Index: head/sys/dev/cxgbe/tom/t4_tom.c =================================================================== --- head/sys/dev/cxgbe/tom/t4_tom.c (revision 356526) +++ head/sys/dev/cxgbe/tom/t4_tom.c (revision 356527) @@ -1,1933 +1,1933 @@ /*- * 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 "opt_kern_tls.h" #include "opt_ratelimit.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 #define TCPSTATES #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_regs_values.h" #include "common/t4_tcb.h" #include "t4_clip.h" #include "tom/t4_tom_l2t.h" #include "tom/t4_tom.h" #include "tom/t4_tls.h" static struct protosw toe_protosw; static struct pr_usrreqs toe_usrreqs; static struct protosw toe6_protosw; static struct pr_usrreqs toe6_usrreqs; /* Module ops */ static int t4_tom_mod_load(void); static int t4_tom_mod_unload(void); static int t4_tom_modevent(module_t, int, void *); /* ULD ops and helpers */ static int t4_tom_activate(struct adapter *); static int t4_tom_deactivate(struct adapter *); static struct uld_info tom_uld_info = { .uld_id = ULD_TOM, .activate = t4_tom_activate, .deactivate = t4_tom_deactivate, }; static void release_offload_resources(struct toepcb *); static int alloc_tid_tabs(struct tid_info *); static void free_tid_tabs(struct tid_info *); static void free_tom_data(struct adapter *, struct tom_data *); static void reclaim_wr_resources(void *, int); struct toepcb * alloc_toepcb(struct vi_info *vi, int flags) { struct port_info *pi = vi->pi; struct adapter *sc = pi->adapter; struct toepcb *toep; int tx_credits, txsd_total, len; /* * The firmware counts tx work request credits in units of 16 bytes * each. Reserve room for an ABORT_REQ so the driver never has to worry * about tx credits if it wants to abort a connection. */ tx_credits = sc->params.ofldq_wr_cred; tx_credits -= howmany(sizeof(struct cpl_abort_req), 16); /* * Shortest possible tx work request is a fw_ofld_tx_data_wr + 1 byte * immediate payload, and firmware counts tx work request credits in * units of 16 byte. Calculate the maximum work requests possible. */ txsd_total = tx_credits / howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16); len = offsetof(struct toepcb, txsd) + txsd_total * sizeof(struct ofld_tx_sdesc); toep = malloc(len, M_CXGBE, M_ZERO | flags); if (toep == NULL) return (NULL); refcount_init(&toep->refcount, 1); toep->td = sc->tom_softc; toep->vi = vi; toep->tid = -1; toep->tx_total = tx_credits; toep->tx_credits = tx_credits; mbufq_init(&toep->ulp_pduq, INT_MAX); mbufq_init(&toep->ulp_pdu_reclaimq, INT_MAX); toep->txsd_total = txsd_total; toep->txsd_avail = txsd_total; toep->txsd_pidx = 0; toep->txsd_cidx = 0; aiotx_init_toep(toep); return (toep); } /* * Initialize a toepcb after its params have been filled out. */ int init_toepcb(struct vi_info *vi, struct toepcb *toep) { struct conn_params *cp = &toep->params; struct port_info *pi = vi->pi; struct adapter *sc = pi->adapter; struct tx_cl_rl_params *tc; if (cp->tc_idx >= 0 && cp->tc_idx < sc->chip_params->nsched_cls) { tc = &pi->sched_params->cl_rl[cp->tc_idx]; mtx_lock(&sc->tc_lock); if (tc->flags & CLRL_ERR) { log(LOG_ERR, "%s: failed to associate traffic class %u with tid %u\n", device_get_nameunit(vi->dev), cp->tc_idx, toep->tid); cp->tc_idx = -1; } else { tc->refcount++; } mtx_unlock(&sc->tc_lock); } toep->ofld_txq = &sc->sge.ofld_txq[cp->txq_idx]; toep->ofld_rxq = &sc->sge.ofld_rxq[cp->rxq_idx]; toep->ctrlq = &sc->sge.ctrlq[pi->port_id]; tls_init_toep(toep); if (ulp_mode(toep) == ULP_MODE_TCPDDP) ddp_init_toep(toep); return (0); } struct toepcb * hold_toepcb(struct toepcb *toep) { refcount_acquire(&toep->refcount); return (toep); } void free_toepcb(struct toepcb *toep) { if (refcount_release(&toep->refcount) == 0) return; KASSERT(!(toep->flags & TPF_ATTACHED), ("%s: attached to an inpcb", __func__)); KASSERT(!(toep->flags & TPF_CPL_PENDING), ("%s: CPL pending", __func__)); if (ulp_mode(toep) == ULP_MODE_TCPDDP) ddp_uninit_toep(toep); tls_uninit_toep(toep); free(toep, M_CXGBE); } /* * Set up the socket for TCP offload. */ void offload_socket(struct socket *so, struct toepcb *toep) { struct tom_data *td = toep->td; struct inpcb *inp = sotoinpcb(so); struct tcpcb *tp = intotcpcb(inp); struct sockbuf *sb; INP_WLOCK_ASSERT(inp); /* Update socket */ sb = &so->so_snd; SOCKBUF_LOCK(sb); sb->sb_flags |= SB_NOCOALESCE; SOCKBUF_UNLOCK(sb); sb = &so->so_rcv; SOCKBUF_LOCK(sb); sb->sb_flags |= SB_NOCOALESCE; if (inp->inp_vflag & INP_IPV6) so->so_proto = &toe6_protosw; else so->so_proto = &toe_protosw; SOCKBUF_UNLOCK(sb); /* Update TCP PCB */ tp->tod = &td->tod; tp->t_toe = toep; tp->t_flags |= TF_TOE; /* Install an extra hold on inp */ toep->inp = inp; toep->flags |= TPF_ATTACHED; in_pcbref(inp); /* Add the TOE PCB to the active list */ mtx_lock(&td->toep_list_lock); TAILQ_INSERT_HEAD(&td->toep_list, toep, link); mtx_unlock(&td->toep_list_lock); } /* This is _not_ the normal way to "unoffload" a socket. */ void undo_offload_socket(struct socket *so) { struct inpcb *inp = sotoinpcb(so); struct tcpcb *tp = intotcpcb(inp); struct toepcb *toep = tp->t_toe; struct tom_data *td = toep->td; struct sockbuf *sb; INP_WLOCK_ASSERT(inp); sb = &so->so_snd; SOCKBUF_LOCK(sb); sb->sb_flags &= ~SB_NOCOALESCE; SOCKBUF_UNLOCK(sb); sb = &so->so_rcv; SOCKBUF_LOCK(sb); sb->sb_flags &= ~SB_NOCOALESCE; SOCKBUF_UNLOCK(sb); tp->tod = NULL; tp->t_toe = NULL; tp->t_flags &= ~TF_TOE; toep->inp = NULL; toep->flags &= ~TPF_ATTACHED; if (in_pcbrele_wlocked(inp)) panic("%s: inp freed.", __func__); mtx_lock(&td->toep_list_lock); TAILQ_REMOVE(&td->toep_list, toep, link); mtx_unlock(&td->toep_list_lock); } static void release_offload_resources(struct toepcb *toep) { struct tom_data *td = toep->td; struct adapter *sc = td_adapter(td); int tid = toep->tid; KASSERT(!(toep->flags & TPF_CPL_PENDING), ("%s: %p has CPL pending.", __func__, toep)); KASSERT(!(toep->flags & TPF_ATTACHED), ("%s: %p is still attached.", __func__, toep)); CTR5(KTR_CXGBE, "%s: toep %p (tid %d, l2te %p, ce %p)", __func__, toep, tid, toep->l2te, toep->ce); /* * These queues should have been emptied at approximately the same time * that a normal connection's socket's so_snd would have been purged or * drained. Do _not_ clean up here. */ MPASS(mbufq_len(&toep->ulp_pduq) == 0); MPASS(mbufq_len(&toep->ulp_pdu_reclaimq) == 0); #ifdef INVARIANTS if (ulp_mode(toep) == ULP_MODE_TCPDDP) ddp_assert_empty(toep); #endif MPASS(TAILQ_EMPTY(&toep->aiotx_jobq)); if (toep->l2te) t4_l2t_release(toep->l2te); if (tid >= 0) { remove_tid(sc, tid, toep->ce ? 2 : 1); release_tid(sc, tid, toep->ctrlq); } if (toep->ce) t4_release_lip(sc, toep->ce); if (toep->params.tc_idx != -1) t4_release_cl_rl(sc, toep->vi->pi->port_id, toep->params.tc_idx); mtx_lock(&td->toep_list_lock); TAILQ_REMOVE(&td->toep_list, toep, link); mtx_unlock(&td->toep_list_lock); free_toepcb(toep); } /* * The kernel is done with the TCP PCB and this is our opportunity to unhook the * toepcb hanging off of it. If the TOE driver is also done with the toepcb (no * pending CPL) then it is time to release all resources tied to the toepcb. * * Also gets called when an offloaded active open fails and the TOM wants the * kernel to take the TCP PCB back. */ static void t4_pcb_detach(struct toedev *tod __unused, struct tcpcb *tp) { #if defined(KTR) || defined(INVARIANTS) struct inpcb *inp = tp->t_inpcb; #endif struct toepcb *toep = tp->t_toe; INP_WLOCK_ASSERT(inp); KASSERT(toep != NULL, ("%s: toep is NULL", __func__)); KASSERT(toep->flags & TPF_ATTACHED, ("%s: not attached", __func__)); #ifdef KTR if (tp->t_state == TCPS_SYN_SENT) { CTR6(KTR_CXGBE, "%s: atid %d, toep %p (0x%x), inp %p (0x%x)", __func__, toep->tid, toep, toep->flags, inp, inp->inp_flags); } else { CTR6(KTR_CXGBE, "t4_pcb_detach: tid %d (%s), toep %p (0x%x), inp %p (0x%x)", toep->tid, tcpstates[tp->t_state], toep, toep->flags, inp, inp->inp_flags); } #endif tp->t_toe = NULL; tp->t_flags &= ~TF_TOE; toep->flags &= ~TPF_ATTACHED; if (!(toep->flags & TPF_CPL_PENDING)) release_offload_resources(toep); } /* * setsockopt handler. */ static void t4_ctloutput(struct toedev *tod, struct tcpcb *tp, int dir, int name) { struct adapter *sc = tod->tod_softc; struct toepcb *toep = tp->t_toe; if (dir == SOPT_GET) return; CTR4(KTR_CXGBE, "%s: tp %p, dir %u, name %u", __func__, tp, dir, name); switch (name) { case TCP_NODELAY: if (tp->t_state != TCPS_ESTABLISHED) break; toep->params.nagle = tp->t_flags & TF_NODELAY ? 0 : 1; t4_set_tcb_field(sc, toep->ctrlq, toep, W_TCB_T_FLAGS, V_TF_NAGLE(1), V_TF_NAGLE(toep->params.nagle), 0, 0); break; default: break; } } static inline uint64_t get_tcb_tflags(const uint64_t *tcb) { return ((be64toh(tcb[14]) << 32) | (be64toh(tcb[15]) >> 32)); } static inline uint32_t get_tcb_field(const uint64_t *tcb, u_int word, uint32_t mask, u_int shift) { #define LAST_WORD ((TCB_SIZE / 4) - 1) uint64_t t1, t2; int flit_idx; MPASS(mask != 0); MPASS(word <= LAST_WORD); MPASS(shift < 32); flit_idx = (LAST_WORD - word) / 2; if (word & 0x1) shift += 32; t1 = be64toh(tcb[flit_idx]) >> shift; t2 = 0; if (fls(mask) > 64 - shift) { /* * Will spill over into the next logical flit, which is the flit * before this one. The flit_idx before this one must be valid. */ MPASS(flit_idx > 0); t2 = be64toh(tcb[flit_idx - 1]) << (64 - shift); } return ((t2 | t1) & mask); #undef LAST_WORD } #define GET_TCB_FIELD(tcb, F) \ get_tcb_field(tcb, W_TCB_##F, M_TCB_##F, S_TCB_##F) /* * Issues a CPL_GET_TCB to read the entire TCB for the tid. */ static int send_get_tcb(struct adapter *sc, u_int tid) { struct cpl_get_tcb *cpl; struct wrq_cookie cookie; MPASS(tid < sc->tids.ntids); cpl = start_wrq_wr(&sc->sge.ctrlq[0], howmany(sizeof(*cpl), 16), &cookie); if (__predict_false(cpl == NULL)) return (ENOMEM); bzero(cpl, sizeof(*cpl)); INIT_TP_WR(cpl, tid); OPCODE_TID(cpl) = htobe32(MK_OPCODE_TID(CPL_GET_TCB, tid)); cpl->reply_ctrl = htobe16(V_REPLY_CHAN(0) | V_QUEUENO(sc->sge.ofld_rxq[0].iq.cntxt_id)); cpl->cookie = 0xff; commit_wrq_wr(&sc->sge.ctrlq[0], cpl, &cookie); return (0); } static struct tcb_histent * alloc_tcb_histent(struct adapter *sc, u_int tid, int flags) { struct tcb_histent *te; MPASS(flags == M_NOWAIT || flags == M_WAITOK); te = malloc(sizeof(*te), M_CXGBE, M_ZERO | flags); if (te == NULL) return (NULL); mtx_init(&te->te_lock, "TCB entry", NULL, MTX_DEF); callout_init_mtx(&te->te_callout, &te->te_lock, 0); te->te_adapter = sc; te->te_tid = tid; return (te); } static void free_tcb_histent(struct tcb_histent *te) { mtx_destroy(&te->te_lock); free(te, M_CXGBE); } /* * Start tracking the tid in the TCB history. */ int add_tid_to_history(struct adapter *sc, u_int tid) { struct tcb_histent *te = NULL; struct tom_data *td = sc->tom_softc; int rc; MPASS(tid < sc->tids.ntids); if (td->tcb_history == NULL) return (ENXIO); rw_wlock(&td->tcb_history_lock); if (td->tcb_history[tid] != NULL) { rc = EEXIST; goto done; } te = alloc_tcb_histent(sc, tid, M_NOWAIT); if (te == NULL) { rc = ENOMEM; goto done; } mtx_lock(&te->te_lock); rc = send_get_tcb(sc, tid); if (rc == 0) { te->te_flags |= TE_RPL_PENDING; td->tcb_history[tid] = te; } else { free(te, M_CXGBE); } mtx_unlock(&te->te_lock); done: rw_wunlock(&td->tcb_history_lock); return (rc); } static void remove_tcb_histent(struct tcb_histent *te) { struct adapter *sc = te->te_adapter; struct tom_data *td = sc->tom_softc; rw_assert(&td->tcb_history_lock, RA_WLOCKED); mtx_assert(&te->te_lock, MA_OWNED); MPASS(td->tcb_history[te->te_tid] == te); td->tcb_history[te->te_tid] = NULL; free_tcb_histent(te); rw_wunlock(&td->tcb_history_lock); } static inline struct tcb_histent * lookup_tcb_histent(struct adapter *sc, u_int tid, bool addrem) { struct tcb_histent *te; struct tom_data *td = sc->tom_softc; MPASS(tid < sc->tids.ntids); if (td->tcb_history == NULL) return (NULL); if (addrem) rw_wlock(&td->tcb_history_lock); else rw_rlock(&td->tcb_history_lock); te = td->tcb_history[tid]; if (te != NULL) { mtx_lock(&te->te_lock); return (te); /* with both locks held */ } if (addrem) rw_wunlock(&td->tcb_history_lock); else rw_runlock(&td->tcb_history_lock); return (te); } static inline void release_tcb_histent(struct tcb_histent *te) { struct adapter *sc = te->te_adapter; struct tom_data *td = sc->tom_softc; mtx_assert(&te->te_lock, MA_OWNED); mtx_unlock(&te->te_lock); rw_assert(&td->tcb_history_lock, RA_RLOCKED); rw_runlock(&td->tcb_history_lock); } static void request_tcb(void *arg) { struct tcb_histent *te = arg; mtx_assert(&te->te_lock, MA_OWNED); /* Noone else is supposed to update the histent. */ MPASS(!(te->te_flags & TE_RPL_PENDING)); if (send_get_tcb(te->te_adapter, te->te_tid) == 0) te->te_flags |= TE_RPL_PENDING; else callout_schedule(&te->te_callout, hz / 100); } static void update_tcb_histent(struct tcb_histent *te, const uint64_t *tcb) { struct tom_data *td = te->te_adapter->tom_softc; uint64_t tflags = get_tcb_tflags(tcb); uint8_t sample = 0; if (GET_TCB_FIELD(tcb, SND_MAX_RAW) != GET_TCB_FIELD(tcb, SND_UNA_RAW)) { if (GET_TCB_FIELD(tcb, T_RXTSHIFT) != 0) sample |= TS_RTO; if (GET_TCB_FIELD(tcb, T_DUPACKS) != 0) sample |= TS_DUPACKS; if (GET_TCB_FIELD(tcb, T_DUPACKS) >= td->dupack_threshold) sample |= TS_FASTREXMT; } if (GET_TCB_FIELD(tcb, SND_MAX_RAW) != 0) { uint32_t snd_wnd; sample |= TS_SND_BACKLOGGED; /* for whatever reason. */ snd_wnd = GET_TCB_FIELD(tcb, RCV_ADV); if (tflags & V_TF_RECV_SCALE(1)) snd_wnd <<= GET_TCB_FIELD(tcb, RCV_SCALE); if (GET_TCB_FIELD(tcb, SND_CWND) < snd_wnd) sample |= TS_CWND_LIMITED; /* maybe due to CWND */ } if (tflags & V_TF_CCTRL_ECN(1)) { /* * CE marker on incoming IP hdr, echoing ECE back in the TCP * hdr. Indicates congestion somewhere on the way from the peer * to this node. */ if (tflags & V_TF_CCTRL_ECE(1)) sample |= TS_ECN_ECE; /* * ECE seen and CWR sent (or about to be sent). Might indicate * congestion on the way to the peer. This node is reducing its * congestion window in response. */ if (tflags & (V_TF_CCTRL_CWR(1) | V_TF_CCTRL_RFR(1))) sample |= TS_ECN_CWR; } te->te_sample[te->te_pidx] = sample; if (++te->te_pidx == nitems(te->te_sample)) te->te_pidx = 0; memcpy(te->te_tcb, tcb, TCB_SIZE); te->te_flags |= TE_ACTIVE; } static int do_get_tcb_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { struct adapter *sc = iq->adapter; const struct cpl_get_tcb_rpl *cpl = mtod(m, const void *); const uint64_t *tcb = (const uint64_t *)(const void *)(cpl + 1); struct tcb_histent *te; const u_int tid = GET_TID(cpl); bool remove; remove = GET_TCB_FIELD(tcb, T_STATE) == TCPS_CLOSED; te = lookup_tcb_histent(sc, tid, remove); if (te == NULL) { /* Not in the history. Who issued the GET_TCB for this? */ device_printf(sc->dev, "tcb %u: flags 0x%016jx, state %u, " "srtt %u, sscale %u, rscale %u, cookie 0x%x\n", tid, (uintmax_t)get_tcb_tflags(tcb), GET_TCB_FIELD(tcb, T_STATE), GET_TCB_FIELD(tcb, T_SRTT), GET_TCB_FIELD(tcb, SND_SCALE), GET_TCB_FIELD(tcb, RCV_SCALE), cpl->cookie); goto done; } MPASS(te->te_flags & TE_RPL_PENDING); te->te_flags &= ~TE_RPL_PENDING; if (remove) { remove_tcb_histent(te); } else { update_tcb_histent(te, tcb); callout_reset(&te->te_callout, hz / 10, request_tcb, te); release_tcb_histent(te); } done: m_freem(m); return (0); } static void fill_tcp_info_from_tcb(struct adapter *sc, uint64_t *tcb, struct tcp_info *ti) { uint32_t v; ti->tcpi_state = GET_TCB_FIELD(tcb, T_STATE); v = GET_TCB_FIELD(tcb, T_SRTT); ti->tcpi_rtt = tcp_ticks_to_us(sc, v); v = GET_TCB_FIELD(tcb, T_RTTVAR); ti->tcpi_rttvar = tcp_ticks_to_us(sc, v); ti->tcpi_snd_ssthresh = GET_TCB_FIELD(tcb, SND_SSTHRESH); ti->tcpi_snd_cwnd = GET_TCB_FIELD(tcb, SND_CWND); ti->tcpi_rcv_nxt = GET_TCB_FIELD(tcb, RCV_NXT); v = GET_TCB_FIELD(tcb, TX_MAX); ti->tcpi_snd_nxt = v - GET_TCB_FIELD(tcb, SND_NXT_RAW); /* Receive window being advertised by us. */ ti->tcpi_rcv_wscale = GET_TCB_FIELD(tcb, SND_SCALE); /* Yes, SND. */ ti->tcpi_rcv_space = GET_TCB_FIELD(tcb, RCV_WND); /* Send window */ ti->tcpi_snd_wscale = GET_TCB_FIELD(tcb, RCV_SCALE); /* Yes, RCV. */ ti->tcpi_snd_wnd = GET_TCB_FIELD(tcb, RCV_ADV); if (get_tcb_tflags(tcb) & V_TF_RECV_SCALE(1)) ti->tcpi_snd_wnd <<= ti->tcpi_snd_wscale; else ti->tcpi_snd_wscale = 0; } static void fill_tcp_info_from_history(struct adapter *sc, struct tcb_histent *te, struct tcp_info *ti) { fill_tcp_info_from_tcb(sc, te->te_tcb, ti); } /* * Reads the TCB for the given tid using a memory window and copies it to 'buf' * in the same format as CPL_GET_TCB_RPL. */ static void read_tcb_using_memwin(struct adapter *sc, u_int tid, uint64_t *buf) { int i, j, k, rc; uint32_t addr; u_char *tcb, tmp; MPASS(tid < sc->tids.ntids); addr = t4_read_reg(sc, A_TP_CMM_TCB_BASE) + tid * TCB_SIZE; rc = read_via_memwin(sc, 2, addr, (uint32_t *)buf, TCB_SIZE); if (rc != 0) return; tcb = (u_char *)buf; for (i = 0, j = TCB_SIZE - 16; i < j; i += 16, j -= 16) { for (k = 0; k < 16; k++) { tmp = tcb[i + k]; tcb[i + k] = tcb[j + k]; tcb[j + k] = tmp; } } } static void fill_tcp_info(struct adapter *sc, u_int tid, struct tcp_info *ti) { uint64_t tcb[TCB_SIZE / sizeof(uint64_t)]; struct tcb_histent *te; ti->tcpi_toe_tid = tid; te = lookup_tcb_histent(sc, tid, false); if (te != NULL) { fill_tcp_info_from_history(sc, te, ti); release_tcb_histent(te); } else { if (!