rc_prr_out += len; } /* First question is it a retransmission or new? */ if (seq_out == snd_max) { /* Its new */ again: rsm = rack_alloc(rack); if (rsm == NULL) { /* * Hmm out of memory and the tcb got destroyed while * we tried to wait. */ return; } if (th_flags & TH_FIN) { rsm->r_flags = RACK_HAS_FIN|add_flag; } else { rsm->r_flags = add_flag; } if (hw_tls) rsm->r_hw_tls = 1; rsm->r_tim_lastsent[0] = cts; rsm->r_rtr_cnt = 1; rsm->r_rtr_bytes = 0; if (th_flags & TH_SYN) { /* The data space is one beyond snd_una */ rsm->r_flags |= RACK_HAS_SYN; } rsm->r_start = seq_out; rsm->r_end = rsm->r_start + len; rsm->r_dupack = 0; /* * save off the mbuf location that * sndmbuf_noadv returned (which is * where we started copying from).. */ rsm->m = s_mb; rsm->soff = s_moff; /* * Here we do add in the len of send, since its not yet * reflected in in snduna <->snd_max */ rsm->r_fas = (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) + (rsm->r_end - rsm->r_start)); /* rsm->m will be NULL if RACK_HAS_SYN or RACK_HAS_FIN is set */ if (rsm->m) { if (rsm->m->m_len <= rsm->soff) { /* * XXXrrs Question, will this happen? * * If sbsndptr is set at the correct place * then s_moff should always be somewhere * within rsm->m. But if the sbsndptr was * off then that won't be true. If it occurs * we need to walkout to the correct location. */ struct mbuf *lm; lm = rsm->m; while (lm->m_len <= rsm->soff) { rsm->soff -= lm->m_len; lm = lm->m_next; KASSERT(lm != NULL, ("%s rack:%p lm goes null orig_off:%u origmb:%p rsm->soff:%u", __func__, rack, s_moff, s_mb, rsm->soff)); } rsm->m = lm; } rsm->orig_m_len = rsm->m->m_len; } else rsm->orig_m_len = 0; rack_log_retran_reason(rack, rsm, __LINE__, 0, 2); /* Log a new rsm */ rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_NEW, 0, __LINE__); #ifndef INVARIANTS (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); #else insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); if (insret != NULL) { panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p", nrsm, insret, rack, rsm); } #endif TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 1; /* * Special case detection, is there just a single * packet outstanding when we are not in recovery? * * If this is true mark it so. */ if ((IN_FASTRECOVERY(tp->t_flags) == 0) && (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) == ctf_fixed_maxseg(tp))) { struct rack_sendmap *prsm; prsm = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); if (prsm) prsm->r_one_out_nr = 1; } return; } /* * If we reach here its a retransmission and we need to find it. */ memset(&fe, 0, sizeof(fe)); more: if (hintrsm && (hintrsm->r_start == seq_out)) { rsm = hintrsm; hintrsm = NULL; } else { /* No hints sorry */ rsm = NULL; } if ((rsm) && (rsm->r_start == seq_out)) { seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag); if (len == 0) { return; } else { goto more; } } /* Ok it was not the last pointer go through it the hard way. */ refind: fe.r_start = seq_out; rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe); if (rsm) { if (rsm->r_start == seq_out) { seq_out = rack_update_entry(tp, rack, rsm, cts, &len, add_flag); if (len == 0) { return; } else { goto refind; } } if (SEQ_GEQ(seq_out, rsm->r_start) && SEQ_LT(seq_out, rsm->r_end)) { /* Transmitted within this piece */ /* * Ok we must split off the front and then let the * update do the rest */ nrsm = rack_alloc_full_limit(rack); if (nrsm == NULL) { rack_update_rsm(tp, rack, rsm, cts, add_flag); return; } /* * copy rsm to nrsm and then trim the front of rsm * to not include this part. */ rack_clone_rsm(rack, nrsm, rsm, seq_out); rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SPLIT, 0, __LINE__); #ifndef INVARIANTS (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); #else insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); if (insret != NULL) { panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p", nrsm, insret, rack, rsm); } #endif if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } rsm->r_flags &= (~RACK_HAS_FIN); seq_out = rack_update_entry(tp, rack, nrsm, cts, &len, add_flag); if (len == 0) { return; } else if (len > 0) goto refind; } } /* * Hmm not found in map did they retransmit both old and on into the * new? */ if (seq_out == tp->snd_max) { goto again; } else if (SEQ_LT(seq_out, tp->snd_max)) { #ifdef INVARIANTS printf("seq_out:%u len:%d snd_una:%u snd_max:%u -- but rsm not found?\n", seq_out, len, tp->snd_una, tp->snd_max); printf("Starting Dump of all rack entries\n"); RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) { printf("rsm:%p start:%u end:%u\n", rsm, rsm->r_start, rsm->r_end); } printf("Dump complete\n"); panic("seq_out not found rack:%p tp:%p", rack, tp); #endif } else { #ifdef INVARIANTS /* * Hmm beyond sndmax? (only if we are using the new rtt-pack * flag) */ panic("seq_out:%u(%d) is beyond snd_max:%u tp:%p", seq_out, len, tp->snd_max, tp); #endif } } /* * Record one of the RTT updates from an ack into * our sample structure. */ static void tcp_rack_xmit_timer(struct tcp_rack *rack, int32_t rtt, uint32_t len, uint32_t us_rtt, int confidence, struct rack_sendmap *rsm, uint16_t rtrcnt) { if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) || (rack->r_ctl.rack_rs.rs_rtt_lowest > rtt)) { rack->r_ctl.rack_rs.rs_rtt_lowest = rtt; } if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) || (rack->r_ctl.rack_rs.rs_rtt_highest < rtt)) { rack->r_ctl.rack_rs.rs_rtt_highest = rtt; } if (rack->rc_tp->t_flags & TF_GPUTINPROG) { if (us_rtt < rack->r_ctl.rc_gp_lowrtt) rack->r_ctl.rc_gp_lowrtt = us_rtt; if (rack->rc_tp->snd_wnd > rack->r_ctl.rc_gp_high_rwnd) rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd; } if ((confidence == 1) && ((rsm == NULL) || (rsm->r_just_ret) || (rsm->r_one_out_nr && len < (ctf_fixed_maxseg(rack->rc_tp) * 2)))) { /* * If the rsm had a just return * hit it then we can't trust the * rtt measurement for buffer deterimination * Note that a confidence of 2, indicates * SACK'd which overrides the r_just_ret or * the r_one_out_nr. If it was a CUM-ACK and * we had only two outstanding, but get an * ack for only 1. Then that also lowers our * confidence. */ confidence = 0; } if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) || (rack->r_ctl.rack_rs.rs_us_rtt > us_rtt)) { if (rack->r_ctl.rack_rs.confidence == 0) { /* * We take anything with no current confidence * saved. */ rack->r_ctl.rack_rs.rs_us_rtt = us_rtt; rack->r_ctl.rack_rs.confidence = confidence; rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt; } else if (confidence || rack->r_ctl.rack_rs.confidence) { /* * Once we have a confident number, * we can update it with a smaller * value since this confident number * may include the DSACK time until * the next segment (the second one) arrived. */ rack->r_ctl.rack_rs.rs_us_rtt = us_rtt; rack->r_ctl.rack_rs.confidence = confidence; rack->r_ctl.rack_rs.rs_us_rtrcnt = rtrcnt; } } rack_log_rtt_upd(rack->rc_tp, rack, us_rtt, len, rsm, confidence); rack->r_ctl.rack_rs.rs_flags = RACK_RTT_VALID; rack->r_ctl.rack_rs.rs_rtt_tot += rtt; rack->r_ctl.rack_rs.rs_rtt_cnt++; } /* * Collect new round-trip time estimate * and update averages and current timeout. */ static void tcp_rack_xmit_timer_commit(struct tcp_rack *rack, struct tcpcb *tp) { int32_t delta; int32_t rtt; if (rack->r_ctl.rack_rs.rs_flags & RACK_RTT_EMPTY) /* No valid sample */ return; if (rack->r_ctl.rc_rate_sample_method == USE_RTT_LOW) { /* We are to use the lowest RTT seen in a single ack */ rtt = rack->r_ctl.rack_rs.rs_rtt_lowest; } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_HIGH) { /* We are to use the highest RTT seen in a single ack */ rtt = rack->r_ctl.rack_rs.rs_rtt_highest; } else if (rack->r_ctl.rc_rate_sample_method == USE_RTT_AVG) { /* We are to use the average RTT seen in a single ack */ rtt = (int32_t)(rack->r_ctl.rack_rs.rs_rtt_tot / (uint64_t)rack->r_ctl.rack_rs.rs_rtt_cnt); } else { #ifdef INVARIANTS panic("Unknown rtt variant %d", rack->r_ctl.rc_rate_sample_method); #endif return; } if (rtt == 0) rtt = 1; if (rack->rc_gp_rtt_set == 0) { /* * With no RTT we have to accept * even one we are not confident of. */ rack->r_ctl.rc_gp_srtt = rack->r_ctl.rack_rs.rs_us_rtt; rack->rc_gp_rtt_set = 1; } else if (rack->r_ctl.rack_rs.confidence) { /* update the running gp srtt */ rack->r_ctl.rc_gp_srtt -= (rack->r_ctl.rc_gp_srtt/8); rack->r_ctl.rc_gp_srtt += rack->r_ctl.rack_rs.rs_us_rtt / 8; } if (rack->r_ctl.rack_rs.confidence) { /* * record the low and high for highly buffered path computation, * we only do this if we are confident (not a retransmission). */ if (rack->r_ctl.rc_highest_us_rtt < rack->r_ctl.rack_rs.rs_us_rtt) { rack->r_ctl.rc_highest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt; } if (rack->rc_highly_buffered == 0) { /* * Currently once we declare a path has * highly buffered there is no going * back, which may be a problem... */ if ((rack->r_ctl.rc_highest_us_rtt / rack->r_ctl.rc_lowest_us_rtt) > rack_hbp_thresh) { rack_log_rtt_shrinks(rack, rack->r_ctl.rack_rs.rs_us_rtt, rack->r_ctl.rc_highest_us_rtt, rack->r_ctl.rc_lowest_us_rtt, RACK_RTTS_SEEHBP); rack->rc_highly_buffered = 1; } } } if ((rack->r_ctl.rack_rs.confidence) || (rack->r_ctl.rack_rs.rs_us_rtrcnt == 1)) { /* * If we are highly confident of it it was * never retransmitted we accept it as the last us_rtt. */ rack->r_ctl.rc_last_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt; /* The lowest rtt can be set if its was not retransmited */ if (rack->r_ctl.rc_lowest_us_rtt > rack->r_ctl.rack_rs.rs_us_rtt) { rack->r_ctl.rc_lowest_us_rtt = rack->r_ctl.rack_rs.rs_us_rtt; if (rack->r_ctl.rc_lowest_us_rtt == 0) rack->r_ctl.rc_lowest_us_rtt = 1; } } rack = (struct tcp_rack *)tp->t_fb_ptr; if (tp->t_srtt != 0) { /* * We keep a simple srtt in microseconds, like our rtt * measurement. We don't need to do any tricks with shifting * etc. Instead we just add in 1/8th of the new measurement * and subtract out 1/8 of the old srtt. We do the same with * the variance after finding the absolute value of the * difference between this sample and the current srtt. */ delta = tp->t_srtt - rtt; /* Take off 1/8th of the current sRTT */ tp->t_srtt -= (tp->t_srtt >> 3); /* Add in 1/8th of the new RTT just measured */ tp->t_srtt += (rtt >> 3); if (tp->t_srtt <= 0) tp->t_srtt = 1; /* Now lets make the absolute value of the variance */ if (delta < 0) delta = -delta; /* Subtract out 1/8th */ tp->t_rttvar -= (tp->t_rttvar >> 3); /* Add in 1/8th of the new variance we just saw */ tp->t_rttvar += (delta >> 3); if (tp->t_rttvar <= 0) tp->t_rttvar = 1; if (tp->t_rttbest > tp->t_srtt + tp->t_rttvar) tp->t_rttbest = tp->t_srtt + tp->t_rttvar; } else { /* * No rtt measurement yet - use the unsmoothed rtt. Set the * variance to half the rtt (so our first retransmit happens * at 3*rtt). */ tp->t_srtt = rtt; tp->t_rttvar = rtt >> 1; tp->t_rttbest = tp->t_srtt + tp->t_rttvar; } rack->rc_srtt_measure_made = 1; KMOD_TCPSTAT_INC(tcps_rttupdated); tp->t_rttupdated++; #ifdef STATS if (rack_stats_gets_ms_rtt == 0) { /* Send in the microsecond rtt used for rxt timeout purposes */ stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt)); } else if (rack_stats_gets_ms_rtt == 1) { /* Send in the millisecond rtt used for rxt timeout purposes */ int32_t ms_rtt; /* Round up */ ms_rtt = (rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC; stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt)); } else if (rack_stats_gets_ms_rtt == 2) { /* Send in the millisecond rtt has close to the path RTT as we can get */ int32_t ms_rtt; /* Round up */ ms_rtt = (rack->r_ctl.rack_rs.rs_us_rtt + HPTS_USEC_IN_MSEC - 1) / HPTS_USEC_IN_MSEC; stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, ms_rtt)); } else { /* Send in the microsecond rtt has close to the path RTT as we can get */ stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rack->r_ctl.rack_rs.rs_us_rtt)); } #endif /* * the retransmit should happen at rtt + 4 * rttvar. Because of the * way we do the smoothing, srtt and rttvar will each average +1/2 * tick of bias. When we compute the retransmit timer, we want 1/2 * tick of rounding and 1 extra tick because of +-1/2 tick * uncertainty in the firing of the timer. The bias will give us * exactly the 1.5 tick we need. But, because the bias is * statistical, we have to test that we don't drop below the minimum * feasible timer (which is 2 ticks). */ tp->t_rxtshift = 0; RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp), max(rack_rto_min, rtt + 2), rack_rto_max, rack->r_ctl.timer_slop); rack_log_rtt_sample(rack, rtt); tp->t_softerror = 0; } static void rack_apply_updated_usrtt(struct tcp_rack *rack, uint32_t us_rtt, uint32_t us_cts) { /* * Apply to filter the inbound us-rtt at us_cts. */ uint32_t old_rtt; old_rtt = get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt); apply_filter_min_small(&rack->r_ctl.rc_gp_min_rtt, us_rtt, us_cts); if (old_rtt > us_rtt) { /* We just hit a new lower rtt time */ rack_log_rtt_shrinks(rack, us_cts, old_rtt, __LINE__, RACK_RTTS_NEWRTT); /* * Only count it if its lower than what we saw within our * calculated range. */ if ((old_rtt - us_rtt) > rack_min_rtt_movement) { if (rack_probertt_lower_within && rack->rc_gp_dyn_mul && (rack->use_fixed_rate == 0) && (rack->rc_always_pace)) { /* * We are seeing a new lower rtt very close * to the time that we would have entered probe-rtt. * This is probably due to the fact that a peer flow * has entered probe-rtt. Lets go in now too. */ uint32_t val; val = rack_probertt_lower_within * rack_time_between_probertt; val /= 100; if ((rack->in_probe_rtt == 0) && ((us_cts - rack->r_ctl.rc_lower_rtt_us_cts) >= (rack_time_between_probertt - val))) { rack_enter_probertt(rack, us_cts); } } rack->r_ctl.rc_lower_rtt_us_cts = us_cts; } } } static int rack_update_rtt(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm, struct tcpopt *to, uint32_t cts, int32_t ack_type, tcp_seq th_ack) { uint32_t us_rtt; int32_t i, all; uint32_t t, len_acked; if ((rsm->r_flags & RACK_ACKED) || (rsm->r_flags & RACK_WAS_ACKED)) /* Already done */ return (0); if (rsm->r_no_rtt_allowed) { /* Not allowed */ return (0); } if (ack_type == CUM_ACKED) { if (SEQ_GT(th_ack, rsm->r_end)) { len_acked = rsm->r_end - rsm->r_start; all = 1; } else { len_acked = th_ack - rsm->r_start; all = 0; } } else { len_acked = rsm->r_end - rsm->r_start; all = 0; } if (rsm->r_rtr_cnt == 1) { t = cts - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; if ((int)t <= 0) t = 1; if (!tp->t_rttlow || tp->t_rttlow > t) tp->t_rttlow = t; if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) { rack->r_ctl.rc_rack_min_rtt = t; if (rack->r_ctl.rc_rack_min_rtt == 0) { rack->r_ctl.rc_rack_min_rtt = 1; } } if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])) us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]; else us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]; if (us_rtt == 0) us_rtt = 1; if (CC_ALGO(tp)->rttsample != NULL) { /* Kick the RTT to the CC */ CC_ALGO(tp)->rttsample(tp->ccv, us_rtt, 1, rsm->r_fas); } rack_apply_updated_usrtt(rack, us_rtt, tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time)); if (ack_type == SACKED) { rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 1); tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, 2 , rsm, rsm->r_rtr_cnt); } else { /* * We need to setup what our confidence * is in this ack. * * If the rsm was app limited and it is * less than a mss in length (the end * of the send) then we have a gap. If we * were app limited but say we were sending * multiple MSS's then we are more confident * int it. * * When we are not app-limited then we see if * the rsm is being included in the current * measurement, we tell this by the app_limited_needs_set * flag. * * Note that being cwnd blocked is not applimited * as well as the pacing delay between packets which * are sending only 1 or 2 MSS's also will show up * in the RTT. We probably need to examine this algorithm * a bit more and enhance it to account for the delay * between rsm's. We could do that by saving off the * pacing delay of each rsm (in an rsm) and then * factoring that in somehow though for now I am * not sure how :) */ int calc_conf = 0; if (rsm->r_flags & RACK_APP_LIMITED) { if (all && (len_acked <= ctf_fixed_maxseg(tp))) calc_conf = 0; else calc_conf = 1; } else if (rack->app_limited_needs_set == 0) { calc_conf = 1; } else { calc_conf = 0; } rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)], cts, 2); tcp_rack_xmit_timer(rack, t + 1, len_acked, us_rtt, calc_conf, rsm, rsm->r_rtr_cnt); } if ((rsm->r_flags & RACK_TLP) && (!IN_FASTRECOVERY(tp->t_flags))) { /* Segment was a TLP and our retrans matched */ if (rack->r_ctl.rc_tlp_cwnd_reduce) { rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__); } } if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) { /* New more recent rack_tmit_time */ rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; rack->rc_rack_rtt = t; } return (1); } /* * We clear the soft/rxtshift since we got an ack. * There is no assurance we will call the commit() function * so we need to clear these to avoid incorrect handling. */ tp->t_rxtshift = 0; RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp), rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop); tp->t_softerror = 0; if (to && (to->to_flags & TOF_TS) && (ack_type == CUM_ACKED) && (to->to_tsecr) && ((rsm->r_flags & RACK_OVERMAX) == 0)) { /* * Now which timestamp does it match? In this block the ACK * must be coming from a previous transmission. */ for (i = 0; i < rsm->r_rtr_cnt; i++) { if (rack_ts_to_msec(rsm->r_tim_lastsent[i]) == to->to_tsecr) { t = cts - (uint32_t)rsm->r_tim_lastsent[i]; if ((int)t <= 0) t = 1; if (CC_ALGO(tp)->rttsample != NULL) { /* * Kick the RTT to the CC, here * we lie a bit in that we know the * retransmission is correct even though * we retransmitted. This is because * we match the timestamps. */ if (TSTMP_GT(tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time), rsm->r_tim_lastsent[i])) us_rtt = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time) - (uint32_t)rsm->r_tim_lastsent[i]; else us_rtt = tcp_get_usecs(NULL) - (uint32_t)rsm->r_tim_lastsent[i]; CC_ALGO(tp)->rttsample(tp->ccv, us_rtt, 1, rsm->r_fas); } if ((i + 1) < rsm->r_rtr_cnt) { /* * The peer ack'd from our previous * transmission. We have a spurious * retransmission and thus we dont * want to update our rack_rtt. * * Hmm should there be a CC revert here? * */ return (0); } if (!tp->t_rttlow || tp->t_rttlow > t) tp->t_rttlow = t; if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) { rack->r_ctl.rc_rack_min_rtt = t; if (rack->r_ctl.rc_rack_min_rtt == 0) { rack->r_ctl.rc_rack_min_rtt = 1; } } if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) { /* New more recent rack_tmit_time */ rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; rack->rc_rack_rtt = t; } rack_log_rtt_sample_calc(rack, t, (uint32_t)rsm->r_tim_lastsent[i], cts, 3); tcp_rack_xmit_timer(rack, t + 1, len_acked, t, 0, rsm, rsm->r_rtr_cnt); return (1); } } goto ts_not_found; } else { /* * Ok its a SACK block that we retransmitted. or a windows * machine without timestamps. We can tell nothing from the * time-stamp since its not there or the time the peer last * recieved a segment that moved forward its cum-ack point. */ ts_not_found: i = rsm->r_rtr_cnt - 1; t = cts - (uint32_t)rsm->r_tim_lastsent[i]; if ((int)t <= 0) t = 1; if (rack->r_ctl.rc_rack_min_rtt && SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) { /* * We retransmitted and the ack came back in less * than the smallest rtt we have observed. We most * likely did an improper retransmit as outlined in * 6.2 Step 2 point 2 in the rack-draft so we * don't want to update our rack_rtt. We in * theory (in future) might want to think about reverting our * cwnd state but we won't for now. */ return (0); } else if (rack->r_ctl.rc_rack_min_rtt) { /* * We retransmitted it and the retransmit did the * job. */ if (!rack->r_ctl.rc_rack_min_rtt || SEQ_LT(t, rack->r_ctl.rc_rack_min_rtt)) { rack->r_ctl.rc_rack_min_rtt = t; if (rack->r_ctl.rc_rack_min_rtt == 0) { rack->r_ctl.rc_rack_min_rtt = 1; } } if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, (uint32_t)rsm->r_tim_lastsent[i])) { /* New more recent rack_tmit_time */ rack->r_ctl.rc_rack_tmit_time = (uint32_t)rsm->r_tim_lastsent[i]; rack->rc_rack_rtt = t; } return (1); } } return (0); } /* * Mark the SACK_PASSED flag on all entries prior to rsm send wise. */ static void rack_log_sack_passed(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm) { struct rack_sendmap *nrsm; nrsm = rsm; TAILQ_FOREACH_REVERSE_FROM(nrsm, &rack->r_ctl.rc_tmap, rack_head, r_tnext) { if (nrsm == rsm) { /* Skip orginal segment he is acked */ continue; } if (nrsm->r_flags & RACK_ACKED) { /* * Skip ack'd segments, though we * should not see these, since tmap * should not have ack'd segments. */ continue; } if (nrsm->r_flags & RACK_RWND_COLLAPSED) { /* * If the peer dropped the rwnd on * these then we don't worry about them. */ continue; } if (nrsm->r_flags & RACK_SACK_PASSED) { /* * We found one that is already marked * passed, we have been here before and * so all others below this are marked. */ break; } nrsm->r_flags |= RACK_SACK_PASSED; nrsm->r_flags &= ~RACK_WAS_SACKPASS; } } static void rack_need_set_test(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack, int line, int use_which) { if ((tp->t_flags & TF_GPUTINPROG) && SEQ_GEQ(rsm->r_end, tp->gput_seq)) { /* * We were app limited, and this ack * butts up or goes beyond the point where we want * to start our next measurement. We need * to record the new gput_ts as here and * possibly update the start sequence. */ uint32_t seq, ts; if (rsm->r_rtr_cnt > 1) { /* * This is a retransmit, can we * really make any assessment at this * point? We are not really sure of * the timestamp, is it this or the * previous transmission? * * Lets wait for something better that * is not retransmitted. */ return; } seq = tp->gput_seq; ts = tp->gput_ts; rack->app_limited_needs_set = 0; tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time); /* Do we start at a new end? */ if ((use_which == RACK_USE_BEG) && SEQ_GEQ(rsm->r_start, tp->gput_seq)) { /* * When we get an ACK that just eats * up some of the rsm, we set RACK_USE_BEG * since whats at r_start (i.e. th_ack) * is left unacked and thats where the * measurement not starts. */ tp->gput_seq = rsm->r_start; rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]; } if ((use_which == RACK_USE_END) && SEQ_GEQ(rsm->r_end, tp->gput_seq)) { /* * We use the end when the cumack * is moving forward and completely * deleting the rsm passed so basically * r_end holds th_ack. * * For SACK's we also want to use the end * since this piece just got sacked and * we want to target anything after that * in our measurement. */ tp->gput_seq = rsm->r_end; rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]; } if (use_which == RACK_USE_END_OR_THACK) { /* * special case for ack moving forward, * not a sack, we need to move all the * way up to where this ack cum-ack moves * to. */ if (SEQ_GT(th_ack, rsm->r_end)) tp->gput_seq = th_ack; else tp->gput_seq = rsm->r_end; rack->r_ctl.rc_gp_output_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]; } if (SEQ_GT(tp->gput_seq, tp->gput_ack)) { /* * We moved beyond this guy's range, re-calculate * the new end point. */ if (rack->rc_gp_filled == 0) { tp->gput_ack = tp->gput_seq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp))); } else { tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack); } } /* * We are moving the goal post, we may be able to clear the * measure_saw_probe_rtt flag. */ if ((rack->in_probe_rtt == 0) && (rack->measure_saw_probe_rtt) && (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit))) rack->measure_saw_probe_rtt = 0; rack_log_pacing_delay_calc(rack, ts, tp->gput_ts, seq, tp->gput_seq, 0, 5, line, NULL, 0); if (rack->rc_gp_filled && ((tp->gput_ack - tp->gput_seq) < max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp))))) { uint32_t ideal_amount; ideal_amount = rack_get_measure_window(tp, rack); if (ideal_amount > sbavail(&tp->t_inpcb->inp_socket->so_snd)) { /* * There is no sense of continuing this measurement * because its too small to gain us anything we * trust. Skip it and that way we can start a new * measurement quicker. */ tp->t_flags &= ~TF_GPUTINPROG; rack_log_pacing_delay_calc(rack, tp->gput_ack, tp->gput_seq, 0, 0, 0, 6, __LINE__, NULL, 0); } else { /* * Reset the window further out. */ tp->gput_ack = tp->gput_seq + ideal_amount; } } } } static inline int is_rsm_inside_declared_tlp_block(struct tcp_rack *rack, struct rack_sendmap *rsm) { if (SEQ_LT(rsm->r_end, rack->r_ctl.last_tlp_acked_start)) { /* Behind our TLP definition or right at */ return (0); } if (SEQ_GT(rsm->r_start, rack->r_ctl.last_tlp_acked_end)) { /* The start is beyond or right at our end of TLP definition */ return (0); } /* It has to be a sub-part of the original TLP recorded */ return (1); } static uint32_t rack_proc_sack_blk(struct tcpcb *tp, struct tcp_rack *rack, struct sackblk *sack, struct tcpopt *to, struct rack_sendmap **prsm, uint32_t cts, int *moved_two) { uint32_t start, end, changed = 0; struct rack_sendmap stack_map; struct rack_sendmap *rsm, *nrsm, fe, *prev, *next; #ifdef INVARIANTS struct rack_sendmap *insret; #endif int32_t used_ref = 1; int moved = 0; start = sack->start; end = sack->end; rsm = *prsm; memset(&fe, 0, sizeof(fe)); do_rest_ofb: if ((rsm == NULL) || (SEQ_LT(end, rsm->r_start)) || (SEQ_GEQ(start, rsm->r_end)) || (SEQ_LT(start, rsm->r_start))) { /* * We are not in the right spot, * find the correct spot in the tree. */ used_ref = 0; fe.r_start = start; rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe); moved++; } if (rsm == NULL) { /* TSNH */ goto out; } /* Ok we have an ACK for some piece of this rsm */ if (rsm->r_start != start) { if ((rsm->r_flags & RACK_ACKED) == 0) { /* * Before any splitting or hookery is * done is it a TLP of interest i.e. rxt? */ if ((rsm->r_flags & RACK_TLP) && (rsm->r_rtr_cnt > 1)) { /* * We are splitting a rxt TLP, check * if we need to save off the start/end */ if (rack->rc_last_tlp_acked_set && (is_rsm_inside_declared_tlp_block(rack, rsm))) { /* * We already turned this on since we are inside * the previous one was a partially sack now we * are getting another one (maybe all of it). * */ rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end); /* * Lets make sure we have all of it though. */ if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) { rack->r_ctl.last_tlp_acked_start = rsm->r_start; rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start, rack->r_ctl.last_tlp_acked_end); } if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) { rack->r_ctl.last_tlp_acked_end = rsm->r_end; rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start, rack->r_ctl.last_tlp_acked_end); } } else { rack->r_ctl.last_tlp_acked_start = rsm->r_start; rack->r_ctl.last_tlp_acked_end = rsm->r_end; rack->rc_last_tlp_past_cumack = 0; rack->rc_last_tlp_acked_set = 1; rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end); } } /** * Need to split this in two pieces the before and after, * the before remains in the map, the after must be * added. In other words we have: * rsm |--------------| * sackblk |-------> * rsm will become * rsm |---| * and nrsm will be the sacked piece * nrsm |----------| * * But before we start down that path lets * see if the sack spans over on top of * the next guy and it is already sacked. * */ next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); if (next && (next->r_flags & RACK_ACKED) && SEQ_GEQ(end, next->r_start)) { /** * So the next one is already acked, and * we can thus by hookery use our stack_map * to reflect the piece being sacked and * then adjust the two tree entries moving * the start and ends around. So we start like: * rsm |------------| (not-acked) * next |-----------| (acked) * sackblk |--------> * We want to end like so: * rsm |------| (not-acked) * next |-----------------| (acked) * nrsm |-----| * Where nrsm is a temporary stack piece we * use to update all the gizmos. */ /* Copy up our fudge block */ nrsm = &stack_map; memcpy(nrsm, rsm, sizeof(struct rack_sendmap)); /* Now adjust our tree blocks */ rsm->r_end = start; next->r_start = start; /* Now we must adjust back where next->m is */ rack_setup_offset_for_rsm(rsm, next); /* We don't need to adjust rsm, it did not change */ /* Clear out the dup ack count of the remainder */ rsm->r_dupack = 0; rsm->r_just_ret = 0; rack_log_retran_reason(rack, rsm, __LINE__, 0, 2); /* Now lets make sure our fudge block is right */ nrsm->r_start = start; /* Now lets update all the stats and such */ rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0); if (rack->app_limited_needs_set) rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END); changed += (nrsm->r_end - nrsm->r_start); rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start); if (nrsm->r_flags & RACK_SACK_PASSED) { rack->r_ctl.rc_reorder_ts = cts; } /* * Now we want to go up from rsm (the * one left un-acked) to the next one * in the tmap. We do this so when * we walk backwards we include marking * sack-passed on rsm (The one passed in * is skipped since it is generally called * on something sacked before removing it * from the tmap). */ if (rsm->r_in_tmap) { nrsm = TAILQ_NEXT(rsm, r_tnext); /* * Now that we have the next * one walk backwards from there. */ if (nrsm && nrsm->r_in_tmap) rack_log_sack_passed(tp, rack, nrsm); } /* Now are we done? */ if (SEQ_LT(end, next->r_end) || (end == next->r_end)) { /* Done with block */ goto out; } rack_log_map_chg(tp, rack, &stack_map, rsm, next, MAP_SACK_M1, end, __LINE__); counter_u64_add(rack_sack_used_next_merge, 1); /* Postion for the next block */ start = next->r_end; rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, next); if (rsm == NULL) goto out; } else { /** * We can't use any hookery here, so we * need to split the map. We enter like * so: * rsm |--------| * sackblk |-----> * We will add the new block nrsm and * that will be the new portion, and then * fall through after reseting rsm. So we * split and look like this: * rsm |----| * sackblk |-----> * nrsm |---| * We then fall through reseting * rsm to nrsm, so the next block * picks it up. */ nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT); if (nrsm == NULL) { /* * failed XXXrrs what can we do but loose the sack * info? */ goto out; } counter_u64_add(rack_sack_splits, 1); rack_clone_rsm(rack, nrsm, rsm, start); rsm->r_just_ret = 0; #ifndef INVARIANTS (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); #else insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); if (insret != NULL) { panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p", nrsm, insret, rack, rsm); } #endif if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M2, end, __LINE__); rsm->r_flags &= (~RACK_HAS_FIN); /* Position us to point to the new nrsm that starts the sack blk */ rsm = nrsm; } } else { /* Already sacked this piece */ counter_u64_add(rack_sack_skipped_acked, 1); moved++; if (end == rsm->r_end) { /* Done with block */ rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); goto out; } else if (SEQ_LT(end, rsm->r_end)) { /* A partial sack to a already sacked block */ moved++; rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); goto out; } else { /* * The end goes beyond this guy * reposition the start to the * next block. */ start = rsm->r_end; rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); if (rsm == NULL) goto out; } } } if (SEQ_GEQ(end, rsm->r_end)) { /** * The end of this block is either beyond this guy or right * at this guy. I.e.: * rsm --- |-----| * end |-----| * * end |---------| */ if ((rsm->r_flags & RACK_ACKED) == 0) { /* * Is it a TLP of interest? */ if ((rsm->r_flags & RACK_TLP) && (rsm->r_rtr_cnt > 1)) { /* * We are splitting a rxt TLP, check * if we need to save off the start/end */ if (rack->rc_last_tlp_acked_set && (is_rsm_inside_declared_tlp_block(rack, rsm))) { /* * We already turned this on since we are inside * the previous one was a partially sack now we * are getting another one (maybe all of it). */ rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end); /* * Lets make sure we have all of it though. */ if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) { rack->r_ctl.last_tlp_acked_start = rsm->r_start; rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start, rack->r_ctl.last_tlp_acked_end); } if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) { rack->r_ctl.last_tlp_acked_end = rsm->r_end; rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start, rack->r_ctl.last_tlp_acked_end); } } else { rack->r_ctl.last_tlp_acked_start = rsm->r_start; rack->r_ctl.last_tlp_acked_end = rsm->r_end; rack->rc_last_tlp_past_cumack = 0; rack->rc_last_tlp_acked_set = 1; rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end); } } rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0); changed += (rsm->r_end - rsm->r_start); rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start); if (rsm->r_in_tmap) /* should be true */ rack_log_sack_passed(tp, rack, rsm); /* Is Reordering occuring? */ if (rsm->r_flags & RACK_SACK_PASSED) { rsm->r_flags &= ~RACK_SACK_PASSED; rack->r_ctl.rc_reorder_ts = cts; } if (rack->app_limited_needs_set) rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END); rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time); rsm->r_flags |= RACK_ACKED; if (rsm->r_in_tmap) { TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 0; } rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_SACK_M3, end, __LINE__); } else { counter_u64_add(rack_sack_skipped_acked, 1); moved++; } if (end == rsm->r_end) { /* This block only - done, setup for next */ goto out; } /* * There is more not coverend by this rsm move on * to the next block in the RB tree. */ nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); start = rsm->r_end; rsm = nrsm; if (rsm == NULL) goto out; goto do_rest_ofb; } /** * The end of this sack block is smaller than * our rsm i.e.: * rsm --- |-----| * end |--| */ if ((rsm->r_flags & RACK_ACKED) == 0) { /* * Is it a TLP of interest? */ if ((rsm->r_flags & RACK_TLP) && (rsm->r_rtr_cnt > 1)) { /* * We are splitting a rxt TLP, check * if we need to save off the start/end */ if (rack->rc_last_tlp_acked_set && (is_rsm_inside_declared_tlp_block(rack, rsm))) { /* * We already turned this on since we are inside * the previous one was a partially sack now we * are getting another one (maybe all of it). */ rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end); /* * Lets make sure we have all of it though. */ if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) { rack->r_ctl.last_tlp_acked_start = rsm->r_start; rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start, rack->r_ctl.last_tlp_acked_end); } if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) { rack->r_ctl.last_tlp_acked_end = rsm->r_end; rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start, rack->r_ctl.last_tlp_acked_end); } } else { rack->r_ctl.last_tlp_acked_start = rsm->r_start; rack->r_ctl.last_tlp_acked_end = rsm->r_end; rack->rc_last_tlp_past_cumack = 0; rack->rc_last_tlp_acked_set = 1; rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end); } } prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); if (prev && (prev->r_flags & RACK_ACKED)) { /** * Goal, we want the right remainder of rsm to shrink * in place and span from (rsm->r_start = end) to rsm->r_end. * We want to expand prev to go all the way * to prev->r_end <- end. * so in the tree we have before: * prev |--------| (acked) * rsm |-------| (non-acked) * sackblk |-| * We churn it so we end up with * prev |----------| (acked) * rsm |-----| (non-acked) * nrsm |-| (temporary) * * Note if either prev/rsm is a TLP we don't * do this. */ nrsm = &stack_map; memcpy(nrsm, rsm, sizeof(struct rack_sendmap)); prev->r_end = end; rsm->r_start = end; /* Now adjust nrsm (stack copy) to be * the one that is the small * piece that was "sacked". */ nrsm->r_end = end; rsm->r_dupack = 0; rack_log_retran_reason(rack, rsm, __LINE__, 0, 2); /* * Now that the rsm has had its start moved forward * lets go ahead and get its new place in the world. */ rack_setup_offset_for_rsm(prev, rsm); /* * Now nrsm is our new little piece * that is acked (which was merged * to prev). Update the rtt and changed * based on that. Also check for reordering. */ rack_update_rtt(tp, rack, nrsm, to, cts, SACKED, 0); if (rack->app_limited_needs_set) rack_need_set_test(tp, rack, nrsm, tp->snd_una, __LINE__, RACK_USE_END); changed += (nrsm->r_end - nrsm->r_start); rack->r_ctl.rc_sacked += (nrsm->r_end - nrsm->r_start); if (nrsm->r_flags & RACK_SACK_PASSED) { rack->r_ctl.rc_reorder_ts = cts; } rack_log_map_chg(tp, rack, prev, &stack_map, rsm, MAP_SACK_M4, end, __LINE__); rsm = prev; counter_u64_add(rack_sack_used_prev_merge, 1); } else { /** * This is the case where our previous * block is not acked either, so we must * split the block in two. */ nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT); if (nrsm == NULL) { /* failed rrs what can we do but loose the sack info? */ goto out; } if ((rsm->r_flags & RACK_TLP) && (rsm->r_rtr_cnt > 1)) { /* * We are splitting a rxt TLP, check * if we need to save off the start/end */ if (rack->rc_last_tlp_acked_set && (is_rsm_inside_declared_tlp_block(rack, rsm))) { /* * We already turned this on since this block is inside * the previous one was a partially sack now we * are getting another one (maybe all of it). */ rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end); /* * Lets make sure we have all of it though. */ if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) { rack->r_ctl.last_tlp_acked_start = rsm->r_start; rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start, rack->r_ctl.last_tlp_acked_end); } if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) { rack->r_ctl.last_tlp_acked_end = rsm->r_end; rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start, rack->r_ctl.last_tlp_acked_end); } } else { rack->r_ctl.last_tlp_acked_start = rsm->r_start; rack->r_ctl.last_tlp_acked_end = rsm->r_end; rack->rc_last_tlp_acked_set = 1; rack->rc_last_tlp_past_cumack = 0; rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end); } } /** * In this case nrsm becomes * nrsm->r_start = end; * nrsm->r_end = rsm->r_end; * which is un-acked. * * rsm->r_end = nrsm->r_start; * i.e. the remaining un-acked * piece is left on the left * hand side. * * So we start like this * rsm |----------| (not acked) * sackblk |---| * build it so we have * rsm |---| (acked) * nrsm |------| (not acked) */ counter_u64_add(rack_sack_splits, 1); rack_clone_rsm(rack, nrsm, rsm, end); rsm->r_flags &= (~RACK_HAS_FIN); rsm->r_just_ret = 0; #ifndef INVARIANTS (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); #else insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); if (insret != NULL) { panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p", nrsm, insret, rack, rsm); } #endif if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } nrsm->r_dupack = 0; rack_log_retran_reason(rack, nrsm, __LINE__, 0, 2); rack_update_rtt(tp, rack, rsm, to, cts, SACKED, 0); changed += (rsm->r_end - rsm->r_start); rack->r_ctl.rc_sacked += (rsm->r_end - rsm->r_start); if (rsm->r_in_tmap) /* should be true */ rack_log_sack_passed(tp, rack, rsm); /* Is Reordering occuring? */ if (rsm->r_flags & RACK_SACK_PASSED) { rsm->r_flags &= ~RACK_SACK_PASSED; rack->r_ctl.rc_reorder_ts = cts; } if (rack->app_limited_needs_set) rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_END); rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time); rsm->r_flags |= RACK_ACKED; rack_log_map_chg(tp, rack, NULL, rsm, nrsm, MAP_SACK_M5, end, __LINE__); if (rsm->r_in_tmap) { TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 0; } } } else if (start != end){ /* * The block was already acked. */ counter_u64_add(rack_sack_skipped_acked, 1); moved++; } out: if (rsm && ((rsm->r_flags & RACK_TLP) == 0) && (rsm->r_flags & RACK_ACKED)) { /* * Now can we merge where we worked * with either the previous or * next block? */ next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); while (next) { if (next->r_flags & RACK_TLP) break; if (next->r_flags & RACK_ACKED) { /* yep this and next can be merged */ rsm = rack_merge_rsm(rack, rsm, next); next = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); } else break; } /* Now what about the previous? */ prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); while (prev) { if (prev->r_flags & RACK_TLP) break; if (prev->r_flags & RACK_ACKED) { /* yep the previous and this can be merged */ rsm = rack_merge_rsm(rack, prev, rsm); prev = RB_PREV(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); } else break; } } if (used_ref == 0) { counter_u64_add(rack_sack_proc_all, 1); } else { counter_u64_add(rack_sack_proc_short, 1); } /* Save off the next one for quick reference. */ if (rsm) nrsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); else nrsm = NULL; *prsm = rack->r_ctl.rc_sacklast = nrsm; /* Pass back the moved. */ *moved_two = moved; return (changed); } static void inline rack_peer_reneges(struct tcp_rack *rack, struct rack_sendmap *rsm, tcp_seq th_ack) { struct rack_sendmap *tmap; tmap = NULL; while (rsm && (rsm->r_flags & RACK_ACKED)) { /* Its no longer sacked, mark it so */ rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); #ifdef INVARIANTS if (rsm->r_in_tmap) { panic("rack:%p rsm:%p flags:0x%x in tmap?", rack, rsm, rsm->r_flags); } #endif rsm->r_flags &= ~(RACK_ACKED|RACK_SACK_PASSED|RACK_WAS_SACKPASS); /* Rebuild it into our tmap */ if (tmap == NULL) { TAILQ_INSERT_HEAD(&rack->r_ctl.rc_tmap, rsm, r_tnext); tmap = rsm; } else { TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, tmap, rsm, r_tnext); tmap = rsm; } tmap->r_in_tmap = 1; rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); } /* * Now lets possibly clear the sack filter so we start * recognizing sacks that cover this area. */ sack_filter_clear(&rack->r_ctl.rack_sf, th_ack); } static void rack_do_decay(struct tcp_rack *rack) { struct timeval res; #define timersub(tvp, uvp, vvp) \ do { \ (vvp)->tv_sec = (tvp)->tv_sec - (uvp)->tv_sec; \ (vvp)->tv_usec = (tvp)->tv_usec - (uvp)->tv_usec; \ if ((vvp)->tv_usec < 0) { \ (vvp)->tv_sec--; \ (vvp)->tv_usec += 1000000; \ } \ } while (0) timersub(&rack->r_ctl.act_rcv_time, &rack->r_ctl.rc_last_time_decay, &res); #undef timersub rack->r_ctl.input_pkt++; if ((rack->rc_in_persist) || (res.tv_sec >= 1) || (rack->rc_tp->snd_max == rack->rc_tp->snd_una)) { /* * Check for decay of non-SAD, * we want all SAD detection metrics to * decay 1/4 per second (or more) passed. */ #ifdef NETFLIX_EXP_DETECTION uint32_t pkt_delta; pkt_delta = rack->r_ctl.input_pkt - rack->r_ctl.saved_input_pkt; #endif /* Update our saved tracking values */ rack->r_ctl.saved_input_pkt = rack->r_ctl.input_pkt; rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time; /* Now do we escape without decay? */ #ifdef NETFLIX_EXP_DETECTION if (rack->rc_in_persist || (rack->rc_tp->snd_max == rack->rc_tp->snd_una) || (pkt_delta < tcp_sad_low_pps)){ /* * We don't decay idle connections * or ones that have a low input pps. */ return; } /* Decay the counters */ rack->r_ctl.ack_count = ctf_decay_count(rack->r_ctl.ack_count, tcp_sad_decay_val); rack->r_ctl.sack_count = ctf_decay_count(rack->r_ctl.sack_count, tcp_sad_decay_val); rack->r_ctl.sack_moved_extra = ctf_decay_count(rack->r_ctl.sack_moved_extra, tcp_sad_decay_val); rack->r_ctl.sack_noextra_move = ctf_decay_count(rack->r_ctl.sack_noextra_move, tcp_sad_decay_val); #endif } } static void rack_process_to_cumack(struct tcpcb *tp, struct tcp_rack *rack, register uint32_t th_ack, uint32_t cts, struct tcpopt *to) { struct rack_sendmap *rsm; #ifdef INVARIANTS struct rack_sendmap *rm; #endif /* * The ACK point is advancing to th_ack, we must drop off * the packets in the rack log and calculate any eligble * RTT's. */ rack->r_wanted_output = 1; /* Tend any TLP that has been marked for 1/2 the seq space (its old) */ if ((rack->rc_last_tlp_acked_set == 1)&& (rack->rc_last_tlp_past_cumack == 1) && (SEQ_GT(rack->r_ctl.last_tlp_acked_start, th_ack))) { /* * We have reached the point where our last rack * tlp retransmit sequence is ahead of the cum-ack. * This can only happen when the cum-ack moves all * the way around (its been a full 2^^31+1 bytes * or more since we sent a retransmitted TLP). Lets * turn off the valid flag since its not really valid. * * Note since sack's also turn on this event we have * a complication, we have to wait to age it out until * the cum-ack is by the TLP before checking which is * what the next else clause does. */ rack_log_dsack_event(rack, 9, __LINE__, rack->r_ctl.last_tlp_acked_start, rack->r_ctl.last_tlp_acked_end); rack->rc_last_tlp_acked_set = 0; rack->rc_last_tlp_past_cumack = 0; } else if ((rack->rc_last_tlp_acked_set == 1) && (rack->rc_last_tlp_past_cumack == 0) && (SEQ_GEQ(th_ack, rack->r_ctl.last_tlp_acked_end))) { /* * It is safe to start aging TLP's out. */ rack->rc_last_tlp_past_cumack = 1; } /* We do the same for the tlp send seq as well */ if ((rack->rc_last_sent_tlp_seq_valid == 1) && (rack->rc_last_sent_tlp_past_cumack == 1) && (SEQ_GT(rack->r_ctl.last_sent_tlp_seq, th_ack))) { rack_log_dsack_event(rack, 9, __LINE__, rack->r_ctl.last_sent_tlp_seq, (rack->r_ctl.last_sent_tlp_seq + rack->r_ctl.last_sent_tlp_len)); rack->rc_last_sent_tlp_seq_valid = 0; rack->rc_last_sent_tlp_past_cumack = 0; } else if ((rack->rc_last_sent_tlp_seq_valid == 1) && (rack->rc_last_sent_tlp_past_cumack == 0) && (SEQ_GEQ(th_ack, rack->r_ctl.last_sent_tlp_seq))) { /* * It is safe to start aging TLP's send. */ rack->rc_last_sent_tlp_past_cumack = 1; } more: rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree); if (rsm == NULL) { if ((th_ack - 1) == tp->iss) { /* * For the SYN incoming case we will not * have called tcp_output for the sending of * the SYN, so there will be no map. All * other cases should probably be a panic. */ return; } if (tp->t_flags & TF_SENTFIN) { /* if we sent a FIN we often will not have map */ return; } #ifdef INVARIANTS panic("No rack map tp:%p for state:%d ack:%u rack:%p snd_una:%u snd_max:%u snd_nxt:%u\n", tp, tp->t_state, th_ack, rack, tp->snd_una, tp->snd_max, tp->snd_nxt); #endif return; } if (SEQ_LT(th_ack, rsm->r_start)) { /* Huh map is missing this */ #ifdef INVARIANTS printf("Rack map starts at r_start:%u for th_ack:%u huh? ts:%d rs:%d\n", rsm->r_start, th_ack, tp->t_state, rack->r_state); #endif return; } rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED, th_ack); /* Now was it a retransmitted TLP? */ if ((rsm->r_flags & RACK_TLP) && (rsm->r_rtr_cnt > 1)) { /* * Yes, this rsm was a TLP and retransmitted, remember that * since if a DSACK comes back on this we don't want * to think of it as a reordered segment. This may * get updated again with possibly even other TLPs * in flight, but thats ok. Only when we don't send * a retransmitted TLP for 1/2 the sequences space * will it get turned off (above). */ if (rack->rc_last_tlp_acked_set && (is_rsm_inside_declared_tlp_block(rack, rsm))) { /* * We already turned this on since the end matches, * the previous one was a partially ack now we * are getting another one (maybe all of it). */ rack_log_dsack_event(rack, 10, __LINE__, rsm->r_start, rsm->r_end); /* * Lets make sure we have all of it though. */ if (SEQ_LT(rsm->r_start, rack->r_ctl.last_tlp_acked_start)) { rack->r_ctl.last_tlp_acked_start = rsm->r_start; rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start, rack->r_ctl.last_tlp_acked_end); } if (SEQ_GT(rsm->r_end, rack->r_ctl.last_tlp_acked_end)) { rack->r_ctl.last_tlp_acked_end = rsm->r_end; rack_log_dsack_event(rack, 11, __LINE__, rack->r_ctl.last_tlp_acked_start, rack->r_ctl.last_tlp_acked_end); } } else { rack->rc_last_tlp_past_cumack = 1; rack->r_ctl.last_tlp_acked_start = rsm->r_start; rack->r_ctl.last_tlp_acked_end = rsm->r_end; rack->rc_last_tlp_acked_set = 1; rack_log_dsack_event(rack, 8, __LINE__, rsm->r_start, rsm->r_end); } } /* Now do we consume the whole thing? */ if (SEQ_GEQ(th_ack, rsm->r_end)) { /* Its all consumed. */ uint32_t left; uint8_t newly_acked; rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_FREE, rsm->r_end, __LINE__); rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; rsm->r_rtr_bytes = 0; /* Record the time of highest cumack sent */ rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]; #ifndef INVARIANTS (void)RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); #else rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); if (rm != rsm) { panic("removing head in rack:%p rsm:%p rm:%p", rack, rsm, rm); } #endif if (rsm->r_in_tmap) { TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 0; } newly_acked = 1; if (rsm->r_flags & RACK_ACKED) { /* * It was acked on the scoreboard -- remove * it from total */ rack->r_ctl.rc_sacked -= (rsm->r_end - rsm->r_start); newly_acked = 0; } else if (rsm->r_flags & RACK_SACK_PASSED) { /* * There are segments ACKED on the * scoreboard further up. We are seeing * reordering. */ rsm->r_flags &= ~RACK_SACK_PASSED; rsm->r_ack_arrival = rack_to_usec_ts(&rack->r_ctl.act_rcv_time); rsm->r_flags |= RACK_ACKED; rack->r_ctl.rc_reorder_ts = cts; if (rack->r_ent_rec_ns) { /* * We have sent no more, and we saw an sack * then ack arrive. */ rack->r_might_revert = 1; } } if ((rsm->r_flags & RACK_TO_REXT) && (tp->t_flags & TF_RCVD_TSTMP) && (to->to_flags & TOF_TS) && (to->to_tsecr != 0) && (tp->t_flags & TF_PREVVALID)) { /* * We can use the timestamp to see * if this retransmission was from the * first transmit. If so we made a mistake. */ tp->t_flags &= ~TF_PREVVALID; if (to->to_tsecr == rack_ts_to_msec(rsm->r_tim_lastsent[0])) { /* The first transmit is what this ack is for */ rack_cong_signal(tp, CC_RTO_ERR, th_ack, __LINE__); } } left = th_ack - rsm->r_end; if (rack->app_limited_needs_set && newly_acked) rack_need_set_test(tp, rack, rsm, th_ack, __LINE__, RACK_USE_END_OR_THACK); /* Free back to zone */ rack_free(rack, rsm); if (left) { goto more; } /* Check for reneging */ rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree); if (rsm && (rsm->r_flags & RACK_ACKED) && (th_ack == rsm->r_start)) { /* * The peer has moved snd_una up to * the edge of this send, i.e. one * that it had previously acked. The only * way that can be true if the peer threw * away data (space issues) that it had * previously sacked (else it would have * given us snd_una up to (rsm->r_end). * We need to undo the acked markings here. * * Note we have to look to make sure th_ack is * our rsm->r_start in case we get an old ack * where th_ack is behind snd_una. */ rack_peer_reneges(rack, rsm, th_ack); } return; } if (rsm->r_flags & RACK_ACKED) { /* * It was acked on the scoreboard -- remove it from * total for the part being cum-acked. */ rack->r_ctl.rc_sacked -= (th_ack - rsm->r_start); } /* * Clear the dup ack count for * the piece that remains. */ rsm->r_dupack = 0; rack_log_retran_reason(rack, rsm, __LINE__, 0, 2); if (rsm->r_rtr_bytes) { /* * It was retransmitted adjust the * sack holes for what was acked. */ int ack_am; ack_am = (th_ack - rsm->r_start); if (ack_am >= rsm->r_rtr_bytes) { rack->r_ctl.rc_holes_rxt -= ack_am; rsm->r_rtr_bytes -= ack_am; } } /* * Update where the piece starts and record * the time of send of highest cumack sent. */ rack->r_ctl.rc_gp_cumack_ts = rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)]; rack_log_map_chg(tp, rack, NULL, rsm, NULL, MAP_TRIM_HEAD, th_ack, __LINE__); /* Now we need to move our offset forward too */ if (rsm->m && (rsm->orig_m_len != rsm->m->m_len)) { /* Fix up the orig_m_len and possibly the mbuf offset */ rack_adjust_orig_mlen(rsm); } rsm->soff += (th_ack - rsm->r_start); rsm->r_start = th_ack; /* Now do we need to move the mbuf fwd too? */ if (rsm->m) { while (rsm->soff >= rsm->m->m_len) { rsm->soff -= rsm->m->m_len; rsm->m = rsm->m->m_next; KASSERT((rsm->m != NULL), (" nrsm:%p hit at soff:%u null m", rsm, rsm->soff)); } rsm->orig_m_len = rsm->m->m_len; } if (rack->app_limited_needs_set) rack_need_set_test(tp, rack, rsm, tp->snd_una, __LINE__, RACK_USE_BEG); } static void rack_handle_might_revert(struct tcpcb *tp, struct tcp_rack *rack) { struct rack_sendmap *rsm; int sack_pass_fnd = 0; if (rack->r_might_revert) { /* * Ok we have reordering, have not sent anything, we * might want to revert the congestion state if nothing * further has SACK_PASSED on it. Lets check. * * We also get here when we have DSACKs come in for * all the data that we FR'd. Note that a rxt or tlp * timer clears this from happening. */ TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) { if (rsm->r_flags & RACK_SACK_PASSED) { sack_pass_fnd = 1; break; } } if (sack_pass_fnd == 0) { /* * We went into recovery * incorrectly due to reordering! */ int orig_cwnd; rack->r_ent_rec_ns = 0; orig_cwnd = tp->snd_cwnd; tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at_erec; tp->snd_recover = tp->snd_una; rack_log_to_prr(rack, 14, orig_cwnd, __LINE__); EXIT_RECOVERY(tp->t_flags); } rack->r_might_revert = 0; } } #ifdef NETFLIX_EXP_DETECTION static void rack_do_detection(struct tcpcb *tp, struct tcp_rack *rack, uint32_t bytes_this_ack, uint32_t segsiz) { if ((rack->do_detection || tcp_force_detection) && tcp_sack_to_ack_thresh && tcp_sack_to_move_thresh && ((rack->r_ctl.rc_num_maps_alloced > tcp_map_minimum) || rack->sack_attack_disable)) { /* * We have thresholds set to find * possible attackers and disable sack. * Check them. */ uint64_t ackratio, moveratio, movetotal; /* Log detecting */ rack_log_sad(rack, 1); ackratio = (uint64_t)(rack->r_ctl.sack_count); ackratio *= (uint64_t)(1000); if (rack->r_ctl.ack_count) ackratio /= (uint64_t)(rack->r_ctl.ack_count); else { /* We really should not hit here */ ackratio = 1000; } if ((rack->sack_attack_disable == 0) && (ackratio > rack_highest_sack_thresh_seen)) rack_highest_sack_thresh_seen = (uint32_t)ackratio; movetotal = rack->r_ctl.sack_moved_extra; movetotal += rack->r_ctl.sack_noextra_move; moveratio = rack->r_ctl.sack_moved_extra; moveratio *= (uint64_t)1000; if (movetotal) moveratio /= movetotal; else { /* No moves, thats pretty good */ moveratio = 0; } if ((rack->sack_attack_disable == 0) && (moveratio > rack_highest_move_thresh_seen)) rack_highest_move_thresh_seen = (uint32_t)moveratio; if (rack->sack_attack_disable == 0) { if ((ackratio > tcp_sack_to_ack_thresh) && (moveratio > tcp_sack_to_move_thresh)) { /* Disable sack processing */ rack->sack_attack_disable = 1; if (rack->r_rep_attack == 0) { rack->r_rep_attack = 1; counter_u64_add(rack_sack_attacks_detected, 1); } if (tcp_attack_on_turns_on_logging) { /* * Turn on logging, used for debugging * false positives. */ rack->rc_tp->t_logstate = tcp_attack_on_turns_on_logging; } /* Clamp the cwnd at flight size */ rack->r_ctl.rc_saved_cwnd = rack->rc_tp->snd_cwnd; rack->rc_tp->snd_cwnd = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); rack_log_sad(rack, 2); } } else { /* We are sack-disabled check for false positives */ if ((ackratio <= tcp_restoral_thresh) || (rack->r_ctl.rc_num_maps_alloced < tcp_map_minimum)) { rack->sack_attack_disable = 0; rack_log_sad(rack, 3); /* Restart counting */ rack->r_ctl.sack_count = 0; rack->r_ctl.sack_moved_extra = 0; rack->r_ctl.sack_noextra_move = 1; rack->r_ctl.ack_count = max(1, (bytes_this_ack / segsiz)); if (rack->r_rep_reverse == 0) { rack->r_rep_reverse = 1; counter_u64_add(rack_sack_attacks_reversed, 1); } /* Restore the cwnd */ if (rack->r_ctl.rc_saved_cwnd > rack->rc_tp->snd_cwnd) rack->rc_tp->snd_cwnd = rack->r_ctl.rc_saved_cwnd; } } } } #endif static int rack_note_dsack(struct tcp_rack *rack, tcp_seq start, tcp_seq end) { uint32_t am, l_end; int was_tlp = 0; if (SEQ_GT(end, start)) am = end - start; else am = 0; if ((rack->rc_last_tlp_acked_set ) && (SEQ_GEQ(start, rack->r_ctl.last_tlp_acked_start)) && (SEQ_LEQ(end, rack->r_ctl.last_tlp_acked_end))) { /* * The DSACK is because of a TLP which we don't * do anything with the reordering window over since * it was not reordering that caused the DSACK but * our previous retransmit TLP. */ rack_log_dsack_event(rack, 7, __LINE__, start, end); was_tlp = 1; goto skip_dsack_round; } if (rack->rc_last_sent_tlp_seq_valid) { l_end = rack->r_ctl.last_sent_tlp_seq + rack->r_ctl.last_sent_tlp_len; if (SEQ_GEQ(start, rack->r_ctl.last_sent_tlp_seq) && (SEQ_LEQ(end, l_end))) { /* * This dsack is from the last sent TLP, ignore it * for reordering purposes. */ rack_log_dsack_event(rack, 7, __LINE__, start, end); was_tlp = 1; goto skip_dsack_round; } } if (rack->rc_dsack_round_seen == 0) { rack->rc_dsack_round_seen = 1; rack->r_ctl.dsack_round_end = rack->rc_tp->snd_max; rack->r_ctl.num_dsack++; rack->r_ctl.dsack_persist = 16; /* 16 is from the standard */ rack_log_dsack_event(rack, 2, __LINE__, 0, 0); } skip_dsack_round: /* * We keep track of how many DSACK blocks we get * after a recovery incident. */ rack->r_ctl.dsack_byte_cnt += am; if (!IN_FASTRECOVERY(rack->rc_tp->t_flags) && rack->r_ctl.retran_during_recovery && (rack->r_ctl.dsack_byte_cnt >= rack->r_ctl.retran_during_recovery)) { /* * False recovery most likely culprit is reordering. If * nothing else is missing we need to revert. */ rack->r_might_revert = 1; rack_handle_might_revert(rack->rc_tp, rack); rack->r_might_revert = 0; rack->r_ctl.retran_during_recovery = 0; rack->r_ctl.dsack_byte_cnt = 0; } return (was_tlp); } static uint32_t do_rack_compute_pipe(struct tcpcb *tp, struct tcp_rack *rack, uint32_t snd_una) { return (((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt); } static int32_t rack_compute_pipe(struct tcpcb *tp) { return ((int32_t)do_rack_compute_pipe(tp, (struct tcp_rack *)tp->t_fb_ptr, tp->snd_una)); } static void rack_update_prr(struct tcpcb *tp, struct tcp_rack *rack, uint32_t changed, tcp_seq th_ack) { /* Deal with changed and PRR here (in recovery only) */ uint32_t pipe, snd_una; rack->r_ctl.rc_prr_delivered += changed; if (sbavail(&rack->rc_inp->inp_socket->so_snd) <= (tp->snd_max - tp->snd_una)) { /* * It is all outstanding, we are application limited * and thus we don't need more room to send anything. * Note we use tp->snd_una here and not th_ack because * the data as yet not been cut from the sb. */ rack->r_ctl.rc_prr_sndcnt = 0; return; } /* Compute prr_sndcnt */ if (SEQ_GT(tp->snd_una, th_ack)) { snd_una = tp->snd_una; } else { snd_una = th_ack; } pipe = do_rack_compute_pipe(tp, rack, snd_una); if (pipe > tp->snd_ssthresh) { long sndcnt; sndcnt = rack->r_ctl.rc_prr_delivered * tp->snd_ssthresh; if (rack->r_ctl.rc_prr_recovery_fs > 0) sndcnt /= (long)rack->r_ctl.rc_prr_recovery_fs; else { rack->r_ctl.rc_prr_sndcnt = 0; rack_log_to_prr(rack, 9, 0, __LINE__); sndcnt = 0; } sndcnt++; if (sndcnt > (long)rack->r_ctl.rc_prr_out) sndcnt -= rack->r_ctl.rc_prr_out; else sndcnt = 0; rack->r_ctl.rc_prr_sndcnt = sndcnt; rack_log_to_prr(rack, 10, 0, __LINE__); } else { uint32_t limit; if (rack->r_ctl.rc_prr_delivered > rack->r_ctl.rc_prr_out) limit = (rack->r_ctl.rc_prr_delivered - rack->r_ctl.rc_prr_out); else limit = 0; if (changed > limit) limit = changed; limit += ctf_fixed_maxseg(tp); if (tp->snd_ssthresh > pipe) { rack->r_ctl.rc_prr_sndcnt = min((tp->snd_ssthresh - pipe), limit); rack_log_to_prr(rack, 11, 0, __LINE__); } else { rack->r_ctl.rc_prr_sndcnt = min(0, limit); rack_log_to_prr(rack, 12, 0, __LINE__); } } } static void rack_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th, int entered_recovery, int dup_ack_struck) { uint32_t changed; struct tcp_rack *rack; struct rack_sendmap *rsm; struct sackblk sack, sack_blocks[TCP_MAX_SACK + 1]; register uint32_t th_ack; int32_t i, j, k, num_sack_blks = 0; uint32_t cts, acked, ack_point; int loop_start = 0, moved_two = 0; uint32_t tsused; INP_WLOCK_ASSERT(tp->t_inpcb); if (tcp_get_flags(th) & TH_RST) { /* We don't log resets */ return; } rack = (struct tcp_rack *)tp->t_fb_ptr; cts = tcp_get_usecs(NULL); rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree); changed = 0; th_ack = th->th_ack; if (rack->sack_attack_disable == 0) rack_do_decay(rack); if (BYTES_THIS_ACK(tp, th) >= ctf_fixed_maxseg(rack->rc_tp)) { /* * You only get credit for * MSS and greater (and you get extra * credit for larger cum-ack moves). */ int ac; ac = BYTES_THIS_ACK(tp, th) / ctf_fixed_maxseg(rack->rc_tp); rack->r_ctl.ack_count += ac; counter_u64_add(rack_ack_total, ac); } if (rack->r_ctl.ack_count > 0xfff00000) { /* * reduce the number to keep us under * a uint32_t. */ rack->r_ctl.ack_count /= 2; rack->r_ctl.sack_count /= 2; } if (SEQ_GT(th_ack, tp->snd_una)) { rack_log_progress_event(rack, tp, ticks, PROGRESS_UPDATE, __LINE__); tp->t_acktime = ticks; } if (rsm && SEQ_GT(th_ack, rsm->r_start)) changed = th_ack - rsm->r_start; if (changed) { rack_process_to_cumack(tp, rack, th_ack, cts, to); } if ((to->to_flags & TOF_SACK) == 0) { /* We are done nothing left and no sack. */ rack_handle_might_revert(tp, rack); /* * For cases where we struck a dup-ack * with no SACK, add to the changes so * PRR will work right. */ if (dup_ack_struck && (changed == 0)) { changed += ctf_fixed_maxseg(rack->rc_tp); } goto out; } /* Sack block processing */ if (SEQ_GT(th_ack, tp->snd_una)) ack_point = th_ack; else ack_point = tp->snd_una; for (i = 0; i < to->to_nsacks; i++) { bcopy((to->to_sacks + i * TCPOLEN_SACK), &sack, sizeof(sack)); sack.start = ntohl(sack.start); sack.end = ntohl(sack.end); if (SEQ_GT(sack.end, sack.start) && SEQ_GT(sack.start, ack_point) && SEQ_LT(sack.start, tp->snd_max) && SEQ_GT(sack.end, ack_point) && SEQ_LEQ(sack.end, tp->snd_max)) { sack_blocks[num_sack_blks] = sack; num_sack_blks++; } else if (SEQ_LEQ(sack.start, th_ack) && SEQ_LEQ(sack.end, th_ack)) { int was_tlp; was_tlp = rack_note_dsack(rack, sack.start, sack.end); /* * Its a D-SACK block. */ tcp_record_dsack(tp, sack.start, sack.end, was_tlp); } } if (rack->rc_dsack_round_seen) { /* Is the dsack roound over? */ if (SEQ_GEQ(th_ack, rack->r_ctl.dsack_round_end)) { /* Yes it is */ rack->rc_dsack_round_seen = 0; rack_log_dsack_event(rack, 3, __LINE__, 0, 0); } } /* * Sort the SACK blocks so we can update the rack scoreboard with * just one pass. */ num_sack_blks = sack_filter_blks(&rack->r_ctl.rack_sf, sack_blocks, num_sack_blks, th->th_ack); ctf_log_sack_filter(rack->rc_tp, num_sack_blks, sack_blocks); if (num_sack_blks == 0) { /* Nothing to sack (DSACKs?) */ goto out_with_totals; } if (num_sack_blks < 2) { /* Only one, we don't need to sort */ goto do_sack_work; } /* Sort the sacks */ for (i = 0; i < num_sack_blks; i++) { for (j = i + 1; j < num_sack_blks; j++) { if (SEQ_GT(sack_blocks[i].end, sack_blocks[j].end)) { sack = sack_blocks[i]; sack_blocks[i] = sack_blocks[j]; sack_blocks[j] = sack; } } } /* * Now are any of the sack block ends the same (yes some * implementations send these)? */ again: if (num_sack_blks == 0) goto out_with_totals; if (num_sack_blks > 1) { for (i = 0; i < num_sack_blks; i++) { for (j = i + 1; j < num_sack_blks; j++) { if (sack_blocks[i].end == sack_blocks[j].end) { /* * Ok these two have the same end we * want the smallest end and then * throw away the larger and start * again. */ if (SEQ_LT(sack_blocks[j].start, sack_blocks[i].start)) { /* * The second block covers * more area use that */ sack_blocks[i].start = sack_blocks[j].start; } /* * Now collapse out the dup-sack and * lower the count */ for (k = (j + 1); k < num_sack_blks; k++) { sack_blocks[j].start = sack_blocks[k].start; sack_blocks[j].end = sack_blocks[k].end; j++; } num_sack_blks--; goto again; } } } } do_sack_work: /* * First lets look to see if * we have retransmitted and * can use the transmit next? */ rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); if (rsm && SEQ_GT(sack_blocks[0].end, rsm->r_start) && SEQ_LT(sack_blocks[0].start, rsm->r_end)) { /* * We probably did the FR and the next * SACK in continues as we would expect. */ acked = rack_proc_sack_blk(tp, rack, &sack_blocks[0], to, &rsm, cts, &moved_two); if (acked) { rack->r_wanted_output = 1; changed += acked; } if (num_sack_blks == 1) { /* * This is what we would expect from * a normal implementation to happen * after we have retransmitted the FR, * i.e the sack-filter pushes down * to 1 block and the next to be retransmitted * is the sequence in the sack block (has more * are acked). Count this as ACK'd data to boost * up the chances of recovering any false positives. */ rack->r_ctl.ack_count += (acked / ctf_fixed_maxseg(rack->rc_tp)); counter_u64_add(rack_ack_total, (acked / ctf_fixed_maxseg(rack->rc_tp))); counter_u64_add(rack_express_sack, 1); if (rack->r_ctl.ack_count > 0xfff00000) { /* * reduce the number to keep us under * a uint32_t. */ rack->r_ctl.ack_count /= 2; rack->r_ctl.sack_count /= 2; } goto out_with_totals; } else { /* * Start the loop through the * rest of blocks, past the first block. */ moved_two = 0; loop_start = 1; } } /* Its a sack of some sort */ rack->r_ctl.sack_count++; if (rack->r_ctl.sack_count > 0xfff00000) { /* * reduce the number to keep us under * a uint32_t. */ rack->r_ctl.ack_count /= 2; rack->r_ctl.sack_count /= 2; } counter_u64_add(rack_sack_total, 1); if (rack->sack_attack_disable) { /* An attacker disablement is in place */ if (num_sack_blks > 1) { rack->r_ctl.sack_count += (num_sack_blks - 1); rack->r_ctl.sack_moved_extra++; counter_u64_add(rack_move_some, 1); if (rack->r_ctl.sack_moved_extra > 0xfff00000) { rack->r_ctl.sack_moved_extra /= 2; rack->r_ctl.sack_noextra_move /= 2; } } goto out; } rsm = rack->r_ctl.rc_sacklast; for (i = loop_start; i < num_sack_blks; i++) { acked = rack_proc_sack_blk(tp, rack, &sack_blocks[i], to, &rsm, cts, &moved_two); if (acked) { rack->r_wanted_output = 1; changed += acked; } if (moved_two) { /* * If we did not get a SACK for at least a MSS and * had to move at all, or if we moved more than our * threshold, it counts against the "extra" move. */ rack->r_ctl.sack_moved_extra += moved_two; counter_u64_add(rack_move_some, 1); } else { /* * else we did not have to move * any more than we would expect. */ rack->r_ctl.sack_noextra_move++; counter_u64_add(rack_move_none, 1); } if (moved_two && (acked < ctf_fixed_maxseg(rack->rc_tp))) { /* * If the SACK was not a full MSS then * we add to sack_count the number of * MSS's (or possibly more than * a MSS if its a TSO send) we had to skip by. */ rack->r_ctl.sack_count += moved_two; counter_u64_add(rack_sack_total, moved_two); } /* * Now we need to setup for the next * round. First we make sure we won't * exceed the size of our uint32_t on * the various counts, and then clear out * moved_two. */ if ((rack->r_ctl.sack_moved_extra > 0xfff00000) || (rack->r_ctl.sack_noextra_move > 0xfff00000)) { rack->r_ctl.sack_moved_extra /= 2; rack->r_ctl.sack_noextra_move /= 2; } if (rack->r_ctl.sack_count > 0xfff00000) { rack->r_ctl.ack_count /= 2; rack->r_ctl.sack_count /= 2; } moved_two = 0; } out_with_totals: if (num_sack_blks > 1) { /* * You get an extra stroke if * you have more than one sack-blk, this * could be where we are skipping forward * and the sack-filter is still working, or * it could be an attacker constantly * moving us. */ rack->r_ctl.sack_moved_extra++; counter_u64_add(rack_move_some, 1); } out: #ifdef NETFLIX_EXP_DETECTION rack_do_detection(tp, rack, BYTES_THIS_ACK(tp, th), ctf_fixed_maxseg(rack->rc_tp)); #endif if (changed) { /* Something changed cancel the rack timer */ rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); } tsused = tcp_get_usecs(NULL); rsm = tcp_rack_output(tp, rack, tsused); if ((!IN_FASTRECOVERY(tp->t_flags)) && rsm && ((rsm->r_flags & RACK_MUST_RXT) == 0)) { /* Enter recovery */ entered_recovery = 1; rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__); /* * When we enter recovery we need to assure we send * one packet. */ if (rack->rack_no_prr == 0) { rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp); rack_log_to_prr(rack, 8, 0, __LINE__); } rack->r_timer_override = 1; rack->r_early = 0; rack->r_ctl.rc_agg_early = 0; } else if (IN_FASTRECOVERY(tp->t_flags) && rsm && (rack->r_rr_config == 3)) { /* * Assure we can output and we get no * remembered pace time except the retransmit. */ rack->r_timer_override = 1; rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT; rack->r_ctl.rc_resend = rsm; } if (IN_FASTRECOVERY(tp->t_flags) && (rack->rack_no_prr == 0) && (entered_recovery == 0)) { rack_update_prr(tp, rack, changed, th_ack); if ((rsm && (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) && ((tcp_in_hpts(rack->rc_inp) == 0) && ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0)))) { /* * If you are pacing output you don't want * to override. */ rack->r_early = 0; rack->r_ctl.rc_agg_early = 0; rack->r_timer_override = 1; } } } static void rack_strike_dupack(struct tcp_rack *rack) { struct rack_sendmap *rsm; rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); while (rsm && (rsm->r_dupack >= DUP_ACK_THRESHOLD)) { rsm = TAILQ_NEXT(rsm, r_tnext); if (rsm->r_flags & RACK_MUST_RXT) { /* Sendmap entries that are marked to * be retransmitted do not need dupack's * struck. We get these marks for a number * of reasons (rxt timeout with no sack, * mtu change, or rwnd collapses). When * these events occur, we know we must retransmit * them and mark the sendmap entries. Dupack counting * is not needed since we are already set to retransmit * it as soon as we can. */ continue; } } if (rsm && (rsm->r_dupack < 0xff)) { rsm->r_dupack++; if (rsm->r_dupack >= DUP_ACK_THRESHOLD) { struct timeval tv; uint32_t cts; /* * Here we see if we need to retransmit. For * a SACK type connection if enough time has passed * we will get a return of the rsm. For a non-sack * connection we will get the rsm returned if the * dupack value is 3 or more. */ cts = tcp_get_usecs(&tv); rack->r_ctl.rc_resend = tcp_rack_output(rack->rc_tp, rack, cts); if (rack->r_ctl.rc_resend != NULL) { if (!IN_FASTRECOVERY(rack->rc_tp->t_flags)) { rack_cong_signal(rack->rc_tp, CC_NDUPACK, rack->rc_tp->snd_una, __LINE__); } rack->r_wanted_output = 1; rack->r_timer_override = 1; rack_log_retran_reason(rack, rsm, __LINE__, 1, 3); } } else { rack_log_retran_reason(rack, rsm, __LINE__, 0, 3); } } } static void rack_check_bottom_drag(struct tcpcb *tp, struct tcp_rack *rack, struct socket *so, int32_t acked) { uint32_t segsiz, minseg; segsiz = ctf_fixed_maxseg(tp); minseg = segsiz; if (tp->snd_max == tp->snd_una) { /* * We are doing dynamic pacing and we are way * under. Basically everything got acked while * we were still waiting on the pacer to expire. * * This means we need to boost the b/w in * addition to any earlier boosting of * the multiplier. */ rack->rc_dragged_bottom = 1; rack_validate_multipliers_at_or_above100(rack); /* * Lets use the segment bytes acked plus * the lowest RTT seen as the basis to * form a b/w estimate. This will be off * due to the fact that the true estimate * should be around 1/2 the time of the RTT * but we can settle for that. */ if ((rack->r_ctl.rack_rs.rs_flags & RACK_RTT_VALID) && acked) { uint64_t bw, calc_bw, rtt; rtt = rack->r_ctl.rack_rs.rs_us_rtt; if (rtt == 0) { /* no us sample is there a ms one? */ if (rack->r_ctl.rack_rs.rs_rtt_lowest) { rtt = rack->r_ctl.rack_rs.rs_rtt_lowest; } else { goto no_measurement; } } bw = acked; calc_bw = bw * 1000000; calc_bw /= rtt; if (rack->r_ctl.last_max_bw && (rack->r_ctl.last_max_bw < calc_bw)) { /* * If we have a last calculated max bw * enforce it. */ calc_bw = rack->r_ctl.last_max_bw; } /* now plop it in */ if (rack->rc_gp_filled == 0) { if (calc_bw > ONE_POINT_TWO_MEG) { /* * If we have no measurement * don't let us set in more than * 1.2Mbps. If we are still too * low after pacing with this we * will hopefully have a max b/w * available to sanity check things. */ calc_bw = ONE_POINT_TWO_MEG; } rack->r_ctl.rc_rtt_diff = 0; rack->r_ctl.gp_bw = calc_bw; rack->rc_gp_filled = 1; if (rack->r_ctl.num_measurements < RACK_REQ_AVG) rack->r_ctl.num_measurements = RACK_REQ_AVG; rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL); } else if (calc_bw > rack->r_ctl.gp_bw) { rack->r_ctl.rc_rtt_diff = 0; if (rack->r_ctl.num_measurements < RACK_REQ_AVG) rack->r_ctl.num_measurements = RACK_REQ_AVG; rack->r_ctl.gp_bw = calc_bw; rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL); } else rack_increase_bw_mul(rack, -1, 0, 0, 1); if ((rack->gp_ready == 0) && (rack->r_ctl.num_measurements >= rack->r_ctl.req_measurements)) { /* We have enough measurements now */ rack->gp_ready = 1; rack_set_cc_pacing(rack); if (rack->defer_options) rack_apply_deferred_options(rack); } /* * For acks over 1mss we do a extra boost to simulate * where we would get 2 acks (we want 110 for the mul). */ if (acked > segsiz) rack_increase_bw_mul(rack, -1, 0, 0, 1); } else { /* * zero rtt possibly?, settle for just an old increase. */ no_measurement: rack_increase_bw_mul(rack, -1, 0, 0, 1); } } else if ((IN_FASTRECOVERY(tp->t_flags) == 0) && (sbavail(&so->so_snd) > max((segsiz * (4 + rack_req_segs)), minseg)) && (rack->r_ctl.cwnd_to_use > max((segsiz * (rack_req_segs + 2)), minseg)) && (tp->snd_wnd > max((segsiz * (rack_req_segs + 2)), minseg)) && (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) <= (segsiz * rack_req_segs))) { /* * We are doing dynamic GP pacing and * we have everything except 1MSS or less * bytes left out. We are still pacing away. * And there is data that could be sent, This * means we are inserting delayed ack time in * our measurements because we are pacing too slow. */ rack_validate_multipliers_at_or_above100(rack); rack->rc_dragged_bottom = 1; rack_increase_bw_mul(rack, -1, 0, 0, 1); } } static void rack_gain_for_fastoutput(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t acked_amount) { /* * The fast output path is enabled and we * have moved the cumack forward. Lets see if * we can expand forward the fast path length by * that amount. What we would ideally like to * do is increase the number of bytes in the * fast path block (left_to_send) by the * acked amount. However we have to gate that * by two factors: * 1) The amount outstanding and the rwnd of the peer * (i.e. we don't want to exceed the rwnd of the peer). * * 2) The amount of data left in the socket buffer (i.e. * we can't send beyond what is in the buffer). * * Note that this does not take into account any increase * in the cwnd. We will only extend the fast path by * what was acked. */ uint32_t new_total, gating_val; new_total = acked_amount + rack->r_ctl.fsb.left_to_send; gating_val = min((sbavail(&so->so_snd) - (tp->snd_max - tp->snd_una)), (tp->snd_wnd - (tp->snd_max - tp->snd_una))); if (new_total <= gating_val) { /* We can increase left_to_send by the acked amount */ counter_u64_add(rack_extended_rfo, 1); rack->r_ctl.fsb.left_to_send = new_total; KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(&rack->rc_inp->inp_socket->so_snd) - (tp->snd_max - tp->snd_una))), ("rack:%p left_to_send:%u sbavail:%u out:%u", rack, rack->r_ctl.fsb.left_to_send, sbavail(&rack->rc_inp->inp_socket->so_snd), (tp->snd_max - tp->snd_una))); } } static void rack_adjust_sendmap(struct tcp_rack *rack, struct sockbuf *sb, tcp_seq snd_una) { /* * Here any sendmap entry that points to the * beginning mbuf must be adjusted to the correct * offset. This must be called with: * 1) The socket buffer locked * 2) snd_una adjusted to its new postion. * * Note that (2) implies rack_ack_received has also * been called. * * We grab the first mbuf in the socket buffer and * then go through the front of the sendmap, recalculating * the stored offset for any sendmap entry that has * that mbuf. We must use the sb functions to do this * since its possible an add was done has well as * the subtraction we may have just completed. This should * not be a penalty though, since we just referenced the sb * to go in and trim off the mbufs that we freed (of course * there will be a penalty for the sendmap references though). */ struct mbuf *m; struct rack_sendmap *rsm; SOCKBUF_LOCK_ASSERT(sb); m = sb->sb_mb; rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree); if ((rsm == NULL) || (m == NULL)) { /* Nothing outstanding */ return; } while (rsm->m && (rsm->m == m)) { /* one to adjust */ #ifdef INVARIANTS struct mbuf *tm; uint32_t soff; tm = sbsndmbuf(sb, (rsm->r_start - snd_una), &soff); if (rsm->orig_m_len != m->m_len) { rack_adjust_orig_mlen(rsm); } if (rsm->soff != soff) { /* * This is not a fatal error, we anticipate it * might happen (the else code), so we count it here * so that under invariant we can see that it really * does happen. */ counter_u64_add(rack_adjust_map_bw, 1); } rsm->m = tm; rsm->soff = soff; if (tm) rsm->orig_m_len = rsm->m->m_len; else rsm->orig_m_len = 0; #else rsm->m = sbsndmbuf(sb, (rsm->r_start - snd_una), &rsm->soff); if (rsm->m) rsm->orig_m_len = rsm->m->m_len; else rsm->orig_m_len = 0; #endif rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); if (rsm == NULL) break; } } /* * Return value of 1, we do not need to call rack_process_data(). * return value of 0, rack_process_data can be called. * For ret_val if its 0 the TCP is locked, if its non-zero * its unlocked and probably unsafe to touch the TCB. */ static int rack_process_ack(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, uint32_t tiwin, int32_t tlen, int32_t * ofia, int32_t thflags, int32_t *ret_val) { int32_t ourfinisacked = 0; int32_t nsegs, acked_amount; int32_t acked; struct mbuf *mfree; struct tcp_rack *rack; int32_t under_pacing = 0; int32_t recovery = 0; rack = (struct tcp_rack *)tp->t_fb_ptr; if (SEQ_GT(th->th_ack, tp->snd_max)) { __ctf_do_dropafterack(m, tp, th, thflags, tlen, ret_val, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt); rack->r_wanted_output = 1; return (1); } if (rack->gp_ready && (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { under_pacing = 1; } if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) { int in_rec, dup_ack_struck = 0; in_rec = IN_FASTRECOVERY(tp->t_flags); if (rack->rc_in_persist) { tp->t_rxtshift = 0; RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp), rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop); } if ((th->th_ack == tp->snd_una) && (tiwin == tp->snd_wnd) && ((to->to_flags & TOF_SACK) == 0)) { rack_strike_dupack(rack); dup_ack_struck = 1; } rack_log_ack(tp, to, th, ((in_rec == 0) && IN_FASTRECOVERY(tp->t_flags)), dup_ack_struck); } if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { /* * Old ack, behind (or duplicate to) the last one rcv'd * Note: We mark reordering is occuring if its * less than and we have not closed our window. */ if (SEQ_LT(th->th_ack, tp->snd_una) && (sbspace(&so->so_rcv) > ctf_fixed_maxseg(tp))) { rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time); } return (0); } /* * If we reach this point, ACK is not a duplicate, i.e., it ACKs * something we sent. */ if (tp->t_flags & TF_NEEDSYN) { /* * T/TCP: Connection was half-synchronized, and our SYN has * been ACK'd (so connection is now fully synchronized). Go * to non-starred state, increment snd_una for ACK of SYN, * and check if we can do window scaling. */ tp->t_flags &= ~TF_NEEDSYN; tp->snd_una++; /* Do window scaling? */ if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == (TF_RCVD_SCALE | TF_REQ_SCALE)) { tp->rcv_scale = tp->request_r_scale; /* Send window already scaled. */ } } nsegs = max(1, m->m_pkthdr.lro_nsegs); INP_WLOCK_ASSERT(tp->t_inpcb); acked = BYTES_THIS_ACK(tp, th); if (acked) { /* * Any time we move the cum-ack forward clear * keep-alive tied probe-not-answered. The * persists clears its own on entry. */ rack->probe_not_answered = 0; } KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs); KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked); /* * If we just performed our first retransmit, and the ACK arrives * within our recovery window, then it was a mistake to do the * retransmit in the first place. Recover our original cwnd and * ssthresh, and proceed to transmit where we left off. */ if ((tp->t_flags & TF_PREVVALID) && ((tp->t_flags & TF_RCVD_TSTMP) == 0)) { tp->t_flags &= ~TF_PREVVALID; if (tp->t_rxtshift == 1 && (int)(ticks - tp->t_badrxtwin) < 0) rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__); } if (acked) { /* assure we are not backed off */ tp->t_rxtshift = 0; RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp), rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop); rack->rc_tlp_in_progress = 0; rack->r_ctl.rc_tlp_cnt_out = 0; /* * If it is the RXT timer we want to * stop it, so we can restart a TLP. */ if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); #ifdef NETFLIX_HTTP_LOGGING tcp_http_check_for_comp(rack->rc_tp, th->th_ack); #endif } /* * If we have a timestamp reply, update smoothed round trip time. If * no timestamp is present but transmit timer is running and timed * sequence number was acked, update smoothed round trip time. Since * we now have an rtt measurement, cancel the timer backoff (cf., * Phil Karn's retransmit alg.). Recompute the initial retransmit * timer. * * Some boxes send broken timestamp replies during the SYN+ACK * phase, ignore timestamps of 0 or we could calculate a huge RTT * and blow up the retransmit timer. */ /* * If all outstanding data is acked, stop retransmit timer and * remember to restart (more output or persist). If there is more * data to be acked, restart retransmit timer, using current * (possibly backed-off) value. */ if (acked == 0) { if (ofia) *ofia = ourfinisacked; return (0); } if (IN_RECOVERY(tp->t_flags)) { if (SEQ_LT(th->th_ack, tp->snd_recover) && (SEQ_LT(th->th_ack, tp->snd_max))) { tcp_rack_partialack(tp); } else { rack_post_recovery(tp, th->th_ack); recovery = 1; } } /* * Let the congestion control algorithm update congestion control * related information. This typically means increasing the * congestion window. */ rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, recovery); SOCKBUF_LOCK(&so->so_snd); acked_amount = min(acked, (int)sbavail(&so->so_snd)); tp->snd_wnd -= acked_amount; mfree = sbcut_locked(&so->so_snd, acked_amount); if ((sbused(&so->so_snd) == 0) && (acked > acked_amount) && (tp->t_state >= TCPS_FIN_WAIT_1) && (tp->t_flags & TF_SENTFIN)) { /* * We must be sure our fin * was sent and acked (we can be * in FIN_WAIT_1 without having * sent the fin). */ ourfinisacked = 1; } tp->snd_una = th->th_ack; if (acked_amount && sbavail(&so->so_snd)) rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una); rack_log_wakeup(tp,rack, &so->so_snd, acked, 2); /* NB: sowwakeup_locked() does an implicit unlock. */ sowwakeup_locked(so); m_freem(mfree); if (SEQ_GT(tp->snd_una, tp->snd_recover)) tp->snd_recover = tp->snd_una; if (SEQ_LT(tp->snd_nxt, tp->snd_una)) { tp->snd_nxt = tp->snd_una; } if (under_pacing && (rack->use_fixed_rate == 0) && (rack->in_probe_rtt == 0) && rack->rc_gp_dyn_mul && rack->rc_always_pace) { /* Check if we are dragging bottom */ rack_check_bottom_drag(tp, rack, so, acked); } if (tp->snd_una == tp->snd_max) { /* Nothing left outstanding */ tp->t_flags &= ~TF_PREVVALID; rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL); rack->r_ctl.retran_during_recovery = 0; rack->r_ctl.dsack_byte_cnt = 0; if (rack->r_ctl.rc_went_idle_time == 0) rack->r_ctl.rc_went_idle_time = 1; rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__); if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0) tp->t_acktime = 0; rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); /* Set need output so persist might get set */ rack->r_wanted_output = 1; sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una); if ((tp->t_state >= TCPS_FIN_WAIT_1) && (sbavail(&so->so_snd) == 0) && (tp->t_flags2 & TF2_DROP_AF_DATA)) { /* * The socket was gone and the * peer sent data (now or in the past), time to * reset him. */ *ret_val = 1; /* tcp_close will kill the inp pre-log the Reset */ tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); tp = tcp_close(tp); ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen); return (1); } } if (ofia) *ofia = ourfinisacked; return (0); } static void rack_log_collapse(struct tcp_rack *rack, uint32_t cnt, uint32_t split, uint32_t out, int line, int dir, uint32_t flags, struct rack_sendmap *rsm) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log, 0, sizeof(log)); log.u_bbr.flex1 = cnt; log.u_bbr.flex2 = split; log.u_bbr.flex3 = out; log.u_bbr.flex4 = line; log.u_bbr.flex5 = rack->r_must_retran; log.u_bbr.flex6 = flags; log.u_bbr.flex7 = rack->rc_has_collapsed; log.u_bbr.flex8 = dir; /* * 1 is collapsed, 0 is uncollapsed, * 2 is log of a rsm being marked, 3 is a split. */ if (rsm == NULL) log.u_bbr.rttProp = 0; else log.u_bbr.rttProp = (uint64_t)rsm; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, TCP_RACK_LOG_COLLAPSE, 0, 0, &log, false, &tv); } } static void rack_collapsed_window(struct tcp_rack *rack, uint32_t out, int line) { /* * Here all we do is mark the collapsed point and set the flag. * This may happen again and again, but there is no * sense splitting our map until we know where the * peer finally lands in the collapse. */ rack_trace_point(rack, RACK_TP_COLLAPSED_WND); if ((rack->rc_has_collapsed == 0) || (rack->r_ctl.last_collapse_point != (rack->rc_tp->snd_una + rack->rc_tp->snd_wnd))) counter_u64_add(rack_collapsed_win_seen, 