(sc->debug_flags & DF_DISABLE_TCB_CACHE)) { /* XXX: tell firmware to flush TCB cache. */ } read_tcb_using_memwin(sc, tid, tcb); fill_tcp_info_from_tcb(sc, tcb, ti); } } /* * Called by the kernel to allow the TOE driver to "refine" values filled up in * the tcp_info for an offloaded connection. */ static void t4_tcp_info(struct toedev *tod, struct tcpcb *tp, struct tcp_info *ti) { struct adapter *sc = tod->tod_softc; struct toepcb *toep = tp->t_toe; INP_WLOCK_ASSERT(tp->t_inpcb); MPASS(ti != NULL); fill_tcp_info(sc, toep->tid, ti); } #ifdef KERN_TLS static int t4_alloc_tls_session(struct toedev *tod, struct tcpcb *tp, struct ktls_session *tls) { struct toepcb *toep = tp->t_toe; INP_WLOCK_ASSERT(tp->t_inpcb); MPASS(tls != NULL); return (tls_alloc_ktls(toep, tls)); } #endif /* * The TOE driver will not receive any more CPLs for the tid associated with the * toepcb; release the hold on the inpcb. */ void final_cpl_received(struct toepcb *toep) { struct inpcb *inp = toep->inp; KASSERT(inp != NULL, ("%s: inp is NULL", __func__)); INP_WLOCK_ASSERT(inp); KASSERT(toep->flags & TPF_CPL_PENDING, ("%s: CPL not pending already?", __func__)); CTR6(KTR_CXGBE, "%s: tid %d, toep %p (0x%x), inp %p (0x%x)", __func__, toep->tid, toep, toep->flags, inp, inp->inp_flags); if (ulp_mode(toep) == ULP_MODE_TCPDDP) release_ddp_resources(toep); toep->inp = NULL; toep->flags &= ~TPF_CPL_PENDING; mbufq_drain(&toep->ulp_pdu_reclaimq); if (!(toep->flags & TPF_ATTACHED)) release_offload_resources(toep); if (!in_pcbrele_wlocked(inp)) INP_WUNLOCK(inp); } void insert_tid(struct adapter *sc, int tid, void *ctx, int ntids) { struct tid_info *t = &sc->tids; MPASS(tid >= t->tid_base); MPASS(tid - t->tid_base < t->ntids); t->tid_tab[tid - t->tid_base] = ctx; atomic_add_int(&t->tids_in_use, ntids); } void * lookup_tid(struct adapter *sc, int tid) { struct tid_info *t = &sc->tids; return (t->tid_tab[tid - t->tid_base]); } void update_tid(struct adapter *sc, int tid, void *ctx) { struct tid_info *t = &sc->tids; t->tid_tab[tid - t->tid_base] = ctx; } void remove_tid(struct adapter *sc, int tid, int ntids) { struct tid_info *t = &sc->tids; t->tid_tab[tid - t->tid_base] = NULL; atomic_subtract_int(&t->tids_in_use, ntids); } /* * What mtu_idx to use, given a 4-tuple. Note that both s->mss and tcp_mssopt * have the MSS that we should advertise in our SYN. Advertised MSS doesn't * account for any TCP options so the effective MSS (only payload, no headers or * options) could be different. */ static int find_best_mtu_idx(struct adapter *sc, struct in_conninfo *inc, struct offload_settings *s) { unsigned short *mtus = &sc->params.mtus[0]; int i, mss, mtu; MPASS(inc != NULL); mss = s->mss > 0 ? s->mss : tcp_mssopt(inc); if (inc->inc_flags & INC_ISIPV6) mtu = mss + sizeof(struct ip6_hdr) + sizeof(struct tcphdr); else mtu = mss + sizeof(struct ip) + sizeof(struct tcphdr); for (i = 0; i < NMTUS - 1 && mtus[i + 1] <= mtu; i++) continue; return (i); } /* * Determine the receive window size for a socket. */ u_long select_rcv_wnd(struct socket *so) { unsigned long wnd; SOCKBUF_LOCK_ASSERT(&so->so_rcv); wnd = sbspace(&so->so_rcv); if (wnd < MIN_RCV_WND) wnd = MIN_RCV_WND; return min(wnd, MAX_RCV_WND); } int select_rcv_wscale(void) { int wscale = 0; unsigned long space = sb_max; if (space > MAX_RCV_WND) space = MAX_RCV_WND; while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < space) wscale++; return (wscale); } __be64 calc_options0(struct vi_info *vi, struct conn_params *cp) { uint64_t opt0 = 0; opt0 |= F_TCAM_BYPASS; MPASS(cp->wscale >= 0 && cp->wscale <= M_WND_SCALE); opt0 |= V_WND_SCALE(cp->wscale); MPASS(cp->mtu_idx >= 0 && cp->mtu_idx < NMTUS); opt0 |= V_MSS_IDX(cp->mtu_idx); MPASS(cp->ulp_mode >= 0 && cp->ulp_mode <= M_ULP_MODE); opt0 |= V_ULP_MODE(cp->ulp_mode); MPASS(cp->opt0_bufsize >= 0 && cp->opt0_bufsize <= M_RCV_BUFSIZ); opt0 |= V_RCV_BUFSIZ(cp->opt0_bufsize); MPASS(cp->l2t_idx >= 0 && cp->l2t_idx < vi->pi->adapter->vres.l2t.size); opt0 |= V_L2T_IDX(cp->l2t_idx); opt0 |= V_SMAC_SEL(vi->smt_idx); opt0 |= V_TX_CHAN(vi->pi->tx_chan); MPASS(cp->keepalive == 0 || cp->keepalive == 1); opt0 |= V_KEEP_ALIVE(cp->keepalive); MPASS(cp->nagle == 0 || cp->nagle == 1); opt0 |= V_NAGLE(cp->nagle); return (htobe64(opt0)); } __be32 calc_options2(struct vi_info *vi, struct conn_params *cp) { uint32_t opt2 = 0; struct port_info *pi = vi->pi; struct adapter *sc = pi->adapter; /* * rx flow control, rx coalesce, congestion control, and tx pace are all * explicitly set by the driver. On T5+ the ISS is also set by the * driver to the value picked by the kernel. */ if (is_t4(sc)) { opt2 |= F_RX_FC_VALID | F_RX_COALESCE_VALID; opt2 |= F_CONG_CNTRL_VALID | F_PACE_VALID; } else { opt2 |= F_T5_OPT_2_VALID; /* all 4 valid */ opt2 |= F_T5_ISS; /* ISS provided in CPL */ } MPASS(cp->sack == 0 || cp->sack == 1); opt2 |= V_SACK_EN(cp->sack); MPASS(cp->tstamp == 0 || cp->tstamp == 1); opt2 |= V_TSTAMPS_EN(cp->tstamp); if (cp->wscale > 0) opt2 |= F_WND_SCALE_EN; MPASS(cp->ecn == 0 || cp->ecn == 1); opt2 |= V_CCTRL_ECN(cp->ecn); /* XXX: F_RX_CHANNEL for multiple rx c-chan support goes here. */ opt2 |= V_TX_QUEUE(sc->params.tp.tx_modq[pi->tx_chan]); opt2 |= V_PACE(0); opt2 |= F_RSS_QUEUE_VALID; opt2 |= V_RSS_QUEUE(sc->sge.ofld_rxq[cp->rxq_idx].iq.abs_id); MPASS(cp->cong_algo >= 0 && cp->cong_algo <= M_CONG_CNTRL); opt2 |= V_CONG_CNTRL(cp->cong_algo); MPASS(cp->rx_coalesce == 0 || cp->rx_coalesce == 1); if (cp->rx_coalesce == 1) opt2 |= V_RX_COALESCE(M_RX_COALESCE); opt2 |= V_RX_FC_DDP(0) | V_RX_FC_DISABLE(0); #ifdef USE_DDP_RX_FLOW_CONTROL if (cp->ulp_mode == ULP_MODE_TCPDDP) opt2 |= F_RX_FC_DDP; #endif if (cp->ulp_mode == ULP_MODE_TLS) opt2 |= F_RX_FC_DISABLE; return (htobe32(opt2)); } uint64_t select_ntuple(struct vi_info *vi, struct l2t_entry *e) { struct adapter *sc = vi->pi->adapter; struct tp_params *tp = &sc->params.tp; uint64_t ntuple = 0; /* * Initialize each of the fields which we care about which are present * in the Compressed Filter Tuple. */ if (tp->vlan_shift >= 0 && EVL_VLANOFTAG(e->vlan) != CPL_L2T_VLAN_NONE) ntuple |= (uint64_t)(F_FT_VLAN_VLD | e->vlan) << tp->vlan_shift; if (tp->port_shift >= 0) ntuple |= (uint64_t)e->lport << tp->port_shift; if (tp->protocol_shift >= 0) ntuple |= (uint64_t)IPPROTO_TCP << tp->protocol_shift; if (tp->vnic_shift >= 0 && tp->ingress_config & F_VNIC) { ntuple |= (uint64_t)(V_FT_VNID_ID_VF(vi->vin) | V_FT_VNID_ID_PF(sc->pf) | V_FT_VNID_ID_VLD(vi->vfvld)) << tp->vnic_shift; } if (is_t4(sc)) return (htobe32((uint32_t)ntuple)); else return (htobe64(V_FILTER_TUPLE(ntuple))); } static int is_tls_sock(struct socket *so, struct adapter *sc) { struct inpcb *inp = sotoinpcb(so); int i, rc; /* XXX: Eventually add a SO_WANT_TLS socket option perhaps? */ rc = 0; ADAPTER_LOCK(sc); for (i = 0; i < sc->tt.num_tls_rx_ports; i++) { if (inp->inp_lport == htons(sc->tt.tls_rx_ports[i]) || inp->inp_fport == htons(sc->tt.tls_rx_ports[i])) { rc = 1; break; } } ADAPTER_UNLOCK(sc); return (rc); } /* * Initialize various connection parameters. */ void init_conn_params(struct vi_info *vi , struct offload_settings *s, struct in_conninfo *inc, struct socket *so, const struct tcp_options *tcpopt, int16_t l2t_idx, struct conn_params *cp) { struct port_info *pi = vi->pi; struct adapter *sc = pi->adapter; struct tom_tunables *tt = &sc->tt; struct inpcb *inp = sotoinpcb(so); struct tcpcb *tp = intotcpcb(inp); u_long wnd; MPASS(s->offload != 0); /* Congestion control algorithm */ if (s->cong_algo >= 0) cp->cong_algo = s->cong_algo & M_CONG_CNTRL; else if (sc->tt.cong_algorithm >= 0) cp->cong_algo = tt->cong_algorithm & M_CONG_CNTRL; else { struct cc_algo *cc = CC_ALGO(tp); if (strcasecmp(cc->name, "reno") == 0) cp->cong_algo = CONG_ALG_RENO; else if (strcasecmp(cc->name, "tahoe") == 0) cp->cong_algo = CONG_ALG_TAHOE; if (strcasecmp(cc->name, "newreno") == 0) cp->cong_algo = CONG_ALG_NEWRENO; if (strcasecmp(cc->name, "highspeed") == 0) cp->cong_algo = CONG_ALG_HIGHSPEED; else { /* * Use newreno in case the algorithm selected by the * host stack is not supported by the hardware. */ cp->cong_algo = CONG_ALG_NEWRENO; } } /* Tx traffic scheduling class. */ if (s->sched_class >= 0 && s->sched_class < sc->chip_params->nsched_cls) { cp->tc_idx = s->sched_class; } else cp->tc_idx = -1; /* Nagle's algorithm. */ if (s->nagle >= 0) cp->nagle = s->nagle > 0 ? 1 : 0; else cp->nagle = tp->t_flags & TF_NODELAY ? 0 : 1; /* TCP Keepalive. */ - if (tcp_always_keepalive || so_options_get(so) & SO_KEEPALIVE) + if (V_tcp_always_keepalive || so_options_get(so) & SO_KEEPALIVE) cp->keepalive = 1; else cp->keepalive = 0; /* Optimization that's specific to T5 @ 40G. */ if (tt->tx_align >= 0) cp->tx_align = tt->tx_align > 0 ? 1 : 0; else if (chip_id(sc) == CHELSIO_T5 && (port_top_speed(pi) > 10 || sc->params.nports > 2)) cp->tx_align = 1; else cp->tx_align = 0; /* ULP mode. */ if (can_tls_offload(sc) && (s->tls > 0 || (s->tls < 0 && is_tls_sock(so, sc)))) cp->ulp_mode = ULP_MODE_TLS; else if (s->ddp > 0 || (s->ddp < 0 && sc->tt.ddp && (so_options_get(so) & SO_NO_DDP) == 0)) cp->ulp_mode = ULP_MODE_TCPDDP; else cp->ulp_mode = ULP_MODE_NONE; /* Rx coalescing. */ if (s->rx_coalesce >= 0) cp->rx_coalesce = s->rx_coalesce > 0 ? 1 : 0; else if (cp->ulp_mode == ULP_MODE_TLS) cp->rx_coalesce = 0; else if (tt->rx_coalesce >= 0) cp->rx_coalesce = tt->rx_coalesce > 0 ? 1 : 0; else cp->rx_coalesce = 1; /* default */ /* * Index in the PMTU table. This controls the MSS that we announce in * our SYN initially, but after ESTABLISHED it controls the MSS that we * use to send data. */ cp->mtu_idx = find_best_mtu_idx(sc, inc, s); /* Tx queue for this connection. */ if (s->txq >= 0 && s->txq < vi->nofldtxq) cp->txq_idx = s->txq; else cp->txq_idx = arc4random() % vi->nofldtxq; cp->txq_idx += vi->first_ofld_txq; /* Rx queue for this connection. */ if (s->rxq >= 0 && s->rxq < vi->nofldrxq) cp->rxq_idx = s->rxq; else cp->rxq_idx = arc4random() % vi->nofldrxq; cp->rxq_idx += vi->first_ofld_rxq; if (SOLISTENING(so)) { /* Passive open */ MPASS(tcpopt != NULL); /* TCP timestamp option */ if (tcpopt->tstamp && (s->tstamp > 0 || (s->tstamp < 0 && V_tcp_do_rfc1323))) cp->tstamp = 1; else cp->tstamp = 0; /* SACK */ if (tcpopt->sack && (s->sack > 0 || (s->sack < 0 && V_tcp_do_sack))) cp->sack = 1; else cp->sack = 0; /* Receive window scaling. */ if (tcpopt->wsf > 0 && tcpopt->wsf < 15 && V_tcp_do_rfc1323) cp->wscale = select_rcv_wscale(); else cp->wscale = 0; /* ECN */ if (tcpopt->ecn && /* XXX: review. */ (s->ecn > 0 || (s->ecn < 0 && V_tcp_do_ecn))) cp->ecn = 1; else cp->ecn = 0; wnd = max(so->sol_sbrcv_hiwat, MIN_RCV_WND); cp->opt0_bufsize = min(wnd >> 10, M_RCV_BUFSIZ); if (tt->sndbuf > 0) cp->sndbuf = tt->sndbuf; else if (so->sol_sbsnd_flags & SB_AUTOSIZE && V_tcp_do_autosndbuf) cp->sndbuf = 256 * 1024; else cp->sndbuf = so->sol_sbsnd_hiwat; } else { /* Active open */ /* TCP timestamp option */ if (s->tstamp > 0 || (s->tstamp < 0 && (tp->t_flags & TF_REQ_TSTMP))) cp->tstamp = 1; else cp->tstamp = 0; /* SACK */ if (s->sack > 0 || (s->sack < 0 && (tp->t_flags & TF_SACK_PERMIT))) cp->sack = 1; else cp->sack = 0; /* Receive window scaling */ if (tp->t_flags & TF_REQ_SCALE) cp->wscale = select_rcv_wscale(); else cp->wscale = 0; /* ECN */ if (s->ecn > 0 || (s->ecn < 0 && V_tcp_do_ecn == 1)) cp->ecn = 1; else cp->ecn = 0; SOCKBUF_LOCK(&so->so_rcv); wnd = max(select_rcv_wnd(so), MIN_RCV_WND); SOCKBUF_UNLOCK(&so->so_rcv); cp->opt0_bufsize = min(wnd >> 10, M_RCV_BUFSIZ); if (tt->sndbuf > 0) cp->sndbuf = tt->sndbuf; else { SOCKBUF_LOCK(&so->so_snd); if (so->so_snd.sb_flags & SB_AUTOSIZE && V_tcp_do_autosndbuf) cp->sndbuf = 256 * 1024; else cp->sndbuf = so->so_snd.sb_hiwat; SOCKBUF_UNLOCK(&so->so_snd); } } cp->l2t_idx = l2t_idx; /* This will be initialized on ESTABLISHED. */ cp->emss = 0; } int negative_advice(int status) { return (status == CPL_ERR_RTX_NEG_ADVICE || status == CPL_ERR_PERSIST_NEG_ADVICE || status == CPL_ERR_KEEPALV_NEG_ADVICE); } static int alloc_tid_tab(struct tid_info *t, int flags) { MPASS(t->ntids > 0); MPASS(t->tid_tab == NULL); t->tid_tab = malloc(t->ntids * sizeof(*t->tid_tab), M_CXGBE, M_ZERO | flags); if (t->tid_tab == NULL) return (ENOMEM); atomic_store_rel_int(&t->tids_in_use, 0); return (0); } static void free_tid_tab(struct tid_info *t) { KASSERT(t->tids_in_use == 0, ("%s: %d tids still in use.", __func__, t->tids_in_use)); free(t->tid_tab, M_CXGBE); t->tid_tab = NULL; } static int alloc_stid_tab(struct tid_info *t, int flags) { MPASS(t->nstids > 0); MPASS(t->stid_tab == NULL); t->stid_tab = malloc(t->nstids * sizeof(*t->stid_tab), M_CXGBE, M_ZERO | flags); if (t->stid_tab == NULL) return (ENOMEM); mtx_init(&t->stid_lock, "stid lock", NULL, MTX_DEF); t->stids_in_use = 0; TAILQ_INIT(&t->stids); t->nstids_free_head = t->nstids; return (0); } static void free_stid_tab(struct tid_info *t) { KASSERT(t->stids_in_use == 0, ("%s: %d tids still in use.", __func__, t->stids_in_use)); if (mtx_initialized(&t->stid_lock)) mtx_destroy(&t->stid_lock); free(t->stid_tab, M_CXGBE); t->stid_tab = NULL; } static void free_tid_tabs(struct tid_info *t) { free_tid_tab(t); free_stid_tab(t); } static int alloc_tid_tabs(struct tid_info *t) { int rc; rc = alloc_tid_tab(t, M_NOWAIT); if (rc != 0) goto failed; rc = alloc_stid_tab(t, M_NOWAIT); if (rc != 0) goto failed; return (0); failed: free_tid_tabs(t); return (rc); } static inline void alloc_tcb_history(struct adapter *sc, struct tom_data *td) { if (sc->tids.ntids == 0 || sc->tids.ntids > 1024) return; rw_init(&td->tcb_history_lock, "TCB history"); td->tcb_history = malloc(sc->tids.ntids * sizeof(*td->tcb_history), M_CXGBE, M_ZERO | M_NOWAIT); td->dupack_threshold = G_DUPACKTHRESH(t4_read_reg(sc, A_TP_PARA_REG0)); } static inline void free_tcb_history(struct adapter *sc, struct tom_data *td) { #ifdef INVARIANTS int i; if (td->tcb_history != NULL) { for (i = 0; i < sc->tids.ntids; i++) { MPASS(td->tcb_history[i] == NULL); } } #endif free(td->tcb_history, M_CXGBE); if (rw_initialized(&td->tcb_history_lock)) rw_destroy(&td->tcb_history_lock); } static void free_tom_data(struct adapter *sc, struct tom_data *td) { ASSERT_SYNCHRONIZED_OP(sc); KASSERT(TAILQ_EMPTY(&td->toep_list), ("%s: TOE PCB list is not empty.", __func__)); KASSERT(td->lctx_count == 0, ("%s: lctx hash table is not empty.", __func__)); t4_free_ppod_region(&td->pr); if (td->listen_mask != 0) hashdestroy(td->listen_hash, M_CXGBE, td->listen_mask); if (mtx_initialized(&td->unsent_wr_lock)) mtx_destroy(&td->unsent_wr_lock); if (mtx_initialized(&td->lctx_hash_lock)) mtx_destroy(&td->lctx_hash_lock); if (mtx_initialized(&td->toep_list_lock)) mtx_destroy(&td->toep_list_lock); free_tcb_history(sc, td); free_tid_tabs(&sc->tids); free(td, M_CXGBE); } static char * prepare_pkt(int open_type, uint16_t vtag, struct inpcb *inp, int *pktlen, int *buflen) { char *pkt; struct tcphdr *th; int ipv6, len; const int maxlen = max(sizeof(struct ether_header), sizeof(struct ether_vlan_header)) + max(sizeof(struct ip), sizeof(struct ip6_hdr)) + sizeof(struct tcphdr); MPASS(open_type == OPEN_TYPE_ACTIVE || open_type == OPEN_TYPE_LISTEN); pkt = malloc(maxlen, M_CXGBE, M_ZERO | M_NOWAIT); if (pkt == NULL) return (NULL); ipv6 = inp->inp_vflag & INP_IPV6; len = 0; if (EVL_VLANOFTAG(vtag) == 0xfff) { struct ether_header *eh = (void *)pkt; if (ipv6) eh->ether_type = htons(ETHERTYPE_IPV6); else eh->ether_type = htons(ETHERTYPE_IP); len += sizeof(*eh); } else { struct ether_vlan_header *evh = (void *)pkt; evh->evl_encap_proto = htons(ETHERTYPE_VLAN); evh->evl_tag = htons(vtag); if (ipv6) evh->evl_proto = htons(ETHERTYPE_IPV6); else evh->evl_proto = htons(ETHERTYPE_IP); len += sizeof(*evh); } if (ipv6) { struct ip6_hdr *ip6 = (void *)&pkt[len]; ip6->ip6_vfc = IPV6_VERSION; ip6->ip6_plen = htons(sizeof(struct tcphdr)); ip6->ip6_nxt = IPPROTO_TCP; if (open_type == OPEN_TYPE_ACTIVE) { ip6->ip6_src = inp->in6p_laddr; ip6->ip6_dst = inp->in6p_faddr; } else if (open_type == OPEN_TYPE_LISTEN) { ip6->ip6_src = inp->in6p_laddr; ip6->ip6_dst = ip6->ip6_src; } len += sizeof(*ip6); } else { struct ip *ip = (void *)&pkt[len]; ip->ip_v = IPVERSION; ip->ip_hl = sizeof(*ip) >> 2; ip->ip_tos = inp->inp_ip_tos; ip->ip_len = htons(sizeof(struct ip) + sizeof(struct tcphdr)); ip->ip_ttl = inp->inp_ip_ttl; ip->ip_p = IPPROTO_TCP; if (open_type == OPEN_TYPE_ACTIVE) { ip->ip_src = inp->inp_laddr; ip->ip_dst = inp->inp_faddr; } else if (open_type == OPEN_TYPE_LISTEN) { ip->ip_src = inp->inp_laddr; ip->ip_dst = ip->ip_src; } len += sizeof(*ip); } th = (void *)&pkt[len]; if (open_type == OPEN_TYPE_ACTIVE) { th->th_sport = inp->inp_lport; /* network byte order already */ th->th_dport = inp->inp_fport; /* ditto */ } else if (open_type == OPEN_TYPE_LISTEN) { th->th_sport = inp->inp_lport; /* network byte order already */ th->th_dport = th->th_sport; } len += sizeof(th); *pktlen = *buflen = len; return (pkt); } const struct offload_settings * lookup_offload_policy(struct adapter *sc, int open_type, struct mbuf *m, uint16_t vtag, struct inpcb *inp) { const struct t4_offload_policy *op; char *pkt; struct offload_rule *r; int i, matched, pktlen, buflen; static const struct offload_settings allow_offloading_settings = { .offload = 1, .rx_coalesce = -1, .cong_algo = -1, .sched_class = -1, .tstamp = -1, .sack = -1, .nagle = -1, .ecn = -1, .ddp = -1, .tls = -1, .txq = -1, .rxq = -1, .mss = -1, }; static const struct offload_settings disallow_offloading_settings = { .offload = 0, /* rest is irrelevant when offload is off. */ }; rw_assert(&sc->policy_lock, RA_LOCKED); /* * If there's no Connection Offloading Policy attached to the device * then we need to return a default static policy. If * "cop_managed_offloading" is true, then we need to disallow * offloading until a COP is attached to the device. Otherwise we * allow offloading ... */ op = sc->policy; if (op == NULL) { if (sc->tt.cop_managed_offloading) return (&disallow_offloading_settings); else return (&allow_offloading_settings); } switch (open_type) { case OPEN_TYPE_ACTIVE: case OPEN_TYPE_LISTEN: pkt = prepare_pkt(open_type, vtag, inp, &pktlen, &buflen); break; case OPEN_TYPE_PASSIVE: MPASS(m != NULL); pkt = mtod(m, char *); MPASS(*pkt == CPL_PASS_ACCEPT_REQ); pkt += sizeof(struct cpl_pass_accept_req); pktlen = m->m_pkthdr.len - sizeof(struct cpl_pass_accept_req); buflen = m->m_len - sizeof(struct cpl_pass_accept_req); break; default: MPASS(0); return (&disallow_offloading_settings); } if (pkt == NULL || pktlen == 0 || buflen == 0) return (&disallow_offloading_settings); matched = 0; r = &op->rule[0]; for (i = 0; i < op->nrules; i++, r++) { if (r->open_type != open_type && r->open_type != OPEN_TYPE_DONTCARE) { continue; } matched = bpf_filter(r->bpf_prog.bf_insns, pkt, pktlen, buflen); if (matched) break; } if (open_type == OPEN_TYPE_ACTIVE || open_type == OPEN_TYPE_LISTEN) free(pkt, M_CXGBE); return (matched ? &r->settings : &disallow_offloading_settings); } static void reclaim_wr_resources(void *arg, int count) { struct tom_data *td = arg; STAILQ_HEAD(, wrqe) twr_list = STAILQ_HEAD_INITIALIZER(twr_list); struct cpl_act_open_req *cpl; u_int opcode, atid, tid; struct wrqe *wr; struct adapter *sc = td_adapter(td); mtx_lock(&td->unsent_wr_lock); STAILQ_SWAP(&td->unsent_wr_list, &twr_list, wrqe); mtx_unlock(&td->unsent_wr_lock); while ((wr = STAILQ_FIRST(&twr_list)) != NULL) { STAILQ_REMOVE_HEAD(&twr_list, link); cpl = wrtod(wr); opcode = GET_OPCODE(cpl); switch (opcode) { case CPL_ACT_OPEN_REQ: case CPL_ACT_OPEN_REQ6: atid = G_TID_TID(be32toh(OPCODE_TID(cpl))); CTR2(KTR_CXGBE, "%s: atid %u ", __func__, atid); act_open_failure_cleanup(sc, atid, EHOSTUNREACH); free(wr, M_CXGBE); break; case CPL_PASS_ACCEPT_RPL: tid = GET_TID(cpl); CTR2(KTR_CXGBE, "%s: tid %u ", __func__, tid); synack_failure_cleanup(sc, tid); free(wr, M_CXGBE); break; default: log(LOG_ERR, "%s: leaked work request %p, wr_len %d, " "opcode %x\n", __func__, wr, wr->wr_len, opcode); /* WR not freed here; go look at it with a debugger. */ } } } /* * Ground control to Major TOM * Commencing countdown, engines on */ static int t4_tom_activate(struct adapter *sc) { struct tom_data *td; struct toedev *tod; struct vi_info *vi; int i, rc, v; ASSERT_SYNCHRONIZED_OP(sc); /* per-adapter softc for TOM */ td = malloc(sizeof(*td), M_CXGBE, M_ZERO | M_NOWAIT); if (td == NULL) return (ENOMEM); /* List of TOE PCBs and associated lock */ mtx_init(&td->toep_list_lock, "PCB list lock", NULL, MTX_DEF); TAILQ_INIT(&td->toep_list); /* Listen context */ mtx_init(&td->lctx_hash_lock, "lctx hash lock", NULL, MTX_DEF); td->listen_hash = hashinit_flags(LISTEN_HASH_SIZE, M_CXGBE, &td->listen_mask, HASH_NOWAIT); /* List of WRs for which L2 resolution failed */ mtx_init(&td->unsent_wr_lock, "Unsent WR list lock", NULL, MTX_DEF); STAILQ_INIT(&td->unsent_wr_list); TASK_INIT(&td->reclaim_wr_resources, 0, reclaim_wr_resources, td); /* TID tables */ rc = alloc_tid_tabs(&sc->tids); if (rc != 0) goto done; rc = t4_init_ppod_region(&td->pr, &sc->vres.ddp, t4_read_reg(sc, A_ULP_RX_TDDP_PSZ), "TDDP page pods"); if (rc != 0) goto done; t4_set_reg_field(sc, A_ULP_RX_TDDP_TAGMASK, V_TDDPTAGMASK(M_TDDPTAGMASK), td->pr.pr_tag_mask); alloc_tcb_history(sc, td); /* toedev ops */ tod = &td->tod; init_toedev(tod); tod->tod_softc = sc; tod->tod_connect = t4_connect; tod->tod_listen_start = t4_listen_start; tod->tod_listen_stop = t4_listen_stop; tod->tod_rcvd = t4_rcvd; tod->tod_output = t4_tod_output; tod->tod_send_rst = t4_send_rst; tod->tod_send_fin = t4_send_fin; tod->tod_pcb_detach = t4_pcb_detach; tod->tod_l2_update = t4_l2_update; tod->tod_syncache_added = t4_syncache_added; tod->tod_syncache_removed = t4_syncache_removed; tod->tod_syncache_respond = t4_syncache_respond; tod->tod_offload_socket = t4_offload_socket; tod->tod_ctloutput = t4_ctloutput; tod->tod_tcp_info = t4_tcp_info; #ifdef KERN_TLS tod->tod_alloc_tls_session = t4_alloc_tls_session; #endif for_each_port(sc, i) { for_each_vi(sc->port[i], v, vi) { TOEDEV(vi->ifp) = &td->tod; } } sc->tom_softc = td; register_toedev(sc->tom_softc); done: if (rc != 0) free_tom_data(sc, td); return (rc); } static int t4_tom_deactivate(struct adapter *sc) { int rc = 0; struct tom_data *td = sc->tom_softc; ASSERT_SYNCHRONIZED_OP(sc); if (td == NULL) return (0); /* XXX. KASSERT? */ if (sc->offload_map != 0) return (EBUSY); /* at least one port has IFCAP_TOE enabled */ if (uld_active(sc, ULD_IWARP) || uld_active(sc, ULD_ISCSI)) return (EBUSY); /* both iWARP and iSCSI rely on the TOE. */ mtx_lock(&td->toep_list_lock); if (!TAILQ_EMPTY(&td->toep_list)) rc = EBUSY; mtx_unlock(&td->toep_list_lock); mtx_lock(&td->lctx_hash_lock); if (td->lctx_count > 0) rc = EBUSY; mtx_unlock(&td->lctx_hash_lock); taskqueue_drain(taskqueue_thread, &td->reclaim_wr_resources); mtx_lock(&td->unsent_wr_lock); if (!STAILQ_EMPTY(&td->unsent_wr_list)) rc = EBUSY; mtx_unlock(&td->unsent_wr_lock); if (rc == 0) { unregister_toedev(sc->tom_softc); free_tom_data(sc, td); sc->tom_softc = NULL; } return (rc); } static int t4_aio_queue_tom(struct socket *so, struct kaiocb *job) { struct tcpcb *tp = so_sototcpcb(so); struct toepcb *toep = tp->t_toe; int error; if (ulp_mode(toep) == ULP_MODE_TCPDDP) { error = t4_aio_queue_ddp(so, job); if (error != EOPNOTSUPP) return (error); } return (t4_aio_queue_aiotx(so, job)); } static int t4_ctloutput_tom(struct socket *so, struct sockopt *sopt) { if (sopt->sopt_level != IPPROTO_TCP) return (tcp_ctloutput(so, sopt)); switch (sopt->sopt_name) { case TCP_TLSOM_SET_TLS_CONTEXT: case TCP_TLSOM_GET_TLS_TOM: case TCP_TLSOM_CLR_TLS_TOM: case TCP_TLSOM_CLR_QUIES: return (t4_ctloutput_tls(so, sopt)); default: return (tcp_ctloutput(so, sopt)); } } static int t4_tom_mod_load(void) { struct protosw *tcp_protosw, *tcp6_protosw; /* CPL handlers */ t4_register_cpl_handler(CPL_GET_TCB_RPL, do_get_tcb_rpl); t4_register_shared_cpl_handler(CPL_L2T_WRITE_RPL, do_l2t_write_rpl2, CPL_COOKIE_TOM); t4_init_connect_cpl_handlers(); t4_init_listen_cpl_handlers(); t4_init_cpl_io_handlers(); t4_ddp_mod_load(); t4_tls_mod_load(); tcp_protosw = pffindproto(PF_INET, IPPROTO_TCP, SOCK_STREAM); if (tcp_protosw == NULL) return (ENOPROTOOPT); bcopy(tcp_protosw, &toe_protosw, sizeof(toe_protosw)); bcopy(tcp_protosw->pr_usrreqs, &toe_usrreqs, sizeof(toe_usrreqs)); toe_usrreqs.pru_aio_queue = t4_aio_queue_tom; toe_protosw.pr_ctloutput = t4_ctloutput_tom; toe_protosw.pr_usrreqs = &toe_usrreqs; tcp6_protosw = pffindproto(PF_INET6, IPPROTO_TCP, SOCK_STREAM); if (tcp6_protosw == NULL) return (ENOPROTOOPT); bcopy(tcp6_protosw, &toe6_protosw, sizeof(toe6_protosw)); bcopy(tcp6_protosw->pr_usrreqs, &toe6_usrreqs, sizeof(toe6_usrreqs)); toe6_usrreqs.pru_aio_queue = t4_aio_queue_tom; toe6_protosw.pr_ctloutput = t4_ctloutput_tom; toe6_protosw.pr_usrreqs = &toe6_usrreqs; return (t4_register_uld(&tom_uld_info)); } static void tom_uninit(struct adapter *sc, void *arg __unused) { if (begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4tomun")) return; /* Try to free resources (works only if no port has IFCAP_TOE) */ if (uld_active(sc, ULD_TOM)) t4_deactivate_uld(sc, ULD_TOM); end_synchronized_op(sc, 0); } static int t4_tom_mod_unload(void) { t4_iterate(tom_uninit, NULL); if (t4_unregister_uld(&tom_uld_info) == EBUSY) return (EBUSY); t4_tls_mod_unload(); t4_ddp_mod_unload(); t4_uninit_connect_cpl_handlers(); t4_uninit_listen_cpl_handlers(); t4_uninit_cpl_io_handlers(); t4_register_shared_cpl_handler(CPL_L2T_WRITE_RPL, NULL, CPL_COOKIE_TOM); return (0); } #endif /* TCP_OFFLOAD */ static int t4_tom_modevent(module_t mod, int cmd, void *arg) { int rc = 0; #ifdef TCP_OFFLOAD switch (cmd) { case MOD_LOAD: rc = t4_tom_mod_load(); break; case MOD_UNLOAD: rc = t4_tom_mod_unload(); break; default: rc = EINVAL; } #else printf("t4_tom: compiled without TCP_OFFLOAD support.\n"); rc = EOPNOTSUPP; #endif return (rc); } static moduledata_t t4_tom_moddata= { "t4_tom", t4_tom_modevent, 0 }; MODULE_VERSION(t4_tom, 1); MODULE_DEPEND(t4_tom, toecore, 1, 1, 1); MODULE_DEPEND(t4_tom, t4nex, 1, 1, 1); DECLARE_MODULE(t4_tom, t4_tom_moddata, SI_SUB_EXEC, SI_ORDER_ANY); Index: head/sys/netinet/tcp_input.c =================================================================== --- head/sys/netinet/tcp_input.c (revision 356526) +++ head/sys/netinet/tcp_input.c (revision 356527) @@ -1,3835 +1,3835 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1994, 1995 * The Regents of the University of California. All rights reserved. * Copyright (c) 2007-2008,2010 * Swinburne University of Technology, Melbourne, Australia. * Copyright (c) 2009-2010 Lawrence Stewart * Copyright (c) 2010 The FreeBSD Foundation * Copyright (c) 2010-2011 Juniper Networks, Inc. * All rights reserved. * * Portions of this software were developed at the Centre for Advanced Internet * Architectures, Swinburne University of Technology, by Lawrence Stewart, * James Healy and David Hayes, made possible in part by a grant from the Cisco * University Research Program Fund at Community Foundation Silicon Valley. * * Portions of this software were developed at the Centre for Advanced * Internet Architectures, Swinburne University of Technology, Melbourne, * Australia by David Hayes under sponsorship from the FreeBSD Foundation. * * 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. * * @(#)tcp_input.c 8.12 (Berkeley) 5/24/95 */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include "opt_ipsec.h" #include "opt_tcpdebug.h" #include #include #include #ifdef TCP_HHOOK #include #endif #include #include #include /* for proc0 declaration */ #include #include #include #include #include #include #include #include #include #include #include /* before tcp_seq.h, for tcp_random18() */ #include #include #include #include #include #define TCPSTATES /* for logging */ #include #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 #include #include #include #include #include #ifdef TCPPCAP #include #endif #include #ifdef TCPDEBUG #include #endif /* TCPDEBUG */ #ifdef TCP_OFFLOAD #include #endif #include #include #include const int tcprexmtthresh = 3; -int tcp_log_in_vain = 0; -SYSCTL_INT(_net_inet_tcp, OID_AUTO, log_in_vain, CTLFLAG_RW, - &tcp_log_in_vain, 0, +VNET_DEFINE(int, tcp_log_in_vain) = 0; +SYSCTL_INT(_net_inet_tcp, OID_AUTO, log_in_vain, CTLFLAG_VNET | CTLFLAG_RW, + &VNET_NAME(tcp_log_in_vain), 0, "Log all incoming TCP segments to closed ports"); VNET_DEFINE(int, blackhole) = 0; #define V_blackhole VNET(blackhole) SYSCTL_INT(_net_inet_tcp, OID_AUTO, blackhole, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(blackhole), 0, "Do not send RST on segments to closed ports"); VNET_DEFINE(int, tcp_delack_enabled) = 1; SYSCTL_INT(_net_inet_tcp, OID_AUTO, delayed_ack, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_delack_enabled), 0, "Delay ACK to try and piggyback it onto a data packet"); VNET_DEFINE(int, drop_synfin) = 0; SYSCTL_INT(_net_inet_tcp, OID_AUTO, drop_synfin, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(drop_synfin), 0, "Drop TCP packets with SYN+FIN set"); VNET_DEFINE(int, tcp_do_newcwv) = 0; SYSCTL_INT(_net_inet_tcp, OID_AUTO, newcwv, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_do_newcwv), 0, "Enable New Congestion Window Validation per RFC7661"); VNET_DEFINE(int, tcp_do_rfc6675_pipe) = 0; SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc6675_pipe, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_do_rfc6675_pipe), 0, "Use calculated pipe/in-flight bytes per RFC 6675"); VNET_DEFINE(int, tcp_do_rfc3042) = 1; SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3042, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_do_rfc3042), 0, "Enable RFC 3042 (Limited Transmit)"); VNET_DEFINE(int, tcp_do_rfc3390) = 1; SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3390, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_do_rfc3390), 0, "Enable RFC 3390 (Increasing TCP's Initial Congestion Window)"); VNET_DEFINE(int, tcp_initcwnd_segments) = 10; SYSCTL_INT(_net_inet_tcp, OID_AUTO, initcwnd_segments, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_initcwnd_segments), 0, "Slow-start flight size (initial congestion window) in number of segments"); VNET_DEFINE(int, tcp_do_rfc3465) = 1; SYSCTL_INT(_net_inet_tcp, OID_AUTO, rfc3465, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_do_rfc3465), 0, "Enable RFC 3465 (Appropriate Byte Counting)"); VNET_DEFINE(int, tcp_abc_l_var) = 2; SYSCTL_INT(_net_inet_tcp, OID_AUTO, abc_l_var, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_abc_l_var), 2, "Cap the max cwnd increment during slow-start to this number of segments"); static SYSCTL_NODE(_net_inet_tcp, OID_AUTO, ecn, CTLFLAG_RW, 0, "TCP ECN"); VNET_DEFINE(int, tcp_do_ecn) = 2; SYSCTL_INT(_net_inet_tcp_ecn, OID_AUTO, enable, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_do_ecn), 0, "TCP ECN support"); VNET_DEFINE(int, tcp_ecn_maxretries) = 1; SYSCTL_INT(_net_inet_tcp_ecn, OID_AUTO, maxretries, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_ecn_maxretries), 0, "Max retries before giving up on ECN"); VNET_DEFINE(int, tcp_insecure_syn) = 0; SYSCTL_INT(_net_inet_tcp, OID_AUTO, insecure_syn, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_insecure_syn), 0, "Follow RFC793 instead of RFC5961 criteria for accepting SYN packets"); VNET_DEFINE(int, tcp_insecure_rst) = 0; SYSCTL_INT(_net_inet_tcp, OID_AUTO, insecure_rst, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_insecure_rst), 0, "Follow RFC793 instead of RFC5961 criteria for accepting RST packets"); VNET_DEFINE(int, tcp_recvspace) = 1024*64; #define V_tcp_recvspace VNET(tcp_recvspace) SYSCTL_INT(_net_inet_tcp, TCPCTL_RECVSPACE, recvspace, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_recvspace), 0, "Initial receive socket buffer size"); VNET_DEFINE(int, tcp_do_autorcvbuf) = 1; SYSCTL_INT(_net_inet_tcp, OID_AUTO, recvbuf_auto, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_do_autorcvbuf), 0, "Enable automatic receive buffer sizing"); VNET_DEFINE(int, tcp_autorcvbuf_max) = 2*1024*1024; SYSCTL_INT(_net_inet_tcp, OID_AUTO, recvbuf_max, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(tcp_autorcvbuf_max), 0, "Max size of automatic receive buffer"); VNET_DEFINE(struct inpcbhead, tcb); #define tcb6 tcb /* for KAME src sync over BSD*'s */ VNET_DEFINE(struct inpcbinfo, tcbinfo); /* * TCP statistics are stored in an array of counter(9)s, which size matches * size of