1); rack->r_ctl.last_collapse_point = rack->rc_tp->snd_una + rack->rc_tp->snd_wnd; rack->r_ctl.high_collapse_point = rack->rc_tp->snd_max; rack->rc_has_collapsed = 1; rack->r_collapse_point_valid = 1; rack_log_collapse(rack, 0, 0, rack->r_ctl.last_collapse_point, line, 1, 0, NULL); } static void rack_un_collapse_window(struct tcp_rack *rack, int line) { struct rack_sendmap *nrsm, *rsm, fe; int cnt = 0, split = 0; #ifdef INVARIANTS struct rack_sendmap *insret; #endif memset(&fe, 0, sizeof(fe)); rack->rc_has_collapsed = 0; fe.r_start = rack->r_ctl.last_collapse_point; rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe); if (rsm == NULL) { /* Nothing to do maybe the peer ack'ed it all */ rack_log_collapse(rack, 0, 0, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL); return; } /* Now do we need to split this one? */ if (SEQ_GT(rack->r_ctl.last_collapse_point, rsm->r_start)) { rack_log_collapse(rack, rsm->r_start, rsm->r_end, rack->r_ctl.last_collapse_point, line, 3, rsm->r_flags, rsm); nrsm = rack_alloc_limit(rack, RACK_LIMIT_TYPE_SPLIT); if (nrsm == NULL) { /* We can't get a rsm, mark all? */ nrsm = rsm; goto no_split; } /* Clone it */ split = 1; rack_clone_rsm(rack, nrsm, rsm, rack->r_ctl.last_collapse_point); #ifndef INVARIANTS (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); #else insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); if (insret != NULL) { panic("Insert in rb tree of %p fails ret:%p rack:%p rsm:%p", nrsm, insret, rack, rsm); } #endif rack_log_map_chg(rack->rc_tp, rack, NULL, rsm, nrsm, MAP_SPLIT, rack->r_ctl.last_collapse_point, __LINE__); if (rsm->r_in_tmap) { TAILQ_INSERT_AFTER(&rack->r_ctl.rc_tmap, rsm, nrsm, r_tnext); nrsm->r_in_tmap = 1; } /* * Set in the new RSM as the * collapsed starting point */ rsm = nrsm; } no_split: RB_FOREACH_FROM(nrsm, rack_rb_tree_head, rsm) { nrsm->r_flags |= RACK_RWND_COLLAPSED; rack_log_collapse(rack, nrsm->r_start, nrsm->r_end, 0, line, 4, nrsm->r_flags, nrsm); cnt++; } if (cnt) { counter_u64_add(rack_collapsed_win, 1); } rack_log_collapse(rack, cnt, split, ctf_outstanding(rack->rc_tp), line, 0, 0, NULL); } static void rack_handle_delayed_ack(struct tcpcb *tp, struct tcp_rack *rack, int32_t tlen, int32_t tfo_syn) { if (DELAY_ACK(tp, tlen) || tfo_syn) { if (rack->rc_dack_mode && (tlen > 500) && (rack->rc_dack_toggle == 1)) { goto no_delayed_ack; } rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); tp->t_flags |= TF_DELACK; } else { no_delayed_ack: rack->r_wanted_output = 1; tp->t_flags |= TF_ACKNOW; if (rack->rc_dack_mode) { if (tp->t_flags & TF_DELACK) rack->rc_dack_toggle = 1; else rack->rc_dack_toggle = 0; } } } static void rack_validate_fo_sendwin_up(struct tcpcb *tp, struct tcp_rack *rack) { /* * If fast output is in progress, lets validate that * the new window did not shrink on us and make it * so fast output should end. */ if (rack->r_fast_output) { uint32_t out; /* * Calculate what we will send if left as is * and compare that to our send window. */ out = ctf_outstanding(tp); if ((out + rack->r_ctl.fsb.left_to_send) > tp->snd_wnd) { /* ok we have an issue */ if (out >= tp->snd_wnd) { /* Turn off fast output the window is met or collapsed */ rack->r_fast_output = 0; } else { /* we have some room left */ rack->r_ctl.fsb.left_to_send = tp->snd_wnd - out; if (rack->r_ctl.fsb.left_to_send < ctf_fixed_maxseg(tp)) { /* If not at least 1 full segment never mind */ rack->r_fast_output = 0; } } } } } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCP is still * locked. */ static int rack_process_data(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt) { /* * Update window information. Don't look at window if no ACK: TAC's * send garbage on first SYN. */ int32_t nsegs; int32_t tfo_syn; struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; INP_WLOCK_ASSERT(tp->t_inpcb); nsegs = max(1, m->m_pkthdr.lro_nsegs); if ((thflags & TH_ACK) && (SEQ_LT(tp->snd_wl1, th->th_seq) || (tp->snd_wl1 == th->th_seq && (SEQ_LT(tp->snd_wl2, th->th_ack) || (tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd))))) { /* keep track of pure window updates */ if (tlen == 0 && tp->snd_wl2 == th->th_ack && tiwin > tp->snd_wnd) KMOD_TCPSTAT_INC(tcps_rcvwinupd); tp->snd_wnd = tiwin; rack_validate_fo_sendwin_up(tp, rack); tp->snd_wl1 = th->th_seq; tp->snd_wl2 = th->th_ack; if (tp->snd_wnd > tp->max_sndwnd) tp->max_sndwnd = tp->snd_wnd; rack->r_wanted_output = 1; } else if (thflags & TH_ACK) { if ((tp->snd_wl2 == th->th_ack) && (tiwin < tp->snd_wnd)) { tp->snd_wnd = tiwin; rack_validate_fo_sendwin_up(tp, rack); tp->snd_wl1 = th->th_seq; tp->snd_wl2 = th->th_ack; } } if (tp->snd_wnd < ctf_outstanding(tp)) /* The peer collapsed the window */ rack_collapsed_window(rack, ctf_outstanding(tp), __LINE__); else if (rack->rc_has_collapsed) rack_un_collapse_window(rack, __LINE__); if ((rack->r_collapse_point_valid) && (SEQ_GT(th->th_ack, rack->r_ctl.high_collapse_point))) rack->r_collapse_point_valid = 0; /* Was persist timer active and now we have window space? */ if ((rack->rc_in_persist != 0) && (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs))) { rack_exit_persist(tp, rack, rack->r_ctl.rc_rcvtime); tp->snd_nxt = tp->snd_max; /* Make sure we output to start the timer */ rack->r_wanted_output = 1; } /* Do we enter persists? */ if ((rack->rc_in_persist == 0) && (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) && TCPS_HAVEESTABLISHED(tp->t_state) && ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) && sbavail(&tp->t_inpcb->inp_socket->so_snd) && (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) { /* * Here the rwnd is less than * the pacing size, we are established, * nothing is outstanding, and there is * data to send. Enter persists. */ rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime); } if (tp->t_flags2 & TF2_DROP_AF_DATA) { m_freem(m); return (0); } /* * don't process the URG bit, ignore them drag * along the up. */ tp->rcv_up = tp->rcv_nxt; INP_WLOCK_ASSERT(tp->t_inpcb); /* * Process the segment text, merging it into the TCP sequencing * queue, and arranging for acknowledgment of receipt if necessary. * This process logically involves adjusting tp->rcv_wnd as data is * presented to the user (this happens in tcp_usrreq.c, case * PRU_RCVD). If a FIN has already been received on this connection * then we just ignore the text. */ tfo_syn = ((tp->t_state == TCPS_SYN_RECEIVED) && IS_FASTOPEN(tp->t_flags)); if ((tlen || (thflags & TH_FIN) || (tfo_syn && tlen > 0)) && TCPS_HAVERCVDFIN(tp->t_state) == 0) { tcp_seq save_start = th->th_seq; tcp_seq save_rnxt = tp->rcv_nxt; int save_tlen = tlen; m_adj(m, drop_hdrlen); /* delayed header drop */ /* * Insert segment which includes th into TCP reassembly * queue with control block tp. Set thflags to whether * reassembly now includes a segment with FIN. This handles * the common case inline (segment is the next to be * received on an established connection, and the queue is * empty), avoiding linkage into and removal from the queue * and repetition of various conversions. Set DELACK for * segments received in order, but ack immediately when * segments are out of order (so fast retransmit can work). */ if (th->th_seq == tp->rcv_nxt && SEGQ_EMPTY(tp) && (TCPS_HAVEESTABLISHED(tp->t_state) || tfo_syn)) { #ifdef NETFLIX_SB_LIMITS u_int mcnt, appended; if (so->so_rcv.sb_shlim) { mcnt = m_memcnt(m); appended = 0; if (counter_fo_get(so->so_rcv.sb_shlim, mcnt, CFO_NOSLEEP, NULL) == false) { counter_u64_add(tcp_sb_shlim_fails, 1); m_freem(m); return (0); } } #endif rack_handle_delayed_ack(tp, rack, tlen, tfo_syn); tp->rcv_nxt += tlen; if (tlen && ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) && (tp->t_fbyte_in == 0)) { tp->t_fbyte_in = ticks; if (tp->t_fbyte_in == 0) tp->t_fbyte_in = 1; if (tp->t_fbyte_out && tp->t_fbyte_in) tp->t_flags2 |= TF2_FBYTES_COMPLETE; } thflags = tcp_get_flags(th) & TH_FIN; KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs); KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen); SOCKBUF_LOCK(&so->so_rcv); if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { m_freem(m); } else #ifdef NETFLIX_SB_LIMITS appended = #endif sbappendstream_locked(&so->so_rcv, m, 0); rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1); /* NB: sorwakeup_locked() does an implicit unlock. */ sorwakeup_locked(so); #ifdef NETFLIX_SB_LIMITS if (so->so_rcv.sb_shlim && appended != mcnt) counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended); #endif } else { /* * XXX: Due to the header drop above "th" is * theoretically invalid by now. Fortunately * m_adj() doesn't actually frees any mbufs when * trimming from the head. */ tcp_seq temp = save_start; thflags = tcp_reass(tp, th, &temp, &tlen, m); tp->t_flags |= TF_ACKNOW; if (tp->t_flags & TF_WAKESOR) { tp->t_flags &= ~TF_WAKESOR; /* NB: sorwakeup_locked() does an implicit unlock. */ sorwakeup_locked(so); } } if ((tp->t_flags & TF_SACK_PERMIT) && (save_tlen > 0) && TCPS_HAVEESTABLISHED(tp->t_state)) { if ((tlen == 0) && (SEQ_LT(save_start, save_rnxt))) { /* * DSACK actually handled in the fastpath * above. */ RACK_OPTS_INC(tcp_sack_path_1); tcp_update_sack_list(tp, save_start, save_start + save_tlen); } else if ((tlen > 0) && SEQ_GT(tp->rcv_nxt, save_rnxt)) { if ((tp->rcv_numsacks >= 1) && (tp->sackblks[0].end == save_start)) { /* * Partial overlap, recorded at todrop * above. */ RACK_OPTS_INC(tcp_sack_path_2a); tcp_update_sack_list(tp, tp->sackblks[0].start, tp->sackblks[0].end); } else { RACK_OPTS_INC(tcp_sack_path_2b); tcp_update_dsack_list(tp, save_start, save_start + save_tlen); } } else if (tlen >= save_tlen) { /* Update of sackblks. */ RACK_OPTS_INC(tcp_sack_path_3); tcp_update_dsack_list(tp, save_start, save_start + save_tlen); } else if (tlen > 0) { RACK_OPTS_INC(tcp_sack_path_4); tcp_update_dsack_list(tp, save_start, save_start + tlen); } } } else { m_freem(m); thflags &= ~TH_FIN; } /* * If FIN is received ACK the FIN and let the user know that the * connection is closing. */ if (thflags & TH_FIN) { if (TCPS_HAVERCVDFIN(tp->t_state) == 0) { /* The socket upcall is handled by socantrcvmore. */ socantrcvmore(so); /* * If connection is half-synchronized (ie NEEDSYN * flag on) then delay ACK, so it may be piggybacked * when SYN is sent. Otherwise, since we received a * FIN then no more input can be expected, send ACK * now. */ if (tp->t_flags & TF_NEEDSYN) { rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); tp->t_flags |= TF_DELACK; } else { tp->t_flags |= TF_ACKNOW; } tp->rcv_nxt++; } switch (tp->t_state) { /* * In SYN_RECEIVED and ESTABLISHED STATES enter the * CLOSE_WAIT state. */ case TCPS_SYN_RECEIVED: tp->t_starttime = ticks; /* FALLTHROUGH */ case TCPS_ESTABLISHED: rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); tcp_state_change(tp, TCPS_CLOSE_WAIT); break; /* * If still in FIN_WAIT_1 STATE FIN has not been * acked so enter the CLOSING state. */ case TCPS_FIN_WAIT_1: rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); tcp_state_change(tp, TCPS_CLOSING); break; /* * In FIN_WAIT_2 state enter the TIME_WAIT state, * starting the time-wait timer, turning off the * other standard timers. */ case TCPS_FIN_WAIT_2: rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); tcp_twstart(tp); return (1); } } /* * Return any desired output. */ if ((tp->t_flags & TF_ACKNOW) || (sbavail(&so->so_snd) > (tp->snd_max - tp->snd_una))) { rack->r_wanted_output = 1; } INP_WLOCK_ASSERT(tp->t_inpcb); return (0); } /* * Here nothing is really faster, its just that we * have broken out the fast-data path also just like * the fast-ack. */ static int rack_do_fastnewdata(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint8_t iptos) { int32_t nsegs; int32_t newsize = 0; /* automatic sockbuf scaling */ struct tcp_rack *rack; #ifdef NETFLIX_SB_LIMITS u_int mcnt, appended; #endif #ifdef TCPDEBUG /* * The size of tcp_saveipgen must be the size of the max ip header, * now IPv6. */ u_char tcp_saveipgen[IP6_HDR_LEN]; struct tcphdr tcp_savetcp; short ostate = 0; #endif /* * If last ACK falls within this segment's sequence numbers, record * the timestamp. NOTE that the test is modified according to the * latest proposal of the tcplw@cray.com list (Braden 1993/04/26). */ if (__predict_false(th->th_seq != tp->rcv_nxt)) { return (0); } if (__predict_false(tp->snd_nxt != tp->snd_max)) { return (0); } if (tiwin && tiwin != tp->snd_wnd) { return (0); } if (__predict_false((tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN)))) { return (0); } if (__predict_false((to->to_flags & TOF_TS) && (TSTMP_LT(to->to_tsval, tp->ts_recent)))) { return (0); } if (__predict_false((th->th_ack != tp->snd_una))) { return (0); } if (__predict_false(tlen > sbspace(&so->so_rcv))) { return (0); } if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to->to_tsval; } rack = (struct tcp_rack *)tp->t_fb_ptr; /* * This is a pure, in-sequence data packet with nothing on the * reassembly queue and we have enough buffer space to take it. */ nsegs = max(1, m->m_pkthdr.lro_nsegs); #ifdef NETFLIX_SB_LIMITS if (so->so_rcv.sb_shlim) { mcnt = m_memcnt(m); appended = 0; if (counter_fo_get(so->so_rcv.sb_shlim, mcnt, CFO_NOSLEEP, NULL) == false) { counter_u64_add(tcp_sb_shlim_fails, 1); m_freem(m); return (1); } } #endif /* Clean receiver SACK report if present */ if (tp->rcv_numsacks) tcp_clean_sackreport(tp); KMOD_TCPSTAT_INC(tcps_preddat); tp->rcv_nxt += tlen; if (tlen && ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) && (tp->t_fbyte_in == 0)) { tp->t_fbyte_in = ticks; if (tp->t_fbyte_in == 0) tp->t_fbyte_in = 1; if (tp->t_fbyte_out && tp->t_fbyte_in) tp->t_flags2 |= TF2_FBYTES_COMPLETE; } /* * Pull snd_wl1 up to prevent seq wrap relative to th_seq. */ tp->snd_wl1 = th->th_seq; /* * Pull rcv_up up to prevent seq wrap relative to rcv_nxt. */ tp->rcv_up = tp->rcv_nxt; KMOD_TCPSTAT_ADD(tcps_rcvpack, nsegs); KMOD_TCPSTAT_ADD(tcps_rcvbyte, tlen); #ifdef TCPDEBUG if (so->so_options & SO_DEBUG) tcp_trace(TA_INPUT, ostate, tp, (void *)tcp_saveipgen, &tcp_savetcp, 0); #endif newsize = tcp_autorcvbuf(m, th, so, tp, tlen); /* Add data to socket buffer. */ SOCKBUF_LOCK(&so->so_rcv); if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { m_freem(m); } else { /* * Set new socket buffer size. Give up when limit is * reached. */ if (newsize) if (!sbreserve_locked(so, SO_RCV, newsize, NULL)) so->so_rcv.sb_flags &= ~SB_AUTOSIZE; m_adj(m, drop_hdrlen); /* delayed header drop */ #ifdef NETFLIX_SB_LIMITS appended = #endif sbappendstream_locked(&so->so_rcv, m, 0); ctf_calc_rwin(so, tp); } rack_log_wakeup(tp,rack, &so->so_rcv, tlen, 1); /* NB: sorwakeup_locked() does an implicit unlock. */ sorwakeup_locked(so); #ifdef NETFLIX_SB_LIMITS if (so->so_rcv.sb_shlim && mcnt != appended) counter_fo_release(so->so_rcv.sb_shlim, mcnt - appended); #endif rack_handle_delayed_ack(tp, rack, tlen, 0); if (tp->snd_una == tp->snd_max) sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una); return (1); } /* * This subfunction is used to try to highly optimize the * fast path. We again allow window updates that are * in sequence to remain in the fast-path. We also add * in the __predict's to attempt to help the compiler. * Note that if we return a 0, then we can *not* process * it and the caller should push the packet into the * slow-path. */ static int rack_fastack(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t nxt_pkt, uint32_t cts) { int32_t acked; int32_t nsegs; #ifdef TCPDEBUG /* * The size of tcp_saveipgen must be the size of the max ip header, * now IPv6. */ u_char tcp_saveipgen[IP6_HDR_LEN]; struct tcphdr tcp_savetcp; short ostate = 0; #endif int32_t under_pacing = 0; struct tcp_rack *rack; if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { /* Old ack, behind (or duplicate to) the last one rcv'd */ return (0); } if (__predict_false(SEQ_GT(th->th_ack, tp->snd_max))) { /* Above what we have sent? */ return (0); } if (__predict_false(tp->snd_nxt != tp->snd_max)) { /* We are retransmitting */ return (0); } if (__predict_false(tiwin == 0)) { /* zero window */ return (0); } if (__predict_false(tp->t_flags & (TF_NEEDSYN | TF_NEEDFIN))) { /* We need a SYN or a FIN, unlikely.. */ return (0); } if ((to->to_flags & TOF_TS) && __predict_false(TSTMP_LT(to->to_tsval, tp->ts_recent))) { /* Timestamp is behind .. old ack with seq wrap? */ return (0); } if (__predict_false(IN_RECOVERY(tp->t_flags))) { /* Still recovering */ return (0); } rack = (struct tcp_rack *)tp->t_fb_ptr; if (rack->r_ctl.rc_sacked) { /* We have sack holes on our scoreboard */ return (0); } /* Ok if we reach here, we can process a fast-ack */ if (rack->gp_ready && (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { under_pacing = 1; } nsegs = max(1, m->m_pkthdr.lro_nsegs); rack_log_ack(tp, to, th, 0, 0); /* Did the window get updated? */ if (tiwin != tp->snd_wnd) { tp->snd_wnd = tiwin; rack_validate_fo_sendwin_up(tp, rack); tp->snd_wl1 = th->th_seq; if (tp->snd_wnd > tp->max_sndwnd) tp->max_sndwnd = tp->snd_wnd; } /* Do we exit persists? */ if ((rack->rc_in_persist != 0) && (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs))) { rack_exit_persist(tp, rack, cts); } /* Do we enter persists? */ if ((rack->rc_in_persist == 0) && (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) && TCPS_HAVEESTABLISHED(tp->t_state) && ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) && sbavail(&tp->t_inpcb->inp_socket->so_snd) && (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) { /* * Here the rwnd is less than * the pacing size, we are established, * nothing is outstanding, and there is * data to send. Enter persists. */ rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime); } /* * If last ACK falls within this segment's sequence numbers, record * the timestamp. NOTE that the test is modified according to the * latest proposal of the tcplw@cray.com list (Braden 1993/04/26). */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent)) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to->to_tsval; } /* * This is a pure ack for outstanding data. */ KMOD_TCPSTAT_INC(tcps_predack); /* * "bad retransmit" recovery. */ if ((tp->t_flags & TF_PREVVALID) && ((tp->t_flags & TF_RCVD_TSTMP) == 0)) { tp->t_flags &= ~TF_PREVVALID; if (tp->t_rxtshift == 1 && (int)(ticks - tp->t_badrxtwin) < 0) rack_cong_signal(tp, CC_RTO_ERR, th->th_ack, __LINE__); } /* * Recalculate the transmit timer / rtt. * * Some boxes send broken timestamp replies during the SYN+ACK * phase, ignore timestamps of 0 or we could calculate a huge RTT * and blow up the retransmit timer. */ acked = BYTES_THIS_ACK(tp, th); #ifdef TCP_HHOOK /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */ hhook_run_tcp_est_in(tp, th, to); #endif KMOD_TCPSTAT_ADD(tcps_rcvackpack, nsegs); KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked); if (acked) { struct mbuf *mfree; rack_ack_received(tp, rack, th->th_ack, nsegs, CC_ACK, 0); SOCKBUF_LOCK(&so->so_snd); mfree = sbcut_locked(&so->so_snd, acked); tp->snd_una = th->th_ack; /* Note we want to hold the sb lock through the sendmap adjust */ rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una); /* Wake up the socket if we have room to write more */ rack_log_wakeup(tp,rack, &so->so_snd, acked, 2); sowwakeup_locked(so); m_freem(mfree); tp->t_rxtshift = 0; RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp), rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop); rack->rc_tlp_in_progress = 0; rack->r_ctl.rc_tlp_cnt_out = 0; /* * If it is the RXT timer we want to * stop it, so we can restart a TLP. */ if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); #ifdef NETFLIX_HTTP_LOGGING tcp_http_check_for_comp(rack->rc_tp, th->th_ack); #endif } /* * Let the congestion control algorithm update congestion control * related information. This typically means increasing the * congestion window. */ if (tp->snd_wnd < ctf_outstanding(tp)) { /* The peer collapsed the window */ rack_collapsed_window(rack, ctf_outstanding(tp), __LINE__); } else if (rack->rc_has_collapsed) rack_un_collapse_window(rack, __LINE__); if ((rack->r_collapse_point_valid) && (SEQ_GT(tp->snd_una, rack->r_ctl.high_collapse_point))) rack->r_collapse_point_valid = 0; /* * Pull snd_wl2 up to prevent seq wrap relative to th_ack. */ tp->snd_wl2 = th->th_ack; tp->t_dupacks = 0; m_freem(m); /* ND6_HINT(tp); *//* Some progress has been made. */ /* * If all outstanding data are acked, stop retransmit timer, * otherwise restart timer using current (possibly backed-off) * value. If process is waiting for space, wakeup/selwakeup/signal. * If data are ready to send, let tcp_output decide between more * output or persist. */ #ifdef TCPDEBUG if (so->so_options & SO_DEBUG) tcp_trace(TA_INPUT, ostate, tp, (void *)tcp_saveipgen, &tcp_savetcp, 0); #endif if (under_pacing && (rack->use_fixed_rate == 0) && (rack->in_probe_rtt == 0) && rack->rc_gp_dyn_mul && rack->rc_always_pace) { /* Check if we are dragging bottom */ rack_check_bottom_drag(tp, rack, so, acked); } if (tp->snd_una == tp->snd_max) { tp->t_flags &= ~TF_PREVVALID; rack->r_ctl.retran_during_recovery = 0; rack->r_ctl.dsack_byte_cnt = 0; rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL); if (rack->r_ctl.rc_went_idle_time == 0) rack->r_ctl.rc_went_idle_time = 1; rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__); if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0) tp->t_acktime = 0; rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); } if (acked && rack->r_fast_output) rack_gain_for_fastoutput(rack, tp, so, (uint32_t)acked); if (sbavail(&so->so_snd)) { rack->r_wanted_output = 1; } return (1); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCP is still * locked. */ static int rack_do_syn_sent(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t ret_val = 0; int32_t todrop; int32_t ourfinisacked = 0; struct tcp_rack *rack; ctf_calc_rwin(so, tp); /* * If the state is SYN_SENT: if seg contains an ACK, but not for our * SYN, drop the input. if seg contains a RST, then drop the * connection. if seg does not contain SYN, then drop it. Otherwise * this is an acceptable SYN segment initialize tp->rcv_nxt and * tp->irs if seg contains ack then advance tp->snd_una if seg * contains an ECE and ECN support is enabled, the stream is ECN * capable. if SYN has been acked change to ESTABLISHED else * SYN_RCVD state arrange for segment to be acked (eventually) * continue processing rest of data/controls. */ if ((thflags & TH_ACK) && (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) { tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } if ((thflags & (TH_ACK | TH_RST)) == (TH_ACK | TH_RST)) { TCP_PROBE5(connect__refused, NULL, tp, mtod(m, const char *), tp, th); tp = tcp_drop(tp, ECONNREFUSED); ctf_do_drop(m, tp); return (1); } if (thflags & TH_RST) { ctf_do_drop(m, tp); return (1); } if (!(thflags & TH_SYN)) { ctf_do_drop(m, tp); return (1); } tp->irs = th->th_seq; tcp_rcvseqinit(tp); rack = (struct tcp_rack *)tp->t_fb_ptr; if (thflags & TH_ACK) { int tfo_partial = 0; KMOD_TCPSTAT_INC(tcps_connects); soisconnected(so); #ifdef MAC mac_socketpeer_set_from_mbuf(m, so); #endif /* Do window scaling on this connection? */ if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == (TF_RCVD_SCALE | TF_REQ_SCALE)) { tp->rcv_scale = tp->request_r_scale; } tp->rcv_adv += min(tp->rcv_wnd, TCP_MAXWIN << tp->rcv_scale); /* * If not all the data that was sent in the TFO SYN * has been acked, resend the remainder right away. */ if (IS_FASTOPEN(tp->t_flags) && (tp->snd_una != tp->snd_max)) { tp->snd_nxt = th->th_ack; tfo_partial = 1; } /* * If there's data, delay ACK; if there's also a FIN ACKNOW * will be turned on later. */ if (DELAY_ACK(tp, tlen) && tlen != 0 && !tfo_partial) { rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); tp->t_flags |= TF_DELACK; } else { rack->r_wanted_output = 1; tp->t_flags |= TF_ACKNOW; rack->rc_dack_toggle = 0; } tcp_ecn_input_syn_sent(tp, thflags, iptos); if (SEQ_GT(th->th_ack, tp->snd_una)) { /* * We advance snd_una for the * fast open case. If th_ack is * acknowledging data beyond * snd_una we can't just call * ack-processing since the * data stream in our send-map * will start at snd_una + 1 (one * beyond the SYN). If its just * equal we don't need to do that * and there is no send_map. */ tp->snd_una++; } /* * Received in SYN_SENT[*] state. Transitions: * SYN_SENT --> ESTABLISHED SYN_SENT* --> FIN_WAIT_1 */ tp->t_starttime = ticks; if (tp->t_flags & TF_NEEDFIN) { tcp_state_change(tp, TCPS_FIN_WAIT_1); tp->t_flags &= ~TF_NEEDFIN; thflags &= ~TH_SYN; } else { tcp_state_change(tp, TCPS_ESTABLISHED); TCP_PROBE5(connect__established, NULL, tp, mtod(m, const char *), tp, th); rack_cc_conn_init(tp); } } else { /* * Received initial SYN in SYN-SENT[*] state => simultaneous * open. If segment contains CC option and there is a * cached CC, apply TAO test. If it succeeds, connection is * * half-synchronized. Otherwise, do 3-way handshake: * SYN-SENT -> SYN-RECEIVED SYN-SENT* -> SYN-RECEIVED* If * there was no CC option, clear cached CC value. */ tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN | TF_SONOTCONN); tcp_state_change(tp, TCPS_SYN_RECEIVED); } INP_WLOCK_ASSERT(tp->t_inpcb); /* * Advance th->th_seq to correspond to first data byte. If data, * trim to stay within window, dropping FIN if necessary. */ th->th_seq++; if (tlen > tp->rcv_wnd) { todrop = tlen - tp->rcv_wnd; m_adj(m, -todrop); tlen = tp->rcv_wnd; thflags &= ~TH_FIN; KMOD_TCPSTAT_INC(tcps_rcvpackafterwin); KMOD_TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop); } tp->snd_wl1 = th->th_seq - 1; tp->rcv_up = th->th_seq; /* * Client side of transaction: already sent SYN and data. If the * remote host used T/TCP to validate the SYN, our data will be * ACK'd; if so, enter normal data segment processing in the middle * of step 5, ack processing. Otherwise, goto step 6. */ if (thflags & TH_ACK) { /* For syn-sent we need to possibly update the rtt */ if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) { uint32_t t, mcts; mcts = tcp_ts_getticks(); t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC; if (!tp->t_rttlow || tp->t_rttlow > t) tp->t_rttlow = t; rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 4); tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2); tcp_rack_xmit_timer_commit(rack, tp); } if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) return (ret_val); /* We may have changed to FIN_WAIT_1 above */ if (tp->t_state == TCPS_FIN_WAIT_1) { /* * In FIN_WAIT_1 STATE in addition to the processing * for the ESTABLISHED state if our FIN is now * acknowledged then enter FIN_WAIT_2. */ if (ourfinisacked) { /* * If we can't receive any more data, then * closing user can proceed. Starting the * timer is contrary to the specification, * but if we don't get a FIN we'll hang * forever. * * XXXjl: we should release the tp also, and * use a compressed state. */ if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { soisdisconnected(so); tcp_timer_activate(tp, TT_2MSL, (tcp_fast_finwait2_recycle ? tcp_finwait2_timeout : TP_MAXIDLE(tp))); } tcp_state_change(tp, TCPS_FIN_WAIT_2); } } } return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCP is still * locked. */ static int rack_do_syn_recv(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { struct tcp_rack *rack; int32_t ret_val = 0; int32_t ourfinisacked = 0; ctf_calc_rwin(so, tp); if ((thflags & TH_ACK) && (SEQ_LEQ(th->th_ack, tp->snd_una) || SEQ_GT(th->th_ack, tp->snd_max))) { tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } rack = (struct tcp_rack *)tp->t_fb_ptr; if (IS_FASTOPEN(tp->t_flags)) { /* * When a TFO connection is in SYN_RECEIVED, the * only valid packets are the initial SYN, a * retransmit/copy of the initial SYN (possibly with * a subset of the original data), a valid ACK, a * FIN, or a RST. */ if ((thflags & (TH_SYN | TH_ACK)) == (TH_SYN | TH_ACK)) { tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } else if (thflags & TH_SYN) { /* non-initial SYN is ignored */ if ((rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) || (rack->r_ctl.rc_hpts_flags & PACE_TMR_TLP) || (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK)) { ctf_do_drop(m, NULL); return (0); } } else if (!(thflags & (TH_ACK | TH_FIN | TH_RST))) { ctf_do_drop(m, NULL); return (0); } } if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (__ctf_process_rst(m, th, so, tp, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt)); /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } /* * In the SYN-RECEIVED state, validate that the packet belongs to * this connection before trimming the data to fit the receive * window. Check the sequence number versus IRS since we know the * sequence numbers haven't wrapped. This is a partial fix for the * "LAND" DoS attack. */ if (SEQ_LT(th->th_seq, tp->irs)) { tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt)) { return (ret_val); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to->to_tsval; } tp->snd_wnd = tiwin; rack_validate_fo_sendwin_up(tp, rack); /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (IS_FASTOPEN(tp->t_flags)) { rack_cc_conn_init(tp); } return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } KMOD_TCPSTAT_INC(tcps_connects); if (tp->t_flags & TF_SONOTCONN) { tp->t_flags &= ~TF_SONOTCONN; soisconnected(so); } /* Do window scaling? */ if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == (TF_RCVD_SCALE | TF_REQ_SCALE)) { tp->rcv_scale = tp->request_r_scale; } /* * Make transitions: SYN-RECEIVED -> ESTABLISHED SYN-RECEIVED* -> * FIN-WAIT-1 */ tp->t_starttime = ticks; if (IS_FASTOPEN(tp->t_flags) && tp->t_tfo_pending) { tcp_fastopen_decrement_counter(tp->t_tfo_pending); tp->t_tfo_pending = NULL; } if (tp->t_flags & TF_NEEDFIN) { tcp_state_change(tp, TCPS_FIN_WAIT_1); tp->t_flags &= ~TF_NEEDFIN; } else { tcp_state_change(tp, TCPS_ESTABLISHED); TCP_PROBE5(accept__established, NULL, tp, mtod(m, const char *), tp, th); /* * TFO connections call cc_conn_init() during SYN * processing. Calling it again here for such connections * is not harmless as it would undo the snd_cwnd reduction * that occurs when a TFO SYN|ACK is retransmitted. */ if (!IS_FASTOPEN(tp->t_flags)) rack_cc_conn_init(tp); } /* * Account for the ACK of our SYN prior to * regular ACK processing below, except for * simultaneous SYN, which is handled later. */ if (SEQ_GT(th->th_ack, tp->snd_una) && !(tp->t_flags & TF_NEEDSYN)) tp->snd_una++; /* * If segment contains data or ACK, will call tcp_reass() later; if * not, do so now to pass queued data to user. */ if (tlen == 0 && (thflags & TH_FIN) == 0) { (void) tcp_reass(tp, (struct tcphdr *)0, NULL, 0, (struct mbuf *)0); if (tp->t_flags & TF_WAKESOR) { tp->t_flags &= ~TF_WAKESOR; /* NB: sorwakeup_locked() does an implicit unlock. */ sorwakeup_locked(so); } } tp->snd_wl1 = th->th_seq - 1; /* For syn-recv we need to possibly update the rtt */ if ((to->to_flags & TOF_TS) != 0 && to->to_tsecr) { uint32_t t, mcts; mcts = tcp_ts_getticks(); t = (mcts - to->to_tsecr) * HPTS_USEC_IN_MSEC; if (!tp->t_rttlow || tp->t_rttlow > t) tp->t_rttlow = t; rack_log_rtt_sample_calc(rack, t, (to->to_tsecr * 1000), (mcts * 1000), 5); tcp_rack_xmit_timer(rack, t + 1, 1, t, 0, NULL, 2); tcp_rack_xmit_timer_commit(rack, tp); } if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { return (ret_val); } if (tp->t_state == TCPS_FIN_WAIT_1) { /* We could have went to FIN_WAIT_1 (or EST) above */ /* * In FIN_WAIT_1 STATE in addition to the processing for the * ESTABLISHED state if our FIN is now acknowledged then * enter FIN_WAIT_2. */ if (ourfinisacked) { /* * If we can't receive any more data, then closing * user can proceed. Starting the timer is contrary * to the specification, but if we don't get a FIN * we'll hang forever. * * XXXjl: we should release the tp also, and use a * compressed state. */ if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { soisdisconnected(so); tcp_timer_activate(tp, TT_2MSL, (tcp_fast_finwait2_recycle ? tcp_finwait2_timeout : TP_MAXIDLE(tp))); } tcp_state_change(tp, TCPS_FIN_WAIT_2); } } return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCP is still * locked. */ static int rack_do_established(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t ret_val = 0; struct tcp_rack *rack; /* * Header prediction: check for the two common cases of a * uni-directional data xfer. If the packet has no control flags, * is in-sequence, the window didn't change and we're not * retransmitting, it's a candidate. If the length is zero and the * ack moved forward, we're the sender side of the xfer. Just free * the data acked & wake any higher level process that was blocked * waiting for space. If the length is non-zero and the ack didn't * move, we're the receiver side. If we're getting packets in-order * (the reassembly queue is empty), add the data toc The socket * buffer and note that we need a delayed ack. Make sure that the * hidden state-flags are also off. Since we check for * TCPS_ESTABLISHED first, it can only be TH_NEEDSYN. */ rack = (struct tcp_rack *)tp->t_fb_ptr; if (__predict_true(((to->to_flags & TOF_SACK) == 0)) && __predict_true((thflags & (TH_SYN | TH_FIN | TH_RST | TH_ACK)) == TH_ACK) && __predict_true(SEGQ_EMPTY(tp)) && __predict_true(th->th_seq == tp->rcv_nxt)) { if (tlen == 0) { if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen, tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime)) { return (0); } } else { if (rack_do_fastnewdata(m, th, so, tp, to, drop_hdrlen, tlen, tiwin, nxt_pkt, iptos)) { return (0); } } } ctf_calc_rwin(so, tp); if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (__ctf_process_rst(m, th, so, tp, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt)) { return (ret_val); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1; return (ret_val); } else { ctf_do_drop(m, NULL); return (0); } } /* * Ack processing. */ if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) { return (ret_val); } if (sbavail(&so->so_snd)) { if (ctf_progress_timeout_check(tp, true)) { rack_log_progress_event(rack, tp, tick, PROGRESS_DROP, __LINE__); ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } } /* State changes only happen in rack_process_data() */ return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCP is still * locked. */ static int rack_do_close_wait(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t ret_val = 0; struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; ctf_calc_rwin(so, tp); if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (__ctf_process_rst(m, th, so, tp, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt)) { return (ret_val); } /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1; return (ret_val); } else { ctf_do_drop(m, NULL); return (0); } } /* * Ack processing. */ if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, NULL, thflags, &ret_val)) { return (ret_val); } if (sbavail(&so->so_snd)) { if (ctf_progress_timeout_check(tp, true)) { rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr, tp, tick, PROGRESS_DROP, __LINE__); ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } } return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } static int rack_check_data_after_close(struct mbuf *m, struct tcpcb *tp, int32_t *tlen, struct tcphdr *th, struct socket *so) { struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; if (rack->rc_allow_data_af_clo == 0) { close_now: tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE); /* tcp_close will kill the inp pre-log the Reset */ tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); tp = tcp_close(tp); KMOD_TCPSTAT_INC(tcps_rcvafterclose); ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, (*tlen)); return (1); } if (sbavail(&so->so_snd) == 0) goto close_now; /* Ok we allow data that is ignored and a followup reset */ tcp_log_end_status(tp, TCP_EI_STATUS_DATA_A_CLOSE); tp->rcv_nxt = th->th_seq + *tlen; tp->t_flags2 |= TF2_DROP_AF_DATA; rack->r_wanted_output = 1; *tlen = 0; return (0); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCP is still * locked. */ static int rack_do_fin_wait_1(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t ret_val = 0; int32_t ourfinisacked = 0; struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; ctf_calc_rwin(so, tp); if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (__ctf_process_rst(m, th, so, tp, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt)) { return (ret_val); } /* * If new data are received on a connection after the user processes * are gone, then RST the other end. */ if ((tp->t_flags & TF_CLOSED) && tlen && rack_check_data_after_close(m, tp, &tlen, th, so)) return (1); /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1; return (ret_val); } else { ctf_do_drop(m, NULL); return (0); } } /* * Ack processing. */ if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { return (ret_val); } if (ourfinisacked) { /* * If we can't receive any more data, then closing user can * proceed. Starting the timer is contrary to the * specification, but if we don't get a FIN we'll hang * forever. * * XXXjl: we should release the tp also, and use a * compressed state. */ if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { soisdisconnected(so); tcp_timer_activate(tp, TT_2MSL, (tcp_fast_finwait2_recycle ? tcp_finwait2_timeout : TP_MAXIDLE(tp))); } tcp_state_change(tp, TCPS_FIN_WAIT_2); } if (sbavail(&so->so_snd)) { if (ctf_progress_timeout_check(tp, true)) { rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr, tp, tick, PROGRESS_DROP, __LINE__); ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } } return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCP is still * locked. */ static int rack_do_closing(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t ret_val = 0; int32_t ourfinisacked = 0; struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; ctf_calc_rwin(so, tp); if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (__ctf_process_rst(m, th, so, tp, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt)) { return (ret_val); } /* * If new data are received on a connection after the user processes * are gone, then RST the other end. */ if ((tp->t_flags & TF_CLOSED) && tlen && rack_check_data_after_close(m, tp, &tlen, th, so)) return (1); /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1; return (ret_val); } else { ctf_do_drop(m, NULL); return (0); } } /* * Ack processing. */ if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { return (ret_val); } if (ourfinisacked) { tcp_twstart(tp); m_freem(m); return (1); } if (sbavail(&so->so_snd)) { if (ctf_progress_timeout_check(tp, true)) { rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr, tp, tick, PROGRESS_DROP, __LINE__); ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } } return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCP is still * locked. */ static int rack_do_lastack(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t ret_val = 0; int32_t ourfinisacked = 0; struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; ctf_calc_rwin(so, tp); if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (__ctf_process_rst(m, th, so, tp, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt)) { return (ret_val); } /* * If new data are received on a connection after the user processes * are gone, then RST the other end. */ if ((tp->t_flags & TF_CLOSED) && tlen && rack_check_data_after_close(m, tp, &tlen, th, so)) return (1); /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1; return (ret_val); } else { ctf_do_drop(m, NULL); return (0); } } /* * case TCPS_LAST_ACK: Ack processing. */ if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { return (ret_val); } if (ourfinisacked) { tp = tcp_close(tp); ctf_do_drop(m, tp); return (1); } if (sbavail(&so->so_snd)) { if (ctf_progress_timeout_check(tp, true)) { rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr, tp, tick, PROGRESS_DROP, __LINE__); ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } } return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } /* * Return value of 1, the TCB is unlocked and most * likely gone, return value of 0, the TCP is still * locked. */ static int rack_do_fin_wait_2(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, struct tcpopt *to, int32_t drop_hdrlen, int32_t tlen, uint32_t tiwin, int32_t thflags, int32_t nxt_pkt, uint8_t iptos) { int32_t ret_val = 0; int32_t ourfinisacked = 0; struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; ctf_calc_rwin(so, tp); /* Reset receive buffer auto scaling when not in bulk receive mode. */ if ((thflags & TH_RST) || (tp->t_fin_is_rst && (thflags & TH_FIN))) return (__ctf_process_rst(m, th, so, tp, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt)); /* * RFC5961 Section 4.2 Send challenge ACK for any SYN in * synchronized state. */ if (thflags & TH_SYN) { ctf_challenge_ack(m, th, tp, &ret_val); return (ret_val); } /* * RFC 1323 PAWS: If we have a timestamp reply on this segment and * it's less than ts_recent, drop it. */ if ((to->to_flags & TOF_TS) != 0 && tp->ts_recent && TSTMP_LT(to->to_tsval, tp->ts_recent)) { if (ctf_ts_check(m, th, tp, tlen, thflags, &ret_val)) return (ret_val); } if (_ctf_drop_checks(to, m, th, tp, &tlen, &thflags, &drop_hdrlen, &ret_val, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt)) { return (ret_val); } /* * If new data are received on a connection after the user processes * are gone, then RST the other end. */ if ((tp->t_flags & TF_CLOSED) && tlen && rack_check_data_after_close(m, tp, &tlen, th, so)) return (1); /* * If last ACK falls within this segment's sequence numbers, record * its timestamp. NOTE: 1) That the test incorporates suggestions * from the latest proposal of the tcplw@cray.com list (Braden * 1993/04/26). 2) That updating only on newer timestamps interferes * with our earlier PAWS tests, so this check should be solely * predicated on the sequence space of this segment. 3) That we * modify the segment boundary check to be Last.ACK.Sent <= SEG.SEQ * + SEG.Len instead of RFC1323's Last.ACK.Sent < SEG.SEQ + * SEG.Len, This modified check allows us to overcome RFC1323's * limitations as described in Stevens TCP/IP Illustrated Vol. 2 * p.869. In such cases, we can still calculate the RTT correctly * when RCV.NXT == Last.ACK.Sent. */ if ((to->to_flags & TOF_TS) != 0 && SEQ_LEQ(th->th_seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, th->th_seq + tlen + ((thflags & (TH_SYN | TH_FIN)) != 0))) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = to->to_tsval; } /* * If the ACK bit is off: if in SYN-RECEIVED state or SENDSYN flag * is on (half-synchronized state), then queue data for later * processing; else drop segment and return. */ if ((thflags & TH_ACK) == 0) { if (tp->t_flags & TF_NEEDSYN) { return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } else if (tp->t_flags & TF_ACKNOW) { ctf_do_dropafterack(m, tp, th, thflags, tlen, &ret_val); ((struct tcp_rack *)tp->t_fb_ptr)->r_wanted_output = 1; return (ret_val); } else { ctf_do_drop(m, NULL); return (0); } } /* * Ack processing. */ if (rack_process_ack(m, th, so, tp, to, tiwin, tlen, &ourfinisacked, thflags, &ret_val)) { return (ret_val); } if (sbavail(&so->so_snd)) { if (ctf_progress_timeout_check(tp, true)) { rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr, tp, tick, PROGRESS_DROP, __LINE__); ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } } return (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } static void inline rack_clear_rate_sample(struct tcp_rack *rack) { rack->r_ctl.rack_rs.rs_flags = RACK_RTT_EMPTY; rack->r_ctl.rack_rs.rs_rtt_cnt = 0; rack->r_ctl.rack_rs.rs_rtt_tot = 0; } static void rack_set_pace_segments(struct tcpcb *tp, struct tcp_rack *rack, uint32_t line, uint64_t *fill_override) { uint64_t bw_est, rate_wanted; int chged = 0; uint32_t user_max, orig_min, orig_max; orig_min = rack->r_ctl.rc_pace_min_segs; orig_max = rack->r_ctl.rc_pace_max_segs; user_max = ctf_fixed_maxseg(tp) * rack->rc_user_set_max_segs; if (ctf_fixed_maxseg(tp) != rack->r_ctl.rc_pace_min_segs) chged = 1; rack->r_ctl.rc_pace_min_segs = ctf_fixed_maxseg(tp); if (rack->use_fixed_rate || rack->rc_force_max_seg) { if (user_max != rack->r_ctl.rc_pace_max_segs) chged = 1; } if (rack->rc_force_max_seg) { rack->r_ctl.rc_pace_max_segs = user_max; } else if (rack->use_fixed_rate) { bw_est = rack_get_bw(rack); if ((rack->r_ctl.crte == NULL) || (bw_est != rack->r_ctl.crte->rate)) { rack->r_ctl.rc_pace_max_segs = user_max; } else { /* We are pacing right at the hardware rate */ uint32_t segsiz; segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs); rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size( tp, bw_est, segsiz, 0, rack->r_ctl.crte, NULL); } } else if (rack->rc_always_pace) { if (rack->r_ctl.gp_bw || #ifdef NETFLIX_PEAKRATE rack->rc_tp->t_maxpeakrate || #endif rack->r_ctl.init_rate) { /* We have a rate of some sort set */ uint32_t orig; bw_est = rack_get_bw(rack); orig = rack->r_ctl.rc_pace_max_segs; if (fill_override) rate_wanted = *fill_override; else rate_wanted = rack_get_output_bw(rack, bw_est, NULL, NULL); if (rate_wanted) { /* We have something */ rack->r_ctl.rc_pace_max_segs = rack_get_pacing_len(rack, rate_wanted, ctf_fixed_maxseg(rack->rc_tp)); } else rack->r_ctl.rc_pace_max_segs = rack->r_ctl.rc_pace_min_segs; if (orig != rack->r_ctl.rc_pace_max_segs) chged = 1; } else if ((rack->r_ctl.gp_bw == 0) && (rack->r_ctl.rc_pace_max_segs == 0)) { /* * If we have nothing limit us to bursting * out IW sized pieces. */ chged = 1; rack->r_ctl.rc_pace_max_segs = rc_init_window(rack); } } if (rack->r_ctl.rc_pace_max_segs > PACE_MAX_IP_BYTES) { chged = 1; rack->r_ctl.rc_pace_max_segs = PACE_MAX_IP_BYTES; } if (chged) rack_log_type_pacing_sizes(tp, rack, orig_min, orig_max, line, 2); } static void rack_init_fsb_block(struct tcpcb *tp, struct tcp_rack *rack) { #ifdef INET6 struct ip6_hdr *ip6 = NULL; #endif #ifdef INET struct ip *ip = NULL; #endif struct udphdr *udp = NULL; /* Ok lets fill in the fast block, it can only be used with no IP options! */ #ifdef INET6 if (rack->r_is_v6) { rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr; if (tp->t_port) { rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr); udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr)); udp->uh_sport = htons(V_tcp_udp_tunneling_port); udp->uh_dport = tp->t_port; rack->r_ctl.fsb.udp = udp; rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1); } else { rack->r_ctl.fsb.th = (struct tcphdr *)(ip6 + 1); rack->r_ctl.fsb.udp = NULL; } tcpip_fillheaders(rack->rc_inp, tp->t_port, ip6, rack->r_ctl.fsb.th); } else #endif /* INET6 */ { rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr); ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr; if (tp->t_port) { rack->r_ctl.fsb.tcp_ip_hdr_len += sizeof(struct udphdr); udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip)); udp->uh_sport = htons(V_tcp_udp_tunneling_port); udp->uh_dport = tp->t_port; rack->r_ctl.fsb.udp = udp; rack->r_ctl.fsb.th = (struct tcphdr *)(udp + 1); } else { rack->r_ctl.fsb.udp = NULL; rack->r_ctl.fsb.th = (struct tcphdr *)(ip + 1); } tcpip_fillheaders(rack->rc_inp, tp->t_port, ip, rack->r_ctl.fsb.th); } rack->r_fsb_inited = 1; } static int rack_init_fsb(struct tcpcb *tp, struct tcp_rack *rack) { /* * Allocate the larger of spaces V6 if available else just * V4 and include udphdr (overbook) */ #ifdef INET6 rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct ip6_hdr) + sizeof(struct tcphdr) + sizeof(struct udphdr); #else rack->r_ctl.fsb.tcp_ip_hdr_len = sizeof(struct tcpiphdr) + sizeof(struct udphdr); #endif rack->r_ctl.fsb.tcp_ip_hdr = malloc(rack->r_ctl.fsb.tcp_ip_hdr_len, M_TCPFSB, M_NOWAIT|M_ZERO); if (rack->r_ctl.fsb.tcp_ip_hdr == NULL) { return (ENOMEM); } rack->r_fsb_inited = 0; return (0); } static int rack_init(struct tcpcb *tp) { struct tcp_rack *rack = NULL; #ifdef INVARIANTS struct rack_sendmap *insret; #endif uint32_t iwin, snt, us_cts; int err; tp->t_fb_ptr = uma_zalloc(rack_pcb_zone, M_NOWAIT); if (tp->t_fb_ptr == NULL) { /* * We need to allocate memory but cant. The INP and INP_INFO * locks and they are recursive (happens during setup. So a * scheme to drop the locks fails :( * */ return (ENOMEM); } memset(tp->t_fb_ptr, 0, sizeof(struct tcp_rack)); rack = (struct tcp_rack *)tp->t_fb_ptr; RB_INIT(&rack->r_ctl.rc_mtree); TAILQ_INIT(&rack->r_ctl.rc_free); TAILQ_INIT(&rack->r_ctl.rc_tmap); rack->rc_tp = tp; rack->rc_inp = tp->t_inpcb; /* Set the flag */ rack->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0; /* Probably not needed but lets be sure */ rack_clear_rate_sample(rack); /* * Save off the default values, socket options will poke * at these if pacing is not on or we have not yet * reached where pacing is on (gp_ready/fixed enabled). * When they get set into the CC module (when gp_ready * is enabled or we enable fixed) then we will set these * values into the CC and place in here the old values * so we have a restoral. Then we will set the flag * rc_pacing_cc_set. That way whenever we turn off pacing * or switch off this stack, we will know to go restore * the saved values. */ rack->r_ctl.rc_saved_beta.beta = V_newreno_beta_ecn; rack->r_ctl.rc_saved_beta.beta_ecn = V_newreno_beta_ecn; /* We want abe like behavior as well */ rack->r_ctl.rc_saved_beta.newreno_flags |= CC_NEWRENO_BETA_ECN_ENABLED; rack->r_ctl.rc_reorder_fade = rack_reorder_fade; rack->rc_allow_data_af_clo = rack_ignore_data_after_close; rack->r_ctl.rc_tlp_threshold = rack_tlp_thresh; rack->r_ctl.roundends = tp->snd_max; if (use_rack_rr) rack->use_rack_rr = 1; if (V_tcp_delack_enabled) tp->t_delayed_ack = 1; else tp->t_delayed_ack = 0; #ifdef TCP_ACCOUNTING if (rack_tcp_accounting) { tp->t_flags2 |= TF2_TCP_ACCOUNTING; } #endif if (rack_enable_shared_cwnd) rack->rack_enable_scwnd = 1; rack->rc_user_set_max_segs = rack_hptsi_segments; rack->rc_force_max_seg = 0; if (rack_use_imac_dack) rack->rc_dack_mode = 1; TAILQ_INIT(&rack->r_ctl.opt_list); rack->r_ctl.rc_reorder_shift = rack_reorder_thresh; rack->r_ctl.rc_pkt_delay = rack_pkt_delay; rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp; rack->r_ctl.rc_lowest_us_rtt = 0xffffffff; rack->r_ctl.rc_highest_us_rtt = 0; rack->r_ctl.bw_rate_cap = rack_bw_rate_cap; rack->r_ctl.timer_slop = TICKS_2_USEC(tcp_rexmit_slop); if (rack_use_cmp_acks) rack->r_use_cmp_ack = 1; if (rack_disable_prr) rack->rack_no_prr = 1; if (rack_gp_no_rec_chg) rack->rc_gp_no_rec_chg = 1; if (rack_pace_every_seg && tcp_can_enable_pacing()) { rack->rc_always_pace = 1; if (rack->use_fixed_rate || rack->gp_ready) rack_set_cc_pacing(rack); } else rack->rc_always_pace = 0; if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack) rack->r_mbuf_queue = 1; else rack->r_mbuf_queue = 0; if (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack) tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ; else tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ; rack_set_pace_segments(tp, rack, __LINE__, NULL); if (rack_limits_scwnd) rack->r_limit_scw = 1; else rack->r_limit_scw = 0; rack->rc_labc = V_tcp_abc_l_var; rack->r_ctl.rc_high_rwnd = tp->snd_wnd; rack->r_ctl.cwnd_to_use = tp->snd_cwnd; rack->r_ctl.rc_rate_sample_method = rack_rate_sample_method; rack->rack_tlp_threshold_use = rack_tlp_threshold_use; rack->r_ctl.rc_prr_sendalot = rack_send_a_lot_in_prr; rack->r_ctl.rc_min_to = rack_min_to; microuptime(&rack->r_ctl.act_rcv_time); rack->r_ctl.rc_last_time_decay = rack->r_ctl.act_rcv_time; rack->rc_init_win = rack_default_init_window; rack->r_ctl.rack_per_of_gp_ss = rack_per_of_gp_ss; if (rack_hw_up_only) rack->r_up_only = 1; if (rack_do_dyn_mul) { /* When dynamic adjustment is on CA needs to start at 100% */ rack->rc_gp_dyn_mul = 1; if (rack_do_dyn_mul >= 100) rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul; } else rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca; rack->r_ctl.rack_per_of_gp_rec = rack_per_of_gp_rec; rack->r_ctl.rack_per_of_gp_probertt = rack_per_of_gp_probertt; rack->r_ctl.rc_tlp_rxt_last_time = tcp_tv_to_mssectick(&rack->r_ctl.act_rcv_time); setup_time_filter_small(&rack->r_ctl.rc_gp_min_rtt, FILTER_TYPE_MIN, rack_probertt_filter_life); us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time); rack->r_ctl.rc_lower_rtt_us_cts = us_cts; rack->r_ctl.rc_time_of_last_probertt = us_cts; rack->r_ctl.challenge_ack_ts = tcp_ts_getticks(); rack->r_ctl.rc_time_probertt_starts = 0; if (rack_dsack_std_based & 0x1) { /* Basically this means all rack timers are at least (srtt + 1/4 srtt) */ rack->rc_rack_tmr_std_based = 1; } if (rack_dsack_std_based & 0x2) { /* Basically this means rack timers are extended based on dsack by up to (2 * srtt) */ rack->rc_rack_use_dsack = 1; } /* We require at least one measurement, even if the sysctl is 0 */ if (rack_req_measurements) rack->r_ctl.req_measurements = rack_req_measurements; else rack->r_ctl.req_measurements = 1; if (rack_enable_hw_pacing) rack->rack_hdw_pace_ena = 1; if (rack_hw_rate_caps) rack->r_rack_hw_rate_caps = 1; /* Do we force on detection? */ #ifdef NETFLIX_EXP_DETECTION if (tcp_force_detection) rack->do_detection = 1; else #endif rack->do_detection = 0; if (rack_non_rxt_use_cr) rack->rack_rec_nonrxt_use_cr = 1; err = rack_init_fsb(tp, rack); if (err) { uma_zfree(rack_pcb_zone, tp->t_fb_ptr); tp->t_fb_ptr = NULL; return (err); } if (tp->snd_una != tp->snd_max) { /* Create a send map for the current outstanding data */ struct rack_sendmap *rsm; rsm = rack_alloc(rack); if (rsm == NULL) { uma_zfree(rack_pcb_zone, tp->t_fb_ptr); tp->t_fb_ptr = NULL; return (ENOMEM); } rsm->r_no_rtt_allowed = 1; rsm->r_tim_lastsent[0] = rack_to_usec_ts(&rack->r_ctl.act_rcv_time); rsm->r_rtr_cnt = 1; rsm->r_rtr_bytes = 0; if (tp->t_flags & TF_SENTFIN) rsm->r_flags |= RACK_HAS_FIN; if ((tp->snd_una == tp->iss) && !TCPS_HAVEESTABLISHED(tp->t_state)) rsm->r_flags |= RACK_HAS_SYN; rsm->r_start = tp->snd_una; rsm->r_end = tp->snd_max; rsm->r_dupack = 0; if (rack->rc_inp->inp_socket->so_snd.sb_mb != NULL) { rsm->m = sbsndmbuf(&rack->rc_inp->inp_socket->so_snd, 0, &rsm->soff); if (rsm->m) rsm->orig_m_len = rsm->m->m_len; else rsm->orig_m_len = 0; } else { /* * This can happen if we have a stand-alone FIN or * SYN. */ rsm->m = NULL; rsm->orig_m_len = 0; rsm->soff = 0; } #ifndef INVARIANTS (void)RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); #else insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); if (insret != NULL) { panic("Insert in rb tree fails ret:%p rack:%p rsm:%p", insret, rack, rsm); } #endif TAILQ_INSERT_TAIL(&rack->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 1; } /* * Timers in Rack are kept in microseconds so lets * convert any initial incoming variables * from ticks into usecs. Note that we * also change the values of t_srtt and t_rttvar, if * they are non-zero. They are kept with a 5 * bit decimal so we have to carefully convert * these to get the full precision. */ rack_convert_rtts(tp); tp->t_rttlow = TICKS_2_USEC(tp->t_rttlow); if (rack_do_hystart) { tp->ccv->flags |= CCF_HYSTART_ALLOWED; if (rack_do_hystart > 1) tp->ccv->flags |= CCF_HYSTART_CAN_SH_CWND; if (rack_do_hystart > 2) tp->ccv->flags |= CCF_HYSTART_CONS_SSTH; } if (rack_def_profile) rack_set_profile(rack, rack_def_profile); /* Cancel the GP measurement in progress */ tp->t_flags &= ~TF_GPUTINPROG; if (SEQ_GT(tp->snd_max, tp->iss)) snt = tp->snd_max - tp->iss; else snt = 0; iwin = rc_init_window(rack); if (snt < iwin) { /* We are not past the initial window * so we need to make sure cwnd is * correct. */ if (tp->snd_cwnd < iwin) tp->snd_cwnd = iwin; /* * If we are within the initial window * we want ssthresh to be unlimited. Setting * it to the rwnd (which the default stack does * and older racks) is not really a good idea * since we want to be in SS and grow both the * cwnd and the rwnd (via dynamic rwnd growth). If * we set it to the rwnd then as the peer grows its * rwnd we will be stuck in CA and never hit SS. * * Its far better to raise it up high (this takes the * risk that there as been a loss already, probably * we should have an indicator in all stacks of loss * but we don't), but considering the normal use this * is a risk worth taking. The consequences of not * hitting SS are far worse than going one more time * into it early on (before we have sent even a IW). * It is highly unlikely that we will have had a loss * before getting the IW out. */ tp->snd_ssthresh = 0xffffffff; } rack_stop_all_timers(tp); /* Lets setup the fsb block */ rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0); rack_log_rtt_shrinks(rack, us_cts, tp->t_rxtcur, __LINE__, RACK_RTTS_INIT); return (0); } static int rack_handoff_ok(struct tcpcb *tp) { if ((tp->t_state == TCPS_CLOSED) || (tp->t_state == TCPS_LISTEN)) { /* Sure no problem though it may not stick */ return (0); } if ((tp->t_state == TCPS_SYN_SENT) || (tp->t_state == TCPS_SYN_RECEIVED)) { /* * We really don't know if you support sack, * you have to get to ESTAB or beyond to tell. */ return (EAGAIN); } if ((tp->t_flags & TF_SENTFIN) && ((tp->snd_max - tp->snd_una) > 1)) { /* * Rack will only send a FIN after all data is acknowledged. * So in this case we have more data outstanding. We can't * switch stacks until either all data and only the FIN * is left (in which case rack_init() now knows how * to deal with that) all is acknowledged and we * are only left with incoming data, though why you * would want to switch to rack after all data is acknowledged * I have no idea (rrs)! */ return (EAGAIN); } if ((tp->t_flags & TF_SACK_PERMIT) || rack_sack_not_required){ return (0); } /* * If we reach here we don't do SACK on this connection so we can * never do rack. */ return (EINVAL); } static void rack_fini(struct tcpcb *tp, int32_t tcb_is_purged) { if (tp->t_fb_ptr) { struct tcp_rack *rack; struct rack_sendmap *rsm, *nrsm; #ifdef INVARIANTS struct rack_sendmap *rm; #endif rack = (struct tcp_rack *)tp->t_fb_ptr; if (tp->t_in_pkt) { /* * It is unsafe to process the packets since a * reset may be lurking in them (its rare but it * can occur). If we were to find a RST, then we * would end up dropping the connection and the * INP lock, so when we return the caller (tcp_usrreq) * will blow up when it trys to unlock the inp. */ struct mbuf *save, *m; m = tp->t_in_pkt; tp->t_in_pkt = NULL; tp->t_tail_pkt = NULL; while (m) { save = m->m_nextpkt; m->m_nextpkt = NULL; m_freem(m); m = save; } } tp->t_flags &= ~TF_FORCEDATA; #ifdef NETFLIX_SHARED_CWND if (rack->r_ctl.rc_scw) { uint32_t limit; if (rack->r_limit_scw) limit = max(1, rack->r_ctl.rc_lowest_us_rtt); else limit = 0; tcp_shared_cwnd_free_full(tp, rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index, limit); rack->r_ctl.rc_scw = NULL; } #endif if (rack->r_ctl.fsb.tcp_ip_hdr) { free(rack->r_ctl.fsb.tcp_ip_hdr, M_TCPFSB); rack->r_ctl.fsb.tcp_ip_hdr = NULL; rack->r_ctl.fsb.th = NULL; } /* Convert back to ticks, with */ if (tp->t_srtt > 1) { uint32_t val, frac; val = USEC_2_TICKS(tp->t_srtt); frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz); tp->t_srtt = val << TCP_RTT_SHIFT; /* * frac is the fractional part here is left * over from converting to hz and shifting. * We need to convert this to the 5 bit * remainder. */ if (frac) { if (hz == 1000) { frac = (((uint64_t)frac * (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC); } else { frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC); } tp->t_srtt += frac; } } if (tp->t_rttvar) { uint32_t val, frac; val = USEC_2_TICKS(tp->t_rttvar); frac = tp->t_srtt % (HPTS_USEC_IN_SEC / hz); tp->t_rttvar = val << TCP_RTTVAR_SHIFT; /* * frac is the fractional part here is left * over from converting to hz and shifting. * We need to convert this to the 5 bit * remainder. */ if (frac) { if (hz == 1000) { frac = (((uint64_t)frac * (uint64_t)TCP_RTT_SCALE) / (uint64_t)HPTS_USEC_IN_MSEC); } else { frac = (((uint64_t)frac * (uint64_t)(hz) * (uint64_t)TCP_RTT_SCALE) /(uint64_t)HPTS_USEC_IN_SEC); } tp->t_rttvar += frac; } } tp->t_rxtcur = USEC_2_TICKS(tp->t_rxtcur); tp->t_rttlow = USEC_2_TICKS(tp->t_rttlow); if (rack->rc_always_pace) { tcp_decrement_paced_conn(); rack_undo_cc_pacing(rack); rack->rc_always_pace = 0; } /* Clean up any options if they were not applied */ while (!TAILQ_EMPTY(&rack->r_ctl.opt_list)) { struct deferred_opt_list *dol; dol = TAILQ_FIRST(&rack->r_ctl.opt_list); TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next); free(dol, M_TCPDO); } /* rack does not use force data but other stacks may clear it */ if (rack->r_ctl.crte != NULL) { tcp_rel_pacing_rate(rack->r_ctl.crte, tp); rack->rack_hdrw_pacing = 0; rack->r_ctl.crte = NULL; } #ifdef TCP_BLACKBOX tcp_log_flowend(tp); #endif RB_FOREACH_SAFE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm) { #ifndef INVARIANTS (void)RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); #else rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); if (rm != rsm) { panic("At fini, rack:%p rsm:%p rm:%p", rack, rsm, rm); } #endif uma_zfree(rack_zone, rsm); } rsm = TAILQ_FIRST(&rack->r_ctl.rc_free); while (rsm) { TAILQ_REMOVE(&rack->r_ctl.rc_free, rsm, r_tnext); uma_zfree(rack_zone, rsm); rsm = TAILQ_FIRST(&rack->r_ctl.rc_free); } rack->rc_free_cnt = 0; uma_zfree(rack_pcb_zone, tp->t_fb_ptr); tp->t_fb_ptr = NULL; } if (tp->t_inpcb) { tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ; tp->t_inpcb->inp_flags2 &= ~INP_MBUF_QUEUE_READY; tp->t_inpcb->inp_flags2 &= ~INP_DONT_SACK_QUEUE; tp->t_inpcb->inp_flags2 &= ~INP_MBUF_ACKCMP; /* Cancel the GP measurement in progress */ tp->t_flags &= ~TF_GPUTINPROG; tp->t_inpcb->inp_flags2 &= ~INP_MBUF_L_ACKS; } /* Make sure snd_nxt is correctly set */ tp->snd_nxt = tp->snd_max; } static void rack_set_state(struct tcpcb *tp, struct tcp_rack *rack) { if ((rack->r_state == TCPS_CLOSED) && (tp->t_state != TCPS_CLOSED)) { rack->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0; } switch (tp->t_state) { case TCPS_SYN_SENT: rack->r_state = TCPS_SYN_SENT; rack->r_substate = rack_do_syn_sent; break; case TCPS_SYN_RECEIVED: rack->r_state = TCPS_SYN_RECEIVED; rack->r_substate = rack_do_syn_recv; break; case TCPS_ESTABLISHED: rack_set_pace_segments(tp, rack, __LINE__, NULL); rack->r_state = TCPS_ESTABLISHED; rack->r_substate = rack_do_established; break; case TCPS_CLOSE_WAIT: rack_set_pace_segments(tp, rack, __LINE__, NULL); rack->r_state = TCPS_CLOSE_WAIT; rack->r_substate = rack_do_close_wait; break; case TCPS_FIN_WAIT_1: rack_set_pace_segments(tp, rack, __LINE__, NULL); rack->r_state = TCPS_FIN_WAIT_1; rack->r_substate = rack_do_fin_wait_1; break; case TCPS_CLOSING: rack_set_pace_segments(tp, rack, __LINE__, NULL); rack->r_state = TCPS_CLOSING; rack->r_substate = rack_do_closing; break; case TCPS_LAST_ACK: rack_set_pace_segments(tp, rack, __LINE__, NULL); rack->r_state = TCPS_LAST_ACK; rack->r_substate = rack_do_lastack; break; case TCPS_FIN_WAIT_2: rack_set_pace_segments(tp, rack, __LINE__, NULL); rack->r_state = TCPS_FIN_WAIT_2; rack->r_substate = rack_do_fin_wait_2; break; case TCPS_LISTEN: case TCPS_CLOSED: case TCPS_TIME_WAIT: default: break; }; if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state)) rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP; } static void rack_timer_audit(struct tcpcb *tp, struct tcp_rack *rack, struct sockbuf *sb) { /* * We received an ack, and then did not * call send or were bounced out due to the * hpts was running. Now a timer is up as well, is * it the right timer? */ struct rack_sendmap *rsm; int tmr_up; tmr_up = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK; if (rack->rc_in_persist && (tmr_up == PACE_TMR_PERSIT)) return; rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); if (((rsm == NULL) || (tp->t_state < TCPS_ESTABLISHED)) && (tmr_up == PACE_TMR_RXT)) { /* Should be an RXT */ return; } if (rsm == NULL) { /* Nothing outstanding? */ if (tp->t_flags & TF_DELACK) { if (tmr_up == PACE_TMR_DELACK) /* We are supposed to have delayed ack up and we do */ return; } else if (sbavail(&tp->t_inpcb->inp_socket->so_snd) && (tmr_up == PACE_TMR_RXT)) { /* * if we hit enobufs then we would expect the possibility * of nothing outstanding and the RXT up (and the hptsi timer). */ return; } else if (((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) && (tp->t_state <= TCPS_CLOSING)) && (tmr_up == PACE_TMR_KEEP) && (tp->snd_max == tp->snd_una)) { /* We should have keep alive up and we do */ return; } } if (SEQ_GT(tp->snd_max, tp->snd_una) && ((tmr_up == PACE_TMR_TLP) || (tmr_up == PACE_TMR_RACK) || (tmr_up == PACE_TMR_RXT))) { /* * Either a Rack, TLP or RXT is fine if we * have outstanding data. */ return; } else if (tmr_up == PACE_TMR_DELACK) { /* * If the delayed ack was going to go off * before the rtx/tlp/rack timer were going to * expire, then that would be the timer in control. * Note we don't check the time here trusting the * code is correct. */ return; } /* * Ok the timer originally started is not what we want now. * We will force the hpts to be stopped if any, and restart * with the slot set to what was in the saved slot. */ if (tcp_in_hpts(rack->rc_inp)) { if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) { uint32_t us_cts; us_cts = tcp_get_usecs(NULL); if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) { rack->r_early = 1; rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts); } rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT; } tcp_hpts_remove(tp->t_inpcb); } rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0); } static void rack_do_win_updates(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tiwin, uint32_t seq, uint32_t ack, uint32_t cts, uint32_t high_seq) { if ((SEQ_LT(tp->snd_wl1, seq) || (tp->snd_wl1 == seq && (SEQ_LT(tp->snd_wl2, ack) || (tp->snd_wl2 == ack && tiwin > tp->snd_wnd))))) { /* keep track of pure window updates */ if ((tp->snd_wl2 == ack) && (tiwin > tp->snd_wnd)) KMOD_TCPSTAT_INC(tcps_rcvwinupd); tp->snd_wnd = tiwin; rack_validate_fo_sendwin_up(tp, rack); tp->snd_wl1 = seq; tp->snd_wl2 = ack; if (tp->snd_wnd > tp->max_sndwnd) tp->max_sndwnd = tp->snd_wnd; rack->r_wanted_output = 1; } else if ((tp->snd_wl2 == ack) && (tiwin < tp->snd_wnd)) { tp->snd_wnd = tiwin; rack_validate_fo_sendwin_up(tp, rack); tp->snd_wl1 = seq; tp->snd_wl2 = ack; } else { /* Not a valid win update */ return; } /* Do we exit persists? */ if ((rack->rc_in_persist != 0) && (tp->snd_wnd >= min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs))) { rack_exit_persist(tp, rack, cts); } /* Do we enter persists? */ if ((rack->rc_in_persist == 0) && (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) && TCPS_HAVEESTABLISHED(tp->t_state) && ((tp->snd_max == tp->snd_una) || rack->rc_has_collapsed) && sbavail(&tp->t_inpcb->inp_socket->so_snd) && (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd)) { /* * Here the rwnd is less than * the pacing size, we are established, * nothing is outstanding, and there is * data to send. Enter persists. */ rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime); } } static void rack_log_input_packet(struct tcpcb *tp, struct tcp_rack *rack, struct tcp_ackent *ae, int ackval, uint32_t high_seq) { if (tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval ltv; char tcp_hdr_buf[60]; struct tcphdr *th; struct timespec ts; uint32_t orig_snd_una; uint8_t xx = 0; #ifdef NETFLIX_HTTP_LOGGING struct http_sendfile_track *http_req; if (SEQ_GT(ae->ack, tp->snd_una)) { http_req = tcp_http_find_req_for_seq(tp, (ae->ack-1)); } else { http_req = tcp_http_find_req_for_seq(tp, ae->ack); } #endif memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp); if (rack->rack_no_prr == 0) log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt; else log.u_bbr.flex1 = 0; log.u_bbr.use_lt_bw = rack->r_ent_rec_ns; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->r_might_revert; log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced; log.u_bbr.inflight = ctf_flight_size(tp, rack->r_ctl.rc_sacked); log.u_bbr.pkts_out = tp->t_maxseg; log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags; log.u_bbr.flex7 = 1; log.u_bbr.lost = ae->flags; log.u_bbr.cwnd_gain = ackval; log.u_bbr.pacing_gain = 0x2; if (ae->flags & TSTMP_HDWR) { /* Record the hardware timestamp if present */ log.u_bbr.flex3 = M_TSTMP; ts.tv_sec = ae->timestamp / 1000000000; ts.tv_nsec = ae->timestamp % 1000000000; ltv.tv_sec = ts.tv_sec; ltv.tv_usec = ts.tv_nsec / 1000; log.u_bbr.lt_epoch = tcp_tv_to_usectick(<v); } else if (ae->flags & TSTMP_LRO) { /* Record the LRO the arrival timestamp */ log.u_bbr.flex3 = M_TSTMP_LRO; ts.tv_sec = ae->timestamp / 1000000000; ts.tv_nsec = ae->timestamp % 1000000000; ltv.tv_sec = ts.tv_sec; ltv.tv_usec = ts.tv_nsec / 1000; log.u_bbr.flex5 = tcp_tv_to_usectick(<v); } log.u_bbr.timeStamp = tcp_get_usecs(<v); /* Log the rcv time */ log.u_bbr.delRate = ae->timestamp; #ifdef NETFLIX_HTTP_LOGGING log.u_bbr.applimited = tp->t_http_closed; log.u_bbr.applimited <<= 8; log.u_bbr.applimited |= tp->t_http_open; log.u_bbr.applimited <<= 8; log.u_bbr.applimited |= tp->t_http_req; if (http_req) { /* Copy out any client req info */ /* seconds */ log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC); /* useconds */ log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC); log.u_bbr.rttProp = http_req->timestamp; log.u_bbr.cur_del_rate = http_req->start; if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) { log.u_bbr.flex8 |= 1; } else { log.u_bbr.flex8 |= 2; log.u_bbr.bw_inuse = http_req->end; } log.u_bbr.flex6 = http_req->start_seq; if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) { log.u_bbr.flex8 |= 4; log.u_bbr.epoch = http_req->end_seq; } } #endif memset(tcp_hdr_buf, 0, sizeof(tcp_hdr_buf)); th = (struct tcphdr *)tcp_hdr_buf; th->th_seq = ae->seq; th->th_ack = ae->ack; th->th_win = ae->win; /* Now fill in the ports */ th->th_sport = tp->t_inpcb->inp_fport; th->th_dport = tp->t_inpcb->inp_lport; tcp_set_flags(th, ae->flags); /* Now do we have a timestamp option? */ if (ae->flags & HAS_TSTMP) { u_char *cp; uint32_t val; th->th_off = ((sizeof(struct tcphdr) + TCPOLEN_TSTAMP_APPA) >> 2); cp = (u_char *)(th + 1); *cp = TCPOPT_NOP; cp++; *cp = TCPOPT_NOP; cp++; *cp = TCPOPT_TIMESTAMP; cp++; *cp = TCPOLEN_TIMESTAMP; cp++; val = htonl(ae->ts_value); bcopy((char *)&val, (char *)cp, sizeof(uint32_t)); val = htonl(ae->ts_echo); bcopy((char *)&val, (char *)(cp + 4), sizeof(uint32_t)); } else th->th_off = (sizeof(struct tcphdr) >> 2); /* * For sane logging we need to play a little trick. * If the ack were fully processed we would have moved * snd_una to high_seq, but since compressed acks are * processed in two phases, at this point (logging) snd_una * won't be advanced. So we would see multiple acks showing * the advancement. We can prevent that by "pretending" that * snd_una was advanced and then un-advancing it so that the * logging code has the right value for tlb_snd_una. */ if (tp->snd_una != high_seq) { orig_snd_una = tp->snd_una; tp->snd_una = high_seq; xx = 1; } else xx = 0; TCP_LOG_EVENTP(tp, th, &tp->t_inpcb->inp_socket->so_rcv, &tp->t_inpcb->inp_socket->so_snd, TCP_LOG_IN, 0, 0, &log, true, <v); if (xx) { tp->snd_una = orig_snd_una; } } } static void rack_handle_probe_response(struct tcp_rack *rack, uint32_t tiwin, uint32_t us_cts) { uint32_t us_rtt; /* * A persist or keep-alive was forced out, update our * min rtt time. Note now worry about lost responses. * When a subsequent keep-alive or persist times out * and forced_ack is still on, then the last probe * was not responded to. In such cases we have a * sysctl that controls the behavior. Either we apply * the rtt but with reduced confidence (0). Or we just * plain don't apply the rtt estimate. Having data flow * will clear the probe_not_answered flag i.e. cum-ack * move forward exiting and reentering persists. */ rack->forced_ack = 0; rack->rc_tp->t_rxtshift = 0; if ((rack->rc_in_persist && (tiwin == rack->rc_tp->snd_wnd)) || (rack->rc_in_persist == 0)) { /* * In persists only apply the RTT update if this is * a response to our window probe. And that * means the rwnd sent must match the current * snd_wnd. If it does not, then we got a * window update ack instead. For keepalive * we allow the answer no matter what the window. * * Note that if the probe_not_answered is set then * the forced_ack_ts is the oldest one i.e. the first * probe sent that might have been lost. This assures * us that if we do calculate an RTT it is longer not * some short thing. */ if (rack->rc_in_persist) counter_u64_add(rack_persists_acks, 1); us_rtt = us_cts - rack->r_ctl.forced_ack_ts; if (us_rtt == 0) us_rtt = 1; if (rack->probe_not_answered == 0) { rack_apply_updated_usrtt(rack, us_rtt, us_cts); tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 3, NULL, 1); } else { /* We have a retransmitted probe here too */ if (rack_apply_rtt_with_reduced_conf) { rack_apply_updated_usrtt(rack, us_rtt, us_cts); tcp_rack_xmit_timer(rack, us_rtt, 0, us_rtt, 0, NULL, 1); } } } } static int rack_do_compressed_ack_processing(struct tcpcb *tp, struct socket *so, struct mbuf *m, int nxt_pkt, struct timeval *tv) { /* * Handle a "special" compressed ack mbuf. Each incoming * ack has only four possible dispositions: * * A) It moves the cum-ack forward * B) It is behind the cum-ack. * C) It is a window-update ack. * D) It is a dup-ack. * * Note that we can have between 1 -> TCP_COMP_ACK_ENTRIES * in the incoming mbuf. We also need to still pay attention * to nxt_pkt since there may be another packet after this * one. */ #ifdef TCP_ACCOUNTING uint64_t ts_val; uint64_t rdstc; #endif int segsiz; struct timespec ts; struct tcp_rack *rack; struct tcp_ackent *ae; uint32_t tiwin, ms_cts, cts, acked, acked_amount, high_seq, win_seq, the_win, win_upd_ack; int cnt, i, did_out, ourfinisacked = 0; struct tcpopt to_holder, *to = NULL; #ifdef TCP_ACCOUNTING int win_up_req = 0; #endif int nsegs = 0; int under_pacing = 1; int recovery = 0; #ifdef TCP_ACCOUNTING sched_pin(); #endif rack = (struct tcp_rack *)tp->t_fb_ptr; if (rack->gp_ready && (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) under_pacing = 0; else under_pacing = 1; if (rack->r_state != tp->t_state) rack_set_state(tp, rack); if ((tp->t_state >= TCPS_FIN_WAIT_1) && (tp->t_flags & TF_GPUTINPROG)) { /* * We have a goodput in progress * and we have entered a late state. * Do we have enough data in the sb * to handle the GPUT request? */ uint32_t bytes; bytes = tp->gput_ack - tp->gput_seq; if (SEQ_GT(tp->gput_seq, tp->snd_una)) bytes += tp->gput_seq - tp->snd_una; if (bytes > sbavail(&tp->t_inpcb->inp_socket->so_snd)) { /* * There are not enough bytes in the socket * buffer that have been sent to cover this * measurement. Cancel it. */ rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/, rack->r_ctl.rc_gp_srtt /*flex1*/, tp->gput_seq, 0, 0, 18, __LINE__, NULL, 0); tp->t_flags &= ~TF_GPUTINPROG; } } to = &to_holder; to->to_flags = 0; KASSERT((m->m_len >= sizeof(struct tcp_ackent)), ("tp:%p m_cmpack:%p with invalid len:%u", tp, m, m->m_len)); cnt = m->m_len / sizeof(struct tcp_ackent); counter_u64_add(rack_multi_single_eq, cnt); high_seq = tp->snd_una; the_win = tp->snd_wnd; win_seq = tp->snd_wl1; win_upd_ack = tp->snd_wl2; cts = tcp_tv_to_usectick(tv); ms_cts = tcp_tv_to_mssectick(tv); rack->r_ctl.rc_rcvtime = cts; segsiz = ctf_fixed_maxseg(tp); if ((rack->rc_gp_dyn_mul) && (rack->use_fixed_rate == 0) && (rack->rc_always_pace)) { /* Check in on probertt */ rack_check_probe_rtt(rack, cts); } for (i = 0; i < cnt; i++) { #ifdef TCP_ACCOUNTING ts_val = get_cyclecount(); #endif rack_clear_rate_sample(rack); ae = ((mtod(m, struct tcp_ackent *)) + i); /* Setup the window */ tiwin = ae->win << tp->snd_scale; if (tiwin > rack->r_ctl.rc_high_rwnd) rack->r_ctl.rc_high_rwnd = tiwin; /* figure out the type of ack */ if (SEQ_LT(ae->ack, high_seq)) { /* Case B*/ ae->ack_val_set = ACK_BEHIND; } else if (SEQ_GT(ae->ack, high_seq)) { /* Case A */ ae->ack_val_set = ACK_CUMACK; } else if ((tiwin == the_win) && (rack->rc_in_persist == 0)){ /* Case D */ ae->ack_val_set = ACK_DUPACK; } else { /* Case C */ ae->ack_val_set = ACK_RWND; } rack_log_input_packet(tp, rack, ae, ae->ack_val_set, high_seq); /* Validate timestamp */ if (ae->flags & HAS_TSTMP) { /* Setup for a timestamp */ to->to_flags = TOF_TS; ae->ts_echo -= tp->ts_offset; to->to_tsecr = ae->ts_echo; to->to_tsval = ae->ts_value; /* * If echoed timestamp is later than the current time, fall back to * non RFC1323 RTT calculation. Normalize timestamp if syncookies * were used when this connection was established. */ if (TSTMP_GT(ae->ts_echo, ms_cts)) to->to_tsecr = 0; if (tp->ts_recent && TSTMP_LT(ae->ts_value, tp->ts_recent)) { if (ctf_ts_check_ac(tp, (ae->flags & 0xff))) { #ifdef TCP_ACCOUNTING rdstc = get_cyclecount(); if (rdstc > ts_val) { counter_u64_add(tcp_proc_time[ae->ack_val_set] , (rdstc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val); } } #endif continue; } } if (SEQ_LEQ(ae->seq, tp->last_ack_sent) && SEQ_LEQ(tp->last_ack_sent, ae->seq)) { tp->ts_recent_age = tcp_ts_getticks(); tp->ts_recent = ae->ts_value; } } else { /* Setup for a no options */ to->to_flags = 0; } /* Update the rcv time and perform idle reduction possibly */ if (tp->t_idle_reduce && (tp->snd_max == tp->snd_una) && (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) { counter_u64_add(rack_input_idle_reduces, 1); rack_cc_after_idle(rack, tp); } tp->t_rcvtime = ticks; /* Now what about ECN? */ if (tcp_ecn_input_segment(tp, ae->flags, ae->codepoint)) rack_cong_signal(tp, CC_ECN, ae->ack, __LINE__); #ifdef TCP_ACCOUNTING /* Count for the specific type of ack in */ counter_u64_add(tcp_cnt_counters[ae->ack_val_set], 1); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[ae->ack_val_set]++; } #endif /* * Note how we could move up these in the determination * above, but we don't so that way the timestamp checks (and ECN) * is done first before we do any processing on the ACK. * The non-compressed path through the code has this * weakness (noted by @jtl) that it actually does some * processing before verifying the timestamp information. * We don't take that path here which is why we set * the ack_val_set first, do the timestamp and ecn * processing, and then look at what we have setup. */ if (ae->ack_val_set == ACK_BEHIND) { /* * Case B flag reordering, if window is not closed * or it could be a keep-alive or persists */ if (SEQ_LT(ae->ack, tp->snd_una) && (sbspace(&so->so_rcv) > segsiz)) { rack->r_ctl.rc_reorder_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time); } } else if (ae->ack_val_set == ACK_DUPACK) { /* Case D */ rack_strike_dupack(rack); } else if (ae->ack_val_set == ACK_RWND) { /* Case C */ if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) { ts.tv_sec = ae->timestamp / 1000000000; ts.tv_nsec = ae->timestamp % 1000000000; rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec; rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000; } else { rack->r_ctl.act_rcv_time = *tv; } if (rack->forced_ack) { rack_handle_probe_response(rack, tiwin, tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time)); } #ifdef TCP_ACCOUNTING win_up_req = 1; #endif win_upd_ack = ae->ack; win_seq = ae->seq; the_win = tiwin; rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts, high_seq); } else { /* Case A */ if (SEQ_GT(ae->ack, tp->snd_max)) { /* * We just send an ack since the incoming * ack is beyond the largest seq we sent. */ if ((tp->t_flags & TF_ACKNOW) == 0) { ctf_ack_war_checks(tp, &rack->r_ctl.challenge_ack_ts, &rack->r_ctl.challenge_ack_cnt); if (tp->t_flags && TF_ACKNOW) rack->r_wanted_output = 1; } } else { nsegs++; /* If the window changed setup to update */ if (tiwin != tp->snd_wnd) { win_upd_ack = ae->ack; win_seq = ae->seq; the_win = tiwin; rack_do_win_updates(tp, rack, the_win, win_seq, win_upd_ack, cts, high_seq); } #ifdef TCP_ACCOUNTING /* Account for the acks */ if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[CNT_OF_ACKS_IN] += (((ae->ack - high_seq) + segsiz - 1) / segsiz); } counter_u64_add(tcp_cnt_counters[CNT_OF_ACKS_IN], (((ae->ack - high_seq) + segsiz - 1) / segsiz)); #endif high_seq = ae->ack; if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.flex1 = high_seq; log.u_bbr.flex2 = rack->r_ctl.roundends; log.u_bbr.flex3 = rack->r_ctl.current_round; log.u_bbr.rttProp = (uint64_t)CC_ALGO(tp)->newround; log.u_bbr.flex8 = 8; tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0, 0, &log, false, NULL, NULL, 0, &tv); } /* * The draft (v3) calls for us to use SEQ_GEQ, but that * causes issues when we are just going app limited. Lets * instead use SEQ_GT where its equal but more data * is outstanding. */ if ((SEQ_GT(high_seq, rack->r_ctl.roundends)) || ((high_seq == rack->r_ctl.roundends) && SEQ_GT(tp->snd_max, tp->snd_una))) { rack->r_ctl.current_round++; rack->r_ctl.roundends = tp->snd_max; if (CC_ALGO(tp)->newround != NULL) { CC_ALGO(tp)->newround(tp->ccv, rack->r_ctl.current_round); } } /* Setup our act_rcv_time */ if ((ae->flags & TSTMP_LRO) || (ae->flags & TSTMP_HDWR)) { ts.tv_sec = ae->timestamp / 1000000000; ts.tv_nsec = ae->timestamp % 1000000000; rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec; rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000; } else { rack->r_ctl.act_rcv_time = *tv; } rack_process_to_cumack(tp, rack, ae->ack, cts, to); if (rack->rc_dsack_round_seen) { /* Is the dsack round over? */ if (SEQ_GEQ(ae->ack, rack->r_ctl.dsack_round_end)) { /* Yes it is */ rack->rc_dsack_round_seen = 0; rack_log_dsack_event(rack, 3, __LINE__, 0, 0); } } } } /* And lets be sure to commit the rtt measurements for this ack */ tcp_rack_xmit_timer_commit(rack, tp); #ifdef TCP_ACCOUNTING rdstc = get_cyclecount(); if (rdstc > ts_val) { counter_u64_add(tcp_proc_time[ae->ack_val_set] , (rdstc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[ae->ack_val_set] += (rdstc - ts_val); if (ae->ack_val_set == ACK_CUMACK) tp->tcp_proc_time[CYC_HANDLE_MAP] += (rdstc - ts_val); } } #endif } #ifdef TCP_ACCOUNTING ts_val = get_cyclecount(); #endif /* Tend to any collapsed window */ if (SEQ_GT(tp->snd_max, high_seq) && (tp->snd_wnd < (tp->snd_max - high_seq))) { /* The peer collapsed the window */ rack_collapsed_window(rack, (tp->snd_max - high_seq), __LINE__); } else if (rack->rc_has_collapsed) rack_un_collapse_window(rack, __LINE__); if ((rack->r_collapse_point_valid) && (SEQ_GT(high_seq, rack->r_ctl.high_collapse_point))) rack->r_collapse_point_valid = 0; acked_amount = acked = (high_seq - tp->snd_una); if (acked) { /* * Clear the probe not answered flag * since cum-ack moved forward. */ rack->probe_not_answered = 0; if (rack->sack_attack_disable == 0) rack_do_decay(rack); if (acked >= segsiz) { /* * You only get credit for * MSS and greater (and you get extra * credit for larger cum-ack moves). */ int ac; ac = acked / segsiz; rack->r_ctl.ack_count += ac; counter_u64_add(rack_ack_total, ac); } if (rack->r_ctl.ack_count > 0xfff00000) { /* * reduce the number to keep us under * a uint32_t. */ rack->r_ctl.ack_count /= 2; rack->r_ctl.sack_count /= 2; } if (tp->t_flags & TF_NEEDSYN) { /* * T/TCP: Connection was half-synchronized, and our SYN has * been ACK'd (so connection is now fully synchronized). Go * to non-starred state, increment snd_una for ACK of SYN, * and check if we can do window scaling. */ tp->t_flags &= ~TF_NEEDSYN; tp->snd_una++; acked_amount = acked = (high_seq - tp->snd_una); } if (acked > sbavail(&so->so_snd)) acked_amount = sbavail(&so->so_snd); #ifdef NETFLIX_EXP_DETECTION /* * We only care on a cum-ack move if we are in a sack-disabled * state. We have already added in to the ack_count, and we never * would disable on a cum-ack move, so we only care to do the * detection if it may "undo" it, i.e. we were in disabled already. */ if (rack->sack_attack_disable) rack_do_detection(tp, rack, acked_amount, segsiz); #endif if (IN_FASTRECOVERY(tp->t_flags) && (rack->rack_no_prr == 0)) rack_update_prr(tp, rack, acked_amount, high_seq); if (IN_RECOVERY(tp->t_flags)) { if (SEQ_LT(high_seq, tp->snd_recover) && (SEQ_LT(high_seq, tp->snd_max))) { tcp_rack_partialack(tp); } else { rack_post_recovery(tp, high_seq); recovery = 1; } } /* Handle the rack-log-ack part (sendmap) */ if ((sbused(&so->so_snd) == 0) && (acked > acked_amount) && (tp->t_state >= TCPS_FIN_WAIT_1) && (tp->t_flags & TF_SENTFIN)) { /* * We must be sure our fin * was sent and acked (we can be * in FIN_WAIT_1 without having * sent the fin). */ ourfinisacked = 1; /* * Lets make sure snd_una is updated * since most likely acked_amount = 0 (it * should be). */ tp->snd_una = high_seq; } /* Did we make a RTO error? */ if ((tp->t_flags & TF_PREVVALID) && ((tp->t_flags & TF_RCVD_TSTMP) == 0)) { tp->t_flags &= ~TF_PREVVALID; if (tp->t_rxtshift == 1 && (int)(ticks - tp->t_badrxtwin) < 0) rack_cong_signal(tp, CC_RTO_ERR, high_seq, __LINE__); } /* Handle the data in the socket buffer */ KMOD_TCPSTAT_ADD(tcps_rcvackpack, 1); KMOD_TCPSTAT_ADD(tcps_rcvackbyte, acked); if (acked_amount > 0) { struct mbuf *mfree; rack_ack_received(tp, rack, high_seq, nsegs, CC_ACK, recovery); SOCKBUF_LOCK(&so->so_snd); mfree = sbcut_locked(&so->so_snd, acked_amount); tp->snd_una = high_seq; /* Note we want to hold the sb lock through the sendmap adjust */ rack_adjust_sendmap(rack, &so->so_snd, tp->snd_una); /* Wake up the socket if we have room to write more */ rack_log_wakeup(tp,rack, &so->so_snd, acked, 2); sowwakeup_locked(so); m_freem(mfree); } /* update progress */ tp->t_acktime = ticks; rack_log_progress_event(rack, tp, tp->t_acktime, PROGRESS_UPDATE, __LINE__); /* Clear