struct tcpstat. TCP running connection count is a regular array. */ VNET_PCPUSTAT_DEFINE(struct tcpstat, tcpstat); SYSCTL_VNET_PCPUSTAT(_net_inet_tcp, TCPCTL_STATS, stats, struct tcpstat, tcpstat, "TCP statistics (struct tcpstat, netinet/tcp_var.h)"); VNET_DEFINE(counter_u64_t, tcps_states[TCP_NSTATES]); SYSCTL_COUNTER_U64_ARRAY(_net_inet_tcp, TCPCTL_STATES, states, CTLFLAG_RD | CTLFLAG_VNET, &VNET_NAME(tcps_states)[0], TCP_NSTATES, "TCP connection counts by TCP state"); static void tcp_vnet_init(const void *unused) { COUNTER_ARRAY_ALLOC(V_tcps_states, TCP_NSTATES, M_WAITOK); VNET_PCPUSTAT_ALLOC(tcpstat, M_WAITOK); } VNET_SYSINIT(tcp_vnet_init, SI_SUB_PROTO_IFATTACHDOMAIN, SI_ORDER_ANY, tcp_vnet_init, NULL); #ifdef VIMAGE static void tcp_vnet_uninit(const void *unused) { COUNTER_ARRAY_FREE(V_tcps_states, TCP_NSTATES); VNET_PCPUSTAT_FREE(tcpstat); } VNET_SYSUNINIT(tcp_vnet_uninit, SI_SUB_PROTO_IFATTACHDOMAIN, SI_ORDER_ANY, tcp_vnet_uninit, NULL); #endif /* VIMAGE */ /* * Kernel module interface for updating tcpstat. The argument is an index * into tcpstat treated as an array. */ void kmod_tcpstat_inc(int statnum) { counter_u64_add(VNET(tcpstat)[statnum], 1); } #ifdef TCP_HHOOK /* * Wrapper for the TCP established input helper hook. */ void hhook_run_tcp_est_in(struct tcpcb *tp, struct tcphdr *th, struct tcpopt *to) { struct tcp_hhook_data hhook_data; if (V_tcp_hhh[HHOOK_TCP_EST_IN]->hhh_nhooks > 0) { hhook_data.tp = tp; hhook_data.th = th; hhook_data.to = to; hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_IN], &hhook_data, tp->osd); } } #endif /* * CC wrapper hook functions */ void cc_ack_received(struct tcpcb *tp, struct tcphdr *th, uint16_t nsegs, uint16_t type) { #ifdef STATS int32_t gput; #endif INP_WLOCK_ASSERT(tp->t_inpcb); tp->ccv->nsegs = nsegs; tp->ccv->bytes_this_ack = BYTES_THIS_ACK(tp, th); if ((!V_tcp_do_newcwv && (tp->snd_cwnd <= tp->snd_wnd)) || (V_tcp_do_newcwv && (tp->snd_cwnd <= tp->snd_wnd) && (tp->snd_cwnd < (tcp_compute_pipe(tp) * 2)))) tp->ccv->flags |= CCF_CWND_LIMITED; else tp->ccv->flags &= ~CCF_CWND_LIMITED; if (type == CC_ACK) { #ifdef STATS stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF, ((int32_t)tp->snd_cwnd) - tp->snd_wnd); if (!IN_RECOVERY(tp->t_flags)) stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_ACKLEN, tp->ccv->bytes_this_ack / (tcp_maxseg(tp) * nsegs)); if ((tp->t_flags & TF_GPUTINPROG) && SEQ_GEQ(th->th_ack, tp->gput_ack)) { /* * Compute goodput in bits per millisecond. */ gput = (((int64_t)(th->th_ack - tp->gput_seq)) << 3) / max(1, tcp_ts_getticks() - tp->gput_ts); 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); tp->t_flags &= ~TF_GPUTINPROG; tp->t_stats_gput_prev = gput; } #endif /* STATS */ if (tp->snd_cwnd > tp->snd_ssthresh) { tp->t_bytes_acked += min(tp->ccv->bytes_this_ack, nsegs * V_tcp_abc_l_var * tcp_maxseg(tp)); if (tp->t_bytes_acked >= tp->snd_cwnd) { tp->t_bytes_acked -= tp->snd_cwnd; tp->ccv->flags |= CCF_ABC_SENTAWND; } } else { tp->ccv->flags &= ~CCF_ABC_SENTAWND; tp->t_bytes_acked = 0; } } if (CC_ALGO(tp)->ack_received != NULL) { /* XXXLAS: Find a way to live without this */ tp->ccv->curack = th->th_ack; CC_ALGO(tp)->ack_received(tp->ccv, type); } #ifdef STATS stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, tp->snd_cwnd); #endif } void cc_conn_init(struct tcpcb *tp) { struct hc_metrics_lite metrics; struct inpcb *inp = tp->t_inpcb; u_int maxseg; int rtt; INP_WLOCK_ASSERT(tp->t_inpcb); tcp_hc_get(&inp->inp_inc, &metrics); maxseg = tcp_maxseg(tp); if (tp->t_srtt == 0 && (rtt = metrics.rmx_rtt)) { tp->t_srtt = rtt; tp->t_rttbest = tp->t_srtt + TCP_RTT_SCALE; TCPSTAT_INC(tcps_usedrtt); if (metrics.rmx_rttvar) { tp->t_rttvar = metrics.rmx_rttvar; TCPSTAT_INC(tcps_usedrttvar); } else { /* default variation is +- 1 rtt */ tp->t_rttvar = tp->t_srtt * TCP_RTTVAR_SCALE / TCP_RTT_SCALE; } TCPT_RANGESET(tp->t_rxtcur, ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1, tp->t_rttmin, TCPTV_REXMTMAX); } if (metrics.rmx_ssthresh) { /* * There's some sort of gateway or interface * buffer limit on the path. Use this to set * the slow start threshold, but set the * threshold to no less than 2*mss. */ tp->snd_ssthresh = max(2 * maxseg, metrics.rmx_ssthresh); TCPSTAT_INC(tcps_usedssthresh); } /* * Set the initial slow-start flight size. * * If a SYN or SYN/ACK was lost and retransmitted, we have to * reduce the initial CWND to one segment as congestion is likely * requiring us to be cautious. */ if (tp->snd_cwnd == 1) tp->snd_cwnd = maxseg; /* SYN(-ACK) lost */ else tp->snd_cwnd = tcp_compute_initwnd(maxseg); if (CC_ALGO(tp)->conn_init != NULL) CC_ALGO(tp)->conn_init(tp->ccv); } void inline cc_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type) { u_int maxseg; INP_WLOCK_ASSERT(tp->t_inpcb); #ifdef STATS stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_CSIG, type); #endif switch(type) { case CC_NDUPACK: if (!IN_FASTRECOVERY(tp->t_flags)) { 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)) { TCPSTAT_INC(tcps_ecn_rcwnd); tp->snd_recover = tp->snd_max; if (tp->t_flags2 & TF2_ECN_PERMIT) tp->t_flags2 |= TF2_ECN_SND_CWR; } break; case CC_RTO: maxseg = tcp_maxseg(tp); tp->t_dupacks = 0; tp->t_bytes_acked = 0; EXIT_RECOVERY(tp->t_flags); tp->snd_ssthresh = max(2, min(tp->snd_wnd, tp->snd_cwnd) / 2 / maxseg) * maxseg; tp->snd_cwnd = maxseg; break; case CC_RTO_ERR: 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); if (tp->t_flags & TF_WASCRECOVERY) ENTER_CONGRECOVERY(tp->t_flags); tp->snd_nxt = tp->snd_max; tp->t_flags &= ~TF_PREVVALID; tp->t_badrxtwin = 0; break; } if (CC_ALGO(tp)->cong_signal != NULL) { if (th != NULL) tp->ccv->curack = th->th_ack; CC_ALGO(tp)->cong_signal(tp->ccv, type); } } void inline cc_post_recovery(struct tcpcb *tp, struct tcphdr *th) { INP_WLOCK_ASSERT(tp->t_inpcb); /* XXXLAS: KASSERT that we're in recovery? */ if (CC_ALGO(tp)->post_recovery != NULL) { tp->ccv->curack = th->th_ack; CC_ALGO(tp)->post_recovery(tp->ccv); } /* XXXLAS: EXIT_RECOVERY ? */ tp->t_bytes_acked = 0; } /* * 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. */ #define DELAY_ACK(tp, tlen) \ ((!tcp_timer_active(tp, TT_DELACK) && \ (tp->t_flags & TF_RXWIN0SENT) == 0) && \ (tlen <= tp->t_maxseg) && \ (V_tcp_delack_enabled || (tp->t_flags & TF_NEEDSYN))) void inline cc_ecnpkt_handler(struct tcpcb *tp, struct tcphdr *th, uint8_t iptos) { INP_WLOCK_ASSERT(tp->t_inpcb); if (CC_ALGO(tp)->ecnpkt_handler != NULL) { switch (iptos & IPTOS_ECN_MASK) { case IPTOS_ECN_CE: tp->ccv->flags |= CCF_IPHDR_CE; break; case IPTOS_ECN_ECT0: /* FALLTHROUGH */ case IPTOS_ECN_ECT1: /* FALLTHROUGH */ case IPTOS_ECN_NOTECT: tp->ccv->flags &= ~CCF_IPHDR_CE; break; } if (th->th_flags & TH_CWR) tp->ccv->flags |= CCF_TCPHDR_CWR; else tp->ccv->flags &= ~CCF_TCPHDR_CWR; CC_ALGO(tp)->ecnpkt_handler(tp->ccv); if (tp->ccv->flags & CCF_ACKNOW) { tcp_timer_activate(tp, TT_DELACK, tcp_delacktime); tp->t_flags |= TF_ACKNOW; } } } /* * TCP input handling is split into multiple parts: * tcp6_input is a thin wrapper around tcp_input for the extended * ip6_protox[] call format in ip6_input * tcp_input handles primary segment validation, inpcb lookup and * SYN processing on listen sockets * tcp_do_segment processes the ACK and text of the segment for * establishing, established and closing connections */ #ifdef INET6 int tcp6_input(struct mbuf **mp, int *offp, int proto) { struct mbuf *m; struct in6_ifaddr *ia6; struct ip6_hdr *ip6; m = *mp; if (m->m_len < *offp + sizeof(struct tcphdr)) { m = m_pullup(m, *offp + sizeof(struct tcphdr)); if (m == NULL) { *mp = m; TCPSTAT_INC(tcps_rcvshort); return (IPPROTO_DONE); } } /* * draft-itojun-ipv6-tcp-to-anycast * better place to put this in? */ ip6 = mtod(m, struct ip6_hdr *); ia6 = in6ifa_ifwithaddr(&ip6->ip6_dst, 0 /* XXX */); if (ia6 && (ia6->ia6_flags & IN6_IFF_ANYCAST)) { ifa_free(&ia6->ia_ifa); icmp6_error(m, ICMP6_DST_UNREACH, ICMP6_DST_UNREACH_ADDR, (caddr_t)&ip6->ip6_dst - (caddr_t)ip6); *mp = NULL; return (IPPROTO_DONE); } if (ia6) ifa_free(&ia6->ia_ifa); *mp = m; return (tcp_input(mp, offp, proto)); } #endif /* INET6 */ int tcp_input(struct mbuf **mp, int *offp, int proto) { struct mbuf *m = *mp; struct tcphdr *th = NULL; struct ip *ip = NULL; struct inpcb *inp = NULL; struct tcpcb *tp = NULL; struct socket *so = NULL; u_char *optp = NULL; int off0; int optlen = 0; #ifdef INET int len; uint8_t ipttl; #endif int tlen = 0, off; int drop_hdrlen; int thflags; int rstreason = 0; /* For badport_bandlim accounting purposes */ uint8_t iptos; struct m_tag *fwd_tag = NULL; #ifdef INET6 struct ip6_hdr *ip6 = NULL; int isipv6; #else const void *ip6 = NULL; #endif /* INET6 */ struct tcpopt to; /* options in this segment */ char *s = NULL; /* address and port logging */ #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 NET_EPOCH_ASSERT(); #ifdef INET6 isipv6 = (mtod(m, struct ip *)->ip_v == 6) ? 1 : 0; #endif off0 = *offp; m = *mp; *mp = NULL; to.to_flags = 0; TCPSTAT_INC(tcps_rcvtotal); #ifdef INET6 if (isipv6) { ip6 = mtod(m, struct ip6_hdr *); th = (struct tcphdr *)((caddr_t)ip6 + off0); tlen = sizeof(*ip6) + ntohs(ip6->ip6_plen) - off0; if (m->m_pkthdr.csum_flags & CSUM_DATA_VALID_IPV6) { if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR) th->th_sum = m->m_pkthdr.csum_data; else th->th_sum = in6_cksum_pseudo(ip6, tlen, IPPROTO_TCP, m->m_pkthdr.csum_data); th->th_sum ^= 0xffff; } else th->th_sum = in6_cksum(m, IPPROTO_TCP, off0, tlen); if (th->th_sum) { TCPSTAT_INC(tcps_rcvbadsum); goto drop; } /* * 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 (IN6_IS_ADDR_UNSPECIFIED(&ip6->ip6_src)) { /* XXX stat */ goto drop; } iptos = (ntohl(ip6->ip6_flow) >> 20) & 0xff; } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { /* * Get IP and TCP header together in first mbuf. * Note: IP leaves IP header in first mbuf. */ if (off0 > sizeof (struct ip)) { ip_stripoptions(m); off0 = sizeof(struct ip); } if (m->m_len < sizeof (struct tcpiphdr)) { if ((m = m_pullup(m, sizeof (struct tcpiphdr))) == NULL) { TCPSTAT_INC(tcps_rcvshort); return (IPPROTO_DONE); } } ip = mtod(m, struct ip *); th = (struct tcphdr *)((caddr_t)ip + off0); tlen = ntohs(ip->ip_len) - off0; iptos = ip->ip_tos; if (m->m_pkthdr.csum_flags & CSUM_DATA_VALID) { if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR) th->th_sum = m->m_pkthdr.csum_data; else th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, htonl(m->m_pkthdr.csum_data + tlen + IPPROTO_TCP)); th->th_sum ^= 0xffff; } else { struct ipovly *ipov = (struct ipovly *)ip; /* * Checksum extended TCP header and data. */ len = off0 + tlen; ipttl = ip->ip_ttl; bzero(ipov->ih_x1, sizeof(ipov->ih_x1)); ipov->ih_len = htons(tlen); th->th_sum = in_cksum(m, len); /* Reset length for SDT probes. */ ip->ip_len = htons(len); /* Reset TOS bits */ ip->ip_tos = iptos; /* Re-initialization for later version check */ ip->ip_ttl = ipttl; ip->ip_v = IPVERSION; ip->ip_hl = off0 >> 2; } if (th->th_sum) { TCPSTAT_INC(tcps_rcvbadsum); goto drop; } } #endif /* INET */ /* * Check that TCP offset makes sense, * pull out TCP options and adjust length. XXX */ off = th->th_off << 2; if (off < sizeof (struct tcphdr) || off > tlen) { TCPSTAT_INC(tcps_rcvbadoff); goto drop; } tlen -= off; /* tlen is used instead of ti->ti_len */ if (off > sizeof (struct tcphdr)) { #ifdef INET6 if (isipv6) { if (m->m_len < off0 + off) { m = m_pullup(m, off0 + off); if (m == NULL) { TCPSTAT_INC(tcps_rcvshort); return (IPPROTO_DONE); } } ip6 = mtod(m, struct ip6_hdr *); th = (struct tcphdr *)((caddr_t)ip6 + off0); } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { if (m->m_len < sizeof(struct ip) + off) { if ((m = m_pullup(m, sizeof (struct ip) + off)) == NULL) { TCPSTAT_INC(tcps_rcvshort); return (IPPROTO_DONE); } ip = mtod(m, struct ip *); th = (struct tcphdr *)((caddr_t)ip + off0); } } #endif optlen = off - sizeof (struct tcphdr); optp = (u_char *)(th + 1); } thflags = th->th_flags; /* * Convert TCP protocol specific fields to host format. */ tcp_fields_to_host(th); /* * Delay dropping TCP, IP headers, IPv6 ext headers, and TCP options. */ drop_hdrlen = off0 + off; /* * Grab info from PACKET_TAG_IPFORWARD tag prepended to the chain. */ if ( #ifdef INET6 (isipv6 && (m->m_flags & M_IP6_NEXTHOP)) #ifdef INET || (!isipv6 && (m->m_flags & M_IP_NEXTHOP)) #endif #endif #if defined(INET) && !defined(INET6) (m->m_flags & M_IP_NEXTHOP) #endif ) fwd_tag = m_tag_find(m, PACKET_TAG_IPFORWARD, NULL); findpcb: #ifdef INET6 if (isipv6 && fwd_tag != NULL) { struct sockaddr_in6 *next_hop6; next_hop6 = (struct sockaddr_in6 *)(fwd_tag + 1); /* * Transparently forwarded. Pretend to be the destination. * Already got one like this? */ inp = in6_pcblookup_mbuf(&V_tcbinfo, &ip6->ip6_src, th->th_sport, &ip6->ip6_dst, th->th_dport, INPLOOKUP_WLOCKPCB, m->m_pkthdr.rcvif, m); if (!inp) { /* * It's new. Try to find the ambushing socket. * Because we've rewritten the destination address, * any hardware-generated hash is ignored. */ inp = in6_pcblookup(&V_tcbinfo, &ip6->ip6_src, th->th_sport, &next_hop6->sin6_addr, next_hop6->sin6_port ? ntohs(next_hop6->sin6_port) : th->th_dport, INPLOOKUP_WILDCARD | INPLOOKUP_WLOCKPCB, m->m_pkthdr.rcvif); } } else if (isipv6) { inp = in6_pcblookup_mbuf(&V_tcbinfo, &ip6->ip6_src, th->th_sport, &ip6->ip6_dst, th->th_dport, INPLOOKUP_WILDCARD | INPLOOKUP_WLOCKPCB, m->m_pkthdr.rcvif, m); } #endif /* INET6 */ #if defined(INET6) && defined(INET) else #endif #ifdef INET if (fwd_tag != NULL) { struct sockaddr_in *next_hop; next_hop = (struct sockaddr_in *)(fwd_tag+1); /* * Transparently forwarded. Pretend to be the destination. * already got one like this? */ inp = in_pcblookup_mbuf(&V_tcbinfo, ip->ip_src, th->th_sport, ip->ip_dst, th->th_dport, INPLOOKUP_WLOCKPCB, m->m_pkthdr.rcvif, m); if (!inp) { /* * It's new. Try to find the ambushing socket. * Because we've rewritten the destination address, * any hardware-generated hash is ignored. */ inp = in_pcblookup(&V_tcbinfo, ip->ip_src, th->th_sport, next_hop->sin_addr, next_hop->sin_port ? ntohs(next_hop->sin_port) : th->th_dport, INPLOOKUP_WILDCARD | INPLOOKUP_WLOCKPCB, m->m_pkthdr.rcvif); } } else inp = in_pcblookup_mbuf(&V_tcbinfo, ip->ip_src, th->th_sport, ip->ip_dst, th->th_dport, INPLOOKUP_WILDCARD | INPLOOKUP_WLOCKPCB, m->m_pkthdr.rcvif, m); #endif /* INET */ /* * If the INPCB does not exist then all data in the incoming * segment is discarded and an appropriate RST is sent back. * XXX MRT Send RST using which routing table? */ if (inp == NULL) { /* * Log communication attempts to ports that are not * in use. */ - if ((tcp_log_in_vain == 1 && (thflags & TH_SYN)) || - tcp_log_in_vain == 2) { + if ((V_tcp_log_in_vain == 1 && (thflags & TH_SYN)) || + V_tcp_log_in_vain == 2) { if ((s = tcp_log_vain(NULL, th, (void *)ip, ip6))) log(LOG_INFO, "%s; %s: Connection attempt " "to closed port\n", s, __func__); } /* * When blackholing do not respond with a RST but * completely ignore the segment and drop it. */ if ((V_blackhole == 1 && (thflags & TH_SYN)) || V_blackhole == 2) goto dropunlock; rstreason = BANDLIM_RST_CLOSEDPORT; goto dropwithreset; } INP_WLOCK_ASSERT(inp); /* * While waiting for inp lock during the lookup, another thread * can have dropped the inpcb, in which case we need to loop back * and try to find a new inpcb to deliver to. */ if (inp->inp_flags & INP_DROPPED) { INP_WUNLOCK(inp); inp = NULL; goto findpcb; } if ((inp->inp_flowtype == M_HASHTYPE_NONE) && (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) && ((inp->inp_socket == NULL) || (inp->inp_socket->so_options & SO_ACCEPTCONN) == 0)) { inp->inp_flowid = m->m_pkthdr.flowid; inp->inp_flowtype = M_HASHTYPE_GET(m); } #if defined(IPSEC) || defined(IPSEC_SUPPORT) #ifdef INET6 if (isipv6 && IPSEC_ENABLED(ipv6) && IPSEC_CHECK_POLICY(ipv6, m, inp) != 0) { goto dropunlock; } #ifdef INET else #endif #endif /* INET6 */ #ifdef INET if (IPSEC_ENABLED(ipv4) && IPSEC_CHECK_POLICY(ipv4, m, inp) != 0) { goto dropunlock; } #endif /* INET */ #endif /* IPSEC */ /* * Check the minimum TTL for socket. */ if (inp->inp_ip_minttl != 0) { #ifdef INET6 if (isipv6) { if (inp->inp_ip_minttl > ip6->ip6_hlim) goto dropunlock; } else #endif if (inp->inp_ip_minttl > ip->ip_ttl) goto dropunlock; } /* * A previous connection in TIMEWAIT state is supposed to catch stray * or duplicate segments arriving late. If this segment was a * legitimate new connection attempt, the old INPCB gets removed and * we can try again to find a listening socket. * * At this point, due to earlier optimism, we may hold only an inpcb * lock, and not the inpcbinfo write lock. If so, we need to try to * acquire it, or if that fails, acquire a reference on the inpcb, * drop all locks, acquire a global write lock, and then re-acquire * the inpcb lock. We may at that point discover that another thread * has tried to free the inpcb, in which case we need to loop back * and try to find a new inpcb to deliver to. * * XXXRW: It may be time to rethink timewait locking. */ if (inp->inp_flags & INP_TIMEWAIT) { if (thflags & TH_SYN) tcp_dooptions(&to, optp, optlen, TO_SYN); /* * NB: tcp_twcheck unlocks the INP and frees the mbuf. */ if (tcp_twcheck(inp, &to, th, m, tlen)) goto findpcb; return (IPPROTO_DONE); } /* * The TCPCB may no longer exist if the connection is winding * down or it is in the CLOSED state. Either way we drop the * segment and send an appropriate response. */ tp = intotcpcb(inp); if (tp == NULL || tp->t_state == TCPS_CLOSED) { rstreason = BANDLIM_RST_CLOSEDPORT; goto dropwithreset; } #ifdef TCP_OFFLOAD if (tp->t_flags & TF_TOE) { tcp_offload_input(tp, m); m = NULL; /* consumed by the TOE driver */ goto dropunlock; } #endif #ifdef MAC INP_WLOCK_ASSERT(inp); if (mac_inpcb_check_deliver(inp, m)) goto dropunlock; #endif so = inp->inp_socket; KASSERT(so != NULL, ("%s: so == NULL", __func__)); #ifdef TCPDEBUG if (so->so_options & SO_DEBUG) { ostate = tp->t_state; #ifdef INET6 if (isipv6) { bcopy((char *)ip6, (char *)tcp_saveipgen, sizeof(*ip6)); } else #endif bcopy((char *)ip, (char *)tcp_saveipgen, sizeof(*ip)); tcp_savetcp = *th; } #endif /* TCPDEBUG */ /* * When the socket is accepting connections (the INPCB is in LISTEN * state) we look into the SYN cache if this is a new connection * attempt or the completion of a previous one. */ KASSERT(tp->t_state == TCPS_LISTEN || !(so->so_options & SO_ACCEPTCONN), ("%s: so accepting but tp %p not listening", __func__, tp)); if (tp->t_state == TCPS_LISTEN && (so->so_options & SO_ACCEPTCONN)) { struct in_conninfo inc; bzero(&inc, sizeof(inc)); #ifdef INET6 if (isipv6) { inc.inc_flags |= INC_ISIPV6; if (inp->inp_inc.inc_flags & INC_IPV6MINMTU) inc.inc_flags |= INC_IPV6MINMTU; inc.inc6_faddr = ip6->ip6_src; inc.inc6_laddr = ip6->ip6_dst; } else #endif { inc.inc_faddr = ip->ip_src; inc.inc_laddr = ip->ip_dst; } inc.inc_fport = th->th_sport; inc.inc_lport = th->th_dport; inc.inc_fibnum = so->so_fibnum; /* * Check for an existing connection attempt in syncache if * the flag is only ACK. A successful lookup creates a new * socket appended to the listen queue in SYN_RECEIVED state. */ if ((thflags & (TH_RST|TH_ACK|TH_SYN)) == TH_ACK) { /* * Parse the TCP options here because * syncookies need access to the reflected * timestamp. */ tcp_dooptions(&to, optp, optlen, 0); /* * NB: syncache_expand() doesn't unlock * inp and tcpinfo locks. */ rstreason = syncache_expand(&inc, &to, th, &so, m); if (rstreason < 0) { /* * A failing TCP MD5 signature comparison * must result in the segment being dropped * and must not produce any response back * to the sender. */ goto dropunlock; } else if (rstreason == 0) { /* * No syncache entry or ACK was not * for our SYN/ACK. Send a RST. * NB: syncache did its own logging * of the failure cause. */ rstreason = BANDLIM_RST_OPENPORT; goto dropwithreset; } tfo_socket_result: if (so == NULL) { /* * We completed the 3-way handshake * but could not allocate a socket * either due to memory shortage, * listen queue length limits or * global socket limits. Send RST * or wait and have the remote end * retransmit the ACK for another * try. */ if ((s = tcp_log_addrs(&inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: Listen socket: " "Socket allocation failed due to " "limits or memory shortage, %s\n", s, __func__, V_tcp_sc_rst_sock_fail ? "sending RST" : "try again"); if (V_tcp_sc_rst_sock_fail) { rstreason = BANDLIM_UNLIMITED; goto dropwithreset; } else goto dropunlock; } /* * Socket is created in state SYN_RECEIVED. * Unlock the listen socket, lock the newly * created socket and update the tp variable. */ INP_WUNLOCK(inp); /* listen socket */ inp = sotoinpcb(so); /* * New connection inpcb is already locked by * syncache_expand(). */ INP_WLOCK_ASSERT(inp); tp = intotcpcb(inp); KASSERT(tp->t_state == TCPS_SYN_RECEIVED, ("%s: ", __func__)); /* * Process the segment and the data it * contains. tcp_do_segment() consumes * the mbuf chain and unlocks the inpcb. */ TCP_PROBE5(receive, NULL, tp, m, tp, th); tp->t_fb->tfb_tcp_do_segment(m, th, so, tp, drop_hdrlen, tlen, iptos); return (IPPROTO_DONE); } /* * Segment flag validation for new connection attempts: * * Our (SYN|ACK) response was rejected. * Check with syncache and remove entry to prevent * retransmits. * * NB: syncache_chkrst does its own logging of failure * causes. */ if (thflags & TH_RST) { syncache_chkrst(&inc, th, m); goto dropunlock; } /* * We can't do anything without SYN. */ if ((thflags & TH_SYN) == 0) { if ((s = tcp_log_addrs(&inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: Listen socket: " "SYN is missing, segment ignored\n", s, __func__); TCPSTAT_INC(tcps_badsyn); goto dropunlock; } /* * (SYN|ACK) is bogus on a listen socket. */ if (thflags & TH_ACK) { if ((s = tcp_log_addrs(&inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: Listen socket: " "SYN|ACK invalid, segment rejected\n", s, __func__); syncache_badack(&inc); /* XXX: Not needed! */ TCPSTAT_INC(tcps_badsyn); rstreason = BANDLIM_RST_OPENPORT; goto dropwithreset; } /* * If the drop_synfin option is enabled, drop all * segments with both the SYN and FIN bits set. * This prevents e.g. nmap from identifying the * TCP/IP stack. * XXX: Poor reasoning. nmap has other methods * and is constantly refining its stack detection * strategies. * XXX: This is a violation of the TCP specification * and was used by RFC1644. */ if ((thflags & TH_FIN) && V_drop_synfin) { if ((s = tcp_log_addrs(&inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: Listen socket: " "SYN|FIN segment ignored (based on " "sysctl setting)\n", s, __func__); TCPSTAT_INC(tcps_badsyn); goto dropunlock; } /* * Segment's flags are (SYN) or (SYN|FIN). * * TH_PUSH, TH_URG, TH_ECE, TH_CWR are ignored * as they do not affect the state of the TCP FSM. * The data pointed to by TH_URG and th_urp is ignored. */ KASSERT((thflags & (TH_RST|TH_ACK)) == 0, ("%s: Listen socket: TH_RST or TH_ACK set", __func__)); KASSERT(thflags & (TH_SYN), ("%s: Listen socket: TH_SYN not set", __func__)); #ifdef INET6 /* * If deprecated address is forbidden, * we do not accept SYN to deprecated interface * address to prevent any new inbound connection from * getting established. * When we do not accept SYN, we send a TCP RST, * with deprecated source address (instead of dropping * it). We compromise it as it is much better for peer * to send a RST, and RST will be the final packet * for the exchange. * * If we do not forbid deprecated addresses, we accept * the SYN packet. RFC2462 does not suggest dropping * SYN in this case. * If we decipher RFC2462 5.5.4, it says like this: * 1. use of deprecated addr with existing * communication is okay - "SHOULD continue to be * used" * 2. use of it with new communication: * (2a) "SHOULD NOT be used if alternate address * with sufficient scope is available" * (2b) nothing mentioned otherwise. * Here we fall into (2b) case as we have no choice in * our source address selection - we must obey the peer. * * The wording in RFC2462 is confusing, and there are * multiple description text for deprecated address * handling - worse, they are not exactly the same. * I believe 5.5.4 is the best one, so we follow 5.5.4. */ if (isipv6 && !V_ip6_use_deprecated) { struct in6_ifaddr *ia6; ia6 = in6ifa_ifwithaddr(&ip6->ip6_dst, 0 /* XXX */); if (ia6 != NULL && (ia6->ia6_flags & IN6_IFF_DEPRECATED)) { ifa_free(&ia6->ia_ifa); if ((s = tcp_log_addrs(&inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: Listen socket: " "Connection attempt to deprecated " "IPv6 address rejected\n", s, __func__); rstreason = BANDLIM_RST_OPENPORT; goto dropwithreset; } if (ia6) ifa_free(&ia6->ia_ifa); } #endif /* INET6 */ /* * Basic sanity checks on incoming SYN requests: * Don't respond if the destination is a link layer * broadcast according to RFC1122 4.2.3.10, p. 104. * If it is from this socket it must be forged. * Don't respond if the source or destination is a * global or subnet broad- or multicast address. * Note that it is quite possible to receive unicast * link-layer packets with a broadcast IP address. Use * in_broadcast() to find them. */ if (m->m_flags & (M_BCAST|M_MCAST)) { if ((s = tcp_log_addrs(&inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: Listen socket: " "Connection attempt from broad- or multicast " "link layer address ignored\n", s, __func__); goto dropunlock; } #ifdef INET6 if (isipv6) { if (th->th_dport == th->th_sport && IN6_ARE_ADDR_EQUAL(&ip6->ip6_dst, &ip6->ip6_src)) { if ((s = tcp_log_addrs(&inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: Listen socket: " "Connection attempt to/from self " "ignored\n", s, __func__); goto dropunlock; } if (IN6_IS_ADDR_MULTICAST(&ip6->ip6_dst) || IN6_IS_ADDR_MULTICAST(&ip6->ip6_src)) { if ((s = tcp_log_addrs(&inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: Listen socket: " "Connection attempt from/to multicast " "address ignored\n", s, __func__); goto dropunlock; } } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { if (th->th_dport == th->th_sport && ip->ip_dst.s_addr == ip->ip_src.s_addr) { if ((s = tcp_log_addrs(&inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: Listen socket: " "Connection attempt from/to self " "ignored\n", s, __func__); goto dropunlock; } if (IN_MULTICAST(ntohl(ip->ip_dst.s_addr)) || IN_MULTICAST(ntohl(ip->ip_src.s_addr)) || ip->ip_src.s_addr == htonl(INADDR_BROADCAST) || in_broadcast(ip->ip_dst, m->m_pkthdr.rcvif)) { if ((s = tcp_log_addrs(&inc, th, NULL, NULL))) log(LOG_DEBUG, "%s; %s: Listen socket: " "Connection attempt from/to broad- " "or multicast address ignored\n", s, __func__); goto dropunlock; } } #endif /* * SYN appears to be valid. Create compressed TCP state * for syncache. */ #ifdef TCPDEBUG if (so->so_options & SO_DEBUG) tcp_trace(TA_INPUT, ostate, tp, (void *)tcp_saveipgen, &tcp_savetcp, 0); #endif TCP_PROBE3(debug__input, tp, th, m); tcp_dooptions(&to, optp, optlen, TO_SYN); if (syncache_add(&inc, &to, th, inp, &so, m, NULL, NULL, iptos)) goto tfo_socket_result; /* * Entry added to syncache and mbuf consumed. * Only the listen socket is unlocked by syncache_add(). */ INP_INFO_WUNLOCK_ASSERT(&V_tcbinfo); return (IPPROTO_DONE); } else if (tp->t_state == TCPS_LISTEN) { /* * When a listen socket is torn down the SO_ACCEPTCONN * flag is removed first while connections are drained * from the accept queue in a unlock/lock cycle of the * ACCEPT_LOCK, opening a race condition allowing a SYN * attempt go through unhandled. */ goto dropunlock; } #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) if (tp->t_flags & TF_SIGNATURE) { tcp_dooptions(&to, optp, optlen, thflags); if ((to.to_flags & TOF_SIGNATURE) == 0) { TCPSTAT_INC(tcps_sig_err_nosigopt); goto dropunlock; } if (!TCPMD5_ENABLED() || TCPMD5_INPUT(m, th, to.to_signature) != 0) goto dropunlock; } #endif TCP_PROBE5(receive, NULL, tp, m, tp, th); /* * Segment belongs to a connection in SYN_SENT, ESTABLISHED or later * state. tcp_do_segment() always consumes the mbuf chain, unlocks * the inpcb, and unlocks pcbinfo. */ tp->t_fb->tfb_tcp_do_segment(m, th, so, tp, drop_hdrlen, tlen, iptos); return (IPPROTO_DONE); dropwithreset: TCP_PROBE5(receive, NULL, tp, m, tp, th); if (inp != NULL) { tcp_dropwithreset(m, th, tp, tlen, rstreason); INP_WUNLOCK(inp); } else tcp_dropwithreset(m, th, NULL, tlen, rstreason); m = NULL; /* mbuf chain got consumed. */ goto drop; dropunlock: if (m != NULL) TCP_PROBE5(receive, NULL, tp, m, tp, th); if (inp != NULL) INP_WUNLOCK(inp); drop: INP_INFO_WUNLOCK_ASSERT(&V_tcbinfo); if (s != NULL) free(s, M_TCPLOG); if (m != NULL) m_freem(m); return (IPPROTO_DONE); } /* * Automatic sizing of receive 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). * * On the receive side the socket buffer memory is only rarely * used to any significant extent. This allows us to be much * more aggressive in scaling the receive socket buffer. For * the case that the buffer space is actually used to a large * extent and we run out of kernel memory we can simply drop * the new segments; TCP on the sender will just retransmit it * later. Setting the buffer size too big may only consume too * much kernel memory if the application doesn't read() from * the socket or packet loss or reordering makes use of the * reassembly queue. * * The criteria to step up the receive buffer one notch are: * 1. Application has not set receive buffer size with * SO_RCVBUF. Setting SO_RCVBUF clears SB_AUTOSIZE. * 2. the number of bytes received during 1/2 of an sRTT * is at least 3/8 of the current socket buffer size. * 3. receive buffer size has not hit maximal automatic size; * * If all of the criteria are met we increaset the socket buffer * by a 1/2 (bounded by the max). This allows us to keep ahead * of slow-start but also makes it so our peer never gets limited * by our rwnd which we then open up causing a burst. * * This algorithm does two steps per RTT at most and only if * we receive a bulk stream w/o packet losses or reorderings. * Shrinking the buffer during idle times is not necessary as * it doesn't consume any memory when idle. * * TODO: Only step up if the application is actually serving * the buffer to better manage the socket buffer resources. */ int tcp_autorcvbuf(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int tlen) { int newsize = 0; if (V_tcp_do_autorcvbuf && (so->so_rcv.sb_flags & SB_AUTOSIZE) && tp->t_srtt != 0 && tp->rfbuf_ts != 0 && TCP_TS_TO_TICKS(tcp_ts_getticks() - tp->rfbuf_ts) > ((tp->t_srtt >> TCP_RTT_SHIFT)/2)) { if (tp->rfbuf_cnt > ((so->so_rcv.sb_hiwat / 2)/ 4 * 3) && so->so_rcv.sb_hiwat < V_tcp_autorcvbuf_max) { newsize = min((so->so_rcv.sb_hiwat + (so->so_rcv.sb_hiwat/2)), V_tcp_autorcvbuf_max); } TCP_PROBE6(receive__autoresize, NULL, tp, m, tp, th, newsize); /* Start over with next RTT. */ tp->rfbuf_ts = 0; tp->rfbuf_cnt = 0; } else { tp->rfbuf_cnt += tlen; /* add up */ } return (newsize); } void tcp_do_segment(struct mbuf *m, struct tcphdr *th, struct socket *so, struct tcpcb *tp, int drop_hdrlen, int tlen, uint8_t iptos) { int thflags, acked, ourfinisacked, needoutput = 0, sack_changed; int rstreason, todrop, win; uint32_t tiwin; uint16_t nsegs; char *s; struct in_conninfo *inc; struct mbuf *mfree; struct tcpopt to; int tfo_syn; #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 thflags = th->th_flags; inc = &tp->t_inpcb->inp_inc; tp->sackhint.last_sack_ack = 0; sack_changed = 0; nsegs = max(1, m->m_pkthdr.lro_nsegs); NET_EPOCH_ASSERT(); INP_WLOCK_ASSERT(tp->t_inpcb); KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN", __func__)); KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT", __func__)); #ifdef TCPPCAP /* Save segment, if requested. */ tcp_pcap_add(th, m, &(tp->t_inpkts)); #endif TCP_LOG_EVENT(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0, tlen, NULL, true); if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) { if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { log(LOG_DEBUG, "%s; %s: " "SYN|FIN segment ignored (based on " "sysctl setting)\n", s, __func__); free(s, M_TCPLOG); } goto drop; } /* * If a segment with the ACK-bit set arrives in the SYN-SENT state * check SEQ.ACK first. */ 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))) { rstreason = BANDLIM_UNLIMITED; goto dropwithreset; } /* * Segment received on connection. * Reset idle time and keep-alive timer. * XXX: This should be done after segment * validation to ignore broken/spoofed segs. */ tp->t_rcvtime = ticks; /* * Scale up 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 /* * TCP ECN processing. */ if (tp->t_flags2 & TF2_ECN_PERMIT) { if (thflags & TH_CWR) tp->t_flags2 &= ~TF2_ECN_SND_ECE; switch (iptos & IPTOS_ECN_MASK) { case IPTOS_ECN_CE: tp->t_flags2 |= TF2_ECN_SND_ECE; TCPSTAT_INC(tcps_ecn_ce); break; case IPTOS_ECN_ECT0: TCPSTAT_INC(tcps_ecn_ect0); break; case IPTOS_ECN_ECT1: TCPSTAT_INC(tcps_ecn_ect1); break; } /* Process a packet differently from RFC3168. */ cc_ecnpkt_handler(tp, th, iptos); /* Congestion experienced. */ if (thflags & TH_ECE) { cc_cong_signal(tp, th, CC_ECN); } } /* * 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 defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE) if ((tp->t_flags & TF_SIGNATURE) != 0 && (to.to_flags & TOF_SIGNATURE) == 0) { TCPSTAT_INC(tcps_sig_err_sigopt); /* XXX: should drop? */ } #endif /* * 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_ts_getticks())) to.to_tsecr = 0; else if (tp->t_flags & TF_PREVVALID && tp->t_badrxtwin != 0 && SEQ_LT(to.to_tsecr, tp->t_badrxtwin)) cc_cong_signal(tp, th, CC_RTO_ERR); } /* * 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)) { if ((to.to_flags & TOF_SCALE) && (tp->t_flags & TF_REQ_SCALE)) { tp->t_flags |= TF_RCVD_SCALE; tp->snd_scale = to.to_wscale; } /* * 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_RCVD_TSTMP; tp->ts_recent = to.to_tsval; tp->ts_recent_age = tcp_ts_getticks(); } 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); } } /* * If timestamps were negotiated during SYN/ACK they should * appear on every segment during this session and vice versa. */ if ((tp->t_flags & TF_RCVD_TSTMP) && !