out shifts and such */ tp->t_rxtshift = 0; RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp), rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop); rack->rc_tlp_in_progress = 0; rack->r_ctl.rc_tlp_cnt_out = 0; /* Send recover and snd_nxt must be dragged along */ if (SEQ_GT(tp->snd_una, tp->snd_recover)) tp->snd_recover = tp->snd_una; if (SEQ_LT(tp->snd_nxt, tp->snd_una)) tp->snd_nxt = tp->snd_una; /* * If the RXT timer is running we want to * stop it, so we can restart a TLP (or new RXT). */ if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RXT) rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); #ifdef NETFLIX_HTTP_LOGGING tcp_http_check_for_comp(rack->rc_tp, high_seq); #endif tp->snd_wl2 = high_seq; tp->t_dupacks = 0; if (under_pacing && (rack->use_fixed_rate == 0) && (rack->in_probe_rtt == 0) && rack->rc_gp_dyn_mul && rack->rc_always_pace) { /* Check if we are dragging bottom */ rack_check_bottom_drag(tp, rack, so, acked); } if (tp->snd_una == tp->snd_max) { tp->t_flags &= ~TF_PREVVALID; rack->r_ctl.retran_during_recovery = 0; rack->r_ctl.dsack_byte_cnt = 0; rack->r_ctl.rc_went_idle_time = tcp_get_usecs(NULL); if (rack->r_ctl.rc_went_idle_time == 0) rack->r_ctl.rc_went_idle_time = 1; rack_log_progress_event(rack, tp, 0, PROGRESS_CLEAR, __LINE__); if (sbavail(&tp->t_inpcb->inp_socket->so_snd) == 0) tp->t_acktime = 0; /* Set so we might enter persists... */ rack->r_wanted_output = 1; rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una); if ((tp->t_state >= TCPS_FIN_WAIT_1) && (sbavail(&so->so_snd) == 0) && (tp->t_flags2 & TF2_DROP_AF_DATA)) { /* * The socket was gone and the * peer sent data (not now in the past), time to * reset him. */ rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); /* tcp_close will kill the inp pre-log the Reset */ tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); #ifdef TCP_ACCOUNTING rdstc = get_cyclecount(); if (rdstc > ts_val) { counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val); tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val); } } #endif m_freem(m); tp = tcp_close(tp); if (tp == NULL) { #ifdef TCP_ACCOUNTING sched_unpin(); #endif return (1); } /* * We would normally do drop-with-reset which would * send back a reset. We can't since we don't have * all the needed bits. Instead lets arrange for * a call to tcp_output(). That way since we * are in the closed state we will generate a reset. * * Note if tcp_accounting is on we don't unpin since * we do that after the goto label. */ goto send_out_a_rst; } if ((sbused(&so->so_snd) == 0) && (tp->t_state >= TCPS_FIN_WAIT_1) && (tp->t_flags & TF_SENTFIN)) { /* * If we can't receive any more data, then closing user can * proceed. Starting the timer is contrary to the * specification, but if we don't get a FIN we'll hang * forever. * */ if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { soisdisconnected(so); tcp_timer_activate(tp, TT_2MSL, (tcp_fast_finwait2_recycle ? tcp_finwait2_timeout : TP_MAXIDLE(tp))); } if (ourfinisacked == 0) { /* * We don't change to fin-wait-2 if we have our fin acked * which means we are probably in TCPS_CLOSING. */ tcp_state_change(tp, TCPS_FIN_WAIT_2); } } } /* Wake up the socket if we have room to write more */ if (sbavail(&so->so_snd)) { rack->r_wanted_output = 1; if (ctf_progress_timeout_check(tp, true)) { rack_log_progress_event((struct tcp_rack *)tp->t_fb_ptr, tp, tick, PROGRESS_DROP, __LINE__); /* * We cheat here and don't send a RST, we should send one * when the pacer drops the connection. */ #ifdef TCP_ACCOUNTING rdstc = get_cyclecount(); if (rdstc > ts_val) { counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val); tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val); } } sched_unpin(); #endif (void)tcp_drop(tp, ETIMEDOUT); m_freem(m); return (1); } } if (ourfinisacked) { switch(tp->t_state) { case TCPS_CLOSING: #ifdef TCP_ACCOUNTING rdstc = get_cyclecount(); if (rdstc > ts_val) { counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val); tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val); } } sched_unpin(); #endif tcp_twstart(tp); m_freem(m); return (1); break; case TCPS_LAST_ACK: #ifdef TCP_ACCOUNTING rdstc = get_cyclecount(); if (rdstc > ts_val) { counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val); tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val); } } sched_unpin(); #endif tp = tcp_close(tp); ctf_do_drop(m, tp); return (1); break; case TCPS_FIN_WAIT_1: #ifdef TCP_ACCOUNTING rdstc = get_cyclecount(); if (rdstc > ts_val) { counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val); tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val); } } #endif if (so->so_rcv.sb_state & SBS_CANTRCVMORE) { soisdisconnected(so); tcp_timer_activate(tp, TT_2MSL, (tcp_fast_finwait2_recycle ? tcp_finwait2_timeout : TP_MAXIDLE(tp))); } tcp_state_change(tp, TCPS_FIN_WAIT_2); break; default: break; } } if (rack->r_fast_output) { /* * We re doing fast output.. can we expand that? */ rack_gain_for_fastoutput(rack, tp, so, acked_amount); } #ifdef TCP_ACCOUNTING rdstc = get_cyclecount(); if (rdstc > ts_val) { counter_u64_add(tcp_proc_time[ACK_CUMACK] , (rdstc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[ACK_CUMACK] += (rdstc - ts_val); tp->tcp_proc_time[CYC_HANDLE_ACK] += (rdstc - ts_val); } } } else if (win_up_req) { rdstc = get_cyclecount(); if (rdstc > ts_val) { counter_u64_add(tcp_proc_time[ACK_RWND] , (rdstc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[ACK_RWND] += (rdstc - ts_val); } } #endif } /* Now is there a next packet, if so we are done */ m_freem(m); did_out = 0; if (nxt_pkt) { #ifdef TCP_ACCOUNTING sched_unpin(); #endif rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 5, nsegs); return (0); } rack_handle_might_revert(tp, rack); ctf_calc_rwin(so, tp); if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) { send_out_a_rst: if (tcp_output(tp) < 0) { #ifdef TCP_ACCOUNTING sched_unpin(); #endif return (1); } did_out = 1; } rack_free_trim(rack); #ifdef TCP_ACCOUNTING sched_unpin(); #endif rack_timer_audit(tp, rack, &so->so_snd); rack_log_doseg_done(rack, cts, nxt_pkt, did_out, 6, nsegs); return (0); } static int rack_do_segment_nounlock(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos, int32_t nxt_pkt, struct timeval *tv) { #ifdef TCP_ACCOUNTING uint64_t ts_val; #endif int32_t thflags, retval, did_out = 0; int32_t way_out = 0; /* * cts - is the current time from tv (caller gets ts) in microseconds. * ms_cts - is the current time from tv in milliseconds. * us_cts - is the time that LRO or hardware actually got the packet in microseconds. */ uint32_t cts, us_cts, ms_cts; uint32_t tiwin, high_seq; struct timespec ts; struct tcpopt to; struct tcp_rack *rack; struct rack_sendmap *rsm; int32_t prev_state = 0; #ifdef TCP_ACCOUNTING int ack_val_set = 0xf; #endif int nsegs; /* * tv passed from common code is from either M_TSTMP_LRO or * tcp_get_usecs() if no LRO m_pkthdr timestamp is present. */ rack = (struct tcp_rack *)tp->t_fb_ptr; if (m->m_flags & M_ACKCMP) { /* * All compressed ack's are ack's by definition so * remove any ack required flag and then do the processing. */ rack->rc_ack_required = 0; return (rack_do_compressed_ack_processing(tp, so, m, nxt_pkt, tv)); } if (m->m_flags & M_ACKCMP) { panic("Impossible reach m has ackcmp? m:%p tp:%p", m, tp); } cts = tcp_tv_to_usectick(tv); ms_cts = tcp_tv_to_mssectick(tv); nsegs = m->m_pkthdr.lro_nsegs; counter_u64_add(rack_proc_non_comp_ack, 1); thflags = tcp_get_flags(th); #ifdef TCP_ACCOUNTING sched_pin(); if (thflags & TH_ACK) ts_val = get_cyclecount(); #endif if ((m->m_flags & M_TSTMP) || (m->m_flags & M_TSTMP_LRO)) { mbuf_tstmp2timespec(m, &ts); rack->r_ctl.act_rcv_time.tv_sec = ts.tv_sec; rack->r_ctl.act_rcv_time.tv_usec = ts.tv_nsec/1000; } else rack->r_ctl.act_rcv_time = *tv; kern_prefetch(rack, &prev_state); prev_state = 0; /* * Unscale the window into a 32-bit value. For the SYN_SENT state * the scale is zero. */ tiwin = th->th_win << tp->snd_scale; #ifdef TCP_ACCOUNTING if (thflags & TH_ACK) { /* * We have a tradeoff here. We can either do what we are * doing i.e. pinning to this CPU and then doing the accounting * we could do a critical enter, setup the rdtsc and cpu * as in below, and then validate we are on the same CPU on * exit. I have choosen to not do the critical enter since * that often will gain you a context switch, and instead lock * us (line above this if) to the same CPU with sched_pin(). This * means we may be context switched out for a higher priority * interupt but we won't be moved to another CPU. * * If this occurs (which it won't very often since we most likely * are running this code in interupt context and only a higher * priority will bump us ... clock?) we will falsely add in * to the time the interupt processing time plus the ack processing * time. This is ok since its a rare event. */ ack_val_set = tcp_do_ack_accounting(tp, th, &to, tiwin, ctf_fixed_maxseg(tp)); } #endif /* * Parse options on any incoming segment. */ memset(&to, 0, sizeof(to)); tcp_dooptions(&to, (u_char *)(th + 1), (th->th_off << 2) - sizeof(struct tcphdr), (thflags & TH_SYN) ? TO_SYN : 0); NET_EPOCH_ASSERT(); INP_WLOCK_ASSERT(tp->t_inpcb); KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN", __func__)); if ((tp->t_state >= TCPS_FIN_WAIT_1) && (tp->t_flags & TF_GPUTINPROG)) { /* * We have a goodput in progress * and we have entered a late state. * Do we have enough data in the sb * to handle the GPUT request? */ uint32_t bytes; bytes = tp->gput_ack - tp->gput_seq; if (SEQ_GT(tp->gput_seq, tp->snd_una)) bytes += tp->gput_seq - tp->snd_una; if (bytes > sbavail(&tp->t_inpcb->inp_socket->so_snd)) { /* * There are not enough bytes in the socket * buffer that have been sent to cover this * measurement. Cancel it. */ rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/, rack->r_ctl.rc_gp_srtt /*flex1*/, tp->gput_seq, 0, 0, 18, __LINE__, NULL, 0); tp->t_flags &= ~TF_GPUTINPROG; } } high_seq = th->th_ack; if (tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval ltv; #ifdef NETFLIX_HTTP_LOGGING struct http_sendfile_track *http_req; if (SEQ_GT(th->th_ack, tp->snd_una)) { http_req = tcp_http_find_req_for_seq(tp, (th->th_ack-1)); } else { http_req = tcp_http_find_req_for_seq(tp, th->th_ack); } #endif memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp); if (rack->rack_no_prr == 0) log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt; else log.u_bbr.flex1 = 0; log.u_bbr.use_lt_bw = rack->r_ent_rec_ns; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->r_might_revert; log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced; log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.pkts_out = rack->rc_tp->t_maxseg; log.u_bbr.flex3 = m->m_flags; log.u_bbr.flex4 = rack->r_ctl.rc_hpts_flags; log.u_bbr.lost = thflags; log.u_bbr.pacing_gain = 0x1; #ifdef TCP_ACCOUNTING log.u_bbr.cwnd_gain = ack_val_set; #endif log.u_bbr.flex7 = 2; if (m->m_flags & M_TSTMP) { /* Record the hardware timestamp if present */ mbuf_tstmp2timespec(m, &ts); ltv.tv_sec = ts.tv_sec; ltv.tv_usec = ts.tv_nsec / 1000; log.u_bbr.lt_epoch = tcp_tv_to_usectick(<v); } else if (m->m_flags & M_TSTMP_LRO) { /* Record the LRO the arrival timestamp */ mbuf_tstmp2timespec(m, &ts); ltv.tv_sec = ts.tv_sec; ltv.tv_usec = ts.tv_nsec / 1000; log.u_bbr.flex5 = tcp_tv_to_usectick(<v); } log.u_bbr.timeStamp = tcp_get_usecs(<v); /* Log the rcv time */ log.u_bbr.delRate = m->m_pkthdr.rcv_tstmp; #ifdef NETFLIX_HTTP_LOGGING log.u_bbr.applimited = tp->t_http_closed; log.u_bbr.applimited <<= 8; log.u_bbr.applimited |= tp->t_http_open; log.u_bbr.applimited <<= 8; log.u_bbr.applimited |= tp->t_http_req; if (http_req) { /* Copy out any client req info */ /* seconds */ log.u_bbr.pkt_epoch = (http_req->localtime / HPTS_USEC_IN_SEC); /* useconds */ log.u_bbr.delivered = (http_req->localtime % HPTS_USEC_IN_SEC); log.u_bbr.rttProp = http_req->timestamp; log.u_bbr.cur_del_rate = http_req->start; if (http_req->flags & TCP_HTTP_TRACK_FLG_OPEN) { log.u_bbr.flex8 |= 1; } else { log.u_bbr.flex8 |= 2; log.u_bbr.bw_inuse = http_req->end; } log.u_bbr.flex6 = http_req->start_seq; if (http_req->flags & TCP_HTTP_TRACK_FLG_COMP) { log.u_bbr.flex8 |= 4; log.u_bbr.epoch = http_req->end_seq; } } #endif TCP_LOG_EVENTP(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0, tlen, &log, true, <v); } /* Remove ack required flag if set, we have one */ if (thflags & TH_ACK) rack->rc_ack_required = 0; if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) { way_out = 4; retval = 0; m_freem(m); goto done_with_input; } /* * If a segment with the ACK-bit set arrives in the SYN-SENT state * check SEQ.ACK first as described on page 66 of RFC 793, section 3.9. */ if ((tp->t_state == TCPS_SYN_SENT) && (thflags & TH_ACK) && (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) { tcp_log_end_status(tp, TCP_EI_STATUS_RST_IN_FRONT); ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); #ifdef TCP_ACCOUNTING sched_unpin(); #endif return (1); } /* * If timestamps were negotiated during SYN/ACK and a * segment without a timestamp is received, silently drop * the segment, unless it is a RST segment or missing timestamps are * tolerated. * See section 3.2 of RFC 7323. */ if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS) && ((thflags & TH_RST) == 0) && (V_tcp_tolerate_missing_ts == 0)) { way_out = 5; retval = 0; m_freem(m); goto done_with_input; } /* * Segment received on connection. Reset idle time and keep-alive * timer. XXX: This should be done after segment validation to * ignore broken/spoofed segs. */ if (tp->t_idle_reduce && (tp->snd_max == tp->snd_una) && (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) { counter_u64_add(rack_input_idle_reduces, 1); rack_cc_after_idle(rack, tp); } tp->t_rcvtime = ticks; #ifdef STATS stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_FRWIN, tiwin); #endif if (tiwin > rack->r_ctl.rc_high_rwnd) rack->r_ctl.rc_high_rwnd = tiwin; /* * TCP ECN processing. XXXJTL: If we ever use ECN, we need to move * this to occur after we've validated the segment. */ if (tcp_ecn_input_segment(tp, thflags, iptos)) rack_cong_signal(tp, CC_ECN, th->th_ack, __LINE__); /* * If echoed timestamp is later than the current time, fall back to * non RFC1323 RTT calculation. Normalize timestamp if syncookies * were used when this connection was established. */ if ((to.to_flags & TOF_TS) && (to.to_tsecr != 0)) { to.to_tsecr -= tp->ts_offset; if (TSTMP_GT(to.to_tsecr, ms_cts)) to.to_tsecr = 0; } /* * If its the first time in we need to take care of options and * verify we can do SACK for rack! */ if (rack->r_state == 0) { /* Should be init'd by rack_init() */ KASSERT(rack->rc_inp != NULL, ("%s: rack->rc_inp unexpectedly NULL", __func__)); if (rack->rc_inp == NULL) { rack->rc_inp = tp->t_inpcb; } /* * Process options only when we get SYN/ACK back. The SYN * case for incoming connections is handled in tcp_syncache. * According to RFC1323 the window field in a SYN (i.e., a * or ) segment itself is never scaled. XXX * this is traditional behavior, may need to be cleaned up. */ if (tp->t_state == TCPS_SYN_SENT && (thflags & TH_SYN)) { /* Handle parallel SYN for ECN */ tcp_ecn_input_parallel_syn(tp, thflags, iptos); if ((to.to_flags & TOF_SCALE) && (tp->t_flags & TF_REQ_SCALE)) { tp->t_flags |= TF_RCVD_SCALE; tp->snd_scale = to.to_wscale; } else tp->t_flags &= ~TF_REQ_SCALE; /* * Initial send window. It will be updated with the * next incoming segment to the scaled value. */ tp->snd_wnd = th->th_win; rack_validate_fo_sendwin_up(tp, rack); if ((to.to_flags & TOF_TS) && (tp->t_flags & TF_REQ_TSTMP)) { tp->t_flags |= TF_RCVD_TSTMP; tp->ts_recent = to.to_tsval; tp->ts_recent_age = cts; } else tp->t_flags &= ~TF_REQ_TSTMP; if (to.to_flags & TOF_MSS) { tcp_mss(tp, to.to_mss); } if ((tp->t_flags & TF_SACK_PERMIT) && (to.to_flags & TOF_SACKPERM) == 0) tp->t_flags &= ~TF_SACK_PERMIT; if (IS_FASTOPEN(tp->t_flags)) { if (to.to_flags & TOF_FASTOPEN) { uint16_t mss; if (to.to_flags & TOF_MSS) mss = to.to_mss; else if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0) mss = TCP6_MSS; else mss = TCP_MSS; tcp_fastopen_update_cache(tp, mss, to.to_tfo_len, to.to_tfo_cookie); } else tcp_fastopen_disable_path(tp); } } /* * At this point we are at the initial call. Here we decide * if we are doing RACK or not. We do this by seeing if * TF_SACK_PERMIT is set and the sack-not-required is clear. * The code now does do dup-ack counting so if you don't * switch back you won't get rack & TLP, but you will still * get this stack. */ if ((rack_sack_not_required == 0) && ((tp->t_flags & TF_SACK_PERMIT) == 0)) { tcp_switch_back_to_default(tp); (*tp->t_fb->tfb_tcp_do_segment) (m, th, so, tp, drop_hdrlen, tlen, iptos); #ifdef TCP_ACCOUNTING sched_unpin(); #endif return (1); } tcp_set_hpts(tp->t_inpcb); sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack); } if (thflags & TH_FIN) tcp_log_end_status(tp, TCP_EI_STATUS_CLIENT_FIN); us_cts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time); if ((rack->rc_gp_dyn_mul) && (rack->use_fixed_rate == 0) && (rack->rc_always_pace)) { /* Check in on probertt */ rack_check_probe_rtt(rack, us_cts); } rack_clear_rate_sample(rack); if ((rack->forced_ack) && ((tcp_get_flags(th) & TH_RST) == 0)) { rack_handle_probe_response(rack, tiwin, us_cts); } /* * This is the one exception case where we set the rack state * always. All other times (timers etc) we must have a rack-state * set (so we assure we have done the checks above for SACK). */ rack->r_ctl.rc_rcvtime = cts; if (rack->r_state != tp->t_state) rack_set_state(tp, rack); if (SEQ_GT(th->th_ack, tp->snd_una) && (rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree)) != NULL) kern_prefetch(rsm, &prev_state); prev_state = rack->r_state; retval = (*rack->r_substate) (m, th, so, tp, &to, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt, iptos); #ifdef INVARIANTS if ((retval == 0) && (tp->t_inpcb == NULL)) { panic("retval:%d tp:%p t_inpcb:NULL state:%d", retval, tp, prev_state); } #endif if (retval == 0) { /* * If retval is 1 the tcb is unlocked and most likely the tp * is gone. */ INP_WLOCK_ASSERT(tp->t_inpcb); if ((rack->rc_gp_dyn_mul) && (rack->rc_always_pace) && (rack->use_fixed_rate == 0) && rack->in_probe_rtt && (rack->r_ctl.rc_time_probertt_starts == 0)) { /* * If we are going for target, lets recheck before * we output. */ rack_check_probe_rtt(rack, us_cts); } if (rack->set_pacing_done_a_iw == 0) { /* How much has been acked? */ if ((tp->snd_una - tp->iss) > (ctf_fixed_maxseg(tp) * 10)) { /* We have enough to set in the pacing segment size */ rack->set_pacing_done_a_iw = 1; rack_set_pace_segments(tp, rack, __LINE__, NULL); } } tcp_rack_xmit_timer_commit(rack, tp); #ifdef TCP_ACCOUNTING /* * If we set the ack_val_se to what ack processing we are doing * we also want to track how many cycles we burned. Note * the bits after tcp_output we let be "free". This is because * we are also tracking the tcp_output times as well. Note the * use of 0xf here since we only have 11 counter (0 - 0xa) and * 0xf cannot be returned and is what we initialize it too to * indicate we are not doing the tabulations. */ if (ack_val_set != 0xf) { uint64_t crtsc; crtsc = get_cyclecount(); counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[ack_val_set] += (crtsc - ts_val); } } #endif if (nxt_pkt == 0) { if ((rack->r_wanted_output != 0) || (rack->r_fast_output != 0)) { do_output_now: if (tcp_output(tp) < 0) return (1); did_out = 1; } rack_start_hpts_timer(rack, tp, cts, 0, 0, 0); rack_free_trim(rack); } /* Update any rounds needed */ if (rack_verbose_logging && (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF)) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.flex1 = high_seq; log.u_bbr.flex2 = rack->r_ctl.roundends; log.u_bbr.flex3 = rack->r_ctl.current_round; log.u_bbr.rttProp = (uint64_t)CC_ALGO(tp)->newround; log.u_bbr.flex8 = 9; tcp_log_event_(tp, NULL, NULL, NULL, BBR_LOG_CWND, 0, 0, &log, false, NULL, NULL, 0, &tv); } /* * The draft (v3) calls for us to use SEQ_GEQ, but that * causes issues when we are just going app limited. Lets * instead use SEQ_GT where its equal but more data * is outstanding. */ if ((SEQ_GT(tp->snd_una, rack->r_ctl.roundends)) || ((tp->snd_una == rack->r_ctl.roundends) && SEQ_GT(tp->snd_max, tp->snd_una))) { rack->r_ctl.current_round++; rack->r_ctl.roundends = tp->snd_max; if (CC_ALGO(tp)->newround != NULL) { CC_ALGO(tp)->newround(tp->ccv, rack->r_ctl.current_round); } } if ((nxt_pkt == 0) && ((rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) == 0) && (SEQ_GT(tp->snd_max, tp->snd_una) || (tp->t_flags & TF_DELACK) || ((V_tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) && (tp->t_state <= TCPS_CLOSING)))) { /* We could not send (probably in the hpts but stopped the timer earlier)? */ if ((tp->snd_max == tp->snd_una) && ((tp->t_flags & TF_DELACK) == 0) && (tcp_in_hpts(rack->rc_inp)) && (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { /* keep alive not needed if we are hptsi output yet */ ; } else { int late = 0; if (tcp_in_hpts(rack->rc_inp)) { if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) { us_cts = tcp_get_usecs(NULL); if (TSTMP_GT(rack->r_ctl.rc_last_output_to, us_cts)) { rack->r_early = 1; rack->r_ctl.rc_agg_early += (rack->r_ctl.rc_last_output_to - us_cts); } else late = 1; rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT; } tcp_hpts_remove(tp->t_inpcb); } if (late && (did_out == 0)) { /* * We are late in the sending * and we did not call the output * (this probably should not happen). */ goto do_output_now; } rack_start_hpts_timer(rack, tp, tcp_get_usecs(NULL), 0, 0, 0); } way_out = 1; } else if (nxt_pkt == 0) { /* Do we have the correct timer running? */ rack_timer_audit(tp, rack, &so->so_snd); way_out = 2; } done_with_input: rack_log_doseg_done(rack, cts, nxt_pkt, did_out, way_out, max(1, nsegs)); if (did_out) rack->r_wanted_output = 0; #ifdef INVARIANTS if (tp->t_inpcb == NULL) { panic("OP:%d retval:%d tp:%p t_inpcb:NULL state:%d", did_out, retval, tp, prev_state); } #endif #ifdef TCP_ACCOUNTING } else { /* * Track the time (see above). */ if (ack_val_set != 0xf) { uint64_t crtsc; crtsc = get_cyclecount(); counter_u64_add(tcp_proc_time[ack_val_set] , (crtsc - ts_val)); /* * Note we *DO NOT* increment the per-tcb counters since * in the else the TP may be gone!! */ } #endif } #ifdef TCP_ACCOUNTING sched_unpin(); #endif return (retval); } void rack_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int32_t drop_hdrlen, int32_t tlen, uint8_t iptos) { struct timeval tv; /* First lets see if we have old packets */ if (tp->t_in_pkt) { if (ctf_do_queued_segments(so, tp, 1)) { m_freem(m); return; } } if (m->m_flags & M_TSTMP_LRO) { mbuf_tstmp2timeval(m, &tv); } else { /* Should not be should we kassert instead? */ tcp_get_usecs(&tv); } if (rack_do_segment_nounlock(m, th, so, tp, drop_hdrlen, tlen, iptos, 0, &tv) == 0) { INP_WUNLOCK(tp->t_inpcb); } } struct rack_sendmap * tcp_rack_output(struct tcpcb *tp, struct tcp_rack *rack, uint32_t tsused) { struct rack_sendmap *rsm = NULL; int32_t idx; uint32_t srtt = 0, thresh = 0, ts_low = 0; /* Return the next guy to be re-transmitted */ if (RB_EMPTY(&rack->r_ctl.rc_mtree)) { return (NULL); } if (tp->t_flags & TF_SENTFIN) { /* retran the end FIN? */ return (NULL); } /* ok lets look at this one */ rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); if (rack->r_must_retran && rsm && (rsm->r_flags & RACK_MUST_RXT)) { return (rsm); } if (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) { goto check_it; } rsm = rack_find_lowest_rsm(rack); if (rsm == NULL) { return (NULL); } check_it: if (((rack->rc_tp->t_flags & TF_SACK_PERMIT) == 0) && (rsm->r_dupack >= DUP_ACK_THRESHOLD)) { /* * No sack so we automatically do the 3 strikes and * retransmit (no rack timer would be started). */ return (rsm); } if (rsm->r_flags & RACK_ACKED) { return (NULL); } if (((rsm->r_flags & RACK_SACK_PASSED) == 0) && (rsm->r_dupack < DUP_ACK_THRESHOLD)) { /* Its not yet ready */ return (NULL); } srtt = rack_grab_rtt(tp, rack); idx = rsm->r_rtr_cnt - 1; ts_low = (uint32_t)rsm->r_tim_lastsent[idx]; thresh = rack_calc_thresh_rack(rack, srtt, tsused); if ((tsused == ts_low) || (TSTMP_LT(tsused, ts_low))) { /* No time since sending */ return (NULL); } if ((tsused - ts_low) < thresh) { /* It has not been long enough yet */ return (NULL); } if ((rsm->r_dupack >= DUP_ACK_THRESHOLD) || ((rsm->r_flags & RACK_SACK_PASSED) && (rack->sack_attack_disable == 0))) { /* * We have passed the dup-ack threshold * a SACK has indicated this is missing. * Note that if you are a declared attacker * it is only the dup-ack threshold that * will cause retransmits. */ /* log retransmit reason */ rack_log_retran_reason(rack, rsm, (tsused - ts_low), thresh, 1); rack->r_fast_output = 0; return (rsm); } return (NULL); } static void rack_log_pacing_delay_calc(struct tcp_rack *rack, uint32_t len, uint32_t slot, uint64_t bw_est, uint64_t bw, uint64_t len_time, int method, int line, struct rack_sendmap *rsm, uint8_t quality) { if (rack->rc_tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log, 0, sizeof(log)); log.u_bbr.flex1 = slot; log.u_bbr.flex2 = len; log.u_bbr.flex3 = rack->r_ctl.rc_pace_min_segs; log.u_bbr.flex4 = rack->r_ctl.rc_pace_max_segs; log.u_bbr.flex5 = rack->r_ctl.rack_per_of_gp_ss; log.u_bbr.flex6 = rack->r_ctl.rack_per_of_gp_ca; log.u_bbr.use_lt_bw = rack->rc_ack_can_sendout_data; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->r_late; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->r_early; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->app_limited_needs_set; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->rc_gp_filled; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->measure_saw_probe_rtt; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->in_probe_rtt; log.u_bbr.use_lt_bw <<= 1; log.u_bbr.use_lt_bw |= rack->gp_ready; log.u_bbr.pkt_epoch = line; log.u_bbr.epoch = rack->r_ctl.rc_agg_delayed; log.u_bbr.lt_epoch = rack->r_ctl.rc_agg_early; log.u_bbr.applimited = rack->r_ctl.rack_per_of_gp_rec; log.u_bbr.bw_inuse = bw_est; log.u_bbr.delRate = bw; if (rack->r_ctl.gp_bw == 0) log.u_bbr.cur_del_rate = 0; else log.u_bbr.cur_del_rate = rack_get_bw(rack); log.u_bbr.rttProp = len_time; log.u_bbr.pkts_out = rack->r_ctl.rc_rack_min_rtt; log.u_bbr.lost = rack->r_ctl.rc_probertt_sndmax_atexit; log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm); if (rack->r_ctl.cwnd_to_use < rack->rc_tp->snd_ssthresh) { /* We are in slow start */ log.u_bbr.flex7 = 1; } else { /* we are on congestion avoidance */ log.u_bbr.flex7 = 0; } log.u_bbr.flex8 = method; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.cwnd_gain = rack->rc_gp_saw_rec; log.u_bbr.cwnd_gain <<= 1; log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ss; log.u_bbr.cwnd_gain <<= 1; log.u_bbr.cwnd_gain |= rack->rc_gp_saw_ca; log.u_bbr.bbr_substate = quality; TCP_LOG_EVENTP(rack->rc_tp, NULL, &rack->rc_inp->inp_socket->so_rcv, &rack->rc_inp->inp_socket->so_snd, BBR_LOG_HPTSI_CALC, 0, 0, &log, false, &tv); } } static uint32_t rack_get_pacing_len(struct tcp_rack *rack, uint64_t bw, uint32_t mss) { uint32_t new_tso, user_max; user_max = rack->rc_user_set_max_segs * mss; if (rack->rc_force_max_seg) { return (user_max); } if (rack->use_fixed_rate && ((rack->r_ctl.crte == NULL) || (bw != rack->r_ctl.crte->rate))) { /* Use the user mss since we are not exactly matched */ return (user_max); } new_tso = tcp_get_pacing_burst_size(rack->rc_tp, bw, mss, rack_pace_one_seg, rack->r_ctl.crte, NULL); if (new_tso > user_max) new_tso = user_max; return (new_tso); } static int32_t pace_to_fill_cwnd(struct tcp_rack *rack, int32_t slot, uint32_t len, uint32_t segsiz, int *capped, uint64_t *rate_wanted, uint8_t non_paced) { uint64_t lentim, fill_bw; /* Lets first see if we are full, if so continue with normal rate */ rack->r_via_fill_cw = 0; if (ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked) > rack->r_ctl.cwnd_to_use) return (slot); if ((ctf_outstanding(rack->rc_tp) + (segsiz-1)) > rack->rc_tp->snd_wnd) return (slot); if (rack->r_ctl.rc_last_us_rtt == 0) return (slot); if (rack->rc_pace_fill_if_rttin_range && (rack->r_ctl.rc_last_us_rtt >= (get_filter_value_small(&rack->r_ctl.rc_gp_min_rtt) * rack->rtt_limit_mul))) { /* The rtt is huge, N * smallest, lets not fill */ return (slot); } /* * first lets calculate the b/w based on the last us-rtt * and the sndwnd. */ fill_bw = rack->r_ctl.cwnd_to_use; /* Take the rwnd if its smaller */ if (fill_bw > rack->rc_tp->snd_wnd) fill_bw = rack->rc_tp->snd_wnd; if (rack->r_fill_less_agg) { /* * Now take away the inflight (this will reduce our * aggressiveness and yeah, if we get that much out in 1RTT * we will have had acks come back and still be behind). */ fill_bw -= ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); } /* Now lets make it into a b/w */ fill_bw *= (uint64_t)HPTS_USEC_IN_SEC; fill_bw /= (uint64_t)rack->r_ctl.rc_last_us_rtt; /* We are below the min b/w */ if (non_paced) *rate_wanted = fill_bw; if ((fill_bw < RACK_MIN_BW) || (fill_bw < *rate_wanted)) return (slot); if (rack->r_ctl.bw_rate_cap && (fill_bw > rack->r_ctl.bw_rate_cap)) fill_bw = rack->r_ctl.bw_rate_cap; rack->r_via_fill_cw = 1; if (rack->r_rack_hw_rate_caps && (rack->r_ctl.crte != NULL)) { uint64_t high_rate; high_rate = tcp_hw_highest_rate(rack->r_ctl.crte); if (fill_bw > high_rate) { /* We are capping bw at the highest rate table entry */ if (*rate_wanted > high_rate) { /* The original rate was also capped */ rack->r_via_fill_cw = 0; } rack_log_hdwr_pacing(rack, fill_bw, high_rate, __LINE__, 0, 3); fill_bw = high_rate; if (capped) *capped = 1; } } else if ((rack->r_ctl.crte == NULL) && (rack->rack_hdrw_pacing == 0) && (rack->rack_hdw_pace_ena) && rack->r_rack_hw_rate_caps && (rack->rack_attempt_hdwr_pace == 0) && (rack->rc_inp->inp_route.ro_nh != NULL) && (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) { /* * Ok we may have a first attempt that is greater than our top rate * lets check. */ uint64_t high_rate; high_rate = tcp_hw_highest_rate_ifp(rack->rc_inp->inp_route.ro_nh->nh_ifp, rack->rc_inp); if (high_rate) { if (fill_bw > high_rate) { fill_bw = high_rate; if (capped) *capped = 1; } } } /* * Ok fill_bw holds our mythical b/w to fill the cwnd * in a rtt, what does that time wise equate too? */ lentim = (uint64_t)(len) * (uint64_t)HPTS_USEC_IN_SEC; lentim /= fill_bw; *rate_wanted = fill_bw; if (non_paced || (lentim < slot)) { rack_log_pacing_delay_calc(rack, len, slot, fill_bw, 0, lentim, 12, __LINE__, NULL, 0); return ((int32_t)lentim); } else return (slot); } static int32_t rack_get_pacing_delay(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len, struct rack_sendmap *rsm, uint32_t segsiz) { uint64_t srtt; int32_t slot = 0; int can_start_hw_pacing = 1; int err; if (rack->rc_always_pace == 0) { /* * We use the most optimistic possible cwnd/srtt for * sending calculations. This will make our * calculation anticipate getting more through * quicker then possible. But thats ok we don't want * the peer to have a gap in data sending. */ uint64_t cwnd, tr_perms = 0; int32_t reduce = 0; old_method: /* * We keep no precise pacing with the old method * instead we use the pacer to mitigate bursts. */ if (rack->r_ctl.rc_rack_min_rtt) srtt = rack->r_ctl.rc_rack_min_rtt; else srtt = max(tp->t_srtt, 1); if (rack->r_ctl.rc_rack_largest_cwnd) cwnd = rack->r_ctl.rc_rack_largest_cwnd; else cwnd = rack->r_ctl.cwnd_to_use; /* Inflate cwnd by 1000 so srtt of usecs is in ms */ tr_perms = (cwnd * 1000) / srtt; if (tr_perms == 0) { tr_perms = ctf_fixed_maxseg(tp); } /* * Calculate how long this will take to drain, if * the calculation comes out to zero, thats ok we * will use send_a_lot to possibly spin around for * more increasing tot_len_this_send to the point * that its going to require a pace, or we hit the * cwnd. Which in that case we are just waiting for * a ACK. */ slot = len / tr_perms; /* Now do we reduce the time so we don't run dry? */ if (slot && rack_slot_reduction) { reduce = (slot / rack_slot_reduction); if (reduce < slot) { slot -= reduce; } else slot = 0; } slot *= HPTS_USEC_IN_MSEC; if (rack->rc_pace_to_cwnd) { uint64_t rate_wanted = 0; slot = pace_to_fill_cwnd(rack, slot, len, segsiz, NULL, &rate_wanted, 1); rack->rc_ack_can_sendout_data = 1; rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, 0, 0, 14, __LINE__, NULL, 0); } else rack_log_pacing_delay_calc(rack, len, slot, tr_perms, reduce, 0, 7, __LINE__, NULL, 0); } else { uint64_t bw_est, res, lentim, rate_wanted; uint32_t orig_val, segs, oh; int capped = 0; int prev_fill; if ((rack->r_rr_config == 1) && rsm) { return (rack->r_ctl.rc_min_to); } if (rack->use_fixed_rate) { rate_wanted = bw_est = rack_get_fixed_pacing_bw(rack); } else if ((rack->r_ctl.init_rate == 0) && #ifdef NETFLIX_PEAKRATE (rack->rc_tp->t_maxpeakrate == 0) && #endif (rack->r_ctl.gp_bw == 0)) { /* no way to yet do an estimate */ bw_est = rate_wanted = 0; } else { bw_est = rack_get_bw(rack); rate_wanted = rack_get_output_bw(rack, bw_est, rsm, &capped); } if ((bw_est == 0) || (rate_wanted == 0) || ((rack->gp_ready == 0) && (rack->use_fixed_rate == 0))) { /* * No way yet to make a b/w estimate or * our raise is set incorrectly. */ goto old_method; } /* We need to account for all the overheads */ segs = (len + segsiz - 1) / segsiz; /* * We need the diff between 1514 bytes (e-mtu with e-hdr) * and how much data we put in each packet. Yes this * means we may be off if we are larger than 1500 bytes * or smaller. But this just makes us more conservative. */ if (rack_hw_rate_min && (bw_est < rack_hw_rate_min)) can_start_hw_pacing = 0; if (ETHERNET_SEGMENT_SIZE > segsiz) oh = ETHERNET_SEGMENT_SIZE - segsiz; else oh = 0; segs *= oh; lentim = (uint64_t)(len + segs) * (uint64_t)HPTS_USEC_IN_SEC; res = lentim / rate_wanted; slot = (uint32_t)res; orig_val = rack->r_ctl.rc_pace_max_segs; if (rack->r_ctl.crte == NULL) { /* * Only do this if we are not hardware pacing * since if we are doing hw-pacing below we will * set make a call after setting up or changing * the rate. */ rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL); } else if (rack->rc_inp->inp_snd_tag == NULL) { /* * We lost our rate somehow, this can happen * if the interface changed underneath us. */ tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp); rack->r_ctl.crte = NULL; /* Lets re-allow attempting to setup pacing */ rack->rack_hdrw_pacing = 0; rack->rack_attempt_hdwr_pace = 0; rack_log_hdwr_pacing(rack, rate_wanted, bw_est, __LINE__, 0, 6); } /* Did we change the TSO size, if so log it */ if (rack->r_ctl.rc_pace_max_segs != orig_val) rack_log_pacing_delay_calc(rack, len, slot, orig_val, 0, 0, 15, __LINE__, NULL, 0); prev_fill = rack->r_via_fill_cw; if ((rack->rc_pace_to_cwnd) && (capped == 0) && (rack->use_fixed_rate == 0) && (rack->in_probe_rtt == 0) && (IN_FASTRECOVERY(rack->rc_tp->t_flags) == 0)) { /* * We want to pace at our rate *or* faster to * fill the cwnd to the max if its not full. */ slot = pace_to_fill_cwnd(rack, slot, (len+segs), segsiz, &capped, &rate_wanted, 0); } if ((rack->rc_inp->inp_route.ro_nh != NULL) && (rack->rc_inp->inp_route.ro_nh->nh_ifp != NULL)) { if ((rack->rack_hdw_pace_ena) && (can_start_hw_pacing > 0) && (rack->rack_hdrw_pacing == 0) && (rack->rack_attempt_hdwr_pace == 0)) { /* * Lets attempt to turn on hardware pacing * if we can. */ rack->rack_attempt_hdwr_pace = 1; rack->r_ctl.crte = tcp_set_pacing_rate(rack->rc_tp, rack->rc_inp->inp_route.ro_nh->nh_ifp, rate_wanted, RS_PACING_GEQ, &err, &rack->r_ctl.crte_prev_rate); if (rack->r_ctl.crte) { rack->rack_hdrw_pacing = 1; rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted, segsiz, 0, rack->r_ctl.crte, NULL); rack_log_hdwr_pacing(rack, rate_wanted, rack->r_ctl.crte->rate, __LINE__, err, 0); rack->r_ctl.last_hw_bw_req = rate_wanted; } else { counter_u64_add(rack_hw_pace_init_fail, 1); } } else if (rack->rack_hdrw_pacing && (rack->r_ctl.last_hw_bw_req != rate_wanted)) { /* Do we need to adjust our rate? */ const struct tcp_hwrate_limit_table *nrte; if (rack->r_up_only && (rate_wanted < rack->r_ctl.crte->rate)) { /** * We have four possible states here * having to do with the previous time * and this time. * previous | this-time * A) 0 | 0 -- fill_cw not in the picture * B) 1 | 0 -- we were doing a fill-cw but now are not * C) 1 | 1 -- all rates from fill_cw * D) 0 | 1 -- we were doing non-fill and now we are filling * * For case A, C and D we don't allow a drop. But for * case B where we now our on our steady rate we do * allow a drop. * */ if (!((prev_fill == 1) && (rack->r_via_fill_cw == 0))) goto done_w_hdwr; } if ((rate_wanted > rack->r_ctl.crte->rate) || (rate_wanted <= rack->r_ctl.crte_prev_rate)) { if (rack_hw_rate_to_low && (bw_est < rack_hw_rate_to_low)) { /* * The pacing rate is too low for hardware, but * do allow hardware pacing to be restarted. */ rack_log_hdwr_pacing(rack, bw_est, rack->r_ctl.crte->rate, __LINE__, 0, 5); tcp_rel_pacing_rate(rack->r_ctl.crte, rack->rc_tp); rack->r_ctl.crte = NULL; rack->rack_attempt_hdwr_pace = 0; rack->rack_hdrw_pacing = 0; rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted); goto done_w_hdwr; } nrte = tcp_chg_pacing_rate(rack->r_ctl.crte, rack->rc_tp, rack->rc_inp->inp_route.ro_nh->nh_ifp, rate_wanted, RS_PACING_GEQ, &err, &rack->r_ctl.crte_prev_rate); if (nrte == NULL) { /* Lost the rate */ rack->rack_hdrw_pacing = 0; rack->r_ctl.crte = NULL; rack_log_hdwr_pacing(rack, rate_wanted, 0, __LINE__, err, 1); rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted); counter_u64_add(rack_hw_pace_lost, 1); } else if (nrte != rack->r_ctl.crte) { rack->r_ctl.crte = nrte; rack->r_ctl.rc_pace_max_segs = tcp_get_pacing_burst_size(tp, rate_wanted, segsiz, 0, rack->r_ctl.crte, NULL); rack_log_hdwr_pacing(rack, rate_wanted, rack->r_ctl.crte->rate, __LINE__, err, 2); rack->r_ctl.last_hw_bw_req = rate_wanted; } } else { /* We just need to adjust the segment size */ rack_set_pace_segments(rack->rc_tp, rack, __LINE__, &rate_wanted); rack_log_hdwr_pacing(rack, rate_wanted, rack->r_ctl.crte->rate, __LINE__, 0, 4); rack->r_ctl.last_hw_bw_req = rate_wanted; } } } if ((rack->r_ctl.crte != NULL) && (rack->r_ctl.crte->rate == rate_wanted)) { /* * We need to add a extra if the rates * are exactly matched. The idea is * we want the software to make sure the * queue is empty before adding more, this * gives us N MSS extra pace times where * N is our sysctl */ slot += (rack->r_ctl.crte->time_between * rack_hw_pace_extra_slots); } done_w_hdwr: if (rack_limit_time_with_srtt && (rack->use_fixed_rate == 0) && #ifdef NETFLIX_PEAKRATE (rack->rc_tp->t_maxpeakrate == 0) && #endif (rack->rack_hdrw_pacing == 0)) { /* * Sanity check, we do not allow the pacing delay * to be longer than the SRTT of the path. If it is * a slow path, then adding a packet should increase * the RTT and compensate for this i.e. the srtt will * be greater so the allowed pacing time will be greater. * * Note this restriction is not for where a peak rate * is set, we are doing fixed pacing or hardware pacing. */ if (rack->rc_tp->t_srtt) srtt = rack->rc_tp->t_srtt; else srtt = RACK_INITIAL_RTO * HPTS_USEC_IN_MSEC; /* its in ms convert */ if (srtt < (uint64_t)slot) { rack_log_pacing_delay_calc(rack, srtt, slot, rate_wanted, bw_est, lentim, 99, __LINE__, NULL, 0); slot = srtt; } } rack_log_pacing_delay_calc(rack, len, slot, rate_wanted, bw_est, lentim, 2, __LINE__, rsm, 0); } if (rack->r_ctl.crte && (rack->r_ctl.crte->rs_num_enobufs > 0)) { /* * If this rate is seeing enobufs when it * goes to send then either the nic is out * of gas or we are mis-estimating the time * somehow and not letting the queue empty * completely. Lets add to the pacing time. */ int hw_boost_delay; hw_boost_delay = rack->r_ctl.crte->time_between * rack_enobuf_hw_boost_mult; if (hw_boost_delay > rack_enobuf_hw_max) hw_boost_delay = rack_enobuf_hw_max; else if (hw_boost_delay < rack_enobuf_hw_min) hw_boost_delay = rack_enobuf_hw_min; slot += hw_boost_delay; } return (slot); } static void rack_start_gp_measurement(struct tcpcb *tp, struct tcp_rack *rack, tcp_seq startseq, uint32_t sb_offset) { struct rack_sendmap *my_rsm = NULL; struct rack_sendmap fe; if (tp->t_state < TCPS_ESTABLISHED) { /* * We don't start any measurements if we are * not at least established. */ return; } if (tp->t_state >= TCPS_FIN_WAIT_1) { /* * We will get no more data into the SB * this means we need to have the data available * before we start a measurement. */ if (sbavail(&tp->t_inpcb->inp_socket->so_snd) < max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp)))) { /* Nope not enough data */ return; } } tp->t_flags |= TF_GPUTINPROG; rack->r_ctl.rc_gp_lowrtt = 0xffffffff; rack->r_ctl.rc_gp_high_rwnd = rack->rc_tp->snd_wnd; tp->gput_seq = startseq; rack->app_limited_needs_set = 0; if (rack->in_probe_rtt) rack->measure_saw_probe_rtt = 1; else if ((rack->measure_saw_probe_rtt) && (SEQ_GEQ(tp->gput_seq, rack->r_ctl.rc_probertt_sndmax_atexit))) rack->measure_saw_probe_rtt = 0; if (rack->rc_gp_filled) tp->gput_ts = tcp_tv_to_usectick(&rack->r_ctl.act_rcv_time); else { /* Special case initial measurement */ struct timeval tv; tp->gput_ts = tcp_get_usecs(&tv); rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv); } /* * We take a guess out into the future, * if we have no measurement and no * initial rate, we measure the first * initial-windows worth of data to * speed up getting some GP measurement and * thus start pacing. */ if ((rack->rc_gp_filled == 0) && (rack->r_ctl.init_rate == 0)) { rack->app_limited_needs_set = 1; tp->gput_ack = startseq + max(rc_init_window(rack), (MIN_GP_WIN * ctf_fixed_maxseg(tp))); rack_log_pacing_delay_calc(rack, tp->gput_seq, tp->gput_ack, 0, tp->gput_ts, rack->r_ctl.rc_app_limited_cnt, 9, __LINE__, NULL, 0); return; } if (sb_offset) { /* * We are out somewhere in the sb * can we use the already outstanding data? */ if (rack->r_ctl.rc_app_limited_cnt == 0) { /* * Yes first one is good and in this case * the tp->gput_ts is correctly set based on * the last ack that arrived (no need to * set things up when an ack comes in). */ my_rsm = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree); if ((my_rsm == NULL) || (my_rsm->r_rtr_cnt != 1)) { /* retransmission? */ goto use_latest; } } else { if (rack->r_ctl.rc_first_appl == NULL) { /* * If rc_first_appl is NULL * then the cnt should be 0. * This is probably an error, maybe * a KASSERT would be approprate. */ goto use_latest; } /* * If we have a marker pointer to the last one that is * app limited we can use that, but we need to set * things up so that when it gets ack'ed we record * the ack time (if its not already acked). */ rack->app_limited_needs_set = 1; /* * We want to get to the rsm that is either * next with space i.e. over 1 MSS or the one * after that (after the app-limited). */ my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rack->r_ctl.rc_first_appl); if (my_rsm) { if ((my_rsm->r_end - my_rsm->r_start) <= ctf_fixed_maxseg(tp)) /* Have to use the next one */ my_rsm = RB_NEXT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, my_rsm); else { /* Use after the first MSS of it is acked */ tp->gput_seq = my_rsm->r_start + ctf_fixed_maxseg(tp); goto start_set; } } if ((my_rsm == NULL) || (my_rsm->r_rtr_cnt != 1)) { /* * Either its a retransmit or * the last is the app-limited one. */ goto use_latest; } } tp->gput_seq = my_rsm->r_start; start_set: if (my_rsm->r_flags & RACK_ACKED) { /* * This one has been acked use the arrival ack time */ tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival; rack->app_limited_needs_set = 0; } rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)]; tp->gput_ack = tp->gput_seq + rack_get_measure_window(tp, rack); rack_log_pacing_delay_calc(rack, tp->gput_seq, tp->gput_ack, (uint64_t)my_rsm, tp->gput_ts, rack->r_ctl.rc_app_limited_cnt, 9, __LINE__, NULL, 0); return; } use_latest: /* * We don't know how long we may have been * idle or if this is the first-send. Lets * setup the flag so we will trim off * the first ack'd data so we get a true * measurement. */ rack->app_limited_needs_set = 1; tp->gput_ack = startseq + rack_get_measure_window(tp, rack); /* Find this guy so we can pull the send time */ fe.r_start = startseq; my_rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe); if (my_rsm) { rack->r_ctl.rc_gp_output_ts = my_rsm->r_tim_lastsent[(my_rsm->r_rtr_cnt-1)]; if (my_rsm->r_flags & RACK_ACKED) { /* * Unlikely since its probably what was * just transmitted (but I am paranoid). */ tp->gput_ts = (uint32_t)my_rsm->r_ack_arrival; rack->app_limited_needs_set = 0; } if (SEQ_LT(my_rsm->r_start, tp->gput_seq)) { /* This also is unlikely */ tp->gput_seq = my_rsm->r_start; } } else { /* * TSNH unless we have some send-map limit, * and even at that it should not be hitting * that limit (we should have stopped sending). */ struct timeval tv; microuptime(&tv); rack->r_ctl.rc_gp_output_ts = rack_to_usec_ts(&tv); } rack_log_pacing_delay_calc(rack, tp->gput_seq, tp->gput_ack, (uint64_t)my_rsm, tp->gput_ts, rack->r_ctl.rc_app_limited_cnt, 9, __LINE__, NULL, 0); } static inline uint32_t rack_what_can_we_send(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cwnd_to_use, uint32_t avail, int32_t sb_offset) { uint32_t len; uint32_t sendwin; if (tp->snd_wnd > cwnd_to_use) sendwin = cwnd_to_use; else sendwin = tp->snd_wnd; if (ctf_outstanding(tp) >= tp->snd_wnd) { /* We never want to go over our peers rcv-window */ len = 0; } else { uint32_t flight; flight = ctf_flight_size(tp, rack->r_ctl.rc_sacked); if (flight >= sendwin) { /* * We have in flight what we are allowed by cwnd (if * it was rwnd blocking it would have hit above out * >= tp->snd_wnd). */ return (0); } len = sendwin - flight; if ((len + ctf_outstanding(tp)) > tp->snd_wnd) { /* We would send too much (beyond the rwnd) */ len = tp->snd_wnd - ctf_outstanding(tp); } if ((len + sb_offset) > avail) { /* * We don't have that much in the SB, how much is * there? */ len = avail - sb_offset; } } return (len); } static void rack_log_fsb(struct tcp_rack *rack, struct tcpcb *tp, struct socket *so, uint32_t flags, unsigned ipoptlen, int32_t orig_len, int32_t len, int error, int rsm_is_null, int optlen, int line, uint16_t mode) { if (tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp); log.u_bbr.flex1 = error; log.u_bbr.flex2 = flags; log.u_bbr.flex3 = rsm_is_null; log.u_bbr.flex4 = ipoptlen; log.u_bbr.flex5 = tp->rcv_numsacks; log.u_bbr.flex6 = rack->r_ctl.rc_agg_early; log.u_bbr.flex7 = optlen; log.u_bbr.flex8 = rack->r_fsb_inited; log.u_bbr.applimited = rack->r_fast_output; log.u_bbr.bw_inuse = rack_get_bw(rack); log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL); log.u_bbr.cwnd_gain = mode; log.u_bbr.pkts_out = orig_len; log.u_bbr.lt_epoch = len; log.u_bbr.delivered = line; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); tcp_log_event_(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_FSB, 0, len, &log, false, NULL, NULL, 0, &tv); } } static struct mbuf * rack_fo_base_copym(struct mbuf *the_m, uint32_t the_off, int32_t *plen, struct rack_fast_send_blk *fsb, int32_t seglimit, int32_t segsize, int hw_tls) { #ifdef KERN_TLS struct ktls_session *tls, *ntls; #ifdef INVARIANTS struct mbuf *start; #endif #endif struct mbuf *m, *n, **np, *smb; struct mbuf *top; int32_t off, soff; int32_t len = *plen; int32_t fragsize; int32_t len_cp = 0; uint32_t mlen, frags; soff = off = the_off; smb = m = the_m; np = ⊤ top = NULL; #ifdef KERN_TLS if (hw_tls && (m->m_flags & M_EXTPG)) tls = m->m_epg_tls; else tls = NULL; #ifdef INVARIANTS start = m; #endif #endif while (len > 0) { if (m == NULL) { *plen = len_cp; break; } #ifdef KERN_TLS if (hw_tls) { if (m->m_flags & M_EXTPG) ntls = m->m_epg_tls; else ntls = NULL; /* * Avoid mixing TLS records with handshake * data or TLS records from different * sessions. */ if (tls != ntls) { MPASS(m != start); *plen = len_cp; break; } } #endif mlen = min(len, m->m_len - off); if (seglimit) { /* * For M_EXTPG mbufs, add 3 segments * + 1 in case we are crossing page boundaries * + 2 in case the TLS hdr/trailer are used * It is cheaper to just add the segments * than it is to take the cache miss to look * at the mbuf ext_pgs state in detail. */ if (m->m_flags & M_EXTPG) { fragsize = min(segsize, PAGE_SIZE); frags = 3; } else { fragsize = segsize; frags = 0; } /* Break if we really can't fit anymore. */ if ((frags + 1) >= seglimit) { *plen = len_cp; break; } /* * Reduce size if you can't copy the whole * mbuf. If we can't copy the whole mbuf, also * adjust len so the loop will end after this * mbuf. */ if ((frags + howmany(mlen, fragsize)) >= seglimit) { mlen = (seglimit - frags - 1) * fragsize; len = mlen; *plen = len_cp + len; } frags += howmany(mlen, fragsize); if (frags == 0) frags++; seglimit -= frags; KASSERT(seglimit > 0, ("%s: seglimit went too low", __func__)); } n = m_get(M_NOWAIT, m->m_type); *np = n; if (n == NULL) goto nospace; n->m_len = mlen; soff += mlen; len_cp += n->m_len; if (m->m_flags & (M_EXT|M_EXTPG)) { n->m_data = m->m_data + off; mb_dupcl(n, m); } else { bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t), (u_int)n->m_len); } len -= n->m_len; off = 0; m = m->m_next; np = &n->m_next; if (len || (soff == smb->m_len)) { /* * We have more so we move forward or * we have consumed the entire mbuf and * len has fell to 0. */ soff = 0; smb = m; } } if (fsb != NULL) { fsb->m = smb; fsb->off = soff; if (smb) { /* * Save off the size of the mbuf. We do * this so that we can recognize when it * has been trimmed by sbcut() as acks * come in. */ fsb->o_m_len = smb->m_len; } else { /* * This is the case where the next mbuf went to NULL. This * means with this copy we have sent everything in the sb. * In theory we could clear the fast_output flag, but lets * not since its possible that we could get more added * and acks that call the extend function which would let * us send more. */ fsb->o_m_len = 0; } } return (top); nospace: if (top) m_freem(top); return (NULL); } /* * This is a copy of m_copym(), taking the TSO segment size/limit * constraints into account, and advancing the sndptr as it goes. */ static struct mbuf * rack_fo_m_copym(struct tcp_rack *rack, int32_t *plen, int32_t seglimit, int32_t segsize, struct mbuf **s_mb, int *s_soff) { struct mbuf *m, *n; int32_t soff; soff = rack->r_ctl.fsb.off; m = rack->r_ctl.fsb.m; if (rack->r_ctl.fsb.o_m_len > m->m_len) { /* * The mbuf had the front of it chopped off by an ack * we need to adjust the soff/off by that difference. */ uint32_t delta; delta = rack->r_ctl.fsb.o_m_len - m->m_len; soff -= delta; } else if (rack->r_ctl.fsb.o_m_len < m->m_len) { /* * The mbuf was expanded probably by * a m_compress. Just update o_m_len. */ rack->r_ctl.fsb.o_m_len = m->m_len; } KASSERT(soff >= 0, ("%s, negative off %d", __FUNCTION__, soff)); KASSERT(*plen >= 0, ("%s, negative len %d", __FUNCTION__, *plen)); KASSERT(soff < m->m_len, ("%s rack:%p len:%u m:%p m->m_len:%u < off?", __FUNCTION__, rack, *plen, m, m->m_len)); /* Save off the right location before we copy and advance */ *s_soff = soff; *s_mb = rack->r_ctl.fsb.m; n = rack_fo_base_copym(m, soff, plen, &rack->r_ctl.fsb, seglimit, segsize, rack->r_ctl.fsb.hw_tls); return (n); } static int rack_fast_rsm_output(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm, uint64_t ts_val, uint32_t cts, uint32_t ms_cts, struct timeval *tv, int len, uint8_t doing_tlp) { /* * Enter the fast retransmit path. We are given that a sched_pin is * in place (if accounting is compliled in) and the cycle count taken * at the entry is in the ts_val. The concept her is that the rsm * now holds the mbuf offsets and such so we can directly transmit * without a lot of overhead, the len field is already set for * us to prohibit us from sending too much (usually its 1MSS). */ struct ip *ip = NULL; struct udphdr *udp = NULL; struct tcphdr *th = NULL; struct mbuf *m = NULL; struct inpcb *inp; uint8_t *cpto; struct tcp_log_buffer *lgb; #ifdef TCP_ACCOUNTING uint64_t crtsc; int cnt_thru = 1; #endif struct tcpopt to; u_char opt[TCP_MAXOLEN]; uint32_t hdrlen, optlen; int32_t slot, segsiz, max_val, tso = 0, error, ulen = 0; uint16_t flags; uint32_t if_hw_tsomaxsegcount = 0, startseq; uint32_t if_hw_tsomaxsegsize; #ifdef INET6 struct ip6_hdr *ip6 = NULL; if (rack->r_is_v6) { ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr; hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); } else #endif /* INET6 */ { ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr; hdrlen = sizeof(struct tcpiphdr); } if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) { goto failed; } if (doing_tlp) { /* Its a TLP add the flag, it may already be there but be sure */ rsm->r_flags |= RACK_TLP; } else { /* If it was a TLP it is not not on this retransmit */ rsm->r_flags &= ~RACK_TLP; } startseq = rsm->r_start; segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs); inp = rack->rc_inp; to.to_flags = 0; flags = tcp_outflags[tp->t_state]; if (flags & (TH_SYN|TH_RST)) { goto failed; } if (rsm->r_flags & RACK_HAS_FIN) { /* We can't send a FIN here */ goto failed; } if (flags & TH_FIN) { /* We never send a FIN */ flags &= ~TH_FIN; } if (tp->t_flags & TF_RCVD_TSTMP) { to.to_tsval = ms_cts + tp->ts_offset; to.to_tsecr = tp->ts_recent; to.to_flags = TOF_TS; } optlen = tcp_addoptions(&to, opt); hdrlen += optlen; udp = rack->r_ctl.fsb.udp; if (udp) hdrlen += sizeof(struct udphdr); if (rack->r_ctl.rc_pace_max_segs) max_val = rack->r_ctl.rc_pace_max_segs; else if (rack->rc_user_set_max_segs) max_val = rack->rc_user_set_max_segs * segsiz; else max_val = len; if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && (len > segsiz) && (tp->t_port == 0)) tso = 1; #ifdef INET6 if (MHLEN < hdrlen + max_linkhdr) m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); else #endif m = m_gethdr(M_NOWAIT, MT_DATA); if (m == NULL) goto failed; m->m_data += max_linkhdr; m->m_len = hdrlen; th = rack->r_ctl.fsb.th; /* Establish the len to send */ if (len > max_val) len = max_val; if ((tso) && (len + optlen > tp->t_maxseg)) { uint32_t if_hw_tsomax; int32_t max_len; /* extract TSO information */ if_hw_tsomax = tp->t_tsomax; if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; /* * Check if we should limit by maximum payload * length: */ if (if_hw_tsomax != 0) { /* compute maximum TSO length */ max_len = (if_hw_tsomax - hdrlen - max_linkhdr); if (max_len <= 0) { goto failed; } else if (len > max_len) { len = max_len; } } if (len <= segsiz) { /* * In case there are too many small fragments don't * use TSO: */ tso = 0; } } else { tso = 0; } if ((tso == 0) && (len > segsiz)) len = segsiz; if ((len == 0) || (len <= MHLEN - hdrlen - max_linkhdr)) { goto failed; } th->th_seq = htonl(rsm->r_start); th->th_ack = htonl(tp->rcv_nxt); /* * The PUSH bit should only be applied * if the full retransmission is made. If * we are sending less than this is the * left hand edge and should not have * the PUSH bit. */ if ((rsm->r_flags & RACK_HAD_PUSH) && (len == (rsm->r_end - rsm->r_start))) flags |= TH_PUSH; th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale)); if (th->th_win == 0) { tp->t_sndzerowin++; tp->t_flags |= TF_RXWIN0SENT; } else tp->t_flags &= ~TF_RXWIN0SENT; if (rsm->r_flags & RACK_TLP) { /* * TLP should not count in retran count, but * in its own bin */ counter_u64_add(rack_tlp_retran, 1); counter_u64_add(rack_tlp_retran_bytes, len); } else { tp->t_sndrexmitpack++; KMOD_TCPSTAT_INC(tcps_sndrexmitpack); KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len); } #ifdef STATS stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB, len); #endif if (rsm->m == NULL) goto failed; if (rsm->orig_m_len != rsm->m->m_len) { /* Fix up the orig_m_len and possibly the mbuf offset */ rack_adjust_orig_mlen(rsm); } m->m_next = rack_fo_base_copym(rsm->m, rsm->soff, &len, NULL, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, rsm->r_hw_tls); if (len <= segsiz) { /* * Must have ran out of mbufs for the copy * shorten it to no longer need tso. Lets * not put on sendalot since we are low on * mbufs. */ tso = 0; } if ((m->m_next == NULL) || (len <= 0)){ goto failed; } if (udp) { if (rack->r_is_v6) ulen = hdrlen + len - sizeof(struct ip6_hdr); else ulen = hdrlen + len - sizeof(struct ip); udp->uh_ulen = htons(ulen); } m->m_pkthdr.rcvif = (struct ifnet *)0; if (TCPS_HAVERCVDSYN(tp->t_state) && (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) { int ect = tcp_ecn_output_established(tp, &flags, len, true); if ((tp->t_state == TCPS_SYN_RECEIVED) && (tp->t_flags2 & TF2_ECN_SND_ECE)) tp->t_flags2 &= ~TF2_ECN_SND_ECE; #ifdef INET6 if (rack->r_is_v6) { ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20); ip6->ip6_flow |= htonl(ect << 20); } else #endif { ip->ip_tos &= ~IPTOS_ECN_MASK; ip->ip_tos |= ect; } } tcp_set_flags(th, flags); m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ #ifdef INET6 if (rack->r_is_v6) { if (tp->t_port) { m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0); th->th_sum = htons(0); UDPSTAT_INC(udps_opackets); } else { m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP, 0); } } #endif #if defined(INET6) && defined(INET) else #endif #ifdef INET { if (tp->t_port) { m->m_pkthdr.csum_flags = CSUM_UDP; m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); udp->uh_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); th->th_sum = htons(0); UDPSTAT_INC(udps_opackets); } else { m->m_pkthdr.csum_flags = CSUM_TCP; m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen)); } /* IP version must be set here for ipv4/ipv6 checking later */ KASSERT(ip->ip_v == IPVERSION, ("%s: IP version incorrect: %d", __func__, ip->ip_v)); } #endif if (tso) { KASSERT(len > tp->t_maxseg - optlen, ("%s: len <= tso_segsz tp:%p", __func__, tp)); m->m_pkthdr.csum_flags |= CSUM_TSO; m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen; } #ifdef INET6 if (rack->r_is_v6) { ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit; ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) tp->t_flags2 |= TF2_PLPMTU_PMTUD; else tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { ip->ip_len = htons(m->m_pkthdr.len); ip->ip_ttl = rack->r_ctl.fsb.hoplimit; if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) { tp->t_flags2 |= TF2_PLPMTU_PMTUD; if (tp->t_port == 0 || len < V_tcp_minmss) { ip->ip_off |= htons(IP_DF); } } else { tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; } } #endif /* Time to copy in our header */ cpto = mtod(m, uint8_t *); memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len); th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr)); if (optlen) { bcopy(opt, th + 1, optlen); th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; } else { th->th_off = sizeof(struct tcphdr) >> 2; } if (tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; if (rsm->r_flags & RACK_RWND_COLLAPSED) { rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm); counter_u64_add(rack_collapsed_win_rxt, 1); counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start)); } memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp); if (rack->rack_no_prr) log.u_bbr.flex1 = 0; else log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs; log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs; log.u_bbr.flex4 = max_val; log.u_bbr.flex5 = 0; /* Save off the early/late values */ log.u_bbr.flex6 = rack->r_ctl.rc_agg_early; log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed; log.u_bbr.bw_inuse = rack_get_bw(rack); if (doing_tlp == 0) log.u_bbr.flex8 = 1; else log.u_bbr.flex8 = 2; log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL); log.u_bbr.flex7 = 55; log.u_bbr.pkts_out = tp->t_maxseg; log.u_bbr.timeStamp = cts; log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use; log.u_bbr.delivered = 0; lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK, len, &log, false, NULL, NULL, 0, tv); } else lgb = NULL; #ifdef INET6 if (rack->r_is_v6) { error = ip6_output(m, NULL, &inp->inp_route6, 0, NULL, NULL, inp); } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { error = ip_output(m, NULL, &inp->inp_route, 0, 0, inp); } #endif m = NULL; if (lgb) { lgb->tlb_errno = error; lgb = NULL; } if (error) { goto failed; } rack_log_output(tp, &to, len, rsm->r_start, flags, error, rack_to_usec_ts(tv), rsm, RACK_SENT_FP, rsm->m, rsm->soff, rsm->r_hw_tls); if (doing_tlp && (rack->fast_rsm_hack == 0)) { rack->rc_tlp_in_progress = 1; rack->r_ctl.rc_tlp_cnt_out++; } if (error == 0) { tcp_account_for_send(tp, len, 1, doing_tlp, rsm->r_hw_tls); if (doing_tlp) { rack->rc_last_sent_tlp_past_cumack = 0; rack->rc_last_sent_tlp_seq_valid = 1; rack->r_ctl.last_sent_tlp_seq = rsm->r_start; rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start; } } tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); rack->forced_ack = 0; /* If we send something zap the FA flag */ if (IN_FASTRECOVERY(tp->t_flags) && rsm) rack->r_ctl.retran_during_recovery += len; { int idx; idx = (len / segsiz) + 3; if (idx >= TCP_MSS_ACCT_ATIMER) counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1); else counter_u64_add(rack_out_size[idx], 1); } if (tp->t_rtttime == 0) { tp->t_rtttime = ticks; tp->t_rtseq = startseq; KMOD_TCPSTAT_INC(tcps_segstimed); } counter_u64_add(rack_fto_rsm_send, 1); if (error && (error == ENOBUFS)) { if (rack->r_ctl.crte != NULL) { rack_trace_point(rack, RACK_TP_HWENOBUF); } else rack_trace_point(rack, RACK_TP_ENOBUF); slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC); if (rack->rc_enobuf < 0x7f) rack->rc_enobuf++; if (slot < (10 * HPTS_USEC_IN_MSEC)) slot = 10 * HPTS_USEC_IN_MSEC; } else slot = rack_get_pacing_delay(rack, tp, len, NULL, segsiz); if ((slot == 0) || (rack->rc_always_pace == 0) || (rack->r_rr_config == 1)) { /* * We have no pacing set or we * are using old-style rack or * we are overridden to use the old 1ms pacing. */ slot = rack->r_ctl.rc_min_to; } rack_start_hpts_timer(rack, tp, cts, slot, len, 0); #ifdef TCP_ACCOUNTING crtsc = get_cyclecount(); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru; } counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val); } counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((len + segsiz - 1) / segsiz); } counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((len + segsiz - 1) / segsiz)); sched_unpin(); #endif return (0); failed: if (m) m_free(m); return (-1); } static void rack_sndbuf_autoscale(struct tcp_rack *rack) { /* * Automatic sizing of send socket buffer. Often the send buffer * size is not optimally adjusted to the actual network conditions * at hand (delay bandwidth product). Setting the buffer size too * small limits throughput on links with high bandwidth and high * delay (eg. trans-continental/oceanic links). Setting the * buffer size too big consumes too much real kernel memory, * especially with many connections on busy servers. * * The criteria to step up the send buffer one notch are: * 1. receive window of remote host is larger than send buffer * (with a fudge factor of 5/4th); * 2. send buffer is filled to 7/8th with data (so we actually * have data to make use of it); * 3. send buffer fill has not hit maximal automatic size; * 4. our send window (slow start and cogestion controlled) is * larger than sent but unacknowledged data in send buffer. * * Note that the rack version moves things much faster since * we want to avoid hitting cache lines in the rack_fast_output() * path so this is called much less often and thus moves * the SB forward by a percentage. */ struct socket *so; struct tcpcb *tp; uint32_t sendwin, scaleup; tp = rack->rc_tp; so = rack->rc_inp->inp_socket; sendwin = min(rack->r_ctl.cwnd_to_use, tp->snd_wnd); if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) { if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat && sbused(&so->so_snd) >= (so->so_snd.sb_hiwat / 8 * 7) && sbused(&so->so_snd) < V_tcp_autosndbuf_max && sendwin >= (sbused(&so->so_snd) - (tp->snd_nxt - tp->snd_una))) { if (rack_autosndbuf_inc) scaleup = (rack_autosndbuf_inc * so->so_snd.sb_hiwat) / 100; else scaleup = V_tcp_autosndbuf_inc; if (scaleup < V_tcp_autosndbuf_inc) scaleup = V_tcp_autosndbuf_inc; scaleup += so->so_snd.sb_hiwat; if (scaleup > V_tcp_autosndbuf_max) scaleup = V_tcp_autosndbuf_max; if (!sbreserve_locked(so, SO_SND, scaleup, curthread)) so->so_snd.sb_flags &= ~SB_AUTOSIZE; } } } static int rack_fast_output(struct tcpcb *tp, struct tcp_rack *rack, uint64_t ts_val, uint32_t cts, uint32_t ms_cts, struct timeval *tv, long tot_len, int *send_err) { /* * Enter to do fast output. We are given that the sched_pin is * in place (if accounting is compiled in) and the cycle count taken * at entry is in place in ts_val. The idea here is that * we know how many more bytes needs to be sent (presumably either * during pacing or to fill the cwnd and that was greater than * the max-burst). We have how much to send and all the info we * need to just send. */ struct ip *ip = NULL; struct udphdr *udp = NULL; struct tcphdr *th = NULL; struct mbuf *m, *s_mb; struct inpcb *inp; uint8_t *cpto; struct tcp_log_buffer *lgb; #ifdef TCP_ACCOUNTING uint64_t crtsc; #endif struct tcpopt to; u_char opt[TCP_MAXOLEN]; uint32_t hdrlen, optlen; #ifdef TCP_ACCOUNTING int cnt_thru = 1; #endif int32_t slot, segsiz, len, max_val, tso = 0, sb_offset, error, ulen = 0; uint16_t flags; uint32_t s_soff; uint32_t if_hw_tsomaxsegcount = 0, startseq; uint32_t if_hw_tsomaxsegsize; uint16_t add_flag = RACK_SENT_FP; #ifdef INET6 struct ip6_hdr *ip6 = NULL; if (rack->r_is_v6) { ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr; hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); } else #endif /* INET6 */ { ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr; hdrlen = sizeof(struct tcpiphdr); } if (tp->t_port && (V_tcp_udp_tunneling_port == 0)) { m = NULL; goto failed; } startseq = tp->snd_max; segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs); inp = rack->rc_inp; len = rack->r_ctl.fsb.left_to_send; to.to_flags = 0; flags = rack->r_ctl.fsb.tcp_flags; if (tp->t_flags & TF_RCVD_TSTMP) { to.to_tsval = ms_cts + tp->ts_offset; to.to_tsecr = tp->ts_recent; to.to_flags = TOF_TS; } optlen = tcp_addoptions(&to, opt); hdrlen += optlen; udp = rack->r_ctl.fsb.udp; if (udp) hdrlen += sizeof(struct udphdr); if (rack->r_ctl.rc_pace_max_segs) max_val = rack->r_ctl.rc_pace_max_segs; else if (rack->rc_user_set_max_segs) max_val = rack->rc_user_set_max_segs * segsiz; else max_val = len; if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && (len > segsiz) && (tp->t_port == 0)) tso = 1; again: #ifdef INET6 if (MHLEN < hdrlen + max_linkhdr) m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); else #endif m = m_gethdr(M_NOWAIT, MT_DATA); if (m == NULL) goto failed; m->m_data += max_linkhdr; m->m_len = hdrlen; th = rack->r_ctl.fsb.th; /* Establish the len to send */ if (len > max_val) len = max_val; if ((tso) && (len + optlen > tp->t_maxseg)) { uint32_t if_hw_tsomax; int32_t max_len; /* extract TSO information */ if_hw_tsomax = tp->t_tsomax; if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; /* * Check if we should limit by maximum payload * length: */ if (if_hw_tsomax != 0) { /* compute maximum TSO length */ max_len = (if_hw_tsomax - hdrlen - max_linkhdr); if (max_len <= 0) { goto failed; } else if (len > max_len) { len = max_len; } } if (len <= segsiz) { /* * In case there are too many small fragments don't * use TSO: */ tso = 0; } } else { tso = 0; } if ((tso == 0) && (len > segsiz)) len = segsiz; if ((len == 0) || (len <= MHLEN - hdrlen - max_linkhdr)) { goto failed; } sb_offset = tp->snd_max - tp->snd_una; th->th_seq = htonl(tp->snd_max); th->th_ack = htonl(tp->rcv_nxt); th->th_win = htons((u_short)(rack->r_ctl.fsb.recwin >> tp->rcv_scale)); if (th->th_win == 0) { tp->t_sndzerowin++; tp->t_flags |= TF_RXWIN0SENT; } else tp->t_flags &= ~TF_RXWIN0SENT; tp->snd_up = tp->snd_una; /* drag it along, its deprecated */ KMOD_TCPSTAT_INC(tcps_sndpack); KMOD_TCPSTAT_ADD(tcps_sndbyte, len); #ifdef STATS stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB, len); #endif if (rack->r_ctl.fsb.m == NULL) goto failed; /* s_mb and s_soff are saved for rack_log_output */ m->m_next = rack_fo_m_copym(rack, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, &s_mb, &s_soff); if (len <= segsiz) { /* * Must have ran out of mbufs for the copy * shorten it to no longer need tso. Lets * not put on sendalot since we are low on * mbufs. */ tso = 0; } if (rack->r_ctl.fsb.rfo_apply_push && (len == rack->r_ctl.fsb.left_to_send)) { flags |= TH_PUSH; add_flag |= RACK_HAD_PUSH; } if ((m->m_next == NULL) || (len <= 0)){ goto failed; } if (udp) { if (rack->r_is_v6) ulen = hdrlen + len - sizeof(struct ip6_hdr); else ulen = hdrlen + len - sizeof(struct ip); udp->uh_ulen = htons(ulen); } m->m_pkthdr.rcvif = (struct ifnet *)0; if (TCPS_HAVERCVDSYN(tp->t_state) && (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) { int ect = tcp_ecn_output_established(tp, &flags, len, false); if ((tp->t_state == TCPS_SYN_RECEIVED) && (tp->t_flags2 & TF2_ECN_SND_ECE)) tp->t_flags2 &= ~TF2_ECN_SND_ECE; #ifdef INET6 if (rack->r_is_v6) { ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20); ip6->ip6_flow |= htonl(ect << 20); } else #endif { ip->ip_tos &= ~IPTOS_ECN_MASK; ip->ip_tos |= ect; } } tcp_set_flags(th, flags); m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ #ifdef INET6 if (rack->r_is_v6) { if (tp->t_port) { m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0); th->th_sum = htons(0); UDPSTAT_INC(udps_opackets); } else { m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP, 0); } } #endif #if defined(INET6) && defined(INET) else #endif #ifdef INET { if (tp->t_port) { m->m_pkthdr.csum_flags = CSUM_UDP; m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); udp->uh_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); th->th_sum = htons(0); UDPSTAT_INC(udps_opackets); } else { m->m_pkthdr.csum_flags = CSUM_TCP; m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen)); } /* IP version must be set here for ipv4/ipv6 checking later */ KASSERT(ip->ip_v == IPVERSION, ("%s: IP version incorrect: %d", __func__, ip->ip_v)); } #endif if (tso) { KASSERT(len > tp->t_maxseg - optlen, ("%s: len <= tso_segsz tp:%p", __func__, tp)); m->m_pkthdr.csum_flags |= CSUM_TSO; m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen; } #ifdef INET6 if (rack->r_is_v6) { ip6->ip6_hlim = rack->r_ctl.fsb.hoplimit; ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) tp->t_flags2 |= TF2_PLPMTU_PMTUD; else tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { ip->ip_len = htons(m->m_pkthdr.len); ip->ip_ttl = rack->r_ctl.fsb.hoplimit; if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) { tp->t_flags2 |= TF2_PLPMTU_PMTUD; if (tp->t_port == 0 || len < V_tcp_minmss) { ip->ip_off |= htons(IP_DF); } } else { tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; } } #endif /* Time to copy in our header */ cpto = mtod(m, uint8_t *); memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len); th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr)); if (optlen) { bcopy(opt, th + 1, optlen); th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; } else { th->th_off = sizeof(struct tcphdr) >> 2; } if (tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp); if (rack->rack_no_prr) log.u_bbr.flex1 = 0; else log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs; log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs; log.u_bbr.flex4 = max_val; log.u_bbr.flex5 = 0; /* Save off the early/late values */ log.u_bbr.flex6 = rack->r_ctl.rc_agg_early; log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed; log.u_bbr.bw_inuse = rack_get_bw(rack); log.u_bbr.flex8 = 0; log.u_bbr.pacing_gain = rack_get_output_gain(rack, NULL); log.u_bbr.flex7 = 44; log.u_bbr.pkts_out = tp->t_maxseg; log.u_bbr.timeStamp = cts; log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.lt_epoch = rack->r_ctl.cwnd_to_use; log.u_bbr.delivered = 0; lgb = tcp_log_event_(tp, th, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK, len, &log, false, NULL, NULL, 0, tv); } else lgb = NULL; #ifdef INET6 if (rack->r_is_v6) { error = ip6_output(m, NULL, &inp->inp_route6, 0, NULL, NULL, inp); } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { error = ip_output(m, NULL, &inp->inp_route, 0, 0, inp); } #endif if (lgb) { lgb->tlb_errno = error; lgb = NULL; } if (error) { *send_err = error; m = NULL; goto failed; } rack_log_output(tp, &to, len, tp->snd_max, flags, error, rack_to_usec_ts(tv), NULL, add_flag, s_mb, s_soff, rack->r_ctl.fsb.hw_tls); m = NULL; if (tp->snd_una == tp->snd_max) { rack->r_ctl.rc_tlp_rxt_last_time = cts; rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__); tp->t_acktime = ticks; } if (error == 0) tcp_account_for_send(tp, len, 0, 0, rack->r_ctl.fsb.hw_tls); rack->forced_ack = 0; /* If we send something zap the FA flag */ tot_len += len; if ((tp->t_flags & TF_GPUTINPROG) == 0) rack_start_gp_measurement(tp, rack, tp->snd_max, sb_offset); tp->snd_max += len; tp->snd_nxt = tp->snd_max; { int idx; idx = (len / segsiz) + 3; if (idx >= TCP_MSS_ACCT_ATIMER) counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1); else counter_u64_add(rack_out_size[idx], 1); } if (len <= rack->r_ctl.fsb.left_to_send) rack->r_ctl.fsb.left_to_send -= len; else rack->r_ctl.fsb.left_to_send = 0; if (rack->r_ctl.fsb.left_to_send < segsiz) { rack->r_fast_output = 0; rack->r_ctl.fsb.left_to_send = 0; /* At the end of fast_output scale up the sb */ SOCKBUF_LOCK(&rack->rc_inp->inp_socket->so_snd); rack_sndbuf_autoscale(rack); SOCKBUF_UNLOCK(&rack->rc_inp->inp_socket->so_snd); } if (tp->t_rtttime == 0) { tp->t_rtttime = ticks; tp->t_rtseq = startseq; KMOD_TCPSTAT_INC(tcps_segstimed); } if ((rack->r_ctl.fsb.left_to_send >= segsiz) && (max_val > len) && (tso == 0)) { max_val -= len; len = segsiz; th = rack->r_ctl.fsb.th; #ifdef TCP_ACCOUNTING cnt_thru++; #endif goto again; } tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); counter_u64_add(rack_fto_send, 1); slot = rack_get_pacing_delay(rack, tp, tot_len, NULL, segsiz); rack_start_hpts_timer(rack, tp, cts, slot, tot_len, 0); #ifdef TCP_ACCOUNTING crtsc = get_cyclecount(); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[SND_OUT_DATA] += cnt_thru; } counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], cnt_thru); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val); } counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len + segsiz - 1) / segsiz); } counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len + segsiz - 1) / segsiz)); sched_unpin(); #endif return (0); failed: if (m) m_free(m); rack->r_fast_output = 0; return (-1); } static struct rack_sendmap * rack_check_collapsed(struct tcp_rack *rack, uint32_t cts) { struct rack_sendmap *rsm = NULL; struct rack_sendmap fe; int thresh; restart: fe.r_start = rack->r_ctl.last_collapse_point; rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe); if ((rsm == NULL) || ((rsm->r_flags & RACK_RWND_COLLAPSED) == 0)) { /* Nothing, strange turn off validity */ rack->r_collapse_point_valid = 0; return (NULL); } /* Can we send it yet? */ if (rsm->r_end > (rack->rc_tp->snd_una + rack->rc_tp->snd_wnd)) { /* * Receiver window has not grown enough for * the segment to be put on the wire. */ return (NULL); } if (rsm->r_flags & RACK_ACKED) { /* * It has been sacked, lets move to the * next one if possible. */ rack->r_ctl.last_collapse_point = rsm->r_end; /* Are we done? */ if (SEQ_GEQ(rack->r_ctl.last_collapse_point, rack->r_ctl.high_collapse_point)) { rack->r_collapse_point_valid = 0; return (NULL); } goto restart; } /* Now has it been long enough ? */ thresh = rack_calc_thresh_rack(rack, rack_grab_rtt(rack->rc_tp, rack), cts); if ((cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])) > thresh) { rack_log_collapse(rack, rsm->r_start, (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])), thresh, __LINE__, 6, rsm->r_flags, rsm); return (rsm); } /* Not enough time */ rack_log_collapse(rack, rsm->r_start, (cts - ((uint32_t)rsm->r_tim_lastsent[(rsm->r_rtr_cnt-1)])), thresh, __LINE__, 7, rsm->r_flags, rsm); return (NULL); } static int rack_output(struct tcpcb *tp) { struct socket *so; uint32_t recwin; uint32_t sb_offset, s_moff = 0; int32_t len, error = 0; uint16_t flags; struct mbuf *m, *s_mb = NULL; struct mbuf *mb; uint32_t if_hw_tsomaxsegcount = 0; uint32_t if_hw_tsomaxsegsize; int32_t segsiz, minseg; long tot_len_this_send = 0; #ifdef INET struct ip *ip = NULL; #endif struct udphdr *udp = NULL; struct tcp_rack *rack; struct tcphdr *th; uint8_t pass = 0; uint8_t mark = 0; uint8_t wanted_cookie = 0; u_char opt[TCP_MAXOLEN]; unsigned ipoptlen, optlen, hdrlen, ulen=0; uint32_t rack_seq; #if defined(IPSEC) || defined(IPSEC_SUPPORT) unsigned ipsec_optlen = 0; #endif int32_t idle, sendalot; int32_t sub_from_prr = 0; volatile int32_t sack_rxmit; struct rack_sendmap *rsm = NULL; int32_t tso, mtu; struct tcpopt to; int32_t slot = 0; int32_t sup_rack = 0; uint32_t cts, ms_cts, delayed, early; uint16_t add_flag = RACK_SENT_SP; /* The doing_tlp flag will be set by the actual rack_timeout_tlp() */ uint8_t hpts_calling, doing_tlp = 0; uint32_t cwnd_to_use, pace_max_seg; int32_t do_a_prefetch = 0; int32_t prefetch_rsm = 0; int32_t orig_len = 0; struct timeval tv; int32_t prefetch_so_done = 0; struct tcp_log_buffer *lgb; struct inpcb *inp; struct sockbuf *sb; uint64_t ts_val = 0; #ifdef TCP_ACCOUNTING uint64_t crtsc; #endif #ifdef INET6 struct ip6_hdr *ip6 = NULL; int32_t isipv6; #endif bool hw_tls = false; /* setup and take the cache hits here */ rack = (struct tcp_rack *)tp->t_fb_ptr; #ifdef TCP_ACCOUNTING sched_pin(); ts_val = get_cyclecount(); #endif hpts_calling = rack->rc_inp->inp_hpts_calls; NET_EPOCH_ASSERT(); INP_WLOCK_ASSERT(rack->rc_inp); #ifdef TCP_OFFLOAD if (tp->t_flags & TF_TOE) { #ifdef TCP_ACCOUNTING sched_unpin(); #endif return (tcp_offload_output(tp)); } #endif /* * For TFO connections in SYN_RECEIVED, only allow the initial * SYN|ACK and those sent by the retransmit timer. */ if (IS_FASTOPEN(tp->t_flags) && (tp->t_state == TCPS_SYN_RECEIVED) && SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN|ACK sent */ (rack->r_ctl.rc_resend == NULL)) { /* not a retransmit */ #ifdef TCP_ACCOUNTING sched_unpin(); #endif return (0); } #ifdef INET6 if (rack->r_state) { /* Use the cache line loaded if possible */ isipv6 = rack->r_is_v6; } else { isipv6 = (rack->rc_inp->inp_vflag & INP_IPV6) != 0; } #endif early = 0; cts = tcp_get_usecs(&tv); ms_cts = tcp_tv_to_mssectick(&tv); if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) && tcp_in_hpts(rack->rc_inp)) { /* * We are on the hpts for some timer but not hptsi output. * Remove from the hpts unconditionally. */ rack_timer_cancel(tp, rack, cts, __LINE__); } /* Are we pacing and late? */ if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) && TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) { /* We are delayed */ delayed = cts - rack->r_ctl.rc_last_output_to; } else { delayed = 0; } /* Do the timers, which may override the pacer */ if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { int retval; retval = rack_process_timers(tp, rack, cts, hpts_calling, &doing_tlp); if (retval != 0) { counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1); #ifdef TCP_ACCOUNTING sched_unpin(); #endif /* * If timers want tcp_drop(), then pass error out, * otherwise suppress it. */ return (retval < 0 ? retval : 0); } } if (rack->rc_in_persist) { if (tcp_in_hpts(rack->rc_inp) == 0) { /* Timer is not running */ rack_start_hpts_timer(rack, tp, cts, 0, 0, 0); } #ifdef TCP_ACCOUNTING sched_unpin(); #endif return (0); } if ((rack->rc_ack_required == 1) && (rack->r_timer_override == 0)){ /* A timeout occurred and no ack has arrived */ if (tcp_in_hpts(rack->rc_inp) == 0) { /* Timer is not running */ rack_start_hpts_timer(rack, tp, cts, 0, 0, 0); } #ifdef TCP_ACCOUNTING sched_unpin(); #endif return (0); } if ((rack->r_timer_override) || (rack->rc_ack_can_sendout_data) || (delayed) || (tp->t_state < TCPS_ESTABLISHED)) { rack->rc_ack_can_sendout_data = 0; if (tcp_in_hpts(rack->rc_inp)) tcp_hpts_remove(rack->rc_inp); } else if (tcp_in_hpts(rack->rc_inp)) { /* * On the hpts you can't pass even if ACKNOW is on, we will * when the hpts fires. */ #ifdef TCP_ACCOUNTING crtsc = get_cyclecount(); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[SND_BLOCKED] += (crtsc - ts_val); } counter_u64_add(tcp_proc_time[SND_BLOCKED], (crtsc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[SND_BLOCKED]++; } counter_u64_add(tcp_cnt_counters[SND_BLOCKED], 1); sched_unpin(); #endif counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1); return (0); } rack->rc_inp->inp_hpts_calls = 0; /* Finish out both pacing early and late accounting */ if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) && TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) { early = rack->r_ctl.rc_last_output_to - cts; } else early = 0; if (delayed) { rack->r_ctl.rc_agg_delayed += delayed; rack->r_late = 1; } else if (early) { rack->r_ctl.rc_agg_early += early; rack->r_early = 1; } /* Now that early/late accounting is done turn off the flag */ rack->r_ctl.rc_hpts_flags &= ~PACE_PKT_OUTPUT; rack->r_wanted_output = 0; rack->r_timer_override = 0; if ((tp->t_state != rack->r_state) && TCPS_HAVEESTABLISHED(tp->t_state)) { rack_set_state(tp, rack); } if ((rack->r_fast_output) && (doing_tlp == 0) && (tp->rcv_numsacks == 0)) { int ret; error = 0; ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error); if (ret >= 0) return(ret); else if (error) { inp = rack->rc_inp; so = inp->inp_socket; sb = &so->so_snd; goto nomore; } } inp = rack->rc_inp; /* * For TFO connections in SYN_SENT or SYN_RECEIVED, * only allow the initial SYN or SYN|ACK and those sent * by the retransmit timer. */ if (IS_FASTOPEN(tp->t_flags) && ((tp->t_state == TCPS_SYN_RECEIVED) || (tp->t_state == TCPS_SYN_SENT)) && SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */ (tp->t_rxtshift == 0)) { /* not a retransmit */ cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd; so = inp->inp_socket; sb = &so->so_snd; goto just_return_nolock; } /* * Determine length of data that should be transmitted, and flags * that will be used. If there is some data or critical controls * (SYN, RST) to send, then transmit; otherwise, investigate * further. */ idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una); if (tp->t_idle_reduce) { if (idle && (TICKS_2_USEC(ticks - tp->t_rcvtime) >= tp->t_rxtcur)) rack_cc_after_idle(rack, tp); } tp->t_flags &= ~TF_LASTIDLE; if (idle) { if (tp->t_flags & TF_MORETOCOME) { tp->t_flags |= TF_LASTIDLE; idle = 0; } } if ((tp->snd_una == tp->snd_max) && rack->r_ctl.rc_went_idle_time && TSTMP_GT(cts, rack->r_ctl.rc_went_idle_time)) { idle = cts - rack->r_ctl.rc_went_idle_time; if (idle > rack_min_probertt_hold) { /* Count as a probe rtt */ if (rack->in_probe_rtt == 0) { rack->r_ctl.rc_lower_rtt_us_cts = cts; rack->r_ctl.rc_time_probertt_entered = rack->r_ctl.rc_lower_rtt_us_cts; rack->r_ctl.rc_time_probertt_starts = rack->r_ctl.rc_lower_rtt_us_cts; rack->r_ctl.rc_time_of_last_probertt = rack->r_ctl.rc_lower_rtt_us_cts; } else { rack_exit_probertt(rack, cts); } } idle = 0; } if (rack_use_fsb && (rack->r_fsb_inited == 0) && (rack->r_state != TCPS_CLOSED)) rack_init_fsb_block(tp, rack); again: /* * If we've recently taken a timeout, snd_max will be greater than * snd_nxt. There may be SACK information that allows us to avoid * resending already delivered data. Adjust snd_nxt accordingly. */ sendalot = 0; cts = tcp_get_usecs(&tv); ms_cts = tcp_tv_to_mssectick(&tv); tso = 0; mtu = 0; segsiz = min(ctf_fixed_maxseg(tp), rack->r_ctl.rc_pace_min_segs); minseg = segsiz; if (rack->r_ctl.rc_pace_max_segs == 0) pace_max_seg = rack->rc_user_set_max_segs * segsiz; else pace_max_seg = rack->r_ctl.rc_pace_max_segs; sb_offset = tp->snd_max - tp->snd_una; cwnd_to_use = rack->r_ctl.cwnd_to_use = tp->snd_cwnd; flags = tcp_outflags[tp->t_state]; while (rack->rc_free_cnt < rack_free_cache) { rsm = rack_alloc(rack); if (rsm == NULL) { if (inp->inp_hpts_calls) /* Retry in a ms */ slot = (1 * HPTS_USEC_IN_MSEC); so = inp->inp_socket; sb = &so->so_snd; goto just_return_nolock; } TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext); rack->rc_free_cnt++; rsm = NULL; } if (inp->inp_hpts_calls) inp->inp_hpts_calls = 0; sack_rxmit = 0; len = 0; rsm = NULL; if (flags & TH_RST) { SOCKBUF_LOCK(&inp->inp_socket->so_snd); so = inp->inp_socket; sb = &so->so_snd; goto send; } if (rack->r_ctl.rc_resend) { /* Retransmit timer */ rsm = rack->r_ctl.rc_resend; rack->r_ctl.rc_resend = NULL; len = rsm->r_end - rsm->r_start; sack_rxmit = 1; sendalot = 0; KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start), ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p", __func__, __LINE__, rsm->r_start, tp->snd_una, tp, rack, rsm)); sb_offset = rsm->r_start - tp->snd_una; if (len >= segsiz) len = segsiz; } else if (rack->r_collapse_point_valid && ((rsm = rack_check_collapsed(rack, cts)) != NULL)) { /* * If an RSM is returned then enough time has passed * for us to retransmit it. Move up the collapse point, * since this rsm has its chance to retransmit now. */ rack_trace_point(rack, RACK_TP_COLLAPSED_RXT); rack->r_ctl.last_collapse_point = rsm->r_end; /* Are we done? */ if (SEQ_GEQ(rack->r_ctl.last_collapse_point, rack->r_ctl.high_collapse_point)) rack->r_collapse_point_valid = 0; sack_rxmit = 1; /* We are not doing a TLP */ doing_tlp = 0; len = rsm->r_end - rsm->r_start; sb_offset = rsm->r_start - tp->snd_una; sendalot = 0; if ((rack->full_size_rxt == 0) && (rack->shape_rxt_to_pacing_min == 0) && (len >= segsiz)) len = segsiz; } else if ((rsm = tcp_rack_output(tp, rack, cts)) != NULL) { /* We have a retransmit that takes precedence */ if ((!IN_FASTRECOVERY(tp->t_flags)) && ((rsm->r_flags & RACK_MUST_RXT) == 0) && ((tp->t_flags & TF_WASFRECOVERY) == 0)) { /* Enter recovery if not induced by a time-out */ rack_cong_signal(tp, CC_NDUPACK, tp->snd_una, __LINE__); } #ifdef INVARIANTS if (SEQ_LT(rsm->r_start, tp->snd_una)) { panic("Huh, tp:%p rack:%p rsm:%p start:%u < snd_una:%u\n", tp, rack, rsm, rsm->r_start, tp->snd_una); } #endif len = rsm->r_end - rsm->r_start; KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start), ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p", __func__, __LINE__, rsm->r_start, tp->snd_una, tp, rack, rsm)); sb_offset = rsm->r_start - tp->snd_una; sendalot = 0; if (len >= segsiz) len = segsiz; if (len > 0) { sack_rxmit = 1; KMOD_TCPSTAT_INC(tcps_sack_rexmits); KMOD_TCPSTAT_ADD(tcps_sack_rexmit_bytes, min(len, segsiz)); } } else if (rack->r_ctl.rc_tlpsend) { /* Tail loss probe */ long cwin; long tlen; /* * Check if we can do a TLP with a RACK'd packet * this can happen if we are not doing the rack * cheat and we skipped to a TLP and it * went off. */ rsm = rack->r_ctl.rc_tlpsend; /* We are doing a TLP make sure the flag is preent */ rsm->r_flags |= RACK_TLP; rack->r_ctl.rc_tlpsend = NULL; sack_rxmit = 1; tlen = rsm->r_end - rsm->r_start; if (tlen > segsiz) tlen = segsiz; KASSERT(SEQ_LEQ(tp->snd_una, rsm->r_start), ("%s:%d: r.start:%u < SND.UNA:%u; tp:%p, rack:%p, rsm:%p", __func__, __LINE__, rsm->r_start, tp->snd_una, tp, rack, rsm)); sb_offset = rsm->r_start - tp->snd_una; cwin = min(tp->snd_wnd, tlen); len = cwin; } if (rack->r_must_retran && (doing_tlp == 0) && (SEQ_GT(tp->snd_max, tp->snd_una)) && (rsm == NULL)) { /* * There are two different ways that we * can get into this block: * a) This is a non-sack connection, we had a time-out * and thus r_must_retran was set and everything * left outstanding as been marked for retransmit. * b) The MTU of the path shrank, so that everything * was marked to be retransmitted with the smaller * mtu and r_must_retran was set. * * This means that we expect the sendmap (outstanding) * to all be marked must. We can use the tmap to * look at them. * */ int sendwin, flight; sendwin = min(tp->snd_wnd, tp->snd_cwnd); flight = ctf_flight_size(tp, rack->r_ctl.rc_out_at_rto); if (flight >= sendwin) { /* * We can't send yet. */ so = inp->inp_socket; sb = &so->so_snd; goto just_return_nolock; } /* * This is the case a/b mentioned above. All * outstanding/not-acked should be marked. * We can use the tmap to find them. */ rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); if (rsm == NULL) { /* TSNH */ rack->r_must_retran = 0; rack->r_ctl.rc_out_at_rto = 0; so = inp->inp_socket; sb = &so->so_snd; goto just_return_nolock; } if ((rsm->r_flags & RACK_MUST_RXT) == 0) { /* * The first one does not have the flag, did we collapse * further up in our list? */ rack->r_must_retran = 0; rack->r_ctl.rc_out_at_rto = 0; rsm = NULL; sack_rxmit = 0; } else { sack_rxmit = 1; len = rsm->r_end - rsm->r_start; sb_offset = rsm->r_start - tp->snd_una; sendalot = 0; if ((rack->full_size_rxt == 0) && (rack->shape_rxt_to_pacing_min == 0) && (len >= segsiz)) len = segsiz; /* * Delay removing the flag RACK_MUST_RXT so * that the fastpath for retransmit will * work with this rsm. */ } } /* * Enforce a connection sendmap count limit if set * as long as we are not retransmiting. */ if ((rsm == NULL) && (rack->do_detection == 0) && (V_tcp_map_entries_limit > 0) && (rack->r_ctl.rc_num_maps_alloced >= V_tcp_map_entries_limit)) { counter_u64_add(rack_to_alloc_limited, 1); if (!rack->alloc_limit_reported) { rack->alloc_limit_reported = 1; counter_u64_add(rack_alloc_limited_conns, 1); } so = inp->inp_socket; sb = &so->so_snd; goto just_return_nolock; } if (rsm && (rsm->r_flags & RACK_HAS_FIN)) { /* we are retransmitting the fin */ len--; if (len) { /* * When retransmitting data do *not* include the * FIN. This could happen from a TLP probe. */ flags &= ~TH_FIN; } } if (rsm && rack->r_fsb_inited && rack_use_rsm_rfo && ((rsm->r_flags & RACK_HAS_FIN) == 0)) { int ret; ret = rack_fast_rsm_output(tp, rack, rsm, ts_val, cts, ms_cts, &tv, len, doing_tlp); if (ret == 0) return (0); } so = inp->inp_socket; sb = &so->so_snd; if (do_a_prefetch == 0) { kern_prefetch(sb, &do_a_prefetch); do_a_prefetch = 1; } #ifdef NETFLIX_SHARED_CWND if ((tp->t_flags2 & TF2_TCP_SCWND_ALLOWED) && rack->rack_enable_scwnd) { /* We are doing cwnd sharing */ if (rack->gp_ready && (rack->rack_attempted_scwnd == 0) && (rack->r_ctl.rc_scw == NULL) && tp->t_lib) { /* The pcbid is in, lets make an attempt */ counter_u64_add(rack_try_scwnd, 1); rack->rack_attempted_scwnd = 1; rack->r_ctl.rc_scw = tcp_shared_cwnd_alloc(tp, &rack->r_ctl.rc_scw_index, segsiz); } if (rack->r_ctl.rc_scw && (rack->rack_scwnd_is_idle == 1) && sbavail(&so->so_snd)) { /* we are no longer out of data */ tcp_shared_cwnd_active(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index); rack->rack_scwnd_is_idle = 0; } if (rack->r_ctl.rc_scw) { /* First lets update and get the cwnd */ rack->r_ctl.cwnd_to_use = cwnd_to_use = tcp_shared_cwnd_update(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index, tp->snd_cwnd, tp->snd_wnd, segsiz); } } #endif /* * Get standard flags, and add SYN or FIN if requested by 'hidden' * state flags. */ if (tp->t_flags & TF_NEEDFIN) flags |= TH_FIN; if (tp->t_flags & TF_NEEDSYN) flags |= TH_SYN; if ((sack_rxmit == 0) && (prefetch_rsm == 0)) { void *end_rsm; end_rsm = TAILQ_LAST_FAST(&rack->r_ctl.rc_tmap, rack_sendmap, r_tnext); if (end_rsm) kern_prefetch(end_rsm, &prefetch_rsm); prefetch_rsm = 1; } SOCKBUF_LOCK(sb); /* * If snd_nxt == snd_max and we have transmitted a FIN, the * sb_offset will be > 0 even if so_snd.sb_cc is 0, resulting in a * negative length. This can also occur when TCP opens up its * congestion window while receiving additional duplicate acks after * fast-retransmit because TCP will reset snd_nxt to snd_max after * the fast-retransmit. * * In the normal retransmit-FIN-only case, however, snd_nxt will be * set to snd_una, the sb_offset will be 0, and the length may wind * up 0. * * If sack_rxmit is true we are retransmitting from the scoreboard * in which case len is already set. */ if ((sack_rxmit == 0) && (TCPS_HAVEESTABLISHED(tp->t_state) || IS_FASTOPEN(tp->t_flags))) { uint32_t avail; avail = sbavail(sb); if (SEQ_GT(tp->snd_nxt, tp->snd_una) && avail) sb_offset = tp->snd_nxt - tp->snd_una; else sb_offset = 0; if ((IN_FASTRECOVERY(tp->t_flags) == 0) || rack->rack_no_prr) { if (rack->r_ctl.rc_tlp_new_data) { /* TLP is forcing out new data */ if (rack->r_ctl.rc_tlp_new_data > (uint32_t) (avail - sb_offset)) { rack->r_ctl.rc_tlp_new_data = (uint32_t) (avail - sb_offset); } if ((rack->r_ctl.rc_tlp_new_data + sb_offset) > tp->snd_wnd) { if (tp->snd_wnd > sb_offset) len = tp->snd_wnd - sb_offset; else len = 0; } else { len = rack->r_ctl.rc_tlp_new_data; } rack->r_ctl.rc_tlp_new_data = 0; } else { len = rack_what_can_we_send(tp, rack, cwnd_to_use, avail, sb_offset); } if ((rack->r_ctl.crte == NULL) && IN_FASTRECOVERY(tp->t_flags) && (len > segsiz)) { /* * For prr=off, we need to send only 1 MSS * at a time. We do this because another sack could * be arriving that causes us to send retransmits and * we don't want to be on a long pace due to a larger send * that keeps us from sending out the retransmit. */ len = segsiz; } } else { uint32_t outstanding; /* * We are inside of a Fast recovery episode, this * is caused by a SACK or 3 dup acks. At this point * we have sent all the retransmissions and we rely * on PRR to dictate what we will send in the form of * new data. */ outstanding = tp->snd_max - tp->snd_una; if ((rack->r_ctl.rc_prr_sndcnt + outstanding) > tp->snd_wnd) { if (tp->snd_wnd > outstanding) { len = tp->snd_wnd - outstanding; /* Check to see if we have the data */ if ((sb_offset + len) > avail) { /* It does not all fit */ if (avail > sb_offset) len = avail - sb_offset; else len = 0; } } else { len = 0; } } else if (avail > sb_offset) { len = avail - sb_offset; } else { len = 0; } if (len > 0) { if (len > rack->r_ctl.rc_prr_sndcnt) { len = rack->r_ctl.rc_prr_sndcnt; } if (len > 0) { sub_from_prr = 1; } } if (len > segsiz) { /* * We should never send more than a MSS when * retransmitting or sending new data in prr * mode unless the override flag is on. Most * likely the PRR algorithm is not going to * let us send a lot as well :-) */ if (rack->r_ctl.rc_prr_sendalot == 0) { len = segsiz; } } else if (len < segsiz) { /* * Do we send any? The idea here is if the * send empty's the socket buffer we want to * do it. However if not then lets just wait * for our prr_sndcnt to get bigger. */ long leftinsb; leftinsb = sbavail(sb) - sb_offset; if (leftinsb > len) { /* This send does not empty the sb */ len = 0; } } } } else if (!TCPS_HAVEESTABLISHED(tp->t_state)) { /* * If you have not established * and are not doing FAST OPEN * no data please. */ if ((sack_rxmit == 0) && (!IS_FASTOPEN(tp->t_flags))){ len = 0; sb_offset = 0; } } if (prefetch_so_done == 0) { kern_prefetch(so, &prefetch_so_done); prefetch_so_done = 1; } /* * Lop off SYN bit if it has already been sent. However, if this is * SYN-SENT state and if segment contains data and if we don't know * that foreign host supports TAO, suppress sending segment. */ if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una) && ((sack_rxmit == 0) && (tp->t_rxtshift == 0))) { /* * When sending additional segments following a TFO SYN|ACK, * do not include the SYN bit. */ if (IS_FASTOPEN(tp->t_flags) && (tp->t_state == TCPS_SYN_RECEIVED)) flags &= ~TH_SYN; } /* * Be careful not to send data and/or FIN on SYN segments. This * measure is needed to prevent interoperability problems with not * fully conformant TCP implementations. */ if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) { len = 0; flags &= ~TH_FIN; } /* * On TFO sockets, ensure no data is sent in the following cases: * * - When retransmitting SYN|ACK on a passively-created socket * * - When retransmitting SYN on an actively created socket * * - When sending a zero-length cookie (cookie request) on an * actively created socket * * - When the socket is in the CLOSED state (RST is being sent) */ if (IS_FASTOPEN(tp->t_flags) && (((flags & TH_SYN) && (tp->t_rxtshift > 0)) || ((tp->t_state == TCPS_SYN_SENT) && (tp->t_tfo_client_cookie_len == 0)) || (flags & TH_RST))) { sack_rxmit = 0; len = 0; } /* Without fast-open there should never be data sent on a SYN */ if ((flags & TH_SYN) && (!IS_FASTOPEN(tp->t_flags))) { tp->snd_nxt = tp->iss; len = 0; } if ((len > segsiz) && (tcp_dsack_block_exists(tp))) { /* We only send 1 MSS if we have a DSACK block */ add_flag |= RACK_SENT_W_DSACK; len = segsiz; } orig_len = len; if (len <= 0) { /* * If FIN has been sent but not acked, but we haven't been * called to retransmit, len will be < 0. Otherwise, window * shrank after we sent into it. If window shrank to 0, * cancel pending retransmit, pull snd_nxt back to (closed) * window, and set the persist timer if it isn't already * going. If the window didn't close completely, just wait * for an ACK. * * We also do a general check here to ensure that we will * set the persist timer when we have data to send, but a * 0-byte window. This makes sure the persist timer is set * even if the packet hits one of the "goto send" lines * below. */ len = 0; if ((tp->snd_wnd == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) && (tp->snd_una == tp->snd_max) && (sb_offset < (int)sbavail(sb))) { rack_enter_persist(tp, rack, cts); } } else if ((rsm == NULL) && (doing_tlp == 0) && (len < pace_max_seg)) { /* * We are not sending a maximum sized segment for * some reason. Should we not send anything (think * sws or persists)? */ if ((tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg)) && (TCPS_HAVEESTABLISHED(tp->t_state)) && (len < minseg) && (len < (int)(sbavail(sb) - sb_offset))) { /* * Here the rwnd is less than * the minimum pacing size, this is not a retransmit, * we are established and * the send is not the last in the socket buffer * we send nothing, and we may enter persists * if nothing is outstanding. */ len = 0; if (tp->snd_max == tp->snd_una) { /* * Nothing out we can * go into persists. */ rack_enter_persist(tp, rack, cts); } } else if ((cwnd_to_use >= max(minseg, (segsiz * 4))) && (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) && (len < (int)(sbavail(sb) - sb_offset)) && (len < minseg)) { /* * Here we are not retransmitting, and * the cwnd is not so small that we could * not send at least a min size (rxt timer * not having gone off), We have 2 segments or * more already in flight, its not the tail end * of the socket buffer and the cwnd is blocking * us from sending out a minimum pacing segment size. * Lets not send anything. */ len = 0; } else if (((tp->snd_wnd - ctf_outstanding(tp)) < min((rack->r_ctl.rc_high_rwnd/2), minseg)) && (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * segsiz)) && (len < (int)(sbavail(sb) - sb_offset)) && (TCPS_HAVEESTABLISHED(tp->t_state))) { /* * Here we have a send window but we have * filled it up and we can't send another pacing segment. * We also have in flight more than 2 segments * and we are not completing the sb i.e. we allow * the last bytes of the sb to go out even if * its not a full pacing segment. */ len = 0; } else if ((rack->r_ctl.crte != NULL) && (tp->snd_wnd >= (pace_max_seg * max(1, rack_hw_rwnd_factor))) && (cwnd_to_use >= (pace_max_seg + (4 * segsiz))) && (ctf_flight_size(tp, rack->r_ctl.rc_sacked) >= (2 * segsiz)) && (len < (int)(sbavail(sb) - sb_offset))) { /* * Here we are doing hardware pacing, this is not a TLP, * we are not sending a pace max segment size, there is rwnd * room to send at least N pace_max_seg, the cwnd is greater * than or equal to a full pacing segments plus 4 mss and we have 2 or * more segments in flight and its not the tail of the socket buffer. * * We don't want to send instead we need to get more ack's in to * allow us to send a full pacing segment. Normally, if we are pacing * about the right speed, we should have finished our pacing * send as most of the acks have come back if we are at the * right rate. This is a bit fuzzy since return path delay * can delay the acks, which is why we want to make sure we * have cwnd space to have a bit more than a max pace segments in flight. * * If we have not gotten our acks back we are pacing at too high a * rate delaying will not hurt and will bring our GP estimate down by * injecting the delay. If we don't do this we will send * 2 MSS out in response to the acks being clocked in which * defeats the point of hw-pacing (i.e. to help us get * larger TSO's out). */ len = 0; } } /* len will be >= 0 after this point. */ KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); rack_sndbuf_autoscale(rack); /* * Decide if we can use TCP Segmentation Offloading (if supported by * hardware). * * TSO may only be used if we are in a pure bulk sending state. The * presence of TCP-MD5, SACK retransmits, SACK advertizements and IP * options prevent using TSO. With TSO the TCP header is the same * (except for the sequence number) for all generated packets. This * makes it impossible to transmit any options which vary per * generated segment or packet. * * IPv4 handling has a clear separation of ip options and ip header * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does * the right thing below to provide length of just ip options and thus * checking for ipoptlen is enough to decide if ip options are present. */ ipoptlen = 0; #if defined(IPSEC) || defined(IPSEC_SUPPORT) /* * Pre-calculate here as we save another lookup into the darknesses * of IPsec that way and can actually decide if TSO is ok. */ #ifdef INET6 if (isipv6 && IPSEC_ENABLED(ipv6)) ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb); #ifdef INET else #endif #endif /* INET6 */ #ifdef INET if (IPSEC_ENABLED(ipv4)) ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb); #endif /* INET */ #endif #if defined(IPSEC) || defined(IPSEC_SUPPORT) ipoptlen += ipsec_optlen; #endif if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > segsiz && (tp->t_port == 0) && ((tp->t_flags & TF_SIGNATURE) == 0) && tp->rcv_numsacks == 0 && sack_rxmit == 0 && ipoptlen == 0) tso = 1; { uint32_t outstanding __unused; outstanding = tp->snd_max - tp->snd_una; if (tp->t_flags & TF_SENTFIN) { /* * If we sent a fin, snd_max is 1 higher than * snd_una */ outstanding--; } if (sack_rxmit) { if ((rsm->r_flags & RACK_HAS_FIN) == 0) flags &= ~TH_FIN; } else { if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + sbused(sb))) flags &= ~TH_FIN; } } recwin = lmin(lmax(sbspace(&so->so_rcv), 0), (long)TCP_MAXWIN << tp->rcv_scale); /* * Sender silly window avoidance. We transmit under the following * conditions when len is non-zero: * * - We have a full segment (or more with TSO) - This is the last * buffer in a write()/send() and we are either idle or running * NODELAY - we've timed out (e.g. persist timer) - we have more * then 1/2 the maximum send window's worth of data (receiver may be * limited the window size) - we need to retransmit */ if (len) { if (len >= segsiz) { goto send; } /* * NOTE! on localhost connections an 'ack' from the remote * end may occur synchronously with the output and cause us * to flush a buffer queued with moretocome. XXX * */ if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */ (idle || (tp->t_flags & TF_NODELAY)) && ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) && (tp->t_flags & TF_NOPUSH) == 0) { pass = 2; goto send; } if ((tp->snd_una == tp->snd_max) && len) { /* Nothing outstanding */ pass = 22; goto send; } if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0) { pass = 4; goto send; } if (SEQ_LT(tp->snd_nxt, tp->snd_max)) { /* retransmit case */ pass = 5; goto send; } if (sack_rxmit) { pass = 6; goto send; } if (((tp->snd_wnd - ctf_outstanding(tp)) < segsiz) && (ctf_outstanding(tp) < (segsiz * 2))) { /* * We have less than two MSS outstanding (delayed ack) * and our rwnd will not let us send a full sized * MSS. Lets go ahead and let this small segment * out because we want to try to have at least two * packets inflight to not be caught by delayed ack. */ pass = 12; goto send; } } /* * Sending of standalone window updates. * * Window updates are important when we close our window due to a * full socket buffer and are opening it again after the application * reads data from it. Once the window has opened again and the * remote end starts to send again the ACK clock takes over and * provides the most current window information. * * We must avoid the silly window syndrome whereas every read from * the receive buffer, no matter how small, causes a window update * to be sent. We also should avoid sending a flurry of window * updates when the socket buffer had queued a lot of data and the * application is doing small reads. * * Prevent a flurry of pointless window updates by only sending an * update when we can increase the advertized window by more than * 1/4th of the socket buffer capacity. When the buffer is getting * full or is very small be more aggressive and send an update * whenever we can increase by two mss sized segments. In all other * situations the ACK's to new incoming data will carry further * window increases. * * Don't send an independent window update if a delayed ACK is * pending (it will get piggy-backed on it) or the remote side * already has done a half-close and won't send more data. Skip * this if the connection is in T/TCP half-open state. */ if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) && !(tp->t_flags & TF_DELACK) && !TCPS_HAVERCVDFIN(tp->t_state)) { /* * "adv" is the amount we could increase the window, taking * into account that we are limited by TCP_MAXWIN << * tp->rcv_scale. */ int32_t adv; int oldwin; adv = recwin; if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { oldwin = (tp->rcv_adv - tp->rcv_nxt); if (adv > oldwin) adv -= oldwin; else { /* We can't increase the window */ adv = 0; } } else oldwin = 0; /* * If the new window size ends up being the same as or less * than the old size when it is scaled, then don't force * a window update. */ if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale) goto dontupdate; if (adv >= (int32_t)(2 * segsiz) && (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) || recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) || so->so_rcv.sb_hiwat <= 8 * segsiz)) { pass = 7; goto send; } if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) { pass = 23; goto send; } } dontupdate: /* * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW * is also a catch-all for the retransmit timer timeout case. */ if (tp->t_flags & TF_ACKNOW) { pass = 8; goto send; } if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) { pass = 9; goto send; } /* * If our state indicates that FIN should be sent and we have not * yet done so, then we need to send. */ if ((flags & TH_FIN) && (tp->snd_nxt == tp->snd_una)) { pass = 11; goto send; } /* * No reason to send a segment, just return. */ just_return: SOCKBUF_UNLOCK(sb); just_return_nolock: { int app_limited = CTF_JR_SENT_DATA; if (tot_len_this_send > 0) { /* Make sure snd_nxt is up to max */ rack->r_ctl.fsb.recwin = recwin; slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, NULL, segsiz); if ((error == 0) && rack_use_rfo && ((flags & (TH_SYN|TH_FIN)) == 0) && (ipoptlen == 0) && (tp->snd_nxt == tp->snd_max) && (tp->rcv_numsacks == 0) && rack->r_fsb_inited && TCPS_HAVEESTABLISHED(tp->t_state) && (rack->r_must_retran == 0) && ((tp->t_flags & TF_NEEDFIN) == 0) && (len > 0) && (orig_len > 0) && (orig_len > len) && ((orig_len - len) >= segsiz) && ((optlen == 0) || ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) { /* We can send at least one more MSS using our fsb */ rack->r_fast_output = 1; rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off); rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len; rack->r_ctl.fsb.tcp_flags = flags; rack->r_ctl.fsb.left_to_send = orig_len - len; if (hw_tls) rack->r_ctl.fsb.hw_tls = 1; else rack->r_ctl.fsb.hw_tls = 0; KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))), ("rack:%p left_to_send:%u sbavail:%u out:%u", rack, rack->r_ctl.fsb.left_to_send, sbavail(sb), (tp->snd_max - tp->snd_una))); if (rack->r_ctl.fsb.left_to_send < segsiz) rack->r_fast_output = 0; else { if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una))) rack->r_ctl.fsb.rfo_apply_push = 1; else rack->r_ctl.fsb.rfo_apply_push = 0; } } else rack->r_fast_output = 0; rack_log_fsb(rack, tp, so, flags, ipoptlen, orig_len, len, 0, 1, optlen, __LINE__, 1); if (SEQ_GT(tp->snd_max, tp->snd_nxt)) tp->snd_nxt = tp->snd_max; } else { int end_window = 0; uint32_t seq = tp->gput_ack; rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree); if (rsm) { /* * Mark the last sent that we just-returned (hinting * that delayed ack may play a role in any rtt measurement). */ rsm->r_just_ret = 1; } counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1); rack->r_ctl.rc_agg_delayed = 0; rack->r_early = 0; rack->r_late = 0; rack->r_ctl.rc_agg_early = 0; if ((ctf_outstanding(tp) + min(max(segsiz, (rack->r_ctl.rc_high_rwnd/2)), minseg)) >= tp->snd_wnd) { /* We are limited by the rwnd */ app_limited = CTF_JR_RWND_LIMITED; if (IN_FASTRECOVERY(tp->t_flags)) rack->r_ctl.rc_prr_sndcnt = 0; } else if (ctf_outstanding(tp) >= sbavail(sb)) { /* We are limited by whats available -- app limited */ app_limited = CTF_JR_APP_LIMITED; if (IN_FASTRECOVERY(tp->t_flags)) rack->r_ctl.rc_prr_sndcnt = 0; } else if ((idle == 0) && ((tp->t_flags & TF_NODELAY) == 0) && ((uint32_t)len + (uint32_t)sb_offset >= sbavail(sb)) && (len < segsiz)) { /* * No delay is not on and the * user is sending less than 1MSS. This * brings out SWS avoidance so we * don't send. Another app-limited case. */ app_limited = CTF_JR_APP_LIMITED; } else if (tp->t_flags & TF_NOPUSH) { /* * The user has requested no push of * the last segment and we are * at the last segment. Another app * limited case. */ app_limited = CTF_JR_APP_LIMITED; } else if ((ctf_outstanding(tp) + minseg) > cwnd_to_use) { /* Its the cwnd */ app_limited = CTF_JR_CWND_LIMITED; } else if (IN_FASTRECOVERY(tp->t_flags) && (rack->rack_no_prr == 0) && (rack->r_ctl.rc_prr_sndcnt < segsiz)) { app_limited = CTF_JR_PRR; } else { /* Now why here are we not sending? */ #ifdef NOW #ifdef INVARIANTS panic("rack:%p hit JR_ASSESSING case cwnd_to_use:%u?", rack, cwnd_to_use); #endif #endif app_limited = CTF_JR_ASSESSING; } /* * App limited in some fashion, for our pacing GP * measurements we don't want any gap (even cwnd). * Close down the measurement window. */ if (rack_cwnd_block_ends_measure && ((app_limited == CTF_JR_CWND_LIMITED) || (app_limited == CTF_JR_PRR))) { /* * The reason we are not sending is * the cwnd (or prr). We have been configured * to end the measurement window in * this case. */ end_window = 1; } else if (rack_rwnd_block_ends_measure && (app_limited == CTF_JR_RWND_LIMITED)) { /* * We are rwnd limited and have been * configured to end the measurement * window in this case. */ end_window = 1; } else if (app_limited == CTF_JR_APP_LIMITED) { /* * A true application limited period, we have * ran out of data. */ end_window = 1; } else if (app_limited == CTF_JR_ASSESSING) { /* * In the assessing case we hit the end of * the if/else and had no known reason * This will panic us under invariants.. * * If we get this out in logs we need to * investagate which reason we missed. */ end_window = 1; } if (end_window) { uint8_t log = 0; /* Adjust the Gput measurement */ if ((tp->t_flags & TF_GPUTINPROG) && SEQ_GT(tp->gput_ack, tp->snd_max)) { tp->gput_ack = tp->snd_max; if ((tp->gput_ack - tp->gput_seq) < (MIN_GP_WIN * segsiz)) { /* * There is not enough to measure. */ tp->t_flags &= ~TF_GPUTINPROG; rack_log_pacing_delay_calc(rack, (tp->gput_ack - tp->gput_seq) /*flex2*/, rack->r_ctl.rc_gp_srtt /*flex1*/, tp->gput_seq, 0, 0, 18, __LINE__, NULL, 0); } else log = 1; } /* Mark the last packet has app limited */ rsm = RB_MAX(rack_rb_tree_head, &rack->r_ctl.rc_mtree); if (rsm && ((rsm->r_flags & RACK_APP_LIMITED) == 0)) { if (rack->r_ctl.rc_app_limited_cnt == 0) rack->r_ctl.rc_end_appl = rack->r_ctl.rc_first_appl = rsm; else { /* * Go out to the end app limited and mark * this new one as next and move the end_appl up * to this guy. */ if (rack->r_ctl.rc_end_appl) rack->r_ctl.rc_end_appl->r_nseq_appl = rsm->r_start; rack->r_ctl.rc_end_appl = rsm; } rsm->r_flags |= RACK_APP_LIMITED; rack->r_ctl.rc_app_limited_cnt++; } if (log) rack_log_pacing_delay_calc(rack, rack->r_ctl.rc_app_limited_cnt, seq, tp->gput_ack, 0, 0, 4, __LINE__, NULL, 0); } } /* Check if we need to go into persists or not */ if ((tp->snd_max == tp->snd_una) && TCPS_HAVEESTABLISHED(tp->t_state) && sbavail(sb) && (sbavail(sb) > tp->snd_wnd) && (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), minseg))) { /* Yes lets make sure to move to persist before timer-start */ rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime); } rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, sup_rack); rack_log_type_just_return(rack, cts, tot_len_this_send, slot, hpts_calling, app_limited, cwnd_to_use); } #ifdef NETFLIX_SHARED_CWND if ((sbavail(sb) == 0) && rack->r_ctl.rc_scw) { tcp_shared_cwnd_idle(rack->r_ctl.rc_scw, rack->r_ctl.rc_scw_index); rack->rack_scwnd_is_idle = 1; } #endif #ifdef TCP_ACCOUNTING if (tot_len_this_send > 0) { crtsc = get_cyclecount(); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[SND_OUT_DATA]++; } counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[SND_OUT_DATA] += (crtsc - ts_val); } counter_u64_add(tcp_proc_time[SND_OUT_DATA], (crtsc - ts_val)); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) / segsiz); } counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) / segsiz)); } else { crtsc = get_cyclecount(); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[SND_LIMITED]++; } counter_u64_add(tcp_cnt_counters[SND_LIMITED], 1); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[SND_LIMITED] += (crtsc - ts_val); } counter_u64_add(tcp_proc_time[SND_LIMITED], (crtsc - ts_val)); } sched_unpin(); #endif return (0); send: if (rsm || sack_rxmit) counter_u64_add(rack_nfto_resend, 1); else counter_u64_add(rack_non_fto_send, 1); if ((flags & TH_FIN) && sbavail(sb)) { /* * We do not transmit a FIN * with data outstanding. We * need to make it so all data * is acked first. */ flags &= ~TH_FIN; } /* Enforce stack imposed max seg size if we have one */ if (rack->r_ctl.rc_pace_max_segs && (len > rack->r_ctl.rc_pace_max_segs)) { mark = 1; len = rack->r_ctl.rc_pace_max_segs; } SOCKBUF_LOCK_ASSERT(sb); if (len > 0) { if (len >= segsiz) tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT; else tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT; } /* * Before ESTABLISHED, force sending of initial options unless TCP * set not to do any options. NOTE: we assume that the IP/TCP header * plus TCP options always fit in a single mbuf, leaving room for a * maximum link header, i.e. max_linkhdr + sizeof (struct tcpiphdr) * + optlen <= MCLBYTES */ optlen = 0; #ifdef INET6 if (isipv6) hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); else #endif hdrlen = sizeof(struct tcpiphdr); /* * Compute options for segment. We only have to care about SYN and * established connection segments. Options for SYN-ACK segments * are handled in TCP syncache. */ to.to_flags = 0; if ((tp->t_flags & TF_NOOPT) == 0) { /* Maximum segment size. */ if (flags & TH_SYN) { tp->snd_nxt = tp->iss; to.to_mss = tcp_mssopt(&inp->inp_inc); if (tp->t_port) to.to_mss -= V_tcp_udp_tunneling_overhead; to.to_flags |= TOF_MSS; /* * On SYN or SYN|ACK transmits on TFO connections, * only include the TFO option if it is not a * retransmit, as the presence of the TFO option may * have caused the original SYN or SYN|ACK to have * been dropped by a middlebox. */ if (IS_FASTOPEN(tp->t_flags) && (tp->t_rxtshift == 0)) { if (tp->t_state == TCPS_SYN_RECEIVED) { to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN; to.to_tfo_cookie = (u_int8_t *)&tp->t_tfo_cookie.server; to.to_flags |= TOF_FASTOPEN; wanted_cookie = 1; } else if (tp->t_state == TCPS_SYN_SENT) { to.to_tfo_len = tp->t_tfo_client_cookie_len; to.to_tfo_cookie = tp->t_tfo_cookie.client; to.to_flags |= TOF_FASTOPEN; wanted_cookie = 1; /* * If we wind up having more data to * send with the SYN than can fit in * one segment, don't send any more * until the SYN|ACK comes back from * the other end. */ sendalot = 0; } } } /* Window scaling. */ if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) { to.to_wscale = tp->request_r_scale; to.to_flags |= TOF_SCALE; } /* Timestamps. */ if ((tp->t_flags & TF_RCVD_TSTMP) || ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) { to.to_tsval = ms_cts + tp->ts_offset; to.to_tsecr = tp->ts_recent; to.to_flags |= TOF_TS; } /* Set receive buffer autosizing timestamp. */ if (tp->rfbuf_ts == 0 && (so->so_rcv.sb_flags & SB_AUTOSIZE)) tp->rfbuf_ts = tcp_ts_getticks(); /* Selective ACK's. */ if (tp->t_flags & TF_SACK_PERMIT) { if (flags & TH_SYN) to.to_flags |= TOF_SACKPERM; else if (TCPS_HAVEESTABLISHED(tp->t_state) && tp->rcv_numsacks > 0) { to.to_flags |= TOF_SACK; to.to_nsacks = tp->rcv_numsacks; to.to_sacks = (u_char *)tp->sackblks; } } #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) /* TCP-MD5 (RFC2385). */ if (tp->t_flags & TF_SIGNATURE) to.to_flags |= TOF_SIGNATURE; #endif /* TCP_SIGNATURE */ /* Processing the options. */ hdrlen += optlen = tcp_addoptions(&to, opt); /* * If we wanted a TFO option to be added, but it was unable * to fit, ensure no data is sent. */ if (IS_FASTOPEN(tp->t_flags) && wanted_cookie && !(to.to_flags & TOF_FASTOPEN)) len = 0; } if (tp->t_port) { if (V_tcp_udp_tunneling_port == 0) { /* The port was removed?? */ SOCKBUF_UNLOCK(&so->so_snd); #ifdef TCP_ACCOUNTING crtsc = get_cyclecount(); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[SND_OUT_FAIL]++; } counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val); } counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val)); sched_unpin(); #endif return (EHOSTUNREACH); } hdrlen += sizeof(struct udphdr); } #ifdef INET6 if (isipv6) ipoptlen = ip6_optlen(tp->t_inpcb); else #endif if (tp->t_inpcb->inp_options) ipoptlen = tp->t_inpcb->inp_options->m_len - offsetof(struct ipoption, ipopt_list); else ipoptlen = 0; #if defined(IPSEC) || defined(IPSEC_SUPPORT) ipoptlen += ipsec_optlen; #endif /* * Adjust data length if insertion of options will bump the packet * length beyond the t_maxseg length. Clear the FIN bit because we * cut off the tail of the segment. */ if (len + optlen + ipoptlen > tp->t_maxseg) { if (tso) { uint32_t if_hw_tsomax; uint32_t moff; int32_t max_len; /* extract TSO information */ if_hw_tsomax = tp->t_tsomax; if_hw_tsomaxsegcount = tp->t_tsomaxsegcount; if_hw_tsomaxsegsize = tp->t_tsomaxsegsize; KASSERT(ipoptlen == 0, ("%s: TSO can't do IP options", __func__)); /* * Check if we should limit by maximum payload * length: */ if (if_hw_tsomax != 0) { /* compute maximum TSO length */ max_len = (if_hw_tsomax - hdrlen - max_linkhdr); if (max_len <= 0) { len = 0; } else if (len > max_len) { sendalot = 1; len = max_len; mark = 2; } } /* * Prevent the last segment from being fractional * unless the send sockbuf can be emptied: */ max_len = (tp->t_maxseg - optlen); if ((sb_offset + len) < sbavail(sb)) { moff = len % (u_int)max_len; if (moff != 0) { mark = 3; len -= moff; } } /* * In case there are too many small fragments don't * use TSO: */ if (len <= segsiz) { mark = 4; tso = 0; } /* * Send the FIN in a separate segment after the bulk * sending is done. We don't trust the TSO * implementations to clear the FIN flag on all but * the last segment. */ if (tp->t_flags & TF_NEEDFIN) { sendalot = 4; } } else { mark = 5; if (optlen + ipoptlen >= tp->t_maxseg) { /* * Since we don't have enough space to put * the IP header chain and the TCP header in * one packet as required by RFC 7112, don't * send it. Also ensure that at least one * byte of the payload can be put into the * TCP segment. */ SOCKBUF_UNLOCK(&so->so_snd); error = EMSGSIZE; sack_rxmit = 0; goto out; } len = tp->t_maxseg - optlen - ipoptlen; sendalot = 5; } } else { tso = 0; mark = 6; } KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET, ("%s: len > IP_MAXPACKET", __func__)); #ifdef DIAGNOSTIC #ifdef INET6 if (max_linkhdr + hdrlen > MCLBYTES) #else if (max_linkhdr + hdrlen > MHLEN) #endif panic("tcphdr too big"); #endif /* * This KASSERT is here to catch edge cases at a well defined place. * Before, those had triggered (random) panic conditions further * down. */ KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__)); if ((len == 0) && (flags & TH_FIN) && (sbused(sb))) { /* * We have outstanding data, don't send a fin by itself!. */ goto just_return; } /* * Grab a header mbuf, attaching a copy of data to be transmitted, * and initialize the header from the template for sends on this * connection. */ hw_tls = (sb->sb_flags & SB_TLS_IFNET) != 0; if (len) { uint32_t max_val; uint32_t moff; if (rack->r_ctl.rc_pace_max_segs) max_val = rack->r_ctl.rc_pace_max_segs; else if (rack->rc_user_set_max_segs) max_val = rack->rc_user_set_max_segs * segsiz; else max_val = len; /* * We allow a limit on sending with hptsi. */ if (len > max_val) { mark = 7; len = max_val; } #ifdef INET6 if (MHLEN < hdrlen + max_linkhdr) m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); else #endif m = m_gethdr(M_NOWAIT, MT_DATA); if (m == NULL) { SOCKBUF_UNLOCK(sb); error = ENOBUFS; sack_rxmit = 0; goto out; } m->m_data += max_linkhdr; m->m_len = hdrlen; /* * Start the m_copy functions from the closest mbuf to the * sb_offset in the socket buffer chain. */ mb = sbsndptr_noadv(sb, sb_offset, &moff); s_mb = mb; s_moff = moff; if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) { m_copydata(mb, moff, (int)len, mtod(m, caddr_t)+hdrlen); if (SEQ_LT(tp->snd_nxt, tp->snd_max)) sbsndptr_adv(sb, mb, len); m->m_len += len; } else { struct sockbuf *msb; if (SEQ_LT(tp->snd_nxt, tp->snd_max)) msb = NULL; else msb = sb; m->m_next = tcp_m_copym( mb, moff, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb, ((rsm == NULL) ? hw_tls : 0) #ifdef NETFLIX_COPY_ARGS , &s_mb, &s_moff #endif ); if (len <= (tp->t_maxseg - optlen)) { /* * Must have ran out of mbufs for the copy * shorten it to no longer need tso. Lets * not put on sendalot since we are low on * mbufs. */ tso = 0; } if (m->m_next == NULL) { SOCKBUF_UNLOCK(sb); (void)m_free(m); error = ENOBUFS; sack_rxmit = 0; goto out; } } if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) { if (rsm && (rsm->r_flags & RACK_TLP)) { /* * TLP should not count in retran count, but * in its own bin */ counter_u64_add(rack_tlp_retran, 1); counter_u64_add(rack_tlp_retran_bytes, len); } else { tp->t_sndrexmitpack++; KMOD_TCPSTAT_INC(tcps_sndrexmitpack); KMOD_TCPSTAT_ADD(tcps_sndrexmitbyte, len); } #ifdef STATS stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB, len); #endif } else { KMOD_TCPSTAT_INC(tcps_sndpack); KMOD_TCPSTAT_ADD(tcps_sndbyte, len); #ifdef STATS stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB, len); #endif } /* * If we're sending everything we've got, set PUSH. (This * will keep happy those implementations which only give * data to the user when a buffer fills or a PUSH comes in.) */ if (sb_offset + len == sbused(sb) && sbused(sb) && !(flags & TH_SYN)) { flags |= TH_PUSH; add_flag |= RACK_HAD_PUSH; } SOCKBUF_UNLOCK(sb); } else { SOCKBUF_UNLOCK(sb); if (tp->t_flags & TF_ACKNOW) KMOD_TCPSTAT_INC(tcps_sndacks); else if (flags & (TH_SYN | TH_FIN | TH_RST)) KMOD_TCPSTAT_INC(tcps_sndctrl); else KMOD_TCPSTAT_INC(tcps_sndwinup); m = m_gethdr(M_NOWAIT, MT_DATA); if (m == NULL) { error = ENOBUFS; sack_rxmit = 0; goto out; } #ifdef INET6 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) && MHLEN >= hdrlen) { M_ALIGN(m, hdrlen); } else #endif m->m_data += max_linkhdr; m->m_len = hdrlen; } SOCKBUF_UNLOCK_ASSERT(sb); m->m_pkthdr.rcvif = (struct ifnet *)0; #ifdef MAC mac_inpcb_create_mbuf(inp, m); #endif if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) { #ifdef INET6 if (isipv6) ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr; else #endif /* INET6 */ ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr; th = rack->r_ctl.fsb.th; udp = rack->r_ctl.fsb.udp; if (udp) { #ifdef INET6 if (isipv6) ulen = hdrlen + len - sizeof(struct ip6_hdr); else #endif /* INET6 */ ulen = hdrlen + len - sizeof(struct ip); udp->uh_ulen = htons(ulen); } } else { #ifdef INET6 if (isipv6) { ip6 = mtod(m, struct ip6_hdr *); if (tp->t_port) { udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr)); udp->uh_sport = htons(V_tcp_udp_tunneling_port); udp->uh_dport = tp->t_port; ulen = hdrlen + len - sizeof(struct ip6_hdr); udp->uh_ulen = htons(ulen); th = (struct tcphdr *)(udp + 1); } else th = (struct tcphdr *)(ip6 + 1); tcpip_fillheaders(inp, tp->t_port, ip6, th); } else #endif /* INET6 */ { ip = mtod(m, struct ip *); if (tp->t_port) { udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip)); udp->uh_sport = htons(V_tcp_udp_tunneling_port); udp->uh_dport = tp->t_port; ulen = hdrlen + len - sizeof(struct ip); udp->uh_ulen = htons(ulen); th = (struct tcphdr *)(udp + 1); } else th = (struct tcphdr *)(ip + 1); tcpip_fillheaders(inp, tp->t_port, ip, th); } } /* * Fill in fields, remembering maximum advertised window for use in * delaying messages about window sizes. If resending a FIN, be sure * not to use a new sequence number. */ if (flags & TH_FIN && tp->t_flags & TF_SENTFIN && tp->snd_nxt == tp->snd_max) tp->snd_nxt--; /* * If we are starting a connection, send ECN setup SYN packet. If we * are on a retransmit, we may resend those bits a number of times * as per RFC 3168. */ if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) { flags |= tcp_ecn_output_syn_sent(tp); } /* Also handle parallel SYN for ECN */ if (TCPS_HAVERCVDSYN(tp->t_state) && (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) { int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit); if ((tp->t_state == TCPS_SYN_RECEIVED) && (tp->t_flags2 & TF2_ECN_SND_ECE)) tp->t_flags2 &= ~TF2_ECN_SND_ECE; #ifdef INET6 if (isipv6) { ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << 20); ip6->ip6_flow |= htonl(ect << 20); } else #endif { ip->ip_tos &= ~IPTOS_ECN_MASK; ip->ip_tos |= ect; } } /* * If we are doing retransmissions, then snd_nxt will not reflect * the first unsent octet. For ACK only packets, we do not want the * sequence number of the retransmitted packet, we want the sequence * number of the next unsent octet. So, if there is no data (and no * SYN or FIN), use snd_max instead of snd_nxt when filling in * ti_seq. But if we are in persist state, snd_max might reflect * one byte beyond the right edge of the window, so use snd_nxt in * that case, since we know we aren't doing a retransmission. * (retransmit and persist are mutually exclusive...) */ if (sack_rxmit == 0) { if (len || (flags & (TH_SYN | TH_FIN))) { th->th_seq = htonl(tp->snd_nxt); rack_seq = tp->snd_nxt; } else { th->th_seq = htonl(tp->snd_max); rack_seq = tp->snd_max; } } else { th->th_seq = htonl(rsm->r_start); rack_seq = rsm->r_start; } th->th_ack = htonl(tp->rcv_nxt); tcp_set_flags(th, flags); /* * Calculate receive window. Don't shrink window, but avoid silly * window syndrome. * If a RST segment is sent, advertise a window of zero. */ if (flags & TH_RST) { recwin = 0; } else { if (recwin < (long)(so->so_rcv.sb_hiwat / 4) && recwin < (long)segsiz) { recwin = 0; } if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) && recwin < (long)(tp->rcv_adv - tp->rcv_nxt)) recwin = (long)(tp->rcv_adv - tp->rcv_nxt); } /* * According to RFC1323 the window field in a SYN (i.e., a or * ) segment itself is never scaled. The case is * handled in syncache. */ if (flags & TH_SYN) th->th_win = htons((u_short) (min(sbspace(&so->so_rcv), TCP_MAXWIN))); else { /* Avoid shrinking window with window scaling. */ recwin = roundup2(recwin, 1 << tp->rcv_scale); th->th_win = htons((u_short)(recwin >> tp->rcv_scale)); } /* * Adjust the RXWIN0SENT flag - indicate that we have advertised a 0 * window. This may cause the remote transmitter to stall. This * flag tells soreceive() to disable delayed acknowledgements when * draining the buffer. This can occur if the receiver is * attempting to read more data than can be buffered prior to * transmitting on the connection. */ if (th->th_win == 0) { tp->t_sndzerowin++; tp->t_flags |= TF_RXWIN0SENT; } else tp->t_flags &= ~TF_RXWIN0SENT; tp->snd_up = tp->snd_una; /* drag it along, its deprecated */ /* Now are we using fsb?, if so copy the template data to the mbuf */ if ((ipoptlen == 0) && (rack->r_ctl.fsb.tcp_ip_hdr) && rack->r_fsb_inited) { uint8_t *cpto; cpto = mtod(m, uint8_t *); memcpy(cpto, rack->r_ctl.fsb.tcp_ip_hdr, rack->r_ctl.fsb.tcp_ip_hdr_len); /* * We have just copied in: * IP/IP6 * * tcphdr (no options) * * We need to grab the correct pointers into the mbuf * for both the tcp header, and possibly the udp header (if tunneling). * We do this by using the offset in the copy buffer and adding it * to the mbuf base pointer (cpto). */ #ifdef INET6 if (isipv6) ip6 = mtod(m, struct ip6_hdr *); else #endif /* INET6 */ ip = mtod(m, struct ip *); th = (struct tcphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.th - rack->r_ctl.fsb.tcp_ip_hdr)); /* If we have a udp header lets set it into the mbuf as well */ if (udp) udp = (struct udphdr *)(cpto + ((uint8_t *)rack->r_ctl.fsb.udp - rack->r_ctl.fsb.tcp_ip_hdr)); } #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) if (to.to_flags & TOF_SIGNATURE) { /* * Calculate MD5 signature and put it into the place * determined before. * NOTE: since TCP options buffer doesn't point into * mbuf's data, calculate offset and use it. */ if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th, (u_char *)(th + 1) + (to.to_signature - opt)) != 0) { /* * Do not send segment if the calculation of MD5 * digest has failed. */ goto out; } } #endif if (optlen) { bcopy(opt, th + 1, optlen); th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; } /* * Put TCP length in extended header, and then checksum extended * header and data. */ m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */ #ifdef INET6 if (isipv6) { /* * ip6_plen is not need to be filled now, and will be filled * in ip6_output. */ if (tp->t_port) { m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0); th->th_sum = htons(0); UDPSTAT_INC(udps_opackets); } else { m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP, 0); } } #endif #if defined(INET6) && defined(INET) else #endif #ifdef INET { if (tp->t_port) { m->m_pkthdr.csum_flags = CSUM_UDP; m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); udp->uh_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); th->th_sum = htons(0); UDPSTAT_INC(udps_opackets); } else { m->m_pkthdr.csum_flags = CSUM_TCP; m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen)); } /* IP version must be set here for ipv4/ipv6 checking later */ KASSERT(ip->ip_v == IPVERSION, ("%s: IP version incorrect: %d", __func__, ip->ip_v)); } #endif /* * Enable TSO and specify the size of the segments. The TCP pseudo * header checksum is always provided. XXX: Fixme: This is currently * not the case for IPv6. */ if (tso) { KASSERT(len > tp->t_maxseg - optlen, ("%s: len <= tso_segsz", __func__)); m->m_pkthdr.csum_flags |= CSUM_TSO; m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen; } KASSERT(len + hdrlen == m_length(m, NULL), ("%s: mbuf chain different than expected: %d + %u != %u", __func__, len, hdrlen, m_length(m, NULL))); #ifdef TCP_HHOOK /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */ hhook_run_tcp_est_out(tp, th, &to, len, tso); #endif /* We're getting ready to send; log now. */ if (tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = tcp_in_hpts(rack->rc_inp); if (rack->rack_no_prr) log.u_bbr.flex1 = 0; else log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.flex2 = rack->r_ctl.rc_pace_min_segs; log.u_bbr.flex3 = rack->r_ctl.rc_pace_max_segs; log.u_bbr.flex4 = orig_len; /* Save off the early/late values */ log.u_bbr.flex6 = rack->r_ctl.rc_agg_early; log.u_bbr.applimited = rack->r_ctl.rc_agg_delayed; log.u_bbr.bw_inuse = rack_get_bw(rack); log.u_bbr.flex8 = 0; if (rsm) { if (rsm->r_flags & RACK_RWND_COLLAPSED) { rack_log_collapse(rack, rsm->r_start, rsm->r_end, 0, __LINE__, 5, rsm->r_flags, rsm); counter_u64_add(rack_collapsed_win_rxt, 1); counter_u64_add(rack_collapsed_win_rxt_bytes, (rsm->r_end - rsm->r_start)); } if (doing_tlp) log.u_bbr.flex8 = 2; else log.u_bbr.flex8 = 1; } else { if (doing_tlp) log.u_bbr.flex8 = 3; else log.u_bbr.flex8 = 0; } log.u_bbr.pacing_gain = rack_get_output_gain(rack, rsm); log.u_bbr.flex7 = mark; log.u_bbr.flex7 <<= 8; log.u_bbr.flex7 |= pass; log.u_bbr.pkts_out = tp->t_maxseg; log.u_bbr.timeStamp = cts; log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); log.u_bbr.lt_epoch = cwnd_to_use; log.u_bbr.delivered = sendalot; lgb = tcp_log_event_(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK, len, &log, false, NULL, NULL, 0, &tv); } else lgb = NULL; /* * Fill in IP length and desired time to live and send to IP level. * There should be a better way to handle ttl and tos; we could keep * them in the template, but need a way to checksum without them. */ /* * m->m_pkthdr.len should have been set before cksum calcuration, * because in6_cksum() need it. */ #ifdef INET6 if (isipv6) { /* * we separately set hoplimit for every segment, since the * user might want to change the value via setsockopt. Also, * desired default hop limit might be changed via Neighbor * Discovery. */ rack->r_ctl.fsb.hoplimit = ip6->ip6_hlim = in6_selecthlim(inp, NULL); /* * Set the packet size here for the benefit of DTrace * probes. ip6_output() will set it properly; it's supposed * to include the option header lengths as well. */ ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6)); if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) tp->t_flags2 |= TF2_PLPMTU_PMTUD; else tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; if (tp->t_state == TCPS_SYN_SENT) TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th); TCP_PROBE5(send, NULL, tp, ip6, tp, th); /* TODO: IPv6 IP6TOS_ECT bit on */ error = ip6_output(m, #if defined(IPSEC) || defined(IPSEC_SUPPORT) inp->in6p_outputopts, #else NULL, #endif &inp->inp_route6, ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), NULL, NULL, inp); if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL) mtu = inp->inp_route6.ro_nh->nh_mtu; } #endif /* INET6 */ #if defined(INET) && defined(INET6) else #endif #ifdef INET { ip->ip_len = htons(m->m_pkthdr.len); #ifdef INET6 if (inp->inp_vflag & INP_IPV6PROTO) ip->ip_ttl = in6_selecthlim(inp, NULL); #endif /* INET6 */ rack->r_ctl.fsb.hoplimit = ip->ip_ttl; /* * If we do path MTU discovery, then we set DF on every * packet. This might not be the best thing to do according * to RFC3390 Section 2. However the tcp hostcache migitates * the problem so it affects only the first tcp connection * with a host. * * NB: Don't set DF on small MTU/MSS to have a safe * fallback. */ if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) { tp->t_flags2 |= TF2_PLPMTU_PMTUD; if (tp->t_port == 0 || len < V_tcp_minmss) { ip->ip_off |= htons(IP_DF); } } else { tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; } if (tp->t_state == TCPS_SYN_SENT) TCP_PROBE5(connect__request, NULL, tp, ip, tp, th); TCP_PROBE5(send, NULL, tp, ip, tp, th); error = ip_output(m, #if defined(IPSEC) || defined(IPSEC_SUPPORT) inp->inp_options, #else NULL, #endif &inp->inp_route, ((rsm || sack_rxmit) ? IP_NO_SND_TAG_RL : 0), 0, inp); if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL) mtu = inp->inp_route.ro_nh->nh_mtu; } #endif /* INET */ out: if (lgb) { lgb->tlb_errno = error; lgb = NULL; } /* * In transmit state, time the transmission and arrange for the * retransmit. In persist state, just set snd_max. */ if (error == 0) { tcp_account_for_send(tp, len, (rsm != NULL), doing_tlp, hw_tls); if (rsm && doing_tlp) { rack->rc_last_sent_tlp_past_cumack = 0; rack->rc_last_sent_tlp_seq_valid = 1; rack->r_ctl.last_sent_tlp_seq = rsm->r_start; rack->r_ctl.last_sent_tlp_len = rsm->r_end - rsm->r_start; } rack->forced_ack = 0; /* If we send something zap the FA flag */ if (rsm && (doing_tlp == 0)) { /* Set we retransmitted */ rack->rc_gp_saw_rec = 1; } else { if (cwnd_to_use > tp->snd_ssthresh) { /* Set we sent in CA */ rack->rc_gp_saw_ca = 1; } else { /* Set we sent in SS */ rack->rc_gp_saw_ss = 1; } } if (TCPS_HAVEESTABLISHED(tp->t_state) && (tp->t_flags & TF_SACK_PERMIT) && tp->rcv_numsacks > 0) tcp_clean_dsack_blocks(tp); tot_len_this_send += len; if (len == 0) counter_u64_add(rack_out_size[TCP_MSS_ACCT_SNDACK], 1); else if (len == 1) { counter_u64_add(rack_out_size[TCP_MSS_ACCT_PERSIST], 1); } else if (len > 1) { int idx; idx = (len / segsiz) + 3; if (idx >= TCP_MSS_ACCT_ATIMER) counter_u64_add(rack_out_size[(TCP_MSS_ACCT_ATIMER-1)], 1); else counter_u64_add(rack_out_size[idx], 1); } } if ((rack->rack_no_prr == 0) && sub_from_prr && (error == 0)) { if (rack->r_ctl.rc_prr_sndcnt >= len) rack->r_ctl.rc_prr_sndcnt -= len; else rack->r_ctl.rc_prr_sndcnt = 0; } sub_from_prr = 0; if (doing_tlp) { /* Make sure the TLP is added */ add_flag |= RACK_TLP; } else if (rsm) { /* If its a resend without TLP then it must not have the flag */ rsm->r_flags &= ~RACK_TLP; } rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error, rack_to_usec_ts(&tv), rsm, add_flag, s_mb, s_moff, hw_tls); if ((error == 0) && (len > 0) && (tp->snd_una == tp->snd_max)) rack->r_ctl.rc_tlp_rxt_last_time = cts; { tcp_seq startseq = tp->snd_nxt; /* Track our lost count */ if (rsm && (doing_tlp == 0)) rack->r_ctl.rc_loss_count += rsm->r_end - rsm->r_start; /* * Advance snd_nxt over sequence space of this segment. */ if (error) /* We don't log or do anything with errors */ goto nomore; if (doing_tlp == 0) { if (rsm == NULL) { /* * Not a retransmission of some * sort, new data is going out so * clear our TLP count and flag. */ rack->rc_tlp_in_progress = 0; rack->r_ctl.rc_tlp_cnt_out = 0; } } else { /* * We have just sent a TLP, mark that it is true * and make sure our in progress is set so we * continue to check the count. */ rack->rc_tlp_in_progress = 1; rack->r_ctl.rc_tlp_cnt_out++; } if (flags & (TH_SYN | TH_FIN)) { if (flags & TH_SYN) tp->snd_nxt++; if (flags & TH_FIN) { tp->snd_nxt++; tp->t_flags |= TF_SENTFIN; } } /* In the ENOBUFS case we do *not* update snd_max */ if (sack_rxmit) goto nomore; tp->snd_nxt += len; if (SEQ_GT(tp->snd_nxt, tp->snd_max)) { if (tp->snd_una == tp->snd_max) { /* * Update the time we just added data since * none was outstanding. */ rack_log_progress_event(rack, tp, ticks, PROGRESS_START, __LINE__); tp->t_acktime = ticks; } tp->snd_max = tp->snd_nxt; /* * Time this transmission if not a retransmission and * not currently timing anything. * This is only relevant in case of switching back to * the base stack. */ if (tp->t_rtttime == 0) { tp->t_rtttime = ticks; tp->t_rtseq = startseq; KMOD_TCPSTAT_INC(tcps_segstimed); } if (len && ((tp->t_flags & TF_GPUTINPROG) == 0)) rack_start_gp_measurement(tp, rack, startseq, sb_offset); } /* * If we are doing FO we need to update the mbuf position and subtract * this happens when the peer sends us duplicate information and * we thus want to send a DSACK. * * XXXRRS: This brings to mind a ?, when we send a DSACK block is TSO * turned off? If not then we are going to echo multiple DSACK blocks * out (with the TSO), which we should not be doing. */ if (rack->r_fast_output && len) { if (rack->r_ctl.fsb.left_to_send > len) rack->r_ctl.fsb.left_to_send -= len; else rack->r_ctl.fsb.left_to_send = 0; if (rack->r_ctl.fsb.left_to_send < segsiz) rack->r_fast_output = 0; if (rack->r_fast_output) { rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off); rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len; } } } nomore: if (error) { rack->r_ctl.rc_agg_delayed = 0; rack->r_early = 0; rack->r_late = 0; rack->r_ctl.rc_agg_early = 0; SOCKBUF_UNLOCK_ASSERT(sb); /* Check gotos. */ /* * Failures do not advance the seq counter above. For the * case of ENOBUFS we will fall out and retry in 1ms with * the hpts. Everything else will just have to retransmit * with the timer. * * In any case, we do not want to loop around for another * send without a good reason. */ sendalot = 0; switch (error) { case EPERM: tp->t_softerror = error; #ifdef TCP_ACCOUNTING crtsc = get_cyclecount(); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[SND_OUT_FAIL]++; } counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val); } counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val)); sched_unpin(); #endif return (error); case ENOBUFS: /* * Pace us right away to retry in a some * time */ if (rack->r_ctl.crte != NULL) { rack_trace_point(rack, RACK_TP_HWENOBUF); } else rack_trace_point(rack, RACK_TP_ENOBUF); slot = ((1 + rack->rc_enobuf) * HPTS_USEC_IN_MSEC); if (rack->rc_enobuf < 0x7f) rack->rc_enobuf++; if (slot < (10 * HPTS_USEC_IN_MSEC)) slot = 10 * HPTS_USEC_IN_MSEC; if (rack->r_ctl.crte != NULL) { counter_u64_add(rack_saw_enobuf_hw, 1); tcp_rl_log_enobuf(rack->r_ctl.crte); } counter_u64_add(rack_saw_enobuf, 1); goto enobufs; case EMSGSIZE: /* * For some reason the interface we used initially * to send segments changed to another or lowered * its MTU. If TSO was active we either got an * interface without TSO capabilits or TSO was * turned off. If we obtained mtu from ip_output() * then update it and try again. */ if (tso) tp->t_flags &= ~TF_TSO; if (mtu != 0) { tcp_mss_update(tp, -1, mtu, NULL, NULL); goto again; } slot = 10 * HPTS_USEC_IN_MSEC; rack_start_hpts_timer(rack, tp, cts, slot, 0, 0); #ifdef TCP_ACCOUNTING crtsc = get_cyclecount(); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[SND_OUT_FAIL]++; } counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val); } counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val)); sched_unpin(); #endif return (error); case ENETUNREACH: counter_u64_add(rack_saw_enetunreach, 1); case EHOSTDOWN: case EHOSTUNREACH: case ENETDOWN: if (TCPS_HAVERCVDSYN(tp->t_state)) { tp->t_softerror = error; } /* FALLTHROUGH */ default: slot = 10 * HPTS_USEC_IN_MSEC; rack_start_hpts_timer(rack, tp, cts, slot, 0, 0); #ifdef TCP_ACCOUNTING crtsc = get_cyclecount(); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[SND_OUT_FAIL]++; } counter_u64_add(tcp_cnt_counters[SND_OUT_FAIL], 1); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[SND_OUT_FAIL] += (crtsc - ts_val); } counter_u64_add(tcp_proc_time[SND_OUT_FAIL], (crtsc - ts_val)); sched_unpin(); #endif return (error); } } else { rack->rc_enobuf = 0; if (IN_FASTRECOVERY(tp->t_flags) && rsm) rack->r_ctl.retran_during_recovery += len; } KMOD_TCPSTAT_INC(tcps_sndtotal); /* * Data sent (as far as we can tell). If this advertises a larger * window than any other segment, then remember the size of the * advertised window. Any pending ACK has now been sent. */ if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv)) tp->rcv_adv = tp->rcv_nxt + recwin; tp->last_ack_sent = tp->rcv_nxt; tp->t_flags &= ~(TF_ACKNOW | TF_DELACK); enobufs: if (sendalot) { /* Do we need to turn off sendalot? */ if (rack->r_ctl.rc_pace_max_segs && (tot_len_this_send >= rack->r_ctl.rc_pace_max_segs)) { /* We hit our max. */ sendalot = 0; } else if ((rack->rc_user_set_max_segs) && (tot_len_this_send >= (rack->rc_user_set_max_segs * segsiz))) { /* We hit the user defined max */ sendalot = 0; } } if ((error == 0) && (flags & TH_FIN)) tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_FIN); if (flags & TH_RST) { /* * We don't send again after sending a RST. */ slot = 0; sendalot = 0; if (error == 0) tcp_log_end_status(tp, TCP_EI_STATUS_SERVER_RST); } else if ((slot == 0) && (sendalot == 0) && tot_len_this_send) { /* * Get our pacing rate, if an error * occurred in sending (ENOBUF) we would * hit the else if with slot preset. Other * errors return. */ slot = rack_get_pacing_delay(rack, tp, tot_len_this_send, rsm, segsiz); } if (rsm && (rsm->r_flags & RACK_HAS_SYN) == 0 && rack->use_rack_rr) { /* Its a retransmit and we use the rack cheat? */ if ((slot == 0) || (rack->rc_always_pace == 0) || (rack->r_rr_config == 1)) { /* * We have no pacing set or we * are using old-style rack or * we are overridden to use the old 1ms pacing. */ slot = rack->r_ctl.rc_min_to; } } /* We have sent clear the flag */ rack->r_ent_rec_ns = 0; if (rack->r_must_retran) { if (rsm) { rack->r_ctl.rc_out_at_rto -= (rsm->r_end - rsm->r_start); if (SEQ_GEQ(rsm->r_end, rack->r_ctl.rc_snd_max_at_rto)) { /* * We have retransmitted all. */ rack->r_must_retran = 0; rack->r_ctl.rc_out_at_rto = 0; } } else if (SEQ_GEQ(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) { /* * Sending new data will also kill * the loop. */ rack->r_must_retran = 0; rack->r_ctl.rc_out_at_rto = 0; } } rack->r_ctl.fsb.recwin = recwin; if ((tp->t_flags & (TF_WASCRECOVERY|TF_WASFRECOVERY)) && SEQ_GT(tp->snd_max, rack->r_ctl.rc_snd_max_at_rto)) { /* * We hit an RTO and now have past snd_max at the RTO * clear all the WAS flags. */ tp->t_flags &= ~(TF_WASCRECOVERY|TF_WASFRECOVERY); } if (slot) { /* set the rack tcb into the slot N */ if ((error == 0) && rack_use_rfo && ((flags & (TH_SYN|TH_FIN)) == 0) && (rsm == NULL) && (tp->snd_nxt == tp->snd_max) && (ipoptlen == 0) && (tp->rcv_numsacks == 0) && rack->r_fsb_inited && TCPS_HAVEESTABLISHED(tp->t_state) && (rack->r_must_retran == 0) && ((tp->t_flags & TF_NEEDFIN) == 0) && (len > 0) && (orig_len > 0) && (orig_len > len) && ((orig_len - len) >= segsiz) && ((optlen == 0) || ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) { /* We can send at least one more MSS using our fsb */ rack->r_fast_output = 1; rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off); rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len; rack->r_ctl.fsb.tcp_flags = flags; rack->r_ctl.fsb.left_to_send = orig_len - len; if (hw_tls) rack->r_ctl.fsb.hw_tls = 1; else rack->r_ctl.fsb.hw_tls = 0; KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))), ("rack:%p left_to_send:%u sbavail:%u out:%u", rack, rack->r_ctl.fsb.left_to_send, sbavail(sb), (tp->snd_max - tp->snd_una))); if (rack->r_ctl.fsb.left_to_send < segsiz) rack->r_fast_output = 0; else { if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una))) rack->r_ctl.fsb.rfo_apply_push = 1; else rack->r_ctl.fsb.rfo_apply_push = 0; } } else rack->r_fast_output = 0; rack_log_fsb(rack, tp, so, flags, ipoptlen, orig_len, len, error, (rsm == NULL), optlen, __LINE__, 2); } else if (sendalot) { int ret; sack_rxmit = 0; if ((error == 0) && rack_use_rfo && ((flags & (TH_SYN|TH_FIN)) == 0) && (rsm == NULL) && (ipoptlen == 0) && (tp->rcv_numsacks == 0) && (tp->snd_nxt == tp->snd_max) && (rack->r_must_retran == 0) && rack->r_fsb_inited && TCPS_HAVEESTABLISHED(tp->t_state) && ((tp->t_flags & TF_NEEDFIN) == 0) && (len > 0) && (orig_len > 0) && (orig_len > len) && ((orig_len - len) >= segsiz) && ((optlen == 0) || ((optlen == TCPOLEN_TSTAMP_APPA) && (to.to_flags & TOF_TS)))) { /* we can use fast_output for more */ rack->r_fast_output = 1; rack->r_ctl.fsb.m = sbsndmbuf(sb, (tp->snd_max - tp->snd_una), &rack->r_ctl.fsb.off); rack->r_ctl.fsb.o_m_len = rack->r_ctl.fsb.m->m_len; rack->r_ctl.fsb.tcp_flags = flags; rack->r_ctl.fsb.left_to_send = orig_len - len; if (hw_tls) rack->r_ctl.fsb.hw_tls = 1; else rack->r_ctl.fsb.hw_tls = 0; KASSERT((rack->r_ctl.fsb.left_to_send <= (sbavail(sb) - (tp->snd_max - tp->snd_una))), ("rack:%p left_to_send:%u sbavail:%u out:%u", rack, rack->r_ctl.fsb.left_to_send, sbavail(sb), (tp->snd_max - tp->snd_una))); if (rack->r_ctl.fsb.left_to_send < segsiz) { rack->r_fast_output = 0; } if (rack->r_fast_output) { if (rack->r_ctl.fsb.left_to_send == (sbavail(sb) - (tp->snd_max - tp->snd_una))) rack->r_ctl.fsb.rfo_apply_push = 1; else rack->r_ctl.fsb.rfo_apply_push = 0; rack_log_fsb(rack, tp, so, flags, ipoptlen, orig_len, len, error, (rsm == NULL), optlen, __LINE__, 3); error = 0; ret = rack_fast_output(tp, rack, ts_val, cts, ms_cts, &tv, tot_len_this_send, &error); if (ret >= 0) return (ret); else if (error) goto nomore; } } goto again; } /* Assure when we leave that snd_nxt will point to top */ if (SEQ_GT(tp->snd_max, tp->snd_nxt)) tp->snd_nxt = tp->snd_max; rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0); #ifdef TCP_ACCOUNTING crtsc = get_cyclecount() - ts_val; if (tot_len_this_send) { if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[SND_OUT_DATA]++; } counter_u64_add(tcp_cnt_counters[SND_OUT_DATA], 1); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[SND_OUT_DATA] += crtsc; } counter_u64_add(tcp_proc_time[SND_OUT_DATA], crtsc); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[CNT_OF_MSS_OUT] += ((tot_len_this_send + segsiz - 1) /segsiz); } counter_u64_add(tcp_cnt_counters[CNT_OF_MSS_OUT], ((tot_len_this_send + segsiz - 1) /segsiz)); } else { if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_cnt_counters[SND_OUT_ACK]++; } counter_u64_add(tcp_cnt_counters[SND_OUT_ACK], 1); if (tp->t_flags2 & TF2_TCP_ACCOUNTING) { tp->tcp_proc_time[SND_OUT_ACK] += crtsc; } counter_u64_add(tcp_proc_time[SND_OUT_ACK], crtsc); } sched_unpin(); #endif if (error == ENOBUFS) error = 0; return (error); } static void rack_update_seg(struct tcp_rack *rack) { uint32_t orig_val; orig_val = rack->r_ctl.rc_pace_max_segs; rack_set_pace_segments(rack->rc_tp, rack, __LINE__, NULL); if (orig_val != rack->r_ctl.rc_pace_max_segs) rack_log_pacing_delay_calc(rack, 0, 0, orig_val, 0, 0, 15, __LINE__, NULL, 0); } static void rack_mtu_change(struct tcpcb *tp) { /* * The MSS may have changed */ struct tcp_rack *rack; struct rack_sendmap *rsm; rack = (struct tcp_rack *)tp->t_fb_ptr; if (rack->r_ctl.rc_pace_min_segs != ctf_fixed_maxseg(tp)) { /* * The MTU has changed we need to resend everything * since all we have sent is lost. We first fix * up the mtu though. */ rack_set_pace_segments(tp, rack, __LINE__, NULL); /* We treat this like a full retransmit timeout without the cwnd adjustment */ rack_remxt_tmr(tp); rack->r_fast_output = 0; rack->r_ctl.rc_out_at_rto = ctf_flight_size(tp, rack->r_ctl.rc_sacked); rack->r_ctl.rc_snd_max_at_rto = tp->snd_max; rack->r_must_retran = 1; /* Mark all inflight to needing to be rxt'd */ TAILQ_FOREACH(rsm, &rack->r_ctl.rc_tmap, r_tnext) { rsm->r_flags |= RACK_MUST_RXT; } } sack_filter_clear(&rack->r_ctl.rack_sf, tp->snd_una); /* We don't use snd_nxt to retransmit */ tp->snd_nxt = tp->snd_max; } static int rack_set_profile(struct tcp_rack *rack, int prof) { int err = EINVAL; if (prof == 1) { /* pace_always=1 */ if (rack->rc_always_pace == 0) { if (tcp_can_enable_pacing() == 0) return (EBUSY); } rack->rc_always_pace = 1; if (rack->use_fixed_rate || rack->gp_ready) rack_set_cc_pacing(rack); rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ; rack->rack_attempt_hdwr_pace = 0; /* cmpack=1 */ if (rack_use_cmp_acks) rack->r_use_cmp_ack = 1; if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) && rack->r_use_cmp_ack) rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP; /* scwnd=1 */ rack->rack_enable_scwnd = 1; /* dynamic=100 */ rack->rc_gp_dyn_mul = 1; /* gp_inc_ca */ rack->r_ctl.rack_per_of_gp_ca = 100; /* rrr_conf=3 */ rack->r_rr_config = 3; /* npush=2 */ rack->r_ctl.rc_no_push_at_mrtt = 2; /* fillcw=1 */ rack->rc_pace_to_cwnd = 1; rack->rc_pace_fill_if_rttin_range = 0; rack->rtt_limit_mul = 0; /* noprr=1 */ rack->rack_no_prr = 1; /* lscwnd=1 */ rack->r_limit_scw = 1; /* gp_inc_rec */ rack->r_ctl.rack_per_of_gp_rec = 90; err = 0; } else if (prof == 3) { /* Same as profile one execept fill_cw becomes 2 (less aggressive set) */ /* pace_always=1 */ if (rack->rc_always_pace == 0) { if (tcp_can_enable_pacing() == 0) return (EBUSY); } rack->rc_always_pace = 1; if (rack->use_fixed_rate || rack->gp_ready) rack_set_cc_pacing(rack); rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ; rack->rack_attempt_hdwr_pace = 0; /* cmpack=1 */ if (rack_use_cmp_acks) rack->r_use_cmp_ack = 1; if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state) && rack->r_use_cmp_ack) rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP; /* scwnd=1 */ rack->rack_enable_scwnd = 1; /* dynamic=100 */ rack->rc_gp_dyn_mul = 1; /* gp_inc_ca */ rack->r_ctl.rack_per_of_gp_ca = 100; /* rrr_conf=3 */ rack->r_rr_config = 3; /* npush=2 */ rack->r_ctl.rc_no_push_at_mrtt = 2; /* fillcw=2 */ rack->rc_pace_to_cwnd = 1; rack->r_fill_less_agg = 1; rack->rc_pace_fill_if_rttin_range = 0; rack->rtt_limit_mul = 0; /* noprr=1 */ rack->rack_no_prr = 1; /* lscwnd=1 */ rack->r_limit_scw = 1; /* gp_inc_rec */ rack->r_ctl.rack_per_of_gp_rec = 90; err = 0; } else if (prof == 2) { /* cmpack=1 */ if (rack->rc_always_pace == 0) { if (tcp_can_enable_pacing() == 0) return (EBUSY); } rack->rc_always_pace = 1; if (rack->use_fixed_rate || rack->gp_ready) rack_set_cc_pacing(rack); rack->r_use_cmp_ack = 1; if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state)) rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP; /* pace_always=1 */ rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ; /* scwnd=1 */ rack->rack_enable_scwnd = 1; /* dynamic=100 */ rack->rc_gp_dyn_mul = 1; rack->r_ctl.rack_per_of_gp_ca = 100; /* rrr_conf=3 */ rack->r_rr_config = 3; /* npush=2 */ rack->r_ctl.rc_no_push_at_mrtt = 2; /* fillcw=1 */ rack->rc_pace_to_cwnd = 1; rack->rc_pace_fill_if_rttin_range = 0; rack->rtt_limit_mul = 0; /* noprr=1 */ rack->rack_no_prr = 1; /* lscwnd=0 */ rack->r_limit_scw = 0; err = 0; } else if (prof == 0) { /* This changes things back to the default settings */ err = 0; if (rack->rc_always_pace) { tcp_decrement_paced_conn(); rack_undo_cc_pacing(rack); rack->rc_always_pace = 0; } if (rack_pace_every_seg && tcp_can_enable_pacing()) { rack->rc_always_pace = 1; if (rack->use_fixed_rate || rack->gp_ready) rack_set_cc_pacing(rack); } else rack->rc_always_pace = 0; if (rack_dsack_std_based & 0x1) { /* Basically this means all rack timers are at least (srtt + 1/4 srtt) */ rack->rc_rack_tmr_std_based = 1; } if (rack_dsack_std_based & 0x2) { /* Basically this means rack timers are extended based on dsack by up to (2 * srtt) */ rack->rc_rack_use_dsack = 1; } if (rack_use_cmp_acks) rack->r_use_cmp_ack = 1; else rack->r_use_cmp_ack = 0; if (rack_disable_prr) rack->rack_no_prr = 1; else rack->rack_no_prr = 0; if (rack_gp_no_rec_chg) rack->rc_gp_no_rec_chg = 1; else rack->rc_gp_no_rec_chg = 0; if (rack_enable_mqueue_for_nonpaced || rack->r_use_cmp_ack) { rack->r_mbuf_queue = 1; if (TCPS_HAVEESTABLISHED(rack->rc_tp->t_state)) rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP; rack->rc_inp->inp_flags2 |= INP_SUPPORTS_MBUFQ; } else { rack->r_mbuf_queue = 0; rack->rc_inp->inp_flags2 &= ~INP_SUPPORTS_MBUFQ; } if (rack_enable_shared_cwnd) rack->rack_enable_scwnd = 1; else rack->rack_enable_scwnd = 0; if (rack_do_dyn_mul) { /* When dynamic adjustment is on CA needs to start at 100% */ rack->rc_gp_dyn_mul = 1; if (rack_do_dyn_mul >= 100) rack->r_ctl.rack_per_of_gp_ca = rack_do_dyn_mul; } else { rack->r_ctl.rack_per_of_gp_ca = rack_per_of_gp_ca; rack->rc_gp_dyn_mul = 0; } rack->r_rr_config = 0; rack->r_ctl.rc_no_push_at_mrtt = 0; rack->rc_pace_to_cwnd = 0; rack->rc_pace_fill_if_rttin_range = 0; rack->rtt_limit_mul = 0; if (rack_enable_hw_pacing) rack->rack_hdw_pace_ena = 1; else rack->rack_hdw_pace_ena = 0; if (rack_disable_prr) rack->rack_no_prr = 1; else rack->rack_no_prr = 0; if (rack_limits_scwnd) rack->r_limit_scw = 1; else rack->r_limit_scw = 0; err = 0; } return (err); } static int rack_add_deferred_option(struct tcp_rack *rack, int sopt_name, uint64_t loptval) { struct deferred_opt_list *dol; dol = malloc(sizeof(struct deferred_opt_list), M_TCPFSB, M_NOWAIT|M_ZERO); if (dol == NULL) { /* * No space yikes -- fail out.. */ return (0); } dol->optname = sopt_name; dol->optval = loptval; TAILQ_INSERT_TAIL(&rack->r_ctl.opt_list, dol, next); return (1); } static int rack_process_option(struct tcpcb *tp, struct tcp_rack *rack, int sopt_name, uint32_t optval, uint64_t loptval) { struct epoch_tracker et; struct sockopt sopt; struct cc_newreno_opts opt; uint64_t val; int error = 0; uint16_t ca, ss; switch (sopt_name) { case TCP_RACK_DSACK_OPT: RACK_OPTS_INC(tcp_rack_dsack_opt); if (optval & 0x1) { rack->rc_rack_tmr_std_based = 1; } else { rack->rc_rack_tmr_std_based = 0; } if (optval & 0x2) { rack->rc_rack_use_dsack = 1; } else { rack->rc_rack_use_dsack = 0; } rack_log_dsack_event(rack, 5, __LINE__, 0, 0); break; case TCP_RACK_PACING_BETA: RACK_OPTS_INC(tcp_rack_beta); if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) { /* This only works for newreno. */ error = EINVAL; break; } if (rack->rc_pacing_cc_set) { /* * Set them into the real CC module * whats in the rack pcb is the old values * to be used on restoral/ */ sopt.sopt_dir = SOPT_SET; opt.name = CC_NEWRENO_BETA; opt.val = optval; if (CC_ALGO(tp)->ctl_output != NULL) error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt); else { error = ENOENT; break; } } else { /* * Not pacing yet so set it into our local * rack pcb storage. */ rack->r_ctl.rc_saved_beta.beta = optval; } break; case TCP_RACK_TIMER_SLOP: RACK_OPTS_INC(tcp_rack_timer_slop); rack->r_ctl.timer_slop = optval; if (rack->rc_tp->t_srtt) { /* * If we have an SRTT lets update t_rxtcur * to have the new slop. */ RACK_TCPT_RANGESET(tp->t_rxtcur, RACK_REXMTVAL(tp), rack_rto_min, rack_rto_max, rack->r_ctl.timer_slop); } break; case TCP_RACK_PACING_BETA_ECN: RACK_OPTS_INC(tcp_rack_beta_ecn); if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) { /* This only works for newreno. */ error = EINVAL; break; } if (rack->rc_pacing_cc_set) { /* * Set them into the real CC module * whats in the rack pcb is the old values * to be used on restoral/ */ sopt.sopt_dir = SOPT_SET; opt.name = CC_NEWRENO_BETA_ECN; opt.val = optval; if (CC_ALGO(tp)->ctl_output != NULL) error = CC_ALGO(tp)->ctl_output(tp->ccv, &sopt, &opt); else error = ENOENT; } else { /* * Not pacing yet so set it into our local * rack pcb storage. */ rack->r_ctl.rc_saved_beta.beta_ecn = optval; rack->r_ctl.rc_saved_beta.newreno_flags = CC_NEWRENO_BETA_ECN_ENABLED; } break; case TCP_DEFER_OPTIONS: RACK_OPTS_INC(tcp_defer_opt); if (optval) { if (rack->gp_ready) { /* Too late */ error = EINVAL; break; } rack->defer_options = 1; } else rack->defer_options = 0; break; case TCP_RACK_MEASURE_CNT: RACK_OPTS_INC(tcp_rack_measure_cnt); if (optval && (optval <= 0xff)) { rack->r_ctl.req_measurements = optval; } else error = EINVAL; break; case TCP_REC_ABC_VAL: RACK_OPTS_INC(tcp_rec_abc_val); if (optval > 0) rack->r_use_labc_for_rec = 1; else rack->r_use_labc_for_rec = 0; break; case TCP_RACK_ABC_VAL: RACK_OPTS_INC(tcp_rack_abc_val); if ((optval > 0) && (optval < 255)) rack->rc_labc = optval; else error = EINVAL; break; case TCP_HDWR_UP_ONLY: RACK_OPTS_INC(tcp_pacing_up_only); if (optval) rack->r_up_only = 1; else rack->r_up_only = 0; break; case TCP_PACING_RATE_CAP: RACK_OPTS_INC(tcp_pacing_rate_cap); rack->r_ctl.bw_rate_cap = loptval; break; case TCP_RACK_PROFILE: RACK_OPTS_INC(tcp_profile); error = rack_set_profile(rack, optval); break; case TCP_USE_CMP_ACKS: RACK_OPTS_INC(tcp_use_cmp_acks); if ((optval == 0) && (rack->rc_inp->inp_flags2 & INP_MBUF_ACKCMP)) { /* You can't turn it off once its on! */ error = EINVAL; } else if ((optval == 1) && (rack->r_use_cmp_ack == 0)) { rack->r_use_cmp_ack = 1; rack->r_mbuf_queue = 1; tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ; } if (rack->r_use_cmp_ack && TCPS_HAVEESTABLISHED(tp->t_state)) rack->rc_inp->inp_flags2 |= INP_MBUF_ACKCMP; break; case TCP_SHARED_CWND_TIME_LIMIT: RACK_OPTS_INC(tcp_lscwnd); if (optval) rack->r_limit_scw = 1; else rack->r_limit_scw = 0; break; case TCP_RACK_PACE_TO_FILL: RACK_OPTS_INC(tcp_fillcw); if (optval == 0) rack->rc_pace_to_cwnd = 0; else { rack->rc_pace_to_cwnd = 1; if (optval > 1) rack->r_fill_less_agg = 1; } if ((optval >= rack_gp_rtt_maxmul) && rack_gp_rtt_maxmul && (optval < 0xf)) { rack->rc_pace_fill_if_rttin_range = 1; rack->rtt_limit_mul = optval; } else { rack->rc_pace_fill_if_rttin_range = 0; rack->rtt_limit_mul = 0; } break; case TCP_RACK_NO_PUSH_AT_MAX: RACK_OPTS_INC(tcp_npush); if (optval == 0) rack->r_ctl.rc_no_push_at_mrtt = 0; else if (optval < 0xff) rack->r_ctl.rc_no_push_at_mrtt = optval; else error = EINVAL; break; case TCP_SHARED_CWND_ENABLE: RACK_OPTS_INC(tcp_rack_scwnd); if (optval == 0) rack->rack_enable_scwnd = 0; else rack->rack_enable_scwnd = 1; break; case TCP_RACK_MBUF_QUEUE: /* Now do we use the LRO mbuf-queue feature */ RACK_OPTS_INC(tcp_rack_mbufq); if (optval || rack->r_use_cmp_ack) rack->r_mbuf_queue = 1; else rack->r_mbuf_queue = 0; if (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack) tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ; else tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ; break; case TCP_RACK_NONRXT_CFG_RATE: RACK_OPTS_INC(tcp_rack_cfg_rate); if (optval == 0) rack->rack_rec_nonrxt_use_cr = 0; else rack->rack_rec_nonrxt_use_cr = 1; break; case TCP_NO_PRR: RACK_OPTS_INC(tcp_rack_noprr); if (optval == 0) rack->rack_no_prr = 0; else if (optval == 1) rack->rack_no_prr = 1; else if (optval == 2) rack->no_prr_addback = 1; else error = EINVAL; break; case TCP_TIMELY_DYN_ADJ: RACK_OPTS_INC(tcp_timely_dyn); if (optval == 0) rack->rc_gp_dyn_mul = 0; else { rack->rc_gp_dyn_mul = 1; if (optval >= 100) { /* * If the user sets something 100 or more * its the gp_ca value. */ rack->r_ctl.rack_per_of_gp_ca = optval; } } break; case TCP_RACK_DO_DETECTION: RACK_OPTS_INC(tcp_rack_do_detection); if (optval == 0) rack->do_detection = 0; else rack->do_detection = 1; break; case TCP_RACK_TLP_USE: if ((optval < TLP_USE_ID) || (optval > TLP_USE_TWO_TWO)) { error = EINVAL; break; } RACK_OPTS_INC(tcp_tlp_use); rack->rack_tlp_threshold_use = optval; break; case TCP_RACK_TLP_REDUCE: /* RACK TLP cwnd reduction (bool) */ RACK_OPTS_INC(tcp_rack_tlp_reduce); rack->r_ctl.rc_tlp_cwnd_reduce = optval; break; /* Pacing related ones */ case TCP_RACK_PACE_ALWAYS: /* * zero is old rack method, 1 is new * method using a pacing rate. */ RACK_OPTS_INC(tcp_rack_pace_always); if (optval > 0) { if (rack->rc_always_pace) { error = EALREADY; break; } else if (tcp_can_enable_pacing()) { rack->rc_always_pace = 1; if (rack->use_fixed_rate || rack->gp_ready) rack_set_cc_pacing(rack); } else { error = ENOSPC; break; } } else { if (rack->rc_always_pace) { tcp_decrement_paced_conn(); rack->rc_always_pace = 0; rack_undo_cc_pacing(rack); } } if (rack->r_mbuf_queue || rack->rc_always_pace || rack->r_use_cmp_ack) tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ; else tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ; /* A rate may be set irate or other, if so set seg size */ rack_update_seg(rack); break; case TCP_BBR_RACK_INIT_RATE: RACK_OPTS_INC(tcp_initial_rate); val = optval; /* Change from kbits per second to bytes per second */ val *= 1000; val /= 8; rack->r_ctl.init_rate = val; if (rack->rc_init_win != rack_default_init_window) { uint32_t win, snt; /* * Options don't always get applied * in the order you think. So in order * to assure we update a cwnd we need * to check and see if we are still * where we should raise the cwnd. */ win = rc_init_window(rack); if (SEQ_GT(tp->snd_max, tp->iss)) snt = tp->snd_max - tp->iss; else snt = 0; if ((snt < win) && (tp->snd_cwnd < win)) tp->snd_cwnd = win; } if (rack->rc_always_pace) rack_update_seg(rack); break; case TCP_BBR_IWINTSO: RACK_OPTS_INC(tcp_initial_win); if (optval && (optval <= 0xff)) { uint32_t win, snt; rack->rc_init_win = optval; win = rc_init_window(rack); if (SEQ_GT(tp->snd_max, tp->iss)) snt = tp->snd_max - tp->iss; else snt = 0; if ((snt < win) && (tp->t_srtt | #ifdef NETFLIX_PEAKRATE tp->t_maxpeakrate | #endif rack->r_ctl.init_rate)) { /* * We are not past the initial window * and we have some bases for pacing, * so we need to possibly adjust up * the cwnd. Note even if we don't set * the cwnd, its still ok to raise the rc_init_win * which can be used coming out of idle when we * would have a rate. */ if (tp->snd_cwnd < win) tp->snd_cwnd = win; } if (rack->rc_always_pace) rack_update_seg(rack); } else error = EINVAL; break; case TCP_RACK_FORCE_MSEG: RACK_OPTS_INC(tcp_rack_force_max_seg); if (optval) rack->rc_force_max_seg = 1; else rack->rc_force_max_seg = 0; break; case TCP_RACK_PACE_MAX_SEG: /* Max segments size in a pace in bytes */ RACK_OPTS_INC(tcp_rack_max_seg); rack->rc_user_set_max_segs = optval; rack_set_pace_segments(tp, rack, __LINE__, NULL); break; case TCP_RACK_PACE_RATE_REC: /* Set the fixed pacing rate in Bytes per second ca */ RACK_OPTS_INC(tcp_rack_pace_rate_rec); rack->r_ctl.rc_fixed_pacing_rate_rec = optval; if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0) rack->r_ctl.rc_fixed_pacing_rate_ca = optval; if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0) rack->r_ctl.rc_fixed_pacing_rate_ss = optval; rack->use_fixed_rate = 1; if (rack->rc_always_pace) rack_set_cc_pacing(rack); rack_log_pacing_delay_calc(rack, rack->r_ctl.rc_fixed_pacing_rate_ss, rack->r_ctl.rc_fixed_pacing_rate_ca, rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8, __LINE__, NULL,0); break; case TCP_RACK_PACE_RATE_SS: /* Set the fixed pacing rate in Bytes per second ca */ RACK_OPTS_INC(tcp_rack_pace_rate_ss); rack->r_ctl.rc_fixed_pacing_rate_ss = optval; if (rack->r_ctl.rc_fixed_pacing_rate_ca == 0) rack->r_ctl.rc_fixed_pacing_rate_ca = optval; if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0) rack->r_ctl.rc_fixed_pacing_rate_rec = optval; rack->use_fixed_rate = 1; if (rack->rc_always_pace) rack_set_cc_pacing(rack); rack_log_pacing_delay_calc(rack, rack->r_ctl.rc_fixed_pacing_rate_ss, rack->r_ctl.rc_fixed_pacing_rate_ca, rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8, __LINE__, NULL, 0); break; case TCP_RACK_PACE_RATE_CA: /* Set the fixed pacing rate in Bytes per second ca */ RACK_OPTS_INC(tcp_rack_pace_rate_ca); rack->r_ctl.rc_fixed_pacing_rate_ca = optval; if (rack->r_ctl.rc_fixed_pacing_rate_ss == 0) rack->r_ctl.rc_fixed_pacing_rate_ss = optval; if (rack->r_ctl.rc_fixed_pacing_rate_rec == 0) rack->r_ctl.rc_fixed_pacing_rate_rec = optval; rack->use_fixed_rate = 1; if (rack->rc_always_pace) rack_set_cc_pacing(rack); rack_log_pacing_delay_calc(rack, rack->r_ctl.rc_fixed_pacing_rate_ss, rack->r_ctl.rc_fixed_pacing_rate_ca, rack->r_ctl.rc_fixed_pacing_rate_rec, 0, 0, 8, __LINE__, NULL, 0); break; case TCP_RACK_GP_INCREASE_REC: RACK_OPTS_INC(tcp_gp_inc_rec); rack->r_ctl.rack_per_of_gp_rec = optval; rack_log_pacing_delay_calc(rack, rack->r_ctl.rack_per_of_gp_ss, rack->r_ctl.rack_per_of_gp_ca, rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1, __LINE__, NULL, 0); break; case TCP_RACK_GP_INCREASE_CA: RACK_OPTS_INC(tcp_gp_inc_ca); ca = optval; if (ca < 100) { /* * We don't allow any reduction * over the GP b/w. */ error = EINVAL; break; } rack->r_ctl.rack_per_of_gp_ca = ca; rack_log_pacing_delay_calc(rack, rack->r_ctl.rack_per_of_gp_ss, rack->r_ctl.rack_per_of_gp_ca, rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1, __LINE__, NULL, 0); break; case TCP_RACK_GP_INCREASE_SS: RACK_OPTS_INC(tcp_gp_inc_ss); ss = optval; if (ss < 100) { /* * We don't allow any reduction * over the GP b/w. */ error = EINVAL; break; } rack->r_ctl.rack_per_of_gp_ss = ss; rack_log_pacing_delay_calc(rack, rack->r_ctl.rack_per_of_gp_ss, rack->r_ctl.rack_per_of_gp_ca, rack->r_ctl.rack_per_of_gp_rec, 0, 0, 1, __LINE__, NULL, 0); break; case TCP_RACK_RR_CONF: RACK_OPTS_INC(tcp_rack_rrr_no_conf_rate); if (optval && optval <= 3) rack->r_rr_config = optval; else rack->r_rr_config = 0; break; case TCP_HDWR_RATE_CAP: RACK_OPTS_INC(tcp_hdwr_rate_cap); if (optval) { if (rack->r_rack_hw_rate_caps == 0) rack->r_rack_hw_rate_caps = 1; else error = EALREADY; } else { rack->r_rack_hw_rate_caps = 0; } break; case TCP_BBR_HDWR_PACE: RACK_OPTS_INC(tcp_hdwr_pacing); if (optval){ if (rack->rack_hdrw_pacing == 0) { rack->rack_hdw_pace_ena = 1; rack->rack_attempt_hdwr_pace = 0; } else error = EALREADY; } else { rack->rack_hdw_pace_ena = 0; #ifdef RATELIMIT if (rack->r_ctl.crte != NULL) { rack->rack_hdrw_pacing = 0; rack->rack_attempt_hdwr_pace = 0; tcp_rel_pacing_rate(rack->r_ctl.crte, tp); rack->r_ctl.crte = NULL; } #endif } break; /* End Pacing related ones */ case TCP_RACK_PRR_SENDALOT: /* Allow PRR to send more than one seg */ RACK_OPTS_INC(tcp_rack_prr_sendalot); rack->r_ctl.rc_prr_sendalot = optval; break; case TCP_RACK_MIN_TO: /* Minimum time between rack t-o's in ms */ RACK_OPTS_INC(tcp_rack_min_to); rack->r_ctl.rc_min_to = optval; break; case TCP_RACK_EARLY_SEG: /* If early recovery max segments */ RACK_OPTS_INC(tcp_rack_early_seg); rack->r_ctl.rc_early_recovery_segs = optval; break; case TCP_RACK_ENABLE_HYSTART: { if (optval) { tp->ccv->flags |= CCF_HYSTART_ALLOWED; if (rack_do_hystart > RACK_HYSTART_ON) tp->ccv->flags |= CCF_HYSTART_CAN_SH_CWND; if (rack_do_hystart > RACK_HYSTART_ON_W_SC) tp->ccv->flags |= CCF_HYSTART_CONS_SSTH; } else { tp->ccv->flags &= ~(CCF_HYSTART_ALLOWED|CCF_HYSTART_CAN_SH_CWND|CCF_HYSTART_CONS_SSTH); } } break; case TCP_RACK_REORD_THRESH: /* RACK reorder threshold (shift amount) */ RACK_OPTS_INC(tcp_rack_reord_thresh); if ((optval > 0) && (optval < 31)) rack->r_ctl.rc_reorder_shift = optval; else error = EINVAL; break; case TCP_RACK_REORD_FADE: /* Does reordering fade after ms time */ RACK_OPTS_INC(tcp_rack_reord_fade); rack->r_ctl.rc_reorder_fade = optval; break; case TCP_RACK_TLP_THRESH: /* RACK TLP theshold i.e. srtt+(srtt/N) */ RACK_OPTS_INC(tcp_rack_tlp_thresh); if (optval) rack->r_ctl.rc_tlp_threshold = optval; else error = EINVAL; break; case TCP_BBR_USE_RACK_RR: RACK_OPTS_INC(tcp_rack_rr); if (optval) rack->use_rack_rr = 1; else rack->use_rack_rr = 0; break; case TCP_FAST_RSM_HACK: RACK_OPTS_INC(tcp_rack_fastrsm_hack); if (optval) rack->fast_rsm_hack = 1; else rack->fast_rsm_hack = 0; break; case TCP_RACK_PKT_DELAY: /* RACK added ms i.e. rack-rtt + reord + N */ RACK_OPTS_INC(tcp_rack_pkt_delay); rack->r_ctl.rc_pkt_delay = optval; break; case TCP_DELACK: RACK_OPTS_INC(tcp_rack_delayed_ack); if (optval == 0) tp->t_delayed_ack = 0; else tp->t_delayed_ack = 1; if (tp->t_flags & TF_DELACK) { tp->t_flags &= ~TF_DELACK; tp->t_flags |= TF_ACKNOW; NET_EPOCH_ENTER(et); rack_output(tp); NET_EPOCH_EXIT(et); } break; case TCP_BBR_RACK_RTT_USE: RACK_OPTS_INC(tcp_rack_rtt_use); if ((optval != USE_RTT_HIGH) && (optval != USE_RTT_LOW) && (optval != USE_RTT_AVG)) error = EINVAL; else rack->r_ctl.rc_rate_sample_method = optval; break; case TCP_DATA_AFTER_CLOSE: RACK_OPTS_INC(tcp_data_after_close); if (optval) rack->rc_allow_data_af_clo = 1; else rack->rc_allow_data_af_clo = 0; break; default: break; } #ifdef NETFLIX_STATS tcp_log_socket_option(tp, sopt_name, optval, error); #endif return (error); } static void rack_apply_deferred_options(struct tcp_rack *rack) { struct deferred_opt_list *dol, *sdol; uint32_t s_optval; TAILQ_FOREACH_SAFE(dol, &rack->r_ctl.opt_list, next, sdol) { TAILQ_REMOVE(&rack->r_ctl.opt_list, dol, next); /* Disadvantage of deferal is you loose the error return */ s_optval = (uint32_t)dol->optval; (void)rack_process_option(rack->rc_tp, rack, dol->optname, s_optval, dol->optval); free(dol, M_TCPDO); } } static void rack_hw_tls_change(struct tcpcb *tp, int chg) { /* * HW tls state has changed.. fix all * rsm's in flight. */ struct tcp_rack *rack; struct rack_sendmap *rsm; rack = (struct tcp_rack *)tp->t_fb_ptr; RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) { if (chg) rsm->r_hw_tls = 1; else rsm->r_hw_tls = 0; } if (chg) rack->r_ctl.fsb.hw_tls = 1; else rack->r_ctl.fsb.hw_tls = 0; } static int rack_pru_options(struct tcpcb *tp, int flags) { if (flags & PRUS_OOB) return (EOPNOTSUPP); return (0); } static struct tcp_function_block __tcp_rack = { .tfb_tcp_block_name = __XSTRING(STACKNAME), .tfb_tcp_output = rack_output, .tfb_do_queued_segments = ctf_do_queued_segments, .tfb_do_segment_nounlock = rack_do_segment_nounlock, .tfb_tcp_do_segment = rack_do_segment, .tfb_tcp_ctloutput = rack_ctloutput, .tfb_tcp_fb_init = rack_init, .tfb_tcp_fb_fini = rack_fini, .tfb_tcp_timer_stop_all = rack_stopall, .tfb_tcp_timer_activate = rack_timer_activate, .tfb_tcp_timer_active = rack_timer_active, .tfb_tcp_timer_stop = rack_timer_stop, .tfb_tcp_rexmit_tmr = rack_remxt_tmr, .tfb_tcp_handoff_ok = rack_handoff_ok, .tfb_tcp_mtu_chg = rack_mtu_change, .tfb_pru_options = rack_pru_options, .tfb_hwtls_change = rack_hw_tls_change, .tfb_compute_pipe = rack_compute_pipe, .tfb_flags = TCP_FUNC_OUTPUT_CANDROP, }; /* * rack_ctloutput() must drop the inpcb lock before performing copyin on * socket option arguments. When it re-acquires the lock after the copy, it * has to revalidate that the connection is still valid for the socket * option. */ static int rack_set_sockopt(struct inpcb *inp, struct sockopt *sopt) { #ifdef INET6 struct ip6_hdr *ip6; #endif #ifdef INET struct ip *ip; #endif struct tcpcb *tp; struct tcp_rack *rack; uint64_t loptval; int32_t error = 0, optval; tp = intotcpcb(inp); rack = (struct tcp_rack *)tp->t_fb_ptr; if (rack == NULL) { INP_WUNLOCK(inp); return (EINVAL); } #ifdef INET6 ip6 = (struct ip6_hdr *)rack->r_ctl.fsb.tcp_ip_hdr; #endif #ifdef INET ip = (struct ip *)rack->r_ctl.fsb.tcp_ip_hdr; #endif switch (sopt->sopt_level) { #ifdef INET6 case IPPROTO_IPV6: MPASS(inp->inp_vflag & INP_IPV6PROTO); switch (sopt->sopt_name) { case IPV6_USE_MIN_MTU: tcp6_use_min_mtu(tp); break; case IPV6_TCLASS: /* * The DSCP codepoint has changed, update the fsb. */ ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) | (rack->rc_inp->inp_flow & IPV6_FLOWINFO_MASK); break; } INP_WUNLOCK(inp); return (0); #endif #ifdef INET case IPPROTO_IP: switch (sopt->sopt_name) { case IP_TOS: /* * The DSCP codepoint has changed, update the fsb. */ ip->ip_tos = rack->rc_inp->inp_ip_tos; break; case IP_TTL: /* * The TTL has changed, update the fsb. */ ip->ip_ttl = rack->rc_inp->inp_ip_ttl; break; } INP_WUNLOCK(inp); return (0); #endif } switch (sopt->sopt_name) { case TCP_RACK_TLP_REDUCE: /* URL:tlp_reduce */ /* Pacing related ones */ case TCP_RACK_PACE_ALWAYS: /* URL:pace_always */ case TCP_BBR_RACK_INIT_RATE: /* URL:irate */ case TCP_BBR_IWINTSO: /* URL:tso_iwin */ case TCP_RACK_PACE_MAX_SEG: /* URL:pace_max_seg */ case TCP_RACK_FORCE_MSEG: /* URL:force_max_seg */ case TCP_RACK_PACE_RATE_CA: /* URL:pr_ca */ case TCP_RACK_PACE_RATE_SS: /* URL:pr_ss*/ case TCP_RACK_PACE_RATE_REC: /* URL:pr_rec */ case TCP_RACK_GP_INCREASE_CA: /* URL:gp_inc_ca */ case TCP_RACK_GP_INCREASE_SS: /* URL:gp_inc_ss */ case TCP_RACK_GP_INCREASE_REC: /* URL:gp_inc_rec */ case TCP_RACK_RR_CONF: /* URL:rrr_conf */ case TCP_BBR_HDWR_PACE: /* URL:hdwrpace */ case TCP_HDWR_RATE_CAP: /* URL:hdwrcap boolean */ case TCP_PACING_RATE_CAP: /* URL:cap -- used by side-channel */ case TCP_HDWR_UP_ONLY: /* URL:uponly -- hardware pacing boolean */ /* End pacing related */ case TCP_FAST_RSM_HACK: /* URL:frsm_hack */ case TCP_DELACK: /* URL:delack (in base TCP i.e. tcp_hints along with cc etc ) */ case TCP_RACK_PRR_SENDALOT: /* URL:prr_sendalot */ case TCP_RACK_MIN_TO: /* URL:min_to */ case TCP_RACK_EARLY_SEG: /* URL:early_seg */ case TCP_RACK_REORD_THRESH: /* URL:reord_thresh */ case TCP_RACK_REORD_FADE: /* URL:reord_fade */ case TCP_RACK_TLP_THRESH: /* URL:tlp_thresh */ case TCP_RACK_PKT_DELAY: /* URL:pkt_delay */ case TCP_RACK_TLP_USE: /* URL:tlp_use */ case TCP_BBR_RACK_RTT_USE: /* URL:rttuse */ case TCP_BBR_USE_RACK_RR: /* URL:rackrr */ case TCP_RACK_DO_DETECTION: /* URL:detect */ case TCP_NO_PRR: /* URL:noprr */ case TCP_TIMELY_DYN_ADJ: /* URL:dynamic */ case TCP_DATA_AFTER_CLOSE: /* no URL */ case TCP_RACK_NONRXT_CFG_RATE: /* URL:nonrxtcr */ case TCP_SHARED_CWND_ENABLE: /* URL:scwnd */ case TCP_RACK_MBUF_QUEUE: /* URL:mqueue */ case TCP_RACK_NO_PUSH_AT_MAX: /* URL:npush */ case TCP_RACK_PACE_TO_FILL: /* URL:fillcw */ case TCP_SHARED_CWND_TIME_LIMIT: /* URL:lscwnd */ case TCP_RACK_PROFILE: /* URL:profile */ case TCP_USE_CMP_ACKS: /* URL:cmpack */ case TCP_RACK_ABC_VAL: /* URL:labc */ case TCP_REC_ABC_VAL: /* URL:reclabc */ case TCP_RACK_MEASURE_CNT: /* URL:measurecnt */ case TCP_DEFER_OPTIONS: /* URL:defer */ case TCP_RACK_DSACK_OPT: /* URL:dsack */ case TCP_RACK_PACING_BETA: /* URL:pacing_beta */ case TCP_RACK_PACING_BETA_ECN: /* URL:pacing_beta_ecn */ case TCP_RACK_TIMER_SLOP: /* URL:timer_slop */ case TCP_RACK_ENABLE_HYSTART: /* URL:hystart */ break; default: /* Filter off all unknown options to the base stack */ return (tcp_default_ctloutput(inp, sopt)); break; } INP_WUNLOCK(inp); if (sopt->sopt_name == TCP_PACING_RATE_CAP) { error = sooptcopyin(sopt, &loptval, sizeof(loptval), sizeof(loptval)); /* * We truncate it down to 32 bits for the socket-option trace this * means rates > 34Gbps won't show right, but thats probably ok. */ optval = (uint32_t)loptval; } else { error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval)); /* Save it in 64 bit form too */ loptval = optval; } if (error) return (error); INP_WLOCK(inp); - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { INP_WUNLOCK(inp); return (ECONNRESET); } if (tp->t_fb != &__tcp_rack) { INP_WUNLOCK(inp); return (ENOPROTOOPT); } if (rack->defer_options && (rack->gp_ready == 0) && (sopt->sopt_name != TCP_DEFER_OPTIONS) && (sopt->sopt_name != TCP_RACK_PACING_BETA) && (sopt->sopt_name != TCP_RACK_PACING_BETA_ECN) && (sopt->sopt_name != TCP_RACK_MEASURE_CNT)) { /* Options are beind deferred */ if (rack_add_deferred_option(rack, sopt->sopt_name, loptval)) { INP_WUNLOCK(inp); return (0); } else { /* No memory to defer, fail */ INP_WUNLOCK(inp); return (ENOMEM); } } error = rack_process_option(tp, rack, sopt->sopt_name, optval, loptval); INP_WUNLOCK(inp); return (error); } static void rack_fill_info(struct tcpcb *tp, struct tcp_info *ti) { INP_WLOCK_ASSERT(tp->t_inpcb); bzero(ti, sizeof(*ti)); ti->tcpi_state = tp->t_state; if ((tp->t_flags & TF_REQ_TSTMP) && (tp->t_flags & TF_RCVD_TSTMP)) ti->tcpi_options |= TCPI_OPT_TIMESTAMPS; if (tp->t_flags & TF_SACK_PERMIT) ti->tcpi_options |= TCPI_OPT_SACK; if ((tp->t_flags & TF_REQ_SCALE) && (tp->t_flags & TF_RCVD_SCALE)) { ti->tcpi_options |= TCPI_OPT_WSCALE; ti->tcpi_snd_wscale = tp->snd_scale; ti->tcpi_rcv_wscale = tp->rcv_scale; } if (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT)) ti->tcpi_options |= TCPI_OPT_ECN; if (tp->t_flags & TF_FASTOPEN) ti->tcpi_options |= TCPI_OPT_TFO; /* still kept in ticks is t_rcvtime */ ti->tcpi_last_data_recv = ((uint32_t)ticks - tp->t_rcvtime) * tick; /* Since we hold everything in precise useconds this is easy */ ti->tcpi_rtt = tp->t_srtt; ti->tcpi_rttvar = tp->t_rttvar; ti->tcpi_rto = tp->t_rxtcur; ti->tcpi_snd_ssthresh = tp->snd_ssthresh; ti->tcpi_snd_cwnd = tp->snd_cwnd; /* * FreeBSD-specific extension fields for tcp_info. */ ti->tcpi_rcv_space = tp->rcv_wnd; ti->tcpi_rcv_nxt = tp->rcv_nxt; ti->tcpi_snd_wnd = tp->snd_wnd; ti->tcpi_snd_bwnd = 0; /* Unused, kept for compat. */ ti->tcpi_snd_nxt = tp->snd_nxt; ti->tcpi_snd_mss = tp->t_maxseg; ti->tcpi_rcv_mss = tp->t_maxseg; ti->tcpi_snd_rexmitpack = tp->t_sndrexmitpack; ti->tcpi_rcv_ooopack = tp->t_rcvoopack; ti->tcpi_snd_zerowin = tp->t_sndzerowin; #ifdef NETFLIX_STATS ti->tcpi_total_tlp = tp->t_sndtlppack; ti->tcpi_total_tlp_bytes = tp->t_sndtlpbyte; memcpy(&ti->tcpi_rxsyninfo, &tp->t_rxsyninfo, sizeof(struct tcpsyninfo)); #endif #ifdef TCP_OFFLOAD if (tp->t_flags & TF_TOE) { ti->tcpi_options |= TCPI_OPT_TOE; tcp_offload_tcp_info(tp, ti); } #endif } static int rack_get_sockopt(struct inpcb *inp, struct sockopt *sopt) { struct tcpcb *tp; struct tcp_rack *rack; int32_t error, optval; uint64_t val, loptval; struct tcp_info ti; /* * Because all our options are either boolean or an int, we can just * pull everything into optval and then unlock and copy. If we ever * add a option that is not a int, then this will have quite an * impact to this routine. */ error = 0; tp = intotcpcb(inp); rack = (struct tcp_rack *)tp->t_fb_ptr; if (rack == NULL) { INP_WUNLOCK(inp); return (EINVAL); } switch (sopt->sopt_name) { case TCP_INFO: /* First get the info filled */ rack_fill_info(tp, &ti); /* Fix up the rtt related fields if needed */ INP_WUNLOCK(inp); error = sooptcopyout(sopt, &ti, sizeof ti); return (error); /* * Beta is the congestion control value for NewReno that influences how * much of a backoff happens when loss is detected. It is normally set * to 50 for 50% i.e. the cwnd is reduced to 50% of its previous value * when you exit recovery. */ case TCP_RACK_PACING_BETA: if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) error = EINVAL; else if (rack->rc_pacing_cc_set == 0) optval = rack->r_ctl.rc_saved_beta.beta; else { /* * Reach out into the CC data and report back what * I have previously set. Yeah it looks hackish but * we don't want to report the saved values. */ if (tp->ccv->cc_data) optval = ((struct newreno *)tp->ccv->cc_data)->beta; else error = EINVAL; } break; /* * Beta_ecn is the congestion control value for NewReno that influences how * much of a backoff happens when a ECN mark is detected. It is normally set * to 80 for 80% i.e. the cwnd is reduced by 20% of its previous value when * you exit recovery. Note that classic ECN has a beta of 50, it is only * ABE Ecn that uses this "less" value, but we do too with pacing :) */ case TCP_RACK_PACING_BETA_ECN: if (strcmp(tp->cc_algo->name, CCALGONAME_NEWRENO) != 0) error = EINVAL; else if (rack->rc_pacing_cc_set == 0) optval = rack->r_ctl.rc_saved_beta.beta_ecn; else { /* * Reach out into the CC data and report back what * I have previously set. Yeah it looks hackish but * we don't want to report the saved values. */ if (tp->ccv->cc_data) optval = ((struct newreno *)tp->ccv->cc_data)->beta_ecn; else error = EINVAL; } break; case TCP_RACK_DSACK_OPT: optval = 0; if (rack->rc_rack_tmr_std_based) { optval |= 1; } if (rack->rc_rack_use_dsack) { optval |= 2; } break; case TCP_RACK_ENABLE_HYSTART: { if (tp->ccv->flags & CCF_HYSTART_ALLOWED) { optval = RACK_HYSTART_ON; if (tp->ccv->flags & CCF_HYSTART_CAN_SH_CWND) optval = RACK_HYSTART_ON_W_SC; if (tp->ccv->flags & CCF_HYSTART_CONS_SSTH) optval = RACK_HYSTART_ON_W_SC_C; } else { optval = RACK_HYSTART_OFF; } } break; case TCP_FAST_RSM_HACK: optval = rack->fast_rsm_hack; break; case TCP_DEFER_OPTIONS: optval = rack->defer_options; break; case TCP_RACK_MEASURE_CNT: optval = rack->r_ctl.req_measurements; break; case TCP_REC_ABC_VAL: optval = rack->r_use_labc_for_rec; break; case TCP_RACK_ABC_VAL: optval = rack->rc_labc; break; case TCP_HDWR_UP_ONLY: optval= rack->r_up_only; break; case TCP_PACING_RATE_CAP: loptval = rack->r_ctl.bw_rate_cap; break; case TCP_RACK_PROFILE: /* You cannot retrieve a profile, its write only */ error = EINVAL; break; case TCP_USE_CMP_ACKS: optval = rack->r_use_cmp_ack; break; case TCP_RACK_PACE_TO_FILL: optval = rack->rc_pace_to_cwnd; if (optval && rack->r_fill_less_agg) optval++; break; case TCP_RACK_NO_PUSH_AT_MAX: optval = rack->r_ctl.rc_no_push_at_mrtt; break; case TCP_SHARED_CWND_ENABLE: optval = rack->rack_enable_scwnd; break; case TCP_RACK_NONRXT_CFG_RATE: optval = rack->rack_rec_nonrxt_use_cr; break; case TCP_NO_PRR: if (rack->rack_no_prr == 1) optval = 1; else if (rack->no_prr_addback == 1) optval = 2; else optval = 0; break; case TCP_RACK_DO_DETECTION: optval = rack->do_detection; break; case TCP_RACK_MBUF_QUEUE: /* Now do we use the LRO mbuf-queue feature */ optval = rack->r_mbuf_queue; break; case TCP_TIMELY_DYN_ADJ: optval = rack->rc_gp_dyn_mul; break; case TCP_BBR_IWINTSO: optval = rack->rc_init_win; break; case TCP_RACK_TLP_REDUCE: /* RACK TLP cwnd reduction (bool) */ optval = rack->r_ctl.rc_tlp_cwnd_reduce; break; case TCP_BBR_RACK_INIT_RATE: val = rack->r_ctl.init_rate; /* convert to kbits per sec */ val *= 8; val /= 1000; optval = (uint32_t)val; break; case TCP_RACK_FORCE_MSEG: optval = rack->rc_force_max_seg; break; case TCP_RACK_PACE_MAX_SEG: /* Max segments in a pace */ optval = rack->rc_user_set_max_segs; break; case TCP_RACK_PACE_ALWAYS: /* Use the always pace method */ optval = rack->rc_always_pace; break; case TCP_RACK_PRR_SENDALOT: /* Allow PRR to send more than one seg */ optval = rack->r_ctl.rc_prr_sendalot; break; case TCP_RACK_MIN_TO: /* Minimum time between rack t-o's in ms */ optval = rack->r_ctl.rc_min_to; break; case TCP_RACK_EARLY_SEG: /* If early recovery max segments */ optval = rack->r_ctl.rc_early_recovery_segs; break; case TCP_RACK_REORD_THRESH: /* RACK reorder threshold (shift amount) */ optval = rack->r_ctl.rc_reorder_shift; break; case TCP_RACK_REORD_FADE: /* Does reordering fade after ms time */ optval = rack->r_ctl.rc_reorder_fade; break; case TCP_BBR_USE_RACK_RR: /* Do we use the rack cheat for rxt */ optval = rack->use_rack_rr; break; case TCP_RACK_RR_CONF: optval = rack->r_rr_config; break; case TCP_HDWR_RATE_CAP: optval = rack->r_rack_hw_rate_caps; break; case TCP_BBR_HDWR_PACE: optval = rack->rack_hdw_pace_ena; break; case TCP_RACK_TLP_THRESH: /* RACK TLP theshold i.e. srtt+(srtt/N) */ optval = rack->r_ctl.rc_tlp_threshold; break; case TCP_RACK_PKT_DELAY: /* RACK added ms i.e. rack-rtt + reord + N */ optval = rack->r_ctl.rc_pkt_delay; break; case TCP_RACK_TLP_USE: optval = rack->rack_tlp_threshold_use; break; case TCP_RACK_PACE_RATE_CA: optval = rack->r_ctl.rc_fixed_pacing_rate_ca; break; case TCP_RACK_PACE_RATE_SS: optval = rack->r_ctl.rc_fixed_pacing_rate_ss; break; case TCP_RACK_PACE_RATE_REC: optval = rack->r_ctl.rc_fixed_pacing_rate_rec; break; case TCP_RACK_GP_INCREASE_SS: optval = rack->r_ctl.rack_per_of_gp_ca; break; case TCP_RACK_GP_INCREASE_CA: optval = rack->r_ctl.rack_per_of_gp_ss; break; case TCP_BBR_RACK_RTT_USE: optval = rack->r_ctl.rc_rate_sample_method; break; case TCP_DELACK: optval = tp->t_delayed_ack; break; case TCP_DATA_AFTER_CLOSE: optval = rack->rc_allow_data_af_clo; break; case TCP_SHARED_CWND_TIME_LIMIT: optval = rack->r_limit_scw; break; case TCP_RACK_TIMER_SLOP: optval = rack->r_ctl.timer_slop; break; default: return (tcp_default_ctloutput(inp, sopt)); break; } INP_WUNLOCK(inp); if (error == 0) { if (TCP_PACING_RATE_CAP) error = sooptcopyout(sopt, &loptval, sizeof loptval); else error = sooptcopyout(sopt, &optval, sizeof optval); } return (error); } static int rack_ctloutput(struct inpcb *inp, struct sockopt *sopt) { if (sopt->sopt_dir == SOPT_SET) { return (rack_set_sockopt(inp, sopt)); } else if (sopt->sopt_dir == SOPT_GET) { return (rack_get_sockopt(inp, sopt)); } else { panic("%s: sopt_dir $%d", __func__, sopt->sopt_dir); } } static const char *rack_stack_names[] = { __XSTRING(STACKNAME), #ifdef STACKALIAS __XSTRING(STACKALIAS), #endif }; static int rack_ctor(void *mem, int32_t size, void *arg, int32_t how) { memset(mem, 0, size); return (0); } static void rack_dtor(void *mem, int32_t size, void *arg) { } static bool rack_mod_inited = false; static int tcp_addrack(module_t mod, int32_t type, void *data) { int32_t err = 0; int num_stacks; switch (type) { case MOD_LOAD: rack_zone = uma_zcreate(__XSTRING(MODNAME) "_map", sizeof(struct rack_sendmap), rack_ctor, rack_dtor, NULL, NULL, UMA_ALIGN_PTR, 0); rack_pcb_zone = uma_zcreate(__XSTRING(MODNAME) "_pcb", sizeof(struct tcp_rack), rack_ctor, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0); sysctl_ctx_init(&rack_sysctl_ctx); rack_sysctl_root = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_STATIC_CHILDREN(_net_inet_tcp), OID_AUTO, #ifdef STACKALIAS __XSTRING(STACKALIAS), #else __XSTRING(STACKNAME), #endif CTLFLAG_RW | CTLFLAG_MPSAFE, 0, ""); if (rack_sysctl_root == NULL) { printf("Failed to add sysctl node\n"); err = EFAULT; goto free_uma; } rack_init_sysctls(); num_stacks = nitems(rack_stack_names); err = register_tcp_functions_as_names(&__tcp_rack, M_WAITOK, rack_stack_names, &num_stacks); if (err) { printf("Failed to register %s stack name for " "%s module\n", rack_stack_names[num_stacks], __XSTRING(MODNAME)); sysctl_ctx_free(&rack_sysctl_ctx); free_uma: uma_zdestroy(rack_zone); uma_zdestroy(rack_pcb_zone); rack_counter_destroy(); printf("Failed to register rack module -- err:%d\n", err); return (err); } tcp_lro_reg_mbufq(); rack_mod_inited = true; break; case MOD_QUIESCE: err = deregister_tcp_functions(&__tcp_rack, true, false); break; case MOD_UNLOAD: err = deregister_tcp_functions(&__tcp_rack, false, true); if (err == EBUSY) break; if (rack_mod_inited) { uma_zdestroy(rack_zone); uma_zdestroy(rack_pcb_zone); sysctl_ctx_free(&rack_sysctl_ctx); rack_counter_destroy(); rack_mod_inited = false; } tcp_lro_dereg_mbufq(); err = 0; break; default: return (EOPNOTSUPP); } return (err); } static moduledata_t tcp_rack = { .name = __XSTRING(MODNAME), .evhand = tcp_addrack, .priv = 0 }; MODULE_VERSION(MODNAME, 1); DECLARE_MODULE(MODNAME, tcp_rack, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY); MODULE_DEPEND(MODNAME, tcphpts, 1, 1, 1); diff --git a/sys/netinet/tcp_subr.c b/sys/netinet/tcp_subr.c index c50f416351c3..8fdaab35fb19 100644 --- a/sys/netinet/tcp_subr.c +++ b/sys/netinet/tcp_subr.c @@ -1,4111 +1,4102 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995 * The Regents of the University of California. All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)tcp_subr.c 8.2 (Berkeley) 5/24/95 */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include "opt_ipsec.h" #include "opt_kern_tls.h" #include "opt_tcpdebug.h" #include #include #include #include #include #ifdef TCP_HHOOK #include #endif #include #ifdef TCP_HHOOK #include #endif #ifdef KERN_TLS #include #endif #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef INET6 #include #include #include #include #include #include #include #endif #include #ifdef INVARIANTS #define TCPSTATES #endif #include #include #include #include #include #include #include #include #include #include #ifdef TCPPCAP #include #endif #ifdef TCPDEBUG #include #endif #ifdef TCP_OFFLOAD #include #endif #include #include #ifdef INET6 #include #endif #include #include #include #include #ifdef INET6 static ip6proto_ctlinput_t tcp6_ctlinput; static udp_tun_icmp_t tcp6_ctlinput_viaudp; #endif VNET_DEFINE(int, tcp_mssdflt) = TCP_MSS; #ifdef INET6 VNET_DEFINE(int, tcp_v6mssdflt) = TCP6_MSS; #endif #ifdef NETFLIX_EXP_DETECTION /* Sack attack detection thresholds and such */ SYSCTL_NODE(_net_inet_tcp, OID_AUTO, sack_attack, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Sack Attack detection thresholds"); int32_t tcp_force_detection = 0; SYSCTL_INT(_net_inet_tcp_sack_attack, OID_AUTO, force_detection, CTLFLAG_RW, &tcp_force_detection, 0, "Do we force detection even if the INP has it off?"); int32_t tcp_sack_to_ack_thresh = 700; /* 70 % */ SYSCTL_INT(_net_inet_tcp_sack_attack, OID_AUTO, sack_to_ack_thresh, CTLFLAG_RW, &tcp_sack_to_ack_thresh, 700, "Percentage of sacks to acks we must see above (10.1 percent is 101)?"); int32_t tcp_sack_to_move_thresh = 600; /* 60 % */ SYSCTL_INT(_net_inet_tcp_sack_attack, OID_AUTO, move_thresh, CTLFLAG_RW, &tcp_sack_to_move_thresh, 600, "Percentage of sack moves we must see above (10.1 percent is 101)"); int32_t tcp_restoral_thresh = 650; /* 65 % (sack:2:ack -5%) */ SYSCTL_INT(_net_inet_tcp_sack_attack, OID_AUTO, restore_thresh, CTLFLAG_RW, &tcp_restoral_thresh, 550, "Percentage of sack to ack percentage we must see below to restore(10.1 percent is 101)"); int32_t tcp_sad_decay_val = 800; SYSCTL_INT(_net_inet_tcp_sack_attack, OID_AUTO, decay_per, CTLFLAG_RW, &tcp_sad_decay_val, 800, "The decay percentage (10.1 percent equals 101 )"); int32_t tcp_map_minimum = 500; SYSCTL_INT(_net_inet_tcp_sack_attack, OID_AUTO, nummaps, CTLFLAG_RW, &tcp_map_minimum, 500, "Number of Map enteries before we start detection"); int32_t tcp_attack_on_turns_on_logging = 0; SYSCTL_INT(_net_inet_tcp_sack_attack, OID_AUTO, attacks_logged, CTLFLAG_RW, &tcp_attack_on_turns_on_logging, 0, "When we have a positive hit on attack, do we turn on logging?"); int32_t tcp_sad_pacing_interval = 2000; SYSCTL_INT(_net_inet_tcp_sack_attack, OID_AUTO, sad_pacing_int, CTLFLAG_RW, &tcp_sad_pacing_interval, 2000, "What is the minimum pacing interval for a classified attacker?"); int32_t tcp_sad_low_pps = 100; SYSCTL_INT(_net_inet_tcp_sack_attack, OID_AUTO, sad_low_pps, CTLFLAG_RW, &tcp_sad_low_pps, 100, "What is the input pps that below which we do not decay?"); #endif uint32_t tcp_ack_war_time_window = 1000; SYSCTL_UINT(_net_inet_tcp, OID_AUTO, ack_war_timewindow, CTLFLAG_RW, &tcp_ack_war_time_window, 1000, "If the tcp_stack does ack-war prevention how many milliseconds are in its time window?"); uint32_t tcp_ack_war_cnt = 5; SYSCTL_UINT(_net_inet_tcp, OID_AUTO, ack_war_cnt, CTLFLAG_RW, &tcp_ack_war_cnt, 5, "If the tcp_stack does ack-war prevention how many acks can be sent in its time window?"); struct rwlock tcp_function_lock; static int sysctl_net_inet_tcp_mss_check(SYSCTL_HANDLER_ARGS) { int error, new; new = V_tcp_mssdflt; error = sysctl_handle_int(oidp, &new, 0, req); if (error == 0 && req->newptr) { if (new < TCP_MINMSS) error = EINVAL; else V_tcp_mssdflt = new; } return (error); } SYSCTL_PROC(_net_inet_tcp, TCPCTL_MSSDFLT, mssdflt, CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &VNET_NAME(tcp_mssdflt), 0, &sysctl_net_inet_tcp_mss_check, "I", "Default TCP Maximum Segment Size"); #ifdef INET6 static int sysctl_net_inet_tcp_mss_v6_check(SYSCTL_HANDLER_ARGS) { int error, new; new = V_tcp_v6mssdflt; error = sysctl_handle_int(oidp, &new, 0, req); if (error == 0 && req->newptr) { if (new < TCP_MINMSS) error = EINVAL; else V_tcp_v6mssdflt = new; } return (error); } SYSCTL_PROC(_net_inet_tcp, TCPCTL_V6MSSDFLT, v6mssdflt, CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &VNET_NAME(tcp_v6mssdflt), 0, &sysctl_net_inet_tcp_mss_v6_check, "I", "Default TCP Maximum Segment Size for IPv6"); #endif /* INET6 */ /* * Minimum MSS we accept and use. This prevents DoS attacks where * we are forced to a ridiculous low MSS like 20 and send hundreds * of packets instead of one. The effect scales with the available * bandwidth and quickly saturates the CPU and network interface * with packet generation and sending. Set to zero to disable MINMSS * checking. This setting prevents us from sending too small packets. */ VNET_DEFINE(int, tcp_minmss) = TCP_MINMSS; SYSCTL_INT(_net_inet_tcp, OID_AUTO, minmss, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_minmss), 0, "Minimum TCP Maximum Segment Size"); VNET_DEFINE(int, tcp_do_rfc1323) = 1; SYSCTL_INT(_net_inet_tcp, TCPCTL_DO_RFC1323, rfc1323, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_do_rfc1323), 0, "Enable rfc1323 (high performance TCP) extensions"); /* * As of June 2021, several TCP stacks violate RFC 7323 from September 2014. * Some stacks negotiate TS, but never send them after connection setup. Some * stacks negotiate TS, but don't send them when sending keep-alive segments. * These include modern widely deployed TCP stacks. * Therefore tolerating violations for now... */ VNET_DEFINE(int, tcp_tolerate_missing_ts) = 1; SYSCTL_INT(_net_inet_tcp, OID_AUTO, tolerate_missing_ts, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_tolerate_missing_ts), 0, "Tolerate missing TCP timestamps"); VNET_DEFINE(int, tcp_ts_offset_per_conn) = 1; SYSCTL_INT(_net_inet_tcp, OID_AUTO, ts_offset_per_conn, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_ts_offset_per_conn), 0, "Initialize TCP timestamps per connection instead of per host pair"); /* How many connections are pacing */ static volatile uint32_t number_of_tcp_connections_pacing = 0; static uint32_t shadow_num_connections = 0; static int tcp_pacing_limit = 10000; SYSCTL_INT(_net_inet_tcp, OID_AUTO, pacing_limit, CTLFLAG_RW, &tcp_pacing_limit, 1000, "If the TCP stack does pacing, is there a limit (-1 = no, 0 = no pacing N = number of connections)"); SYSCTL_UINT(_net_inet_tcp, OID_AUTO, pacing_count, CTLFLAG_RD, &shadow_num_connections, 0, "Number of TCP connections being paced"); static int tcp_log_debug = 0; SYSCTL_INT(_net_inet_tcp, OID_AUTO, log_debug, CTLFLAG_RW, &tcp_log_debug, 0, "Log errors caused by incoming TCP segments"); static int tcp_tcbhashsize; SYSCTL_INT(_net_inet_tcp, OID_AUTO, tcbhashsize, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &tcp_tcbhashsize, 0, "Size of TCP control-block hashtable"); static int do_tcpdrain = 1; SYSCTL_INT(_net_inet_tcp, OID_AUTO, do_tcpdrain, CTLFLAG_RW, &do_tcpdrain, 0, "Enable tcp_drain routine for extra help when low on mbufs"); SYSCTL_UINT(_net_inet_tcp, OID_AUTO, pcbcount, CTLFLAG_VNET | CTLFLAG_RD, &VNET_NAME(tcbinfo.ipi_count), 0, "Number of active PCBs"); VNET_DEFINE_STATIC(int, icmp_may_rst) = 1; #define V_icmp_may_rst VNET(icmp_may_rst) SYSCTL_INT(_net_inet_tcp, OID_AUTO, icmp_may_rst, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(icmp_may_rst), 0, "Certain ICMP unreachable messages may abort connections in SYN_SENT"); VNET_DEFINE_STATIC(int, tcp_isn_reseed_interval) = 0; #define V_tcp_isn_reseed_interval VNET(tcp_isn_reseed_interval) SYSCTL_INT(_net_inet_tcp, OID_AUTO, isn_reseed_interval, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_isn_reseed_interval), 0, "Seconds between reseeding of ISN secret"); static int tcp_soreceive_stream; SYSCTL_INT(_net_inet_tcp, OID_AUTO, soreceive_stream, CTLFLAG_RDTUN, &tcp_soreceive_stream, 0, "Using soreceive_stream for TCP sockets"); VNET_DEFINE(uma_zone_t, sack_hole_zone); #define V_sack_hole_zone VNET(sack_hole_zone) VNET_DEFINE(uint32_t, tcp_map_entries_limit) = 0; /* unlimited */ static int sysctl_net_inet_tcp_map_limit_check(SYSCTL_HANDLER_ARGS) { int error; uint32_t new; new = V_tcp_map_entries_limit; error = sysctl_handle_int(oidp, &new, 0, req); if (error == 0 && req->newptr) { /* only allow "0" and value > minimum */ if (new > 0 && new < TCP_MIN_MAP_ENTRIES_LIMIT) error = EINVAL; else V_tcp_map_entries_limit = new; } return (error); } SYSCTL_PROC(_net_inet_tcp, OID_AUTO, map_limit, CTLFLAG_VNET | CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &VNET_NAME(tcp_map_entries_limit), 0, &sysctl_net_inet_tcp_map_limit_check, "IU", "Total sendmap entries limit"); VNET_DEFINE(uint32_t, tcp_map_split_limit) = 0; /* unlimited */ SYSCTL_UINT(_net_inet_tcp, OID_AUTO, split_limit, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_map_split_limit), 0, "Total sendmap split entries limit"); #ifdef TCP_HHOOK VNET_DEFINE(struct hhook_head *, tcp_hhh[HHOOK_TCP_LAST+1]); #endif #define TS_OFFSET_SECRET_LENGTH SIPHASH_KEY_LENGTH VNET_DEFINE_STATIC(u_char, ts_offset_secret[TS_OFFSET_SECRET_LENGTH]); #define V_ts_offset_secret VNET(ts_offset_secret) static int tcp_default_fb_init(struct tcpcb *tp); static void tcp_default_fb_fini(struct tcpcb *tp, int tcb_is_purged); static int tcp_default_handoff_ok(struct tcpcb *tp); static struct inpcb *tcp_notify(struct inpcb *, int); static struct inpcb *tcp_mtudisc_notify(struct inpcb *, int); static struct inpcb *tcp_mtudisc(struct inpcb *, int); static struct inpcb *tcp_drop_syn_sent(struct inpcb *, int); static char * tcp_log_addr(struct in_conninfo *inc, struct tcphdr *th, const void *ip4hdr, const void *ip6hdr); static ipproto_ctlinput_t tcp_ctlinput; static udp_tun_icmp_t tcp_ctlinput_viaudp; static struct tcp_function_block tcp_def_funcblk = { .tfb_tcp_block_name = "freebsd", .tfb_tcp_output = tcp_default_output, .tfb_tcp_do_segment = tcp_do_segment, .tfb_tcp_ctloutput = tcp_default_ctloutput, .tfb_tcp_handoff_ok = tcp_default_handoff_ok, .tfb_tcp_fb_init = tcp_default_fb_init, .tfb_tcp_fb_fini = tcp_default_fb_fini, }; static int tcp_fb_cnt = 0; struct tcp_funchead t_functions; static struct tcp_function_block *tcp_func_set_ptr = &tcp_def_funcblk; void tcp_record_dsack(struct tcpcb *tp, tcp_seq start, tcp_seq end, int tlp) { TCPSTAT_INC(tcps_dsack_count); tp->t_dsack_pack++; if (tlp == 0) { if (SEQ_GT(end, start)) { tp->t_dsack_bytes += (end - start); TCPSTAT_ADD(tcps_dsack_bytes, (end - start)); } else { tp->t_dsack_tlp_bytes += (start - end); TCPSTAT_ADD(tcps_dsack_bytes, (start - end)); } } else { if (SEQ_GT(end, start)) { tp->t_dsack_bytes += (end - start); TCPSTAT_ADD(tcps_dsack_tlp_bytes, (end - start)); } else { tp->t_dsack_tlp_bytes += (start - end); TCPSTAT_ADD(tcps_dsack_tlp_bytes, (start - end)); } } } static struct tcp_function_block * find_tcp_functions_locked(struct tcp_function_set *fs) { struct tcp_function *f; struct tcp_function_block *blk=NULL; TAILQ_FOREACH(f, &t_functions, tf_next) { if (strcmp(f->tf_name, fs->function_set_name) == 0) { blk = f->tf_fb; break; } } return(blk); } static struct tcp_function_block * find_tcp_fb_locked(struct tcp_function_block *blk, struct tcp_function **s) { struct tcp_function_block *rblk=NULL; struct tcp_function *f; TAILQ_FOREACH(f, &t_functions, tf_next) { if (f->tf_fb == blk) { rblk = blk; if (s) { *s = f; } break; } } return (rblk); } struct tcp_function_block * find_and_ref_tcp_functions(struct tcp_function_set *fs) { struct tcp_function_block *blk; rw_rlock(&tcp_function_lock); blk = find_tcp_functions_locked(fs); if (blk) refcount_acquire(&blk->tfb_refcnt); rw_runlock(&tcp_function_lock); return(blk); } struct tcp_function_block * find_and_ref_tcp_fb(struct tcp_function_block *blk) { struct tcp_function_block *rblk; rw_rlock(&tcp_function_lock); rblk = find_tcp_fb_locked(blk, NULL); if (rblk) refcount_acquire(&rblk->tfb_refcnt); rw_runlock(&tcp_function_lock); return(rblk); } /* Find a matching alias for the given tcp_function_block. */ int find_tcp_function_alias(struct tcp_function_block *blk, struct tcp_function_set *fs) { struct tcp_function *f; int found; found = 0; rw_rlock(&tcp_function_lock); TAILQ_FOREACH(f, &t_functions, tf_next) { if ((f->tf_fb == blk) && (strncmp(f->tf_name, blk->tfb_tcp_block_name, TCP_FUNCTION_NAME_LEN_MAX) != 0)) { /* Matching function block with different name. */ strncpy(fs->function_set_name, f->tf_name, TCP_FUNCTION_NAME_LEN_MAX); found = 1; break; } } /* Null terminate the string appropriately. */ if (found) { fs->function_set_name[TCP_FUNCTION_NAME_LEN_MAX - 1] = '\0'; } else { fs->function_set_name[0] = '\0'; } rw_runlock(&tcp_function_lock); return (found); } static struct tcp_function_block * find_and_ref_tcp_default_fb(void) { struct tcp_function_block *rblk; rw_rlock(&tcp_function_lock); rblk = tcp_func_set_ptr; refcount_acquire(&rblk->tfb_refcnt); rw_runlock(&tcp_function_lock); return (rblk); } void tcp_switch_back_to_default(struct tcpcb *tp) { struct tcp_function_block *tfb; KASSERT(tp->t_fb != &tcp_def_funcblk, ("%s: called by the built-in default stack", __func__)); /* * Release the old stack. This function will either find a new one * or panic. */ if (tp->t_fb->tfb_tcp_fb_fini != NULL) (*tp->t_fb->tfb_tcp_fb_fini)(tp, 0); refcount_release(&tp->t_fb->tfb_refcnt); /* * Now, we'll find a new function block to use. * Start by trying the current user-selected * default, unless this stack is the user-selected * default. */ tfb = find_and_ref_tcp_default_fb(); if (tfb == tp->t_fb) { refcount_release(&tfb->tfb_refcnt); tfb = NULL; } /* Does the stack accept this connection? */ if (tfb != NULL && tfb->tfb_tcp_handoff_ok != NULL && (*tfb->tfb_tcp_handoff_ok)(tp)) { refcount_release(&tfb->tfb_refcnt); tfb = NULL; } /* Try to use that stack. */ if (tfb != NULL) { /* Initialize the new stack. If it succeeds, we are done. */ tp->t_fb = tfb; if (tp->t_fb->tfb_tcp_fb_init == NULL || (*tp->t_fb->tfb_tcp_fb_init)(tp) == 0) return; /* * Initialization failed. Release the reference count on * the stack. */ refcount_release(&tfb->tfb_refcnt); } /* * If that wasn't feasible, use the built-in default * stack which is not allowed to reject anyone. */ tfb = find_and_ref_tcp_fb(&tcp_def_funcblk); if (tfb == NULL) { /* there always should be a default */ panic("Can't refer to tcp_def_funcblk"); } if (tfb->tfb_tcp_handoff_ok != NULL) { if ((*tfb->tfb_tcp_handoff_ok) (tp)) { /* The default stack cannot say no */ panic("Default stack rejects a new session?"); } } tp->t_fb = tfb; if (tp->t_fb->tfb_tcp_fb_init != NULL && (*tp->t_fb->tfb_tcp_fb_init)(tp)) { /* The default stack cannot fail */ panic("Default stack initialization failed"); } } static bool tcp_recv_udp_tunneled_packet(struct mbuf *m, int off, struct inpcb *inp, const struct sockaddr *sa, void *ctx) { struct ip *iph; #ifdef INET6 struct ip6_hdr *ip6; #endif struct udphdr *uh; struct tcphdr *th; int thlen; uint16_t port; TCPSTAT_INC(tcps_tunneled_pkts); if ((m->m_flags & M_PKTHDR) == 0) { /* Can't handle one that is not a pkt hdr */ TCPSTAT_INC(tcps_tunneled_errs); goto out; } thlen = sizeof(struct tcphdr); if (m->m_len < off + sizeof(struct udphdr) + thlen && (m = m_pullup(m, off + sizeof(struct udphdr) + thlen)) == NULL) { TCPSTAT_INC(tcps_tunneled_errs); goto out; } iph = mtod(m, struct ip *); uh = (struct udphdr *)((caddr_t)iph + off); th = (struct tcphdr *)(uh + 1); thlen = th->th_off << 2; if (m->m_len < off + sizeof(struct udphdr) + thlen) { m = m_pullup(m, off + sizeof(struct udphdr) + thlen); if (m == NULL) { TCPSTAT_INC(tcps_tunneled_errs); goto out; } else { iph = mtod(m, struct ip *); uh = (struct udphdr *)((caddr_t)iph + off); th = (struct tcphdr *)(uh + 1); } } m->m_pkthdr.tcp_tun_port = port = uh->uh_sport; bcopy(th, uh, m->m_len - off); m->m_len -= sizeof(struct udphdr); m->m_pkthdr.len -= sizeof(struct udphdr); /* * We use the same algorithm for * both UDP and TCP for c-sum. So * the code in tcp_input will skip * the checksum. So we do nothing * with the flag (m->m_pkthdr.csum_flags). */ switch (iph->ip_v) { #ifdef INET case IPVERSION: iph->ip_len = htons(ntohs(iph->ip_len) - sizeof(struct udphdr)); tcp_input_with_port(&m, &off, IPPROTO_TCP, port); break; #endif #ifdef INET6 case IPV6_VERSION >> 4: ip6 = mtod(m, struct ip6_hdr *); ip6->ip6_plen = htons(ntohs(ip6->ip6_plen) - sizeof(struct udphdr)); tcp6_input_with_port(&m, &off, IPPROTO_TCP, port); break; #endif default: goto out; break; } return (true); out: m_freem(m); return (true); } static int sysctl_net_inet_default_tcp_functions(SYSCTL_HANDLER_ARGS) { int error=ENOENT; struct tcp_function_set fs; struct tcp_function_block *blk; memset(&fs, 0, sizeof(fs)); rw_rlock(&tcp_function_lock); blk = find_tcp_fb_locked(tcp_func_set_ptr, NULL); if (blk) { /* Found him */ strcpy(fs.function_set_name, blk->tfb_tcp_block_name); fs.pcbcnt = blk->tfb_refcnt; } rw_runlock(&tcp_function_lock); error = sysctl_handle_string(oidp, fs.function_set_name, sizeof(fs.function_set_name), req); /* Check for error or no change */ if (error != 0 || req->newptr == NULL) return(error); rw_wlock(&tcp_function_lock); blk = find_tcp_functions_locked(&fs); if ((blk == NULL) || (blk->tfb_flags & TCP_FUNC_BEING_REMOVED)) { error = ENOENT; goto done; } tcp_func_set_ptr = blk; done: rw_wunlock(&tcp_function_lock); return (error); } SYSCTL_PROC(_net_inet_tcp, OID_AUTO, functions_default, CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_NEEDGIANT, NULL, 0, sysctl_net_inet_default_tcp_functions, "A", "Set/get the default TCP functions"); static int sysctl_net_inet_list_available(SYSCTL_HANDLER_ARGS) { int error, cnt, linesz; struct tcp_function *f; char *buffer, *cp; size_t bufsz, outsz; bool alias; cnt = 0; rw_rlock(&tcp_function_lock); TAILQ_FOREACH(f, &t_functions, tf_next) { cnt++; } rw_runlock(&tcp_function_lock); bufsz = (cnt+2) * ((TCP_FUNCTION_NAME_LEN_MAX * 2) + 13) + 1; buffer = malloc(bufsz, M_TEMP, M_WAITOK); error = 0; cp = buffer; linesz = snprintf(cp, bufsz, "\n%-32s%c %-32s %s\n", "Stack", 'D', "Alias", "PCB count"); cp += linesz; bufsz -= linesz; outsz = linesz; rw_rlock(&tcp_function_lock); TAILQ_FOREACH(f, &t_functions, tf_next) { alias = (f->tf_name != f->tf_fb->tfb_tcp_block_name); linesz = snprintf(cp, bufsz, "%-32s%c %-32s %u\n", f->tf_fb->tfb_tcp_block_name, (f->tf_fb == tcp_func_set_ptr) ? '*' : ' ', alias ? f->tf_name : "-", f->tf_fb->tfb_refcnt); if (linesz >= bufsz) { error = EOVERFLOW; break; } cp += linesz; bufsz -= linesz; outsz += linesz; } rw_runlock(&tcp_function_lock); if (error == 0) error = sysctl_handle_string(oidp, buffer, outsz + 1, req); free(buffer, M_TEMP); return (error); } SYSCTL_PROC(_net_inet_tcp, OID_AUTO, functions_available, CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT, NULL, 0, sysctl_net_inet_list_available, "A", "list available TCP Function sets"); VNET_DEFINE(int, tcp_udp_tunneling_port) = TCP_TUNNELING_PORT_DEFAULT; #ifdef INET VNET_DEFINE(struct socket *, udp4_tun_socket) = NULL; #define V_udp4_tun_socket VNET(udp4_tun_socket) #endif #ifdef INET6 VNET_DEFINE(struct socket *, udp6_tun_socket) = NULL; #define V_udp6_tun_socket VNET(udp6_tun_socket) #endif static void tcp_over_udp_stop(void) { /* * This function assumes sysctl caller holds inp_rinfo_lock() * for writing! */ #ifdef INET if (V_udp4_tun_socket != NULL) { soclose(V_udp4_tun_socket); V_udp4_tun_socket = NULL; } #endif #ifdef INET6 if (V_udp6_tun_socket != NULL) { soclose(V_udp6_tun_socket); V_udp6_tun_socket = NULL; } #endif } static int tcp_over_udp_start(void) { uint16_t port; int ret; #ifdef INET struct sockaddr_in sin; #endif #ifdef INET6 struct sockaddr_in6 sin6; #endif /* * This function assumes sysctl caller holds inp_info_rlock() * for writing! */ port = V_tcp_udp_tunneling_port; if (ntohs(port) == 0) { /* Must have a port set */ return (EINVAL); } #ifdef INET if (V_udp4_tun_socket != NULL) { /* Already running -- must stop first */ return (EALREADY); } #endif #ifdef INET6 if (V_udp6_tun_socket != NULL) { /* Already running -- must stop first */ return (EALREADY); } #endif #ifdef INET if ((ret = socreate(PF_INET, &V_udp4_tun_socket, SOCK_DGRAM, IPPROTO_UDP, curthread->td_ucred, curthread))) { tcp_over_udp_stop(); return (ret); } /* Call the special UDP hook. */ if ((ret = udp_set_kernel_tunneling(V_udp4_tun_socket, tcp_recv_udp_tunneled_packet, tcp_ctlinput_viaudp, NULL))) { tcp_over_udp_stop(); return (ret); } /* Ok, we have a socket, bind it to the port. */ memset(&sin, 0, sizeof(struct sockaddr_in)); sin.sin_len = sizeof(struct sockaddr_in); sin.sin_family = AF_INET; sin.sin_port = htons(port); if ((ret = sobind(V_udp4_tun_socket, (struct sockaddr *)&sin, curthread))) { tcp_over_udp_stop(); return (ret); } #endif #ifdef INET6 if ((ret = socreate(PF_INET6, &V_udp6_tun_socket, SOCK_DGRAM, IPPROTO_UDP, curthread->td_ucred, curthread))) { tcp_over_udp_stop(); return (ret); } /* Call the special UDP hook. */ if ((ret = udp_set_kernel_tunneling(V_udp6_tun_socket, tcp_recv_udp_tunneled_packet, tcp6_ctlinput_viaudp, NULL))) { tcp_over_udp_stop(); return (ret); } /* Ok, we have a socket, bind it to the port. */ memset(&sin6, 0, sizeof(struct sockaddr_in6)); sin6.sin6_len = sizeof(struct sockaddr_in6); sin6.sin6_family = AF_INET6; sin6.sin6_port = htons(port); if ((ret = sobind(V_udp6_tun_socket, (struct sockaddr *)&sin6, curthread))) { tcp_over_udp_stop(); return (ret); } #endif return (0); } static int sysctl_net_inet_tcp_udp_tunneling_port_check(SYSCTL_HANDLER_ARGS) { int error; uint32_t old, new; old = V_tcp_udp_tunneling_port; new = old; error = sysctl_handle_int(oidp, &new, 0, req); if ((error == 0) && (req->newptr != NULL)) { if ((new < TCP_TUNNELING_PORT_MIN) || (new > TCP_TUNNELING_PORT_MAX)) { error = EINVAL; } else { V_tcp_udp_tunneling_port = new; if (old != 0) { tcp_over_udp_stop(); } if (new != 0) { error = tcp_over_udp_start(); } } } return (error); } SYSCTL_PROC(_net_inet_tcp, OID_AUTO, udp_tunneling_port, CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, &VNET_NAME(tcp_udp_tunneling_port), 0, &sysctl_net_inet_tcp_udp_tunneling_port_check, "IU", "Tunneling port for tcp over udp"); VNET_DEFINE(int, tcp_udp_tunneling_overhead) = TCP_TUNNELING_OVERHEAD_DEFAULT; static int sysctl_net_inet_tcp_udp_tunneling_overhead_check(SYSCTL_HANDLER_ARGS) { int error, new; new = V_tcp_udp_tunneling_overhead; error = sysctl_handle_int(oidp, &new, 0, req); if (error == 0 && req->newptr) { if ((new < TCP_TUNNELING_OVERHEAD_MIN) || (new > TCP_TUNNELING_OVERHEAD_MAX)) error = EINVAL; else V_tcp_udp_tunneling_overhead = new; } return (error); } SYSCTL_PROC(_net_inet_tcp, OID_AUTO, udp_tunneling_overhead, CTLFLAG_VNET | CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, &VNET_NAME(tcp_udp_tunneling_overhead), 0, &sysctl_net_inet_tcp_udp_tunneling_overhead_check, "IU", "MSS reduction when using tcp over udp"); /* * Exports one (struct tcp_function_info) for each alias/name. */ static int sysctl_net_inet_list_func_info(SYSCTL_HANDLER_ARGS) { int cnt, error; struct tcp_function *f; struct tcp_function_info tfi; /* * We don't allow writes. */ if (req->newptr != NULL) return (EINVAL); /* * Wire the old buffer so we can directly copy the functions to * user space without dropping the lock. */ if (req->oldptr != NULL) { error = sysctl_wire_old_buffer(req, 0); if (error) return (error); } /* * Walk the list and copy out matching entries. If INVARIANTS * is compiled in, also walk the list to verify the length of * the list matches what we have recorded. */ rw_rlock(&tcp_function_lock); cnt = 0; #ifndef INVARIANTS if (req->oldptr == NULL) { cnt = tcp_fb_cnt; goto skip_loop; } #endif TAILQ_FOREACH(f, &t_functions, tf_next) { #ifdef INVARIANTS cnt++; #endif if (req->oldptr != NULL) { bzero(&tfi, sizeof(tfi)); tfi.tfi_refcnt = f->tf_fb->tfb_refcnt; tfi.tfi_id = f->tf_fb->tfb_id; (void)strlcpy(tfi.tfi_alias, f->tf_name, sizeof(tfi.tfi_alias)); (void)strlcpy(tfi.tfi_name, f->tf_fb->tfb_tcp_block_name, sizeof(tfi.tfi_name)); error = SYSCTL_OUT(req, &tfi, sizeof(tfi)); /* * Don't stop on error, as that is the * mechanism we use to accumulate length * information if the buffer was too short. */ } } KASSERT(cnt == tcp_fb_cnt, ("%s: cnt (%d) != tcp_fb_cnt (%d)", __func__, cnt, tcp_fb_cnt)); #ifndef INVARIANTS skip_loop: #endif rw_runlock(&tcp_function_lock); if (req->oldptr == NULL) error = SYSCTL_OUT(req, NULL, (cnt + 1) * sizeof(struct tcp_function_info)); return (error); } SYSCTL_PROC(_net_inet_tcp, OID_AUTO, function_info, CTLTYPE_OPAQUE | CTLFLAG_SKIP | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0, sysctl_net_inet_list_func_info, "S,tcp_function_info", "List TCP function block name-to-ID mappings"); /* * tfb_tcp_handoff_ok() function for the default stack. * Note that we'll basically try to take all comers. */ static int tcp_default_handoff_ok(struct tcpcb *tp) { return (0); } /* * tfb_tcp_fb_init() function for the default stack. * * This handles making sure we have appropriate timers set if you are * transitioning a socket that has some amount of setup done. * * The init() fuction from the default can *never* return non-zero i.e. * it is required to always succeed since it is the stack of last resort! */ static int tcp_default_fb_init(struct tcpcb *tp) { struct socket *so; INP_WLOCK_ASSERT(tp->t_inpcb); KASSERT(tp->t_state >= 0 && tp->t_state < TCPS_TIME_WAIT, ("%s: connection %p in unexpected state %d", __func__, tp, tp->t_state)); /* * Nothing to do for ESTABLISHED or LISTEN states. And, we don't * know what to do for unexpected states (which includes TIME_WAIT). */ if (tp->t_state <= TCPS_LISTEN || tp->t_state >= TCPS_TIME_WAIT) return (0); /* * Make sure some kind of transmission timer is set if there is * outstanding data. */ so = tp->t_inpcb->inp_socket; if ((!TCPS_HAVEESTABLISHED(tp->t_state) || sbavail(&so->so_snd) || tp->snd_una != tp->snd_max) && !(tcp_timer_active(tp, TT_REXMT) || tcp_timer_active(tp, TT_PERSIST))) { /* * If the session has established and it looks like it should * be in the persist state, set the persist timer. Otherwise, * set the retransmit timer. */ if (TCPS_HAVEESTABLISHED(tp->t_state) && tp->snd_wnd == 0 && (int32_t)(tp->snd_nxt - tp->snd_una) < (int32_t)sbavail(&so->so_snd)) tcp_setpersist(tp); else tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); } /* All non-embryonic sessions get a keepalive timer. */ if (!tcp_timer_active(tp, TT_KEEP)) tcp_timer_activate(tp, TT_KEEP, TCPS_HAVEESTABLISHED(tp->t_state) ? TP_KEEPIDLE(tp) : TP_KEEPINIT(tp)); /* * Make sure critical variables are initialized * if transitioning while in Recovery. */ if IN_FASTRECOVERY(tp->t_flags) { if (tp->sackhint.recover_fs == 0) tp->sackhint.recover_fs = max(1, tp->snd_nxt - tp->snd_una); } return (0); } /* * tfb_tcp_fb_fini() function for the default stack. * * This changes state as necessary (or prudent) to prepare for another stack * to assume responsibility for the connection. */ static void tcp_default_fb_fini(struct tcpcb *tp, int tcb_is_purged) { INP_WLOCK_ASSERT(tp->t_inpcb); return; } /* * Target size of TCP PCB hash tables. Must be a power of two. * * Note that this can be overridden by the kernel environment * variable net.inet.tcp.tcbhashsize */ #ifndef TCBHASHSIZE #define TCBHASHSIZE 0 #endif /* * XXX * Callouts should be moved into struct tcp directly. They are currently * separate because the tcpcb structure is exported to userland for sysctl * parsing purposes, which do not know about callouts. */ struct tcpcb_mem { struct tcpcb tcb; struct tcp_timer tt; struct cc_var ccv; #ifdef TCP_HHOOK struct osd osd; #endif }; VNET_DEFINE_STATIC(uma_zone_t, tcpcb_zone); #define V_tcpcb_zone VNET(tcpcb_zone) MALLOC_DEFINE(M_TCPLOG, "tcplog", "TCP address and flags print buffers"); MALLOC_DEFINE(M_TCPFUNCTIONS, "tcpfunc", "TCP function set memory"); static struct mtx isn_mtx; #define ISN_LOCK_INIT() mtx_init(&isn_mtx, "isn_mtx", NULL, MTX_DEF) #define ISN_LOCK() mtx_lock(&isn_mtx) #define ISN_UNLOCK() mtx_unlock(&isn_mtx) INPCBSTORAGE_DEFINE(tcpcbstor, "tcpinp", "tcp_inpcb", "tcp", "tcphash"); /* * Take a value and get the next power of 2 that doesn't overflow. * Used to size the tcp_inpcb hash buckets. */ static int maketcp_hashsize(int size) { int hashsize; /* * auto tune. * get the next power of 2 higher than maxsockets. */ hashsize = 1 << fls(size); /* catch overflow, and just go one power of 2 smaller */ if (hashsize < size) { hashsize = 1 << (fls(size) - 1); } return (hashsize); } static volatile int next_tcp_stack_id = 1; /* * Register a TCP function block with the name provided in the names * array. (Note that this function does NOT automatically register * blk->tfb_tcp_block_name as a stack name. Therefore, you should * explicitly include blk->tfb_tcp_block_name in the list of names if * you wish to register the stack with that name.) * * Either all name registrations will succeed or all will fail. If * a name registration fails, the function will update the num_names * argument to point to the array index of the name that encountered * the failure. * * Returns 0 on success, or an error code on failure. */ int register_tcp_functions_as_names(struct tcp_function_block *blk, int wait, const char *names[], int *num_names) { struct tcp_function *n; struct tcp_function_set fs; int error, i; KASSERT(names != NULL && *num_names > 0, ("%s: Called with 0-length name list", __func__)); KASSERT(names != NULL, ("%s: Called with NULL name list", __func__)); KASSERT(rw_initialized(&tcp_function_lock), ("%s: called too early", __func__)); if ((blk->tfb_tcp_output == NULL) || (blk->tfb_tcp_do_segment == NULL) || (blk->tfb_tcp_ctloutput == NULL) || (strlen(blk->tfb_tcp_block_name) == 0)) { /* * These functions are required and you * need a name. */ *num_names = 0; return (EINVAL); } if (blk->tfb_tcp_timer_stop_all || blk->tfb_tcp_timer_activate || blk->tfb_tcp_timer_active || blk->tfb_tcp_timer_stop) { /* * If you define one timer function you * must have them all. */ if ((blk->tfb_tcp_timer_stop_all == NULL) || (blk->tfb_tcp_timer_activate == NULL) || (blk->tfb_tcp_timer_active == NULL) || (blk->tfb_tcp_timer_stop == NULL)) { *num_names = 0; return (EINVAL); } } if (blk->tfb_flags & TCP_FUNC_BEING_REMOVED) { *num_names = 0; return (EINVAL); } refcount_init(&blk->tfb_refcnt, 0); blk->tfb_id = atomic_fetchadd_int(&next_tcp_stack_id, 1); for (i = 0; i < *num_names; i++) { n = malloc(sizeof(struct tcp_function), M_TCPFUNCTIONS, wait); if (n == NULL) { error = ENOMEM; goto cleanup; } n->tf_fb = blk; (void)strlcpy(fs.function_set_name, names[i], sizeof(fs.function_set_name)); rw_wlock(&tcp_function_lock); if (find_tcp_functions_locked(&fs) != NULL) { /* Duplicate name space not allowed */ rw_wunlock(&tcp_function_lock); free(n, M_TCPFUNCTIONS); error = EALREADY; goto cleanup; } (void)strlcpy(n->tf_name, names[i], sizeof(n->tf_name)); TAILQ_INSERT_TAIL(&t_functions, n, tf_next); tcp_fb_cnt++; rw_wunlock(&tcp_function_lock); } return(0); cleanup: /* * Deregister the names we just added. Because registration failed * for names[i], we don't need to deregister that name. */ *num_names = i; rw_wlock(&tcp_function_lock); while (--i >= 0) { TAILQ_FOREACH(n, &t_functions, tf_next) { if (!strncmp(n->tf_name, names[i], TCP_FUNCTION_NAME_LEN_MAX)) { TAILQ_REMOVE(&t_functions, n, tf_next); tcp_fb_cnt--; n->tf_fb = NULL; free(n, M_TCPFUNCTIONS); break; } } } rw_wunlock(&tcp_function_lock); return (error); } /* * Register a TCP function block using the name provided in the name * argument. * * Returns 0 on success, or an error code on failure. */ int register_tcp_functions_as_name(struct tcp_function_block *blk, const char *name, int wait) { const char *name_list[1]; int num_names, rv; num_names = 1; if (name != NULL) name_list[0] = name; else name_list[0] = blk->tfb_tcp_block_name; rv = register_tcp_functions_as_names(blk, wait, name_list, &num_names); return (rv); } /* * Register a TCP function block using the name defined in * blk->tfb_tcp_block_name. * * Returns 0 on success, or an error code on failure. */ int register_tcp_functions(struct tcp_function_block *blk, int wait) { return (register_tcp_functions_as_name(blk, NULL, wait)); } /* * Deregister all names associated with a function block. This * functionally removes the function block from use within the system. * * When called with a true quiesce argument, mark the function block * as being removed so no more stacks will use it and determine * whether the removal would succeed. * * When called with a false quiesce argument, actually attempt the * removal. * * When called with a force argument, attempt to switch all TCBs to * use the default stack instead of returning EBUSY. * * Returns 0 on success (or if the removal would succeed, or an error * code on failure. */ int deregister_tcp_functions(struct tcp_function_block *blk, bool quiesce, bool force) { struct tcp_function *f; if (blk == &tcp_def_funcblk) { /* You can't un-register the default */ return (EPERM); } rw_wlock(&tcp_function_lock); if (blk == tcp_func_set_ptr) { /* You can't free the current default */ rw_wunlock(&tcp_function_lock); return (EBUSY); } /* Mark the block so no more stacks can use it. */ blk->tfb_flags |= TCP_FUNC_BEING_REMOVED; /* * If TCBs are still attached to the stack, attempt to switch them * to the default stack. */ if (force && blk->tfb_refcnt) { struct inpcb_iterator inpi = INP_ALL_ITERATOR(&V_tcbinfo, INPLOOKUP_WLOCKPCB); struct inpcb *inp; struct tcpcb *tp; VNET_ITERATOR_DECL(vnet_iter); rw_wunlock(&tcp_function_lock); VNET_LIST_RLOCK(); VNET_FOREACH(vnet_iter) { CURVNET_SET(vnet_iter); while ((inp = inp_next(&inpi)) != NULL) { - if (inp->inp_flags & INP_TIMEWAIT) - continue; tp = intotcpcb(inp); if (tp == NULL || tp->t_fb != blk) continue; tcp_switch_back_to_default(tp); } CURVNET_RESTORE(); } VNET_LIST_RUNLOCK(); rw_wlock(&tcp_function_lock); } if (blk->tfb_refcnt) { /* TCBs still attached. */ rw_wunlock(&tcp_function_lock); return (EBUSY); } if (quiesce) { /* Skip removal. */ rw_wunlock(&tcp_function_lock); return (0); } /* Remove any function names that map to this function block. */ while (find_tcp_fb_locked(blk, &f) != NULL) { TAILQ_REMOVE(&t_functions, f, tf_next); tcp_fb_cnt--; f->tf_fb = NULL; free(f, M_TCPFUNCTIONS); } rw_wunlock(&tcp_function_lock); return (0); } static void tcp_drain(void) { struct epoch_tracker et; VNET_ITERATOR_DECL(vnet_iter); if (!do_tcpdrain) return; NET_EPOCH_ENTER(et); VNET_LIST_RLOCK_NOSLEEP(); VNET_FOREACH(vnet_iter) { CURVNET_SET(vnet_iter); struct inpcb_iterator inpi = INP_ALL_ITERATOR(&V_tcbinfo, INPLOOKUP_WLOCKPCB); struct inpcb *inpb; struct tcpcb *tcpb; /* * Walk the tcpbs, if existing, and flush the reassembly queue, * if there is one... * XXX: The "Net/3" implementation doesn't imply that the TCP * reassembly queue should be flushed, but in a situation * where we're really low on mbufs, this is potentially * useful. */ while ((inpb = inp_next(&inpi)) != NULL) { - if (inpb->inp_flags & INP_TIMEWAIT) - continue; if ((tcpb = intotcpcb(inpb)) != NULL) { tcp_reass_flush(tcpb); tcp_clean_sackreport(tcpb); #ifdef TCP_BLACKBOX tcp_log_drain(tcpb); #endif #ifdef TCPPCAP if (tcp_pcap_aggressive_free) { /* Free the TCP PCAP queues. */ tcp_pcap_drain(&(tcpb->t_inpkts)); tcp_pcap_drain(&(tcpb->t_outpkts)); } #endif } } CURVNET_RESTORE(); } VNET_LIST_RUNLOCK_NOSLEEP(); NET_EPOCH_EXIT(et); } static void tcp_vnet_init(void *arg __unused) { #ifdef TCP_HHOOK if (hhook_head_register(HHOOK_TYPE_TCP, HHOOK_TCP_EST_IN, &V_tcp_hhh[HHOOK_TCP_EST_IN], HHOOK_NOWAIT|HHOOK_HEADISINVNET) != 0) printf("%s: WARNING: unable to register helper hook\n", __func__); if (hhook_head_register(HHOOK_TYPE_TCP, HHOOK_TCP_EST_OUT, &V_tcp_hhh[HHOOK_TCP_EST_OUT], HHOOK_NOWAIT|HHOOK_HEADISINVNET) != 0) printf("%s: WARNING: unable to register helper hook\n", __func__); #endif #ifdef STATS if (tcp_stats_init()) printf("%s: WARNING: unable to initialise TCP stats\n", __func__); #endif in_pcbinfo_init(&V_tcbinfo, &tcpcbstor, tcp_tcbhashsize, tcp_tcbhashsize); /* * These have to be type stable for the benefit of the timers. */ V_tcpcb_zone = uma_zcreate("tcpcb", sizeof(struct tcpcb_mem), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); uma_zone_set_max(V_tcpcb_zone, maxsockets); uma_zone_set_warning(V_tcpcb_zone, "kern.ipc.maxsockets limit reached"); syncache_init(); tcp_hc_init(); TUNABLE_INT_FETCH("net.inet.tcp.sack.enable", &V_tcp_do_sack); V_sack_hole_zone = uma_zcreate("sackhole", sizeof(struct sackhole), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); tcp_fastopen_init(); COUNTER_ARRAY_ALLOC(V_tcps_states, TCP_NSTATES, M_WAITOK); VNET_PCPUSTAT_ALLOC(tcpstat, M_WAITOK); V_tcp_msl = TCPTV_MSL; } VNET_SYSINIT(tcp_vnet_init, SI_SUB_PROTO_DOMAIN, SI_ORDER_FOURTH, tcp_vnet_init, NULL); static void tcp_init(void *arg __unused) { const char *tcbhash_tuneable; int hashsize; tcp_reass_global_init(); /* XXX virtualize those below? */ tcp_delacktime = TCPTV_DELACK; tcp_keepinit = TCPTV_KEEP_INIT; tcp_keepidle = TCPTV_KEEP_IDLE; tcp_keepintvl = TCPTV_KEEPINTVL; tcp_maxpersistidle = TCPTV_KEEP_IDLE; tcp_rexmit_initial = TCPTV_RTOBASE; if (tcp_rexmit_initial < 1) tcp_rexmit_initial = 1; tcp_rexmit_min = TCPTV_MIN; if (tcp_rexmit_min < 1) tcp_rexmit_min = 1; tcp_persmin = TCPTV_PERSMIN; tcp_persmax = TCPTV_PERSMAX; tcp_rexmit_slop = TCPTV_CPU_VAR; tcp_finwait2_timeout = TCPTV_FINWAIT2_TIMEOUT; /* Setup the tcp function block list */ TAILQ_INIT(&t_functions); rw_init(&tcp_function_lock, "tcp_func_lock"); register_tcp_functions(&tcp_def_funcblk, M_WAITOK); #ifdef TCP_BLACKBOX /* Initialize the TCP logging data. */ tcp_log_init(); #endif arc4rand(&V_ts_offset_secret, sizeof(V_ts_offset_secret), 0); if (tcp_soreceive_stream) { #ifdef INET tcp_protosw.pr_soreceive = soreceive_stream; #endif #ifdef INET6 tcp6_protosw.pr_soreceive = soreceive_stream; #endif /* INET6 */ } #ifdef INET6 max_protohdr_grow(sizeof(struct ip6_hdr) + sizeof(struct tcphdr)); #else /* INET6 */ max_protohdr_grow(sizeof(struct tcpiphdr)); #endif /* INET6 */ ISN_LOCK_INIT(); EVENTHANDLER_REGISTER(shutdown_pre_sync, tcp_fini, NULL, SHUTDOWN_PRI_DEFAULT); EVENTHANDLER_REGISTER(vm_lowmem, tcp_drain, NULL, LOWMEM_PRI_DEFAULT); EVENTHANDLER_REGISTER(mbuf_lowmem, tcp_drain, NULL, LOWMEM_PRI_DEFAULT); tcp_inp_lro_direct_queue = counter_u64_alloc(M_WAITOK); tcp_inp_lro_wokeup_queue = counter_u64_alloc(M_WAITOK); tcp_inp_lro_compressed = counter_u64_alloc(M_WAITOK); tcp_inp_lro_locks_taken = counter_u64_alloc(M_WAITOK); tcp_extra_mbuf = counter_u64_alloc(M_WAITOK); tcp_would_have_but = counter_u64_alloc(M_WAITOK); tcp_comp_total = counter_u64_alloc(M_WAITOK); tcp_uncomp_total = counter_u64_alloc(M_WAITOK); tcp_bad_csums = counter_u64_alloc(M_WAITOK); #ifdef TCPPCAP tcp_pcap_init(); #endif hashsize = TCBHASHSIZE; tcbhash_tuneable = "net.inet.tcp.tcbhashsize"; TUNABLE_INT_FETCH(tcbhash_tuneable, &hashsize); if (hashsize == 0) { /* * Auto tune the hash size based on maxsockets. * A perfect hash would have a 1:1 mapping * (hashsize = maxsockets) however it's been * suggested that O(2) average is better. */ hashsize = maketcp_hashsize(maxsockets / 4); /* * Our historical default is 512, * do not autotune lower than this. */ if (hashsize < 512) hashsize = 512; if (bootverbose) printf("%s: %s auto tuned to %d\n", __func__, tcbhash_tuneable, hashsize); } /* * We require a hashsize to be a power of two. * Previously if it was not a power of two we would just reset it * back to 512, which could be a nasty surprise if you did not notice * the error message. * Instead what we do is clip it to the closest power of two lower * than the specified hash value. */ if (!powerof2(hashsize)) { int oldhashsize = hashsize; hashsize = maketcp_hashsize(hashsize); /* prevent absurdly low value */ if (hashsize < 16) hashsize = 16; printf("%s: WARNING: TCB hash size not a power of 2, " "clipped from %d to %d.\n", __func__, oldhashsize, hashsize); } tcp_tcbhashsize = hashsize; #ifdef INET IPPROTO_REGISTER(IPPROTO_TCP, tcp_input, tcp_ctlinput); #endif #ifdef INET6 IP6PROTO_REGISTER(IPPROTO_TCP, tcp6_input, tcp6_ctlinput); #endif } SYSINIT(tcp_init, SI_SUB_PROTO_DOMAIN, SI_ORDER_THIRD, tcp_init, NULL); #ifdef VIMAGE static void tcp_destroy(void *unused __unused) { int n; #ifdef TCP_HHOOK int error; #endif /* * All our processes are gone, all our sockets should be cleaned * up, which means, we should be past the tcp_discardcb() calls. * Sleep to let all tcpcb timers really disappear and cleanup. */ for (;;) { INP_INFO_WLOCK(&V_tcbinfo); n = V_tcbinfo.ipi_count; INP_INFO_WUNLOCK(&V_tcbinfo); if (n == 0) break; pause("tcpdes", hz / 10); } tcp_hc_destroy(); syncache_destroy(); in_pcbinfo_destroy(&V_tcbinfo); /* tcp_discardcb() clears the sack_holes up. */ uma_zdestroy(V_sack_hole_zone); uma_zdestroy(V_tcpcb_zone); /* * Cannot free the zone until all tcpcbs are released as we attach * the allocations to them. */ tcp_fastopen_destroy(); COUNTER_ARRAY_FREE(V_tcps_states, TCP_NSTATES); VNET_PCPUSTAT_FREE(tcpstat); #ifdef TCP_HHOOK error = hhook_head_deregister(V_tcp_hhh[HHOOK_TCP_EST_IN]); if (error != 0) { printf("%s: WARNING: unable to deregister helper hook " "type=%d, id=%d: error %d returned\n", __func__, HHOOK_TYPE_TCP, HHOOK_TCP_EST_IN, error); } error = hhook_head_deregister(V_tcp_hhh[HHOOK_TCP_EST_OUT]); if (error != 0) { printf("%s: WARNING: unable to deregister helper hook " "type=%d, id=%d: error %d returned\n", __func__, HHOOK_TYPE_TCP, HHOOK_TCP_EST_OUT, error); } #endif } VNET_SYSUNINIT(tcp, SI_SUB_PROTO_DOMAIN, SI_ORDER_FOURTH, tcp_destroy, NULL); #endif void tcp_fini(void *xtp) { } /* * Fill in the IP and TCP headers for an outgoing packet, given the tcpcb. * tcp_template used to store this data in mbufs, but we now recopy it out * of the tcpcb each time to conserve mbufs. */ void tcpip_fillheaders(struct inpcb *inp, uint16_t port, void *ip_ptr, void *tcp_ptr) { struct tcphdr *th = (struct tcphdr *)tcp_ptr; INP_WLOCK_ASSERT(inp); #ifdef INET6 if ((inp->inp_vflag & INP_IPV6) != 0) { struct ip6_hdr *ip6; ip6 = (struct ip6_hdr *)ip_ptr; ip6->ip6_flow = (ip6->ip6_flow & ~IPV6_FLOWINFO_MASK) | (inp->inp_flow & IPV6_FLOWINFO_MASK); ip6->ip6_vfc = (ip6->ip6_vfc & ~IPV6_VERSION_MASK) | (IPV6_VERSION & IPV6_VERSION_MASK); if (port == 0) ip6->ip6_nxt = IPPROTO_TCP; else ip6->ip6_nxt = IPPROTO_UDP; ip6->ip6_plen = htons(sizeof(struct tcphdr)); ip6->ip6_src = inp->in6p_laddr; ip6->ip6_dst = inp->in6p_faddr; } #endif /* INET6 */ #if defined(INET6) && defined(INET) else #endif #ifdef INET { struct ip *ip; ip = (struct ip *)ip_ptr; ip->ip_v = IPVERSION; ip->ip_hl = 5; ip->ip_tos = inp->inp_ip_tos; ip->ip_len = 0; ip->ip_id = 0; ip->ip_off = 0; ip->ip_ttl = inp->inp_ip_ttl; ip->ip_sum = 0; if (port == 0) ip->ip_p = IPPROTO_TCP; else ip->ip_p = IPPROTO_UDP; ip->ip_src = inp->inp_laddr; ip->ip_dst = inp->inp_faddr; } #endif /* INET */ th->th_sport = inp->inp_lport; th->th_dport = inp->inp_fport; th->th_seq = 0; th->th_ack = 0; th->th_off = 5; tcp_set_flags(th, 0); th->th_win = 0; th->th_urp = 0; th->th_sum = 0; /* in_pseudo() is called later for ipv4 */ } /* * Create template to be used to send tcp packets on a connection. * Allocates an mbuf and fills in a skeletal tcp/ip header. The only * use for this function is in keepalives, which use tcp_respond. */ struct tcptemp * tcpip_maketemplate(struct inpcb *inp) { struct tcptemp *t; t = malloc(sizeof(*t), M_TEMP, M_NOWAIT); if (t == NULL) return (NULL); tcpip_fillheaders(inp, 0, (void *)&t->tt_ipgen, (void *)&t->tt_t); return (t); } /* * Send a single message to the TCP at address specified by * the given TCP/IP header. If m == NULL, then we make a copy * of the tcpiphdr at th and send directly to the addressed host. * This is used to force keep alive messages out using the TCP * template for a connection. If flags are given then we send * a message back to the TCP which originated the segment th, * and discard the mbuf containing it and any other attached mbufs. * * In any case the ack and sequence number of the transmitted * segment are as specified by the parameters. * * NOTE: If m != NULL, then th must point to *inside* the mbuf. */ void tcp_respond(struct tcpcb *tp, void *ipgen, struct tcphdr *th, struct mbuf *m, tcp_seq ack, tcp_seq seq, int flags) { struct tcpopt to; struct inpcb *inp; struct ip *ip; struct mbuf *optm; struct udphdr *uh = NULL; struct tcphdr *nth; struct tcp_log_buffer *lgb; u_char *optp; #ifdef INET6 struct ip6_hdr *ip6; int isipv6; #endif /* INET6 */ int optlen, tlen, win, ulen; bool incl_opts; uint16_t port; int output_ret; #ifdef INVARIANTS int thflags = tcp_get_flags(th); #endif KASSERT(tp != NULL || m != NULL, ("tcp_respond: tp and m both NULL")); NET_EPOCH_ASSERT(); #ifdef INET6 isipv6 = ((struct ip *)ipgen)->ip_v == (IPV6_VERSION >> 4); ip6 = ipgen; #endif /* INET6 */ ip = ipgen; if (tp != NULL) { inp = tp->t_inpcb; KASSERT(inp != NULL, ("tcp control block w/o inpcb")); INP_LOCK_ASSERT(inp); } else inp = NULL; if (m != NULL) { #ifdef INET6 if (isipv6 && ip6 && (ip6->ip6_nxt == IPPROTO_UDP)) port = m->m_pkthdr.tcp_tun_port; else #endif if (ip && (ip->ip_p == IPPROTO_UDP)) port = m->m_pkthdr.tcp_tun_port; else port = 0; } else port = tp->t_port; incl_opts = false; win = 0; if (tp != NULL) { if (!(flags & TH_RST)) { win = sbspace(&inp->inp_socket->so_rcv); if (win > TCP_MAXWIN << tp->rcv_scale) win = TCP_MAXWIN << tp->rcv_scale; } if ((tp->t_flags & TF_NOOPT) == 0) incl_opts = true; } if (m == NULL) { m = m_gethdr(M_NOWAIT, MT_DATA); if (m == NULL) return; m->m_data += max_linkhdr; #ifdef INET6 if (isipv6) { bcopy((caddr_t)ip6, mtod(m, caddr_t), sizeof(struct ip6_hdr)); ip6 = mtod(m, struct ip6_hdr *); nth = (struct tcphdr *)(ip6 + 1); if (port) { /* Insert a UDP header */ uh = (struct udphdr *)nth; uh->uh_sport = htons(V_tcp_udp_tunneling_port); uh->uh_dport = port; nth = (struct tcphdr *)(uh + 1); } } else #endif /* INET6 */ { bcopy((caddr_t)ip, mtod(m, caddr_t), sizeof(struct ip)); ip = mtod(m, struct ip *); nth = (struct tcphdr *)(ip + 1); if (port) { /* Insert a UDP header */ uh = (struct udphdr *)nth; uh->uh_sport = htons(V_tcp_udp_tunneling_port); uh->uh_dport = port; nth = (struct tcphdr *)(uh + 1); } } bcopy((caddr_t)th, (caddr_t)nth, sizeof(struct tcphdr)); flags = TH_ACK; } else if ((!M_WRITABLE(m)) || (port != 0)) { struct mbuf *n; /* Can't reuse 'm', allocate a new mbuf. */ n = m_gethdr(M_NOWAIT, MT_DATA); if (n == NULL) { m_freem(m); return; } if (!m_dup_pkthdr(n, m, M_NOWAIT)) { m_freem(m); m_freem(n); return; } n->m_data += max_linkhdr; /* m_len is set later */ #define xchg(a,b,type) { type t; t=a; a=b; b=t; } #ifdef INET6 if (isipv6) { bcopy((caddr_t)ip6, mtod(n, caddr_t), sizeof(struct ip6_hdr)); ip6 = mtod(n, struct ip6_hdr *); xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr); nth = (struct tcphdr *)(ip6 + 1); if (port) { /* Insert a UDP header */ uh = (struct udphdr *)nth; uh->uh_sport = htons(V_tcp_udp_tunneling_port); uh->uh_dport = port; nth = (struct tcphdr *)(uh + 1); } } else #endif /* INET6 */ { bcopy((caddr_t)ip, mtod(n, caddr_t), sizeof(struct ip)); ip = mtod(n, struct ip *); xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, uint32_t); nth = (struct tcphdr *)(ip + 1); if (port) { /* Insert a UDP header */ uh = (struct udphdr *)nth; uh->uh_sport = htons(V_tcp_udp_tunneling_port); uh->uh_dport = port; nth = (struct tcphdr *)(uh + 1); } } bcopy((caddr_t)th, (caddr_t)nth, sizeof(struct tcphdr)); xchg(nth->th_dport, nth->th_sport, uint16_t); th = nth; m_freem(m); m = n; } else { /* * reuse the mbuf. * XXX MRT We inherit the FIB, which is lucky. */ m_freem(m->m_next); m->m_next = NULL; m->m_data = (caddr_t)ipgen; /* m_len is set later */ #ifdef INET6 if (isipv6) { xchg(ip6->ip6_dst, ip6->ip6_src, struct in6_addr); nth = (struct tcphdr *)(ip6 + 1); } else #endif /* INET6 */ { xchg(ip->ip_dst.s_addr, ip->ip_src.s_addr, uint32_t); nth = (struct tcphdr *)(ip + 1); } if (th != nth) { /* * this is usually a case when an extension header * exists between the IPv6 header and the * TCP header. */ nth->th_sport = th->th_sport; nth->th_dport = th->th_dport; } xchg(nth->th_dport, nth->th_sport, uint16_t); #undef xchg } tlen = 0; #ifdef INET6 if (isipv6) tlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr); #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET tlen = sizeof (struct tcpiphdr); #endif if (port) tlen += sizeof (struct udphdr); #ifdef INVARIANTS m->m_len = 0; KASSERT(M_TRAILINGSPACE(m) >= tlen, ("Not enough trailing space for message (m=%p, need=%d, have=%ld)", m, tlen, (long)M_TRAILINGSPACE(m))); #endif m->m_len = tlen; to.to_flags = 0; if (incl_opts) { /* Make sure we have room. */ if (M_TRAILINGSPACE(m) < TCP_MAXOLEN) { m->m_next = m_get(M_NOWAIT, MT_DATA); if (m->m_next) { optp = mtod(m->m_next, u_char *); optm = m->m_next; } else incl_opts = false; } else { optp = (u_char *) (nth + 1); optm = m; } } if (incl_opts) { /* Timestamps. */ if (tp->t_flags & TF_RCVD_TSTMP) { to.to_tsval = tcp_ts_getticks() + tp->ts_offset; to.to_tsecr = tp->ts_recent; to.to_flags |= TOF_TS; } #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) /* TCP-MD5 (RFC2385). */ if (tp->t_flags & TF_SIGNATURE) to.to_flags |= TOF_SIGNATURE; #endif /* Add the options. */ tlen += optlen = tcp_addoptions(&to, optp); /* Update m_len in the correct mbuf. */ optm->m_len += optlen; } else optlen = 0; #ifdef INET6 if (isipv6) { if (uh) { ulen = tlen - sizeof(struct ip6_hdr); uh->uh_ulen = htons(ulen); } ip6->ip6_flow = 0; ip6->ip6_vfc = IPV6_VERSION; if (port) ip6->ip6_nxt = IPPROTO_UDP; else ip6->ip6_nxt = IPPROTO_TCP; ip6->ip6_plen = htons(tlen - sizeof(*ip6)); } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { if (uh) { ulen = tlen - sizeof(struct ip); uh->uh_ulen = htons(ulen); } ip->ip_len = htons(tlen); ip->ip_ttl = V_ip_defttl; if (port) { ip->ip_p = IPPROTO_UDP; } else { ip->ip_p = IPPROTO_TCP; } if (V_path_mtu_discovery) ip->ip_off |= htons(IP_DF); } #endif m->m_pkthdr.len = tlen; m->m_pkthdr.rcvif = NULL; #ifdef MAC if (inp != NULL) { /* * Packet is associated with a socket, so allow the * label of the response to reflect the socket label. */ INP_LOCK_ASSERT(inp); mac_inpcb_create_mbuf(inp, m); } else { /* * Packet is not associated with a socket, so possibly * update the label in place. */ mac_netinet_tcp_reply(m); } #endif nth->th_seq = htonl(seq); nth->th_ack = htonl(ack); nth->th_off = (sizeof (struct tcphdr) + optlen) >> 2; tcp_set_flags(nth, flags); if (tp != NULL) nth->th_win = htons((u_short) (win >> tp->rcv_scale)); else nth->th_win = htons((u_short)win); nth->th_urp = 0; #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) if (to.to_flags & TOF_SIGNATURE) { if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, nth, to.to_signature) != 0) { m_freem(m); return; } } #endif #ifdef INET6 if (isipv6) { if (port) { m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); uh->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0); nth->th_sum = 0; } else { m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); nth->th_sum = in6_cksum_pseudo(ip6, tlen - sizeof(struct ip6_hdr), IPPROTO_TCP, 0); } ip6->ip6_hlim = in6_selecthlim(tp != NULL ? tp->t_inpcb : NULL, NULL); } #endif /* INET6 */ #if defined(INET6) && defined(INET) else #endif #ifdef INET { if (port) { uh->uh_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP)); m->m_pkthdr.csum_flags = CSUM_UDP; m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); nth->th_sum = 0; } else { m->m_pkthdr.csum_flags = CSUM_TCP; m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); nth->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, htons((u_short)(tlen - sizeof(struct ip) + ip->ip_p))); } } #endif /* INET */ #ifdef TCPDEBUG if (tp == NULL || (inp->inp_socket->so_options & SO_DEBUG)) tcp_trace(TA_OUTPUT, 0, tp, mtod(m, void *), th, 0); #endif TCP_PROBE3(debug__output, tp, th, m); if (flags & TH_RST) TCP_PROBE5(accept__refused, NULL, NULL, m, tp, nth); lgb = NULL; if ((tp != NULL) && (tp->t_logstate != TCP_LOG_STATE_OFF)) { if (INP_WLOCKED(inp)) { union tcp_log_stackspecific log; struct timeval tv; memset(&log.u_bbr, 0, sizeof(log.u_bbr)); log.u_bbr.inhpts = tp->t_inpcb->inp_in_hpts; log.u_bbr.flex8 = 4; log.u_bbr.pkts_out = tp->t_maxseg; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.delivered = 0; lgb = tcp_log_event_(tp, nth, NULL, NULL, TCP_LOG_OUT, ERRNO_UNK, 0, &log, false, NULL, NULL, 0, &tv); } else { /* * We can not log the packet, since we only own the * read lock, but a write lock is needed. The read lock * is not upgraded to a write lock, since only getting * the read lock was done intentionally to improve the * handling of SYN flooding attacks. * This happens only for pure SYN segments received in * the initial CLOSED state, or received in a more * advanced state than listen and the UDP encapsulation * port is unexpected. * The incoming SYN segments do not really belong to * the TCP connection and the handling does not change * the state of the TCP connection. Therefore, the * sending of the RST segments is not logged. Please * note that also the incoming SYN segments are not * logged. * * The following code ensures that the above description * is and stays correct. */ KASSERT((thflags & (TH_ACK|TH_SYN)) == TH_SYN && (tp->t_state == TCPS_CLOSED || (tp->t_state > TCPS_LISTEN && tp->t_port != port)), ("%s: Logging of TCP segment with flags 0x%b and " "UDP encapsulation port %u skipped in state %s", __func__, thflags, PRINT_TH_FLAGS, ntohs(port), tcpstates[tp->t_state])); } } if (flags & TH_ACK) TCPSTAT_INC(tcps_sndacks); else if (flags & (TH_SYN|TH_FIN|TH_RST)) TCPSTAT_INC(tcps_sndctrl); TCPSTAT_INC(tcps_sndtotal); #ifdef INET6 if (isipv6) { TCP_PROBE5(send, NULL, tp, ip6, tp, nth); output_ret = ip6_output(m, NULL, NULL, 0, NULL, NULL, inp); } #endif /* INET6 */ #if defined(INET) && defined(INET6) else #endif #ifdef INET { TCP_PROBE5(send, NULL, tp, ip, tp, nth); output_ret = ip_output(m, NULL, NULL, 0, NULL, inp); } #endif if (lgb != NULL) lgb->tlb_errno = output_ret; } /* * Create a new TCP control block, making an * empty reassembly queue and hooking it to the argument * protocol control block. The `inp' parameter must have * come from the zone allocator set up in tcp_init(). */ struct tcpcb * tcp_newtcpcb(struct inpcb *inp) { struct tcpcb_mem *tm; struct tcpcb *tp; #ifdef INET6 int isipv6 = (inp->inp_vflag & INP_IPV6) != 0; #endif /* INET6 */ tm = uma_zalloc(V_tcpcb_zone, M_NOWAIT | M_ZERO); if (tm == NULL) return (NULL); tp = &tm->tcb; /* Initialise cc_var struct for this tcpcb. */ tp->ccv = &tm->ccv; tp->ccv->type = IPPROTO_TCP; tp->ccv->ccvc.tcp = tp; rw_rlock(&tcp_function_lock); tp->t_fb = tcp_func_set_ptr; refcount_acquire(&tp->t_fb->tfb_refcnt); rw_runlock(&tcp_function_lock); /* * Use the current system default CC algorithm. */ cc_attach(tp, CC_DEFAULT_ALGO()); /* * The tcpcb will hold a reference on its inpcb until tcp_discardcb() * is called. */ in_pcbref(inp); /* Reference for tcpcb */ tp->t_inpcb = inp; if (CC_ALGO(tp)->cb_init != NULL) if (CC_ALGO(tp)->cb_init(tp->ccv, NULL) > 0) { cc_detach(tp); if (tp->t_fb->tfb_tcp_fb_fini) (*tp->t_fb->tfb_tcp_fb_fini)(tp, 1); in_pcbrele_wlocked(inp); refcount_release(&tp->t_fb->tfb_refcnt); uma_zfree(V_tcpcb_zone, tm); return (NULL); } #ifdef TCP_HHOOK tp->osd = &tm->osd; if (khelp_init_osd(HELPER_CLASS_TCP, tp->osd)) { if (tp->t_fb->tfb_tcp_fb_fini) (*tp->t_fb->tfb_tcp_fb_fini)(tp, 1); in_pcbrele_wlocked(inp); refcount_release(&tp->t_fb->tfb_refcnt); uma_zfree(V_tcpcb_zone, tm); return (NULL); } #endif #ifdef VIMAGE tp->t_vnet = inp->inp_vnet; #endif tp->t_timers = &tm->tt; TAILQ_INIT(&tp->t_segq); tp->t_maxseg = #ifdef INET6 isipv6 ? V_tcp_v6mssdflt : #endif /* INET6 */ V_tcp_mssdflt; /* Set up our timeouts. */ callout_init(&tp->t_timers->tt_rexmt, 1); callout_init(&tp->t_timers->tt_persist, 1); callout_init(&tp->t_timers->tt_keep, 1); callout_init(&tp->t_timers->tt_2msl, 1); callout_init(&tp->t_timers->tt_delack, 1); switch (V_tcp_do_rfc1323) { case 0: break; default: case 1: tp->t_flags = (TF_REQ_SCALE|TF_REQ_TSTMP); break; case 2: tp->t_flags = TF_REQ_SCALE; break; case 3: tp->t_flags = TF_REQ_TSTMP; break; } if (V_tcp_do_sack) tp->t_flags |= TF_SACK_PERMIT; TAILQ_INIT(&tp->snd_holes); /* * Init srtt to TCPTV_SRTTBASE (0), so we can tell that we have no * rtt estimate. Set rttvar so that srtt + 4 * rttvar gives * reasonable initial retransmit time. */ tp->t_srtt = TCPTV_SRTTBASE; tp->t_rttvar = ((tcp_rexmit_initial - TCPTV_SRTTBASE) << TCP_RTTVAR_SHIFT) / 4; tp->t_rttmin = tcp_rexmit_min; tp->t_rxtcur = tcp_rexmit_initial; tp->snd_cwnd = TCP_MAXWIN << TCP_MAX_WINSHIFT; tp->snd_ssthresh = TCP_MAXWIN << TCP_MAX_WINSHIFT; tp->t_rcvtime = ticks; /* * IPv4 TTL initialization is necessary for an IPv6 socket as well, * because the socket may be bound to an IPv6 wildcard address, * which may match an IPv4-mapped IPv6 address. */ inp->inp_ip_ttl = V_ip_defttl; inp->inp_ppcb = tp; #ifdef TCPPCAP /* * Init the TCP PCAP queues. */ tcp_pcap_tcpcb_init(tp); #endif #ifdef TCP_BLACKBOX /* Initialize the per-TCPCB log data. */ tcp_log_tcpcbinit(tp); #endif tp->t_pacing_rate = -1; if (tp->t_fb->tfb_tcp_fb_init) { if ((*tp->t_fb->tfb_tcp_fb_init)(tp)) { refcount_release(&tp->t_fb->tfb_refcnt); in_pcbrele_wlocked(inp); uma_zfree(V_tcpcb_zone, tm); return (NULL); } } #ifdef STATS if (V_tcp_perconn_stats_enable == 1) tp->t_stats = stats_blob_alloc(V_tcp_perconn_stats_dflt_tpl, 0); #endif if (V_tcp_do_lrd) tp->t_flags |= TF_LRD; return (tp); /* XXX */ } /* * Drop a TCP connection, reporting * the specified error. If connection is synchronized, * then send a RST to peer. */ struct tcpcb * tcp_drop(struct tcpcb *tp, int errno) { struct socket *so = tp->t_inpcb->inp_socket; NET_EPOCH_ASSERT(); INP_WLOCK_ASSERT(tp->t_inpcb); if (TCPS_HAVERCVDSYN(tp->t_state)) { tcp_state_change(tp, TCPS_CLOSED); /* Don't use tcp_output() here due to possible recursion. */ (void)tcp_output_nodrop(tp); TCPSTAT_INC(tcps_drops); } else TCPSTAT_INC(tcps_conndrops); if (errno == ETIMEDOUT && tp->t_softerror) errno = tp->t_softerror; so->so_error = errno; return (tcp_close(tp)); } void tcp_discardcb(struct tcpcb *tp) { struct inpcb *inp = tp->t_inpcb; INP_WLOCK_ASSERT(inp); /* * Make sure that all of our timers are stopped before we delete the * PCB. * * If stopping a timer fails, we schedule a discard function in same * callout, and the last discard function called will take care of * deleting the tcpcb. */ tp->t_timers->tt_draincnt = 0; tcp_timer_stop(tp, TT_REXMT); tcp_timer_stop(tp, TT_PERSIST); tcp_timer_stop(tp, TT_KEEP); tcp_timer_stop(tp, TT_2MSL); tcp_timer_stop(tp, TT_DELACK); if (tp->t_fb->tfb_tcp_timer_stop_all) { /* * Call the stop-all function of the methods, * this function should call the tcp_timer_stop() * method with each of the function specific timeouts. * That stop will be called via the tfb_tcp_timer_stop() * which should use the async drain function of the * callout system (see tcp_var.h). */ tp->t_fb->tfb_tcp_timer_stop_all(tp); } /* free the reassembly queue, if any */ tcp_reass_flush(tp); #ifdef TCP_OFFLOAD /* Disconnect offload device, if any. */ if (tp->t_flags & TF_TOE) tcp_offload_detach(tp); #endif tcp_free_sackholes(tp); #ifdef TCPPCAP /* Free the TCP PCAP queues. */ tcp_pcap_drain(&(tp->t_inpkts)); tcp_pcap_drain(&(tp->t_outpkts)); #endif /* Allow the CC algorithm to clean up after itself. */ if (CC_ALGO(tp)->cb_destroy != NULL) CC_ALGO(tp)->cb_destroy(tp->ccv); CC_DATA(tp) = NULL; /* Detach from the CC algorithm */ cc_detach(tp); #ifdef TCP_HHOOK khelp_destroy_osd(tp->osd); #endif #ifdef STATS stats_blob_destroy(tp->t_stats); #endif CC_ALGO(tp) = NULL; inp->inp_ppcb = NULL; if (tp->t_timers->tt_draincnt == 0) { bool released __diagused; released = tcp_freecb(tp); KASSERT(!released, ("%s: inp %p should not have been released " "here", __func__, inp)); } } bool tcp_freecb(struct tcpcb *tp) { struct inpcb *inp = tp->t_inpcb; struct socket *so = inp->inp_socket; #ifdef INET6 bool isipv6 = (inp->inp_vflag & INP_IPV6) != 0; #endif INP_WLOCK_ASSERT(inp); MPASS(tp->t_timers->tt_draincnt == 0); /* We own the last reference on tcpcb, let's free it. */ #ifdef TCP_BLACKBOX tcp_log_tcpcbfini(tp); #endif TCPSTATES_DEC(tp->t_state); if (tp->t_fb->tfb_tcp_fb_fini) (*tp->t_fb->tfb_tcp_fb_fini)(tp, 1); /* * If we got enough samples through the srtt filter, * save the rtt and rttvar in the routing entry. * 'Enough' is arbitrarily defined as 4 rtt samples. * 4 samples is enough for the srtt filter to converge * to within enough % of the correct value; fewer samples * and we could save a bogus rtt. The danger is not high * as tcp quickly recovers from everything. * XXX: Works very well but needs some more statistics! * * XXXRRS: Updating must be after the stack fini() since * that may be converting some internal representation of * say srtt etc into the general one used by other stacks. * Lets also at least protect against the so being NULL * as RW stated below. */ if ((tp->t_rttupdated >= 4) && (so != NULL)) { struct hc_metrics_lite metrics; uint32_t ssthresh; bzero(&metrics, sizeof(metrics)); /* * Update the ssthresh always when the conditions below * are satisfied. This gives us better new start value * for the congestion avoidance for new connections. * ssthresh is only set if packet loss occurred on a session. * * XXXRW: 'so' may be NULL here, and/or socket buffer may be * being torn down. Ideally this code would not use 'so'. */ ssthresh = tp->snd_ssthresh; if (ssthresh != 0 && ssthresh < so->so_snd.sb_hiwat / 2) { /* * convert the limit from user data bytes to * packets then to packet data bytes. */ ssthresh = (ssthresh + tp->t_maxseg / 2) / tp->t_maxseg; if (ssthresh < 2) ssthresh = 2; ssthresh *= (tp->t_maxseg + #ifdef INET6 (isipv6 ? sizeof (struct ip6_hdr) + sizeof (struct tcphdr) : #endif sizeof (struct tcpiphdr) #ifdef INET6 ) #endif ); } else ssthresh = 0; metrics.rmx_ssthresh = ssthresh; metrics.rmx_rtt = tp->t_srtt; metrics.rmx_rttvar = tp->t_rttvar; metrics.rmx_cwnd = tp->snd_cwnd; metrics.rmx_sendpipe = 0; metrics.rmx_recvpipe = 0; tcp_hc_update(&inp->inp_inc, &metrics); } refcount_release(&tp->t_fb->tfb_refcnt); uma_zfree(V_tcpcb_zone, tp); return (in_pcbrele_wlocked(inp)); } /* * Attempt to close a TCP control block, marking it as dropped, and freeing * the socket if we hold the only reference. */ struct tcpcb * tcp_close(struct tcpcb *tp) { struct inpcb *inp = tp->t_inpcb; struct socket *so; INP_WLOCK_ASSERT(inp); #ifdef TCP_OFFLOAD if (tp->t_state == TCPS_LISTEN) tcp_offload_listen_stop(tp); #endif /* * This releases the TFO pending counter resource for TFO listen * sockets as well as passively-created TFO sockets that transition * from SYN_RECEIVED to CLOSED. */ if (tp->t_tfo_pending) { tcp_fastopen_decrement_counter(tp->t_tfo_pending); tp->t_tfo_pending = NULL; } #ifdef TCPHPTS tcp_hpts_remove(inp); #endif in_pcbdrop(inp); TCPSTAT_INC(tcps_closed); if (tp->t_state != TCPS_CLOSED) tcp_state_change(tp, TCPS_CLOSED); KASSERT(inp->inp_socket != NULL, ("tcp_close: inp_socket NULL")); so = inp->inp_socket; soisdisconnected(so); if (inp->inp_flags & INP_SOCKREF) { inp->inp_flags &= ~INP_SOCKREF; INP_WUNLOCK(inp); sorele(so); return (NULL); } return (tp); } /* * Notify a tcp user of an asynchronous error; * store error as soft error, but wake up user * (for now, won't do anything until can select for soft error). * * Do not wake up user since there currently is no mechanism for * reporting soft errors (yet - a kqueue filter may be added). */ static struct inpcb * tcp_notify(struct inpcb *inp, int error) { struct tcpcb *tp; INP_WLOCK_ASSERT(inp); - if ((inp->inp_flags & INP_TIMEWAIT) || - (inp->inp_flags & INP_DROPPED)) + if (inp->inp_flags & INP_DROPPED) return (inp); tp = intotcpcb(inp); KASSERT(tp != NULL, ("tcp_notify: tp == NULL")); /* * Ignore some errors if we are hooked up. * If connection hasn't completed, has retransmitted several times, * and receives a second error, give up now. This is better * than waiting a long time to establish a connection that * can never complete. */ if (tp->t_state == TCPS_ESTABLISHED && (error == EHOSTUNREACH || error == ENETUNREACH || error == EHOSTDOWN)) { if (inp->inp_route.ro_nh) { NH_FREE(inp->inp_route.ro_nh); inp->inp_route.ro_nh = (struct nhop_object *)NULL; } return (inp); } else if (tp->t_state < TCPS_ESTABLISHED && tp->t_rxtshift > 3 && tp->t_softerror) { tp = tcp_drop(tp, error); if (tp != NULL) return (inp); else return (NULL); } else { tp->t_softerror = error; return (inp); } #if 0 wakeup( &so->so_timeo); sorwakeup(so); sowwakeup(so); #endif } static int tcp_pcblist(SYSCTL_HANDLER_ARGS) { struct inpcb_iterator inpi = INP_ALL_ITERATOR(&V_tcbinfo, INPLOOKUP_RLOCKPCB); struct xinpgen xig; struct inpcb *inp; int error; if (req->newptr != NULL) return (EPERM); if (req->oldptr == NULL) { int n; n = V_tcbinfo.ipi_count + counter_u64_fetch(V_tcps_states[TCPS_SYN_RECEIVED]); n += imax(n / 8, 10); req->oldidx = 2 * (sizeof xig) + n * sizeof(struct xtcpcb); return (0); } if ((error = sysctl_wire_old_buffer(req, 0)) != 0) return (error); bzero(&xig, sizeof(xig)); xig.xig_len = sizeof xig; xig.xig_count = V_tcbinfo.ipi_count + counter_u64_fetch(V_tcps_states[TCPS_SYN_RECEIVED]); xig.xig_gen = V_tcbinfo.ipi_gencnt; xig.xig_sogen = so_gencnt; error = SYSCTL_OUT(req, &xig, sizeof xig); if (error) return (error); error = syncache_pcblist(req); if (error) return (error); while ((inp = inp_next(&inpi)) != NULL) { if (inp->inp_gencnt <= xig.xig_gen && cr_canseeinpcb(req->td->td_ucred, inp) == 0) { struct xtcpcb xt; tcp_inptoxtp(inp, &xt); error = SYSCTL_OUT(req, &xt, sizeof xt); if (error) { INP_RUNLOCK(inp); break; } else continue; } } if (!error) { /* * Give the user an updated idea of our state. * If the generation differs from what we told * her before, she knows that something happened * while we were processing this request, and it * might be necessary to retry. */ xig.xig_gen = V_tcbinfo.ipi_gencnt; xig.xig_sogen = so_gencnt; xig.xig_count = V_tcbinfo.ipi_count + counter_u64_fetch(V_tcps_states[TCPS_SYN_RECEIVED]); error = SYSCTL_OUT(req, &xig, sizeof xig); } return (error); } SYSCTL_PROC(_net_inet_tcp, TCPCTL_PCBLIST, pcblist, CTLTYPE_OPAQUE | CTLFLAG_RD | CTLFLAG_NEEDGIANT, NULL, 0, tcp_pcblist, "S,xtcpcb", "List of active TCP connections"); #ifdef INET static int tcp_getcred(SYSCTL_HANDLER_ARGS) { struct xucred xuc; struct sockaddr_in addrs[2]; struct epoch_tracker et; struct inpcb *inp; int error; error = priv_check(req->td, PRIV_NETINET_GETCRED); if (error) return (error); error = SYSCTL_IN(req, addrs, sizeof(addrs)); if (error) return (error); NET_EPOCH_ENTER(et); inp = in_pcblookup(&V_tcbinfo, addrs[1].sin_addr, addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, INPLOOKUP_RLOCKPCB, NULL); NET_EPOCH_EXIT(et); if (inp != NULL) { if (inp->inp_socket == NULL) error = ENOENT; if (error == 0) error = cr_canseeinpcb(req->td->td_ucred, inp); if (error == 0) cru2x(inp->inp_cred, &xuc); INP_RUNLOCK(inp); } else error = ENOENT; if (error == 0) error = SYSCTL_OUT(req, &xuc, sizeof(struct xucred)); return (error); } SYSCTL_PROC(_net_inet_tcp, OID_AUTO, getcred, CTLTYPE_OPAQUE | CTLFLAG_RW | CTLFLAG_PRISON | CTLFLAG_NEEDGIANT, 0, 0, tcp_getcred, "S,xucred", "Get the xucred of a TCP connection"); #endif /* INET */ #ifdef INET6 static int tcp6_getcred(SYSCTL_HANDLER_ARGS) { struct epoch_tracker et; struct xucred xuc; struct sockaddr_in6 addrs[2]; struct inpcb *inp; int error; #ifdef INET int mapped = 0; #endif error = priv_check(req->td, PRIV_NETINET_GETCRED); if (error) return (error); error = SYSCTL_IN(req, addrs, sizeof(addrs)); if (error) return (error); if ((error = sa6_embedscope(&addrs[0], V_ip6_use_defzone)) != 0 || (error = sa6_embedscope(&addrs[1], V_ip6_use_defzone)) != 0) { return (error); } if (IN6_IS_ADDR_V4MAPPED(&addrs[0].sin6_addr)) { #ifdef INET if (IN6_IS_ADDR_V4MAPPED(&addrs[1].sin6_addr)) mapped = 1; else #endif return (EINVAL); } NET_EPOCH_ENTER(et); #ifdef INET if (mapped == 1) inp = in_pcblookup(&V_tcbinfo, *(struct in_addr *)&addrs[1].sin6_addr.s6_addr[12], addrs[1].sin6_port, *(struct in_addr *)&addrs[0].sin6_addr.s6_addr[12], addrs[0].sin6_port, INPLOOKUP_RLOCKPCB, NULL); else #endif inp = in6_pcblookup(&V_tcbinfo, &addrs[1].sin6_addr, addrs[1].sin6_port, &addrs[0].sin6_addr, addrs[0].sin6_port, INPLOOKUP_RLOCKPCB, NULL); NET_EPOCH_EXIT(et); if (inp != NULL) { if (inp->inp_socket == NULL) error = ENOENT; if (error == 0) error = cr_canseeinpcb(req->td->td_ucred, inp); if (error == 0) cru2x(inp->inp_cred, &xuc); INP_RUNLOCK(inp); } else error = ENOENT; if (error == 0) error = SYSCTL_OUT(req, &xuc, sizeof(struct xucred)); return (error); } SYSCTL_PROC(_net_inet6_tcp6, OID_AUTO, getcred, CTLTYPE_OPAQUE | CTLFLAG_RW | CTLFLAG_PRISON | CTLFLAG_NEEDGIANT, 0, 0, tcp6_getcred, "S,xucred", "Get the xucred of a TCP6 connection"); #endif /* INET6 */ #ifdef INET /* Path MTU to try next when a fragmentation-needed message is received. */ static inline int tcp_next_pmtu(const struct icmp *icp, const struct ip *ip) { int mtu = ntohs(icp->icmp_nextmtu); /* If no alternative MTU was proposed, try the next smaller one. */ if (!mtu) mtu = ip_next_mtu(ntohs(ip->ip_len), 1); if (mtu < V_tcp_minmss + sizeof(struct tcpiphdr)) mtu = V_tcp_minmss + sizeof(struct tcpiphdr); return (mtu); } static void tcp_ctlinput_with_port(struct icmp *icp, uint16_t port) { struct ip *ip; struct tcphdr *th; struct inpcb *inp; struct tcpcb *tp; struct inpcb *(*notify)(struct inpcb *, int); struct in_conninfo inc; tcp_seq icmp_tcp_seq; int errno, mtu; errno = icmp_errmap(icp); switch (errno) { case 0: return; case EMSGSIZE: notify = tcp_mtudisc_notify; break; case ECONNREFUSED: if (V_icmp_may_rst) notify = tcp_drop_syn_sent; else notify = tcp_notify; break; case EHOSTUNREACH: if (V_icmp_may_rst && icp->icmp_type == ICMP_TIMXCEED) notify = tcp_drop_syn_sent; else notify = tcp_notify; break; default: notify = tcp_notify; } ip = &icp->icmp_ip; th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2)); icmp_tcp_seq = th->th_seq; inp = in_pcblookup(&V_tcbinfo, ip->ip_dst, th->th_dport, ip->ip_src, th->th_sport, INPLOOKUP_WLOCKPCB, NULL); if (inp != NULL) { - if (!