(to.to_flags & TOF_TS)) { if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { log(LOG_DEBUG, "%s; %s: Timestamp missing, " "no action\n", s, __func__); free(s, M_TCPLOG); } } if (!(tp->t_flags & TF_RCVD_TSTMP) && (to.to_flags & TOF_TS)) { if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { log(LOG_DEBUG, "%s; %s: Timestamp not expected, " "no action\n", s, __func__); free(s, M_TCPLOG); } } /* * 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 to * 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. */ if (tp->t_state == TCPS_ESTABLISHED && th->th_seq == tp->rcv_nxt && (thflags & (TH_SYN|TH_FIN|TH_RST|TH_URG|TH_ACK)) == TH_ACK && tp->snd_nxt == tp->snd_max && tiwin && tiwin == tp->snd_wnd && ((tp->t_flags & (TF_NEEDSYN|TF_NEEDFIN)) == 0) && SEGQ_EMPTY(tp) && ((to.to_flags & TOF_TS) == 0 || TSTMP_GEQ(to.to_tsval, tp->ts_recent)) ) { /* * 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; } if (tlen == 0) { if (SEQ_GT(th->th_ack, tp->snd_una) && SEQ_LEQ(th->th_ack, tp->snd_max) && !IN_RECOVERY(tp->t_flags) && (to.to_flags & TOF_SACK) == 0 && TAILQ_EMPTY(&tp->snd_holes)) { /* * This is a pure ack for outstanding data. */ TCPSTAT_INC(tcps_predack); /* * "bad retransmit" recovery without timestamps. */ if ((to.to_flags & TOF_TS) == 0 && tp->t_rxtshift == 1 && tp->t_flags & TF_PREVVALID && (int)(ticks - tp->t_badrxtwin) < 0) { cc_cong_signal(tp, th, CC_RTO_ERR); } /* * 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. */ if ((to.to_flags & TOF_TS) != 0 && to.to_tsecr) { uint32_t t; t = tcp_ts_getticks() - to.to_tsecr; if (!tp->t_rttlow || tp->t_rttlow > t) tp->t_rttlow = t; tcp_xmit_timer(tp, TCP_TS_TO_TICKS(t) + 1); } else if (tp->t_rtttime && SEQ_GT(th->th_ack, tp->t_rtseq)) { if (!tp->t_rttlow || tp->t_rttlow > ticks - tp->t_rtttime) tp->t_rttlow = ticks - tp->t_rtttime; tcp_xmit_timer(tp, ticks - tp->t_rtttime); } 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 TCPSTAT_ADD(tcps_rcvackpack, nsegs); TCPSTAT_ADD(tcps_rcvackbyte, acked); sbdrop(&so->so_snd, acked); if (SEQ_GT(tp->snd_una, tp->snd_recover) && SEQ_LEQ(th->th_ack, tp->snd_recover)) tp->snd_recover = th->th_ack - 1; /* * Let the congestion control algorithm update * congestion control related information. This * typically means increasing the congestion * window. */ cc_ack_received(tp, th, nsegs, CC_ACK); tp->snd_una = th->th_ack; /* * 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); /* * 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 TCP_PROBE3(debug__input, tp, th, m); if (tp->snd_una == tp->snd_max) tcp_timer_activate(tp, TT_REXMT, 0); else if (!tcp_timer_active(tp, TT_PERSIST)) tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); sowwakeup(so); if (sbavail(&so->so_snd)) (void) tp->t_fb->tfb_tcp_output(tp); goto check_delack; } } else if (th->th_ack == tp->snd_una && tlen <= sbspace(&so->so_rcv)) { int newsize = 0; /* automatic sockbuf scaling */ /* * This is a pure, in-sequence data packet with * nothing on the reassembly queue and we have enough * buffer space to take it. */ /* Clean receiver SACK report if present */ if ((tp->t_flags & TF_SACK_PERMIT) && tp->rcv_numsacks) tcp_clean_sackreport(tp); TCPSTAT_INC(tcps_preddat); tp->rcv_nxt += tlen; /* * 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; TCPSTAT_ADD(tcps_rcvpack, nsegs); 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 TCP_PROBE3(debug__input, tp, th, m); 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, so, NULL)) so->so_rcv.sb_flags &= ~SB_AUTOSIZE; m_adj(m, drop_hdrlen); /* delayed header drop */ sbappendstream_locked(&so->so_rcv, m, 0); } /* NB: sorwakeup_locked() does an implicit unlock. */ sorwakeup_locked(so); if (DELAY_ACK(tp, tlen)) { tp->t_flags |= TF_DELACK; } else { tp->t_flags |= TF_ACKNOW; tp->t_fb->tfb_tcp_output(tp); } goto check_delack; } } /* * Calculate amount of space in receive window, * and then do TCP input processing. * Receive window is amount of space in rcv queue, * but not less than advertised window. */ win = sbspace(&so->so_rcv); if (win < 0) win = 0; tp->rcv_wnd = imax(win, (int)(tp->rcv_adv - tp->rcv_nxt)); switch (tp->t_state) { /* * If the state is SYN_RECEIVED: * if seg contains an ACK, but not for our SYN/ACK, send a RST. */ case TCPS_SYN_RECEIVED: if ((thflags & TH_ACK) && (SEQ_LEQ(th->th_ack, tp->snd_una) || SEQ_GT(th->th_ack, tp->snd_max))) { rstreason = BANDLIM_RST_OPENPORT; goto dropwithreset; } 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)) { rstreason = BANDLIM_RST_OPENPORT; goto dropwithreset; } else if (thflags & TH_SYN) { /* non-initial SYN is ignored */ if ((tcp_timer_active(tp, TT_DELACK) || tcp_timer_active(tp, TT_REXMT))) goto drop; } else if (!(thflags & (TH_ACK|TH_FIN|TH_RST))) { goto drop; } } break; /* * If the state is SYN_SENT: * if seg contains a RST with valid ACK (SEQ.ACK has already * been verified), then drop the connection. * if seg contains a RST without an ACK, drop the seg. * if seg does not contain SYN, then drop the seg. * 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, beginning with URG */ case TCPS_SYN_SENT: if ((thflags & (TH_ACK|TH_RST)) == (TH_ACK|TH_RST)) { TCP_PROBE5(connect__refused, NULL, tp, m, tp, th); tp = tcp_drop(tp, ECONNREFUSED); } if (thflags & TH_RST) goto drop; if (!(thflags & TH_SYN)) goto drop; tp->irs = th->th_seq; tcp_rcvseqinit(tp); if (thflags & TH_ACK) { int tfo_partial_ack = 0; 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); tp->snd_una++; /* SYN is acked */ /* * 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_ack = 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_ack) tcp_timer_activate(tp, TT_DELACK, tcp_delacktime); else tp->t_flags |= TF_ACKNOW; if (((thflags & (TH_CWR | TH_ECE)) == TH_ECE) && V_tcp_do_ecn) { tp->t_flags2 |= TF2_ECN_PERMIT; TCPSTAT_INC(tcps_ecn_shs); } /* * 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, m, tp, th); cc_conn_init(tp); tcp_timer_activate(tp, TT_KEEP, TP_KEEPIDLE(tp)); } } else { /* * Received initial SYN in SYN-SENT[*] state => * simultaneous open. * If it succeeds, connection is * half-synchronized. * Otherwise, do 3-way handshake: * SYN-SENT -> SYN-RECEIVED * SYN-SENT* -> SYN-RECEIVED* */ tp->t_flags |= (TF_ACKNOW | TF_NEEDSYN); tcp_timer_activate(tp, TT_REXMT, 0); 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; TCPSTAT_INC(tcps_rcvpackafterwin); 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) goto process_ACK; goto step6; /* * If the state is LAST_ACK or CLOSING or TIME_WAIT: * do normal processing. * * NB: Leftover from RFC1644 T/TCP. Cases to be reused later. */ case TCPS_LAST_ACK: case TCPS_CLOSING: break; /* continue normal processing */ } /* * States other than LISTEN or SYN_SENT. * First check the RST flag and sequence number since reset segments * are exempt from the timestamp and connection count tests. This * fixes a bug introduced by the Stevens, vol. 2, p. 960 bugfix * below which allowed reset segments in half the sequence space * to fall though and be processed (which gives forged reset * segments with a random sequence number a 50 percent chance of * killing a connection). * Then check timestamp, if present. * Then check the connection count, if present. * Then check that at least some bytes of segment are within * receive window. If segment begins before rcv_nxt, * drop leading data (and SYN); if nothing left, just ack. */ if (thflags & TH_RST) { /* * RFC5961 Section 3.2 * * - RST drops connection only if SEG.SEQ == RCV.NXT. * - If RST is in window, we send challenge ACK. * * Note: to take into account delayed ACKs, we should * test against last_ack_sent instead of rcv_nxt. * Note 2: we handle special case of closed window, not * covered by the RFC. */ if ((SEQ_GEQ(th->th_seq, tp->last_ack_sent) && SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) || (tp->rcv_wnd == 0 && tp->last_ack_sent == th->th_seq)) { KASSERT(tp->t_state != TCPS_SYN_SENT, ("%s: TH_RST for TCPS_SYN_SENT th %p tp %p", __func__, th, tp)); if (V_tcp_insecure_rst || tp->last_ack_sent == th->th_seq) { TCPSTAT_INC(tcps_drops); /* Drop the connection. */ switch (tp->t_state) { case TCPS_SYN_RECEIVED: so->so_error = ECONNREFUSED; goto close; case TCPS_ESTABLISHED: case TCPS_FIN_WAIT_1: case TCPS_FIN_WAIT_2: case TCPS_CLOSE_WAIT: case TCPS_CLOSING: case TCPS_LAST_ACK: so->so_error = ECONNRESET; close: /* FALLTHROUGH */ default: tp = tcp_close(tp); } } else { TCPSTAT_INC(tcps_badrst); /* Send challenge ACK. */ tcp_respond(tp, mtod(m, void *), th, m, tp->rcv_nxt, tp->snd_nxt, TH_ACK); tp->last_ack_sent = tp->rcv_nxt; m = NULL; } } goto drop; } /* * RFC5961 Section 4.2 * Send challenge ACK for any SYN in synchronized state. */ if ((thflags & TH_SYN) && tp->t_state != TCPS_SYN_SENT && tp->t_state != TCPS_SYN_RECEIVED) { TCPSTAT_INC(tcps_badsyn); if (V_tcp_insecure_syn && SEQ_GEQ(th->th_seq, tp->last_ack_sent) && SEQ_LT(th->th_seq, tp->last_ack_sent + tp->rcv_wnd)) { tp = tcp_drop(tp, ECONNRESET); rstreason = BANDLIM_UNLIMITED; } else { /* Send challenge ACK. */ tcp_respond(tp, mtod(m, void *), th, m, tp->rcv_nxt, tp->snd_nxt, TH_ACK); tp->last_ack_sent = tp->rcv_nxt; m = NULL; } goto drop; } /* * 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)) { /* Check to see if ts_recent is over 24 days old. */ if (tcp_ts_getticks() - tp->ts_recent_age > TCP_PAWS_IDLE) { /* * Invalidate ts_recent. If this segment updates * ts_recent, the age will be reset later and ts_recent * will get a valid value. If it does not, setting * ts_recent to zero will at least satisfy the * requirement that zero be placed in the timestamp * echo reply when ts_recent isn't valid. The * age isn't reset until we get a valid ts_recent * because we don't want out-of-order segments to be * dropped when ts_recent is old. */ tp->ts_recent = 0; } else { TCPSTAT_INC(tcps_rcvduppack); TCPSTAT_ADD(tcps_rcvdupbyte, tlen); TCPSTAT_INC(tcps_pawsdrop); if (tlen) goto dropafterack; goto drop; } } /* * 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 (tp->t_state == TCPS_SYN_RECEIVED && SEQ_LT(th->th_seq, tp->irs)) { rstreason = BANDLIM_RST_OPENPORT; goto dropwithreset; } todrop = tp->rcv_nxt - th->th_seq; if (todrop > 0) { if (thflags & TH_SYN) { thflags &= ~TH_SYN; th->th_seq++; if (th->th_urp > 1) th->th_urp--; else thflags &= ~TH_URG; todrop--; } /* * Following if statement from Stevens, vol. 2, p. 960. */ if (todrop > tlen || (todrop == tlen && (thflags & TH_FIN) == 0)) { /* * Any valid FIN must be to the left of the window. * At this point the FIN must be a duplicate or out * of sequence; drop it. */ thflags &= ~TH_FIN; /* * Send an ACK to resynchronize and drop any data. * But keep on processing for RST or ACK. */ tp->t_flags |= TF_ACKNOW; todrop = tlen; TCPSTAT_INC(tcps_rcvduppack); TCPSTAT_ADD(tcps_rcvdupbyte, todrop); } else { TCPSTAT_INC(tcps_rcvpartduppack); TCPSTAT_ADD(tcps_rcvpartdupbyte, todrop); } /* * DSACK - add SACK block for dropped range */ if (tp->t_flags & TF_SACK_PERMIT) { tcp_update_sack_list(tp, th->th_seq, th->th_seq + todrop); /* * ACK now, as the next in-sequence segment * will clear the DSACK block again */ tp->t_flags |= TF_ACKNOW; } drop_hdrlen += todrop; /* drop from the top afterwards */ th->th_seq += todrop; tlen -= todrop; if (th->th_urp > todrop) th->th_urp -= todrop; else { thflags &= ~TH_URG; th->th_urp = 0; } } /* * If new data are received on a connection after the * user processes are gone, then RST the other end. */ if ((so->so_state & SS_NOFDREF) && tp->t_state > TCPS_CLOSE_WAIT && tlen) { if ((s = tcp_log_addrs(inc, th, NULL, NULL))) { log(LOG_DEBUG, "%s; %s: %s: Received %d bytes of data " "after socket was closed, " "sending RST and removing tcpcb\n", s, __func__, tcpstates[tp->t_state], tlen); free(s, M_TCPLOG); } tp = tcp_close(tp); TCPSTAT_INC(tcps_rcvafterclose); rstreason = BANDLIM_UNLIMITED; goto dropwithreset; } /* * If segment ends after window, drop trailing data * (and PUSH and FIN); if nothing left, just ACK. */ todrop = (th->th_seq + tlen) - (tp->rcv_nxt + tp->rcv_wnd); if (todrop > 0) { TCPSTAT_INC(tcps_rcvpackafterwin); if (todrop >= tlen) { TCPSTAT_ADD(tcps_rcvbyteafterwin, tlen); /* * If window is closed can only take segments at * window edge, and have to drop data and PUSH from * incoming segments. Continue processing, but * remember to ack. Otherwise, drop segment * and ack. */ if (tp->rcv_wnd == 0 && th->th_seq == tp->rcv_nxt) { tp->t_flags |= TF_ACKNOW; TCPSTAT_INC(tcps_rcvwinprobe); } else goto dropafterack; } else TCPSTAT_ADD(tcps_rcvbyteafterwin, todrop); m_adj(m, -todrop); tlen -= todrop; thflags &= ~(TH_PUSH|TH_FIN); } /* * 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_state == TCPS_SYN_RECEIVED || (tp->t_flags & TF_NEEDSYN)) { if (tp->t_state == TCPS_SYN_RECEIVED && IS_FASTOPEN(tp->t_flags)) { tp->snd_wnd = tiwin; cc_conn_init(tp); } goto step6; } else if (tp->t_flags & TF_ACKNOW) goto dropafterack; else goto drop; } /* * Ack processing. */ switch (tp->t_state) { /* * In SYN_RECEIVED state, the ack ACKs our SYN, so enter * ESTABLISHED state and continue processing. * The ACK was checked above. */ case TCPS_SYN_RECEIVED: TCPSTAT_INC(tcps_connects); 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; } tp->snd_wnd = tiwin; /* * 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; /* * Account for the ACK of our SYN prior to * regular ACK processing below. */ tp->snd_una++; } 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, m, 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); tcp_timer_activate(tp, TT_KEEP, TP_KEEPIDLE(tp)); } /* * 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); tp->snd_wl1 = th->th_seq - 1; /* FALLTHROUGH */ /* * In ESTABLISHED state: drop duplicate ACKs; ACK out of range * ACKs. If the ack is in the range * tp->snd_una < th->th_ack <= tp->snd_max * then advance tp->snd_una to th->th_ack and drop * data from the retransmission queue. If this ACK reflects * more up to date window information we update our window information. */ case TCPS_ESTABLISHED: case TCPS_FIN_WAIT_1: case TCPS_FIN_WAIT_2: case TCPS_CLOSE_WAIT: case TCPS_CLOSING: case TCPS_LAST_ACK: if (SEQ_GT(th->th_ack, tp->snd_max)) { TCPSTAT_INC(tcps_rcvacktoomuch); goto dropafterack; } if ((tp->t_flags & TF_SACK_PERMIT) && ((to.to_flags & TOF_SACK) || !TAILQ_EMPTY(&tp->snd_holes))) sack_changed = tcp_sack_doack(tp, &to, th->th_ack); else /* * Reset the value so that previous (valid) value * from the last ack with SACK doesn't get used. */ tp->sackhint.sacked_bytes = 0; #ifdef TCP_HHOOK /* Run HHOOK_TCP_ESTABLISHED_IN helper hooks. */ hhook_run_tcp_est_in(tp, th, &to); #endif if (SEQ_LEQ(th->th_ack, tp->snd_una)) { u_int maxseg; maxseg = tcp_maxseg(tp); if (tlen == 0 && (tiwin == tp->snd_wnd || (tp->t_flags & TF_SACK_PERMIT))) { /* * If this is the first time we've seen a * FIN from the remote, this is not a * duplicate and it needs to be processed * normally. This happens during a * simultaneous close. */ if ((thflags & TH_FIN) && (TCPS_HAVERCVDFIN(tp->t_state) == 0)) { tp->t_dupacks = 0; break; } TCPSTAT_INC(tcps_rcvdupack); /* * If we have outstanding data (other than * a window probe), this is a completely * duplicate ack (ie, window info didn't * change and FIN isn't set), * the ack is the biggest we've * seen and we've seen exactly our rexmt * threshold of them, assume a packet * has been dropped and retransmit it. * Kludge snd_nxt & the congestion * window so we send only this one * packet. * * We know we're losing at the current * window size so do congestion avoidance * (set ssthresh to half the current window * and pull our congestion window back to * the new ssthresh). * * Dup acks mean that packets have left the * network (they're now cached at the receiver) * so bump cwnd by the amount in the receiver * to keep a constant cwnd packets in the * network. * * When using TCP ECN, notify the peer that * we reduced the cwnd. */ /* * Following 2 kinds of acks should not affect * dupack counting: * 1) Old acks * 2) Acks with SACK but without any new SACK * information in them. These could result from * any anomaly in the network like a switch * duplicating packets or a possible DoS attack. */ if (th->th_ack != tp->snd_una || ((tp->t_flags & TF_SACK_PERMIT) && !sack_changed)) break; else if (!tcp_timer_active(tp, TT_REXMT)) tp->t_dupacks = 0; else if (++tp->t_dupacks > tcprexmtthresh || IN_FASTRECOVERY(tp->t_flags)) { cc_ack_received(tp, th, nsegs, CC_DUPACK); if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags)) { int awnd; /* * Compute the amount of data in flight first. * We can inject new data into the pipe iff * we have less than 1/2 the original window's * worth of data in flight. */ if (V_tcp_do_rfc6675_pipe) awnd = tcp_compute_pipe(tp); else awnd = (tp->snd_nxt - tp->snd_fack) + tp->sackhint.sack_bytes_rexmit; if (awnd < tp->snd_ssthresh) { tp->snd_cwnd += maxseg; if (tp->snd_cwnd > tp->snd_ssthresh) tp->snd_cwnd = tp->snd_ssthresh; } } else tp->snd_cwnd += maxseg; (void) tp->t_fb->tfb_tcp_output(tp); goto drop; } else if (tp->t_dupacks == tcprexmtthresh) { tcp_seq onxt = tp->snd_nxt; /* * If we're doing sack, check to * see if we're already in sack * recovery. If we're not doing sack, * check to see if we're in newreno * recovery. */ if (tp->t_flags & TF_SACK_PERMIT) { if (IN_FASTRECOVERY(tp->t_flags)) { tp->t_dupacks = 0; break; } } else { if (SEQ_LEQ(th->th_ack, tp->snd_recover)) { tp->t_dupacks = 0; break; } } /* Congestion signal before ack. */ cc_cong_signal(tp, th, CC_NDUPACK); cc_ack_received(tp, th, nsegs, CC_DUPACK); tcp_timer_activate(tp, TT_REXMT, 0); tp->t_rtttime = 0; if (tp->t_flags & TF_SACK_PERMIT) { TCPSTAT_INC( tcps_sack_recovery_episode); tp->sack_newdata = tp->snd_nxt; tp->snd_cwnd = maxseg; (void) tp->t_fb->tfb_tcp_output(tp); goto drop; } tp->snd_nxt = th->th_ack; tp->snd_cwnd = maxseg; (void) tp->t_fb->tfb_tcp_output(tp); KASSERT(tp->snd_limited <= 2, ("%s: tp->snd_limited too big", __func__)); tp->snd_cwnd = tp->snd_ssthresh + maxseg * (tp->t_dupacks - tp->snd_limited); if (SEQ_GT(onxt, tp->snd_nxt)) tp->snd_nxt = onxt; goto drop; } else if (V_tcp_do_rfc3042) { /* * Process first and second duplicate * ACKs. Each indicates a segment * leaving the network, creating room * for more. Make sure we can send a * packet on reception of each duplicate * ACK by increasing snd_cwnd by one * segment. Restore the original * snd_cwnd after packet transmission. */ cc_ack_received(tp, th, nsegs, CC_DUPACK); uint32_t oldcwnd = tp->snd_cwnd; tcp_seq oldsndmax = tp->snd_max; u_int sent; int avail; KASSERT(tp->t_dupacks == 1 || tp->t_dupacks == 2, ("%s: dupacks not 1 or 2", __func__)); if (tp->t_dupacks == 1) tp->snd_limited = 0; tp->snd_cwnd = (tp->snd_nxt - tp->snd_una) + (tp->t_dupacks - tp->snd_limited) * maxseg; /* * Only call tcp_output when there * is new data available to be sent. * Otherwise we would send pure ACKs. */ SOCKBUF_LOCK(&so->so_snd); avail = sbavail(&so->so_snd) - (tp->snd_nxt - tp->snd_una); SOCKBUF_UNLOCK(&so->so_snd); if (avail > 0) (void) tp->t_fb->tfb_tcp_output(tp); sent = tp->snd_max - oldsndmax; if (sent > maxseg) { KASSERT((tp->t_dupacks == 2 && tp->snd_limited == 0) || (sent == maxseg + 1 && tp->t_flags & TF_SENTFIN), ("%s: sent too much", __func__)); tp->snd_limited = 2; } else if (sent > 0) ++tp->snd_limited; tp->snd_cwnd = oldcwnd; goto drop; } } break; } else { /* * This ack is advancing the left edge, reset the * counter. */ tp->t_dupacks = 0; /* * If this ack also has new SACK info, increment the * counter as per rfc6675. */ if ((tp->t_flags & TF_SACK_PERMIT) && sack_changed) tp->t_dupacks++; } KASSERT(SEQ_GT(th->th_ack, tp->snd_una), ("%s: th_ack <= snd_una", __func__)); /* * If the congestion window was inflated to account * for the other side's cached packets, retract it. */ if (IN_FASTRECOVERY(tp->t_flags)) { if (SEQ_LT(th->th_ack, tp->snd_recover)) { if (tp->t_flags & TF_SACK_PERMIT) tcp_sack_partialack(tp, th); else tcp_newreno_partial_ack(tp, th); } else cc_post_recovery(tp, th); } /* * 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. */ } } process_ACK: INP_WLOCK_ASSERT(tp->t_inpcb); acked = BYTES_THIS_ACK(tp, th); KASSERT(acked >= 0, ("%s: acked unexepectedly negative " "(tp->snd_una=%u, th->th_ack=%u, tp=%p, m=%p)", __func__, tp->snd_una, th->th_ack, tp, m)); TCPSTAT_ADD(tcps_rcvackpack, nsegs); 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_rxtshift == 1 && tp->t_flags & TF_PREVVALID && tp->t_badrxtwin && SEQ_LT(to.to_tsecr, tp->t_badrxtwin)) cc_cong_signal(tp, th, CC_RTO_ERR); /* * 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 ((to.to_flags & TOF_TS) != 0 && to.to_tsecr) { uint32_t t; t = tcp_ts_getticks() - to.to_tsecr; if (!tp->t_rttlow || tp->t_rttlow > t) tp->t_rttlow = t; tcp_xmit_timer(tp, TCP_TS_TO_TICKS(t) + 1); } else if (tp->t_rtttime && SEQ_GT(th->th_ack, tp->t_rtseq)) { if (!tp->t_rttlow || tp->t_rttlow > ticks - tp->t_rtttime) tp->t_rttlow = ticks - tp->t_rtttime; tcp_xmit_timer(tp, ticks - tp->t_rtttime); } /* * 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 (th->th_ack == tp->snd_max) { tcp_timer_activate(tp, TT_REXMT, 0); needoutput = 1; } else if (!tcp_timer_active(tp, TT_PERSIST)) tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur); /* * If no data (only SYN) was ACK'd, * skip rest of ACK processing. */ if (acked == 0) goto step6; /* * Let the congestion control algorithm update congestion * control related information. This typically means increasing * the congestion window. */ cc_ack_received(tp, th, nsegs, CC_ACK); SOCKBUF_LOCK(&so->so_snd); if (acked > sbavail(&so->so_snd)) { if (tp->snd_wnd >= sbavail(&so->so_snd)) tp->snd_wnd -= sbavail(&so->so_snd); else tp->snd_wnd = 0; mfree = sbcut_locked(&so->so_snd, (int)sbavail(&so->so_snd)); ourfinisacked = 1; } else { mfree = sbcut_locked(&so->so_snd, acked); if (tp->snd_wnd >= (uint32_t) acked) tp->snd_wnd -= acked; else tp->snd_wnd = 0; ourfinisacked = 0; } /* NB: sowwakeup_locked() does an implicit unlock. */ sowwakeup_locked(so); m_freem(mfree); /* Detect una wraparound. */ if (!IN_RECOVERY(tp->t_flags) && SEQ_GT(tp->snd_una, tp->snd_recover) && SEQ_LEQ(th->th_ack, tp->snd_recover)) tp->snd_recover = th->th_ack - 1; /* XXXLAS: Can this be moved up into cc_post_recovery? */ if (IN_RECOVERY(tp->t_flags) && SEQ_GEQ(th->th_ack, tp->snd_recover)) { EXIT_RECOVERY(tp->t_flags); } tp->snd_una = th->th_ack; if (tp->t_flags & TF_SACK_PERMIT) { 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; switch (tp->t_state) { /* * 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. */ case TCPS_FIN_WAIT_1: 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); } break; /* * In CLOSING STATE in addition to the processing for * the ESTABLISHED state if the ACK acknowledges our FIN * then enter the TIME-WAIT state, otherwise ignore * the segment. */ case TCPS_CLOSING: if (ourfinisacked) { tcp_twstart(tp); m_freem(m); return; } break; /* * In LAST_ACK, we may still be waiting for data to drain * and/or to be acked, as well as for the ack of our FIN. * If our FIN is now acknowledged, delete the TCB, * enter the closed state and return. */ case TCPS_LAST_ACK: if (ourfinisacked) { tp = tcp_close(tp); goto drop; } break; } } step6: INP_WLOCK_ASSERT(tp->t_inpcb); /* * Update window information. * Don't look at window if no ACK: TAC's send garbage on first SYN. */ 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) 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; needoutput = 1; } /* * Process segments with URG. */ if ((thflags & TH_URG) && th->th_urp && TCPS_HAVERCVDFIN(tp->t_state) == 0) { /* * This is a kludge, but if we receive and accept * random urgent pointers, we'll crash in * soreceive. It's hard to imagine someone * actually wanting to send this much urgent data. */ SOCKBUF_LOCK(&so->so_rcv); if (th->th_urp + sbavail(&so->so_rcv) > sb_max) { th->th_urp = 0; /* XXX */ thflags &= ~TH_URG; /* XXX */ SOCKBUF_UNLOCK(&so->so_rcv); /* XXX */ goto dodata; /* XXX */ } /* * If this segment advances the known urgent pointer, * then mark the data stream. This should not happen * in CLOSE_WAIT, CLOSING, LAST_ACK or TIME_WAIT STATES since * a FIN has been received from the remote side. * In these states we ignore the URG. * * 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 as the original * spec states (in one of two places). */ if (SEQ_GT(th->th_seq+th->th_urp, tp->rcv_up)) { tp->rcv_up = th->th_seq + th->th_urp; so->so_oobmark = sbavail(&so->so_rcv) + (tp->rcv_up - tp->rcv_nxt) - 1; if (so->so_oobmark == 0) so->so_rcv.sb_state |= SBS_RCVATMARK; sohasoutofband(so); tp->t_oobflags &= ~(TCPOOB_HAVEDATA | TCPOOB_HADDATA); } SOCKBUF_UNLOCK(&so->so_rcv); /* * Remove out of band data so doesn't get presented to user. * This can happen independent of advancing the URG pointer, * but if two URG's are pending at once, some out-of-band * data may creep in... ick. */ if (th->th_urp <= (uint32_t)tlen && !(so->so_options & SO_OOBINLINE)) { /* hdr drop is delayed */ tcp_pulloutofband(so, th, m, drop_hdrlen); } } else { /* * If no out of band data is expected, * pull receive urgent pointer along * with the receive window. */ if (SEQ_GT(tp->rcv_nxt, tp->rcv_up)) tp->rcv_up = tp->rcv_nxt; } dodata: /* XXX */ 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) && 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)) { if (DELAY_ACK(tp, tlen) || tfo_syn) tp->t_flags |= TF_DELACK; else tp->t_flags |= TF_ACKNOW; tp->rcv_nxt += tlen; thflags = th->th_flags & TH_FIN; TCPSTAT_INC(tcps_rcvpack); TCPSTAT_ADD(tcps_rcvbyte, tlen); SOCKBUF_LOCK(&so->so_rcv); if (so->so_rcv.sb_state & SBS_CANTRCVMORE) m_freem(m); else sbappendstream_locked(&so->so_rcv, m, 0); /* NB: sorwakeup_locked() does an implicit unlock. */ sorwakeup_locked(so); } 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_SACK_PERMIT) && (save_tlen > 0)) { 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); } } #if 0 /* * Note the amount of data that peer has sent into * our window, in order to estimate the sender's * buffer size. * XXX: Unused. */ if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) len = so->so_rcv.sb_hiwat - (tp->rcv_adv - tp->rcv_nxt); else len = so->so_rcv.sb_hiwat; #endif } 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) { 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; 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: tcp_twstart(tp); return; } } #ifdef TCPDEBUG if (so->so_options & SO_DEBUG) tcp_trace(TA_INPUT, ostate, tp, (void *)tcp_saveipgen, &tcp_savetcp, 0); #endif TCP_PROBE3(debug__input, tp, th, m); /* * Return any desired output. */ if (needoutput || (tp->t_flags & TF_ACKNOW)) (void) tp->t_fb->tfb_tcp_output(tp); check_delack: INP_WLOCK_ASSERT(tp->t_inpcb); if (tp->t_flags & TF_DELACK) { tp->t_flags &= ~TF_DELACK; tcp_timer_activate(tp, TT_DELACK, tcp_delacktime); } INP_WUNLOCK(tp->t_inpcb); return; dropafterack: /* * Generate an ACK dropping incoming segment if it occupies * sequence space, where the ACK reflects our state. * * We can now skip the test for the RST flag since all * paths to this code happen after packets containing * RST have been dropped. * * In the SYN-RECEIVED state, don't send an ACK unless the * segment we received passes the SYN-RECEIVED ACK test. * If it fails send a RST. This breaks the loop in the * "LAND" DoS attack, and also prevents an ACK storm * between two listening ports that have been sent forged * SYN segments, each with the source address of the other. */ if (tp->t_state == TCPS_SYN_RECEIVED && (thflags & TH_ACK) && (SEQ_GT(tp->snd_una, th->th_ack) || SEQ_GT(th->th_ack, tp->snd_max)) ) { rstreason = BANDLIM_RST_OPENPORT; goto dropwithreset; } #ifdef TCPDEBUG if (so->so_options & SO_DEBUG) tcp_trace(TA_DROP, ostate, tp, (void *)tcp_saveipgen, &tcp_savetcp, 0); #endif TCP_PROBE3(debug__input, tp, th, m); tp->t_flags |= TF_ACKNOW; (void) tp->t_fb->tfb_tcp_output(tp); INP_WUNLOCK(tp->t_inpcb); m_freem(m); return; dropwithreset: if (tp != NULL) { tcp_dropwithreset(m, th, tp, tlen, rstreason); INP_WUNLOCK(tp->t_inpcb); } else tcp_dropwithreset(m, th, NULL, tlen, rstreason); return; drop: /* * Drop space held by incoming segment and return. */ #ifdef TCPDEBUG if (tp == NULL || (tp->t_inpcb->inp_socket->so_options & SO_DEBUG)) tcp_trace(TA_DROP, ostate, tp, (void *)tcp_saveipgen, &tcp_savetcp, 0); #endif TCP_PROBE3(debug__input, tp, th, m); if (tp != NULL) INP_WUNLOCK(tp->t_inpcb); m_freem(m); } /* * Issue RST and make ACK acceptable to originator of segment. * The mbuf must still include the original packet header. * tp may be NULL. */ void tcp_dropwithreset(struct mbuf *m, struct tcphdr *th, struct tcpcb *tp, int tlen, int rstreason) { #ifdef INET struct ip *ip; #endif #ifdef INET6 struct ip6_hdr *ip6; #endif if (tp != NULL) { INP_WLOCK_ASSERT(tp->t_inpcb); } /* Don't bother if destination was broadcast/multicast. */ if ((th->th_flags & TH_RST) || m->m_flags & (M_BCAST|M_MCAST)) goto drop; #ifdef INET6 if (mtod(m, struct ip *)->ip_v == 6) { ip6 = mtod(m, struct ip6_hdr *); if (IN6_IS_ADDR_MULTICAST(&ip6->ip6_dst) || IN6_IS_ADDR_MULTICAST(&ip6->ip6_src)) goto drop; /* IPv6 anycast check is done at tcp6_input() */ } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { ip = mtod(m, struct ip *); if (IN_MULTICAST(ntohl(ip->ip_dst.s_addr)) || IN_MULTICAST(ntohl(ip->ip_src.s_addr)) || ip->ip_src.s_addr == htonl(INADDR_BROADCAST) || in_broadcast(ip->ip_dst, m->m_pkthdr.rcvif)) goto drop; } #endif /* Perform bandwidth limiting. */ if (badport_bandlim(rstreason) < 0) goto drop; /* tcp_respond consumes the mbuf chain. */ if (th->th_flags & TH_ACK) { tcp_respond(tp, mtod(m, void *), th, m, (tcp_seq)0, th->th_ack, TH_RST); } else { if (th->th_flags & TH_SYN) tlen++; if (th->th_flags & TH_FIN) tlen++; tcp_respond(tp, mtod(m, void *), th, m, th->th_seq+tlen, (tcp_seq)0, TH_RST|TH_ACK); } return; drop: m_freem(m); } /* * Parse TCP options and place in tcpopt. */ void tcp_dooptions(struct tcpopt *to, u_char *cp, int cnt, int flags) { int opt, optlen; to->to_flags = 0; 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_MAXSEG: if (optlen != TCPOLEN_MAXSEG) continue; if (!(flags & TO_SYN)) continue; to->to_flags |= TOF_MSS; bcopy((char *)cp + 2, (char *)&to->to_mss, sizeof(to->to_mss)); to->to_mss = ntohs(to->to_mss); break; case TCPOPT_WINDOW: if (optlen != TCPOLEN_WINDOW) continue; if (!(flags & TO_SYN)) continue; to->to_flags |= TOF_SCALE; to->to_wscale = min(cp[2], TCP_MAX_WINSHIFT); break; case TCPOPT_TIMESTAMP: if (optlen != TCPOLEN_TIMESTAMP) continue; to->to_flags |= TOF_TS; bcopy((char *)cp + 2, (char *)&to->to_tsval, sizeof(to->to_tsval)); to->to_tsval = ntohl(to->to_tsval); bcopy((char *)cp + 6, (char *)&to->to_tsecr, sizeof(to->to_tsecr)); to->to_tsecr = ntohl(to->to_tsecr); break; case TCPOPT_SIGNATURE: /* * In order to reply to a host which has set the * TCP_SIGNATURE option in its initial SYN, we have * to record the fact that the option was observed * here for the syncache code to perform the correct * response. */ if (optlen != TCPOLEN_SIGNATURE) continue; to->to_flags |= TOF_SIGNATURE; to->to_signature = cp + 2; break; case TCPOPT_SACK_PERMITTED: if (optlen != TCPOLEN_SACK_PERMITTED) continue; if (!(flags & TO_SYN)) continue; if (!V_tcp_do_sack) continue; to->to_flags |= TOF_SACKPERM; break; case TCPOPT_SACK: if (optlen <= 2 || (optlen - 2) % TCPOLEN_SACK != 0) continue; if (flags & TO_SYN) continue; to->to_flags |= TOF_SACK; to->to_nsacks = (optlen - 2) / TCPOLEN_SACK; to->to_sacks = cp + 2; TCPSTAT_INC(tcps_sack_rcv_blocks); break; case TCPOPT_FAST_OPEN: /* * Cookie length validation is performed by the * server side cookie checking code or the client * side cookie cache update code. */ if (!