(inp->inp_flags & INP_TIMEWAIT) && - !(inp->inp_flags & INP_DROPPED) && + if (!(inp->inp_flags & INP_DROPPED) && !(inp->inp_socket == NULL)) { tp = intotcpcb(inp); #ifdef TCP_OFFLOAD if (tp->t_flags & TF_TOE && errno == EMSGSIZE) { /* * MTU discovery for offloaded connections. Let * the TOE driver verify seq# and process it. */ mtu = tcp_next_pmtu(icp, ip); tcp_offload_pmtu_update(tp, icmp_tcp_seq, mtu); goto out; } #endif if (tp->t_port != port) { goto out; } if (SEQ_GEQ(ntohl(icmp_tcp_seq), tp->snd_una) && SEQ_LT(ntohl(icmp_tcp_seq), tp->snd_max)) { if (errno == EMSGSIZE) { /* * MTU discovery: we got a needfrag and * will potentially try a lower MTU. */ mtu = tcp_next_pmtu(icp, ip); /* * Only process the offered MTU if it * is smaller than the current one. */ if (mtu < tp->t_maxseg + sizeof(struct tcpiphdr)) { bzero(&inc, sizeof(inc)); inc.inc_faddr = ip->ip_dst; inc.inc_fibnum = inp->inp_inc.inc_fibnum; tcp_hc_updatemtu(&inc, mtu); inp = tcp_mtudisc(inp, mtu); } } else inp = (*notify)(inp, errno); } } } else { bzero(&inc, sizeof(inc)); inc.inc_fport = th->th_dport; inc.inc_lport = th->th_sport; inc.inc_faddr = ip->ip_dst; inc.inc_laddr = ip->ip_src; syncache_unreach(&inc, icmp_tcp_seq, port); } out: if (inp != NULL) INP_WUNLOCK(inp); } static void tcp_ctlinput(struct icmp *icmp) { tcp_ctlinput_with_port(icmp, htons(0)); } static void tcp_ctlinput_viaudp(udp_tun_icmp_param_t param) { /* Its a tunneled TCP over UDP icmp */ struct icmp *icmp = param.icmp; struct ip *outer_ip, *inner_ip; struct udphdr *udp; struct tcphdr *th, ttemp; int i_hlen, o_len; uint16_t port; outer_ip = (struct ip *)((caddr_t)icmp - sizeof(struct ip)); inner_ip = &icmp->icmp_ip; i_hlen = inner_ip->ip_hl << 2; o_len = ntohs(outer_ip->ip_len); if (o_len < (sizeof(struct ip) + 8 + i_hlen + sizeof(struct udphdr) + offsetof(struct tcphdr, th_ack))) { /* Not enough data present */ return; } /* Ok lets strip out the inner udphdr header by copying up on top of it the tcp hdr */ udp = (struct udphdr *)(((caddr_t)inner_ip) + i_hlen); if (ntohs(udp->uh_sport) != V_tcp_udp_tunneling_port) { return; } port = udp->uh_dport; th = (struct tcphdr *)(udp + 1); memcpy(&ttemp, th, sizeof(struct tcphdr)); memcpy(udp, &ttemp, sizeof(struct tcphdr)); /* Now adjust down the size of the outer IP header */ o_len -= sizeof(struct udphdr); outer_ip->ip_len = htons(o_len); /* Now call in to the normal handling code */ tcp_ctlinput_with_port(icmp, port); } #endif /* INET */ #ifdef INET6 static inline int tcp6_next_pmtu(const struct icmp6_hdr *icmp6) { int mtu = ntohl(icmp6->icmp6_mtu); /* * If no alternative MTU was proposed, or the proposed MTU was too * small, set to the min. */ if (mtu < IPV6_MMTU) mtu = IPV6_MMTU - 8; /* XXXNP: what is the adjustment for? */ return (mtu); } static void tcp6_ctlinput_with_port(struct ip6ctlparam *ip6cp, uint16_t port) { struct in6_addr *dst; struct inpcb *(*notify)(struct inpcb *, int); struct ip6_hdr *ip6; struct mbuf *m; struct inpcb *inp; struct tcpcb *tp; struct icmp6_hdr *icmp6; struct in_conninfo inc; struct tcp_ports { uint16_t th_sport; uint16_t th_dport; } t_ports; tcp_seq icmp_tcp_seq; unsigned int mtu; unsigned int off; int errno; icmp6 = ip6cp->ip6c_icmp6; m = ip6cp->ip6c_m; ip6 = ip6cp->ip6c_ip6; off = ip6cp->ip6c_off; dst = &ip6cp->ip6c_finaldst->sin6_addr; errno = icmp6_errmap(icmp6); switch (errno) { case 0: return; case EMSGSIZE: notify = tcp_mtudisc_notify; break; case ECONNREFUSED: if (V_icmp_may_rst) notify = tcp_drop_syn_sent; else notify = tcp_notify; break; case EHOSTUNREACH: /* * There are only four ICMPs that may reset connection: * - administratively prohibited * - port unreachable * - time exceeded in transit * - unknown next header */ if (V_icmp_may_rst && ((icmp6->icmp6_type == ICMP6_DST_UNREACH && (icmp6->icmp6_code == ICMP6_DST_UNREACH_ADMIN || icmp6->icmp6_code == ICMP6_DST_UNREACH_NOPORT)) || (icmp6->icmp6_type == ICMP6_TIME_EXCEEDED && icmp6->icmp6_code == ICMP6_TIME_EXCEED_TRANSIT) || (icmp6->icmp6_type == ICMP6_PARAM_PROB && icmp6->icmp6_code == ICMP6_PARAMPROB_NEXTHEADER))) notify = tcp_drop_syn_sent; else notify = tcp_notify; break; default: notify = tcp_notify; } /* Check if we can safely get the ports from the tcp hdr */ if (m == NULL || (m->m_pkthdr.len < (int32_t) (off + sizeof(struct tcp_ports)))) { return; } bzero(&t_ports, sizeof(struct tcp_ports)); m_copydata(m, off, sizeof(struct tcp_ports), (caddr_t)&t_ports); inp = in6_pcblookup(&V_tcbinfo, &ip6->ip6_dst, t_ports.th_dport, &ip6->ip6_src, t_ports.th_sport, INPLOOKUP_WLOCKPCB, NULL); off += sizeof(struct tcp_ports); if (m->m_pkthdr.len < (int32_t) (off + sizeof(tcp_seq))) { goto out; } m_copydata(m, off, sizeof(tcp_seq), (caddr_t)&icmp_tcp_seq); if (inp != NULL) { - if (!(inp->inp_flags & INP_TIMEWAIT) && - !(inp->inp_flags & INP_DROPPED) && + if (!(inp->inp_flags & INP_DROPPED) && !(inp->inp_socket == NULL)) { tp = intotcpcb(inp); #ifdef TCP_OFFLOAD if (tp->t_flags & TF_TOE && errno == EMSGSIZE) { /* MTU discovery for offloaded connections. */ mtu = tcp6_next_pmtu(icmp6); tcp_offload_pmtu_update(tp, icmp_tcp_seq, mtu); goto out; } #endif if (tp->t_port != port) { goto out; } if (SEQ_GEQ(ntohl(icmp_tcp_seq), tp->snd_una) && SEQ_LT(ntohl(icmp_tcp_seq), tp->snd_max)) { if (errno == EMSGSIZE) { /* * MTU discovery: * If we got a needfrag set the MTU * in the route to the suggested new * value (if given) and then notify. */ mtu = tcp6_next_pmtu(icmp6); bzero(&inc, sizeof(inc)); inc.inc_fibnum = M_GETFIB(m); inc.inc_flags |= INC_ISIPV6; inc.inc6_faddr = *dst; if (in6_setscope(&inc.inc6_faddr, m->m_pkthdr.rcvif, NULL)) goto out; /* * Only process the offered MTU if it * is smaller than the current one. */ if (mtu < tp->t_maxseg + sizeof (struct tcphdr) + sizeof (struct ip6_hdr)) { tcp_hc_updatemtu(&inc, mtu); tcp_mtudisc(inp, mtu); ICMP6STAT_INC(icp6s_pmtuchg); } } else inp = (*notify)(inp, errno); } } } else { bzero(&inc, sizeof(inc)); inc.inc_fibnum = M_GETFIB(m); inc.inc_flags |= INC_ISIPV6; inc.inc_fport = t_ports.th_dport; inc.inc_lport = t_ports.th_sport; inc.inc6_faddr = *dst; inc.inc6_laddr = ip6->ip6_src; syncache_unreach(&inc, icmp_tcp_seq, port); } out: if (inp != NULL) INP_WUNLOCK(inp); } static void tcp6_ctlinput(struct ip6ctlparam *ctl) { tcp6_ctlinput_with_port(ctl, htons(0)); } static void tcp6_ctlinput_viaudp(udp_tun_icmp_param_t param) { struct ip6ctlparam *ip6cp = param.ip6cp; struct mbuf *m; struct udphdr *udp; uint16_t port; m = m_pulldown(ip6cp->ip6c_m, ip6cp->ip6c_off, sizeof(struct udphdr), NULL); if (m == NULL) { return; } udp = mtod(m, struct udphdr *); if (ntohs(udp->uh_sport) != V_tcp_udp_tunneling_port) { return; } port = udp->uh_dport; m_adj(m, sizeof(struct udphdr)); if ((m->m_flags & M_PKTHDR) == 0) { ip6cp->ip6c_m->m_pkthdr.len -= sizeof(struct udphdr); } /* Now call in to the normal handling code */ tcp6_ctlinput_with_port(ip6cp, port); } #endif /* INET6 */ static uint32_t tcp_keyed_hash(struct in_conninfo *inc, u_char *key, u_int len) { SIPHASH_CTX ctx; uint32_t hash[2]; KASSERT(len >= SIPHASH_KEY_LENGTH, ("%s: keylen %u too short ", __func__, len)); SipHash24_Init(&ctx); SipHash_SetKey(&ctx, (uint8_t *)key); SipHash_Update(&ctx, &inc->inc_fport, sizeof(uint16_t)); SipHash_Update(&ctx, &inc->inc_lport, sizeof(uint16_t)); switch (inc->inc_flags & INC_ISIPV6) { #ifdef INET case 0: SipHash_Update(&ctx, &inc->inc_faddr, sizeof(struct in_addr)); SipHash_Update(&ctx, &inc->inc_laddr, sizeof(struct in_addr)); break; #endif #ifdef INET6 case INC_ISIPV6: SipHash_Update(&ctx, &inc->inc6_faddr, sizeof(struct in6_addr)); SipHash_Update(&ctx, &inc->inc6_laddr, sizeof(struct in6_addr)); break; #endif } SipHash_Final((uint8_t *)hash, &ctx); return (hash[0] ^ hash[1]); } uint32_t tcp_new_ts_offset(struct in_conninfo *inc) { struct in_conninfo inc_store, *local_inc; if (!V_tcp_ts_offset_per_conn) { memcpy(&inc_store, inc, sizeof(struct in_conninfo)); inc_store.inc_lport = 0; inc_store.inc_fport = 0; local_inc = &inc_store; } else { local_inc = inc; } return (tcp_keyed_hash(local_inc, V_ts_offset_secret, sizeof(V_ts_offset_secret))); } /* * Following is where TCP initial sequence number generation occurs. * * There are two places where we must use initial sequence numbers: * 1. In SYN-ACK packets. * 2. In SYN packets. * * All ISNs for SYN-ACK packets are generated by the syncache. See * tcp_syncache.c for details. * * The ISNs in SYN packets must be monotonic; TIME_WAIT recycling * depends on this property. In addition, these ISNs should be * unguessable so as to prevent connection hijacking. To satisfy * the requirements of this situation, the algorithm outlined in * RFC 1948 is used, with only small modifications. * * Implementation details: * * Time is based off the system timer, and is corrected so that it * increases by one megabyte per second. This allows for proper * recycling on high speed LANs while still leaving over an hour * before rollover. * * As reading the *exact* system time is too expensive to be done * whenever setting up a TCP connection, we increment the time * offset in two ways. First, a small random positive increment * is added to isn_offset for each connection that is set up. * Second, the function tcp_isn_tick fires once per clock tick * and increments isn_offset as necessary so that sequence numbers * are incremented at approximately ISN_BYTES_PER_SECOND. The * random positive increments serve only to ensure that the same * exact sequence number is never sent out twice (as could otherwise * happen when a port is recycled in less than the system tick * interval.) * * net.inet.tcp.isn_reseed_interval controls the number of seconds * between seeding of isn_secret. This is normally set to zero, * as reseeding should not be necessary. * * Locking of the global variables isn_secret, isn_last_reseed, isn_offset, * isn_offset_old, and isn_ctx is performed using the ISN lock. In * general, this means holding an exclusive (write) lock. */ #define ISN_BYTES_PER_SECOND 1048576 #define ISN_STATIC_INCREMENT 4096 #define ISN_RANDOM_INCREMENT (4096 - 1) #define ISN_SECRET_LENGTH SIPHASH_KEY_LENGTH VNET_DEFINE_STATIC(u_char, isn_secret[ISN_SECRET_LENGTH]); VNET_DEFINE_STATIC(int, isn_last); VNET_DEFINE_STATIC(int, isn_last_reseed); VNET_DEFINE_STATIC(u_int32_t, isn_offset); VNET_DEFINE_STATIC(u_int32_t, isn_offset_old); #define V_isn_secret VNET(isn_secret) #define V_isn_last VNET(isn_last) #define V_isn_last_reseed VNET(isn_last_reseed) #define V_isn_offset VNET(isn_offset) #define V_isn_offset_old VNET(isn_offset_old) tcp_seq tcp_new_isn(struct in_conninfo *inc) { tcp_seq new_isn; u_int32_t projected_offset; ISN_LOCK(); /* Seed if this is the first use, reseed if requested. */ if ((V_isn_last_reseed == 0) || ((V_tcp_isn_reseed_interval > 0) && (((u_int)V_isn_last_reseed + (u_int)V_tcp_isn_reseed_interval*hz) < (u_int)ticks))) { arc4rand(&V_isn_secret, sizeof(V_isn_secret), 0); V_isn_last_reseed = ticks; } /* Compute the hash and return the ISN. */ new_isn = (tcp_seq)tcp_keyed_hash(inc, V_isn_secret, sizeof(V_isn_secret)); V_isn_offset += ISN_STATIC_INCREMENT + (arc4random() & ISN_RANDOM_INCREMENT); if (ticks != V_isn_last) { projected_offset = V_isn_offset_old + ISN_BYTES_PER_SECOND / hz * (ticks - V_isn_last); if (SEQ_GT(projected_offset, V_isn_offset)) V_isn_offset = projected_offset; V_isn_offset_old = V_isn_offset; V_isn_last = ticks; } new_isn += V_isn_offset; ISN_UNLOCK(); return (new_isn); } /* * When a specific ICMP unreachable message is received and the * connection state is SYN-SENT, drop the connection. This behavior * is controlled by the icmp_may_rst sysctl. */ static struct inpcb * tcp_drop_syn_sent(struct inpcb *inp, int errno) { struct tcpcb *tp; NET_EPOCH_ASSERT(); INP_WLOCK_ASSERT(inp); - if ((inp->inp_flags & INP_TIMEWAIT) || - (inp->inp_flags & INP_DROPPED)) + if (inp->inp_flags & INP_DROPPED) return (inp); tp = intotcpcb(inp); if (tp->t_state != TCPS_SYN_SENT) return (inp); if (IS_FASTOPEN(tp->t_flags)) tcp_fastopen_disable_path(tp); tp = tcp_drop(tp, errno); if (tp != NULL) return (inp); else return (NULL); } /* * When `need fragmentation' ICMP is received, update our idea of the MSS * based on the new value. Also nudge TCP to send something, since we * know the packet we just sent was dropped. * This duplicates some code in the tcp_mss() function in tcp_input.c. */ static struct inpcb * tcp_mtudisc_notify(struct inpcb *inp, int error) { return (tcp_mtudisc(inp, -1)); } static struct inpcb * tcp_mtudisc(struct inpcb *inp, int mtuoffer) { struct tcpcb *tp; struct socket *so; INP_WLOCK_ASSERT(inp); - if ((inp->inp_flags & INP_TIMEWAIT) || - (inp->inp_flags & INP_DROPPED)) + if (inp->inp_flags & INP_DROPPED) return (inp); tp = intotcpcb(inp); KASSERT(tp != NULL, ("tcp_mtudisc: tp == NULL")); tcp_mss_update(tp, -1, mtuoffer, NULL, NULL); so = inp->inp_socket; SOCKBUF_LOCK(&so->so_snd); /* If the mss is larger than the socket buffer, decrease the mss. */ if (so->so_snd.sb_hiwat < tp->t_maxseg) tp->t_maxseg = so->so_snd.sb_hiwat; SOCKBUF_UNLOCK(&so->so_snd); TCPSTAT_INC(tcps_mturesent); tp->t_rtttime = 0; tp->snd_nxt = tp->snd_una; tcp_free_sackholes(tp); tp->snd_recover = tp->snd_max; if (tp->t_flags & TF_SACK_PERMIT) EXIT_FASTRECOVERY(tp->t_flags); if (tp->t_fb->tfb_tcp_mtu_chg != NULL) { /* * Conceptually the snd_nxt setting * and freeing sack holes should * be done by the default stacks * own tfb_tcp_mtu_chg(). */ tp->t_fb->tfb_tcp_mtu_chg(tp); } if (tcp_output(tp) < 0) return (NULL); else return (inp); } #ifdef INET /* * Look-up the routing entry to the peer of this inpcb. If no route * is found and it cannot be allocated, then return 0. This routine * is called by TCP routines that access the rmx structure and by * tcp_mss_update to get the peer/interface MTU. */ uint32_t tcp_maxmtu(struct in_conninfo *inc, struct tcp_ifcap *cap) { struct nhop_object *nh; struct ifnet *ifp; uint32_t maxmtu = 0; KASSERT(inc != NULL, ("tcp_maxmtu with NULL in_conninfo pointer")); if (inc->inc_faddr.s_addr != INADDR_ANY) { nh = fib4_lookup(inc->inc_fibnum, inc->inc_faddr, 0, NHR_NONE, 0); if (nh == NULL) return (0); ifp = nh->nh_ifp; maxmtu = nh->nh_mtu; /* Report additional interface capabilities. */ if (cap != NULL) { if (ifp->if_capenable & IFCAP_TSO4 && ifp->if_hwassist & CSUM_TSO) { cap->ifcap |= CSUM_TSO; cap->tsomax = ifp->if_hw_tsomax; cap->tsomaxsegcount = ifp->if_hw_tsomaxsegcount; cap->tsomaxsegsize = ifp->if_hw_tsomaxsegsize; } } } return (maxmtu); } #endif /* INET */ #ifdef INET6 uint32_t tcp_maxmtu6(struct in_conninfo *inc, struct tcp_ifcap *cap) { struct nhop_object *nh; struct in6_addr dst6; uint32_t scopeid; struct ifnet *ifp; uint32_t maxmtu = 0; KASSERT(inc != NULL, ("tcp_maxmtu6 with NULL in_conninfo pointer")); if (inc->inc_flags & INC_IPV6MINMTU) return (IPV6_MMTU); if (!IN6_IS_ADDR_UNSPECIFIED(&inc->inc6_faddr)) { in6_splitscope(&inc->inc6_faddr, &dst6, &scopeid); nh = fib6_lookup(inc->inc_fibnum, &dst6, scopeid, NHR_NONE, 0); if (nh == NULL) return (0); ifp = nh->nh_ifp; maxmtu = nh->nh_mtu; /* Report additional interface capabilities. */ if (cap != NULL) { if (ifp->if_capenable & IFCAP_TSO6 && ifp->if_hwassist & CSUM_TSO) { cap->ifcap |= CSUM_TSO; cap->tsomax = ifp->if_hw_tsomax; cap->tsomaxsegcount = ifp->if_hw_tsomaxsegcount; cap->tsomaxsegsize = ifp->if_hw_tsomaxsegsize; } } } return (maxmtu); } /* * Handle setsockopt(IPV6_USE_MIN_MTU) by a TCP stack. * * XXXGL: we are updating inpcb here with INC_IPV6MINMTU flag. * The right place to do that is ip6_setpktopt() that has just been * executed. By the way it just filled ip6po_minmtu for us. */ void tcp6_use_min_mtu(struct tcpcb *tp) { struct inpcb *inp = tp->t_inpcb; INP_WLOCK_ASSERT(inp); /* * In case of the IPV6_USE_MIN_MTU socket * option, the INC_IPV6MINMTU flag to announce * a corresponding MSS during the initial * handshake. If the TCP connection is not in * the front states, just reduce the MSS being * used. This avoids the sending of TCP * segments which will be fragmented at the * IPv6 layer. */ inp->inp_inc.inc_flags |= INC_IPV6MINMTU; if ((tp->t_state >= TCPS_SYN_SENT) && (inp->inp_inc.inc_flags & INC_ISIPV6)) { struct ip6_pktopts *opt; opt = inp->in6p_outputopts; if (opt != NULL && opt->ip6po_minmtu == IP6PO_MINMTU_ALL && tp->t_maxseg > TCP6_MSS) tp->t_maxseg = TCP6_MSS; } } #endif /* INET6 */ /* * Calculate effective SMSS per RFC5681 definition for a given TCP * connection at its current state, taking into account SACK and etc. */ u_int tcp_maxseg(const struct tcpcb *tp) { u_int optlen; if (tp->t_flags & TF_NOOPT) return (tp->t_maxseg); /* * Here we have a simplified code from tcp_addoptions(), * without a proper loop, and having most of paddings hardcoded. * We might make mistakes with padding here in some edge cases, * but this is harmless, since result of tcp_maxseg() is used * only in cwnd and ssthresh estimations. */ if (TCPS_HAVEESTABLISHED(tp->t_state)) { if (tp->t_flags & TF_RCVD_TSTMP) optlen = TCPOLEN_TSTAMP_APPA; else optlen = 0; #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) if (tp->t_flags & TF_SIGNATURE) optlen += PADTCPOLEN(TCPOLEN_SIGNATURE); #endif if ((tp->t_flags & TF_SACK_PERMIT) && tp->rcv_numsacks > 0) { optlen += TCPOLEN_SACKHDR; optlen += tp->rcv_numsacks * TCPOLEN_SACK; optlen = PADTCPOLEN(optlen); } } else { if (tp->t_flags & TF_REQ_TSTMP) optlen = TCPOLEN_TSTAMP_APPA; else optlen = PADTCPOLEN(TCPOLEN_MAXSEG); if (tp->t_flags & TF_REQ_SCALE) optlen += PADTCPOLEN(TCPOLEN_WINDOW); #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) if (tp->t_flags & TF_SIGNATURE) optlen += PADTCPOLEN(TCPOLEN_SIGNATURE); #endif if (tp->t_flags & TF_SACK_PERMIT) optlen += PADTCPOLEN(TCPOLEN_SACK_PERMITTED); } #undef PAD optlen = min(optlen, TCP_MAXOLEN); return (tp->t_maxseg - optlen); } u_int tcp_fixed_maxseg(const struct tcpcb *tp) { int optlen; if (tp->t_flags & TF_NOOPT) return (tp->t_maxseg); /* * Here we have a simplified code from tcp_addoptions(), * without a proper loop, and having most of paddings hardcoded. * We only consider fixed options that we would send every * time I.e. SACK is not considered. This is important * for cc modules to figure out what the modulo of the * cwnd should be. */ #define PAD(len) ((((len) / 4) + !!((len) % 4)) * 4) if (TCPS_HAVEESTABLISHED(tp->t_state)) { if (tp->t_flags & TF_RCVD_TSTMP) optlen = TCPOLEN_TSTAMP_APPA; else optlen = 0; #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) if (tp->t_flags & TF_SIGNATURE) optlen += PAD(TCPOLEN_SIGNATURE); #endif } else { if (tp->t_flags & TF_REQ_TSTMP) optlen = TCPOLEN_TSTAMP_APPA; else optlen = PAD(TCPOLEN_MAXSEG); if (tp->t_flags & TF_REQ_SCALE) optlen += PAD(TCPOLEN_WINDOW); #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) if (tp->t_flags & TF_SIGNATURE) optlen += PAD(TCPOLEN_SIGNATURE); #endif if (tp->t_flags & TF_SACK_PERMIT) optlen += PAD(TCPOLEN_SACK_PERMITTED); } #undef PAD optlen = min(optlen, TCP_MAXOLEN); return (tp->t_maxseg - optlen); } static int sysctl_drop(SYSCTL_HANDLER_ARGS) { /* addrs[0] is a foreign socket, addrs[1] is a local one. */ struct sockaddr_storage addrs[2]; struct inpcb *inp; struct tcpcb *tp; #ifdef INET struct sockaddr_in *fin = NULL, *lin = NULL; #endif struct epoch_tracker et; #ifdef INET6 struct sockaddr_in6 *fin6, *lin6; #endif int error; inp = NULL; #ifdef INET6 fin6 = lin6 = NULL; #endif error = 0; if (req->oldptr != NULL || req->oldlen != 0) return (EINVAL); if (req->newptr == NULL) return (EPERM); if (req->newlen < sizeof(addrs)) return (ENOMEM); error = SYSCTL_IN(req, &addrs, sizeof(addrs)); if (error) return (error); switch (addrs[0].ss_family) { #ifdef INET6 case AF_INET6: fin6 = (struct sockaddr_in6 *)&addrs[0]; lin6 = (struct sockaddr_in6 *)&addrs[1]; if (fin6->sin6_len != sizeof(struct sockaddr_in6) || lin6->sin6_len != sizeof(struct sockaddr_in6)) return (EINVAL); if (IN6_IS_ADDR_V4MAPPED(&fin6->sin6_addr)) { if (!IN6_IS_ADDR_V4MAPPED(&lin6->sin6_addr)) return (EINVAL); in6_sin6_2_sin_in_sock((struct sockaddr *)&addrs[0]); in6_sin6_2_sin_in_sock((struct sockaddr *)&addrs[1]); #ifdef INET fin = (struct sockaddr_in *)&addrs[0]; lin = (struct sockaddr_in *)&addrs[1]; #endif break; } error = sa6_embedscope(fin6, V_ip6_use_defzone); if (error) return (error); error = sa6_embedscope(lin6, V_ip6_use_defzone); if (error) return (error); break; #endif #ifdef INET case AF_INET: fin = (struct sockaddr_in *)&addrs[0]; lin = (struct sockaddr_in *)&addrs[1]; if (fin->sin_len != sizeof(struct sockaddr_in) || lin->sin_len != sizeof(struct sockaddr_in)) return (EINVAL); break; #endif default: return (EINVAL); } NET_EPOCH_ENTER(et); switch (addrs[0].ss_family) { #ifdef INET6 case AF_INET6: inp = in6_pcblookup(&V_tcbinfo, &fin6->sin6_addr, fin6->sin6_port, &lin6->sin6_addr, lin6->sin6_port, INPLOOKUP_WLOCKPCB, NULL); break; #endif #ifdef INET case AF_INET: inp = in_pcblookup(&V_tcbinfo, fin->sin_addr, fin->sin_port, lin->sin_addr, lin->sin_port, INPLOOKUP_WLOCKPCB, NULL); break; #endif } if (inp != NULL) { if ((inp->inp_flags & INP_DROPPED) == 0 && !SOLISTENING(inp->inp_socket)) { tp = intotcpcb(inp); tp = tcp_drop(tp, ECONNABORTED); if (tp != NULL) INP_WUNLOCK(inp); } else INP_WUNLOCK(inp); } else error = ESRCH; NET_EPOCH_EXIT(et); return (error); } SYSCTL_PROC(_net_inet_tcp, TCPCTL_DROP, drop, CTLFLAG_VNET | CTLTYPE_STRUCT | CTLFLAG_WR | CTLFLAG_SKIP | CTLFLAG_NEEDGIANT, NULL, 0, sysctl_drop, "", "Drop TCP connection"); static int tcp_sysctl_setsockopt(SYSCTL_HANDLER_ARGS) { return (sysctl_setsockopt(oidp, arg1, arg2, req, &V_tcbinfo, &tcp_ctloutput_set)); } SYSCTL_PROC(_net_inet_tcp, OID_AUTO, setsockopt, CTLFLAG_VNET | CTLTYPE_STRUCT | CTLFLAG_WR | CTLFLAG_SKIP | CTLFLAG_MPSAFE, NULL, 0, tcp_sysctl_setsockopt, "", "Set socket option for TCP endpoint"); #ifdef KERN_TLS static int sysctl_switch_tls(SYSCTL_HANDLER_ARGS) { /* addrs[0] is a foreign socket, addrs[1] is a local one. */ struct sockaddr_storage addrs[2]; struct inpcb *inp; #ifdef INET struct sockaddr_in *fin = NULL, *lin = NULL; #endif struct epoch_tracker et; #ifdef INET6 struct sockaddr_in6 *fin6, *lin6; #endif int error; inp = NULL; #ifdef INET6 fin6 = lin6 = NULL; #endif error = 0; if (req->oldptr != NULL || req->oldlen != 0) return (EINVAL); if (req->newptr == NULL) return (EPERM); if (req->newlen < sizeof(addrs)) return (ENOMEM); error = SYSCTL_IN(req, &addrs, sizeof(addrs)); if (error) return (error); switch (addrs[0].ss_family) { #ifdef INET6 case AF_INET6: fin6 = (struct sockaddr_in6 *)&addrs[0]; lin6 = (struct sockaddr_in6 *)&addrs[1]; if (fin6->sin6_len != sizeof(struct sockaddr_in6) || lin6->sin6_len != sizeof(struct sockaddr_in6)) return (EINVAL); if (IN6_IS_ADDR_V4MAPPED(&fin6->sin6_addr)) { if (!IN6_IS_ADDR_V4MAPPED(&lin6->sin6_addr)) return (EINVAL); in6_sin6_2_sin_in_sock((struct sockaddr *)&addrs[0]); in6_sin6_2_sin_in_sock((struct sockaddr *)&addrs[1]); #ifdef INET fin = (struct sockaddr_in *)&addrs[0]; lin = (struct sockaddr_in *)&addrs[1]; #endif break; } error = sa6_embedscope(fin6, V_ip6_use_defzone); if (error) return (error); error = sa6_embedscope(lin6, V_ip6_use_defzone); if (error) return (error); break; #endif #ifdef INET case AF_INET: fin = (struct sockaddr_in *)&addrs[0]; lin = (struct sockaddr_in *)&addrs[1]; if (fin->sin_len != sizeof(struct sockaddr_in) || lin->sin_len != sizeof(struct sockaddr_in)) return (EINVAL); break; #endif default: return (EINVAL); } NET_EPOCH_ENTER(et); switch (addrs[0].ss_family) { #ifdef INET6 case AF_INET6: inp = in6_pcblookup(&V_tcbinfo, &fin6->sin6_addr, fin6->sin6_port, &lin6->sin6_addr, lin6->sin6_port, INPLOOKUP_WLOCKPCB, NULL); break; #endif #ifdef INET case AF_INET: inp = in_pcblookup(&V_tcbinfo, fin->sin_addr, fin->sin_port, lin->sin_addr, lin->sin_port, INPLOOKUP_WLOCKPCB, NULL); break; #endif } NET_EPOCH_EXIT(et); if (inp != NULL) { - if ((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) != 0 || + if ((inp->inp_flags & INP_DROPPED) != 0 || inp->inp_socket == NULL) { error = ECONNRESET; INP_WUNLOCK(inp); } else { struct socket *so; so = inp->inp_socket; soref(so); error = ktls_set_tx_mode(so, arg2 == 0 ? TCP_TLS_MODE_SW : TCP_TLS_MODE_IFNET); INP_WUNLOCK(inp); sorele(so); } } else error = ESRCH; return (error); } SYSCTL_PROC(_net_inet_tcp, OID_AUTO, switch_to_sw_tls, CTLFLAG_VNET | CTLTYPE_STRUCT | CTLFLAG_WR | CTLFLAG_SKIP | CTLFLAG_NEEDGIANT, NULL, 0, sysctl_switch_tls, "", "Switch TCP connection to SW TLS"); SYSCTL_PROC(_net_inet_tcp, OID_AUTO, switch_to_ifnet_tls, CTLFLAG_VNET | CTLTYPE_STRUCT | CTLFLAG_WR | CTLFLAG_SKIP | CTLFLAG_NEEDGIANT, NULL, 1, sysctl_switch_tls, "", "Switch TCP connection to ifnet TLS"); #endif /* * Generate a standardized TCP log line for use throughout the * tcp subsystem. Memory allocation is done with M_NOWAIT to * allow use in the interrupt context. * * NB: The caller MUST free(s, M_TCPLOG) the returned string. * NB: The function may return NULL if memory allocation failed. * * Due to header inclusion and ordering limitations the struct ip * and ip6_hdr pointers have to be passed as void pointers. */ char * tcp_log_vain(struct in_conninfo *inc, struct tcphdr *th, const void *ip4hdr, const void *ip6hdr) { /* Is logging enabled? */ if (V_tcp_log_in_vain == 0) return (NULL); return (tcp_log_addr(inc, th, ip4hdr, ip6hdr)); } char * tcp_log_addrs(struct in_conninfo *inc, struct tcphdr *th, const void *ip4hdr, const void *ip6hdr) { /* Is logging enabled? */ if (tcp_log_debug == 0) return (NULL); return (tcp_log_addr(inc, th, ip4hdr, ip6hdr)); } static char * tcp_log_addr(struct in_conninfo *inc, struct tcphdr *th, const void *ip4hdr, const void *ip6hdr) { char *s, *sp; size_t size; #ifdef INET const struct ip *ip = (const struct ip *)ip4hdr; #endif #ifdef INET6 const struct ip6_hdr *ip6 = (const struct ip6_hdr *)ip6hdr; #endif /* INET6 */ /* * The log line looks like this: * "TCP: [1.2.3.4]:50332 to [1.2.3.4]:80 tcpflags 0x2" */ size = sizeof("TCP: []:12345 to []:12345 tcpflags 0x2<>") + sizeof(PRINT_TH_FLAGS) + 1 + #ifdef INET6 2 * INET6_ADDRSTRLEN; #else 2 * INET_ADDRSTRLEN; #endif /* INET6 */ s = malloc(size, M_TCPLOG, M_ZERO|M_NOWAIT); if (s == NULL) return (NULL); strcat(s, "TCP: ["); sp = s + strlen(s); if (inc && ((inc->inc_flags & INC_ISIPV6) == 0)) { inet_ntoa_r(inc->inc_faddr, sp); sp = s + strlen(s); sprintf(sp, "]:%i to [", ntohs(inc->inc_fport)); sp = s + strlen(s); inet_ntoa_r(inc->inc_laddr, sp); sp = s + strlen(s); sprintf(sp, "]:%i", ntohs(inc->inc_lport)); #ifdef INET6 } else if (inc) { ip6_sprintf(sp, &inc->inc6_faddr); sp = s + strlen(s); sprintf(sp, "]:%i to [", ntohs(inc->inc_fport)); sp = s + strlen(s); ip6_sprintf(sp, &inc->inc6_laddr); sp = s + strlen(s); sprintf(sp, "]:%i", ntohs(inc->inc_lport)); } else if (ip6 && th) { ip6_sprintf(sp, &ip6->ip6_src); sp = s + strlen(s); sprintf(sp, "]:%i to [", ntohs(th->th_sport)); sp = s + strlen(s); ip6_sprintf(sp, &ip6->ip6_dst); sp = s + strlen(s); sprintf(sp, "]:%i", ntohs(th->th_dport)); #endif /* INET6 */ #ifdef INET } else if (ip && th) { inet_ntoa_r(ip->ip_src, sp); sp = s + strlen(s); sprintf(sp, "]:%i to [", ntohs(th->th_sport)); sp = s + strlen(s); inet_ntoa_r(ip->ip_dst, sp); sp = s + strlen(s); sprintf(sp, "]:%i", ntohs(th->th_dport)); #endif /* INET */ } else { free(s, M_TCPLOG); return (NULL); } sp = s + strlen(s); if (th) sprintf(sp, " tcpflags 0x%b", tcp_get_flags(th), PRINT_TH_FLAGS); if (*(s + size - 1) != '\0') panic("%s: string too long", __func__); return (s); } /* * A subroutine which makes it easy to track TCP state changes with DTrace. * This function shouldn't be called for t_state initializations that don't * correspond to actual TCP state transitions. */ void tcp_state_change(struct tcpcb *tp, int newstate) { #if defined(KDTRACE_HOOKS) int pstate = tp->t_state; #endif TCPSTATES_DEC(tp->t_state); TCPSTATES_INC(newstate); tp->t_state = newstate; TCP_PROBE6(state__change, NULL, tp, NULL, tp, NULL, pstate); } /* * Create an external-format (``xtcpcb'') structure using the information in * the kernel-format tcpcb structure pointed to by tp. This is done to * reduce the spew of irrelevant information over this interface, to isolate * user code from changes in the kernel structure, and potentially to provide * information-hiding if we decide that some of this information should be * hidden from users. */ void tcp_inptoxtp(const struct inpcb *inp, struct xtcpcb *xt) { struct tcpcb *tp = intotcpcb(inp); sbintime_t now; bzero(xt, sizeof(*xt)); xt->t_state = tp->t_state; xt->t_logstate = tp->t_logstate; xt->t_flags = tp->t_flags; xt->t_sndzerowin = tp->t_sndzerowin; xt->t_sndrexmitpack = tp->t_sndrexmitpack; xt->t_rcvoopack = tp->t_rcvoopack; xt->t_rcv_wnd = tp->rcv_wnd; xt->t_snd_wnd = tp->snd_wnd; xt->t_snd_cwnd = tp->snd_cwnd; xt->t_snd_ssthresh = tp->snd_ssthresh; xt->t_dsack_bytes = tp->t_dsack_bytes; xt->t_dsack_tlp_bytes = tp->t_dsack_tlp_bytes; xt->t_dsack_pack = tp->t_dsack_pack; xt->t_maxseg = tp->t_maxseg; xt->xt_ecn = (tp->t_flags2 & TF2_ECN_PERMIT) ? 1 : 0 + (tp->t_flags2 & TF2_ACE_PERMIT) ? 2 : 0; now = getsbinuptime(); #define COPYTIMER(ttt) do { \ if (callout_active(&tp->t_timers->ttt)) \ xt->ttt = (tp->t_timers->ttt.c_time - now) / \ SBT_1MS; \ else \ xt->ttt = 0; \ } while (0) COPYTIMER(tt_delack); COPYTIMER(tt_rexmt); COPYTIMER(tt_persist); COPYTIMER(tt_keep); COPYTIMER(tt_2msl); #undef COPYTIMER xt->t_rcvtime = 1000 * (ticks - tp->t_rcvtime) / hz; xt->xt_encaps_port = tp->t_port; bcopy(tp->t_fb->tfb_tcp_block_name, xt->xt_stack, TCP_FUNCTION_NAME_LEN_MAX); bcopy(CC_ALGO(tp)->name, xt->xt_cc, TCP_CA_NAME_MAX); #ifdef TCP_BLACKBOX (void)tcp_log_get_id(tp, xt->xt_logid); #endif xt->xt_len = sizeof(struct xtcpcb); in_pcbtoxinpcb(inp, &xt->xt_inp); if (inp->inp_socket == NULL) xt->xt_inp.xi_socket.xso_protocol = IPPROTO_TCP; } void tcp_log_end_status(struct tcpcb *tp, uint8_t status) { uint32_t bit, i; if ((tp == NULL) || (status > TCP_EI_STATUS_MAX_VALUE) || (status == 0)) { /* Invalid */ return; } if (status > (sizeof(uint32_t) * 8)) { /* Should this be a KASSERT? */ return; } bit = 1U << (status - 1); if (bit & tp->t_end_info_status) { /* already logged */ return; } for (i = 0; i < TCP_END_BYTE_INFO; i++) { if (tp->t_end_info_bytes[i] == TCP_EI_EMPTY_SLOT) { tp->t_end_info_bytes[i] = status; tp->t_end_info_status |= bit; break; } } } int tcp_can_enable_pacing(void) { if ((tcp_pacing_limit == -1) || (tcp_pacing_limit > number_of_tcp_connections_pacing)) { atomic_fetchadd_int(&number_of_tcp_connections_pacing, 1); shadow_num_connections = number_of_tcp_connections_pacing; return (1); } else { return (0); } } static uint8_t tcp_pacing_warning = 0; void tcp_decrement_paced_conn(void) { uint32_t ret; ret = atomic_fetchadd_int(&number_of_tcp_connections_pacing, -1); shadow_num_connections = number_of_tcp_connections_pacing; KASSERT(ret != 0, ("tcp_paced_connection_exits -1 would cause wrap?")); if (ret == 0) { if (tcp_pacing_limit != -1) { printf("Warning all pacing is now disabled, count decrements invalidly!\n"); tcp_pacing_limit = 0; } else if (tcp_pacing_warning == 0) { printf("Warning pacing count is invalid, invalid decrement\n"); tcp_pacing_warning = 1; } } } diff --git a/sys/netinet/tcp_timer.c b/sys/netinet/tcp_timer.c index f4915da6e77c..391b9dfdbc05 100644 --- a/sys/netinet/tcp_timer.c +++ b/sys/netinet/tcp_timer.c @@ -1,1148 +1,1148 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995 * The Regents of the University of California. All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)tcp_timer.c 8.2 (Berkeley) 5/24/95 */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include "opt_tcpdebug.h" #include "opt_rss.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef INET6 #include #endif #include #include #include #include #include #include #include #include #ifdef INET6 #include #endif #include #include int tcp_persmin; SYSCTL_PROC(_net_inet_tcp, OID_AUTO, persmin, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &tcp_persmin, 0, sysctl_msec_to_ticks, "I", "minimum persistence interval"); int tcp_persmax; SYSCTL_PROC(_net_inet_tcp, OID_AUTO, persmax, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &tcp_persmax, 0, sysctl_msec_to_ticks, "I", "maximum persistence interval"); int tcp_keepinit; SYSCTL_PROC(_net_inet_tcp, TCPCTL_KEEPINIT, keepinit, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &tcp_keepinit, 0, sysctl_msec_to_ticks, "I", "time to establish connection"); int tcp_keepidle; SYSCTL_PROC(_net_inet_tcp, TCPCTL_KEEPIDLE, keepidle, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &tcp_keepidle, 0, sysctl_msec_to_ticks, "I", "time before keepalive probes begin"); int tcp_keepintvl; SYSCTL_PROC(_net_inet_tcp, TCPCTL_KEEPINTVL, keepintvl, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &tcp_keepintvl, 0, sysctl_msec_to_ticks, "I", "time between keepalive probes"); int tcp_delacktime; SYSCTL_PROC(_net_inet_tcp, TCPCTL_DELACKTIME, delacktime, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &tcp_delacktime, 0, sysctl_msec_to_ticks, "I", "Time before a delayed ACK is sent"); VNET_DEFINE(int, tcp_msl); SYSCTL_PROC(_net_inet_tcp, OID_AUTO, msl, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_VNET, &VNET_NAME(tcp_msl), 0, sysctl_msec_to_ticks, "I", "Maximum segment lifetime"); int tcp_rexmit_initial; SYSCTL_PROC(_net_inet_tcp, OID_AUTO, rexmit_initial, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &tcp_rexmit_initial, 0, sysctl_msec_to_ticks, "I", "Initial Retransmission Timeout"); int tcp_rexmit_min; SYSCTL_PROC(_net_inet_tcp, OID_AUTO, rexmit_min, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &tcp_rexmit_min, 0, sysctl_msec_to_ticks, "I", "Minimum Retransmission Timeout"); int tcp_rexmit_slop; SYSCTL_PROC(_net_inet_tcp, OID_AUTO, rexmit_slop, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &tcp_rexmit_slop, 0, sysctl_msec_to_ticks, "I", "Retransmission Timer Slop"); VNET_DEFINE(int, tcp_always_keepalive) = 1; SYSCTL_INT(_net_inet_tcp, OID_AUTO, always_keepalive, CTLFLAG_VNET|CTLFLAG_RW, &VNET_NAME(tcp_always_keepalive) , 0, "Assume SO_KEEPALIVE on all TCP connections"); int tcp_fast_finwait2_recycle = 0; SYSCTL_INT(_net_inet_tcp, OID_AUTO, fast_finwait2_recycle, CTLFLAG_RW, &tcp_fast_finwait2_recycle, 0, "Recycle closed FIN_WAIT_2 connections faster"); int tcp_finwait2_timeout; SYSCTL_PROC(_net_inet_tcp, OID_AUTO, finwait2_timeout, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, &tcp_finwait2_timeout, 0, sysctl_msec_to_ticks, "I", "FIN-WAIT2 timeout"); int tcp_keepcnt = TCPTV_KEEPCNT; SYSCTL_INT(_net_inet_tcp, OID_AUTO, keepcnt, CTLFLAG_RW, &tcp_keepcnt, 0, "Number of keepalive probes to send"); /* max idle probes */ int tcp_maxpersistidle; int tcp_rexmit_drop_options = 0; SYSCTL_INT(_net_inet_tcp, OID_AUTO, rexmit_drop_options, CTLFLAG_RW, &tcp_rexmit_drop_options, 0, "Drop TCP options from 3rd and later retransmitted SYN"); int tcp_maxunacktime = TCPTV_MAXUNACKTIME; SYSCTL_PROC(_net_inet_tcp, OID_AUTO, maxunacktime, CTLTYPE_INT|CTLFLAG_RW | CTLFLAG_NEEDGIANT, &tcp_maxunacktime, 0, sysctl_msec_to_ticks, "I", "Maximum time (in ms) that a session can linger without making progress"); VNET_DEFINE(int, tcp_pmtud_blackhole_detect); SYSCTL_INT(_net_inet_tcp, OID_AUTO, pmtud_blackhole_detection, CTLFLAG_RW|CTLFLAG_VNET, &VNET_NAME(tcp_pmtud_blackhole_detect), 0, "Path MTU Discovery Black Hole Detection Enabled"); #ifdef INET VNET_DEFINE(int, tcp_pmtud_blackhole_mss) = 1200; SYSCTL_INT(_net_inet_tcp, OID_AUTO, pmtud_blackhole_mss, CTLFLAG_RW|CTLFLAG_VNET, &VNET_NAME(tcp_pmtud_blackhole_mss), 0, "Path MTU Discovery Black Hole Detection lowered MSS"); #endif #ifdef INET6 VNET_DEFINE(int, tcp_v6pmtud_blackhole_mss) = 1220; SYSCTL_INT(_net_inet_tcp, OID_AUTO, v6pmtud_blackhole_mss, CTLFLAG_RW|CTLFLAG_VNET, &VNET_NAME(tcp_v6pmtud_blackhole_mss), 0, "Path MTU Discovery IPv6 Black Hole Detection lowered MSS"); #endif #ifdef RSS static int per_cpu_timers = 1; #else static int per_cpu_timers = 0; #endif SYSCTL_INT(_net_inet_tcp, OID_AUTO, per_cpu_timers, CTLFLAG_RW, &per_cpu_timers , 0, "run tcp timers on all cpus"); /* * Map the given inp to a CPU id. * * This queries RSS if it's compiled in, else it defaults to the current * CPU ID. */ inline int inp_to_cpuid(struct inpcb *inp) { u_int cpuid; if (per_cpu_timers) { #ifdef RSS cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype); if (cpuid == NETISR_CPUID_NONE) return (curcpu); /* XXX */ else return (cpuid); #endif /* * We don't have a flowid -> cpuid mapping, so cheat and * just map unknown cpuids to curcpu. Not the best, but * apparently better than defaulting to swi 0. */ cpuid = inp->inp_flowid % (mp_maxid + 1); if (! CPU_ABSENT(cpuid)) return (cpuid); return (curcpu); } else { return (0); } } int tcp_backoff[TCP_MAXRXTSHIFT + 1] = { 1, 2, 4, 8, 16, 32, 64, 128, 256, 512, 512, 512, 512 }; int tcp_totbackoff = 2559; /* sum of tcp_backoff[] */ /* * TCP timer processing. */ void tcp_timer_delack(void *xtp) { struct epoch_tracker et; struct tcpcb *tp = xtp; struct inpcb *inp; CURVNET_SET(tp->t_vnet); inp = tp->t_inpcb; KASSERT(inp != NULL, ("%s: tp %p tp->t_inpcb == NULL", __func__, tp)); INP_WLOCK(inp); if (callout_pending(&tp->t_timers->tt_delack) || !callout_active(&tp->t_timers->tt_delack)) { INP_WUNLOCK(inp); CURVNET_RESTORE(); return; } callout_deactivate(&tp->t_timers->tt_delack); if ((inp->inp_flags & INP_DROPPED) != 0) { INP_WUNLOCK(inp); CURVNET_RESTORE(); return; } tp->t_flags |= TF_ACKNOW; TCPSTAT_INC(tcps_delack); NET_EPOCH_ENTER(et); (void) tcp_output_unlock(tp); NET_EPOCH_EXIT(et); CURVNET_RESTORE(); } /* * Call tcp_close() from a callout context. */ static void tcp_timer_close(struct tcpcb *tp) { struct epoch_tracker et; struct inpcb *inp = tp->t_inpcb; INP_WLOCK_ASSERT(inp); NET_EPOCH_ENTER(et); tp = tcp_close(tp); NET_EPOCH_EXIT(et); if (tp != NULL) INP_WUNLOCK(inp); } /* * Call tcp_drop() from a callout context. */ static void tcp_timer_drop(struct tcpcb *tp) { struct epoch_tracker et; struct inpcb *inp = tp->t_inpcb; INP_WLOCK_ASSERT(inp); NET_EPOCH_ENTER(et); tp = tcp_drop(tp, ETIMEDOUT); NET_EPOCH_EXIT(et); if (tp != NULL) INP_WUNLOCK(inp); } void tcp_timer_2msl(void *xtp) { struct tcpcb *tp = xtp; struct inpcb *inp; CURVNET_SET(tp->t_vnet); #ifdef TCPDEBUG int ostate; ostate = tp->t_state; #endif inp = tp->t_inpcb; KASSERT(inp != NULL, ("%s: tp %p tp->t_inpcb == NULL", __func__, tp)); INP_WLOCK(inp); tcp_log_end_status(tp, TCP_EI_STATUS_2MSL); tcp_free_sackholes(tp); if (callout_pending(&tp->t_timers->tt_2msl) || !callout_active(&tp->t_timers->tt_2msl)) { INP_WUNLOCK(tp->t_inpcb); CURVNET_RESTORE(); return; } callout_deactivate(&tp->t_timers->tt_2msl); - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { INP_WUNLOCK(inp); CURVNET_RESTORE(); return; } KASSERT((tp->t_timers->tt_flags & TT_STOPPED) == 0, ("%s: tp %p tcpcb can't be stopped here", __func__, tp)); /* * 2 MSL timeout in shutdown went off. If we're closed but * still waiting for peer to close and connection has been idle * too long delete connection control block. Otherwise, check * again in a bit. * * If fastrecycle of FIN_WAIT_2, in FIN_WAIT_2 and receiver has closed, * there's no point in hanging onto FIN_WAIT_2 socket. Just close it. * Ignore fact that there were recent incoming segments. */ if (tp->t_state == TCPS_TIME_WAIT) { tcp_timer_close(tp); CURVNET_RESTORE(); return; } else if (tp->t_state == TCPS_FIN_WAIT_2 && tcp_fast_finwait2_recycle && tp->t_inpcb->inp_socket && (tp->t_inpcb->inp_socket->so_rcv.sb_state & SBS_CANTRCVMORE)) { TCPSTAT_INC(tcps_finwait2_drops); tcp_timer_close(tp); CURVNET_RESTORE(); return; } else { if (ticks - tp->t_rcvtime <= TP_MAXIDLE(tp)) { callout_reset(&tp->t_timers->tt_2msl, TP_KEEPINTVL(tp), tcp_timer_2msl, tp); } else { tcp_timer_close(tp); CURVNET_RESTORE(); return; } } #ifdef TCPDEBUG if (tp->t_inpcb->inp_socket->so_options & SO_DEBUG) tcp_trace(TA_USER, ostate, tp, (void *)0, (struct tcphdr *)0, PRU_SLOWTIMO); #endif TCP_PROBE2(debug__user, tp, PRU_SLOWTIMO); INP_WUNLOCK(inp); CURVNET_RESTORE(); } void tcp_timer_keep(void *xtp) { struct tcpcb *tp = xtp; struct tcptemp *t_template; struct inpcb *inp; struct epoch_tracker et; CURVNET_SET(tp->t_vnet); #ifdef TCPDEBUG int ostate; ostate = tp->t_state; #endif inp = tp->t_inpcb; KASSERT(inp != NULL, ("%s: tp %p tp->t_inpcb == NULL", __func__, tp)); INP_WLOCK(inp); if (callout_pending(&tp->t_timers->tt_keep) || !callout_active(&tp->t_timers->tt_keep)) { INP_WUNLOCK(inp); CURVNET_RESTORE(); return; } callout_deactivate(&tp->t_timers->tt_keep); - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { INP_WUNLOCK(inp); CURVNET_RESTORE(); return; } KASSERT((tp->t_timers->tt_flags & TT_STOPPED) == 0, ("%s: tp %p tcpcb can't be stopped here", __func__, tp)); /* * Because we don't regularly reset the keepalive callout in * the ESTABLISHED state, it may be that we don't actually need * to send a keepalive yet. If that occurs, schedule another * call for the next time the keepalive timer might expire. */ if (TCPS_HAVEESTABLISHED(tp->t_state)) { u_int idletime; idletime = ticks - tp->t_rcvtime; if (idletime < TP_KEEPIDLE(tp)) { callout_reset(&tp->t_timers->tt_keep, TP_KEEPIDLE(tp) - idletime, tcp_timer_keep, tp); INP_WUNLOCK(inp); CURVNET_RESTORE(); return; } } /* * Keep-alive timer went off; send something * or drop connection if idle for too long. */ TCPSTAT_INC(tcps_keeptimeo); if (tp->t_state < TCPS_ESTABLISHED) goto dropit; if ((V_tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && tp->t_state <= TCPS_CLOSING) { if (ticks - tp->t_rcvtime >= TP_KEEPIDLE(tp) + TP_MAXIDLE(tp)) goto dropit; /* * Send a packet designed to force a response * if the peer is up and reachable: * either an ACK if the connection is still alive, * or an RST if the peer has closed the connection * due to timeout or reboot. * Using sequence number tp->snd_una-1 * causes the transmitted zero-length segment * to lie outside the receive window; * by the protocol spec, this requires the * correspondent TCP to respond. */ TCPSTAT_INC(tcps_keepprobe); t_template = tcpip_maketemplate(inp); if (t_template) { NET_EPOCH_ENTER(et); tcp_respond(tp, t_template->tt_ipgen, &t_template->tt_t, (struct mbuf *)NULL, tp->rcv_nxt, tp->snd_una - 1, 0); NET_EPOCH_EXIT(et); free(t_template, M_TEMP); } callout_reset(&tp->t_timers->tt_keep, TP_KEEPINTVL(tp), tcp_timer_keep, tp); } else callout_reset(&tp->t_timers->tt_keep, TP_KEEPIDLE(tp), tcp_timer_keep, tp); #ifdef TCPDEBUG if (inp->inp_socket->so_options & SO_DEBUG) tcp_trace(TA_USER, ostate, tp, (void *)0, (struct tcphdr *)0, PRU_SLOWTIMO); #endif TCP_PROBE2(debug__user, tp, PRU_SLOWTIMO); INP_WUNLOCK(inp); CURVNET_RESTORE(); return; dropit: TCPSTAT_INC(tcps_keepdrops); NET_EPOCH_ENTER(et); tcp_log_end_status(tp, TCP_EI_STATUS_KEEP_MAX); tp = tcp_drop(tp, ETIMEDOUT); #ifdef TCPDEBUG if (tp != NULL && (tp->t_inpcb->inp_socket->so_options & SO_DEBUG)) tcp_trace(TA_USER, ostate, tp, (void *)0, (struct tcphdr *)0, PRU_SLOWTIMO); #endif TCP_PROBE2(debug__user, tp, PRU_SLOWTIMO); NET_EPOCH_EXIT(et); if (tp != NULL) INP_WUNLOCK(inp); CURVNET_RESTORE(); } /* * Has this session exceeded the maximum time without seeing a substantive * acknowledgement? If so, return true; otherwise false. */ static bool tcp_maxunacktime_check(struct tcpcb *tp) { /* Are we tracking this timer for this session? */ if (TP_MAXUNACKTIME(tp) == 0) return false; /* Do we have a current measurement. */ if (tp->t_acktime == 0) return false; /* Are we within the acceptable range? */ if (TSTMP_GT(TP_MAXUNACKTIME(tp) + tp->t_acktime, (u_int)ticks)) return false; /* We exceeded the timer. */ TCPSTAT_INC(tcps_progdrops); return true; } void tcp_timer_persist(void *xtp) { struct tcpcb *tp = xtp; struct inpcb *inp; struct epoch_tracker et; bool progdrop; int outrv; CURVNET_SET(tp->t_vnet); #ifdef TCPDEBUG int ostate; ostate = tp->t_state; #endif inp = tp->t_inpcb; KASSERT(inp != NULL, ("%s: tp %p tp->t_inpcb == NULL", __func__, tp)); INP_WLOCK(inp); if (callout_pending(&tp->t_timers->tt_persist) || !callout_active(&tp->t_timers->tt_persist)) { INP_WUNLOCK(inp); CURVNET_RESTORE(); return; } callout_deactivate(&tp->t_timers->tt_persist); - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { INP_WUNLOCK(inp); CURVNET_RESTORE(); return; } KASSERT((tp->t_timers->tt_flags & TT_STOPPED) == 0, ("%s: tp %p tcpcb can't be stopped here", __func__, tp)); /* * Persistence timer into zero window. * Force a byte to be output, if possible. */ TCPSTAT_INC(tcps_persisttimeo); /* * Hack: if the peer is dead/unreachable, we do not * time out if the window is closed. After a full * backoff, drop the connection if the idle time * (no responses to probes) reaches the maximum * backoff that we would use if retransmitting. * Also, drop the connection if we haven't been making * progress. */ progdrop = tcp_maxunacktime_check(tp); if (progdrop || (tp->t_rxtshift == TCP_MAXRXTSHIFT && (ticks - tp->t_rcvtime >= tcp_maxpersistidle || ticks - tp->t_rcvtime >= TCP_REXMTVAL(tp) * tcp_totbackoff))) { if (!progdrop) TCPSTAT_INC(tcps_persistdrop); tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX); tcp_timer_drop(tp); CURVNET_RESTORE(); return; } /* * If the user has closed the socket then drop a persisting * connection after a much reduced timeout. */ if (tp->t_state > TCPS_CLOSE_WAIT && (ticks - tp->t_rcvtime) >= TCPTV_PERSMAX) { TCPSTAT_INC(tcps_persistdrop); tcp_log_end_status(tp, TCP_EI_STATUS_PERSIST_MAX); tcp_timer_drop(tp); CURVNET_RESTORE(); return; } tcp_setpersist(tp); tp->t_flags |= TF_FORCEDATA; NET_EPOCH_ENTER(et); outrv = tcp_output_nodrop(tp); tp->t_flags &= ~TF_FORCEDATA; #ifdef TCPDEBUG if (tp != NULL && tp->t_inpcb->inp_socket->so_options & SO_DEBUG) tcp_trace(TA_USER, ostate, tp, NULL, NULL, PRU_SLOWTIMO); #endif TCP_PROBE2(debug__user, tp, PRU_SLOWTIMO); (void) tcp_unlock_or_drop(tp, outrv); NET_EPOCH_EXIT(et); CURVNET_RESTORE(); } void tcp_timer_rexmt(void * xtp) { struct tcpcb *tp = xtp; CURVNET_SET(tp->t_vnet); int rexmt, outrv; struct inpcb *inp; struct epoch_tracker et; bool isipv6; #ifdef TCPDEBUG int ostate; ostate = tp->t_state; #endif inp = tp->t_inpcb; KASSERT(inp != NULL, ("%s: tp %p tp->t_inpcb == NULL", __func__, tp)); INP_WLOCK(inp); if (callout_pending(&tp->t_timers->tt_rexmt) || !callout_active(&tp->t_timers->tt_rexmt)) { INP_WUNLOCK(inp); CURVNET_RESTORE(); return; } callout_deactivate(&tp->t_timers->tt_rexmt); - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { INP_WUNLOCK(inp); CURVNET_RESTORE(); return; } KASSERT((tp->t_timers->tt_flags & TT_STOPPED) == 0, ("%s: tp %p tcpcb can't be stopped here", __func__, tp)); tcp_free_sackholes(tp); TCP_LOG_EVENT(tp, NULL, NULL, NULL, TCP_LOG_RTO, 0, 0, NULL, false); if (tp->t_fb->tfb_tcp_rexmit_tmr) { /* The stack has a timer action too. */ (*tp->t_fb->tfb_tcp_rexmit_tmr)(tp); } /* * Retransmission timer went off. Message has not * been acked within retransmit interval. Back off * to a longer retransmit interval and retransmit one segment. * * If we've either exceeded the maximum number of retransmissions, * or we've gone long enough without making progress, then drop * the session. */ if (++tp->t_rxtshift > TCP_MAXRXTSHIFT || tcp_maxunacktime_check(tp)) { if (tp->t_rxtshift > TCP_MAXRXTSHIFT) TCPSTAT_INC(tcps_timeoutdrop); tp->t_rxtshift = TCP_MAXRXTSHIFT; tcp_log_end_status(tp, TCP_EI_STATUS_RETRAN); tcp_timer_drop(tp); CURVNET_RESTORE(); return; } if (tp->t_state == TCPS_SYN_SENT) { /* * If the SYN was retransmitted, indicate CWND to be * limited to 1 segment in cc_conn_init(). */ tp->snd_cwnd = 1; } else if (tp->t_rxtshift == 1) { /* * first retransmit; record ssthresh and cwnd so they can * be recovered if this turns out to be a "bad" retransmit. * A retransmit is considered "bad" if an ACK for this * segment is received within RTT/2 interval; the assumption * here is that the ACK was already in flight. See * "On Estimating End-to-End Network Path Properties" by * Allman and Paxson for more details. */ tp->snd_cwnd_prev = tp->snd_cwnd; tp->snd_ssthresh_prev = tp->snd_ssthresh; tp->snd_recover_prev = tp->snd_recover; if (IN_FASTRECOVERY(tp->t_flags)) tp->t_flags |= TF_WASFRECOVERY; else tp->t_flags &= ~TF_WASFRECOVERY; if (IN_CONGRECOVERY(tp->t_flags)) tp->t_flags |= TF_WASCRECOVERY; else tp->t_flags &= ~TF_WASCRECOVERY; if ((tp->t_flags & TF_RCVD_TSTMP) == 0) tp->t_badrxtwin = ticks + (tp->t_srtt >> (TCP_RTT_SHIFT + 1)); /* In the event that we've negotiated timestamps * badrxtwin will be set to the value that we set * the retransmitted packet's to_tsval to by tcp_output */ tp->t_flags |= TF_PREVVALID; } else tp->t_flags &= ~TF_PREVVALID; TCPSTAT_INC(tcps_rexmttimeo); if ((tp->t_state == TCPS_SYN_SENT) || (tp->t_state == TCPS_SYN_RECEIVED)) rexmt = tcp_rexmit_initial * tcp_backoff[tp->t_rxtshift]; else rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift]; TCPT_RANGESET(tp->t_rxtcur, rexmt, tp->t_rttmin, TCPTV_REXMTMAX); /* * We enter the path for PLMTUD if connection is established or, if * connection is FIN_WAIT_1 status, reason for the last is that if * amount of data we send is very small, we could send it in couple of * packets and process straight to FIN. In that case we won't catch * ESTABLISHED state. */ #ifdef INET6 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? true : false; #else isipv6 = false; #endif if (((V_tcp_pmtud_blackhole_detect == 1) || (V_tcp_pmtud_blackhole_detect == 2 && !isipv6) || (V_tcp_pmtud_blackhole_detect == 3 && isipv6)) && ((tp->t_state == TCPS_ESTABLISHED) || (tp->t_state == TCPS_FIN_WAIT_1))) { if (tp->t_rxtshift == 1) { /* * We enter blackhole detection after the first * unsuccessful timer based retransmission. * Then we reduce up to two times the MSS, each * candidate giving two tries of retransmissions. * But we give a candidate only two tries, if it * actually reduces the MSS. */ tp->t_blackhole_enter = 2; tp->t_blackhole_exit = tp->t_blackhole_enter; if (isipv6) { #ifdef INET6 if (tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) tp->t_blackhole_exit += 2; if (tp->t_maxseg > V_tcp_v6mssdflt && V_tcp_v6pmtud_blackhole_mss > V_tcp_v6mssdflt) tp->t_blackhole_exit += 2; #endif } else { #ifdef INET if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss) tp->t_blackhole_exit += 2; if (tp->t_maxseg > V_tcp_mssdflt && V_tcp_pmtud_blackhole_mss > V_tcp_mssdflt) tp->t_blackhole_exit += 2; #endif } } if (((tp->t_flags2 & (TF2_PLPMTU_PMTUD|TF2_PLPMTU_MAXSEGSNT)) == (TF2_PLPMTU_PMTUD|TF2_PLPMTU_MAXSEGSNT)) && (tp->t_rxtshift >= tp->t_blackhole_enter && tp->t_rxtshift < tp->t_blackhole_exit && (tp->t_rxtshift - tp->t_blackhole_enter) % 2 == 0)) { /* * Enter Path MTU Black-hole Detection mechanism: * - Disable Path MTU Discovery (IP "DF" bit). * - Reduce MTU to lower value than what we * negotiated with peer. */ if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) == 0) { /* Record that we may have found a black hole. */ tp->t_flags2 |= TF2_PLPMTU_BLACKHOLE; /* Keep track of previous MSS. */ tp->t_pmtud_saved_maxseg = tp->t_maxseg; } /* * Reduce the MSS to blackhole value or to the default * in an attempt to retransmit. */ #ifdef INET6 if (isipv6 && tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss && V_tcp_v6pmtud_blackhole_mss > V_tcp_v6mssdflt) { /* Use the sysctl tuneable blackhole MSS. */ tp->t_maxseg = V_tcp_v6pmtud_blackhole_mss; TCPSTAT_INC(tcps_pmtud_blackhole_activated); } else if (isipv6) { /* Use the default MSS. */ tp->t_maxseg = V_tcp_v6mssdflt; /* * Disable Path MTU Discovery when we switch to * minmss. */ tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); } #endif #if defined(INET6) && defined(INET) else #endif #ifdef INET if (tp->t_maxseg > V_tcp_pmtud_blackhole_mss && V_tcp_pmtud_blackhole_mss > V_tcp_mssdflt) { /* Use the sysctl tuneable blackhole MSS. */ tp->t_maxseg = V_tcp_pmtud_blackhole_mss; TCPSTAT_INC(tcps_pmtud_blackhole_activated); } else { /* Use the default MSS. */ tp->t_maxseg = V_tcp_mssdflt; /* * Disable Path MTU Discovery when we switch to * minmss. */ tp->t_flags2 &= ~TF2_PLPMTU_PMTUD; TCPSTAT_INC(tcps_pmtud_blackhole_activated_min_mss); } #endif /* * Reset the slow-start flight size * as it may depend on the new MSS. */ if (CC_ALGO(tp)->conn_init != NULL) CC_ALGO(tp)->conn_init(tp->ccv); } else { /* * If further retransmissions are still unsuccessful * with a lowered MTU, maybe this isn't a blackhole and * we restore the previous MSS and blackhole detection * flags. */ if ((tp->t_flags2 & TF2_PLPMTU_BLACKHOLE) && (tp->t_rxtshift >= tp->t_blackhole_exit)) { tp->t_flags2 |= TF2_PLPMTU_PMTUD; tp->t_flags2 &= ~TF2_PLPMTU_BLACKHOLE; tp->t_maxseg = tp->t_pmtud_saved_maxseg; TCPSTAT_INC(tcps_pmtud_blackhole_failed); /* * Reset the slow-start flight size as it * may depend on the new MSS. */ if (CC_ALGO(tp)->conn_init != NULL) CC_ALGO(tp)->conn_init(tp->ccv); } } } /* * Disable RFC1323 and SACK if we haven't got any response to * our third SYN to work-around some broken terminal servers * (most of which have hopefully been retired) that have bad VJ * header compression code which trashes TCP segments containing * unknown-to-them TCP options. */ if (tcp_rexmit_drop_options && (tp->t_state == TCPS_SYN_SENT) && (tp->t_rxtshift == 3)) tp->t_flags &= ~(TF_REQ_SCALE|TF_REQ_TSTMP|TF_SACK_PERMIT); /* * If we backed off this far, notify the L3 protocol that we're having * connection problems. */ if (tp->t_rxtshift > TCP_RTT_INVALIDATE) { #ifdef INET6 if ((tp->t_inpcb->inp_vflag & INP_IPV6) != 0) in6_losing(tp->t_inpcb); else #endif in_losing(tp->t_inpcb); } tp->snd_nxt = tp->snd_una; tp->snd_recover = tp->snd_max; /* * Force a segment to be sent. */ tp->t_flags |= TF_ACKNOW; /* * If timing a segment in this window, stop the timer. */ tp->t_rtttime = 0; cc_cong_signal(tp, NULL, CC_RTO); NET_EPOCH_ENTER(et); outrv = tcp_output_nodrop(tp); #ifdef TCPDEBUG if (tp != NULL && (tp->t_inpcb->inp_socket->so_options & SO_DEBUG)) tcp_trace(TA_USER, ostate, tp, (void *)0, (struct tcphdr *)0, PRU_SLOWTIMO); #endif TCP_PROBE2(debug__user, tp, PRU_SLOWTIMO); (void) tcp_unlock_or_drop(tp, outrv); NET_EPOCH_EXIT(et); CURVNET_RESTORE(); } void tcp_timer_activate(struct tcpcb *tp, uint32_t timer_type, u_int delta) { struct callout *t_callout; callout_func_t *f_callout; struct inpcb *inp = tp->t_inpcb; int cpu = inp_to_cpuid(inp); #ifdef TCP_OFFLOAD if (tp->t_flags & TF_TOE) return; #endif if (tp->t_timers->tt_flags & TT_STOPPED) return; switch (timer_type) { case TT_DELACK: t_callout = &tp->t_timers->tt_delack; f_callout = tcp_timer_delack; break; case TT_REXMT: t_callout = &tp->t_timers->tt_rexmt; f_callout = tcp_timer_rexmt; break; case TT_PERSIST: t_callout = &tp->t_timers->tt_persist; f_callout = tcp_timer_persist; break; case TT_KEEP: t_callout = &tp->t_timers->tt_keep; f_callout = tcp_timer_keep; break; case TT_2MSL: t_callout = &tp->t_timers->tt_2msl; f_callout = tcp_timer_2msl; break; default: if (tp->t_fb->tfb_tcp_timer_activate) { tp->t_fb->tfb_tcp_timer_activate(tp, timer_type, delta); return; } panic("tp %p bad timer_type %#x", tp, timer_type); } if (delta == 0) { callout_stop(t_callout); } else { callout_reset_on(t_callout, delta, f_callout, tp, cpu); } } int tcp_timer_active(struct tcpcb *tp, uint32_t timer_type) { struct callout *t_callout; switch (timer_type) { case TT_DELACK: t_callout = &tp->t_timers->tt_delack; break; case TT_REXMT: t_callout = &tp->t_timers->tt_rexmt; break; case TT_PERSIST: t_callout = &tp->t_timers->tt_persist; break; case TT_KEEP: t_callout = &tp->t_timers->tt_keep; break; case TT_2MSL: t_callout = &tp->t_timers->tt_2msl; break; default: if (tp->t_fb->tfb_tcp_timer_active) { return(tp->t_fb->tfb_tcp_timer_active(tp, timer_type)); } panic("tp %p bad timer_type %#x", tp, timer_type); } return callout_active(t_callout); } /* * Stop the timer from running, and apply a flag * against the timer_flags that will force the * timer never to run. The flag is needed to assure * a race does not leave it running and cause * the timer to possibly restart itself (keep and persist * especially do this). */ int tcp_timer_suspend(struct tcpcb *tp, uint32_t timer_type) { struct callout *t_callout; uint32_t t_flags; switch (timer_type) { case TT_DELACK: t_flags = TT_DELACK_SUS; t_callout = &tp->t_timers->tt_delack; break; case TT_REXMT: t_flags = TT_REXMT_SUS; t_callout = &tp->t_timers->tt_rexmt; break; case TT_PERSIST: t_flags = TT_PERSIST_SUS; t_callout = &tp->t_timers->tt_persist; break; case TT_KEEP: t_flags = TT_KEEP_SUS; t_callout = &tp->t_timers->tt_keep; break; case TT_2MSL: t_flags = TT_2MSL_SUS; t_callout = &tp->t_timers->tt_2msl; break; default: panic("tp:%p bad timer_type 0x%x", tp, timer_type); } tp->t_timers->tt_flags |= t_flags; return (callout_stop(t_callout)); } void tcp_timers_unsuspend(struct tcpcb *tp, uint32_t timer_type) { switch (timer_type) { case TT_DELACK: if (tp->t_timers->tt_flags & TT_DELACK_SUS) { tp->t_timers->tt_flags &= ~TT_DELACK_SUS; if (tp->t_flags & TF_DELACK) { /* Delayed ack timer should be up activate a timer */ tp->t_flags &= ~TF_DELACK; tcp_timer_activate(tp, TT_DELACK, tcp_delacktime); } } break; case TT_REXMT: if (tp->t_timers->tt_flags & TT_REXMT_SUS) { tp->t_timers->tt_flags &= ~TT_REXMT_SUS; if (SEQ_GT(tp->snd_max, tp->snd_una) && (tcp_timer_active((tp), TT_PERSIST) == 0) && tp->snd_wnd) { /* We have outstanding data activate a timer */ tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); } } break; case TT_PERSIST: if (tp->t_timers->tt_flags & TT_PERSIST_SUS) { tp->t_timers->tt_flags &= ~TT_PERSIST_SUS; if (tp->snd_wnd == 0) { /* Activate the persists timer */ tp->t_rxtshift = 0; tcp_setpersist(tp); } } break; case TT_KEEP: if (tp->t_timers->tt_flags & TT_KEEP_SUS) { tp->t_timers->tt_flags &= ~TT_KEEP_SUS; tcp_timer_activate(tp, TT_KEEP, TCPS_HAVEESTABLISHED(tp->t_state) ? TP_KEEPIDLE(tp) : TP_KEEPINIT(tp)); } break; case TT_2MSL: if (tp->t_timers->tt_flags &= TT_2MSL_SUS) { tp->t_timers->tt_flags &= ~TT_2MSL_SUS; if ((tp->t_state == TCPS_FIN_WAIT_2) && ((tp->t_inpcb->inp_socket == NULL) || (tp->t_inpcb->inp_socket->so_rcv.sb_state & SBS_CANTRCVMORE))) { /* Star the 2MSL timer */ tcp_timer_activate(tp, TT_2MSL, (tcp_fast_finwait2_recycle) ? tcp_finwait2_timeout : TP_MAXIDLE(tp)); } } break; default: panic("tp:%p bad timer_type 0x%x", tp, timer_type); } } static void tcp_timer_discard(void *ptp) { struct inpcb *inp; struct tcpcb *tp; struct epoch_tracker et; tp = (struct tcpcb *)ptp; CURVNET_SET(tp->t_vnet); NET_EPOCH_ENTER(et); inp = tp->t_inpcb; KASSERT(inp != NULL, ("%s: tp %p tp->t_inpcb == NULL", __func__, tp)); INP_WLOCK(inp); KASSERT((tp->t_timers->tt_flags & TT_STOPPED) != 0, ("%s: tcpcb has to be stopped here", __func__)); if (--tp->t_timers->tt_draincnt > 0 || tcp_freecb(tp) == false) INP_WUNLOCK(inp); NET_EPOCH_EXIT(et); CURVNET_RESTORE(); } void tcp_timer_stop(struct tcpcb *tp, uint32_t timer_type) { struct callout *t_callout; tp->t_timers->tt_flags |= TT_STOPPED; switch (timer_type) { case TT_DELACK: t_callout = &tp->t_timers->tt_delack; break; case TT_REXMT: t_callout = &tp->t_timers->tt_rexmt; break; case TT_PERSIST: t_callout = &tp->t_timers->tt_persist; break; case TT_KEEP: t_callout = &tp->t_timers->tt_keep; break; case TT_2MSL: t_callout = &tp->t_timers->tt_2msl; break; default: if (tp->t_fb->tfb_tcp_timer_stop) { /* * XXXrrs we need to look at this with the * stop case below (flags). */ tp->t_fb->tfb_tcp_timer_stop(tp, timer_type); return; } panic("tp %p bad timer_type %#x", tp, timer_type); } if (callout_async_drain(t_callout, tcp_timer_discard) == 0) { /* * Can't stop the callout, defer tcpcb actual deletion * to the last one. We do this using the async drain * function and incrementing the count in */ tp->t_timers->tt_draincnt++; } } diff --git a/sys/netinet/tcp_usrreq.c b/sys/netinet/tcp_usrreq.c index 43acc0ad1719..d069c804bcc4 100644 --- a/sys/netinet/tcp_usrreq.c +++ b/sys/netinet/tcp_usrreq.c @@ -1,3190 +1,3178 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1982, 1986, 1988, 1993 * The Regents of the University of California. * Copyright (c) 2006-2007 Robert N. M. Watson * Copyright (c) 2010-2011 Juniper Networks, Inc. * All rights reserved. * * Portions of this software were developed by Robert N. M. Watson under * contract to Juniper Networks, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * From: @(#)tcp_usrreq.c 8.2 (Berkeley) 1/3/94 */ #include __FBSDID("$FreeBSD$"); #include "opt_ddb.h" #include "opt_inet.h" #include "opt_inet6.h" #include "opt_ipsec.h" #include "opt_kern_tls.h" #include "opt_tcpdebug.h" #include #include #include #include #include #include #include #include #include #include #include #ifdef INET6 #include #endif /* INET6 */ #include #include #include #include #include #include #ifdef DDB #include #endif #include #include #include #include #include #include #include #include #include #include #include #ifdef INET6 #include #include #include #include #endif #include #include #include #include #include #include #include #include #include #include #ifdef TCPPCAP #include #endif #include #ifdef TCP_OFFLOAD #include #endif #include #include #include #include #include #include #include /* * TCP protocol interface to socket abstraction. */ #ifdef INET static int tcp_connect(struct tcpcb *, struct sockaddr *, struct thread *td); #endif /* INET */ #ifdef INET6 static int tcp6_connect(struct tcpcb *, struct sockaddr *, struct thread *td); #endif /* INET6 */ static void tcp_disconnect(struct tcpcb *); static void tcp_usrclosed(struct tcpcb *); static void tcp_fill_info(struct tcpcb *, struct tcp_info *); static int tcp_pru_options_support(struct tcpcb *tp, int flags); #ifdef TCPDEBUG #define TCPDEBUG0 int ostate = 0 #define TCPDEBUG1() ostate = tp ? tp->t_state : 0 #define TCPDEBUG2(req) if (tp && (so->so_options & SO_DEBUG)) \ tcp_trace(TA_USER, ostate, tp, 0, 0, req) #else #define TCPDEBUG0 #define TCPDEBUG1() #define TCPDEBUG2(req) #endif /* * tcp_require_unique port requires a globally-unique source port for each * outgoing connection. The default is to require the 4-tuple to be unique. */ VNET_DEFINE(int, tcp_require_unique_port) = 0; SYSCTL_INT(_net_inet_tcp, OID_AUTO, require_unique_port, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_require_unique_port), 0, "Require globally-unique ephemeral port for outgoing connections"); #define V_tcp_require_unique_port VNET(tcp_require_unique_port) /* * TCP attaches to socket via pru_attach(), reserving space, * and an internet control block. */ static int tcp_usr_attach(struct socket *so, int proto, struct thread *td) { struct inpcb *inp; struct tcpcb *tp = NULL; int error; TCPDEBUG0; inp = sotoinpcb(so); KASSERT(inp == NULL, ("tcp_usr_attach: inp != NULL")); TCPDEBUG1(); error = soreserve(so, V_tcp_sendspace, V_tcp_recvspace); if (error) goto out; so->so_rcv.sb_flags |= SB_AUTOSIZE; so->so_snd.sb_flags |= SB_AUTOSIZE; error = in_pcballoc(so, &V_tcbinfo); if (error) goto out; inp = sotoinpcb(so); tp = tcp_newtcpcb(inp); if (tp == NULL) { error = ENOBUFS; in_pcbdetach(inp); in_pcbfree(inp); goto out; } tp->t_state = TCPS_CLOSED; INP_WUNLOCK(inp); TCPSTATES_INC(TCPS_CLOSED); out: TCPDEBUG2(PRU_ATTACH); TCP_PROBE2(debug__user, tp, PRU_ATTACH); return (error); } /* * tcp_usr_detach is called when the socket layer loses its final reference * to the socket, be it a file descriptor reference, a reference from TCP, * etc. At this point, there is only one case in which we will keep around * inpcb state: time wait. */ static void tcp_usr_detach(struct socket *so) { struct inpcb *inp; struct tcpcb *tp; inp = sotoinpcb(so); KASSERT(inp != NULL, ("%s: inp == NULL", __func__)); INP_WLOCK(inp); KASSERT(so->so_pcb == inp && inp->inp_socket == so, ("%s: socket %p inp %p mismatch", __func__, so, inp)); tp = intotcpcb(inp); KASSERT(inp->inp_flags & INP_DROPPED || tp->t_state < TCPS_SYN_SENT, ("%s: inp %p not dropped or embryonic", __func__, inp)); tcp_discardcb(tp); in_pcbdetach(inp); in_pcbfree(inp); } #ifdef INET /* * Give the socket an address. */ static int tcp_usr_bind(struct socket *so, struct sockaddr *nam, struct thread *td) { int error = 0; struct inpcb *inp; #ifdef KDTRACE_HOOKS struct tcpcb *tp = NULL; #endif struct sockaddr_in *sinp; sinp = (struct sockaddr_in *)nam; if (nam->sa_family != AF_INET) { /* * Preserve compatibility with old programs. */ if (nam->sa_family != AF_UNSPEC || nam->sa_len < offsetof(struct sockaddr_in, sin_zero) || sinp->sin_addr.s_addr != INADDR_ANY) return (EAFNOSUPPORT); nam->sa_family = AF_INET; } if (nam->sa_len != sizeof(*sinp)) return (EINVAL); /* * Must check for multicast addresses and disallow binding * to them. */ if (IN_MULTICAST(ntohl(sinp->sin_addr.s_addr))) return (EAFNOSUPPORT); TCPDEBUG0; inp = sotoinpcb(so); KASSERT(inp != NULL, ("tcp_usr_bind: inp == NULL")); INP_WLOCK(inp); - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { error = EINVAL; goto out; } #ifdef KDTRACE_HOOKS tp = intotcpcb(inp); #endif TCPDEBUG1(); INP_HASH_WLOCK(&V_tcbinfo); error = in_pcbbind(inp, nam, td->td_ucred); INP_HASH_WUNLOCK(&V_tcbinfo); out: TCPDEBUG2(PRU_BIND); TCP_PROBE2(debug__user, tp, PRU_BIND); INP_WUNLOCK(inp); return (error); } #endif /* INET */ #ifdef INET6 static int tcp6_usr_bind(struct socket *so, struct sockaddr *nam, struct thread *td) { int error = 0; struct inpcb *inp; #ifdef KDTRACE_HOOKS struct tcpcb *tp = NULL; #endif struct sockaddr_in6 *sin6; u_char vflagsav; sin6 = (struct sockaddr_in6 *)nam; if (nam->sa_family != AF_INET6) return (EAFNOSUPPORT); if (nam->sa_len != sizeof(*sin6)) return (EINVAL); /* * Must check for multicast addresses and disallow binding * to them. */ if (IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr)) return (EAFNOSUPPORT); TCPDEBUG0; inp = sotoinpcb(so); KASSERT(inp != NULL, ("tcp6_usr_bind: inp == NULL")); INP_WLOCK(inp); vflagsav = inp->inp_vflag; - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { error = EINVAL; goto out; } #ifdef KDTRACE_HOOKS tp = intotcpcb(inp); #endif TCPDEBUG1(); INP_HASH_WLOCK(&V_tcbinfo); inp->inp_vflag &= ~INP_IPV4; inp->inp_vflag |= INP_IPV6; #ifdef INET if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0) { if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) inp->inp_vflag |= INP_IPV4; else if (IN6_IS_ADDR_V4MAPPED(&sin6->sin6_addr)) { struct sockaddr_in sin; in6_sin6_2_sin(&sin, sin6); if (IN_MULTICAST(ntohl(sin.sin_addr.s_addr))) { error = EAFNOSUPPORT; INP_HASH_WUNLOCK(&V_tcbinfo); goto out; } inp->inp_vflag |= INP_IPV4; inp->inp_vflag &= ~INP_IPV6; error = in_pcbbind(inp, (struct sockaddr *)&sin, td->td_ucred); INP_HASH_WUNLOCK(&V_tcbinfo); goto out; } } #endif error = in6_pcbbind(inp, nam, td->td_ucred); INP_HASH_WUNLOCK(&V_tcbinfo); out: if (error != 0) inp->inp_vflag = vflagsav; TCPDEBUG2(PRU_BIND); TCP_PROBE2(debug__user, tp, PRU_BIND); INP_WUNLOCK(inp); return (error); } #endif /* INET6 */ #ifdef INET /* * Prepare to accept connections. */ static int tcp_usr_listen(struct socket *so, int backlog, struct thread *td) { int error = 0; struct inpcb *inp; struct tcpcb *tp = NULL; TCPDEBUG0; inp = sotoinpcb(so); KASSERT(inp != NULL, ("tcp_usr_listen: inp == NULL")); INP_WLOCK(inp); - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { error = EINVAL; goto out; } tp = intotcpcb(inp); TCPDEBUG1(); SOCK_LOCK(so); error = solisten_proto_check(so); if (error != 0) { SOCK_UNLOCK(so); goto out; } if (inp->inp_lport == 0) { INP_HASH_WLOCK(&V_tcbinfo); error = in_pcbbind(inp, NULL, td->td_ucred); INP_HASH_WUNLOCK(&V_tcbinfo); } if (error == 0) { tcp_state_change(tp, TCPS_LISTEN); solisten_proto(so, backlog); #ifdef TCP_OFFLOAD if ((so->so_options & SO_NO_OFFLOAD) == 0) tcp_offload_listen_start(tp); #endif } else { solisten_proto_abort(so); } SOCK_UNLOCK(so); if (IS_FASTOPEN(tp->t_flags)) tp->t_tfo_pending = tcp_fastopen_alloc_counter(); out: TCPDEBUG2(PRU_LISTEN); TCP_PROBE2(debug__user, tp, PRU_LISTEN); INP_WUNLOCK(inp); return (error); } #endif /* INET */ #ifdef INET6 static int tcp6_usr_listen(struct socket *so, int backlog, struct thread *td) { int error = 0; struct inpcb *inp; struct tcpcb *tp = NULL; u_char vflagsav; TCPDEBUG0; inp = sotoinpcb(so); KASSERT(inp != NULL, ("tcp6_usr_listen: inp == NULL")); INP_WLOCK(inp); - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { error = EINVAL; goto out; } vflagsav = inp->inp_vflag; tp = intotcpcb(inp); TCPDEBUG1(); SOCK_LOCK(so); error = solisten_proto_check(so); if (error != 0) { SOCK_UNLOCK(so); goto out; } INP_HASH_WLOCK(&V_tcbinfo); if (inp->inp_lport == 0) { inp->inp_vflag &= ~INP_IPV4; if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0) inp->inp_vflag |= INP_IPV4; error = in6_pcbbind(inp, NULL, td->td_ucred); } INP_HASH_WUNLOCK(&V_tcbinfo); if (error == 0) { tcp_state_change(tp, TCPS_LISTEN); solisten_proto(so, backlog); #ifdef TCP_OFFLOAD if ((so->so_options & SO_NO_OFFLOAD) == 0) tcp_offload_listen_start(tp); #endif } else { solisten_proto_abort(so); } SOCK_UNLOCK(so); if (IS_FASTOPEN(tp->t_flags)) tp->t_tfo_pending = tcp_fastopen_alloc_counter(); if (error != 0) inp->inp_vflag = vflagsav; out: TCPDEBUG2(PRU_LISTEN); TCP_PROBE2(debug__user, tp, PRU_LISTEN); INP_WUNLOCK(inp); return (error); } #endif /* INET6 */ #ifdef INET /* * Initiate connection to peer. * Create a template for use in transmissions on this connection. * Enter SYN_SENT state, and mark socket as connecting. * Start keep-alive timer, and seed output sequence space. * Send initial segment on connection. */ static int tcp_usr_connect(struct socket *so, struct sockaddr *nam, struct thread *td) { struct epoch_tracker et; int error = 0; struct inpcb *inp; struct tcpcb *tp = NULL; struct sockaddr_in *sinp; sinp = (struct sockaddr_in *)nam; if (nam->sa_family != AF_INET) return (EAFNOSUPPORT); if (nam->sa_len != sizeof (*sinp)) return (EINVAL); /* * Must disallow TCP ``connections'' to multicast addresses. */ if (IN_MULTICAST(ntohl(sinp->sin_addr.s_addr))) return (EAFNOSUPPORT); if (ntohl(sinp->sin_addr.s_addr) == INADDR_BROADCAST) return (EACCES); if ((error = prison_remote_ip4(td->td_ucred, &sinp->sin_addr)) != 0) return (error); TCPDEBUG0; inp = sotoinpcb(so); KASSERT(inp != NULL, ("tcp_usr_connect: inp == NULL")); INP_WLOCK(inp); - if (inp->inp_flags & INP_TIMEWAIT) { - error = EADDRINUSE; - goto out; - } if (inp->inp_flags & INP_DROPPED) { error = ECONNREFUSED; goto out; } if (SOLISTENING(so)) { error = EOPNOTSUPP; goto out; } tp = intotcpcb(inp); TCPDEBUG1(); NET_EPOCH_ENTER(et); if ((error = tcp_connect(tp, nam, td)) != 0) goto out_in_epoch; #ifdef TCP_OFFLOAD if (registered_toedevs > 0 && (so->so_options & SO_NO_OFFLOAD) == 0 && (error = tcp_offload_connect(so, nam)) == 0) goto out_in_epoch; #endif tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp)); error = tcp_output(tp); KASSERT(error >= 0, ("TCP stack %s requested tcp_drop(%p) at connect()" ", error code %d", tp->t_fb->tfb_tcp_block_name, tp, -error)); out_in_epoch: NET_EPOCH_EXIT(et); out: TCPDEBUG2(PRU_CONNECT); TCP_PROBE2(debug__user, tp, PRU_CONNECT); INP_WUNLOCK(inp); return (error); } #endif /* INET */ #ifdef INET6 static int tcp6_usr_connect(struct socket *so, struct sockaddr *nam, struct thread *td) { struct epoch_tracker et; int error = 0; struct inpcb *inp; struct tcpcb *tp = NULL; struct sockaddr_in6 *sin6; u_int8_t incflagsav; u_char vflagsav; TCPDEBUG0; sin6 = (struct sockaddr_in6 *)nam; if (nam->sa_family != AF_INET6) return (EAFNOSUPPORT); if (nam->sa_len != sizeof (*sin6)) return (EINVAL); /* * Must disallow TCP ``connections'' to multicast addresses. */ if (IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr)) return (EAFNOSUPPORT); inp = sotoinpcb(so); KASSERT(inp != NULL, ("tcp6_usr_connect: inp == NULL")); INP_WLOCK(inp); vflagsav = inp->inp_vflag; incflagsav = inp->inp_inc.inc_flags; - if (inp->inp_flags & INP_TIMEWAIT) { - error = EADDRINUSE; - goto out; - } if (inp->inp_flags & INP_DROPPED) { error = ECONNREFUSED; goto out; } if (SOLISTENING(so)) { error = EINVAL; goto out; } tp = intotcpcb(inp); TCPDEBUG1(); #ifdef INET /* * XXXRW: Some confusion: V4/V6 flags relate to binding, and * therefore probably require the hash lock, which isn't held here. * Is this a significant problem? */ if (IN6_IS_ADDR_V4MAPPED(&sin6->sin6_addr)) { struct sockaddr_in sin; if ((inp->inp_flags & IN6P_IPV6_V6ONLY) != 0) { error = EINVAL; goto out; } if ((inp->inp_vflag & INP_IPV4) == 0) { error = EAFNOSUPPORT; goto out; } in6_sin6_2_sin(&sin, sin6); if (IN_MULTICAST(ntohl(sin.sin_addr.s_addr))) { error = EAFNOSUPPORT; goto out; } if (ntohl(sin.sin_addr.s_addr) == INADDR_BROADCAST) { error = EACCES; goto out; } if ((error = prison_remote_ip4(td->td_ucred, &sin.sin_addr)) != 0) goto out; inp->inp_vflag |= INP_IPV4; inp->inp_vflag &= ~INP_IPV6; NET_EPOCH_ENTER(et); if ((error = tcp_connect(tp, (struct sockaddr *)&sin, td)) != 0) goto out_in_epoch; #ifdef TCP_OFFLOAD if (registered_toedevs > 0 && (so->so_options & SO_NO_OFFLOAD) == 0 && (error = tcp_offload_connect(so, nam)) == 0) goto out_in_epoch; #endif error = tcp_output(tp); goto out_in_epoch; } else { if ((inp->inp_vflag & INP_IPV6) == 0) { error = EAFNOSUPPORT; goto out; } } #endif if ((error = prison_remote_ip6(td->td_ucred, &sin6->sin6_addr)) != 0) goto out; inp->inp_vflag &= ~INP_IPV4; inp->inp_vflag |= INP_IPV6; inp->inp_inc.inc_flags |= INC_ISIPV6; NET_EPOCH_ENTER(et); if ((error = tcp6_connect(tp, nam, td)) != 0) goto out_in_epoch; #ifdef TCP_OFFLOAD if (registered_toedevs > 0 && (so->so_options & SO_NO_OFFLOAD) == 0 && (error = tcp_offload_connect(so, nam)) == 0) goto out_in_epoch; #endif tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp)); error = tcp_output(tp); out_in_epoch: NET_EPOCH_EXIT(et); out: KASSERT(error >= 0, ("TCP stack %s requested tcp_drop(%p) at connect()" ", error code %d", tp->t_fb->tfb_tcp_block_name, tp, -error)); /* * If the implicit bind in the connect call fails, restore * the flags we modified. */ if (error != 0 && inp->inp_lport == 0) { inp->inp_vflag = vflagsav; inp->inp_inc.inc_flags = incflagsav; } TCPDEBUG2(PRU_CONNECT); TCP_PROBE2(debug__user, tp, PRU_CONNECT); INP_WUNLOCK(inp); return (error); } #endif /* INET6 */ /* * Initiate disconnect from peer. * If connection never passed embryonic stage, just drop; * else if don't need to let data drain, then can just drop anyways, * else have to begin TCP shutdown process: mark socket disconnecting, * drain unread data, state switch to reflect user close, and * send segment (e.g. FIN) to peer. Socket will be really disconnected * when peer sends FIN and acks ours. * * SHOULD IMPLEMENT LATER PRU_CONNECT VIA REALLOC TCPCB. */ static int tcp_usr_disconnect(struct socket *so) { struct inpcb *inp; struct tcpcb *tp = NULL; struct epoch_tracker et; int error = 0; TCPDEBUG0; NET_EPOCH_ENTER(et); inp = sotoinpcb(so); KASSERT(inp != NULL, ("tcp_usr_disconnect: inp == NULL")); INP_WLOCK(inp); - if (inp->inp_flags & INP_TIMEWAIT) - goto out; if (inp->inp_flags & INP_DROPPED) { error = ECONNRESET; goto out; } tp = intotcpcb(inp); TCPDEBUG1(); tcp_disconnect(tp); out: TCPDEBUG2(PRU_DISCONNECT); TCP_PROBE2(debug__user, tp, PRU_DISCONNECT); INP_WUNLOCK(inp); NET_EPOCH_EXIT(et); return (error); } #ifdef INET /* * Accept a connection. Essentially all the work is done at higher levels; * just return the address of the peer, storing through addr. */ static int tcp_usr_accept(struct socket *so, struct sockaddr **nam) { int error = 0; struct inpcb *inp = NULL; #ifdef KDTRACE_HOOKS struct tcpcb *tp = NULL; #endif struct in_addr addr; in_port_t port = 0; TCPDEBUG0; if (so->so_state & SS_ISDISCONNECTED) return (ECONNABORTED); inp = sotoinpcb(so); KASSERT(inp != NULL, ("tcp_usr_accept: inp == NULL")); INP_WLOCK(inp); - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { error = ECONNABORTED; goto out; } #ifdef KDTRACE_HOOKS tp = intotcpcb(inp); #endif TCPDEBUG1(); /* * We inline in_getpeeraddr and COMMON_END here, so that we can * copy the data of interest and defer the malloc until after we * release the lock. */ port = inp->inp_fport; addr = inp->inp_faddr; out: TCPDEBUG2(PRU_ACCEPT); TCP_PROBE2(debug__user, tp, PRU_ACCEPT); INP_WUNLOCK(inp); if (error == 0) *nam = in_sockaddr(port, &addr); return error; } #endif /* INET */ #ifdef INET6 static int tcp6_usr_accept(struct socket *so, struct sockaddr **nam) { struct inpcb *inp = NULL; int error = 0; #ifdef KDTRACE_HOOKS struct tcpcb *tp = NULL; #endif struct in_addr addr; struct in6_addr addr6; struct epoch_tracker et; in_port_t port = 0; int v4 = 0; TCPDEBUG0; if (so->so_state & SS_ISDISCONNECTED) return (ECONNABORTED); inp = sotoinpcb(so); KASSERT(inp != NULL, ("tcp6_usr_accept: inp == NULL")); NET_EPOCH_ENTER(et); INP_WLOCK(inp); - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { error = ECONNABORTED; goto out; } #ifdef KDTRACE_HOOKS tp = intotcpcb(inp); #endif TCPDEBUG1(); /* * We inline in6_mapped_peeraddr and COMMON_END here, so that we can * copy the data of interest and defer the malloc until after we * release the lock. */ if (inp->inp_vflag & INP_IPV4) { v4 = 1; port = inp->inp_fport; addr = inp->inp_faddr; } else { port = inp->inp_fport; addr6 = inp->in6p_faddr; } out: TCPDEBUG2(PRU_ACCEPT); TCP_PROBE2(debug__user, tp, PRU_ACCEPT); INP_WUNLOCK(inp); NET_EPOCH_EXIT(et); if (error == 0) { if (v4) *nam = in6_v4mapsin6_sockaddr(port, &addr); else *nam = in6_sockaddr(port, &addr6); } return error; } #endif /* INET6 */ /* * Mark the connection as being incapable of further output. */ static int tcp_usr_shutdown(struct socket *so) { int error = 0; struct inpcb *inp; struct tcpcb *tp = NULL; struct epoch_tracker et; TCPDEBUG0; inp = sotoinpcb(so); KASSERT(inp != NULL, ("inp == NULL")); INP_WLOCK(inp); - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { INP_WUNLOCK(inp); return (ECONNRESET); } tp = intotcpcb(inp); NET_EPOCH_ENTER(et); TCPDEBUG1(); socantsendmore(so); tcp_usrclosed(tp); if (!(inp->inp_flags & INP_DROPPED)) error = tcp_output_nodrop(tp); TCPDEBUG2(PRU_SHUTDOWN); TCP_PROBE2(debug__user, tp, PRU_SHUTDOWN); error = tcp_unlock_or_drop(tp, error); NET_EPOCH_EXIT(et); return (error); } /* * After a receive, possibly send window update to peer. */ static int tcp_usr_rcvd(struct socket *so, int flags) { struct epoch_tracker et; struct inpcb *inp; struct tcpcb *tp = NULL; int outrv = 0, error = 0; TCPDEBUG0; inp = sotoinpcb(so); KASSERT(inp != NULL, ("tcp_usr_rcvd: inp == NULL")); INP_WLOCK(inp); - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { INP_WUNLOCK(inp); return (ECONNRESET); } tp = intotcpcb(inp); NET_EPOCH_ENTER(et); TCPDEBUG1(); /* * For passively-created TFO connections, don't attempt a window * update while still in SYN_RECEIVED as this may trigger an early * SYN|ACK. It is preferable to have the SYN|ACK be sent along with * application response data, or failing that, when the DELACK timer * expires. */ if (IS_FASTOPEN(tp->t_flags) && (tp->t_state == TCPS_SYN_RECEIVED)) goto out; #ifdef TCP_OFFLOAD if (tp->t_flags & TF_TOE) tcp_offload_rcvd(tp); else #endif outrv = tcp_output_nodrop(tp); out: TCPDEBUG2(PRU_RCVD); TCP_PROBE2(debug__user, tp, PRU_RCVD); (void) tcp_unlock_or_drop(tp, outrv); NET_EPOCH_EXIT(et); return (error); } /* * Do a send by putting data in output queue and updating urgent * marker if URG set. Possibly send more data. Unlike the other * pru_*() routines, the mbuf chains are our responsibility. We * must either enqueue them or free them. The other pru_* routines * generally are caller-frees. */ static int tcp_usr_send(struct socket *so, int flags, struct mbuf *m, struct sockaddr *nam, struct mbuf *control, struct thread *td) { struct epoch_tracker et; int error = 0; struct inpcb *inp; struct tcpcb *tp = NULL; #ifdef INET #ifdef INET6 struct sockaddr_in sin; #endif struct sockaddr_in *sinp; #endif #ifdef INET6 int isipv6; #endif u_int8_t incflagsav; u_char vflagsav; bool restoreflags; TCPDEBUG0; if (control != NULL) { /* TCP doesn't do control messages (rights, creds, etc) */ if (control->m_len) { m_freem(control); return (EINVAL); } m_freem(control); /* empty control, just free it */ } inp = sotoinpcb(so); KASSERT(inp != NULL, ("tcp_usr_send: inp == NULL")); INP_WLOCK(inp); - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { if (m != NULL && (flags & PRUS_NOTREADY) == 0) m_freem(m); INP_WUNLOCK(inp); return (ECONNRESET); } vflagsav = inp->inp_vflag; incflagsav = inp->inp_inc.inc_flags; restoreflags = false; tp = intotcpcb(inp); NET_EPOCH_ENTER(et); if ((flags & PRUS_OOB) != 0 && (error = tcp_pru_options_support(tp, PRUS_OOB)) != 0) goto out; TCPDEBUG1(); if (nam != NULL && tp->t_state < TCPS_SYN_SENT) { if (tp->t_state == TCPS_LISTEN) { error = EINVAL; goto out; } switch (nam->sa_family) { #ifdef INET case AF_INET: sinp = (struct sockaddr_in *)nam; if (sinp->sin_len != sizeof(struct sockaddr_in)) { error = EINVAL; goto out; } if ((inp->inp_vflag & INP_IPV6) != 0) { error = EAFNOSUPPORT; goto out; } if (IN_MULTICAST(ntohl(sinp->sin_addr.s_addr))) { error = EAFNOSUPPORT; goto out; } if (ntohl(sinp->sin_addr.s_addr) == INADDR_BROADCAST) { error = EACCES; goto out; } if ((error = prison_remote_ip4(td->td_ucred, &sinp->sin_addr))) goto out; #ifdef INET6 isipv6 = 0; #endif break; #endif /* INET */ #ifdef INET6 case AF_INET6: { struct sockaddr_in6 *sin6; sin6 = (struct sockaddr_in6 *)nam; if (sin6->sin6_len != sizeof(*sin6)) { error = EINVAL; goto out; } if ((inp->inp_vflag & INP_IPV6PROTO) == 0) { error = EAFNOSUPPORT; goto out; } if (IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr)) { error = EAFNOSUPPORT; goto out; } if (IN6_IS_ADDR_V4MAPPED(&sin6->sin6_addr)) { #ifdef INET if ((inp->inp_flags & IN6P_IPV6_V6ONLY) != 0) { error = EINVAL; goto out; } if ((inp->inp_vflag & INP_IPV4) == 0) { error = EAFNOSUPPORT; goto out; } restoreflags = true; inp->inp_vflag &= ~INP_IPV6; sinp = &sin; in6_sin6_2_sin(sinp, sin6); if (IN_MULTICAST( ntohl(sinp->sin_addr.s_addr))) { error = EAFNOSUPPORT; goto out; } if ((error = prison_remote_ip4(td->td_ucred, &sinp->sin_addr))) goto out; isipv6 = 0; #else /* !INET */ error = EAFNOSUPPORT; goto out; #endif /* INET */ } else { if ((inp->inp_vflag & INP_IPV6) == 0) { error = EAFNOSUPPORT; goto out; } restoreflags = true; inp->inp_vflag &= ~INP_IPV4; inp->inp_inc.inc_flags |= INC_ISIPV6; if ((error = prison_remote_ip6(td->td_ucred, &sin6->sin6_addr))) goto out; isipv6 = 1; } break; } #endif /* INET6 */ default: error = EAFNOSUPPORT; goto out; } } if (!(flags & PRUS_OOB)) { if (tp->t_acktime == 0) tp->t_acktime = ticks; sbappendstream(&so->so_snd, m, flags); m = NULL; if (nam && tp->t_state < TCPS_SYN_SENT) { KASSERT(tp->t_state == TCPS_CLOSED, ("%s: tp %p is listening", __func__, tp)); /* * Do implied connect if not yet connected, * initialize window to default value, and * initialize maxseg using peer's cached MSS. */ #ifdef INET6 if (isipv6) error = tcp6_connect(tp, nam, td); #endif /* INET6 */ #if defined(INET6) && defined(INET) else #endif #ifdef INET error = tcp_connect(tp, (struct sockaddr *)sinp, td); #endif /* * The bind operation in tcp_connect succeeded. We * no longer want to restore the flags if later * operations fail. */ if (error == 0 || inp->inp_lport != 0) restoreflags = false; if (error) { /* m is freed if PRUS_NOTREADY is unset. */ sbflush(&so->so_snd); goto out; } if (IS_FASTOPEN(tp->t_flags)) tcp_fastopen_connect(tp); else { tp->snd_wnd = TTCP_CLIENT_SND_WND; tcp_mss(tp, -1); } } if (flags & PRUS_EOF) { /* * Close the send side of the connection after * the data is sent. */ socantsendmore(so); tcp_usrclosed(tp); } if (TCPS_HAVEESTABLISHED(tp->t_state) && ((tp->t_flags2 & TF2_FBYTES_COMPLETE) == 0) && (tp->t_fbyte_out == 0) && (so->so_snd.sb_ccc > 0)) { tp->t_fbyte_out = ticks; if (tp->t_fbyte_out == 0) tp->t_fbyte_out = 1; if (tp->t_fbyte_out && tp->t_fbyte_in) tp->t_flags2 |= TF2_FBYTES_COMPLETE; } if (!(inp->inp_flags & INP_DROPPED) && !(flags & PRUS_NOTREADY)) { if (flags & PRUS_MORETOCOME) tp->t_flags |= TF_MORETOCOME; error = tcp_output_nodrop(tp); if (flags & PRUS_MORETOCOME) tp->t_flags &= ~TF_MORETOCOME; } } else { /* * XXXRW: PRUS_EOF not implemented with PRUS_OOB? */ SOCKBUF_LOCK(&so->so_snd); if (sbspace(&so->so_snd) < -512) { SOCKBUF_UNLOCK(&so->so_snd); error = ENOBUFS; goto out; } /* * According to RFC961 (Assigned Protocols), * the urgent pointer points to the last octet * of urgent data. We continue, however, * to consider it to indicate the first octet * of data past the urgent section. * Otherwise, snd_up should be one lower. */ if (tp->t_acktime == 0) tp->t_acktime = ticks; sbappendstream_locked(&so->so_snd, m, flags); SOCKBUF_UNLOCK(&so->so_snd); m = NULL; if (nam && tp->t_state < TCPS_SYN_SENT) { /* * Do implied connect if not yet connected, * initialize window to default value, and * initialize maxseg using peer's cached MSS. */ /* * Not going to contemplate SYN|URG */ if (IS_FASTOPEN(tp->t_flags)) tp->t_flags &= ~TF_FASTOPEN; #ifdef INET6 if (isipv6) error = tcp6_connect(tp, nam, td); #endif /* INET6 */ #if defined(INET6) && defined(INET) else #endif #ifdef INET error = tcp_connect(tp, (struct sockaddr *)sinp, td); #endif /* * The bind operation in tcp_connect succeeded. We * no longer want to restore the flags if later * operations fail. */ if (error == 0 || inp->inp_lport != 0) restoreflags = false; if (error != 0) { /* m is freed if PRUS_NOTREADY is unset. */ sbflush(&so->so_snd); goto out; } tp->snd_wnd = TTCP_CLIENT_SND_WND; tcp_mss(tp, -1); } tp->snd_up = tp->snd_una + sbavail(&so->so_snd); if ((flags & PRUS_NOTREADY) == 0) { tp->t_flags |= TF_FORCEDATA; error = tcp_output_nodrop(tp); tp->t_flags &= ~TF_FORCEDATA; } } TCP_LOG_EVENT(tp, NULL, &inp->inp_socket->so_rcv, &inp->inp_socket->so_snd, TCP_LOG_USERSEND, error, 0, NULL, false); out: /* * In case of PRUS_NOTREADY, the caller or tcp_usr_ready() is * responsible for freeing memory. */ if (m != NULL && (flags & PRUS_NOTREADY) == 0) m_freem(m); /* * If the request was unsuccessful and we changed flags, * restore the original flags. */ if (error != 0 && restoreflags) { inp->inp_vflag = vflagsav; inp->inp_inc.inc_flags = incflagsav; } TCPDEBUG2((flags & PRUS_OOB) ? PRU_SENDOOB : ((flags & PRUS_EOF) ? PRU_SEND_EOF : PRU_SEND)); TCP_PROBE2(debug__user, tp, (flags & PRUS_OOB) ? PRU_SENDOOB : ((flags & PRUS_EOF) ? PRU_SEND_EOF : PRU_SEND)); error = tcp_unlock_or_drop(tp, error); NET_EPOCH_EXIT(et); return (error); } static int tcp_usr_ready(struct socket *so, struct mbuf *m, int count) { struct epoch_tracker et; struct inpcb *inp; struct tcpcb *tp; int error; inp = sotoinpcb(so); INP_WLOCK(inp); - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { INP_WUNLOCK(inp); mb_free_notready(m, count); return (ECONNRESET); } tp = intotcpcb(inp); SOCKBUF_LOCK(&so->so_snd); error = sbready(&so->so_snd, m, count); SOCKBUF_UNLOCK(&so->so_snd); if (error) { INP_WUNLOCK(inp); return (error); } NET_EPOCH_ENTER(et); error = tcp_output_unlock(tp); NET_EPOCH_EXIT(et); return (error); } /* * Abort the TCP. Drop the connection abruptly. */ static void tcp_usr_abort(struct socket *so) { struct inpcb *inp; struct tcpcb *tp = NULL; struct epoch_tracker et; TCPDEBUG0; inp = sotoinpcb(so); KASSERT(inp != NULL, ("tcp_usr_abort: inp == NULL")); NET_EPOCH_ENTER(et); INP_WLOCK(inp); KASSERT(inp->inp_socket != NULL, ("tcp_usr_abort: inp_socket == NULL")); /* * If we still have full TCP state, and we're not dropped, drop. */ - if (!(inp->inp_flags & INP_TIMEWAIT) && - !(inp->inp_flags & INP_DROPPED)) { + if (!(inp->inp_flags & INP_DROPPED)) { tp = intotcpcb(inp); TCPDEBUG1(); tp = tcp_drop(tp, ECONNABORTED); if (tp == NULL) goto dropped; TCPDEBUG2(PRU_ABORT); TCP_PROBE2(debug__user, tp, PRU_ABORT); } if (!(inp->inp_flags & INP_DROPPED)) { soref(so); inp->inp_flags |= INP_SOCKREF; } INP_WUNLOCK(inp); dropped: NET_EPOCH_EXIT(et); } /* * TCP socket is closed. Start friendly disconnect. */ static void tcp_usr_close(struct socket *so) { struct inpcb *inp; struct tcpcb *tp = NULL; struct epoch_tracker et; TCPDEBUG0; inp = sotoinpcb(so); KASSERT(inp != NULL, ("tcp_usr_close: inp == NULL")); NET_EPOCH_ENTER(et); INP_WLOCK(inp); KASSERT(inp->inp_socket != NULL, ("tcp_usr_close: inp_socket == NULL")); /* * If we still have full TCP state, and we're not dropped, initiate * a disconnect. */ - if (!(inp->inp_flags & INP_TIMEWAIT) && - !(inp->inp_flags & INP_DROPPED)) { + if (!(inp->inp_flags & INP_DROPPED)) { tp = intotcpcb(inp); tp->t_flags |= TF_CLOSED; TCPDEBUG1(); tcp_disconnect(tp); TCPDEBUG2(PRU_CLOSE); TCP_PROBE2(debug__user, tp, PRU_CLOSE); } if (!(inp->inp_flags & INP_DROPPED)) { soref(so); inp->inp_flags |= INP_SOCKREF; } INP_WUNLOCK(inp); NET_EPOCH_EXIT(et); } static int tcp_pru_options_support(struct tcpcb *tp, int flags) { /* * If the specific TCP stack has a pru_options * specified then it does not always support * all the PRU_XX options and we must ask it. * If the function is not specified then all * of the PRU_XX options are supported. */ int ret = 0; if (tp->t_fb->tfb_pru_options) { ret = (*tp->t_fb->tfb_pru_options)(tp, flags); } return (ret); } /* * Receive out-of-band data. */ static int tcp_usr_rcvoob(struct socket *so, struct mbuf *m, int flags) { int error = 0; struct inpcb *inp; struct tcpcb *tp = NULL; TCPDEBUG0; inp = sotoinpcb(so); KASSERT(inp != NULL, ("tcp_usr_rcvoob: inp == NULL")); INP_WLOCK(inp); - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { error = ECONNRESET; goto out; } tp = intotcpcb(inp); error = tcp_pru_options_support(tp, PRUS_OOB); if (error) { goto out; } TCPDEBUG1(); if ((so->so_oobmark == 0 && (so->so_rcv.sb_state & SBS_RCVATMARK) == 0) || so->so_options & SO_OOBINLINE || tp->t_oobflags & TCPOOB_HADDATA) { error = EINVAL; goto out; } if ((tp->t_oobflags & TCPOOB_HAVEDATA) == 0) { error = EWOULDBLOCK; goto out; } m->m_len = 1; *mtod(m, caddr_t) = tp->t_iobc; if ((flags & MSG_PEEK) == 0) tp->t_oobflags ^= (TCPOOB_HAVEDATA | TCPOOB_HADDATA); out: TCPDEBUG2(PRU_RCVOOB); TCP_PROBE2(debug__user, tp, PRU_RCVOOB); INP_WUNLOCK(inp); return (error); } #ifdef INET struct protosw tcp_protosw = { .pr_type = SOCK_STREAM, .pr_protocol = IPPROTO_TCP, .pr_flags = PR_CONNREQUIRED | PR_IMPLOPCL | PR_WANTRCVD | PR_CAPATTACH, .pr_ctloutput = tcp_ctloutput, .pr_abort = tcp_usr_abort, .pr_accept = tcp_usr_accept, .pr_attach = tcp_usr_attach, .pr_bind = tcp_usr_bind, .pr_connect = tcp_usr_connect, .pr_control = in_control, .pr_detach = tcp_usr_detach, .pr_disconnect = tcp_usr_disconnect, .pr_listen = tcp_usr_listen, .pr_peeraddr = in_getpeeraddr, .pr_rcvd = tcp_usr_rcvd, .pr_rcvoob = tcp_usr_rcvoob, .pr_send = tcp_usr_send, .pr_ready = tcp_usr_ready, .pr_shutdown = tcp_usr_shutdown, .pr_sockaddr = in_getsockaddr, .pr_sosetlabel = in_pcbsosetlabel, .pr_close = tcp_usr_close, }; #endif /* INET */ #ifdef INET6 struct protosw tcp6_protosw = { .pr_type = SOCK_STREAM, .pr_protocol = IPPROTO_TCP, .pr_flags = PR_CONNREQUIRED | PR_IMPLOPCL |PR_WANTRCVD | PR_CAPATTACH, .pr_ctloutput = tcp_ctloutput, .pr_abort = tcp_usr_abort, .pr_accept = tcp6_usr_accept, .pr_attach = tcp_usr_attach, .pr_bind = tcp6_usr_bind, .pr_connect = tcp6_usr_connect, .pr_control = in6_control, .pr_detach = tcp_usr_detach, .pr_disconnect = tcp_usr_disconnect, .pr_listen = tcp6_usr_listen, .pr_peeraddr = in6_mapped_peeraddr, .pr_rcvd = tcp_usr_rcvd, .pr_rcvoob = tcp_usr_rcvoob, .pr_send = tcp_usr_send, .pr_ready = tcp_usr_ready, .pr_shutdown = tcp_usr_shutdown, .pr_sockaddr = in6_mapped_sockaddr, .pr_sosetlabel = in_pcbsosetlabel, .pr_close = tcp_usr_close, }; #endif /* INET6 */ #ifdef INET /* * Common subroutine to open a TCP connection to remote host specified * by struct sockaddr_in in mbuf *nam. Call in_pcbbind to assign a local * port number if needed. Call in_pcbconnect_setup to do the routing and * to choose a local host address (interface). If there is an existing * incarnation of the same connection in TIME-WAIT state and if the remote * host was sending CC options and if the connection duration was < MSL, then * truncate the previous TIME-WAIT state and proceed. * Initialize connection parameters and enter SYN-SENT state. */ static int tcp_connect(struct tcpcb *tp, struct sockaddr *nam, struct thread *td) { struct inpcb *inp = tp->t_inpcb, *oinp; struct socket *so = inp->inp_socket; struct in_addr laddr; u_short lport; int error; NET_EPOCH_ASSERT(); INP_WLOCK_ASSERT(inp); INP_HASH_WLOCK(&V_tcbinfo); if (V_tcp_require_unique_port && inp->inp_lport == 0) { error = in_pcbbind(inp, (struct sockaddr *)0, td->td_ucred); if (error) goto out; } /* * Cannot simply call in_pcbconnect, because there might be an * earlier incarnation of this same connection still in * TIME_WAIT state, creating an ADDRINUSE error. */ laddr = inp->inp_laddr; lport = inp->inp_lport; error = in_pcbconnect_setup(inp, nam, &laddr.s_addr, &lport, &inp->inp_faddr.s_addr, &inp->inp_fport, &oinp, td->td_ucred); if (error && oinp == NULL) goto out; if (oinp) { error = EADDRINUSE; goto out; } /* Handle initial bind if it hadn't been done in advance. */ if (inp->inp_lport == 0) { inp->inp_lport = lport; if (in_pcbinshash(inp) != 0) { inp->inp_lport = 0; error = EAGAIN; goto out; } } inp->inp_laddr = laddr; in_pcbrehash(inp); INP_HASH_WUNLOCK(&V_tcbinfo); /* * Compute window scaling to request: * Scale to fit into sweet spot. See tcp_syncache.c. * XXX: This should move to tcp_output(). */ while (tp->request_r_scale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << tp->request_r_scale) < sb_max) tp->request_r_scale++; soisconnecting(so); TCPSTAT_INC(tcps_connattempt); tcp_state_change(tp, TCPS_SYN_SENT); tp->iss = tcp_new_isn(&inp->inp_inc); if (tp->t_flags & TF_REQ_TSTMP) tp->ts_offset = tcp_new_ts_offset(&inp->inp_inc); tcp_sendseqinit(tp); return 0; out: INP_HASH_WUNLOCK(&V_tcbinfo); return (error); } #endif /* INET */ #ifdef INET6 static int tcp6_connect(struct tcpcb *tp, struct sockaddr *nam, struct thread *td) { struct inpcb *inp = tp->t_inpcb; int error; INP_WLOCK_ASSERT(inp); INP_HASH_WLOCK(&V_tcbinfo); if (V_tcp_require_unique_port && inp->inp_lport == 0) { error = in6_pcbbind(inp, (struct sockaddr *)0, td->td_ucred); if (error) goto out; } error = in6_pcbconnect(inp, nam, td->td_ucred); if (error != 0) goto out; INP_HASH_WUNLOCK(&V_tcbinfo); /* Compute window scaling to request. */ while (tp->request_r_scale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << tp->request_r_scale) < sb_max) tp->request_r_scale++; soisconnecting(inp->inp_socket); TCPSTAT_INC(tcps_connattempt); tcp_state_change(tp, TCPS_SYN_SENT); tp->iss = tcp_new_isn(&inp->inp_inc); if (tp->t_flags & TF_REQ_TSTMP) tp->ts_offset = tcp_new_ts_offset(&inp->inp_inc); tcp_sendseqinit(tp); return 0; out: INP_HASH_WUNLOCK(&V_tcbinfo); return error; } #endif /* INET6 */ /* * Export TCP internal state information via a struct tcp_info, based on the * Linux 2.6 API. Not ABI compatible as our constants are mapped differently * (TCP state machine, etc). We export all information using FreeBSD-native * constants -- for example, the numeric values for tcpi_state will differ * from Linux. */ static void tcp_fill_info(struct tcpcb *tp, struct tcp_info *ti) { INP_WLOCK_ASSERT(tp->t_inpcb); bzero(ti, sizeof(*ti)); ti->tcpi_state = tp->t_state; if ((tp->t_flags & TF_REQ_TSTMP) && (tp->t_flags & TF_RCVD_TSTMP)) ti->tcpi_options |= TCPI_OPT_TIMESTAMPS; if (tp->t_flags & TF_SACK_PERMIT) ti->tcpi_options |= TCPI_OPT_SACK; if ((tp->t_flags & TF_REQ_SCALE) && (tp->t_flags & TF_RCVD_SCALE)) { ti->tcpi_options |= TCPI_OPT_WSCALE; ti->tcpi_snd_wscale = tp->snd_scale; ti->tcpi_rcv_wscale = tp->rcv_scale; } if (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT)) ti->tcpi_options |= TCPI_OPT_ECN; ti->tcpi_rto = tp->t_rxtcur * tick; ti->tcpi_last_data_recv = ((uint32_t)ticks - tp->t_rcvtime) * tick; ti->tcpi_rtt = ((u_int64_t)tp->t_srtt * tick) >> TCP_RTT_SHIFT; ti->tcpi_rttvar = ((u_int64_t)tp->t_rttvar * tick) >> TCP_RTTVAR_SHIFT; ti->tcpi_snd_ssthresh = tp->snd_ssthresh; ti->tcpi_snd_cwnd = tp->snd_cwnd; /* * FreeBSD-specific extension fields for tcp_info. */ ti->tcpi_rcv_space = tp->rcv_wnd; ti->tcpi_rcv_nxt = tp->rcv_nxt; ti->tcpi_snd_wnd = tp->snd_wnd; ti->tcpi_snd_bwnd = 0; /* Unused, kept for compat. */ ti->tcpi_snd_nxt = tp->snd_nxt; ti->tcpi_snd_mss = tp->t_maxseg; ti->tcpi_rcv_mss = tp->t_maxseg; ti->tcpi_snd_rexmitpack = tp->t_sndrexmitpack; ti->tcpi_rcv_ooopack = tp->t_rcvoopack; ti->tcpi_snd_zerowin = tp->t_sndzerowin; #ifdef TCP_OFFLOAD if (tp->t_flags & TF_TOE) { ti->tcpi_options |= TCPI_OPT_TOE; tcp_offload_tcp_info(tp, ti); } #endif } /* * tcp_ctloutput() must drop the inpcb lock before performing copyin on * socket option arguments. When it re-acquires the lock after the copy, it * has to revalidate that the connection is still valid for the socket * option. */ #define INP_WLOCK_RECHECK_CLEANUP(inp, cleanup) do { \ INP_WLOCK(inp); \ - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { \ + if (inp->inp_flags & INP_DROPPED) { \ INP_WUNLOCK(inp); \ cleanup; \ return (ECONNRESET); \ } \ tp = intotcpcb(inp); \ } while(0) #define INP_WLOCK_RECHECK(inp) INP_WLOCK_RECHECK_CLEANUP((inp), /* noop */) int tcp_ctloutput_set(struct inpcb *inp, struct sockopt *sopt) { struct socket *so = inp->inp_socket; struct tcpcb *tp = intotcpcb(inp); int error = 0; MPASS(sopt->sopt_dir == SOPT_SET); INP_WLOCK_ASSERT(inp); - KASSERT((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) == 0, + KASSERT((inp->inp_flags & INP_DROPPED) == 0, ("inp_flags == %x", inp->inp_flags)); KASSERT(so != NULL, ("inp_socket == NULL")); if (sopt->sopt_level != IPPROTO_TCP) { INP_WUNLOCK(inp); #ifdef INET6 if (inp->inp_vflag & INP_IPV6PROTO) error = ip6_ctloutput(so, sopt); #endif #if defined(INET6) && defined(INET) else #endif #ifdef INET error = ip_ctloutput(so, sopt); #endif /* * When an IP-level socket option affects TCP, pass control * down to stack tfb_tcp_ctloutput, otherwise return what * IP level returned. */ switch (sopt->sopt_level) { #ifdef INET6 case IPPROTO_IPV6: if ((inp->inp_vflag & INP_IPV6PROTO) == 0) return (error); switch (sopt->sopt_name) { case IPV6_TCLASS: /* Notify tcp stacks that care (e.g. RACK). */ break; case IPV6_USE_MIN_MTU: /* Update t_maxseg accordingly. */ break; default: return (error); } break; #endif #ifdef INET case IPPROTO_IP: switch (sopt->sopt_name) { case IP_TOS: inp->inp_ip_tos &= ~IPTOS_ECN_MASK; break; case IP_TTL: /* Notify tcp stacks that care (e.g. RACK). */ break; default: return (error); } break; #endif default: return (error); } INP_WLOCK(inp); - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { INP_WUNLOCK(inp); return (ECONNRESET); } } else if (sopt->sopt_name == TCP_FUNCTION_BLK) { /* * Protect the TCP option TCP_FUNCTION_BLK so * that a sub-function can *never* overwrite this. */ struct tcp_function_set fsn; struct tcp_function_block *blk; INP_WUNLOCK(inp); error = sooptcopyin(sopt, &fsn, sizeof fsn, sizeof fsn); if (error) return (error); INP_WLOCK(inp); - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { INP_WUNLOCK(inp); return (ECONNRESET); } tp = intotcpcb(inp); blk = find_and_ref_tcp_functions(&fsn); if (blk == NULL) { INP_WUNLOCK(inp); return (ENOENT); } if (tp->t_fb == blk) { /* You already have this */ refcount_release(&blk->tfb_refcnt); INP_WUNLOCK(inp); return (0); } if (tp->t_state != TCPS_CLOSED) { /* * The user has advanced the state * past the initial point, we may not * be able to switch. */ if (blk->tfb_tcp_handoff_ok != NULL) { /* * Does the stack provide a * query mechanism, if so it may * still be possible? */ error = (*blk->tfb_tcp_handoff_ok)(tp); } else error = EINVAL; if (error) { refcount_release(&blk->tfb_refcnt); INP_WUNLOCK(inp); return(error); } } if (blk->tfb_flags & TCP_FUNC_BEING_REMOVED) { refcount_release(&blk->tfb_refcnt); INP_WUNLOCK(inp); return (ENOENT); } /* * Release the old refcnt, the * lookup acquired a ref on the * new one already. */ if (tp->t_fb->tfb_tcp_fb_fini) { struct epoch_tracker et; /* * Tell the stack to cleanup with 0 i.e. * the tcb is not going away. */ NET_EPOCH_ENTER(et); (*tp->t_fb->tfb_tcp_fb_fini)(tp, 0); NET_EPOCH_EXIT(et); } #ifdef TCPHPTS /* Assure that we are not on any hpts */ tcp_hpts_remove(tp->t_inpcb); #endif if (blk->tfb_tcp_fb_init) { error = (*blk->tfb_tcp_fb_init)(tp); if (error) { refcount_release(&blk->tfb_refcnt); if (tp->t_fb->tfb_tcp_fb_init) { if((*tp->t_fb->tfb_tcp_fb_init)(tp) != 0) { /* Fall back failed, drop the connection */ INP_WUNLOCK(inp); soabort(so); return (error); } } goto err_out; } } refcount_release(&tp->t_fb->tfb_refcnt); tp->t_fb = blk; #ifdef TCP_OFFLOAD if (tp->t_flags & TF_TOE) { tcp_offload_ctloutput(tp, sopt->sopt_dir, sopt->sopt_name); } #endif err_out: INP_WUNLOCK(inp); return (error); } /* Pass in the INP locked, callee must unlock it. */ return (tp->t_fb->tfb_tcp_ctloutput(inp, sopt)); } static int tcp_ctloutput_get(struct inpcb *inp, struct sockopt *sopt) { struct socket *so = inp->inp_socket; struct tcpcb *tp = intotcpcb(inp); int error = 0; MPASS(sopt->sopt_dir == SOPT_GET); INP_WLOCK_ASSERT(inp); - KASSERT((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) == 0, + KASSERT((inp->inp_flags & INP_DROPPED) == 0, ("inp_flags == %x", inp->inp_flags)); KASSERT(so != NULL, ("inp_socket == NULL")); if (sopt->sopt_level != IPPROTO_TCP) { INP_WUNLOCK(inp); #ifdef INET6 if (inp->inp_vflag & INP_IPV6PROTO) error = ip6_ctloutput(so, sopt); #endif /* INET6 */ #if defined(INET6) && defined(INET) else #endif #ifdef INET error = ip_ctloutput(so, sopt); #endif return (error); } if (((sopt->sopt_name == TCP_FUNCTION_BLK) || (sopt->sopt_name == TCP_FUNCTION_ALIAS))) { struct tcp_function_set fsn; if (sopt->sopt_name == TCP_FUNCTION_ALIAS) { memset(&fsn, 0, sizeof(fsn)); find_tcp_function_alias(tp->t_fb, &fsn); } else { strncpy(fsn.function_set_name, tp->t_fb->tfb_tcp_block_name, TCP_FUNCTION_NAME_LEN_MAX); fsn.function_set_name[TCP_FUNCTION_NAME_LEN_MAX - 1] = '\0'; } fsn.pcbcnt = tp->t_fb->tfb_refcnt; INP_WUNLOCK(inp); error = sooptcopyout(sopt, &fsn, sizeof fsn); return (error); } /* Pass in the INP locked, callee must unlock it. */ return (tp->t_fb->tfb_tcp_ctloutput(inp, sopt)); } int tcp_ctloutput(struct socket *so, struct sockopt *sopt) { struct inpcb *inp; inp = sotoinpcb(so); KASSERT(inp != NULL, ("tcp_ctloutput: inp == NULL")); INP_WLOCK(inp); - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { INP_WUNLOCK(inp); return (ECONNRESET); } if (sopt->sopt_dir == SOPT_SET) return (tcp_ctloutput_set(inp, sopt)); else if (sopt->sopt_dir == SOPT_GET) return (tcp_ctloutput_get(inp, sopt)); else panic("%s: sopt_dir $%d", __func__, sopt->sopt_dir); } /* * If this assert becomes untrue, we need to change the size of the buf * variable in tcp_default_ctloutput(). */ #ifdef CTASSERT CTASSERT(TCP_CA_NAME_MAX <= TCP_LOG_ID_LEN); CTASSERT(TCP_LOG_REASON_LEN <= TCP_LOG_ID_LEN); #endif #ifdef KERN_TLS static int copyin_tls_enable(struct sockopt *sopt, struct tls_enable *tls) { struct tls_enable_v0 tls_v0; int error; if (sopt->sopt_valsize == sizeof(tls_v0)) { error = sooptcopyin(sopt, &tls_v0, sizeof(tls_v0), sizeof(tls_v0)); if (error) return (error); memset(tls, 0, sizeof(*tls)); tls->cipher_key = tls_v0.cipher_key; tls->iv = tls_v0.iv; tls->auth_key = tls_v0.auth_key; tls->cipher_algorithm = tls_v0.cipher_algorithm; tls->cipher_key_len = tls_v0.cipher_key_len; tls->iv_len = tls_v0.iv_len; tls->auth_algorithm = tls_v0.auth_algorithm; tls->auth_key_len = tls_v0.auth_key_len; tls->flags = tls_v0.flags; tls->tls_vmajor = tls_v0.tls_vmajor; tls->tls_vminor = tls_v0.tls_vminor; return (0); } return (sooptcopyin(sopt, tls, sizeof(*tls), sizeof(*tls))); } #endif extern struct cc_algo newreno_cc_algo; static int tcp_set_cc_mod(struct inpcb *inp, struct sockopt *sopt) { struct cc_algo *algo; void *ptr = NULL; struct tcpcb *tp; struct cc_var cc_mem; char buf[TCP_CA_NAME_MAX]; size_t mem_sz; int error; INP_WUNLOCK(inp); error = sooptcopyin(sopt, buf, TCP_CA_NAME_MAX - 1, 1); if (error) return(error); buf[sopt->sopt_valsize] = '\0'; CC_LIST_RLOCK(); STAILQ_FOREACH(algo, &cc_list, entries) { if (strncmp(buf, algo->name, TCP_CA_NAME_MAX) == 0) { if (algo->flags & CC_MODULE_BEING_REMOVED) { /* We can't "see" modules being unloaded */ continue; } break; } } if (algo == NULL) { CC_LIST_RUNLOCK(); return(ESRCH); } /* * With a reference the algorithm cannot be removed * so we hold a reference through the change process. */ cc_refer(algo); CC_LIST_RUNLOCK(); if (algo->cb_init != NULL) { /* We can now pre-get the memory for the CC */ mem_sz = (*algo->cc_data_sz)(); if (mem_sz == 0) { goto no_mem_needed; } ptr = malloc(mem_sz, M_CC_MEM, M_WAITOK); } else { no_mem_needed: mem_sz = 0; ptr = NULL; } /* * Make sure its all clean and zero and also get * back the inplock. */ memset(&cc_mem, 0, sizeof(cc_mem)); INP_WLOCK(inp); - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { INP_WUNLOCK(inp); if (ptr) free(ptr, M_CC_MEM); /* Release our temp reference */ CC_LIST_RLOCK(); cc_release(algo); CC_LIST_RUNLOCK(); return (ECONNRESET); } tp = intotcpcb(inp); if (ptr != NULL) memset(ptr, 0, mem_sz); cc_mem.ccvc.tcp = tp; /* * We once again hold a write lock over the tcb so it's * safe to do these things without ordering concerns. * Note here we init into stack memory. */ if (algo->cb_init != NULL) error = algo->cb_init(&cc_mem, ptr); else error = 0; /* * The CC algorithms, when given their memory * should not fail we could in theory have a * KASSERT here. */ if (error == 0) { /* * Touchdown, lets go ahead and move the * connection to the new CC module by * copying in the cc_mem after we call * the old ones cleanup (if any). */ if (CC_ALGO(tp)->cb_destroy != NULL) CC_ALGO(tp)->cb_destroy(tp->ccv); /* Detach the old CC from the tcpcb */ cc_detach(tp); /* Copy in our temp memory that was inited */ memcpy(tp->ccv, &cc_mem, sizeof(struct cc_var)); /* Now attach the new, which takes a reference */ cc_attach(tp, algo); /* Ok now are we where we have gotten past any conn_init? */ if (TCPS_HAVEESTABLISHED(tp->t_state) && (CC_ALGO(tp)->conn_init != NULL)) { /* Yep run the connection init for the new CC */ CC_ALGO(tp)->conn_init(tp->ccv); } } else if (ptr) free(ptr, M_CC_MEM); INP_WUNLOCK(inp); /* Now lets release our temp reference */ CC_LIST_RLOCK(); cc_release(algo); CC_LIST_RUNLOCK(); return (error); } int tcp_default_ctloutput(struct inpcb *inp, struct sockopt *sopt) { struct tcpcb *tp = intotcpcb(inp); int error, opt, optval; u_int ui; struct tcp_info ti; #ifdef KERN_TLS struct tls_enable tls; struct socket *so = inp->inp_socket; #endif char *pbuf, buf[TCP_LOG_ID_LEN]; #ifdef STATS struct statsblob *sbp; #endif size_t len; INP_WLOCK_ASSERT(inp); - KASSERT((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) == 0, + KASSERT((inp->inp_flags & INP_DROPPED) == 0, ("inp_flags == %x", inp->inp_flags)); KASSERT(inp->inp_socket != NULL, ("inp_socket == NULL")); switch (sopt->sopt_level) { #ifdef INET6 case IPPROTO_IPV6: MPASS(inp->inp_vflag & INP_IPV6PROTO); switch (sopt->sopt_name) { case IPV6_USE_MIN_MTU: tcp6_use_min_mtu(tp); /* FALLTHROUGH */ } INP_WUNLOCK(inp); return (0); #endif #ifdef INET case IPPROTO_IP: INP_WUNLOCK(inp); return (0); #endif } /* * For TCP_CCALGOOPT forward the control to CC module, for both * SOPT_SET and SOPT_GET. */ switch (sopt->sopt_name) { case TCP_CCALGOOPT: INP_WUNLOCK(inp); if (sopt->sopt_valsize > CC_ALGOOPT_LIMIT) return (EINVAL); pbuf = malloc(sopt->sopt_valsize, M_TEMP, M_WAITOK | M_ZERO); error = sooptcopyin(sopt, pbuf, sopt->sopt_valsize, sopt->sopt_valsize); if (error) { free(pbuf, M_TEMP); return (error); } INP_WLOCK_RECHECK_CLEANUP(inp, free(pbuf, M_TEMP)); if (CC_ALGO(tp)->ctl_output != NULL) error = CC_ALGO(tp)->ctl_output(tp->ccv, sopt, pbuf); else error = ENOENT; INP_WUNLOCK(inp); if (error == 0 && sopt->sopt_dir == SOPT_GET) error = sooptcopyout(sopt, pbuf, sopt->sopt_valsize); free(pbuf, M_TEMP); return (error); } switch (sopt->sopt_dir) { case SOPT_SET: switch (sopt->sopt_name) { #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) case TCP_MD5SIG: INP_WUNLOCK(inp); if (!TCPMD5_ENABLED()) return (ENOPROTOOPT); error = TCPMD5_PCBCTL(inp, sopt); if (error) return (error); INP_WLOCK_RECHECK(inp); goto unlock_and_done; #endif /* IPSEC */ case TCP_NODELAY: case TCP_NOOPT: case TCP_LRD: INP_WUNLOCK(inp); error = sooptcopyin(sopt, &optval, sizeof optval, sizeof optval); if (error) return (error); INP_WLOCK_RECHECK(inp); switch (sopt->sopt_name) { case TCP_NODELAY: opt = TF_NODELAY; break; case TCP_NOOPT: opt = TF_NOOPT; break; case TCP_LRD: opt = TF_LRD; break; default: opt = 0; /* dead code to fool gcc */ break; } if (optval) tp->t_flags |= opt; else tp->t_flags &= ~opt; unlock_and_done: #ifdef TCP_OFFLOAD if (tp->t_flags & TF_TOE) { tcp_offload_ctloutput(tp, sopt->sopt_dir, sopt->sopt_name); } #endif INP_WUNLOCK(inp); break; case TCP_NOPUSH: INP_WUNLOCK(inp); error = sooptcopyin(sopt, &optval, sizeof optval, sizeof optval); if (error) return (error); INP_WLOCK_RECHECK(inp); if (optval) tp->t_flags |= TF_NOPUSH; else if (tp->t_flags & TF_NOPUSH) { tp->t_flags &= ~TF_NOPUSH; if (TCPS_HAVEESTABLISHED(tp->t_state)) { struct epoch_tracker et; NET_EPOCH_ENTER(et); error = tcp_output_nodrop(tp); NET_EPOCH_EXIT(et); } } goto unlock_and_done; case TCP_REMOTE_UDP_ENCAPS_PORT: INP_WUNLOCK(inp); error = sooptcopyin(sopt, &optval, sizeof optval, sizeof optval); if (error) return (error); if ((optval < TCP_TUNNELING_PORT_MIN) || (optval > TCP_TUNNELING_PORT_MAX)) { /* Its got to be in range */ return (EINVAL); } if ((V_tcp_udp_tunneling_port == 0) && (optval != 0)) { /* You have to have enabled a UDP tunneling port first */ return (EINVAL); } INP_WLOCK_RECHECK(inp); if (tp->t_state != TCPS_CLOSED) { /* You can't change after you are connected */ error = EINVAL; } else { /* Ok we are all good set the port */ tp->t_port = htons(optval); } goto unlock_and_done; case TCP_MAXSEG: INP_WUNLOCK(inp); error = sooptcopyin(sopt, &optval, sizeof optval, sizeof optval); if (error) return (error); INP_WLOCK_RECHECK(inp); if (optval > 0 && optval <= tp->t_maxseg && optval + 40 >= V_tcp_minmss) tp->t_maxseg = optval; else error = EINVAL; goto unlock_and_done; case TCP_INFO: INP_WUNLOCK(inp); error = EINVAL; break; case TCP_STATS: INP_WUNLOCK(inp); #ifdef STATS error = sooptcopyin(sopt, &optval, sizeof optval, sizeof optval); if (error) return (error); if (optval > 0) sbp = stats_blob_alloc( V_tcp_perconn_stats_dflt_tpl, 0); else sbp = NULL; INP_WLOCK_RECHECK(inp); if ((tp->t_stats != NULL && sbp == NULL) || (tp->t_stats == NULL && sbp != NULL)) { struct statsblob *t = tp->t_stats; tp->t_stats = sbp; sbp = t; } INP_WUNLOCK(inp); stats_blob_destroy(sbp); #else return (EOPNOTSUPP); #endif /* !STATS */ break; case TCP_CONGESTION: error = tcp_set_cc_mod(inp, sopt); break; case TCP_REUSPORT_LB_NUMA: INP_WUNLOCK(inp); error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval)); INP_WLOCK_RECHECK(inp); if (!error) error = in_pcblbgroup_numa(inp, optval); INP_WUNLOCK(inp); break; #ifdef KERN_TLS case TCP_TXTLS_ENABLE: INP_WUNLOCK(inp); error = copyin_tls_enable(sopt, &tls); if (error) break; error = ktls_enable_tx(so, &tls); break; case TCP_TXTLS_MODE: INP_WUNLOCK(inp); error = sooptcopyin(sopt, &ui, sizeof(ui), sizeof(ui)); if (error) return (error); INP_WLOCK_RECHECK(inp); error = ktls_set_tx_mode(so, ui); INP_WUNLOCK(inp); break; case TCP_RXTLS_ENABLE: INP_WUNLOCK(inp); error = sooptcopyin(sopt, &tls, sizeof(tls), sizeof(tls)); if (error) break; error = ktls_enable_rx(so, &tls); break; #endif case TCP_MAXUNACKTIME: case TCP_KEEPIDLE: case TCP_KEEPINTVL: case TCP_KEEPINIT: INP_WUNLOCK(inp); error = sooptcopyin(sopt, &ui, sizeof(ui), sizeof(ui)); if (error) return (error); if (ui > (UINT_MAX / hz)) { error = EINVAL; break; } ui *= hz; INP_WLOCK_RECHECK(inp); switch (sopt->sopt_name) { case TCP_MAXUNACKTIME: tp->t_maxunacktime = ui; break; case TCP_KEEPIDLE: tp->t_keepidle = ui; /* * XXX: better check current remaining * timeout and "merge" it with new value. */ if ((tp->t_state > TCPS_LISTEN) && (tp->t_state <= TCPS_CLOSING)) tcp_timer_activate(tp, TT_KEEP, TP_KEEPIDLE(tp)); break; case TCP_KEEPINTVL: tp->t_keepintvl = ui; if ((tp->t_state == TCPS_FIN_WAIT_2) && (TP_MAXIDLE(tp) > 0)) tcp_timer_activate(tp, TT_2MSL, TP_MAXIDLE(tp)); break; case TCP_KEEPINIT: tp->t_keepinit = ui; if (tp->t_state == TCPS_SYN_RECEIVED || tp->t_state == TCPS_SYN_SENT) tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp)); break; } goto unlock_and_done; case TCP_KEEPCNT: INP_WUNLOCK(inp); error = sooptcopyin(sopt, &ui, sizeof(ui), sizeof(ui)); if (error) return (error); INP_WLOCK_RECHECK(inp); tp->t_keepcnt = ui; if ((tp->t_state == TCPS_FIN_WAIT_2) && (TP_MAXIDLE(tp) > 0)) tcp_timer_activate(tp, TT_2MSL, TP_MAXIDLE(tp)); goto unlock_and_done; #ifdef TCPPCAP case TCP_PCAP_OUT: case TCP_PCAP_IN: INP_WUNLOCK(inp); error = sooptcopyin(sopt, &optval, sizeof optval, sizeof optval); if (error) return (error); INP_WLOCK_RECHECK(inp); if (optval >= 0) tcp_pcap_set_sock_max(TCP_PCAP_OUT ? &(tp->t_outpkts) : &(tp->t_inpkts), optval); else error = EINVAL; goto unlock_and_done; #endif case TCP_FASTOPEN: { struct tcp_fastopen tfo_optval; INP_WUNLOCK(inp); if (!V_tcp_fastopen_client_enable && !V_tcp_fastopen_server_enable) return (EPERM); error = sooptcopyin(sopt, &tfo_optval, sizeof(tfo_optval), sizeof(int)); if (error) return (error); INP_WLOCK_RECHECK(inp); if ((tp->t_state != TCPS_CLOSED) && (tp->t_state != TCPS_LISTEN)) { error = EINVAL; goto unlock_and_done; } if (tfo_optval.enable) { if (tp->t_state == TCPS_LISTEN) { if (!V_tcp_fastopen_server_enable) { error = EPERM; goto unlock_and_done; } if (tp->t_tfo_pending == NULL) tp->t_tfo_pending = tcp_fastopen_alloc_counter(); } else { /* * If a pre-shared key was provided, * stash it in the client cookie * field of the tcpcb for use during * connect. */ if (sopt->sopt_valsize == sizeof(tfo_optval)) { memcpy(tp->t_tfo_cookie.client, tfo_optval.psk, TCP_FASTOPEN_PSK_LEN); tp->t_tfo_client_cookie_len = TCP_FASTOPEN_PSK_LEN; } } tp->t_flags |= TF_FASTOPEN; } else tp->t_flags &= ~TF_FASTOPEN; goto unlock_and_done; } #ifdef TCP_BLACKBOX case TCP_LOG: INP_WUNLOCK(inp); error = sooptcopyin(sopt, &optval, sizeof optval, sizeof optval); if (error) return (error); INP_WLOCK_RECHECK(inp); error = tcp_log_state_change(tp, optval); goto unlock_and_done; case TCP_LOGBUF: INP_WUNLOCK(inp); error = EINVAL; break; case TCP_LOGID: INP_WUNLOCK(inp); error = sooptcopyin(sopt, buf, TCP_LOG_ID_LEN - 1, 0); if (error) break; buf[sopt->sopt_valsize] = '\0'; INP_WLOCK_RECHECK(inp); error = tcp_log_set_id(tp, buf); /* tcp_log_set_id() unlocks the INP. */ break; case TCP_LOGDUMP: case TCP_LOGDUMPID: INP_WUNLOCK(inp); error = sooptcopyin(sopt, buf, TCP_LOG_REASON_LEN - 1, 0); if (error) break; buf[sopt->sopt_valsize] = '\0'; INP_WLOCK_RECHECK(inp); if (sopt->sopt_name == TCP_LOGDUMP) { error = tcp_log_dump_tp_logbuf(tp, buf, M_WAITOK, true); INP_WUNLOCK(inp); } else { tcp_log_dump_tp_bucket_logbufs(tp, buf); /* * tcp_log_dump_tp_bucket_logbufs() drops the * INP lock. */ } break; #endif default: INP_WUNLOCK(inp); error = ENOPROTOOPT; break; } break; case SOPT_GET: tp = intotcpcb(inp); switch (sopt->sopt_name) { #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) case TCP_MD5SIG: INP_WUNLOCK(inp); if (!TCPMD5_ENABLED()) return (ENOPROTOOPT); error = TCPMD5_PCBCTL(inp, sopt); break; #endif case TCP_NODELAY: optval = tp->t_flags & TF_NODELAY; INP_WUNLOCK(inp); error = sooptcopyout(sopt, &optval, sizeof optval); break; case TCP_MAXSEG: optval = tp->t_maxseg; INP_WUNLOCK(inp); error = sooptcopyout(sopt, &optval, sizeof optval); break; case TCP_REMOTE_UDP_ENCAPS_PORT: optval = ntohs(tp->t_port); INP_WUNLOCK(inp); error = sooptcopyout(sopt, &optval, sizeof optval); break; case TCP_NOOPT: optval = tp->t_flags & TF_NOOPT; INP_WUNLOCK(inp); error = sooptcopyout(sopt, &optval, sizeof optval); break; case TCP_NOPUSH: optval = tp->t_flags & TF_NOPUSH; INP_WUNLOCK(inp); error = sooptcopyout(sopt, &optval, sizeof optval); break; case TCP_INFO: tcp_fill_info(tp, &ti); INP_WUNLOCK(inp); error = sooptcopyout(sopt, &ti, sizeof ti); break; case TCP_STATS: { #ifdef STATS int nheld; TYPEOF_MEMBER(struct statsblob, flags) sbflags = 0; error = 0; socklen_t outsbsz = sopt->sopt_valsize; if (tp->t_stats == NULL) error = ENOENT; else if (outsbsz >= tp->t_stats->cursz) outsbsz = tp->t_stats->cursz; else if (outsbsz >= sizeof(struct statsblob)) outsbsz = sizeof(struct statsblob); else error = EINVAL; INP_WUNLOCK(inp); if (error) break; sbp = sopt->sopt_val; nheld = atop(round_page(((vm_offset_t)sbp) + (vm_size_t)outsbsz) - trunc_page((vm_offset_t)sbp)); vm_page_t ma[nheld]; if (vm_fault_quick_hold_pages( &curproc->p_vmspace->vm_map, (vm_offset_t)sbp, outsbsz, VM_PROT_READ | VM_PROT_WRITE, ma, nheld) < 0) { error = EFAULT; break; } if ((error = copyin_nofault(&(sbp->flags), &sbflags, SIZEOF_MEMBER(struct statsblob, flags)))) goto unhold; INP_WLOCK_RECHECK(inp); error = stats_blob_snapshot(&sbp, outsbsz, tp->t_stats, sbflags | SB_CLONE_USRDSTNOFAULT); INP_WUNLOCK(inp); sopt->sopt_valsize = outsbsz; unhold: vm_page_unhold_pages(ma, nheld); #else INP_WUNLOCK(inp); error = EOPNOTSUPP; #endif /* !STATS */ break; } case TCP_CONGESTION: len = strlcpy(buf, CC_ALGO(tp)->name, TCP_CA_NAME_MAX); INP_WUNLOCK(inp); error = sooptcopyout(sopt, buf, len + 1); break; case TCP_MAXUNACKTIME: case TCP_KEEPIDLE: case TCP_KEEPINTVL: case TCP_KEEPINIT: case TCP_KEEPCNT: switch (sopt->sopt_name) { case TCP_MAXUNACKTIME: ui = TP_MAXUNACKTIME(tp) / hz; break; case TCP_KEEPIDLE: ui = TP_KEEPIDLE(tp) / hz; break; case TCP_KEEPINTVL: ui = TP_KEEPINTVL(tp) / hz; break; case TCP_KEEPINIT: ui = TP_KEEPINIT(tp) / hz; break; case TCP_KEEPCNT: ui = TP_KEEPCNT(tp); break; } INP_WUNLOCK(inp); error = sooptcopyout(sopt, &ui, sizeof(ui)); break; #ifdef TCPPCAP case TCP_PCAP_OUT: case TCP_PCAP_IN: optval = tcp_pcap_get_sock_max(TCP_PCAP_OUT ? &(tp->t_outpkts) : &(tp->t_inpkts)); INP_WUNLOCK(inp); error = sooptcopyout(sopt, &optval, sizeof optval); break; #endif case TCP_FASTOPEN: optval = tp->t_flags & TF_FASTOPEN; INP_WUNLOCK(inp); error = sooptcopyout(sopt, &optval, sizeof optval); break; #ifdef TCP_BLACKBOX case TCP_LOG: optval = tp->t_logstate; INP_WUNLOCK(inp); error = sooptcopyout(sopt, &optval, sizeof(optval)); break; case TCP_LOGBUF: /* tcp_log_getlogbuf() does INP_WUNLOCK(inp) */ error = tcp_log_getlogbuf(sopt, tp); break; case TCP_LOGID: len = tcp_log_get_id(tp, buf); INP_WUNLOCK(inp); error = sooptcopyout(sopt, buf, len + 1); break; case TCP_LOGDUMP: case TCP_LOGDUMPID: INP_WUNLOCK(inp); error = EINVAL; break; #endif #ifdef KERN_TLS case TCP_TXTLS_MODE: error = ktls_get_tx_mode(so, &optval); INP_WUNLOCK(inp); if (error == 0) error = sooptcopyout(sopt, &optval, sizeof(optval)); break; case TCP_RXTLS_MODE: error = ktls_get_rx_mode(so, &optval); INP_WUNLOCK(inp); if (error == 0) error = sooptcopyout(sopt, &optval, sizeof(optval)); break; #endif case TCP_LRD: optval = tp->t_flags & TF_LRD; INP_WUNLOCK(inp); error = sooptcopyout(sopt, &optval, sizeof optval); break; default: INP_WUNLOCK(inp); error = ENOPROTOOPT; break; } break; } return (error); } #undef INP_WLOCK_RECHECK #undef INP_WLOCK_RECHECK_CLEANUP /* * Initiate (or continue) disconnect. * If embryonic state, just send reset (once). * If in ``let data drain'' option and linger null, just drop. * Otherwise (hard), mark socket disconnecting and drop * current input data; switch states based on user close, and * send segment to peer (with FIN). */ static void tcp_disconnect(struct tcpcb *tp) { struct inpcb *inp = tp->t_inpcb; struct socket *so = inp->inp_socket; NET_EPOCH_ASSERT(); INP_WLOCK_ASSERT(inp); /* * Neither tcp_close() nor tcp_drop() should return NULL, as the * socket is still open. */ if (tp->t_state < TCPS_ESTABLISHED && !(tp->t_state > TCPS_LISTEN && IS_FASTOPEN(tp->t_flags))) { tp = tcp_close(tp); KASSERT(tp != NULL, ("tcp_disconnect: tcp_close() returned NULL")); } else if ((so->so_options & SO_LINGER) && so->so_linger == 0) { tp = tcp_drop(tp, 0); KASSERT(tp != NULL, ("tcp_disconnect: tcp_drop() returned NULL")); } else { soisdisconnecting(so); sbflush(&so->so_rcv); tcp_usrclosed(tp); if (!(inp->inp_flags & INP_DROPPED)) /* Ignore stack's drop request, we already at it. */ (void)tcp_output_nodrop(tp); } } /* * User issued close, and wish to trail through shutdown states: * if never received SYN, just forget it. If got a SYN from peer, * but haven't sent FIN, then go to FIN_WAIT_1 state to send peer a FIN. * If already got a FIN from peer, then almost done; go to LAST_ACK * state. In all other cases, have already sent FIN to peer (e.g. * after PRU_SHUTDOWN), and just have to play tedious game waiting * for peer to send FIN or not respond to keep-alives, etc. * We can let the user exit from the close as soon as the FIN is acked. */ static void tcp_usrclosed(struct tcpcb *tp) { NET_EPOCH_ASSERT(); INP_WLOCK_ASSERT(tp->t_inpcb); switch (tp->t_state) { case TCPS_LISTEN: #ifdef TCP_OFFLOAD tcp_offload_listen_stop(tp); #endif tcp_state_change(tp, TCPS_CLOSED); /* FALLTHROUGH */ case TCPS_CLOSED: tp = tcp_close(tp); /* * tcp_close() should never return NULL here as the socket is * still open. */ KASSERT(tp != NULL, ("tcp_usrclosed: tcp_close() returned NULL")); break; case TCPS_SYN_SENT: case TCPS_SYN_RECEIVED: tp->t_flags |= TF_NEEDFIN; break; case TCPS_ESTABLISHED: tcp_state_change(tp, TCPS_FIN_WAIT_1); break; case TCPS_CLOSE_WAIT: tcp_state_change(tp, TCPS_LAST_ACK); break; } if (tp->t_acktime == 0) tp->t_acktime = ticks; if (tp->t_state >= TCPS_FIN_WAIT_2) { soisdisconnected(tp->t_inpcb->inp_socket); /* Prevent the connection hanging in FIN_WAIT_2 forever. */ if (tp->t_state == TCPS_FIN_WAIT_2) { int timeout; timeout = (tcp_fast_finwait2_recycle) ? tcp_finwait2_timeout : TP_MAXIDLE(tp); tcp_timer_activate(tp, TT_2MSL, timeout); } } } #ifdef DDB static void db_print_indent(int indent) { int i; for (i = 0; i < indent; i++) db_printf(" "); } static void db_print_tstate(int t_state) { switch (t_state) { case TCPS_CLOSED: db_printf("TCPS_CLOSED"); return; case TCPS_LISTEN: db_printf("TCPS_LISTEN"); return; case TCPS_SYN_SENT: db_printf("TCPS_SYN_SENT"); return; case TCPS_SYN_RECEIVED: db_printf("TCPS_SYN_RECEIVED"); return; case TCPS_ESTABLISHED: db_printf("TCPS_ESTABLISHED"); return; case TCPS_CLOSE_WAIT: db_printf("TCPS_CLOSE_WAIT"); return; case TCPS_FIN_WAIT_1: db_printf("TCPS_FIN_WAIT_1"); return; case TCPS_CLOSING: db_printf("TCPS_CLOSING"); return; case TCPS_LAST_ACK: db_printf("TCPS_LAST_ACK"); return; case TCPS_FIN_WAIT_2: db_printf("TCPS_FIN_WAIT_2"); return; case TCPS_TIME_WAIT: db_printf("TCPS_TIME_WAIT"); return; default: db_printf("unknown"); return; } } static void db_print_tflags(u_int t_flags) { int comma; comma = 0; if (t_flags & TF_ACKNOW) { db_printf("%sTF_ACKNOW", comma ? ", " : ""); comma = 1; } if (t_flags & TF_DELACK) { db_printf("%sTF_DELACK", comma ? ", " : ""); comma = 1; } if (t_flags & TF_NODELAY) { db_printf("%sTF_NODELAY", comma ? ", " : ""); comma = 1; } if (t_flags & TF_NOOPT) { db_printf("%sTF_NOOPT", comma ? ", " : ""); comma = 1; } if (t_flags & TF_SENTFIN) { db_printf("%sTF_SENTFIN", comma ? ", " : ""); comma = 1; } if (t_flags & TF_REQ_SCALE) { db_printf("%sTF_REQ_SCALE", comma ? ", " : ""); comma = 1; } if (t_flags & TF_RCVD_SCALE) { db_printf("%sTF_RECVD_SCALE", comma ? ", " : ""); comma = 1; } if (t_flags & TF_REQ_TSTMP) { db_printf("%sTF_REQ_TSTMP", comma ? ", " : ""); comma = 1; } if (t_flags & TF_RCVD_TSTMP) { db_printf("%sTF_RCVD_TSTMP", comma ? ", " : ""); comma = 1; } if (t_flags & TF_SACK_PERMIT) { db_printf("%sTF_SACK_PERMIT", comma ? ", " : ""); comma = 1; } if (t_flags & TF_NEEDSYN) { db_printf("%sTF_NEEDSYN", comma ? ", " : ""); comma = 1; } if (t_flags & TF_NEEDFIN) { db_printf("%sTF_NEEDFIN", comma ? ", " : ""); comma = 1; } if (t_flags & TF_NOPUSH) { db_printf("%sTF_NOPUSH", comma ? ", " : ""); comma = 1; } if (t_flags & TF_PREVVALID) { db_printf("%sTF_PREVVALID", comma ? ", " : ""); comma = 1; } if (t_flags & TF_MORETOCOME) { db_printf("%sTF_MORETOCOME", comma ? ", " : ""); comma = 1; } if (t_flags & TF_SONOTCONN) { db_printf("%sTF_SONOTCONN", comma ? ", " : ""); comma = 1; } if (t_flags & TF_LASTIDLE) { db_printf("%sTF_LASTIDLE", comma ? ", " : ""); comma = 1; } if (t_flags & TF_RXWIN0SENT) { db_printf("%sTF_RXWIN0SENT", comma ? ", " : ""); comma = 1; } if (t_flags & TF_FASTRECOVERY) { db_printf("%sTF_FASTRECOVERY", comma ? ", " : ""); comma = 1; } if (t_flags & TF_CONGRECOVERY) { db_printf("%sTF_CONGRECOVERY", comma ? ", " : ""); comma = 1; } if (t_flags & TF_WASFRECOVERY) { db_printf("%sTF_WASFRECOVERY", comma ? ", " : ""); comma = 1; } if (t_flags & TF_WASCRECOVERY) { db_printf("%sTF_WASCRECOVERY", comma ? ", " : ""); comma = 1; } if (t_flags & TF_SIGNATURE) { db_printf("%sTF_SIGNATURE", comma ? ", " : ""); comma = 1; } if (t_flags & TF_FORCEDATA) { db_printf("%sTF_FORCEDATA", comma ? ", " : ""); comma = 1; } if (t_flags & TF_TSO) { db_printf("%sTF_TSO", comma ? ", " : ""); comma = 1; } if (t_flags & TF_FASTOPEN) { db_printf("%sTF_FASTOPEN", comma ? ", " : ""); comma = 1; } } static void db_print_tflags2(u_int t_flags2) { int comma; comma = 0; if (t_flags2 & TF2_PLPMTU_BLACKHOLE) { db_printf("%sTF2_PLPMTU_BLACKHOLE", comma ? ", " : ""); comma = 1; } if (t_flags2 & TF2_PLPMTU_PMTUD) { db_printf("%sTF2_PLPMTU_PMTUD", comma ? ", " : ""); comma = 1; } if (t_flags2 & TF2_PLPMTU_MAXSEGSNT) { db_printf("%sTF2_PLPMTU_MAXSEGSNT", comma ? ", " : ""); comma = 1; } if (t_flags2 & TF2_LOG_AUTO) { db_printf("%sTF2_LOG_AUTO", comma ? ", " : ""); comma = 1; } if (t_flags2 & TF2_DROP_AF_DATA) { db_printf("%sTF2_DROP_AF_DATA", comma ? ", " : ""); comma = 1; } if (t_flags2 & TF2_ECN_PERMIT) { db_printf("%sTF2_ECN_PERMIT", comma ? ", " : ""); comma = 1; } if (t_flags2 & TF2_ECN_SND_CWR) { db_printf("%sTF2_ECN_SND_CWR", comma ? ", " : ""); comma = 1; } if (t_flags2 & TF2_ECN_SND_ECE) { db_printf("%sTF2_ECN_SND_ECE", comma ? ", " : ""); comma = 1; } if (t_flags2 & TF2_ACE_PERMIT) { db_printf("%sTF2_ACE_PERMIT", comma ? ", " : ""); comma = 1; } if (t_flags2 & TF2_FBYTES_COMPLETE) { db_printf("%sTF2_FBYTES_COMPLETE", comma ? ", " : ""); comma = 1; } } static void db_print_toobflags(char t_oobflags) { int comma; comma = 0; if (t_oobflags & TCPOOB_HAVEDATA) { db_printf("%sTCPOOB_HAVEDATA", comma ? ", " : ""); comma = 1; } if (t_oobflags & TCPOOB_HADDATA) { db_printf("%sTCPOOB_HADDATA", comma ? ", " : ""); comma = 1; } } static void db_print_tcpcb(struct tcpcb *tp, const char *name, int indent) { db_print_indent(indent); db_printf("%s at %p\n", name, tp); indent += 2; db_print_indent(indent); db_printf("t_segq first: %p t_segqlen: %d t_dupacks: %d\n", TAILQ_FIRST(&tp->t_segq), tp->t_segqlen, tp->t_dupacks); db_print_indent(indent); db_printf("tt_rexmt: %p tt_persist: %p tt_keep: %p\n", &tp->t_timers->tt_rexmt, &tp->t_timers->tt_persist, &tp->t_timers->tt_keep); db_print_indent(indent); db_printf("tt_2msl: %p tt_delack: %p t_inpcb: %p\n", &tp->t_timers->tt_2msl, &tp->t_timers->tt_delack, tp->t_inpcb); db_print_indent(indent); db_printf("t_state: %d (", tp->t_state); db_print_tstate(tp->t_state); db_printf(")\n"); db_print_indent(indent); db_printf("t_flags: 0x%x (", tp->t_flags); db_print_tflags(tp->t_flags); db_printf(")\n"); db_print_indent(indent); db_printf("t_flags2: 0x%x (", tp->t_flags2); db_print_tflags2(tp->t_flags2); db_printf(")\n"); db_print_indent(indent); db_printf("snd_una: 0x%08x snd_max: 0x%08x snd_nxt: x0%08x\n", tp->snd_una, tp->snd_max, tp->snd_nxt); db_print_indent(indent); db_printf("snd_up: 0x%08x snd_wl1: 0x%08x snd_wl2: 0x%08x\n", tp->snd_up, tp->snd_wl1, tp->snd_wl2); db_print_indent(indent); db_printf("iss: 0x%08x irs: 0x%08x rcv_nxt: 0x%08x\n", tp->iss, tp->irs, tp->rcv_nxt); db_print_indent(indent); db_printf("rcv_adv: 0x%08x rcv_wnd: %u rcv_up: 0x%08x\n", tp->rcv_adv, tp->rcv_wnd, tp->rcv_up); db_print_indent(indent); db_printf("snd_wnd: %u snd_cwnd: %u\n", tp->snd_wnd, tp->snd_cwnd); db_print_indent(indent); db_printf("snd_ssthresh: %u snd_recover: " "0x%08x\n", tp->snd_ssthresh, tp->snd_recover); db_print_indent(indent); db_printf("t_rcvtime: %u t_startime: %u\n", tp->t_rcvtime, tp->t_starttime); db_print_indent(indent); db_printf("t_rttime: %u t_rtsq: 0x%08x\n", tp->t_rtttime, tp->t_rtseq); db_print_indent(indent); db_printf("t_rxtcur: %d t_maxseg: %u t_srtt: %d\n", tp->t_rxtcur, tp->t_maxseg, tp->t_srtt); db_print_indent(indent); db_printf("t_rttvar: %d t_rxtshift: %d t_rttmin: %u " "t_rttbest: %u\n", tp->t_rttvar, tp->t_rxtshift, tp->t_rttmin, tp->t_rttbest); db_print_indent(indent); db_printf("t_rttupdated: %lu max_sndwnd: %u t_softerror: %d\n", tp->t_rttupdated, tp->max_sndwnd, tp->t_softerror); db_print_indent(indent); db_printf("t_oobflags: 0x%x (", tp->t_oobflags); db_print_toobflags(tp->t_oobflags); db_printf(") t_iobc: 0x%02x\n", tp->t_iobc); db_print_indent(indent); db_printf("snd_scale: %u rcv_scale: %u request_r_scale: %u\n", tp->snd_scale, tp->rcv_scale, tp->request_r_scale); db_print_indent(indent); db_printf("ts_recent: %u ts_recent_age: %u\n", tp->ts_recent, tp->ts_recent_age); db_print_indent(indent); db_printf("ts_offset: %u last_ack_sent: 0x%08x snd_cwnd_prev: " "%u\n", tp->ts_offset, tp->last_ack_sent, tp->snd_cwnd_prev); db_print_indent(indent); db_printf("snd_ssthresh_prev: %u snd_recover_prev: 0x%08x " "t_badrxtwin: %u\n", tp->snd_ssthresh_prev, tp->snd_recover_prev, tp->t_badrxtwin); db_print_indent(indent); db_printf("snd_numholes: %d snd_holes first: %p\n", tp->snd_numholes, TAILQ_FIRST(&tp->snd_holes)); db_print_indent(indent); db_printf("snd_fack: 0x%08x rcv_numsacks: %d\n", tp->snd_fack, tp->rcv_numsacks); /* Skip sackblks, sackhint. */ db_print_indent(indent); db_printf("t_rttlow: %d rfbuf_ts: %u rfbuf_cnt: %d\n", tp->t_rttlow, tp->rfbuf_ts, tp->rfbuf_cnt); } DB_SHOW_COMMAND(tcpcb, db_show_tcpcb) { struct tcpcb *tp; if (!have_addr) { db_printf("usage: show tcpcb \n"); return; } tp = (struct tcpcb *)addr; db_print_tcpcb(tp, "tcpcb", 0); } #endif diff --git a/sys/netinet/toecore.c b/sys/netinet/toecore.c index a185103df3a7..a078d9cea51f 100644 --- a/sys/netinet/toecore.c +++ b/sys/netinet/toecore.c @@ -1,605 +1,605 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2012 Chelsio Communications, Inc. * All rights reserved. * Written by: Navdeep Parhar * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #define TCPSTATES #include #include #include #include #include #include #include static struct mtx toedev_lock; static TAILQ_HEAD(, toedev) toedev_list; static eventhandler_tag listen_start_eh; static eventhandler_tag listen_stop_eh; static eventhandler_tag lle_event_eh; static int toedev_connect(struct toedev *tod __unused, struct socket *so __unused, struct nhop_object *nh __unused, struct sockaddr *nam __unused) { return (ENOTSUP); } static int toedev_listen_start(struct toedev *tod __unused, struct tcpcb *tp __unused) { return (ENOTSUP); } static int toedev_listen_stop(struct toedev *tod __unused, struct tcpcb *tp __unused) { return (ENOTSUP); } static void toedev_input(struct toedev *tod __unused, struct tcpcb *tp __unused, struct mbuf *m) { m_freem(m); return; } static void toedev_rcvd(struct toedev *tod __unused, struct tcpcb *tp __unused) { return; } static int toedev_output(struct toedev *tod __unused, struct tcpcb *tp __unused) { return (ENOTSUP); } static void toedev_pcb_detach(struct toedev *tod __unused, struct tcpcb *tp __unused) { return; } static void toedev_l2_update(struct toedev *tod __unused, struct ifnet *ifp __unused, struct sockaddr *sa __unused, uint8_t *lladdr __unused, uint16_t vtag __unused) { return; } static void toedev_route_redirect(struct toedev *tod __unused, struct ifnet *ifp __unused, struct nhop_object *nh0 __unused, struct nhop_object *nh1 __unused) { return; } static void toedev_syncache_added(struct toedev *tod __unused, void *ctx __unused) { return; } static void toedev_syncache_removed(struct toedev *tod __unused, void *ctx __unused) { return; } static int toedev_syncache_respond(struct toedev *tod __unused, void *ctx __unused, struct mbuf *m) { m_freem(m); return (0); } static void toedev_offload_socket(struct toedev *tod __unused, void *ctx __unused, struct socket *so __unused) { return; } static void toedev_ctloutput(struct toedev *tod __unused, struct tcpcb *tp __unused, int sopt_dir __unused, int sopt_name __unused) { return; } static void toedev_tcp_info(struct toedev *tod __unused, struct tcpcb *tp __unused, struct tcp_info *ti __unused) { return; } static int toedev_alloc_tls_session(struct toedev *tod __unused, struct tcpcb *tp __unused, struct ktls_session *tls __unused, int direction __unused) { return (EINVAL); } static void toedev_pmtu_update(struct toedev *tod __unused, struct tcpcb *tp __unused, tcp_seq seq __unused, int mtu __unused) { return; } /* * Inform one or more TOE devices about a listening socket. */ static void toe_listen_start(struct inpcb *inp, void *arg) { struct toedev *t, *tod; struct tcpcb *tp; INP_WLOCK_ASSERT(inp); KASSERT(inp->inp_pcbinfo == &V_tcbinfo, ("%s: inp is not a TCP inp", __func__)); - if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) + if (inp->inp_flags & INP_DROPPED) return; tp = intotcpcb(inp); if (tp->t_state != TCPS_LISTEN) return; t = arg; mtx_lock(&toedev_lock); TAILQ_FOREACH(tod, &toedev_list, link) { if (t == NULL || t == tod) tod->tod_listen_start(tod, tp); } mtx_unlock(&toedev_lock); } static void toe_listen_start_event(void *arg __unused, struct tcpcb *tp) { struct inpcb *inp = tp->t_inpcb; INP_WLOCK_ASSERT(inp); KASSERT(tp->t_state == TCPS_LISTEN, ("%s: t_state %s", __func__, tcpstates[tp->t_state])); toe_listen_start(inp, NULL); } static void toe_listen_stop_event(void *arg __unused, struct tcpcb *tp) { struct toedev *tod; #ifdef INVARIANTS struct inpcb *inp = tp->t_inpcb; #endif INP_WLOCK_ASSERT(inp); KASSERT(tp->t_state == TCPS_LISTEN, ("%s: t_state %s", __func__, tcpstates[tp->t_state])); mtx_lock(&toedev_lock); TAILQ_FOREACH(tod, &toedev_list, link) tod->tod_listen_stop(tod, tp); mtx_unlock(&toedev_lock); } /* * Fill up a freshly allocated toedev struct with reasonable defaults. */ void init_toedev(struct toedev *tod) { tod->tod_softc = NULL; /* * Provide no-op defaults so that the kernel can call any toedev * function without having to check whether the TOE driver supplied one * or not. */ tod->tod_connect = toedev_connect; tod->tod_listen_start = toedev_listen_start; tod->tod_listen_stop = toedev_listen_stop; tod->tod_input = toedev_input; tod->tod_rcvd = toedev_rcvd; tod->tod_output = toedev_output; tod->tod_send_rst = toedev_output; tod->tod_send_fin = toedev_output; tod->tod_pcb_detach = toedev_pcb_detach; tod->tod_l2_update = toedev_l2_update; tod->tod_route_redirect = toedev_route_redirect; tod->tod_syncache_added = toedev_syncache_added; tod->tod_syncache_removed = toedev_syncache_removed; tod->tod_syncache_respond = toedev_syncache_respond; tod->tod_offload_socket = toedev_offload_socket; tod->tod_ctloutput = toedev_ctloutput; tod->tod_tcp_info = toedev_tcp_info; tod->tod_alloc_tls_session = toedev_alloc_tls_session; tod->tod_pmtu_update = toedev_pmtu_update; } /* * Register an active TOE device with the system. This allows it to receive * notifications from the kernel. */ int register_toedev(struct toedev *tod) { struct toedev *t; mtx_lock(&toedev_lock); TAILQ_FOREACH(t, &toedev_list, link) { if (t == tod) { mtx_unlock(&toedev_lock); return (EEXIST); } } TAILQ_INSERT_TAIL(&toedev_list, tod, link); registered_toedevs++; mtx_unlock(&toedev_lock); inp_apply_all(toe_listen_start, tod); return (0); } /* * Remove the TOE device from the global list of active TOE devices. It is the * caller's responsibility to ensure that the TOE device is quiesced prior to * this call. */ int unregister_toedev(struct toedev *tod) { struct toedev *t, *t2; int rc = ENODEV; mtx_lock(&toedev_lock); TAILQ_FOREACH_SAFE(t, &toedev_list, link, t2) { if (t == tod) { TAILQ_REMOVE(&toedev_list, tod, link); registered_toedevs--; rc = 0; break; } } KASSERT(registered_toedevs >= 0, ("%s: registered_toedevs (%d) < 0", __func__, registered_toedevs)); mtx_unlock(&toedev_lock); return (rc); } void toe_syncache_add(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, struct inpcb *inp, void *tod, void *todctx, uint8_t iptos) { INP_RLOCK_ASSERT(inp); (void )syncache_add(inc, to, th, inp, inp->inp_socket, NULL, tod, todctx, iptos, htons(0)); } int toe_syncache_expand(struct in_conninfo *inc, struct tcpopt *to, struct tcphdr *th, struct socket **lsop) { NET_EPOCH_ASSERT(); return (syncache_expand(inc, to, th, lsop, NULL, htons(0))); } /* * General purpose check to see if a 4-tuple is in use by the kernel. If a TCP * header (presumably for an incoming SYN) is also provided, an existing 4-tuple * in TIME_WAIT may be assassinated freeing it up for re-use. * * Note that the TCP header must have been run through tcp_fields_to_host() or * equivalent. */ int toe_4tuple_check(struct in_conninfo *inc, struct tcphdr *th, struct ifnet *ifp) { struct inpcb *inp; struct tcpcb *tp; if (inc->inc_flags & INC_ISIPV6) { inp = in6_pcblookup(&V_tcbinfo, &inc->inc6_faddr, inc->inc_fport, &inc->inc6_laddr, inc->inc_lport, INPLOOKUP_RLOCKPCB, ifp); } else { inp = in_pcblookup(&V_tcbinfo, inc->inc_faddr, inc->inc_fport, inc->inc_laddr, inc->inc_lport, INPLOOKUP_RLOCKPCB, ifp); } if (inp != NULL) { INP_RLOCK_ASSERT(inp); tp = intotcpcb(inp); if (tp->t_state == TCPS_TIME_WAIT && th != NULL) { if (!tcp_twcheck(inp, NULL, th, NULL, 0)) return (EADDRINUSE); } else { INP_RUNLOCK(inp); return (EADDRINUSE); } } return (0); } static void toe_lle_event(void *arg __unused, struct llentry *lle, int evt) { struct toedev *tod; struct ifnet *ifp; struct sockaddr *sa; uint8_t *lladdr; uint16_t vid, pcp; int family; struct sockaddr_in6 sin6; LLE_WLOCK_ASSERT(lle); ifp = lltable_get_ifp(lle->lle_tbl); family = lltable_get_af(lle->lle_tbl); if (family != AF_INET && family != AF_INET6) return; /* * Not interested if the interface's TOE capability is not enabled. */ if ((family == AF_INET && !(ifp->if_capenable & IFCAP_TOE4)) || (family == AF_INET6 && !(ifp->if_capenable & IFCAP_TOE6))) return; tod = TOEDEV(ifp); if (tod == NULL) return; sa = (struct sockaddr *)&sin6; lltable_fill_sa_entry(lle, sa); vid = 0xfff; pcp = 0; if (evt != LLENTRY_RESOLVED) { /* * LLENTRY_TIMEDOUT, LLENTRY_DELETED, LLENTRY_EXPIRED all mean * this entry is going to be deleted. */ lladdr = NULL; } else { KASSERT(lle->la_flags & LLE_VALID, ("%s: %p resolved but not valid?", __func__, lle)); lladdr = (uint8_t *)lle->ll_addr; VLAN_TAG(ifp, &vid); VLAN_PCP(ifp, &pcp); } tod->tod_l2_update(tod, ifp, sa, lladdr, EVL_MAKETAG(vid, pcp, 0)); } /* * Returns 0 or EWOULDBLOCK on success (any other value is an error). 0 means * lladdr and vtag are valid on return, EWOULDBLOCK means the TOE driver's * tod_l2_update will be called later, when the entry is resolved or times out. */ int toe_l2_resolve(struct toedev *tod, struct ifnet *ifp, struct sockaddr *sa, uint8_t *lladdr, uint16_t *vtag) { int rc; uint16_t vid, pcp; switch (sa->sa_family) { #ifdef INET case AF_INET: rc = arpresolve(ifp, 0, NULL, sa, lladdr, NULL, NULL); break; #endif #ifdef INET6 case AF_INET6: rc = nd6_resolve(ifp, LLE_SF(AF_INET6, 0), NULL, sa, lladdr, NULL, NULL); break; #endif default: return (EPROTONOSUPPORT); } if (rc == 0) { vid = 0xfff; pcp = 0; if (ifp->if_type == IFT_L2VLAN) { VLAN_TAG(ifp, &vid); VLAN_PCP(ifp, &pcp); } else if (ifp->if_pcp != IFNET_PCP_NONE) { vid = 0; pcp = ifp->if_pcp; } *vtag = EVL_MAKETAG(vid, pcp, 0); } return (rc); } void toe_connect_failed(struct toedev *tod, struct inpcb *inp, int err) { NET_EPOCH_ASSERT(); INP_WLOCK_ASSERT(inp); if (!(inp->inp_flags & INP_DROPPED)) { struct tcpcb *tp = intotcpcb(inp); KASSERT(tp->t_flags & TF_TOE, ("%s: tp %p not offloaded.", __func__, tp)); if (err == EAGAIN) { /* * Temporary failure during offload, take this PCB back. * Detach from the TOE driver and do the rest of what * TCP's pru_connect would have done if the connection * wasn't offloaded. */ tod->tod_pcb_detach(tod, tp); KASSERT(!(tp->t_flags & TF_TOE), ("%s: tp %p still offloaded.", __func__, tp)); tcp_timer_activate(tp, TT_KEEP, TP_KEEPINIT(tp)); if (tcp_output(tp) < 0) INP_WLOCK(inp); /* re-acquire */ } else { tp = tcp_drop(tp, err); if (tp == NULL) INP_WLOCK(inp); /* re-acquire */ } } INP_WLOCK_ASSERT(inp); } static int toecore_load(void) { mtx_init(&toedev_lock, "toedev lock", NULL, MTX_DEF); TAILQ_INIT(&toedev_list); listen_start_eh = EVENTHANDLER_REGISTER(tcp_offload_listen_start, toe_listen_start_event, NULL, EVENTHANDLER_PRI_ANY); listen_stop_eh = EVENTHANDLER_REGISTER(tcp_offload_listen_stop, toe_listen_stop_event, NULL, EVENTHANDLER_PRI_ANY); lle_event_eh = EVENTHANDLER_REGISTER(lle_event, toe_lle_event, NULL, EVENTHANDLER_PRI_ANY); return (0); } static int toecore_unload(void) { mtx_lock(&toedev_lock); if (!TAILQ_EMPTY(&toedev_list)) { mtx_unlock(&toedev_lock); return (EBUSY); } EVENTHANDLER_DEREGISTER(tcp_offload_listen_start, listen_start_eh); EVENTHANDLER_DEREGISTER(tcp_offload_listen_stop, listen_stop_eh); EVENTHANDLER_DEREGISTER(lle_event, lle_event_eh); mtx_unlock(&toedev_lock); mtx_destroy(&toedev_lock); return (0); } static int toecore_mod_handler(module_t mod, int cmd, void *arg) { if (cmd == MOD_LOAD) return (toecore_load()); if (cmd == MOD_UNLOAD) return (toecore_unload()); return (EOPNOTSUPP); } static moduledata_t mod_data= { "toecore", toecore_mod_handler, 0 }; MODULE_VERSION(toecore, 1); DECLARE_MODULE(toecore, mod_data, SI_SUB_EXEC, SI_ORDER_ANY); diff --git a/sys/netinet6/in6_pcb.c b/sys/netinet6/in6_pcb.c index 516d83fa846c..176a5227c9a1 100644 --- a/sys/netinet6/in6_pcb.c +++ b/sys/netinet6/in6_pcb.c @@ -1,1142 +1,1139 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (C) 1995, 1996, 1997, and 1998 WIDE Project. * Copyright (c) 2010-2011 Juniper Networks, Inc. * All rights reserved. * * Portions of this software were developed by Robert N. M. Watson under * contract to Juniper Networks, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the project nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE PROJECT AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE PROJECT OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * $KAME: in6_pcb.c,v 1.31 2001/05/21 05:45:10 jinmei Exp $ */ /*- * Copyright (c) 1982, 1986, 1991, 1993 * The Regents of the University of California. All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)in_pcb.c 8.2 (Berkeley) 1/4/94 */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include "opt_ipsec.h" #include "opt_route.h" #include "opt_rss.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include int in6_pcbsetport(struct in6_addr *laddr, struct inpcb *inp, struct ucred *cred) { struct socket *so = inp->inp_socket; u_int16_t lport = 0; int error, lookupflags = 0; #ifdef INVARIANTS struct inpcbinfo *pcbinfo = inp->inp_pcbinfo; #endif INP_WLOCK_ASSERT(inp); INP_HASH_WLOCK_ASSERT(pcbinfo); error = prison_local_ip6(cred, laddr, ((inp->inp_flags & IN6P_IPV6_V6ONLY) != 0)); if (error) return(error); /* XXX: this is redundant when called from in6_pcbbind */ if ((so->so_options & (SO_REUSEADDR|SO_REUSEPORT|SO_REUSEPORT_LB)) == 0) lookupflags = INPLOOKUP_WILDCARD; inp->inp_flags |= INP_ANONPORT; error = in_pcb_lport(inp, NULL, &lport, cred, lookupflags); if (error != 0) return (error); inp->inp_lport = lport; if (in_pcbinshash(inp) != 0) { inp->in6p_laddr = in6addr_any; inp->inp_lport = 0; return (EAGAIN); } return (0); } int in6_pcbbind(struct inpcb *inp, struct sockaddr *nam, struct ucred *cred) { struct socket *so = inp->inp_socket; struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)NULL; struct inpcbinfo *pcbinfo = inp->inp_pcbinfo; u_short lport = 0; int error, lookupflags = 0; int reuseport = (so->so_options & SO_REUSEPORT); /* * XXX: Maybe we could let SO_REUSEPORT_LB set SO_REUSEPORT bit here * so that we don't have to add to the (already messy) code below. */ int reuseport_lb = (so->so_options & SO_REUSEPORT_LB); INP_WLOCK_ASSERT(inp); INP_HASH_WLOCK_ASSERT(pcbinfo); if (inp->inp_lport || !IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) return (EINVAL); if ((so->so_options & (SO_REUSEADDR|SO_REUSEPORT|SO_REUSEPORT_LB)) == 0) lookupflags = INPLOOKUP_WILDCARD; if (nam == NULL) { if ((error = prison_local_ip6(cred, &inp->in6p_laddr, ((inp->inp_flags & IN6P_IPV6_V6ONLY) != 0))) != 0) return (error); } else { sin6 = (struct sockaddr_in6 *)nam; KASSERT(sin6->sin6_family == AF_INET6, ("%s: invalid address family for %p", __func__, sin6)); KASSERT(sin6->sin6_len == sizeof(*sin6), ("%s: invalid address length for %p", __func__, sin6)); if ((error = sa6_embedscope(sin6, V_ip6_use_defzone)) != 0) return(error); if ((error = prison_local_ip6(cred, &sin6->sin6_addr, ((inp->inp_flags & IN6P_IPV6_V6ONLY) != 0))) != 0) return (error); lport = sin6->sin6_port; if (IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr)) { /* * Treat SO_REUSEADDR as SO_REUSEPORT for multicast; * allow compepte duplication of binding if * SO_REUSEPORT is set, or if SO_REUSEADDR is set * and a multicast address is bound on both * new and duplicated sockets. */ if ((so->so_options & (SO_REUSEADDR|SO_REUSEPORT)) != 0) reuseport = SO_REUSEADDR|SO_REUSEPORT; /* * XXX: How to deal with SO_REUSEPORT_LB here? * Treat same as SO_REUSEPORT for now. */ if ((so->so_options & (SO_REUSEADDR|SO_REUSEPORT_LB)) != 0) reuseport_lb = SO_REUSEADDR|SO_REUSEPORT_LB; } else if (!IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) { struct epoch_tracker et; struct ifaddr *ifa; sin6->sin6_port = 0; /* yech... */ NET_EPOCH_ENTER(et); if ((ifa = ifa_ifwithaddr((struct sockaddr *)sin6)) == NULL && (inp->inp_flags & INP_BINDANY) == 0) { NET_EPOCH_EXIT(et); return (EADDRNOTAVAIL); } /* * XXX: bind to an anycast address might accidentally * cause sending a packet with anycast source address. * We should allow to bind to a deprecated address, since * the application dares to use it. */ if (ifa != NULL && ((struct in6_ifaddr *)ifa)->ia6_flags & (IN6_IFF_ANYCAST|IN6_IFF_NOTREADY|IN6_IFF_DETACHED)) { NET_EPOCH_EXIT(et); return (EADDRNOTAVAIL); } NET_EPOCH_EXIT(et); } if (lport) { struct inpcb *t; /* GROSS */ if (ntohs(lport) <= V_ipport_reservedhigh && ntohs(lport) >= V_ipport_reservedlow && priv_check_cred(cred, PRIV_NETINET_RESERVEDPORT)) return (EACCES); if (!IN6_IS_ADDR_MULTICAST(&sin6->sin6_addr) && priv_check_cred(inp->inp_cred, PRIV_NETINET_REUSEPORT) != 0) { t = in6_pcblookup_local(pcbinfo, &sin6->sin6_addr, lport, INPLOOKUP_WILDCARD, cred); if (t && ((inp->inp_flags2 & INP_BINDMULTI) == 0) && - ((t->inp_flags & INP_TIMEWAIT) == 0) && (so->so_type != SOCK_STREAM || IN6_IS_ADDR_UNSPECIFIED(&t->in6p_faddr)) && (!IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr) || !IN6_IS_ADDR_UNSPECIFIED(&t->in6p_laddr) || (t->inp_flags2 & INP_REUSEPORT) || (t->inp_flags2 & INP_REUSEPORT_LB) == 0) && (inp->inp_cred->cr_uid != t->inp_cred->cr_uid)) return (EADDRINUSE); /* * If the socket is a BINDMULTI socket, then * the credentials need to match and the * original socket also has to have been bound * with BINDMULTI. */ if (t && (! in_pcbbind_check_bindmulti(inp, t))) return (EADDRINUSE); #ifdef INET if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0 && IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) { struct sockaddr_in sin; in6_sin6_2_sin(&sin, sin6); t = in_pcblookup_local(pcbinfo, sin.sin_addr, lport, INPLOOKUP_WILDCARD, cred); if (t && ((inp->inp_flags2 & INP_BINDMULTI) == 0) && - ((t->inp_flags & - INP_TIMEWAIT) == 0) && (so->so_type != SOCK_STREAM || ntohl(t->inp_faddr.s_addr) == INADDR_ANY) && (inp->inp_cred->cr_uid != t->inp_cred->cr_uid)) return (EADDRINUSE); if (t && (! in_pcbbind_check_bindmulti(inp, t))) return (EADDRINUSE); } #endif } t = in6_pcblookup_local(pcbinfo, &sin6->sin6_addr, lport, lookupflags, cred); if (t && (reuseport & inp_so_options(t)) == 0 && (reuseport_lb & inp_so_options(t)) == 0) { return (EADDRINUSE); } #ifdef INET if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0 && IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) { struct sockaddr_in sin; in6_sin6_2_sin(&sin, sin6); t = in_pcblookup_local(pcbinfo, sin.sin_addr, lport, lookupflags, cred); if (t && (reuseport & inp_so_options(t)) == 0 && (reuseport_lb & inp_so_options(t)) == 0 && (ntohl(t->inp_laddr.s_addr) != INADDR_ANY || (t->inp_vflag & INP_IPV6PROTO) != 0)) { return (EADDRINUSE); } } #endif } inp->in6p_laddr = sin6->sin6_addr; } if (lport == 0) { if ((error = in6_pcbsetport(&inp->in6p_laddr, inp, cred)) != 0) { /* Undo an address bind that may have occurred. */ inp->in6p_laddr = in6addr_any; return (error); } } else { inp->inp_lport = lport; if (in_pcbinshash(inp) != 0) { inp->in6p_laddr = in6addr_any; inp->inp_lport = 0; return (EAGAIN); } } return (0); } /* * Transform old in6_pcbconnect() into an inner subroutine for new * in6_pcbconnect(): Do some validity-checking on the remote * address (in mbuf 'nam') and then determine local host address * (i.e., which interface) to use to access that remote host. * * This preserves definition of in6_pcbconnect(), while supporting a * slightly different version for T/TCP. (This is more than * a bit of a kludge, but cleaning up the internal interfaces would * have forced minor changes in every protocol). */ static int in6_pcbladdr(struct inpcb *inp, struct sockaddr_in6 *sin6, struct in6_addr *plocal_addr6) { int error = 0; int scope_ambiguous = 0; struct in6_addr in6a; struct epoch_tracker et; INP_WLOCK_ASSERT(inp); INP_HASH_WLOCK_ASSERT(inp->inp_pcbinfo); /* XXXRW: why? */ if (sin6->sin6_port == 0) return (EADDRNOTAVAIL); if (sin6->sin6_scope_id == 0 && !V_ip6_use_defzone) scope_ambiguous = 1; if ((error = sa6_embedscope(sin6, V_ip6_use_defzone)) != 0) return(error); if (!CK_STAILQ_EMPTY(&V_in6_ifaddrhead)) { /* * If the destination address is UNSPECIFIED addr, * use the loopback addr, e.g ::1. */ if (IN6_IS_ADDR_UNSPECIFIED(&sin6->sin6_addr)) sin6->sin6_addr = in6addr_loopback; } if ((error = prison_remote_ip6(inp->inp_cred, &sin6->sin6_addr)) != 0) return (error); NET_EPOCH_ENTER(et); error = in6_selectsrc_socket(sin6, inp->in6p_outputopts, inp, inp->inp_cred, scope_ambiguous, &in6a, NULL); NET_EPOCH_EXIT(et); if (error) return (error); /* * Do not update this earlier, in case we return with an error. * * XXX: this in6_selectsrc_socket result might replace the bound local * address with the address specified by setsockopt(IPV6_PKTINFO). * Is it the intended behavior? */ *plocal_addr6 = in6a; /* * Don't do pcblookup call here; return interface in * plocal_addr6 * and exit to caller, that will do the lookup. */ return (0); } /* * Outer subroutine: * Connect from a socket to a specified address. * Both address and port must be specified in argument sin. * If don't have a local address for this socket yet, * then pick one. */ int in6_pcbconnect_mbuf(struct inpcb *inp, struct sockaddr *nam, struct ucred *cred, struct mbuf *m, bool rehash) { struct inpcbinfo *pcbinfo = inp->inp_pcbinfo; struct sockaddr_in6 *sin6 = (struct sockaddr_in6 *)nam; struct sockaddr_in6 laddr6; int error; KASSERT(sin6->sin6_family == AF_INET6, ("%s: invalid address family for %p", __func__, sin6)); KASSERT(sin6->sin6_len == sizeof(*sin6), ("%s: invalid address length for %p", __func__, sin6)); bzero(&laddr6, sizeof(laddr6)); laddr6.sin6_family = AF_INET6; INP_WLOCK_ASSERT(inp); INP_HASH_WLOCK_ASSERT(pcbinfo); #ifdef ROUTE_MPATH if (CALC_FLOWID_OUTBOUND) { uint32_t hash_type, hash_val; hash_val = fib6_calc_software_hash(&inp->in6p_laddr, &sin6->sin6_addr, 0, sin6->sin6_port, inp->inp_socket->so_proto->pr_protocol, &hash_type); inp->inp_flowid = hash_val; inp->inp_flowtype = hash_type; } #endif /* * Call inner routine, to assign local interface address. * in6_pcbladdr() may automatically fill in sin6_scope_id. */ if ((error = in6_pcbladdr(inp, sin6, &laddr6.sin6_addr)) != 0) return (error); if (in6_pcblookup_hash_locked(pcbinfo, &sin6->sin6_addr, sin6->sin6_port, IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr) ? &laddr6.sin6_addr : &inp->in6p_laddr, inp->inp_lport, 0, NULL, M_NODOM) != NULL) { return (EADDRINUSE); } if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) { if (inp->inp_lport == 0) { /* * rehash was required to be true in the past for * this case; retain that convention. However, * we now call in_pcb_lport_dest rather than * in6_pcbbind; the former does not insert into * the hash table, the latter does. Change rehash * to false to do the in_pcbinshash below. */ KASSERT(rehash == true, ("Rehashing required for unbound inps")); rehash = false; error = in_pcb_lport_dest(inp, (struct sockaddr *) &laddr6, &inp->inp_lport, (struct sockaddr *) sin6, sin6->sin6_port, cred, INPLOOKUP_WILDCARD); if (error) return (error); } inp->in6p_laddr = laddr6.sin6_addr; } inp->in6p_faddr = sin6->sin6_addr; inp->inp_fport = sin6->sin6_port; /* update flowinfo - draft-itojun-ipv6-flowlabel-api-00 */ inp->inp_flow &= ~IPV6_FLOWLABEL_MASK; if (inp->inp_flags & IN6P_AUTOFLOWLABEL) inp->inp_flow |= (htonl(ip6_randomflowlabel()) & IPV6_FLOWLABEL_MASK); if (rehash) { in_pcbrehash(inp); } else { in_pcbinshash(inp); } return (0); } int in6_pcbconnect(struct inpcb *inp, struct sockaddr *nam, struct ucred *cred) { return (in6_pcbconnect_mbuf(inp, nam, cred, NULL, true)); } void in6_pcbdisconnect(struct inpcb *inp) { INP_WLOCK_ASSERT(inp); INP_HASH_WLOCK_ASSERT(inp->inp_pcbinfo); bzero((caddr_t)&inp->in6p_faddr, sizeof(inp->in6p_faddr)); inp->inp_fport = 0; /* clear flowinfo - draft-itojun-ipv6-flowlabel-api-00 */ inp->inp_flow &= ~IPV6_FLOWLABEL_MASK; in_pcbrehash(inp); } struct sockaddr * in6_sockaddr(in_port_t port, struct in6_addr *addr_p) { struct sockaddr_in6 *sin6; sin6 = malloc(sizeof *sin6, M_SONAME, M_WAITOK); bzero(sin6, sizeof *sin6); sin6->sin6_family = AF_INET6; sin6->sin6_len = sizeof(*sin6); sin6->sin6_port = port; sin6->sin6_addr = *addr_p; (void)sa6_recoverscope(sin6); /* XXX: should catch errors */ return (struct sockaddr *)sin6; } struct sockaddr * in6_v4mapsin6_sockaddr(in_port_t port, struct in_addr *addr_p) { struct sockaddr_in sin; struct sockaddr_in6 *sin6_p; bzero(&sin, sizeof sin); sin.sin_family = AF_INET; sin.sin_len = sizeof(sin); sin.sin_port = port; sin.sin_addr = *addr_p; sin6_p = malloc(sizeof *sin6_p, M_SONAME, M_WAITOK); in6_sin_2_v4mapsin6(&sin, sin6_p); return (struct sockaddr *)sin6_p; } int in6_getsockaddr(struct socket *so, struct sockaddr **nam) { struct inpcb *inp; struct in6_addr addr; in_port_t port; inp = sotoinpcb(so); KASSERT(inp != NULL, ("in6_getsockaddr: inp == NULL")); INP_RLOCK(inp); port = inp->inp_lport; addr = inp->in6p_laddr; INP_RUNLOCK(inp); *nam = in6_sockaddr(port, &addr); return 0; } int in6_getpeeraddr(struct socket *so, struct sockaddr **nam) { struct inpcb *inp; struct in6_addr addr; in_port_t port; inp = sotoinpcb(so); KASSERT(inp != NULL, ("in6_getpeeraddr: inp == NULL")); INP_RLOCK(inp); port = inp->inp_fport; addr = inp->in6p_faddr; INP_RUNLOCK(inp); *nam = in6_sockaddr(port, &addr); return 0; } int in6_mapped_sockaddr(struct socket *so, struct sockaddr **nam) { struct inpcb *inp; int error; inp = sotoinpcb(so); KASSERT(inp != NULL, ("in6_mapped_sockaddr: inp == NULL")); #ifdef INET if ((inp->inp_vflag & (INP_IPV4 | INP_IPV6)) == INP_IPV4) { error = in_getsockaddr(so, nam); if (error == 0) in6_sin_2_v4mapsin6_in_sock(nam); } else #endif { /* scope issues will be handled in in6_getsockaddr(). */ error = in6_getsockaddr(so, nam); } return error; } int in6_mapped_peeraddr(struct socket *so, struct sockaddr **nam) { struct inpcb *inp; int error; inp = sotoinpcb(so); KASSERT(inp != NULL, ("in6_mapped_peeraddr: inp == NULL")); #ifdef INET if ((inp->inp_vflag & (INP_IPV4 | INP_IPV6)) == INP_IPV4) { error = in_getpeeraddr(so, nam); if (error == 0) in6_sin_2_v4mapsin6_in_sock(nam); } else #endif /* scope issues will be handled in in6_getpeeraddr(). */ error = in6_getpeeraddr(so, nam); return error; } /* * Pass some notification to all connections of a protocol * associated with address dst. The local address and/or port numbers * may be specified to limit the search. The "usual action" will be * taken, depending on the ctlinput cmd. The caller must filter any * cmds that are uninteresting (e.g., no error in the map). * Call the protocol specific routine (if any) to report * any errors for each matching socket. */ static bool inp_match6(const struct inpcb *inp, void *v __unused) { return ((inp->inp_vflag & INP_IPV6) != 0); } void in6_pcbnotify(struct inpcbinfo *pcbinfo, struct sockaddr_in6 *sa6_dst, u_int fport_arg, const struct sockaddr_in6 *src, u_int lport_arg, int errno, void *cmdarg, struct inpcb *(*notify)(struct inpcb *, int)) { struct inpcb_iterator inpi = INP_ITERATOR(pcbinfo, INPLOOKUP_WLOCKPCB, inp_match6, NULL); struct inpcb *inp; struct sockaddr_in6 sa6_src; u_short fport = fport_arg, lport = lport_arg; u_int32_t flowinfo; if (IN6_IS_ADDR_UNSPECIFIED(&sa6_dst->sin6_addr)) return; /* * note that src can be NULL when we get notify by local fragmentation. */ sa6_src = (src == NULL) ? sa6_any : *src; flowinfo = sa6_src.sin6_flowinfo; while ((inp = inp_next(&inpi)) != NULL) { INP_WLOCK_ASSERT(inp); /* * If the error designates a new path MTU for a destination * and the application (associated with this socket) wanted to * know the value, notify. * XXX: should we avoid to notify the value to TCP sockets? */ if (errno == EMSGSIZE && cmdarg != NULL) ip6_notify_pmtu(inp, sa6_dst, *(uint32_t *)cmdarg); /* * Detect if we should notify the error. If no source and * destination ports are specified, but non-zero flowinfo and * local address match, notify the error. This is the case * when the error is delivered with an encrypted buffer * by ESP. Otherwise, just compare addresses and ports * as usual. */ if (lport == 0 && fport == 0 && flowinfo && inp->inp_socket != NULL && flowinfo == (inp->inp_flow & IPV6_FLOWLABEL_MASK) && IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, &sa6_src.sin6_addr)) goto do_notify; else if (!IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr, &sa6_dst->sin6_addr) || inp->inp_socket == 0 || (lport && inp->inp_lport != lport) || (!IN6_IS_ADDR_UNSPECIFIED(&sa6_src.sin6_addr) && !IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, &sa6_src.sin6_addr)) || (fport && inp->inp_fport != fport)) { continue; } do_notify: if (notify) (*notify)(inp, errno); } } /* * Lookup a PCB based on the local address and port. Caller must hold the * hash lock. No inpcb locks or references are acquired. */ struct inpcb * in6_pcblookup_local(struct inpcbinfo *pcbinfo, struct in6_addr *laddr, u_short lport, int lookupflags, struct ucred *cred) { struct inpcb *inp; int matchwild = 3, wildcard; KASSERT((lookupflags & ~(INPLOOKUP_WILDCARD)) == 0, ("%s: invalid lookup flags %d", __func__, lookupflags)); INP_HASH_LOCK_ASSERT(pcbinfo); if ((lookupflags & INPLOOKUP_WILDCARD) == 0) { struct inpcbhead *head; /* * Look for an unconnected (wildcard foreign addr) PCB that * matches the local address and port we're looking for. */ head = &pcbinfo->ipi_hashbase[INP_PCBHASH_WILD(lport, pcbinfo->ipi_hashmask)]; CK_LIST_FOREACH(inp, head, inp_hash) { /* XXX inp locking */ if ((inp->inp_vflag & INP_IPV6) == 0) continue; if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) && IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) && inp->inp_lport == lport) { /* Found. */ if (cred == NULL || prison_equal_ip6(cred->cr_prison, inp->inp_cred->cr_prison)) return (inp); } } /* * Not found. */ return (NULL); } else { struct inpcbporthead *porthash; struct inpcbport *phd; struct inpcb *match = NULL; /* * Best fit PCB lookup. * * First see if this local port is in use by looking on the * port hash list. */ porthash = &pcbinfo->ipi_porthashbase[INP_PCBPORTHASH(lport, pcbinfo->ipi_porthashmask)]; CK_LIST_FOREACH(phd, porthash, phd_hash) { if (phd->phd_port == lport) break; } if (phd != NULL) { /* * Port is in use by one or more PCBs. Look for best * fit. */ CK_LIST_FOREACH(inp, &phd->phd_pcblist, inp_portlist) { wildcard = 0; if (cred != NULL && !prison_equal_ip6(cred->cr_prison, inp->inp_cred->cr_prison)) continue; /* XXX inp locking */ if ((inp->inp_vflag & INP_IPV6) == 0) continue; if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr)) wildcard++; if (!IN6_IS_ADDR_UNSPECIFIED( &inp->in6p_laddr)) { if (IN6_IS_ADDR_UNSPECIFIED(laddr)) wildcard++; else if (!IN6_ARE_ADDR_EQUAL( &inp->in6p_laddr, laddr)) continue; } else { if (!IN6_IS_ADDR_UNSPECIFIED(laddr)) wildcard++; } if (wildcard < matchwild) { match = inp; matchwild = wildcard; if (matchwild == 0) break; } } } return (match); } } static bool in6_multi_match(const struct inpcb *inp, void *v __unused) { if ((inp->inp_vflag & INP_IPV6) && inp->in6p_moptions != NULL) return (true); else return (false); } void in6_pcbpurgeif0(struct inpcbinfo *pcbinfo, struct ifnet *ifp) { struct inpcb_iterator inpi = INP_ITERATOR(pcbinfo, INPLOOKUP_RLOCKPCB, in6_multi_match, NULL); struct inpcb *inp; struct in6_multi *inm; struct in6_mfilter *imf; struct ip6_moptions *im6o; IN6_MULTI_LOCK_ASSERT(); while ((inp = inp_next(&inpi)) != NULL) { INP_RLOCK_ASSERT(inp); im6o = inp->in6p_moptions; /* * Unselect the outgoing ifp for multicast if it * is being detached. */ if (im6o->im6o_multicast_ifp == ifp) im6o->im6o_multicast_ifp = NULL; /* * Drop multicast group membership if we joined * through the interface being detached. */ restart: IP6_MFILTER_FOREACH(imf, &im6o->im6o_head) { if ((inm = imf->im6f_in6m) == NULL) continue; if (inm->in6m_ifp != ifp) continue; ip6_mfilter_remove(&im6o->im6o_head, imf); in6_leavegroup_locked(inm, NULL); ip6_mfilter_free(imf); goto restart; } } } /* * Check for alternatives when higher level complains * about service problems. For now, invalidate cached * routing information. If the route was created dynamically * (by a redirect), time to try a default gateway again. */ void in6_losing(struct inpcb *inp) { RO_INVALIDATE_CACHE(&inp->inp_route6); } /* * After a routing change, flush old routing * and allocate a (hopefully) better one. */ struct inpcb * in6_rtchange(struct inpcb *inp, int errno __unused) { RO_INVALIDATE_CACHE(&inp->inp_route6); return inp; } static struct inpcb * in6_pcblookup_lbgroup(const struct inpcbinfo *pcbinfo, const struct in6_addr *laddr, uint16_t lport, const struct in6_addr *faddr, uint16_t fport, int lookupflags, uint8_t numa_domain) { struct inpcb *local_wild, *numa_wild; const struct inpcblbgrouphead *hdr; struct inpcblbgroup *grp; uint32_t idx; INP_HASH_LOCK_ASSERT(pcbinfo); hdr = &pcbinfo->ipi_lbgrouphashbase[ INP_PCBPORTHASH(lport, pcbinfo->ipi_lbgrouphashmask)]; /* * Order of socket selection: * 1. non-wild. * 2. wild (if lookupflags contains INPLOOKUP_WILDCARD). * * NOTE: * - Load balanced group does not contain jailed sockets. * - Load balanced does not contain IPv4 mapped INET6 wild sockets. */ local_wild = NULL; numa_wild = NULL; CK_LIST_FOREACH(grp, hdr, il_list) { #ifdef INET if (!(grp->il_vflag & INP_IPV6)) continue; #endif if (grp->il_lport != lport) continue; idx = INP6_PCBLBGROUP_PKTHASH(faddr, lport, fport) % grp->il_inpcnt; if (IN6_ARE_ADDR_EQUAL(&grp->il6_laddr, laddr)) { if (numa_domain == M_NODOM || grp->il_numa_domain == numa_domain) { return (grp->il_inp[idx]); } else numa_wild = grp->il_inp[idx]; } if (IN6_IS_ADDR_UNSPECIFIED(&grp->il6_laddr) && (lookupflags & INPLOOKUP_WILDCARD) != 0 && (local_wild == NULL || numa_domain == M_NODOM || grp->il_numa_domain == numa_domain)) { local_wild = grp->il_inp[idx]; } } if (numa_wild != NULL) return (numa_wild); return (local_wild); } /* * Lookup PCB in hash list. Used in in_pcb.c as well as here. */ struct inpcb * in6_pcblookup_hash_locked(struct inpcbinfo *pcbinfo, struct in6_addr *faddr, u_int fport_arg, struct in6_addr *laddr, u_int lport_arg, int lookupflags, struct ifnet *ifp, uint8_t numa_domain) { struct inpcbhead *head; struct inpcb *inp, *tmpinp; u_short fport = fport_arg, lport = lport_arg; KASSERT((lookupflags & ~(INPLOOKUP_WILDCARD)) == 0, ("%s: invalid lookup flags %d", __func__, lookupflags)); INP_HASH_LOCK_ASSERT(pcbinfo); /* * First look for an exact match. */ tmpinp = NULL; head = &pcbinfo->ipi_hashbase[INP6_PCBHASH(faddr, lport, fport, pcbinfo->ipi_hashmask)]; CK_LIST_FOREACH(inp, head, inp_hash) { /* XXX inp locking */ if ((inp->inp_vflag & INP_IPV6) == 0) continue; if (IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr, faddr) && IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr) && inp->inp_fport == fport && inp->inp_lport == lport) { /* * XXX We should be able to directly return * the inp here, without any checks. * Well unless both bound with SO_REUSEPORT? */ if (prison_flag(inp->inp_cred, PR_IP6)) return (inp); if (tmpinp == NULL) tmpinp = inp; } } if (tmpinp != NULL) return (tmpinp); /* * Then look in lb group (for wildcard match). */ if ((lookupflags & INPLOOKUP_WILDCARD) != 0) { inp = in6_pcblookup_lbgroup(pcbinfo, laddr, lport, faddr, fport, lookupflags, numa_domain); if (inp != NULL) return (inp); } /* * Then look for a wildcard match, if requested. */ if ((lookupflags & INPLOOKUP_WILDCARD) != 0) { struct inpcb *local_wild = NULL, *local_exact = NULL; struct inpcb *jail_wild = NULL; int injail; /* * Order of socket selection - we always prefer jails. * 1. jailed, non-wild. * 2. jailed, wild. * 3. non-jailed, non-wild. * 4. non-jailed, wild. */ head = &pcbinfo->ipi_hashbase[INP_PCBHASH_WILD(lport, pcbinfo->ipi_hashmask)]; CK_LIST_FOREACH(inp, head, inp_hash) { /* XXX inp locking */ if ((inp->inp_vflag & INP_IPV6) == 0) continue; if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr) || inp->inp_lport != lport) { continue; } injail = prison_flag(inp->inp_cred, PR_IP6); if (injail) { if (prison_check_ip6_locked( inp->inp_cred->cr_prison, laddr) != 0) continue; } else { if (local_exact != NULL) continue; } if (IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, laddr)) { if (injail) return (inp); else local_exact = inp; } else if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) { if (injail) jail_wild = inp; else local_wild = inp; } } /* LIST_FOREACH */ if (jail_wild != NULL) return (jail_wild); if (local_exact != NULL) return (local_exact); if (local_wild != NULL) return (local_wild); } /* if ((lookupflags & INPLOOKUP_WILDCARD) != 0) */ /* * Not found. */ return (NULL); } /* * Lookup PCB in hash list, using pcbinfo tables. This variation locks the * hash list lock, and will return the inpcb locked (i.e., requires * INPLOOKUP_LOCKPCB). */ static struct inpcb * in6_pcblookup_hash(struct inpcbinfo *pcbinfo, struct in6_addr *faddr, u_int fport, struct in6_addr *laddr, u_int lport, int lookupflags, struct ifnet *ifp, uint8_t numa_domain) { struct inpcb *inp; smr_enter(pcbinfo->ipi_smr); inp = in6_pcblookup_hash_locked(pcbinfo, faddr, fport, laddr, lport, lookupflags & INPLOOKUP_WILDCARD, ifp, numa_domain); if (inp != NULL) { if (__predict_false(inp_smr_lock(inp, (lookupflags & INPLOOKUP_LOCKMASK)) == false)) inp = NULL; } else smr_exit(pcbinfo->ipi_smr); return (inp); } /* * Public inpcb lookup routines, accepting a 4-tuple, and optionally, an mbuf * from which a pre-calculated hash value may be extracted. */ struct inpcb * in6_pcblookup(struct inpcbinfo *pcbinfo, struct in6_addr *faddr, u_int fport, struct in6_addr *laddr, u_int lport, int lookupflags, struct ifnet *ifp) { KASSERT((lookupflags & ~INPLOOKUP_MASK) == 0, ("%s: invalid lookup flags %d", __func__, lookupflags)); KASSERT((lookupflags & (INPLOOKUP_RLOCKPCB | INPLOOKUP_WLOCKPCB)) != 0, ("%s: LOCKPCB not set", __func__)); return (in6_pcblookup_hash(pcbinfo, faddr, fport, laddr, lport, lookupflags, ifp, M_NODOM)); } struct inpcb * in6_pcblookup_mbuf(struct inpcbinfo *pcbinfo, struct in6_addr *faddr, u_int fport, struct in6_addr *laddr, u_int lport, int lookupflags, struct ifnet *ifp, struct mbuf *m) { KASSERT((lookupflags & ~INPLOOKUP_MASK) == 0, ("%s: invalid lookup flags %d", __func__, lookupflags)); KASSERT((lookupflags & (INPLOOKUP_RLOCKPCB | INPLOOKUP_WLOCKPCB)) != 0, ("%s: LOCKPCB not set", __func__)); return (in6_pcblookup_hash(pcbinfo, faddr, fport, laddr, lport, lookupflags, ifp, m->m_pkthdr.numa_domain)); } void init_sin6(struct sockaddr_in6 *sin6, struct mbuf *m, int srcordst) { struct ip6_hdr *ip; ip = mtod(m, struct ip6_hdr *); bzero(sin6, sizeof(*sin6)); sin6->sin6_len = sizeof(*sin6); sin6->sin6_family = AF_INET6; sin6->sin6_addr = srcordst ? ip->ip6_dst : ip->ip6_src; (void)sa6_recoverscope(sin6); /* XXX: should catch errors... */ return; } diff --git a/sys/netinet6/ip6_output.c b/sys/netinet6/ip6_output.c index b976b0583515..f00f718445f5 100644 --- a/sys/netinet6/ip6_output.c +++ b/sys/netinet6/ip6_output.c @@ -1,3387 +1,3387 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (C) 1995, 1996, 1997, and 1998 WIDE Project. * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the project nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE PROJECT AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE PROJECT OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * $KAME: ip6_output.c,v 1.279 2002/01/26 06:12:30 jinmei Exp $ */ /*- * Copyright (c) 1982, 1986, 1988, 1990, 1993 * The Regents of the University of California. All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)ip_output.c 8.3 (Berkeley) 1/21/94 */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include "opt_ipsec.h" #include "opt_kern_tls.h" #include "opt_ratelimit.h" #include "opt_route.h" #include "opt_rss.h" #include "opt_sctp.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #if defined(SCTP) || defined(SCTP_SUPPORT) #include #include #endif #include extern int in6_mcast_loop; struct ip6_exthdrs { struct mbuf *ip6e_ip6; struct mbuf *ip6e_hbh; struct mbuf *ip6e_dest1; struct mbuf *ip6e_rthdr; struct mbuf *ip6e_dest2; }; static MALLOC_DEFINE(M_IP6OPT, "ip6opt", "IPv6 options"); static int ip6_pcbopt(int, u_char *, int, struct ip6_pktopts **, struct ucred *, int); static int ip6_pcbopts(struct ip6_pktopts **, struct mbuf *, struct socket *, struct sockopt *); static int ip6_getpcbopt(struct inpcb *, int, struct sockopt *); static int ip6_setpktopt(int, u_char *, int, struct ip6_pktopts *, struct ucred *, int, int, int); static int ip6_copyexthdr(struct mbuf **, caddr_t, int); static int ip6_insertfraghdr(struct mbuf *, struct mbuf *, int, struct ip6_frag **); static int ip6_insert_jumboopt(struct ip6_exthdrs *, u_int32_t); static int ip6_splithdr(struct mbuf *, struct ip6_exthdrs *); static int ip6_getpmtu(struct route_in6 *, int, struct ifnet *, const struct in6_addr *, u_long *, int *, u_int, u_int); static int ip6_calcmtu(struct ifnet *, const struct in6_addr *, u_long, u_long *, int *, u_int); static int ip6_getpmtu_ctl(u_int, const struct in6_addr *, u_long *); static int copypktopts(struct ip6_pktopts *, struct ip6_pktopts *, int); /* * Make an extension header from option data. hp is the source, * mp is the destination, and _ol is the optlen. */ #define MAKE_EXTHDR(hp, mp, _ol) \ do { \ if (hp) { \ struct ip6_ext *eh = (struct ip6_ext *)(hp); \ error = ip6_copyexthdr((mp), (caddr_t)(hp), \ ((eh)->ip6e_len + 1) << 3); \ if (error) \ goto freehdrs; \ (_ol) += (*(mp))->m_len; \ } \ } while (/*CONSTCOND*/ 0) /* * Form a chain of extension headers. * m is the extension header mbuf * mp is the previous mbuf in the chain * p is the next header * i is the type of option. */ #define MAKE_CHAIN(m, mp, p, i)\ do {\ if (m) {\ if (!hdrsplit) \ panic("%s:%d: assumption failed: "\ "hdr not split: hdrsplit %d exthdrs %p",\ __func__, __LINE__, hdrsplit, &exthdrs);\ *mtod((m), u_char *) = *(p);\ *(p) = (i);\ p = mtod((m), u_char *);\ (m)->m_next = (mp)->m_next;\ (mp)->m_next = (m);\ (mp) = (m);\ }\ } while (/*CONSTCOND*/ 0) void in6_delayed_cksum(struct mbuf *m, uint32_t plen, u_short offset) { u_short csum; csum = in_cksum_skip(m, offset + plen, offset); if (m->m_pkthdr.csum_flags & CSUM_UDP_IPV6 && csum == 0) csum = 0xffff; offset += m->m_pkthdr.csum_data; /* checksum offset */ if (offset + sizeof(csum) > m->m_len) m_copyback(m, offset, sizeof(csum), (caddr_t)&csum); else *(u_short *)mtodo(m, offset) = csum; } static void ip6_output_delayed_csum(struct mbuf *m, struct ifnet *ifp, int csum_flags, int plen, int optlen) { KASSERT((plen >= optlen), ("%s:%d: plen %d < optlen %d, m %p, ifp %p " "csum_flags %#x", __func__, __LINE__, plen, optlen, m, ifp, csum_flags)); if (csum_flags & CSUM_DELAY_DATA_IPV6) { in6_delayed_cksum(m, plen - optlen, sizeof(struct ip6_hdr) + optlen); m->m_pkthdr.csum_flags &= ~CSUM_DELAY_DATA_IPV6; } #if defined(SCTP) || defined(SCTP_SUPPORT) if (csum_flags & CSUM_SCTP_IPV6) { sctp_delayed_cksum(m, sizeof(struct ip6_hdr) + optlen); m->m_pkthdr.csum_flags &= ~CSUM_SCTP_IPV6; } #endif } int ip6_fragment(struct ifnet *ifp, struct mbuf *m0, int hlen, u_char nextproto, int fraglen , uint32_t id) { struct mbuf *m, **mnext, *m_frgpart; struct ip6_hdr *ip6, *mhip6; struct ip6_frag *ip6f; int off; int error; int tlen = m0->m_pkthdr.len; KASSERT((fraglen % 8 == 0), ("Fragment length must be a multiple of 8")); m = m0; ip6 = mtod(m, struct ip6_hdr *); mnext = &m->m_nextpkt; for (off = hlen; off < tlen; off += fraglen) { m = m_gethdr(M_NOWAIT, MT_DATA); if (!m) { IP6STAT_INC(ip6s_odropped); return (ENOBUFS); } /* * Make sure the complete packet header gets copied * from the originating mbuf to the newly created * mbuf. This also ensures that existing firewall * classification(s), VLAN tags and so on get copied * to the resulting fragmented packet(s): */ if (m_dup_pkthdr(m, m0, M_NOWAIT) == 0) { m_free(m); IP6STAT_INC(ip6s_odropped); return (ENOBUFS); } *mnext = m; mnext = &m->m_nextpkt; m->m_data += max_linkhdr; mhip6 = mtod(m, struct ip6_hdr *); *mhip6 = *ip6; m->m_len = sizeof(*mhip6); error = ip6_insertfraghdr(m0, m, hlen, &ip6f); if (error) { IP6STAT_INC(ip6s_odropped); return (error); } ip6f->ip6f_offlg = htons((u_short)((off - hlen) & ~7)); if (off + fraglen >= tlen) fraglen = tlen - off; else ip6f->ip6f_offlg |= IP6F_MORE_FRAG; mhip6->ip6_plen = htons((u_short)(fraglen + hlen + sizeof(*ip6f) - sizeof(struct ip6_hdr))); if ((m_frgpart = m_copym(m0, off, fraglen, M_NOWAIT)) == NULL) { IP6STAT_INC(ip6s_odropped); return (ENOBUFS); } m_cat(m, m_frgpart); m->m_pkthdr.len = fraglen + hlen + sizeof(*ip6f); ip6f->ip6f_reserved = 0; ip6f->ip6f_ident = id; ip6f->ip6f_nxt = nextproto; IP6STAT_INC(ip6s_ofragments); in6_ifstat_inc(ifp, ifs6_out_fragcreat); } return (0); } static int ip6_output_send(struct inpcb *inp, struct ifnet *ifp, struct ifnet *origifp, struct mbuf *m, struct sockaddr_in6 *dst, struct route_in6 *ro, bool stamp_tag) { #ifdef KERN_TLS struct ktls_session *tls = NULL; #endif struct m_snd_tag *mst; int error; MPASS((m->m_pkthdr.csum_flags & CSUM_SND_TAG) == 0); mst = NULL; #ifdef KERN_TLS /* * If this is an unencrypted TLS record, save a reference to * the record. This local reference is used to call * ktls_output_eagain after the mbuf has been freed (thus * dropping the mbuf's reference) in if_output. */ if (m->m_next != NULL && mbuf_has_tls_session(m->m_next)) { tls = ktls_hold(m->m_next->m_epg_tls); mst = tls->snd_tag; /* * If a TLS session doesn't have a valid tag, it must * have had an earlier ifp mismatch, so drop this * packet. */ if (mst == NULL) { m_freem(m); error = EAGAIN; goto done; } /* * Always stamp tags that include NIC ktls. */ stamp_tag = true; } #endif #ifdef RATELIMIT if (inp != NULL && mst == NULL) { if ((inp->inp_flags2 & INP_RATE_LIMIT_CHANGED) != 0 || (inp->inp_snd_tag != NULL && inp->inp_snd_tag->ifp != ifp)) in_pcboutput_txrtlmt(inp, ifp, m); if (inp->inp_snd_tag != NULL) mst = inp->inp_snd_tag; } #endif if (stamp_tag && mst != NULL) { KASSERT(m->m_pkthdr.rcvif == NULL, ("trying to add a send tag to a forwarded packet")); if (mst->ifp != ifp) { m_freem(m); error = EAGAIN; goto done; } /* stamp send tag on mbuf */ m->m_pkthdr.snd_tag = m_snd_tag_ref(mst); m->m_pkthdr.csum_flags |= CSUM_SND_TAG; } error = nd6_output_ifp(ifp, origifp, m, dst, (struct route *)ro); done: /* Check for route change invalidating send tags. */ #ifdef KERN_TLS if (tls != NULL) { if (error == EAGAIN) error = ktls_output_eagain(inp, tls); ktls_free(tls); } #endif #ifdef RATELIMIT if (error == EAGAIN) in_pcboutput_eagain(inp); #endif return (error); } /* * IP6 output. * The packet in mbuf chain m contains a skeletal IP6 header (with pri, len, * nxt, hlim, src, dst). * This function may modify ver and hlim only. * The mbuf chain containing the packet will be freed. * The mbuf opt, if present, will not be freed. * If route_in6 ro is present and has ro_nh initialized, route lookup would be * skipped and ro->ro_nh would be used. If ro is present but ro->ro_nh is NULL, * then result of route lookup is stored in ro->ro_nh. * * Type of "mtu": rt_mtu is u_long, ifnet.ifr_mtu is int, and nd_ifinfo.linkmtu * is uint32_t. So we use u_long to hold largest one, which is rt_mtu. * * ifpp - XXX: just for statistics */ int ip6_output(struct mbuf *m0, struct ip6_pktopts *opt, struct route_in6 *ro, int flags, struct ip6_moptions *im6o, struct ifnet **ifpp, struct inpcb *inp) { struct ip6_hdr *ip6; struct ifnet *ifp, *origifp; struct mbuf *m = m0; struct mbuf *mprev; struct route_in6 *ro_pmtu; struct nhop_object *nh; struct sockaddr_in6 *dst, sin6, src_sa, dst_sa; struct in6_addr odst; u_char *nexthdrp; int tlen, len; int error = 0; int vlan_pcp = -1; struct in6_ifaddr *ia = NULL; u_long mtu; int alwaysfrag, dontfrag; u_int32_t optlen, plen = 0, unfragpartlen; struct ip6_exthdrs exthdrs; struct in6_addr src0, dst0; u_int32_t zone; bool hdrsplit; int sw_csum, tso; int needfiblookup; uint32_t fibnum; struct m_tag *fwd_tag = NULL; uint32_t id; NET_EPOCH_ASSERT(); if (inp != NULL) { INP_LOCK_ASSERT(inp); M_SETFIB(m, inp->inp_inc.inc_fibnum); if ((flags & IP_NODEFAULTFLOWID) == 0) { /* Unconditionally set flowid. */ m->m_pkthdr.flowid = inp->inp_flowid; M_HASHTYPE_SET(m, inp->inp_flowtype); } if ((inp->inp_flags2 & INP_2PCP_SET) != 0) vlan_pcp = (inp->inp_flags2 & INP_2PCP_MASK) >> INP_2PCP_SHIFT; #ifdef NUMA m->m_pkthdr.numa_domain = inp->inp_numa_domain; #endif } #if defined(IPSEC) || defined(IPSEC_SUPPORT) /* * IPSec checking which handles several cases. * FAST IPSEC: We re-injected the packet. * XXX: need scope argument. */ if (IPSEC_ENABLED(ipv6)) { if ((error = IPSEC_OUTPUT(ipv6, m, inp)) != 0) { if (error == EINPROGRESS) error = 0; goto done; } } #endif /* IPSEC */ /* Source address validation. */ ip6 = mtod(m, struct ip6_hdr *); if (IN6_IS_ADDR_UNSPECIFIED(&ip6->ip6_src) && (flags & IPV6_UNSPECSRC) == 0) { error = EOPNOTSUPP; IP6STAT_INC(ip6s_badscope); goto bad; } if (IN6_IS_ADDR_MULTICAST(&ip6->ip6_src)) { error = EOPNOTSUPP; IP6STAT_INC(ip6s_badscope); goto bad; } /* * If we are given packet options to add extension headers prepare them. * Calculate the total length of the extension header chain. * Keep the length of the unfragmentable part for fragmentation. */ bzero(&exthdrs, sizeof(exthdrs)); optlen = 0; unfragpartlen = sizeof(struct ip6_hdr); if (opt) { /* Hop-by-Hop options header. */ MAKE_EXTHDR(opt->ip6po_hbh, &exthdrs.ip6e_hbh, optlen); /* Destination options header (1st part). */ if (opt->ip6po_rthdr) { #ifndef RTHDR_SUPPORT_IMPLEMENTED /* * If there is a routing header, discard the packet * right away here. RH0/1 are obsolete and we do not * currently support RH2/3/4. * People trying to use RH253/254 may want to disable * this check. * The moment we do support any routing header (again) * this block should check the routing type more * selectively. */ error = EINVAL; goto bad; #endif /* * Destination options header (1st part). * This only makes sense with a routing header. * See Section 9.2 of RFC 3542. * Disabling this part just for MIP6 convenience is * a bad idea. We need to think carefully about a * way to make the advanced API coexist with MIP6 * options, which might automatically be inserted in * the kernel. */ MAKE_EXTHDR(opt->ip6po_dest1, &exthdrs.ip6e_dest1, optlen); } /* Routing header. */ MAKE_EXTHDR(opt->ip6po_rthdr, &exthdrs.ip6e_rthdr, optlen); unfragpartlen += optlen; /* * NOTE: we don't add AH/ESP length here (done in * ip6_ipsec_output()). */ /* Destination options header (2nd part). */ MAKE_EXTHDR(opt->ip6po_dest2, &exthdrs.ip6e_dest2, optlen); } /* * If there is at least one extension header, * separate IP6 header from the payload. */ hdrsplit = false; if (optlen) { if ((error = ip6_splithdr(m, &exthdrs)) != 0) { m = NULL; goto freehdrs; } m = exthdrs.ip6e_ip6; ip6 = mtod(m, struct ip6_hdr *); hdrsplit = true; } /* Adjust mbuf packet header length. */ m->m_pkthdr.len += optlen; plen = m->m_pkthdr.len - sizeof(*ip6); /* If this is a jumbo payload, insert a jumbo payload option. */ if (plen > IPV6_MAXPACKET) { if (!hdrsplit) { if ((error = ip6_splithdr(m, &exthdrs)) != 0) { m = NULL; goto freehdrs; } m = exthdrs.ip6e_ip6; ip6 = mtod(m, struct ip6_hdr *); hdrsplit = true; } if ((error = ip6_insert_jumboopt(&exthdrs, plen)) != 0) goto freehdrs; ip6->ip6_plen = 0; } else ip6->ip6_plen = htons(plen); nexthdrp = &ip6->ip6_nxt; if (optlen) { /* * Concatenate headers and fill in next header fields. * Here we have, on "m" * IPv6 payload * and we insert headers accordingly. * Finally, we should be getting: * IPv6 hbh dest1 rthdr ah* [esp* dest2 payload]. * * During the header composing process "m" points to IPv6 * header. "mprev" points to an extension header prior to esp. */ mprev = m; /* * We treat dest2 specially. This makes IPsec processing * much easier. The goal here is to make mprev point the * mbuf prior to dest2. * * Result: IPv6 dest2 payload. * m and mprev will point to IPv6 header. */ if (exthdrs.ip6e_dest2) { if (!hdrsplit) panic("%s:%d: assumption failed: " "hdr not split: hdrsplit %d exthdrs %p", __func__, __LINE__, hdrsplit, &exthdrs); exthdrs.ip6e_dest2->m_next = m->m_next; m->m_next = exthdrs.ip6e_dest2; *mtod(exthdrs.ip6e_dest2, u_char *) = ip6->ip6_nxt; ip6->ip6_nxt = IPPROTO_DSTOPTS; } /* * Result: IPv6 hbh dest1 rthdr dest2 payload. * m will point to IPv6 header. mprev will point to the * extension header prior to dest2 (rthdr in the above case). */ MAKE_CHAIN(exthdrs.ip6e_hbh, mprev, nexthdrp, IPPROTO_HOPOPTS); MAKE_CHAIN(exthdrs.ip6e_dest1, mprev, nexthdrp, IPPROTO_DSTOPTS); MAKE_CHAIN(exthdrs.ip6e_rthdr, mprev, nexthdrp, IPPROTO_ROUTING); } IP6STAT_INC(ip6s_localout); /* Route packet. */ ro_pmtu = ro; if (opt && opt->ip6po_rthdr) ro = &opt->ip6po_route; if (ro != NULL) dst = (struct sockaddr_in6 *)&ro->ro_dst; else dst = &sin6; fibnum = (inp != NULL) ? inp->inp_inc.inc_fibnum : M_GETFIB(m); again: /* * If specified, try to fill in the traffic class field. * Do not override if a non-zero value is already set. * We check the diffserv field and the ECN field separately. */ if (opt && opt->ip6po_tclass >= 0) { int mask = 0; if (IPV6_DSCP(ip6) == 0) mask |= 0xfc; if (IPV6_ECN(ip6) == 0) mask |= 0x03; if (mask != 0) ip6->ip6_flow |= htonl((opt->ip6po_tclass & mask) << 20); } /* Fill in or override the hop limit field, if necessary. */ if (opt && opt->ip6po_hlim != -1) ip6->ip6_hlim = opt->ip6po_hlim & 0xff; else if (IN6_IS_ADDR_MULTICAST(&ip6->ip6_dst)) { if (im6o != NULL) ip6->ip6_hlim = im6o->im6o_multicast_hlim; else ip6->ip6_hlim = V_ip6_defmcasthlim; } if (ro == NULL || ro->ro_nh == NULL) { bzero(dst, sizeof(*dst)); dst->sin6_family = AF_INET6; dst->sin6_len = sizeof(*dst); dst->sin6_addr = ip6->ip6_dst; } /* * Validate route against routing table changes. * Make sure that the address family is set in route. */ nh = NULL; ifp = NULL; mtu = 0; if (ro != NULL) { if (ro->ro_nh != NULL && inp != NULL) { ro->ro_dst.sin6_family = AF_INET6; /* XXX KASSERT? */ NH_VALIDATE((struct route *)ro, &inp->inp_rt_cookie, fibnum); } if (ro->ro_nh != NULL && fwd_tag == NULL && (!NH_IS_VALID(ro->ro_nh) || ro->ro_dst.sin6_family != AF_INET6 || !IN6_ARE_ADDR_EQUAL(&ro->ro_dst.sin6_addr, &ip6->ip6_dst))) RO_INVALIDATE_CACHE(ro); if (ro->ro_nh != NULL && fwd_tag == NULL && ro->ro_dst.sin6_family == AF_INET6 && IN6_ARE_ADDR_EQUAL(&ro->ro_dst.sin6_addr, &ip6->ip6_dst)) { /* Nexthop is valid and contains valid ifp */ nh = ro->ro_nh; } else { if (ro->ro_lle) LLE_FREE(ro->ro_lle); /* zeros ro_lle */ ro->ro_lle = NULL; if (fwd_tag == NULL) { bzero(&dst_sa, sizeof(dst_sa)); dst_sa.sin6_family = AF_INET6; dst_sa.sin6_len = sizeof(dst_sa); dst_sa.sin6_addr = ip6->ip6_dst; } error = in6_selectroute(&dst_sa, opt, im6o, ro, &ifp, &nh, fibnum, m->m_pkthdr.flowid); if (error != 0) { IP6STAT_INC(ip6s_noroute); if (ifp != NULL) in6_ifstat_inc(ifp, ifs6_out_discard); goto bad; } /* * At this point at least @ifp is not NULL * Can be the case when dst is multicast, link-local or * interface is explicitly specificed by the caller. */ } if (nh == NULL) { /* * If in6_selectroute() does not return a nexthop * dst may not have been updated. */ *dst = dst_sa; /* XXX */ origifp = ifp; mtu = ifp->if_mtu; } else { ifp = nh->nh_ifp; origifp = nh->nh_aifp; ia = (struct in6_ifaddr *)(nh->nh_ifa); counter_u64_add(nh->nh_pksent, 1); } } else { struct nhop_object *nh; struct in6_addr kdst; uint32_t scopeid; if (fwd_tag == NULL) { bzero(&dst_sa, sizeof(dst_sa)); dst_sa.sin6_family = AF_INET6; dst_sa.sin6_len = sizeof(dst_sa); dst_sa.sin6_addr = ip6->ip6_dst; } if (IN6_IS_ADDR_MULTICAST(&dst_sa.sin6_addr) && im6o != NULL && (ifp = im6o->im6o_multicast_ifp) != NULL) { /* We do not need a route lookup. */ *dst = dst_sa; /* XXX */ origifp = ifp; goto nonh6lookup; } in6_splitscope(&dst_sa.sin6_addr, &kdst, &scopeid); if (IN6_IS_ADDR_MC_LINKLOCAL(&dst_sa.sin6_addr) || IN6_IS_ADDR_MC_NODELOCAL(&dst_sa.sin6_addr)) { if (scopeid > 0) { ifp = in6_getlinkifnet(scopeid); if (ifp == NULL) { error = EHOSTUNREACH; goto bad; } *dst = dst_sa; /* XXX */ origifp = ifp; goto nonh6lookup; } } nh = fib6_lookup(fibnum, &kdst, scopeid, NHR_NONE, m->m_pkthdr.flowid); if (nh == NULL) { IP6STAT_INC(ip6s_noroute); /* No ifp in6_ifstat_inc(ifp, ifs6_out_discard); */ error = EHOSTUNREACH;; goto bad; } ifp = nh->nh_ifp; origifp = nh->nh_aifp; ia = ifatoia6(nh->nh_ifa); if (nh->nh_flags & NHF_GATEWAY) dst->sin6_addr = nh->gw6_sa.sin6_addr; else if (fwd_tag != NULL) dst->sin6_addr = dst_sa.sin6_addr; nonh6lookup: ; } /* * At this point ifp MUST be pointing to the valid transmit ifp. * origifp MUST be valid and pointing to either the same ifp or, * in case of loopback output, to the interface which ip6_src * belongs to. * Examples: * fe80::1%em0 -> fe80::2%em0 -> ifp=em0, origifp=em0 * fe80::1%em0 -> fe80::1%em0 -> ifp=lo0, origifp=em0 * ::1 -> ::1 -> ifp=lo0, origifp=lo0 * * mtu can be 0 and will be refined later. */ KASSERT((ifp != NULL), ("output interface must not be NULL")); KASSERT((origifp != NULL), ("output address interface must not be NULL")); if ((flags & IPV6_FORWARDING) == 0) { /* XXX: the FORWARDING flag can be set for mrouting. */ in6_ifstat_inc(ifp, ifs6_out_request); } /* Setup data structures for scope ID checks. */ src0 = ip6->ip6_src; bzero(&src_sa, sizeof(src_sa)); src_sa.sin6_family = AF_INET6; src_sa.sin6_len = sizeof(src_sa); src_sa.sin6_addr = ip6->ip6_src; dst0 = ip6->ip6_dst; /* Re-initialize to be sure. */ bzero(&dst_sa, sizeof(dst_sa)); dst_sa.sin6_family = AF_INET6; dst_sa.sin6_len = sizeof(dst_sa); dst_sa.sin6_addr = ip6->ip6_dst; /* Check for valid scope ID. */ if (in6_setscope(&src0, origifp, &zone) == 0 && sa6_recoverscope(&src_sa) == 0 && zone == src_sa.sin6_scope_id && in6_setscope(&dst0, origifp, &zone) == 0 && sa6_recoverscope(&dst_sa) == 0 && zone == dst_sa.sin6_scope_id) { /* * The outgoing interface is in the zone of the source * and destination addresses. * */ } else if ((origifp->if_flags & IFF_LOOPBACK) == 0 || sa6_recoverscope(&src_sa) != 0 || sa6_recoverscope(&dst_sa) != 0 || dst_sa.sin6_scope_id == 0 || (src_sa.sin6_scope_id != 0 && src_sa.sin6_scope_id != dst_sa.sin6_scope_id) || ifnet_byindex(dst_sa.sin6_scope_id) == NULL) { /* * If the destination network interface is not a * loopback interface, or the destination network * address has no scope ID, or the source address has * a scope ID set which is different from the * destination address one, or there is no network * interface representing this scope ID, the address * pair is considered invalid. */ IP6STAT_INC(ip6s_badscope); in6_ifstat_inc(origifp, ifs6_out_discard); if (error == 0) error = EHOSTUNREACH; /* XXX */ goto bad; } /* All scope ID checks are successful. */ if (nh && !IN6_IS_ADDR_MULTICAST(&ip6->ip6_dst)) { if (opt && opt->ip6po_nextroute.ro_nh) { /* * The nexthop is explicitly specified by the * application. We assume the next hop is an IPv6 * address. */ dst = (struct sockaddr_in6 *)opt->ip6po_nexthop; } else if ((nh->nh_flags & NHF_GATEWAY)) dst = &nh->gw6_sa; } if (!IN6_IS_ADDR_MULTICAST(&ip6->ip6_dst)) { m->m_flags &= ~(M_BCAST | M_MCAST); /* Just in case. */ } else { m->m_flags = (m->m_flags & ~M_BCAST) | M_MCAST; in6_ifstat_inc(ifp, ifs6_out_mcast); /* Confirm that the outgoing interface supports multicast. */ if (!(ifp->if_flags & IFF_MULTICAST)) { IP6STAT_INC(ip6s_noroute); in6_ifstat_inc(ifp, ifs6_out_discard); error = ENETUNREACH; goto bad; } if ((im6o == NULL && in6_mcast_loop) || (im6o && im6o->im6o_multicast_loop)) { /* * Loop back multicast datagram if not expressly * forbidden to do so, even if we have not joined * the address; protocols will filter it later, * thus deferring a hash lookup and lock acquisition * at the expense of an m_copym(). */ ip6_mloopback(ifp, m); } else { /* * If we are acting as a multicast router, perform * multicast forwarding as if the packet had just * arrived on the interface to which we are about * to send. The multicast forwarding function * recursively calls this function, using the * IPV6_FORWARDING flag to prevent infinite recursion. * * Multicasts that are looped back by ip6_mloopback(), * above, will be forwarded by the ip6_input() routine, * if necessary. */ if (V_ip6_mrouter && (flags & IPV6_FORWARDING) == 0) { /* * XXX: ip6_mforward expects that rcvif is NULL * when it is called from the originating path. * However, it may not always be the case. */ m->m_pkthdr.rcvif = NULL; if (ip6_mforward(ip6, ifp, m) != 0) { m_freem(m); goto done; } } } /* * Multicasts with a hoplimit of zero may be looped back, * above, but must not be transmitted on a network. * Also, multicasts addressed to the loopback interface * are not sent -- the above call to ip6_mloopback() will * loop back a copy if this host actually belongs to the * destination group on the loopback interface. */ if (ip6->ip6_hlim == 0 || (ifp->if_flags & IFF_LOOPBACK) || IN6_IS_ADDR_MC_INTFACELOCAL(&ip6->ip6_dst)) { m_freem(m); goto done; } } /* * Fill the outgoing inteface to tell the upper layer * to increment per-interface statistics. */ if (ifpp) *ifpp = ifp; /* Determine path MTU. */ if ((error = ip6_getpmtu(ro_pmtu, ro != ro_pmtu, ifp, &ip6->ip6_dst, &mtu, &alwaysfrag, fibnum, *nexthdrp)) != 0) goto bad; KASSERT(mtu > 0, ("%s:%d: mtu %ld, ro_pmtu %p ro %p ifp %p " "alwaysfrag %d fibnum %u\n", __func__, __LINE__, mtu, ro_pmtu, ro, ifp, alwaysfrag, fibnum)); /* * The caller of this function may specify to use the minimum MTU * in some cases. * An advanced API option (IPV6_USE_MIN_MTU) can also override MTU * setting. The logic is a bit complicated; by default, unicast * packets will follow path MTU while multicast packets will be sent at * the minimum MTU. If IP6PO_MINMTU_ALL is specified, all packets * including unicast ones will be sent at the minimum MTU. Multicast * packets will always be sent at the minimum MTU unless * IP6PO_MINMTU_DISABLE is explicitly specified. * See RFC 3542 for more details. */ if (mtu > IPV6_MMTU) { if ((flags & IPV6_MINMTU)) mtu = IPV6_MMTU; else if (opt && opt->ip6po_minmtu == IP6PO_MINMTU_ALL) mtu = IPV6_MMTU; else if (IN6_IS_ADDR_MULTICAST(&ip6->ip6_dst) && (opt == NULL || opt->ip6po_minmtu != IP6PO_MINMTU_DISABLE)) { mtu = IPV6_MMTU; } } /* * Clear embedded scope identifiers if necessary. * in6_clearscope() will touch the addresses only when necessary. */ in6_clearscope(&ip6->ip6_src); in6_clearscope(&ip6->ip6_dst); /* * If the outgoing packet contains a hop-by-hop options header, * it must be examined and processed even by the source node. * (RFC 2460, section 4.) */ if (exthdrs.ip6e_hbh) { struct ip6_hbh *hbh = mtod(exthdrs.ip6e_hbh, struct ip6_hbh *); u_int32_t dummy; /* XXX unused */ u_int32_t plen = 0; /* XXX: ip6_process will check the value */ #ifdef DIAGNOSTIC if ((hbh->ip6h_len + 1) << 3 > exthdrs.ip6e_hbh->m_len) panic("ip6e_hbh is not contiguous"); #endif /* * XXX: if we have to send an ICMPv6 error to the sender, * we need the M_LOOP flag since icmp6_error() expects * the IPv6 and the hop-by-hop options header are * contiguous unless the flag is set. */ m->m_flags |= M_LOOP; m->m_pkthdr.rcvif = ifp; if (ip6_process_hopopts(m, (u_int8_t *)(hbh + 1), ((hbh->ip6h_len + 1) << 3) - sizeof(struct ip6_hbh), &dummy, &plen) < 0) { /* m was already freed at this point. */ error = EINVAL;/* better error? */ goto done; } m->m_flags &= ~M_LOOP; /* XXX */ m->m_pkthdr.rcvif = NULL; } /* Jump over all PFIL processing if hooks are not active. */ if (!PFIL_HOOKED_OUT(V_inet6_pfil_head)) goto passout; odst = ip6->ip6_dst; /* Run through list of hooks for output packets. */ switch (pfil_mbuf_out(V_inet6_pfil_head, &m, ifp, inp)) { case PFIL_PASS: ip6 = mtod(m, struct ip6_hdr *); break; case PFIL_DROPPED: error = EACCES; /* FALLTHROUGH */ case PFIL_CONSUMED: goto done; } needfiblookup = 0; /* See if destination IP address was changed by packet filter. */ if (!IN6_ARE_ADDR_EQUAL(&odst, &ip6->ip6_dst)) { m->m_flags |= M_SKIP_FIREWALL; /* If destination is now ourself drop to ip6_input(). */ if (in6_localip(&ip6->ip6_dst)) { m->m_flags |= M_FASTFWD_OURS; if (m->m_pkthdr.rcvif == NULL) m->m_pkthdr.rcvif = V_loif; if (m->m_pkthdr.csum_flags & CSUM_DELAY_DATA_IPV6) { m->m_pkthdr.csum_flags |= CSUM_DATA_VALID_IPV6 | CSUM_PSEUDO_HDR; m->m_pkthdr.csum_data = 0xffff; } #if defined(SCTP) || defined(SCTP_SUPPORT) if (m->m_pkthdr.csum_flags & CSUM_SCTP_IPV6) m->m_pkthdr.csum_flags |= CSUM_SCTP_VALID; #endif error = netisr_queue(NETISR_IPV6, m); goto done; } else { if (ro != NULL) RO_INVALIDATE_CACHE(ro); needfiblookup = 1; /* Redo the routing table lookup. */ } } /* See if fib was changed by packet filter. */ if (fibnum != M_GETFIB(m)) { m->m_flags |= M_SKIP_FIREWALL; fibnum = M_GETFIB(m); if (ro != NULL) RO_INVALIDATE_CACHE(ro); needfiblookup = 1; } if (needfiblookup) goto again; /* See if local, if yes, send it to netisr. */ if (m->m_flags & M_FASTFWD_OURS) { if (m->m_pkthdr.rcvif == NULL) m->m_pkthdr.rcvif = V_loif; if (m->m_pkthdr.csum_flags & CSUM_DELAY_DATA_IPV6) { m->m_pkthdr.csum_flags |= CSUM_DATA_VALID_IPV6 | CSUM_PSEUDO_HDR; m->m_pkthdr.csum_data = 0xffff; } #if defined(SCTP) || defined(SCTP_SUPPORT) if (m->m_pkthdr.csum_flags & CSUM_SCTP_IPV6) m->m_pkthdr.csum_flags |= CSUM_SCTP_VALID; #endif error = netisr_queue(NETISR_IPV6, m); goto done; } /* Or forward to some other address? */ if ((m->m_flags & M_IP6_NEXTHOP) && (fwd_tag = m_tag_find(m, PACKET_TAG_IPFORWARD, NULL)) != NULL) { if (ro != NULL) dst = (struct sockaddr_in6 *)&ro->ro_dst; else dst = &sin6; bcopy((fwd_tag+1), &dst_sa, sizeof(struct sockaddr_in6)); m->m_flags |= M_SKIP_FIREWALL; m->m_flags &= ~M_IP6_NEXTHOP; m_tag_delete(m, fwd_tag); goto again; } passout: if (vlan_pcp > -1) EVL_APPLY_PRI(m, vlan_pcp); /* Ensure the packet data is mapped if the interface requires it. */ if ((ifp->if_capenable & IFCAP_MEXTPG) == 0) { m = mb_unmapped_to_ext(m); if (m == NULL) { IP6STAT_INC(ip6s_odropped); return (ENOBUFS); } } /* * Send the packet to the outgoing interface. * If necessary, do IPv6 fragmentation before sending. * * The logic here is rather complex: * 1: normal case (dontfrag == 0, alwaysfrag == 0) * 1-a: send as is if tlen <= path mtu * 1-b: fragment if tlen > path mtu * * 2: if user asks us not to fragment (dontfrag == 1) * 2-a: send as is if tlen <= interface mtu * 2-b: error if tlen > interface mtu * * 3: if we always need to attach fragment header (alwaysfrag == 1) * always fragment * * 4: if dontfrag == 1 && alwaysfrag == 1 * error, as we cannot handle this conflicting request. */ sw_csum = m->m_pkthdr.csum_flags; if (!hdrsplit) { tso = ((sw_csum & ifp->if_hwassist & (CSUM_TSO | CSUM_INNER_TSO)) != 0) ? 1 : 0; sw_csum &= ~ifp->if_hwassist; } else tso = 0; /* * If we added extension headers, we will not do TSO and calculate the * checksums ourselves for now. * XXX-BZ Need a framework to know when the NIC can handle it, even * with ext. hdrs. */ ip6_output_delayed_csum(m, ifp, sw_csum, plen, optlen); /* XXX-BZ m->m_pkthdr.csum_flags &= ~ifp->if_hwassist; */ tlen = m->m_pkthdr.len; if ((opt && (opt->ip6po_flags & IP6PO_DONTFRAG)) || tso) dontfrag = 1; else dontfrag = 0; if (dontfrag && alwaysfrag) { /* Case 4. */ /* Conflicting request - can't transmit. */ error = EMSGSIZE; goto bad; } if (dontfrag && tlen > IN6_LINKMTU(ifp) && !tso) { /* Case 2-b. */ /* * Even if the DONTFRAG option is specified, we cannot send the * packet when the data length is larger than the MTU of the * outgoing interface. * Notify the error by sending IPV6_PATHMTU ancillary data if * application wanted to know the MTU value. Also return an * error code (this is not described in the API spec). */ if (inp != NULL) ip6_notify_pmtu(inp, &dst_sa, (u_int32_t)mtu); error = EMSGSIZE; goto bad; } /* Transmit packet without fragmentation. */ if (dontfrag || (!alwaysfrag && tlen <= mtu)) { /* Cases 1-a and 2-a. */ struct in6_ifaddr *ia6; ip6 = mtod(m, struct ip6_hdr *); ia6 = in6_ifawithifp(ifp, &ip6->ip6_src); if (ia6) { /* Record statistics for this interface address. */ counter_u64_add(ia6->ia_ifa.ifa_opackets, 1); counter_u64_add(ia6->ia_ifa.ifa_obytes, m->m_pkthdr.len); } error = ip6_output_send(inp, ifp, origifp, m, dst, ro, (flags & IP_NO_SND_TAG_RL) ? false : true); goto done; } /* Try to fragment the packet. Cases 1-b and 3. */ if (mtu < IPV6_MMTU) { /* Path MTU cannot be less than IPV6_MMTU. */ error = EMSGSIZE; in6_ifstat_inc(ifp, ifs6_out_fragfail); goto bad; } else if (ip6->ip6_plen == 0) { /* Jumbo payload cannot be fragmented. */ error = EMSGSIZE; in6_ifstat_inc(ifp, ifs6_out_fragfail); goto bad; } else { u_char nextproto; /* * Too large for the destination or interface; * fragment if possible. * Must be able to put at least 8 bytes per fragment. */ if (mtu > IPV6_MAXPACKET) mtu = IPV6_MAXPACKET; len = (mtu - unfragpartlen - sizeof(struct ip6_frag)) & ~7; if (len < 8) { error = EMSGSIZE; in6_ifstat_inc(ifp, ifs6_out_fragfail); goto bad; } /* * If the interface will not calculate checksums on * fragmented packets, then do it here. * XXX-BZ handle the hw offloading case. Need flags. */ ip6_output_delayed_csum(m, ifp, m->m_pkthdr.csum_flags, plen, optlen); /* * Change the next header field of the last header in the * unfragmentable part. */ if (exthdrs.ip6e_rthdr) { nextproto = *mtod(exthdrs.ip6e_rthdr, u_char *); *mtod(exthdrs.ip6e_rthdr, u_char *) = IPPROTO_FRAGMENT; } else if (exthdrs.ip6e_dest1) { nextproto = *mtod(exthdrs.ip6e_dest1, u_char *); *mtod(exthdrs.ip6e_dest1, u_char *) = IPPROTO_FRAGMENT; } else if (exthdrs.ip6e_hbh) { nextproto = *mtod(exthdrs.ip6e_hbh, u_char *); *mtod(exthdrs.ip6e_hbh, u_char *) = IPPROTO_FRAGMENT; } else { ip6 = mtod(m, struct ip6_hdr *); nextproto = ip6->ip6_nxt; ip6->ip6_nxt = IPPROTO_FRAGMENT; } /* * Loop through length of segment after first fragment, * make new header and copy data of each part and link onto * chain. */ m0 = m; id = htonl(ip6_randomid()); error = ip6_fragment(ifp, m, unfragpartlen, nextproto,len, id); if (error != 0) goto sendorfree; in6_ifstat_inc(ifp, ifs6_out_fragok); } /* Remove leading garbage. */ sendorfree: m = m0->m_nextpkt; m0->m_nextpkt = 0; m_freem(m0); for (; m; m = m0) { m0 = m->m_nextpkt; m->m_nextpkt = 0; if (error == 0) { /* Record statistics for this interface address. */ if (ia) { counter_u64_add(ia->ia_ifa.ifa_opackets, 1); counter_u64_add(ia->ia_ifa.ifa_obytes, m->m_pkthdr.len); } if (vlan_pcp > -1) EVL_APPLY_PRI(m, vlan_pcp); error = ip6_output_send(inp, ifp, origifp, m, dst, ro, true); } else m_freem(m); } if (error == 0) IP6STAT_INC(ip6s_fragmented); done: return (error); freehdrs: m_freem(exthdrs.ip6e_hbh); /* m_freem() checks if mbuf is NULL. */ m_freem(exthdrs.ip6e_dest1); m_freem(exthdrs.ip6e_rthdr); m_freem(exthdrs.ip6e_dest2); /* FALLTHROUGH */ bad: if (m) m_freem(m); goto done; } static int ip6_copyexthdr(struct mbuf **mp, caddr_t hdr, int hlen) { struct mbuf *m; if (hlen > MCLBYTES) return (ENOBUFS); /* XXX */ if (hlen > MLEN) m = m_getcl(M_NOWAIT, MT_DATA, 0); else m = m_get(M_NOWAIT, MT_DATA); if (m == NULL) return (ENOBUFS); m->m_len = hlen; if (hdr) bcopy(hdr, mtod(m, caddr_t), hlen); *mp = m; return (0); } /* * Insert jumbo payload option. */ static int ip6_insert_jumboopt(struct ip6_exthdrs *exthdrs, u_int32_t plen) { struct mbuf *mopt; u_char *optbuf; u_int32_t v; #define JUMBOOPTLEN 8 /* length of jumbo payload option and padding */ /* * If there is no hop-by-hop options header, allocate new one. * If there is one but it doesn't have enough space to store the * jumbo payload option, allocate a cluster to store the whole options. * Otherwise, use it to store the options. */ if (exthdrs->ip6e_hbh == NULL) { mopt = m_get(M_NOWAIT, MT_DATA); if (mopt == NULL) return (ENOBUFS); mopt->m_len = JUMBOOPTLEN; optbuf = mtod(mopt, u_char *); optbuf[1] = 0; /* = ((JUMBOOPTLEN) >> 3) - 1 */ exthdrs->ip6e_hbh = mopt; } else { struct ip6_hbh *hbh; mopt = exthdrs->ip6e_hbh; if (M_TRAILINGSPACE(mopt) < JUMBOOPTLEN) { /* * XXX assumption: * - exthdrs->ip6e_hbh is not referenced from places * other than exthdrs. * - exthdrs->ip6e_hbh is not an mbuf chain. */ int oldoptlen = mopt->m_len; struct mbuf *n; /* * XXX: give up if the whole (new) hbh header does * not fit even in an mbuf cluster. */ if (oldoptlen + JUMBOOPTLEN > MCLBYTES) return (ENOBUFS); /* * As a consequence, we must always prepare a cluster * at this point. */ n = m_getcl(M_NOWAIT, MT_DATA, 0); if (n == NULL) return (ENOBUFS); n->m_len = oldoptlen + JUMBOOPTLEN; bcopy(mtod(mopt, caddr_t), mtod(n, caddr_t), oldoptlen); optbuf = mtod(n, caddr_t) + oldoptlen; m_freem(mopt); mopt = exthdrs->ip6e_hbh = n; } else { optbuf = mtod(mopt, u_char *) + mopt->m_len; mopt->m_len += JUMBOOPTLEN; } optbuf[0] = IP6OPT_PADN; optbuf[1] = 1; /* * Adjust the header length according to the pad and * the jumbo payload option. */ hbh = mtod(mopt, struct ip6_hbh *); hbh->ip6h_len += (JUMBOOPTLEN >> 3); } /* fill in the option. */ optbuf[2] = IP6OPT_JUMBO; optbuf[3] = 4; v = (u_int32_t)htonl(plen + JUMBOOPTLEN); bcopy(&v, &optbuf[4], sizeof(u_int32_t)); /* finally, adjust the packet header length */ exthdrs->ip6e_ip6->m_pkthdr.len += JUMBOOPTLEN; return (0); #undef JUMBOOPTLEN } /* * Insert fragment header and copy unfragmentable header portions. */ static int ip6_insertfraghdr(struct mbuf *m0, struct mbuf *m, int hlen, struct ip6_frag **frghdrp) { struct mbuf *n, *mlast; if (hlen > sizeof(struct ip6_hdr)) { n = m_copym(m0, sizeof(struct ip6_hdr), hlen - sizeof(struct ip6_hdr), M_NOWAIT); if (n == NULL) return (ENOBUFS); m->m_next = n; } else n = m; /* Search for the last mbuf of unfragmentable part. */ for (mlast = n; mlast->m_next; mlast = mlast->m_next) ; if (M_WRITABLE(mlast) && M_TRAILINGSPACE(mlast) >= sizeof(struct ip6_frag)) { /* use the trailing space of the last mbuf for the fragment hdr */ *frghdrp = (struct ip6_frag *)(mtod(mlast, caddr_t) + mlast->m_len); mlast->m_len += sizeof(struct ip6_frag); m->m_pkthdr.len += sizeof(struct ip6_frag); } else { /* allocate a new mbuf for the fragment header */ struct mbuf *mfrg; mfrg = m_get(M_NOWAIT, MT_DATA); if (mfrg == NULL) return (ENOBUFS); mfrg->m_len = sizeof(struct ip6_frag); *frghdrp = mtod(mfrg, struct ip6_frag *); mlast->m_next = mfrg; } return (0); } /* * Calculates IPv6 path mtu for destination @dst. * Resulting MTU is stored in @mtup. * * Returns 0 on success. */ static int ip6_getpmtu_ctl(u_int fibnum, const struct in6_addr *dst, u_long *mtup) { struct epoch_tracker et; struct nhop_object *nh; struct in6_addr kdst; uint32_t scopeid; int error; in6_splitscope(dst, &kdst, &scopeid); NET_EPOCH_ENTER(et); nh = fib6_lookup(fibnum, &kdst, scopeid, NHR_NONE, 0); if (nh != NULL) error = ip6_calcmtu(nh->nh_ifp, dst, nh->nh_mtu, mtup, NULL, 0); else error = EHOSTUNREACH; NET_EPOCH_EXIT(et); return (error); } /* * Calculates IPv6 path MTU for @dst based on transmit @ifp, * and cached data in @ro_pmtu. * MTU from (successful) route lookup is saved (along with dst) * inside @ro_pmtu to avoid subsequent route lookups after packet * filter processing. * * Stores mtu and always-frag value into @mtup and @alwaysfragp. * Returns 0 on success. */ static int ip6_getpmtu(struct route_in6 *ro_pmtu, int do_lookup, struct ifnet *ifp, const struct in6_addr *dst, u_long *mtup, int *alwaysfragp, u_int fibnum, u_int proto) { struct nhop_object *nh; struct in6_addr kdst; uint32_t scopeid; struct sockaddr_in6 *sa6_dst, sin6; u_long mtu; NET_EPOCH_ASSERT(); mtu = 0; if (ro_pmtu == NULL || do_lookup) { /* * Here ro_pmtu has final destination address, while * ro might represent immediate destination. * Use ro_pmtu destination since mtu might differ. */ if (ro_pmtu != NULL) { sa6_dst = (struct sockaddr_in6 *)&ro_pmtu->ro_dst; if (!IN6_ARE_ADDR_EQUAL(&sa6_dst->sin6_addr, dst)) ro_pmtu->ro_mtu = 0; } else sa6_dst = &sin6; if (ro_pmtu == NULL || ro_pmtu->ro_mtu == 0) { bzero(sa6_dst, sizeof(*sa6_dst)); sa6_dst->sin6_family = AF_INET6; sa6_dst->sin6_len = sizeof(struct sockaddr_in6); sa6_dst->sin6_addr = *dst; in6_splitscope(dst, &kdst, &scopeid); nh = fib6_lookup(fibnum, &kdst, scopeid, NHR_NONE, 0); if (nh != NULL) { mtu = nh->nh_mtu; if (ro_pmtu != NULL) ro_pmtu->ro_mtu = mtu; } } else mtu = ro_pmtu->ro_mtu; } if (ro_pmtu != NULL && ro_pmtu->ro_nh != NULL) mtu = ro_pmtu->ro_nh->nh_mtu; return (ip6_calcmtu(ifp, dst, mtu, mtup, alwaysfragp, proto)); } /* * Calculate MTU based on transmit @ifp, route mtu @rt_mtu and * hostcache data for @dst. * Stores mtu and always-frag value into @mtup and @alwaysfragp. * * Returns 0 on success. */ static int ip6_calcmtu(struct ifnet *ifp, const struct in6_addr *dst, u_long rt_mtu, u_long *mtup, int *alwaysfragp, u_int proto) { u_long mtu = 0; int alwaysfrag = 0; int error = 0; if (rt_mtu > 0) { u_int32_t ifmtu; struct in_conninfo inc; bzero(&inc, sizeof(inc)); inc.inc_flags |= INC_ISIPV6; inc.inc6_faddr = *dst; ifmtu = IN6_LINKMTU(ifp); /* TCP is known to react to pmtu changes so skip hc */ if (proto != IPPROTO_TCP) mtu = tcp_hc_getmtu(&inc); if (mtu) mtu = min(mtu, rt_mtu); else mtu = rt_mtu; if (mtu == 0) mtu = ifmtu; else if (mtu < IPV6_MMTU) { /* * RFC2460 section 5, last paragraph: * if we record ICMPv6 too big message with * mtu < IPV6_MMTU, transmit packets sized IPV6_MMTU * or smaller, with framgent header attached. * (fragment header is needed regardless from the * packet size, for translators to identify packets) */ alwaysfrag = 1; mtu = IPV6_MMTU; } } else if (ifp) { mtu = IN6_LINKMTU(ifp); } else error = EHOSTUNREACH; /* XXX */ *mtup = mtu; if (alwaysfragp) *alwaysfragp = alwaysfrag; return (error); } /* * IP6 socket option processing. */ int ip6_ctloutput(struct socket *so, struct sockopt *sopt) { int optdatalen, uproto; void *optdata; struct inpcb *inp = sotoinpcb(so); int error, optval; int level, op, optname; int optlen; struct thread *td; #ifdef RSS uint32_t rss_bucket; int retval; #endif /* * Don't use more than a quarter of mbuf clusters. N.B.: * nmbclusters is an int, but nmbclusters * MCLBYTES may overflow * on LP64 architectures, so cast to u_long to avoid undefined * behavior. ILP32 architectures cannot have nmbclusters * large enough to overflow for other reasons. */ #define IPV6_PKTOPTIONS_MBUF_LIMIT ((u_long)nmbclusters * MCLBYTES / 4) level = sopt->sopt_level; op = sopt->sopt_dir; optname = sopt->sopt_name; optlen = sopt->sopt_valsize; td = sopt->sopt_td; error = 0; optval = 0; uproto = (int)so->so_proto->pr_protocol; if (level != IPPROTO_IPV6) { error = EINVAL; if (sopt->sopt_level == SOL_SOCKET && sopt->sopt_dir == SOPT_SET) { switch (sopt->sopt_name) { case SO_REUSEADDR: INP_WLOCK(inp); if ((so->so_options & SO_REUSEADDR) != 0) inp->inp_flags2 |= INP_REUSEADDR; else inp->inp_flags2 &= ~INP_REUSEADDR; INP_WUNLOCK(inp); error = 0; break; case SO_REUSEPORT: INP_WLOCK(inp); if ((so->so_options & SO_REUSEPORT) != 0) inp->inp_flags2 |= INP_REUSEPORT; else inp->inp_flags2 &= ~INP_REUSEPORT; INP_WUNLOCK(inp); error = 0; break; case SO_REUSEPORT_LB: INP_WLOCK(inp); if ((so->so_options & SO_REUSEPORT_LB) != 0) inp->inp_flags2 |= INP_REUSEPORT_LB; else inp->inp_flags2 &= ~INP_REUSEPORT_LB; INP_WUNLOCK(inp); error = 0; break; case SO_SETFIB: INP_WLOCK(inp); inp->inp_inc.inc_fibnum = so->so_fibnum; INP_WUNLOCK(inp); error = 0; break; case SO_MAX_PACING_RATE: #ifdef RATELIMIT INP_WLOCK(inp); inp->inp_flags2 |= INP_RATE_LIMIT_CHANGED; INP_WUNLOCK(inp); error = 0; #else error = EOPNOTSUPP; #endif break; default: break; } } } else { /* level == IPPROTO_IPV6 */ switch (op) { case SOPT_SET: switch (optname) { case IPV6_2292PKTOPTIONS: #ifdef IPV6_PKTOPTIONS case IPV6_PKTOPTIONS: #endif { struct mbuf *m; if (optlen > IPV6_PKTOPTIONS_MBUF_LIMIT) { printf("ip6_ctloutput: mbuf limit hit\n"); error = ENOBUFS; break; } error = soopt_getm(sopt, &m); /* XXX */ if (error != 0) break; error = soopt_mcopyin(sopt, m); /* XXX */ if (error != 0) break; INP_WLOCK(inp); error = ip6_pcbopts(&inp->in6p_outputopts, m, so, sopt); INP_WUNLOCK(inp); m_freem(m); /* XXX */ break; } /* * Use of some Hop-by-Hop options or some * Destination options, might require special * privilege. That is, normal applications * (without special privilege) might be forbidden * from setting certain options in outgoing packets, * and might never see certain options in received * packets. [RFC 2292 Section 6] * KAME specific note: * KAME prevents non-privileged users from sending or * receiving ANY hbh/dst options in order to avoid * overhead of parsing options in the kernel. */ case IPV6_RECVHOPOPTS: case IPV6_RECVDSTOPTS: case IPV6_RECVRTHDRDSTOPTS: if (td != NULL) { error = priv_check(td, PRIV_NETINET_SETHDROPTS); if (error) break; } /* FALLTHROUGH */ case IPV6_UNICAST_HOPS: case IPV6_HOPLIMIT: case IPV6_RECVPKTINFO: case IPV6_RECVHOPLIMIT: case IPV6_RECVRTHDR: case IPV6_RECVPATHMTU: case IPV6_RECVTCLASS: case IPV6_RECVFLOWID: #ifdef RSS case IPV6_RECVRSSBUCKETID: #endif case IPV6_V6ONLY: case IPV6_AUTOFLOWLABEL: case IPV6_ORIGDSTADDR: case IPV6_BINDANY: case IPV6_BINDMULTI: #ifdef RSS case IPV6_RSS_LISTEN_BUCKET: #endif case IPV6_VLAN_PCP: if (optname == IPV6_BINDANY && td != NULL) { error = priv_check(td, PRIV_NETINET_BINDANY); if (error) break; } if (optlen != sizeof(int)) { error = EINVAL; break; } error = sooptcopyin(sopt, &optval, sizeof optval, sizeof optval); if (error) break; switch (optname) { case IPV6_UNICAST_HOPS: if (optval < -1 || optval >= 256) error = EINVAL; else { /* -1 = kernel default */ inp->in6p_hops = optval; if ((inp->inp_vflag & INP_IPV4) != 0) inp->inp_ip_ttl = optval; } break; #define OPTSET(bit) \ do { \ INP_WLOCK(inp); \ if (optval) \ inp->inp_flags |= (bit); \ else \ inp->inp_flags &= ~(bit); \ INP_WUNLOCK(inp); \ } while (/*CONSTCOND*/ 0) #define OPTSET2292(bit) \ do { \ INP_WLOCK(inp); \ inp->inp_flags |= IN6P_RFC2292; \ if (optval) \ inp->inp_flags |= (bit); \ else \ inp->inp_flags &= ~(bit); \ INP_WUNLOCK(inp); \ } while (/*CONSTCOND*/ 0) #define OPTBIT(bit) (inp->inp_flags & (bit) ? 1 : 0) #define OPTSET2_N(bit, val) do { \ if (val) \ inp->inp_flags2 |= bit; \ else \ inp->inp_flags2 &= ~bit; \ } while (0) #define OPTSET2(bit, val) do { \ INP_WLOCK(inp); \ OPTSET2_N(bit, val); \ INP_WUNLOCK(inp); \ } while (0) #define OPTBIT2(bit) (inp->inp_flags2 & (bit) ? 1 : 0) #define OPTSET2292_EXCLUSIVE(bit) \ do { \ INP_WLOCK(inp); \ if (OPTBIT(IN6P_RFC2292)) { \ error = EINVAL; \ } else { \ if (optval) \ inp->inp_flags |= (bit); \ else \ inp->inp_flags &= ~(bit); \ } \ INP_WUNLOCK(inp); \ } while (/*CONSTCOND*/ 0) case IPV6_RECVPKTINFO: OPTSET2292_EXCLUSIVE(IN6P_PKTINFO); break; case IPV6_HOPLIMIT: { struct ip6_pktopts **optp; /* cannot mix with RFC2292 */ if (OPTBIT(IN6P_RFC2292)) { error = EINVAL; break; } INP_WLOCK(inp); - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { INP_WUNLOCK(inp); return (ECONNRESET); } optp = &inp->in6p_outputopts; error = ip6_pcbopt(IPV6_HOPLIMIT, (u_char *)&optval, sizeof(optval), optp, (td != NULL) ? td->td_ucred : NULL, uproto); INP_WUNLOCK(inp); break; } case IPV6_RECVHOPLIMIT: OPTSET2292_EXCLUSIVE(IN6P_HOPLIMIT); break; case IPV6_RECVHOPOPTS: OPTSET2292_EXCLUSIVE(IN6P_HOPOPTS); break; case IPV6_RECVDSTOPTS: OPTSET2292_EXCLUSIVE(IN6P_DSTOPTS); break; case IPV6_RECVRTHDRDSTOPTS: OPTSET2292_EXCLUSIVE(IN6P_RTHDRDSTOPTS); break; case IPV6_RECVRTHDR: OPTSET2292_EXCLUSIVE(IN6P_RTHDR); break; case IPV6_RECVPATHMTU: /* * We ignore this option for TCP * sockets. * (RFC3542 leaves this case * unspecified.) */ if (uproto != IPPROTO_TCP) OPTSET(IN6P_MTU); break; case IPV6_RECVFLOWID: OPTSET2(INP_RECVFLOWID, optval); break; #ifdef RSS case IPV6_RECVRSSBUCKETID: OPTSET2(INP_RECVRSSBUCKETID, optval); break; #endif case IPV6_V6ONLY: INP_WLOCK(inp); if (inp->inp_lport || !IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) { /* * The socket is already bound. */ INP_WUNLOCK(inp); error = EINVAL; break; } if (optval) { inp->inp_flags |= IN6P_IPV6_V6ONLY; inp->inp_vflag &= ~INP_IPV4; } else { inp->inp_flags &= ~IN6P_IPV6_V6ONLY; inp->inp_vflag |= INP_IPV4; } INP_WUNLOCK(inp); break; case IPV6_RECVTCLASS: /* cannot mix with RFC2292 XXX */ OPTSET2292_EXCLUSIVE(IN6P_TCLASS); break; case IPV6_AUTOFLOWLABEL: OPTSET(IN6P_AUTOFLOWLABEL); break; case IPV6_ORIGDSTADDR: OPTSET2(INP_ORIGDSTADDR, optval); break; case IPV6_BINDANY: OPTSET(INP_BINDANY); break; case IPV6_BINDMULTI: OPTSET2(INP_BINDMULTI, optval); break; #ifdef RSS case IPV6_RSS_LISTEN_BUCKET: if ((optval >= 0) && (optval < rss_getnumbuckets())) { INP_WLOCK(inp); inp->inp_rss_listen_bucket = optval; OPTSET2_N(INP_RSS_BUCKET_SET, 1); INP_WUNLOCK(inp); } else { error = EINVAL; } break; #endif case IPV6_VLAN_PCP: if ((optval >= -1) && (optval <= (INP_2PCP_MASK >> INP_2PCP_SHIFT))) { if (optval == -1) { INP_WLOCK(inp); inp->inp_flags2 &= ~(INP_2PCP_SET | INP_2PCP_MASK); INP_WUNLOCK(inp); } else { INP_WLOCK(inp); inp->inp_flags2 |= INP_2PCP_SET; inp->inp_flags2 &= ~INP_2PCP_MASK; inp->inp_flags2 |= optval << INP_2PCP_SHIFT; INP_WUNLOCK(inp); } } else error = EINVAL; break; } break; case IPV6_TCLASS: case IPV6_DONTFRAG: case IPV6_USE_MIN_MTU: case IPV6_PREFER_TEMPADDR: if (optlen != sizeof(optval)) { error = EINVAL; break; } error = sooptcopyin(sopt, &optval, sizeof optval, sizeof optval); if (error) break; { struct ip6_pktopts **optp; INP_WLOCK(inp); - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { INP_WUNLOCK(inp); return (ECONNRESET); } optp = &inp->in6p_outputopts; error = ip6_pcbopt(optname, (u_char *)&optval, sizeof(optval), optp, (td != NULL) ? td->td_ucred : NULL, uproto); INP_WUNLOCK(inp); break; } case IPV6_2292PKTINFO: case IPV6_2292HOPLIMIT: case IPV6_2292HOPOPTS: case IPV6_2292DSTOPTS: case IPV6_2292RTHDR: /* RFC 2292 */ if (optlen != sizeof(int)) { error = EINVAL; break; } error = sooptcopyin(sopt, &optval, sizeof optval, sizeof optval); if (error) break; switch (optname) { case IPV6_2292PKTINFO: OPTSET2292(IN6P_PKTINFO); break; case IPV6_2292HOPLIMIT: OPTSET2292(IN6P_HOPLIMIT); break; case IPV6_2292HOPOPTS: /* * Check super-user privilege. * See comments for IPV6_RECVHOPOPTS. */ if (td != NULL) { error = priv_check(td, PRIV_NETINET_SETHDROPTS); if (error) return (error); } OPTSET2292(IN6P_HOPOPTS); break; case IPV6_2292DSTOPTS: if (td != NULL) { error = priv_check(td, PRIV_NETINET_SETHDROPTS); if (error) return (error); } OPTSET2292(IN6P_DSTOPTS|IN6P_RTHDRDSTOPTS); /* XXX */ break; case IPV6_2292RTHDR: OPTSET2292(IN6P_RTHDR); break; } break; case IPV6_PKTINFO: case IPV6_HOPOPTS: case IPV6_RTHDR: case IPV6_DSTOPTS: case IPV6_RTHDRDSTOPTS: case IPV6_NEXTHOP: { /* new advanced API (RFC3542) */ u_char *optbuf; u_char optbuf_storage[MCLBYTES]; int optlen; struct ip6_pktopts **optp; /* cannot mix with RFC2292 */ if (OPTBIT(IN6P_RFC2292)) { error = EINVAL; break; } /* * We only ensure valsize is not too large * here. Further validation will be done * later. */ error = sooptcopyin(sopt, optbuf_storage, sizeof(optbuf_storage), 0); if (error) break; optlen = sopt->sopt_valsize; optbuf = optbuf_storage; INP_WLOCK(inp); - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { INP_WUNLOCK(inp); return (ECONNRESET); } optp = &inp->in6p_outputopts; error = ip6_pcbopt(optname, optbuf, optlen, optp, (td != NULL) ? td->td_ucred : NULL, uproto); INP_WUNLOCK(inp); break; } #undef OPTSET case IPV6_MULTICAST_IF: case IPV6_MULTICAST_HOPS: case IPV6_MULTICAST_LOOP: case IPV6_JOIN_GROUP: case IPV6_LEAVE_GROUP: case IPV6_MSFILTER: case MCAST_BLOCK_SOURCE: case MCAST_UNBLOCK_SOURCE: case MCAST_JOIN_GROUP: case MCAST_LEAVE_GROUP: case MCAST_JOIN_SOURCE_GROUP: case MCAST_LEAVE_SOURCE_GROUP: error = ip6_setmoptions(inp, sopt); break; case IPV6_PORTRANGE: error = sooptcopyin(sopt, &optval, sizeof optval, sizeof optval); if (error) break; INP_WLOCK(inp); switch (optval) { case IPV6_PORTRANGE_DEFAULT: inp->inp_flags &= ~(INP_LOWPORT); inp->inp_flags &= ~(INP_HIGHPORT); break; case IPV6_PORTRANGE_HIGH: inp->inp_flags &= ~(INP_LOWPORT); inp->inp_flags |= INP_HIGHPORT; break; case IPV6_PORTRANGE_LOW: inp->inp_flags &= ~(INP_HIGHPORT); inp->inp_flags |= INP_LOWPORT; break; default: error = EINVAL; break; } INP_WUNLOCK(inp); break; #if defined(IPSEC) || defined(IPSEC_SUPPORT) case IPV6_IPSEC_POLICY: if (IPSEC_ENABLED(ipv6)) { error = IPSEC_PCBCTL(ipv6, inp, sopt); break; } /* FALLTHROUGH */ #endif /* IPSEC */ default: error = ENOPROTOOPT; break; } break; case SOPT_GET: switch (optname) { case IPV6_2292PKTOPTIONS: #ifdef IPV6_PKTOPTIONS case IPV6_PKTOPTIONS: #endif /* * RFC3542 (effectively) deprecated the * semantics of the 2292-style pktoptions. * Since it was not reliable in nature (i.e., * applications had to expect the lack of some * information after all), it would make sense * to simplify this part by always returning * empty data. */ sopt->sopt_valsize = 0; break; case IPV6_RECVHOPOPTS: case IPV6_RECVDSTOPTS: case IPV6_RECVRTHDRDSTOPTS: case IPV6_UNICAST_HOPS: case IPV6_RECVPKTINFO: case IPV6_RECVHOPLIMIT: case IPV6_RECVRTHDR: case IPV6_RECVPATHMTU: case IPV6_V6ONLY: case IPV6_PORTRANGE: case IPV6_RECVTCLASS: case IPV6_AUTOFLOWLABEL: case IPV6_BINDANY: case IPV6_FLOWID: case IPV6_FLOWTYPE: case IPV6_RECVFLOWID: #ifdef RSS case IPV6_RSSBUCKETID: case IPV6_RECVRSSBUCKETID: #endif case IPV6_BINDMULTI: case IPV6_VLAN_PCP: switch (optname) { case IPV6_RECVHOPOPTS: optval = OPTBIT(IN6P_HOPOPTS); break; case IPV6_RECVDSTOPTS: optval = OPTBIT(IN6P_DSTOPTS); break; case IPV6_RECVRTHDRDSTOPTS: optval = OPTBIT(IN6P_RTHDRDSTOPTS); break; case IPV6_UNICAST_HOPS: optval = inp->in6p_hops; break; case IPV6_RECVPKTINFO: optval = OPTBIT(IN6P_PKTINFO); break; case IPV6_RECVHOPLIMIT: optval = OPTBIT(IN6P_HOPLIMIT); break; case IPV6_RECVRTHDR: optval = OPTBIT(IN6P_RTHDR); break; case IPV6_RECVPATHMTU: optval = OPTBIT(IN6P_MTU); break; case IPV6_V6ONLY: optval = OPTBIT(IN6P_IPV6_V6ONLY); break; case IPV6_PORTRANGE: { int flags; flags = inp->inp_flags; if (flags & INP_HIGHPORT) optval = IPV6_PORTRANGE_HIGH; else if (flags & INP_LOWPORT) optval = IPV6_PORTRANGE_LOW; else optval = 0; break; } case IPV6_RECVTCLASS: optval = OPTBIT(IN6P_TCLASS); break; case IPV6_AUTOFLOWLABEL: optval = OPTBIT(IN6P_AUTOFLOWLABEL); break; case IPV6_ORIGDSTADDR: optval = OPTBIT2(INP_ORIGDSTADDR); break; case IPV6_BINDANY: optval = OPTBIT(INP_BINDANY); break; case IPV6_FLOWID: optval = inp->inp_flowid; break; case IPV6_FLOWTYPE: optval = inp->inp_flowtype; break; case IPV6_RECVFLOWID: optval = OPTBIT2(INP_RECVFLOWID); break; #ifdef RSS case IPV6_RSSBUCKETID: retval = rss_hash2bucket(inp->inp_flowid, inp->inp_flowtype, &rss_bucket); if (retval == 0) optval = rss_bucket; else error = EINVAL; break; case IPV6_RECVRSSBUCKETID: optval = OPTBIT2(INP_RECVRSSBUCKETID); break; #endif case IPV6_BINDMULTI: optval = OPTBIT2(INP_BINDMULTI); break; case IPV6_VLAN_PCP: if (OPTBIT2(INP_2PCP_SET)) { optval = (inp->inp_flags2 & INP_2PCP_MASK) >> INP_2PCP_SHIFT; } else { optval = -1; } break; } if (error) break; error = sooptcopyout(sopt, &optval, sizeof optval); break; case IPV6_PATHMTU: { u_long pmtu = 0; struct ip6_mtuinfo mtuinfo; struct in6_addr addr; if (!(so->so_state & SS_ISCONNECTED)) return (ENOTCONN); /* * XXX: we dot not consider the case of source * routing, or optional information to specify * the outgoing interface. * Copy faddr out of inp to avoid holding lock * on inp during route lookup. */ INP_RLOCK(inp); bcopy(&inp->in6p_faddr, &addr, sizeof(addr)); INP_RUNLOCK(inp); error = ip6_getpmtu_ctl(so->so_fibnum, &addr, &pmtu); if (error) break; if (pmtu > IPV6_MAXPACKET) pmtu = IPV6_MAXPACKET; bzero(&mtuinfo, sizeof(mtuinfo)); mtuinfo.ip6m_mtu = (u_int32_t)pmtu; optdata = (void *)&mtuinfo; optdatalen = sizeof(mtuinfo); error = sooptcopyout(sopt, optdata, optdatalen); break; } case IPV6_2292PKTINFO: case IPV6_2292HOPLIMIT: case IPV6_2292HOPOPTS: case IPV6_2292RTHDR: case IPV6_2292DSTOPTS: switch (optname) { case IPV6_2292PKTINFO: optval = OPTBIT(IN6P_PKTINFO); break; case IPV6_2292HOPLIMIT: optval = OPTBIT(IN6P_HOPLIMIT); break; case IPV6_2292HOPOPTS: optval = OPTBIT(IN6P_HOPOPTS); break; case IPV6_2292RTHDR: optval = OPTBIT(IN6P_RTHDR); break; case IPV6_2292DSTOPTS: optval = OPTBIT(IN6P_DSTOPTS|IN6P_RTHDRDSTOPTS); break; } error = sooptcopyout(sopt, &optval, sizeof optval); break; case IPV6_PKTINFO: case IPV6_HOPOPTS: case IPV6_RTHDR: case IPV6_DSTOPTS: case IPV6_RTHDRDSTOPTS: case IPV6_NEXTHOP: case IPV6_TCLASS: case IPV6_DONTFRAG: case IPV6_USE_MIN_MTU: case IPV6_PREFER_TEMPADDR: error = ip6_getpcbopt(inp, optname, sopt); break; case IPV6_MULTICAST_IF: case IPV6_MULTICAST_HOPS: case IPV6_MULTICAST_LOOP: case IPV6_MSFILTER: error = ip6_getmoptions(inp, sopt); break; #if defined(IPSEC) || defined(IPSEC_SUPPORT) case IPV6_IPSEC_POLICY: if (IPSEC_ENABLED(ipv6)) { error = IPSEC_PCBCTL(ipv6, inp, sopt); break; } /* FALLTHROUGH */ #endif /* IPSEC */ default: error = ENOPROTOOPT; break; } break; } } return (error); } int ip6_raw_ctloutput(struct socket *so, struct sockopt *sopt) { int error = 0, optval, optlen; const int icmp6off = offsetof(struct icmp6_hdr, icmp6_cksum); struct inpcb *inp = sotoinpcb(so); int level, op, optname; level = sopt->sopt_level; op = sopt->sopt_dir; optname = sopt->sopt_name; optlen = sopt->sopt_valsize; if (level != IPPROTO_IPV6) { return (EINVAL); } switch (optname) { case IPV6_CHECKSUM: /* * For ICMPv6 sockets, no modification allowed for checksum * offset, permit "no change" values to help existing apps. * * RFC3542 says: "An attempt to set IPV6_CHECKSUM * for an ICMPv6 socket will fail." * The current behavior does not meet RFC3542. */ switch (op) { case SOPT_SET: if (optlen != sizeof(int)) { error = EINVAL; break; } error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval)); if (error) break; if (optval < -1 || (optval % 2) != 0) { /* * The API assumes non-negative even offset * values or -1 as a special value. */ error = EINVAL; } else if (inp->inp_ip_p == IPPROTO_ICMPV6) { if (optval != icmp6off) error = EINVAL; } else inp->in6p_cksum = optval; break; case SOPT_GET: if (inp->inp_ip_p == IPPROTO_ICMPV6) optval = icmp6off; else optval = inp->in6p_cksum; error = sooptcopyout(sopt, &optval, sizeof(optval)); break; default: error = EINVAL; break; } break; default: error = ENOPROTOOPT; break; } return (error); } /* * Set up IP6 options in pcb for insertion in output packets or * specifying behavior of outgoing packets. */ static int ip6_pcbopts(struct ip6_pktopts **pktopt, struct mbuf *m, struct socket *so, struct sockopt *sopt) { struct ip6_pktopts *opt = *pktopt; int error = 0; struct thread *td = sopt->sopt_td; struct epoch_tracker et; /* turn off any old options. */ if (opt) { #ifdef DIAGNOSTIC if (opt->ip6po_pktinfo || opt->ip6po_nexthop || opt->ip6po_hbh || opt->ip6po_dest1 || opt->ip6po_dest2 || opt->ip6po_rhinfo.ip6po_rhi_rthdr) printf("ip6_pcbopts: all specified options are cleared.\n"); #endif ip6_clearpktopts(opt, -1); } else { opt = malloc(sizeof(*opt), M_IP6OPT, M_NOWAIT); if (opt == NULL) return (ENOMEM); } *pktopt = NULL; if (!m || m->m_len == 0) { /* * Only turning off any previous options, regardless of * whether the opt is just created or given. */ free(opt, M_IP6OPT); return (0); } /* set options specified by user. */ NET_EPOCH_ENTER(et); if ((error = ip6_setpktopts(m, opt, NULL, (td != NULL) ? td->td_ucred : NULL, so->so_proto->pr_protocol)) != 0) { ip6_clearpktopts(opt, -1); /* XXX: discard all options */ free(opt, M_IP6OPT); NET_EPOCH_EXIT(et); return (error); } NET_EPOCH_EXIT(et); *pktopt = opt; return (0); } /* * initialize ip6_pktopts. beware that there are non-zero default values in * the struct. */ void ip6_initpktopts(struct ip6_pktopts *opt) { bzero(opt, sizeof(*opt)); opt->ip6po_hlim = -1; /* -1 means default hop limit */ opt->ip6po_tclass = -1; /* -1 means default traffic class */ opt->ip6po_minmtu = IP6PO_MINMTU_MCASTONLY; opt->ip6po_prefer_tempaddr = IP6PO_TEMPADDR_SYSTEM; } static int ip6_pcbopt(int optname, u_char *buf, int len, struct ip6_pktopts **pktopt, struct ucred *cred, int uproto) { struct epoch_tracker et; struct ip6_pktopts *opt; int ret; if (*pktopt == NULL) { *pktopt = malloc(sizeof(struct ip6_pktopts), M_IP6OPT, M_NOWAIT); if (*pktopt == NULL) return (ENOBUFS); ip6_initpktopts(*pktopt); } opt = *pktopt; NET_EPOCH_ENTER(et); ret = ip6_setpktopt(optname, buf, len, opt, cred, 1, 0, uproto); NET_EPOCH_EXIT(et); return (ret); } #define GET_PKTOPT_VAR(field, lenexpr) do { \ if (pktopt && pktopt->field) { \ INP_RUNLOCK(inp); \ optdata = malloc(sopt->sopt_valsize, M_TEMP, M_WAITOK); \ malloc_optdata = true; \ INP_RLOCK(inp); \ - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { \ + if (inp->inp_flags & INP_DROPPED) { \ INP_RUNLOCK(inp); \ free(optdata, M_TEMP); \ return (ECONNRESET); \ } \ pktopt = inp->in6p_outputopts; \ if (pktopt && pktopt->field) { \ optdatalen = min(lenexpr, sopt->sopt_valsize); \ bcopy(pktopt->field, optdata, optdatalen); \ } else { \ free(optdata, M_TEMP); \ optdata = NULL; \ malloc_optdata = false; \ } \ } \ } while(0) #define GET_PKTOPT_EXT_HDR(field) GET_PKTOPT_VAR(field, \ (((struct ip6_ext *)pktopt->field)->ip6e_len + 1) << 3) #define GET_PKTOPT_SOCKADDR(field) GET_PKTOPT_VAR(field, \ pktopt->field->sa_len) static int ip6_getpcbopt(struct inpcb *inp, int optname, struct sockopt *sopt) { void *optdata = NULL; bool malloc_optdata = false; int optdatalen = 0; int error = 0; struct in6_pktinfo null_pktinfo; int deftclass = 0, on; int defminmtu = IP6PO_MINMTU_MCASTONLY; int defpreftemp = IP6PO_TEMPADDR_SYSTEM; struct ip6_pktopts *pktopt; INP_RLOCK(inp); pktopt = inp->in6p_outputopts; switch (optname) { case IPV6_PKTINFO: optdata = (void *)&null_pktinfo; if (pktopt && pktopt->ip6po_pktinfo) { bcopy(pktopt->ip6po_pktinfo, &null_pktinfo, sizeof(null_pktinfo)); in6_clearscope(&null_pktinfo.ipi6_addr); } else { /* XXX: we don't have to do this every time... */ bzero(&null_pktinfo, sizeof(null_pktinfo)); } optdatalen = sizeof(struct in6_pktinfo); break; case IPV6_TCLASS: if (pktopt && pktopt->ip6po_tclass >= 0) deftclass = pktopt->ip6po_tclass; optdata = (void *)&deftclass; optdatalen = sizeof(int); break; case IPV6_HOPOPTS: GET_PKTOPT_EXT_HDR(ip6po_hbh); break; case IPV6_RTHDR: GET_PKTOPT_EXT_HDR(ip6po_rthdr); break; case IPV6_RTHDRDSTOPTS: GET_PKTOPT_EXT_HDR(ip6po_dest1); break; case IPV6_DSTOPTS: GET_PKTOPT_EXT_HDR(ip6po_dest2); break; case IPV6_NEXTHOP: GET_PKTOPT_SOCKADDR(ip6po_nexthop); break; case IPV6_USE_MIN_MTU: if (pktopt) defminmtu = pktopt->ip6po_minmtu; optdata = (void *)&defminmtu; optdatalen = sizeof(int); break; case IPV6_DONTFRAG: if (pktopt && ((pktopt->ip6po_flags) & IP6PO_DONTFRAG)) on = 1; else on = 0; optdata = (void *)&on; optdatalen = sizeof(on); break; case IPV6_PREFER_TEMPADDR: if (pktopt) defpreftemp = pktopt->ip6po_prefer_tempaddr; optdata = (void *)&defpreftemp; optdatalen = sizeof(int); break; default: /* should not happen */ #ifdef DIAGNOSTIC panic("ip6_getpcbopt: unexpected option\n"); #endif INP_RUNLOCK(inp); return (ENOPROTOOPT); } INP_RUNLOCK(inp); error = sooptcopyout(sopt, optdata, optdatalen); if (malloc_optdata) free(optdata, M_TEMP); return (error); } void ip6_clearpktopts(struct ip6_pktopts *pktopt, int optname) { if (pktopt == NULL) return; if (optname == -1 || optname == IPV6_PKTINFO) { if (pktopt->ip6po_pktinfo) free(pktopt->ip6po_pktinfo, M_IP6OPT); pktopt->ip6po_pktinfo = NULL; } if (optname == -1 || optname == IPV6_HOPLIMIT) pktopt->ip6po_hlim = -1; if (optname == -1 || optname == IPV6_TCLASS) pktopt->ip6po_tclass = -1; if (optname == -1 || optname == IPV6_NEXTHOP) { if (pktopt->ip6po_nextroute.ro_nh) { NH_FREE(pktopt->ip6po_nextroute.ro_nh); pktopt->ip6po_nextroute.ro_nh = NULL; } if (pktopt->ip6po_nexthop) free(pktopt->ip6po_nexthop, M_IP6OPT); pktopt->ip6po_nexthop = NULL; } if (optname == -1 || optname == IPV6_HOPOPTS) { if (pktopt->ip6po_hbh) free(pktopt->ip6po_hbh, M_IP6OPT); pktopt->ip6po_hbh = NULL; } if (optname == -1 || optname == IPV6_RTHDRDSTOPTS) { if (pktopt->ip6po_dest1) free(pktopt->ip6po_dest1, M_IP6OPT); pktopt->ip6po_dest1 = NULL; } if (optname == -1 || optname == IPV6_RTHDR) { if (pktopt->ip6po_rhinfo.ip6po_rhi_rthdr) free(pktopt->ip6po_rhinfo.ip6po_rhi_rthdr, M_IP6OPT); pktopt->ip6po_rhinfo.ip6po_rhi_rthdr = NULL; if (pktopt->ip6po_route.ro_nh) { NH_FREE(pktopt->ip6po_route.ro_nh); pktopt->ip6po_route.ro_nh = NULL; } } if (optname == -1 || optname == IPV6_DSTOPTS) { if (pktopt->ip6po_dest2) free(pktopt->ip6po_dest2, M_IP6OPT); pktopt->ip6po_dest2 = NULL; } } #define PKTOPT_EXTHDRCPY(type) \ do {\ if (src->type) {\ int hlen = (((struct ip6_ext *)src->type)->ip6e_len + 1) << 3;\ dst->type = malloc(hlen, M_IP6OPT, canwait);\ if (dst->type == NULL)\ goto bad;\ bcopy(src->type, dst->type, hlen);\ }\ } while (/*CONSTCOND*/ 0) static int copypktopts(struct ip6_pktopts *dst, struct ip6_pktopts *src, int canwait) { if (dst == NULL || src == NULL) { printf("ip6_clearpktopts: invalid argument\n"); return (EINVAL); } dst->ip6po_hlim = src->ip6po_hlim; dst->ip6po_tclass = src->ip6po_tclass; dst->ip6po_flags = src->ip6po_flags; dst->ip6po_minmtu = src->ip6po_minmtu; dst->ip6po_prefer_tempaddr = src->ip6po_prefer_tempaddr; if (src->ip6po_pktinfo) { dst->ip6po_pktinfo = malloc(sizeof(*dst->ip6po_pktinfo), M_IP6OPT, canwait); if (dst->ip6po_pktinfo == NULL) goto bad; *dst->ip6po_pktinfo = *src->ip6po_pktinfo; } if (src->ip6po_nexthop) { dst->ip6po_nexthop = malloc(src->ip6po_nexthop->sa_len, M_IP6OPT, canwait); if (dst->ip6po_nexthop == NULL) goto bad; bcopy(src->ip6po_nexthop, dst->ip6po_nexthop, src->ip6po_nexthop->sa_len); } PKTOPT_EXTHDRCPY(ip6po_hbh); PKTOPT_EXTHDRCPY(ip6po_dest1); PKTOPT_EXTHDRCPY(ip6po_dest2); PKTOPT_EXTHDRCPY(ip6po_rthdr); /* not copy the cached route */ return (0); bad: ip6_clearpktopts(dst, -1); return (ENOBUFS); } #undef PKTOPT_EXTHDRCPY struct ip6_pktopts * ip6_copypktopts(struct ip6_pktopts *src, int canwait) { int error; struct ip6_pktopts *dst; dst = malloc(sizeof(*dst), M_IP6OPT, canwait); if (dst == NULL) return (NULL); ip6_initpktopts(dst); if ((error = copypktopts(dst, src, canwait)) != 0) { free(dst, M_IP6OPT); return (NULL); } return (dst); } void ip6_freepcbopts(struct ip6_pktopts *pktopt) { if (pktopt == NULL) return; ip6_clearpktopts(pktopt, -1); free(pktopt, M_IP6OPT); } /* * Set IPv6 outgoing packet options based on advanced API. */ int ip6_setpktopts(struct mbuf *control, struct ip6_pktopts *opt, struct ip6_pktopts *stickyopt, struct ucred *cred, int uproto) { struct cmsghdr *cm = NULL; if (control == NULL || opt == NULL) return (EINVAL); /* * ip6_setpktopt can call ifnet_byindex(), so it's imperative that we * are in the network epoch here. */ NET_EPOCH_ASSERT(); ip6_initpktopts(opt); if (stickyopt) { int error; /* * If stickyopt is provided, make a local copy of the options * for this particular packet, then override them by ancillary * objects. * XXX: copypktopts() does not copy the cached route to a next * hop (if any). This is not very good in terms of efficiency, * but we can allow this since this option should be rarely * used. */ if ((error = copypktopts(opt, stickyopt, M_NOWAIT)) != 0) return (error); } /* * XXX: Currently, we assume all the optional information is stored * in a single mbuf. */ if (control->m_next) return (EINVAL); for (; control->m_len > 0; control->m_data += CMSG_ALIGN(cm->cmsg_len), control->m_len -= CMSG_ALIGN(cm->cmsg_len)) { int error; if (control->m_len < CMSG_LEN(0)) return (EINVAL); cm = mtod(control, struct cmsghdr *); if (cm->cmsg_len == 0 || cm->cmsg_len > control->m_len) return (EINVAL); if (cm->cmsg_level != IPPROTO_IPV6) continue; error = ip6_setpktopt(cm->cmsg_type, CMSG_DATA(cm), cm->cmsg_len - CMSG_LEN(0), opt, cred, 0, 1, uproto); if (error) return (error); } return (0); } /* * Set a particular packet option, as a sticky option or an ancillary data * item. "len" can be 0 only when it's a sticky option. * We have 4 cases of combination of "sticky" and "cmsg": * "sticky=0, cmsg=0": impossible * "sticky=0, cmsg=1": RFC2292 or RFC3542 ancillary data * "sticky=1, cmsg=0": RFC3542 socket option * "sticky=1, cmsg=1": RFC2292 socket option */ static int ip6_setpktopt(int optname, u_char *buf, int len, struct ip6_pktopts *opt, struct ucred *cred, int sticky, int cmsg, int uproto) { int minmtupolicy, preftemp; int error; NET_EPOCH_ASSERT(); if (!sticky && !cmsg) { #ifdef DIAGNOSTIC printf("ip6_setpktopt: impossible case\n"); #endif return (EINVAL); } /* * IPV6_2292xxx is for backward compatibility to RFC2292, and should * not be specified in the context of RFC3542. Conversely, * RFC3542 types should not be specified in the context of RFC2292. */ if (!cmsg) { switch (optname) { case IPV6_2292PKTINFO: case IPV6_2292HOPLIMIT: case IPV6_2292NEXTHOP: case IPV6_2292HOPOPTS: case IPV6_2292DSTOPTS: case IPV6_2292RTHDR: case IPV6_2292PKTOPTIONS: return (ENOPROTOOPT); } } if (sticky && cmsg) { switch (optname) { case IPV6_PKTINFO: case IPV6_HOPLIMIT: case IPV6_NEXTHOP: case IPV6_HOPOPTS: case IPV6_DSTOPTS: case IPV6_RTHDRDSTOPTS: case IPV6_RTHDR: case IPV6_USE_MIN_MTU: case IPV6_DONTFRAG: case IPV6_TCLASS: case IPV6_PREFER_TEMPADDR: /* XXX: not an RFC3542 option */ return (ENOPROTOOPT); } } switch (optname) { case IPV6_2292PKTINFO: case IPV6_PKTINFO: { struct ifnet *ifp = NULL; struct in6_pktinfo *pktinfo; if (len != sizeof(struct in6_pktinfo)) return (EINVAL); pktinfo = (struct in6_pktinfo *)buf; /* * An application can clear any sticky IPV6_PKTINFO option by * doing a "regular" setsockopt with ipi6_addr being * in6addr_any and ipi6_ifindex being zero. * [RFC 3542, Section 6] */ if (optname == IPV6_PKTINFO && opt->ip6po_pktinfo && pktinfo->ipi6_ifindex == 0 && IN6_IS_ADDR_UNSPECIFIED(&pktinfo->ipi6_addr)) { ip6_clearpktopts(opt, optname); break; } if (uproto == IPPROTO_TCP && optname == IPV6_PKTINFO && sticky && !IN6_IS_ADDR_UNSPECIFIED(&pktinfo->ipi6_addr)) { return (EINVAL); } if (IN6_IS_ADDR_MULTICAST(&pktinfo->ipi6_addr)) return (EINVAL); /* validate the interface index if specified. */ if (pktinfo->ipi6_ifindex) { ifp = ifnet_byindex(pktinfo->ipi6_ifindex); if (ifp == NULL) return (ENXIO); } if (ifp != NULL && (ifp->if_afdata[AF_INET6] == NULL || (ND_IFINFO(ifp)->flags & ND6_IFF_IFDISABLED) != 0)) return (ENETDOWN); if (ifp != NULL && !IN6_IS_ADDR_UNSPECIFIED(&pktinfo->ipi6_addr)) { struct in6_ifaddr *ia; in6_setscope(&pktinfo->ipi6_addr, ifp, NULL); ia = in6ifa_ifpwithaddr(ifp, &pktinfo->ipi6_addr); if (ia == NULL) return (EADDRNOTAVAIL); ifa_free(&ia->ia_ifa); } /* * We store the address anyway, and let in6_selectsrc() * validate the specified address. This is because ipi6_addr * may not have enough information about its scope zone, and * we may need additional information (such as outgoing * interface or the scope zone of a destination address) to * disambiguate the scope. * XXX: the delay of the validation may confuse the * application when it is used as a sticky option. */ if (opt->ip6po_pktinfo == NULL) { opt->ip6po_pktinfo = malloc(sizeof(*pktinfo), M_IP6OPT, M_NOWAIT); if (opt->ip6po_pktinfo == NULL) return (ENOBUFS); } bcopy(pktinfo, opt->ip6po_pktinfo, sizeof(*pktinfo)); break; } case IPV6_2292HOPLIMIT: case IPV6_HOPLIMIT: { int *hlimp; /* * RFC 3542 deprecated the usage of sticky IPV6_HOPLIMIT * to simplify the ordering among hoplimit options. */ if (optname == IPV6_HOPLIMIT && sticky) return (ENOPROTOOPT); if (len != sizeof(int)) return (EINVAL); hlimp = (int *)buf; if (*hlimp < -1 || *hlimp > 255) return (EINVAL); opt->ip6po_hlim = *hlimp; break; } case IPV6_TCLASS: { int tclass; if (len != sizeof(int)) return (EINVAL); tclass = *(int *)buf; if (tclass < -1 || tclass > 255) return (EINVAL); opt->ip6po_tclass = tclass; break; } case IPV6_2292NEXTHOP: case IPV6_NEXTHOP: if (cred != NULL) { error = priv_check_cred(cred, PRIV_NETINET_SETHDROPTS); if (error) return (error); } if (len == 0) { /* just remove the option */ ip6_clearpktopts(opt, IPV6_NEXTHOP); break; } /* check if cmsg_len is large enough for sa_len */ if (len < sizeof(struct sockaddr) || len < *buf) return (EINVAL); switch (((struct sockaddr *)buf)->sa_family) { case AF_INET6: { struct sockaddr_in6 *sa6 = (struct sockaddr_in6 *)buf; int error; if (sa6->sin6_len != sizeof(struct sockaddr_in6)) return (EINVAL); if (IN6_IS_ADDR_UNSPECIFIED(&sa6->sin6_addr) || IN6_IS_ADDR_MULTICAST(&sa6->sin6_addr)) { return (EINVAL); } if ((error = sa6_embedscope(sa6, V_ip6_use_defzone)) != 0) { return (error); } break; } case AF_LINK: /* should eventually be supported */ default: return (EAFNOSUPPORT); } /* turn off the previous option, then set the new option. */ ip6_clearpktopts(opt, IPV6_NEXTHOP); opt->ip6po_nexthop = malloc(*buf, M_IP6OPT, M_NOWAIT); if (opt->ip6po_nexthop == NULL) return (ENOBUFS); bcopy(buf, opt->ip6po_nexthop, *buf); break; case IPV6_2292HOPOPTS: case IPV6_HOPOPTS: { struct ip6_hbh *hbh; int hbhlen; /* * XXX: We don't allow a non-privileged user to set ANY HbH * options, since per-option restriction has too much * overhead. */ if (cred != NULL) { error = priv_check_cred(cred, PRIV_NETINET_SETHDROPTS); if (error) return (error); } if (len == 0) { ip6_clearpktopts(opt, IPV6_HOPOPTS); break; /* just remove the option */ } /* message length validation */ if (len < sizeof(struct ip6_hbh)) return (EINVAL); hbh = (struct ip6_hbh *)buf; hbhlen = (hbh->ip6h_len + 1) << 3; if (len != hbhlen) return (EINVAL); /* turn off the previous option, then set the new option. */ ip6_clearpktopts(opt, IPV6_HOPOPTS); opt->ip6po_hbh = malloc(hbhlen, M_IP6OPT, M_NOWAIT); if (opt->ip6po_hbh == NULL) return (ENOBUFS); bcopy(hbh, opt->ip6po_hbh, hbhlen); break; } case IPV6_2292DSTOPTS: case IPV6_DSTOPTS: case IPV6_RTHDRDSTOPTS: { struct ip6_dest *dest, **newdest = NULL; int destlen; if (cred != NULL) { /* XXX: see the comment for IPV6_HOPOPTS */ error = priv_check_cred(cred, PRIV_NETINET_SETHDROPTS); if (error) return (error); } if (len == 0) { ip6_clearpktopts(opt, optname); break; /* just remove the option */ } /* message length validation */ if (len < sizeof(struct ip6_dest)) return (EINVAL); dest = (struct ip6_dest *)buf; destlen = (dest->ip6d_len + 1) << 3; if (len != destlen) return (EINVAL); /* * Determine the position that the destination options header * should be inserted; before or after the routing header. */ switch (optname) { case IPV6_2292DSTOPTS: /* * The old advacned API is ambiguous on this point. * Our approach is to determine the position based * according to the existence of a routing header. * Note, however, that this depends on the order of the * extension headers in the ancillary data; the 1st * part of the destination options header must appear * before the routing header in the ancillary data, * too. * RFC3542 solved the ambiguity by introducing * separate ancillary data or option types. */ if (opt->ip6po_rthdr == NULL) newdest = &opt->ip6po_dest1; else newdest = &opt->ip6po_dest2; break; case IPV6_RTHDRDSTOPTS: newdest = &opt->ip6po_dest1; break; case IPV6_DSTOPTS: newdest = &opt->ip6po_dest2; break; } /* turn off the previous option, then set the new option. */ ip6_clearpktopts(opt, optname); *newdest = malloc(destlen, M_IP6OPT, M_NOWAIT); if (*newdest == NULL) return (ENOBUFS); bcopy(dest, *newdest, destlen); break; } case IPV6_2292RTHDR: case IPV6_RTHDR: { struct ip6_rthdr *rth; int rthlen; if (len == 0) { ip6_clearpktopts(opt, IPV6_RTHDR); break; /* just remove the option */ } /* message length validation */ if (len < sizeof(struct ip6_rthdr)) return (EINVAL); rth = (struct ip6_rthdr *)buf; rthlen = (rth->ip6r_len + 1) << 3; if (len != rthlen) return (EINVAL); switch (rth->ip6r_type) { case IPV6_RTHDR_TYPE_0: if (rth->ip6r_len == 0) /* must contain one addr */ return (EINVAL); if (rth->ip6r_len % 2) /* length must be even */ return (EINVAL); if (rth->ip6r_len / 2 != rth->ip6r_segleft) return (EINVAL); break; default: return (EINVAL); /* not supported */ } /* turn off the previous option */ ip6_clearpktopts(opt, IPV6_RTHDR); opt->ip6po_rthdr = malloc(rthlen, M_IP6OPT, M_NOWAIT); if (opt->ip6po_rthdr == NULL) return (ENOBUFS); bcopy(rth, opt->ip6po_rthdr, rthlen); break; } case IPV6_USE_MIN_MTU: if (len != sizeof(int)) return (EINVAL); minmtupolicy = *(int *)buf; if (minmtupolicy != IP6PO_MINMTU_MCASTONLY && minmtupolicy != IP6PO_MINMTU_DISABLE && minmtupolicy != IP6PO_MINMTU_ALL) { return (EINVAL); } opt->ip6po_minmtu = minmtupolicy; break; case IPV6_DONTFRAG: if (len != sizeof(int)) return (EINVAL); if (uproto == IPPROTO_TCP || *(int *)buf == 0) { /* * we ignore this option for TCP sockets. * (RFC3542 leaves this case unspecified.) */ opt->ip6po_flags &= ~IP6PO_DONTFRAG; } else opt->ip6po_flags |= IP6PO_DONTFRAG; break; case IPV6_PREFER_TEMPADDR: if (len != sizeof(int)) return (EINVAL); preftemp = *(int *)buf; if (preftemp != IP6PO_TEMPADDR_SYSTEM && preftemp != IP6PO_TEMPADDR_NOTPREFER && preftemp != IP6PO_TEMPADDR_PREFER) { return (EINVAL); } opt->ip6po_prefer_tempaddr = preftemp; break; default: return (ENOPROTOOPT); } /* end of switch */ return (0); } /* * Routine called from ip6_output() to loop back a copy of an IP6 multicast * packet to the input queue of a specified interface. Note that this * calls the output routine of the loopback "driver", but with an interface * pointer that might NOT be &loif -- easier than replicating that code here. */ void ip6_mloopback(struct ifnet *ifp, struct mbuf *m) { struct mbuf *copym; struct ip6_hdr *ip6; copym = m_copym(m, 0, M_COPYALL, M_NOWAIT); if (copym == NULL) return; /* * Make sure to deep-copy IPv6 header portion in case the data * is in an mbuf cluster, so that we can safely override the IPv6 * header portion later. */ if (!M_WRITABLE(copym) || copym->m_len < sizeof(struct ip6_hdr)) { copym = m_pullup(copym, sizeof(struct ip6_hdr)); if (copym == NULL) return; } ip6 = mtod(copym, struct ip6_hdr *); /* * clear embedded scope identifiers if necessary. * in6_clearscope will touch the addresses only when necessary. */ in6_clearscope(&ip6->ip6_src); in6_clearscope(&ip6->ip6_dst); if (copym->m_pkthdr.csum_flags & CSUM_DELAY_DATA_IPV6) { copym->m_pkthdr.csum_flags |= CSUM_DATA_VALID_IPV6 | CSUM_PSEUDO_HDR; copym->m_pkthdr.csum_data = 0xffff; } if_simloop(ifp, copym, AF_INET6, 0); } /* * Chop IPv6 header off from the payload. */ static int ip6_splithdr(struct mbuf *m, struct ip6_exthdrs *exthdrs) { struct mbuf *mh; struct ip6_hdr *ip6; ip6 = mtod(m, struct ip6_hdr *); if (m->m_len > sizeof(*ip6)) { mh = m_gethdr(M_NOWAIT, MT_DATA); if (mh == NULL) { m_freem(m); return ENOBUFS; } m_move_pkthdr(mh, m); M_ALIGN(mh, sizeof(*ip6)); m->m_len -= sizeof(*ip6); m->m_data += sizeof(*ip6); mh->m_next = m; m = mh; m->m_len = sizeof(*ip6); bcopy((caddr_t)ip6, mtod(m, caddr_t), sizeof(*ip6)); } exthdrs->ip6e_ip6 = m; return 0; } /* * Compute IPv6 extension header length. */ int ip6_optlen(struct inpcb *inp) { int len; if (!inp->in6p_outputopts) return 0; len = 0; #define elen(x) \ (((struct ip6_ext *)(x)) ? (((struct ip6_ext *)(x))->ip6e_len + 1) << 3 : 0) len += elen(inp->in6p_outputopts->ip6po_hbh); if (inp->in6p_outputopts->ip6po_rthdr) /* dest1 is valid with rthdr only */ len += elen(inp->in6p_outputopts->ip6po_dest1); len += elen(inp->in6p_outputopts->ip6po_rthdr); len += elen(inp->in6p_outputopts->ip6po_dest2); return len; #undef elen } diff --git a/sys/netipsec/xform_tcp.c b/sys/netipsec/xform_tcp.c index ce2552f0a205..42c63813e63c 100644 --- a/sys/netipsec/xform_tcp.c +++ b/sys/netipsec/xform_tcp.c @@ -1,429 +1,429 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 2003 Bruce M. Simpson * Copyright (c) 2016 Andrey V. Elsukov * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. The name of the author may not be used to endorse or promote products * derived from this software without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. */ /* TCP MD5 Signature Option (RFC2385) */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include "opt_ipsec.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef INET6 #include #include #endif #include #include #define TCP_SIGLEN 16 /* length of computed digest in bytes */ #define TCP_KEYLEN_MIN 1 /* minimum length of TCP-MD5 key */ #define TCP_KEYLEN_MAX 80 /* maximum length of TCP-MD5 key */ static int tcp_ipsec_pcbctl(struct inpcb *inp, struct sockopt *sopt) { struct tcpcb *tp; int error, optval; if (sopt->sopt_name != TCP_MD5SIG) { return (ENOPROTOOPT); } if (sopt->sopt_dir == SOPT_GET) { INP_RLOCK(inp); - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { INP_RUNLOCK(inp); return (ECONNRESET); } tp = intotcpcb(inp); optval = (tp->t_flags & TF_SIGNATURE) ? 1 : 0; INP_RUNLOCK(inp); /* On success return with released INP_WLOCK */ return (sooptcopyout(sopt, &optval, sizeof(optval))); } error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval)); if (error != 0) return (error); /* INP_WLOCK_RECHECK */ INP_WLOCK(inp); - if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { + if (inp->inp_flags & INP_DROPPED) { INP_WUNLOCK(inp); return (ECONNRESET); } tp = intotcpcb(inp); if (optval > 0) tp->t_flags |= TF_SIGNATURE; else tp->t_flags &= ~TF_SIGNATURE; INP_WUNLOCK(inp); return (error); } /* * Callback function invoked by m_apply() to digest TCP segment data * contained within an mbuf chain. */ static int tcp_signature_apply(void *fstate, void *data, u_int len) { MD5Update(fstate, (u_char *)data, len); return (0); } #ifdef INET static int ip_pseudo_compute(struct mbuf *m, MD5_CTX *ctx) { struct ippseudo ipp; struct ip *ip; ip = mtod(m, struct ip *); ipp.ippseudo_src.s_addr = ip->ip_src.s_addr; ipp.ippseudo_dst.s_addr = ip->ip_dst.s_addr; ipp.ippseudo_p = IPPROTO_TCP; ipp.ippseudo_pad = 0; ipp.ippseudo_len = htons(m->m_pkthdr.len - (ip->ip_hl << 2)); MD5Update(ctx, (char *)&ipp, sizeof(ipp)); return (ip->ip_hl << 2); } #endif #ifdef INET6 static int ip6_pseudo_compute(struct mbuf *m, MD5_CTX *ctx) { struct ip6_pseudo { struct in6_addr src, dst; uint32_t len; uint32_t nxt; } ip6p __aligned(4); struct ip6_hdr *ip6; ip6 = mtod(m, struct ip6_hdr *); ip6p.src = ip6->ip6_src; ip6p.dst = ip6->ip6_dst; ip6p.len = htonl(m->m_pkthdr.len - sizeof(*ip6)); /* XXX: ext headers */ ip6p.nxt = htonl(IPPROTO_TCP); MD5Update(ctx, (char *)&ip6p, sizeof(ip6p)); return (sizeof(*ip6)); } #endif static int tcp_signature_compute(struct mbuf *m, struct tcphdr *th, struct secasvar *sav, u_char *buf) { MD5_CTX ctx; int len; u_short csum; MD5Init(&ctx); /* Step 1: Update MD5 hash with IP(v6) pseudo-header. */ switch (sav->sah->saidx.dst.sa.sa_family) { #ifdef INET case AF_INET: len = ip_pseudo_compute(m, &ctx); break; #endif #ifdef INET6 case AF_INET6: len = ip6_pseudo_compute(m, &ctx); break; #endif default: return (EAFNOSUPPORT); } /* * Step 2: Update MD5 hash with TCP header, excluding options. * The TCP checksum must be set to zero. */ csum = th->th_sum; th->th_sum = 0; MD5Update(&ctx, (char *)th, sizeof(struct tcphdr)); th->th_sum = csum; /* * Step 3: Update MD5 hash with TCP segment data. * Use m_apply() to avoid an early m_pullup(). */ len += (th->th_off << 2); if (m->m_pkthdr.len - len > 0) m_apply(m, len, m->m_pkthdr.len - len, tcp_signature_apply, &ctx); /* * Step 4: Update MD5 hash with shared secret. */ MD5Update(&ctx, sav->key_auth->key_data, _KEYLEN(sav->key_auth)); MD5Final(buf, &ctx); key_sa_recordxfer(sav, m); return (0); } static void setsockaddrs(const struct mbuf *m, union sockaddr_union *src, union sockaddr_union *dst) { struct ip *ip; IPSEC_ASSERT(m->m_len >= sizeof(*ip), ("unexpected mbuf len")); ip = mtod(m, struct ip *); switch (ip->ip_v) { #ifdef INET case IPVERSION: ipsec4_setsockaddrs(m, src, dst); break; #endif #ifdef INET6 case (IPV6_VERSION >> 4): ipsec6_setsockaddrs(m, src, dst); break; #endif default: bzero(src, sizeof(*src)); bzero(dst, sizeof(*dst)); } } /* * Compute TCP-MD5 hash of an *INBOUND* TCP segment. * Parameters: * m pointer to head of mbuf chain * th pointer to TCP header * buf pointer to storage for computed MD5 digest * * Return 0 if successful, otherwise return error code. */ static int tcp_ipsec_input(struct mbuf *m, struct tcphdr *th, u_char *buf) { char tmpdigest[TCP_SIGLEN]; struct secasindex saidx; struct secasvar *sav; setsockaddrs(m, &saidx.src, &saidx.dst); saidx.proto = IPPROTO_TCP; saidx.mode = IPSEC_MODE_TCPMD5; saidx.reqid = 0; sav = key_allocsa_tcpmd5(&saidx); if (sav == NULL) { KMOD_TCPSTAT_INC(tcps_sig_err_buildsig); return (ENOENT); } if (buf == NULL) { key_freesav(&sav); KMOD_TCPSTAT_INC(tcps_sig_err_nosigopt); return (EACCES); } /* * tcp_input() operates with TCP header fields in host * byte order. We expect them in network byte order. */ tcp_fields_to_net(th); tcp_signature_compute(m, th, sav, tmpdigest); tcp_fields_to_host(th); key_freesav(&sav); if (bcmp(buf, tmpdigest, TCP_SIGLEN) != 0) { KMOD_TCPSTAT_INC(tcps_sig_rcvbadsig); return (EACCES); } KMOD_TCPSTAT_INC(tcps_sig_rcvgoodsig); return (0); } /* * Compute TCP-MD5 hash of an *OUTBOUND* TCP segment. * Parameters: * m pointer to head of mbuf chain * th pointer to TCP header * buf pointer to storage for computed MD5 digest * * Return 0 if successful, otherwise return error code. */ static int tcp_ipsec_output(struct mbuf *m, struct tcphdr *th, u_char *buf) { struct secasindex saidx; struct secasvar *sav; setsockaddrs(m, &saidx.src, &saidx.dst); saidx.proto = IPPROTO_TCP; saidx.mode = IPSEC_MODE_TCPMD5; saidx.reqid = 0; sav = key_allocsa_tcpmd5(&saidx); if (sav == NULL) { KMOD_TCPSTAT_INC(tcps_sig_err_buildsig); return (ENOENT); } tcp_signature_compute(m, th, sav, buf); key_freesav(&sav); return (0); } /* * Initialize a TCP-MD5 SA. Called when the SA is being set up. * * We don't need to set up the tdb prefixed fields, as we don't use the * opencrypto code; we just perform a key length check. * * XXX: Currently we have used single 'magic' SPI and need to still * support this. * * This allows per-host granularity without affecting the userland * interface, which is a simple socket option toggle switch, * TCP_SIGNATURE_ENABLE. * * To allow per-service granularity requires that we have a means * of mapping port to SPI. The mandated way of doing this is to * use SPD entries to specify packet flows which get the TCP-MD5 * treatment, however the code to do this is currently unstable * and unsuitable for production use. * * Therefore we use this compromise in the meantime. */ static int tcpsignature_init(struct secasvar *sav, struct xformsw *xsp) { int keylen; if (sav->alg_auth != SADB_X_AALG_TCP_MD5) { DPRINTF(("%s: unsupported authentication algorithm %u\n", __func__, sav->alg_auth)); return (EINVAL); } if (sav->key_auth == NULL) { DPRINTF(("%s: no authentication key present\n", __func__)); return (EINVAL); } keylen = _KEYLEN(sav->key_auth); if ((keylen < TCP_KEYLEN_MIN) || (keylen > TCP_KEYLEN_MAX)) { DPRINTF(("%s: invalid key length %u\n", __func__, keylen)); return (EINVAL); } sav->tdb_xform = xsp; return (0); } /* * Called when the SA is deleted. */ static void tcpsignature_cleanup(struct secasvar *sav) { } static struct xformsw tcpsignature_xformsw = { .xf_type = XF_TCPSIGNATURE, .xf_name = "TCP-MD5", .xf_init = tcpsignature_init, .xf_cleanup = tcpsignature_cleanup, }; static const struct tcpmd5_methods tcpmd5_methods = { .input = tcp_ipsec_input, .output = tcp_ipsec_output, .pcbctl = tcp_ipsec_pcbctl, }; #ifndef KLD_MODULE /* TCP-MD5 support is build in the kernel */ static const struct tcpmd5_support tcpmd5_ipsec = { .enabled = IPSEC_MODULE_ENABLED, .methods = &tcpmd5_methods }; const struct tcpmd5_support * const tcp_ipsec_support = &tcpmd5_ipsec; #endif /* !KLD_MODULE */ static int tcpmd5_modevent(module_t mod, int type, void *data) { switch (type) { case MOD_LOAD: xform_attach(&tcpsignature_xformsw); #ifdef KLD_MODULE tcpmd5_support_enable(&tcpmd5_methods); #endif break; case MOD_UNLOAD: #ifdef KLD_MODULE tcpmd5_support_disable(); #endif xform_detach(&tcpsignature_xformsw); break; default: return (EOPNOTSUPP); } return (0); } static moduledata_t tcpmd5_mod = { "tcpmd5", tcpmd5_modevent, 0 }; DECLARE_MODULE(tcpmd5, tcpmd5_mod, SI_SUB_PROTO_DOMAIN, SI_ORDER_ANY); MODULE_VERSION(tcpmd5, 1); #ifdef KLD_MODULE MODULE_DEPEND(tcpmd5, ipsec_support, 1, 1, 1); #endif