(flags & TO_SYN)) continue; if (!V_tcp_fastopen_client_enable && !V_tcp_fastopen_server_enable) continue; to->to_flags |= TOF_FASTOPEN; to->to_tfo_len = optlen - 2; to->to_tfo_cookie = to->to_tfo_len ? cp + 2 : NULL; break; default: continue; } } } /* * Pull out of band byte out of a segment so * it doesn't appear in the user's data queue. * It is still reflected in the segment length for * sequencing purposes. */ void tcp_pulloutofband(struct socket *so, struct tcphdr *th, struct mbuf *m, int off) { int cnt = off + th->th_urp - 1; while (cnt >= 0) { if (m->m_len > cnt) { char *cp = mtod(m, caddr_t) + cnt; struct tcpcb *tp = sototcpcb(so); INP_WLOCK_ASSERT(tp->t_inpcb); tp->t_iobc = *cp; tp->t_oobflags |= TCPOOB_HAVEDATA; bcopy(cp+1, cp, (unsigned)(m->m_len - cnt - 1)); m->m_len--; if (m->m_flags & M_PKTHDR) m->m_pkthdr.len--; return; } cnt -= m->m_len; m = m->m_next; if (m == NULL) break; } panic("tcp_pulloutofband"); } /* * Collect new round-trip time estimate * and update averages and current timeout. */ void tcp_xmit_timer(struct tcpcb *tp, int rtt) { int delta; INP_WLOCK_ASSERT(tp->t_inpcb); TCPSTAT_INC(tcps_rttupdated); tp->t_rttupdated++; #ifdef STATS stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt * 1000 / hz)); #endif if ((tp->t_srtt != 0) && (tp->t_rxtshift <= TCP_RTT_INVALIDATE)) { /* * 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 - 1) << TCP_DELTA_SHIFT) - (tp->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT)); if ((tp->t_srtt += delta) <= 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); if ((tp->t_rttvar += delta) <= 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 << TCP_RTT_SHIFT; tp->t_rttvar = rtt << (TCP_RTTVAR_SHIFT - 1); tp->t_rttbest = tp->t_srtt + tp->t_rttvar; } tp->t_rtttime = 0; tp->t_rxtshift = 0; /* * 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(tp->t_rttmin, rtt + 2), TCPTV_REXMTMAX); /* * 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; } /* * Determine a reasonable value for maxseg size. * If the route is known, check route for mtu. * If none, use an mss that can be handled on the outgoing interface * without forcing IP to fragment. If no route is found, route has no mtu, * or the destination isn't local, use a default, hopefully conservative * size (usually 512 or the default IP max size, but no more than the mtu * of the interface), as we can't discover anything about intervening * gateways or networks. We also initialize the congestion/slow start * window to be a single segment if the destination isn't local. * While looking at the routing entry, we also initialize other path-dependent * parameters from pre-set or cached values in the routing entry. * * NOTE that resulting t_maxseg doesn't include space for TCP options or * IP options, e.g. IPSEC data, since length of this data may vary, and * thus it is calculated for every segment separately in tcp_output(). * * NOTE that this routine is only called when we process an incoming * segment, or an ICMP need fragmentation datagram. Outgoing SYN/ACK MSS * settings are handled in tcp_mssopt(). */ void tcp_mss_update(struct tcpcb *tp, int offer, int mtuoffer, struct hc_metrics_lite *metricptr, struct tcp_ifcap *cap) { int mss = 0; uint32_t maxmtu = 0; struct inpcb *inp = tp->t_inpcb; struct hc_metrics_lite metrics; #ifdef INET6 int isipv6 = ((inp->inp_vflag & INP_IPV6) != 0) ? 1 : 0; size_t min_protoh = isipv6 ? sizeof (struct ip6_hdr) + sizeof (struct tcphdr) : sizeof (struct tcpiphdr); #else const size_t min_protoh = sizeof(struct tcpiphdr); #endif INP_WLOCK_ASSERT(tp->t_inpcb); if (mtuoffer != -1) { KASSERT(offer == -1, ("%s: conflict", __func__)); offer = mtuoffer - min_protoh; } /* Initialize. */ #ifdef INET6 if (isipv6) { maxmtu = tcp_maxmtu6(&inp->inp_inc, cap); tp->t_maxseg = V_tcp_v6mssdflt; } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { maxmtu = tcp_maxmtu(&inp->inp_inc, cap); tp->t_maxseg = V_tcp_mssdflt; } #endif /* * No route to sender, stay with default mss and return. */ if (maxmtu == 0) { /* * In case we return early we need to initialize metrics * to a defined state as tcp_hc_get() would do for us * if there was no cache hit. */ if (metricptr != NULL) bzero(metricptr, sizeof(struct hc_metrics_lite)); return; } /* What have we got? */ switch (offer) { case 0: /* * Offer == 0 means that there was no MSS on the SYN * segment, in this case we use tcp_mssdflt as * already assigned to t_maxseg above. */ offer = tp->t_maxseg; break; case -1: /* * Offer == -1 means that we didn't receive SYN yet. */ /* FALLTHROUGH */ default: /* * Prevent DoS attack with too small MSS. Round up * to at least minmss. */ offer = max(offer, V_tcp_minmss); } /* * rmx information is now retrieved from tcp_hostcache. */ tcp_hc_get(&inp->inp_inc, &metrics); if (metricptr != NULL) bcopy(&metrics, metricptr, sizeof(struct hc_metrics_lite)); /* * If there's a discovered mtu in tcp hostcache, use it. * Else, use the link mtu. */ if (metrics.rmx_mtu) mss = min(metrics.rmx_mtu, maxmtu) - min_protoh; else { #ifdef INET6 if (isipv6) { mss = maxmtu - min_protoh; if (!V_path_mtu_discovery && !in6_localaddr(&inp->in6p_faddr)) mss = min(mss, V_tcp_v6mssdflt); } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { mss = maxmtu - min_protoh; if (!V_path_mtu_discovery && !in_localaddr(inp->inp_faddr)) mss = min(mss, V_tcp_mssdflt); } #endif /* * XXX - The above conditional (mss = maxmtu - min_protoh) * probably violates the TCP spec. * The problem is that, since we don't know the * other end's MSS, we are supposed to use a conservative * default. But, if we do that, then MTU discovery will * never actually take place, because the conservative * default is much less than the MTUs typically seen * on the Internet today. For the moment, we'll sweep * this under the carpet. * * The conservative default might not actually be a problem * if the only case this occurs is when sending an initial * SYN with options and data to a host we've never talked * to before. Then, they will reply with an MSS value which * will get recorded and the new parameters should get * recomputed. For Further Study. */ } mss = min(mss, offer); /* * Sanity check: make sure that maxseg will be large * enough to allow some data on segments even if the * all the option space is used (40bytes). Otherwise * funny things may happen in tcp_output. * * XXXGL: shouldn't we reserve space for IP/IPv6 options? */ mss = max(mss, 64); tp->t_maxseg = mss; } void tcp_mss(struct tcpcb *tp, int offer) { int mss; uint32_t bufsize; struct inpcb *inp; struct socket *so; struct hc_metrics_lite metrics; struct tcp_ifcap cap; KASSERT(tp != NULL, ("%s: tp == NULL", __func__)); bzero(&cap, sizeof(cap)); tcp_mss_update(tp, offer, -1, &metrics, &cap); mss = tp->t_maxseg; inp = tp->t_inpcb; /* * If there's a pipesize, change the socket buffer to that size, * don't change if sb_hiwat is different than default (then it * has been changed on purpose with setsockopt). * Make the socket buffers an integral number of mss units; * if the mss is larger than the socket buffer, decrease the mss. */ so = inp->inp_socket; SOCKBUF_LOCK(&so->so_snd); if ((so->so_snd.sb_hiwat == V_tcp_sendspace) && metrics.rmx_sendpipe) bufsize = metrics.rmx_sendpipe; else bufsize = so->so_snd.sb_hiwat; if (bufsize < mss) mss = bufsize; else { bufsize = roundup(bufsize, mss); if (bufsize > sb_max) bufsize = sb_max; if (bufsize > so->so_snd.sb_hiwat) (void)sbreserve_locked(&so->so_snd, bufsize, so, NULL); } SOCKBUF_UNLOCK(&so->so_snd); /* * Sanity check: make sure that maxseg will be large * enough to allow some data on segments even if the * all the option space is used (40bytes). Otherwise * funny things may happen in tcp_output. * * XXXGL: shouldn't we reserve space for IP/IPv6 options? */ tp->t_maxseg = max(mss, 64); SOCKBUF_LOCK(&so->so_rcv); if ((so->so_rcv.sb_hiwat == V_tcp_recvspace) && metrics.rmx_recvpipe) bufsize = metrics.rmx_recvpipe; else bufsize = so->so_rcv.sb_hiwat; if (bufsize > mss) { bufsize = roundup(bufsize, mss); if (bufsize > sb_max) bufsize = sb_max; if (bufsize > so->so_rcv.sb_hiwat) (void)sbreserve_locked(&so->so_rcv, bufsize, so, NULL); } SOCKBUF_UNLOCK(&so->so_rcv); /* Check the interface for TSO capabilities. */ if (cap.ifcap & CSUM_TSO) { tp->t_flags |= TF_TSO; tp->t_tsomax = cap.tsomax; tp->t_tsomaxsegcount = cap.tsomaxsegcount; tp->t_tsomaxsegsize = cap.tsomaxsegsize; } } /* * Determine the MSS option to send on an outgoing SYN. */ int tcp_mssopt(struct in_conninfo *inc) { int mss = 0; uint32_t thcmtu = 0; uint32_t maxmtu = 0; size_t min_protoh; KASSERT(inc != NULL, ("tcp_mssopt with NULL in_conninfo pointer")); #ifdef INET6 if (inc->inc_flags & INC_ISIPV6) { mss = V_tcp_v6mssdflt; maxmtu = tcp_maxmtu6(inc, NULL); min_protoh = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { mss = V_tcp_mssdflt; maxmtu = tcp_maxmtu(inc, NULL); min_protoh = sizeof(struct tcpiphdr); } #endif #if defined(INET6) || defined(INET) thcmtu = tcp_hc_getmtu(inc); /* IPv4 and IPv6 */ #endif if (maxmtu && thcmtu) mss = min(maxmtu, thcmtu) - min_protoh; else if (maxmtu || thcmtu) mss = max(maxmtu, thcmtu) - min_protoh; return (mss); } /* * On a partial ack arrives, force the retransmission of the * next unacknowledged segment. Do not clear tp->t_dupacks. * By setting snd_nxt to ti_ack, this forces retransmission timer to * be started again. */ void tcp_newreno_partial_ack(struct tcpcb *tp, struct tcphdr *th) { tcp_seq onxt = tp->snd_nxt; uint32_t ocwnd = tp->snd_cwnd; u_int maxseg = tcp_maxseg(tp); INP_WLOCK_ASSERT(tp->t_inpcb); tcp_timer_activate(tp, TT_REXMT, 0); tp->t_rtttime = 0; tp->snd_nxt = th->th_ack; /* * Set snd_cwnd to one segment beyond acknowledged offset. * (tp->snd_una has not yet been updated when this function is called.) */ tp->snd_cwnd = maxseg + BYTES_THIS_ACK(tp, th); tp->t_flags |= TF_ACKNOW; (void) tp->t_fb->tfb_tcp_output(tp); tp->snd_cwnd = ocwnd; if (SEQ_GT(onxt, tp->snd_nxt)) tp->snd_nxt = onxt; /* * Partial window deflation. Relies on fact that tp->snd_una * not updated yet. */ if (tp->snd_cwnd > BYTES_THIS_ACK(tp, th)) tp->snd_cwnd -= BYTES_THIS_ACK(tp, th); else tp->snd_cwnd = 0; tp->snd_cwnd += maxseg; } int tcp_compute_pipe(struct tcpcb *tp) { return (tp->snd_max - tp->snd_una + tp->sackhint.sack_bytes_rexmit - tp->sackhint.sacked_bytes); } uint32_t tcp_compute_initwnd(uint32_t maxseg) { /* * Calculate the Initial Window, also used as Restart Window * * RFC5681 Section 3.1 specifies the default conservative values. * RFC3390 specifies slightly more aggressive values. * RFC6928 increases it to ten segments. * Support for user specified value for initial flight size. */ if (V_tcp_initcwnd_segments) return min(V_tcp_initcwnd_segments * maxseg, max(2 * maxseg, V_tcp_initcwnd_segments * 1460)); else if (V_tcp_do_rfc3390) return min(4 * maxseg, max(2 * maxseg, 4380)); else { /* Per RFC5681 Section 3.1 */ if (maxseg > 2190) return (2 * maxseg); else if (maxseg > 1095) return (3 * maxseg); else return (4 * maxseg); } } Index: head/sys/netinet/tcp_stacks/bbr.c =================================================================== --- head/sys/netinet/tcp_stacks/bbr.c (revision 356526) +++ head/sys/netinet/tcp_stacks/bbr.c (revision 356527) @@ -1,15197 +1,15197 @@ /*- * Copyright (c) 2016-9 * Netflix Inc. * 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. * * 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 "opt_kern_tls.h" #include #include #include #include #ifdef TCP_HHOOK #include #endif #include #include #include #include #include #ifdef KERN_TLS #include #endif #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 #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 * measurments 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_segement * 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 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 (inp->inp_in_hpts) { /* 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 ((tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && + 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 * possiblity of nothing outstanding and the RXT up * (and the hptsi timer). */ return; - } else if (((tcp_always_keepalive || + } 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 (inp->inp_in_hpts) tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT); 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 differnt 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, 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, 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, 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 measurment 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, 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, 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, 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 measurment 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, 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, "Maxiumum 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, 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-postive 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 inline int32_t bbr_progress_timeout_check(struct tcp_bbr *bbr) { if (bbr->rc_tp->t_maxunacktime && bbr->rc_tp->t_acktime && TSTMP_GT(ticks, bbr->rc_tp->t_acktime)) { if ((((uint32_t)ticks - bbr->rc_tp->t_acktime)) >= bbr->rc_tp->t_maxunacktime) { /* * There is an assumption here that the caller will * drop the connection, so we increment the * statistics. */ bbr_log_progress_event(bbr, bbr->rc_tp, ticks, PROGRESS_DROP, __LINE__); BBR_STAT_INC(bbr_progress_drops); #ifdef NETFLIX_STATS TCPSTAT_INC(tcps_progdrops); #endif return (1); } } return (0); } 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_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 = bbr->rc_inp->inp_in_hpts; l->ininput = bbr->rc_inp->inp_in_input; 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) { tv.tv_sec = m->m_pkthdr.rcv_tstmp / 1000000000; tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000) / 1000; 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_runningtick; log.u_bbr.bw_inuse = diag->wheel_tick; log.u_bbr.rttProp = diag->wheel_cts; log.u_bbr.delRate = diag->maxticks; 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 measurment. 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); } /* * RRS: Copied from user space! * Calculate a uniformly distributed random number less than upper_bound * avoiding "modulo bias". * * Uniformity is achieved by generating new random numbers until the one * returned is outside the range [0, 2**32 % upper_bound). This * guarantees the selected random number will be inside * [2**32 % upper_bound, 2**32) which maps back to [0, upper_bound) * after reduction modulo upper_bound. */ static uint32_t arc4random_uniform(uint32_t upper_bound) { uint32_t r, min; if (upper_bound < 2) return 0; /* 2**32 % x == (2**32 - x) % x */ min = -upper_bound % upper_bound; /* * This could theoretically loop forever but each retry has * p > 0.5 (worst case, usually far better) of selecting a * number inside the range we need, so it should rarely need * to re-roll. */ for (;;) { r = arc4random(); if (r >= min) break; } return r % upper_bound; } 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-postive 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 normaly * 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 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); 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 (bbr->rc_inp->inp_in_hpts && ((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: 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) && \ ((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_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 postion * 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 (bbr_progress_timeout_check(bbr)) { tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT); return (1); } /* 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; #ifdef KERN_TLS if (bbr->rc_inp->inp_socket->so_snd.sb_flags & SB_TLS_IFNET) { /* * For hardware TLS we do *not* want to send * new data. */ goto need_retran; } #endif /* * 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; TCPSTAT_INC(tcps_delack); bbr->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK; return (0); } /* * Persists timer, here we simply need to setup the * FORCE-DATA flag the output routine will send * the one byte send. * * 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; TCPSTAT_INC(tcps_persisttimeo); /* * Have we exceeded the user specified progress time? */ if (bbr_progress_timeout_check(bbr)) { tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT); goto out; } /* * 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)) { TCPSTAT_INC(tcps_persistdrop); tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT); goto out; } 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) { TCPSTAT_INC(tcps_persistdrop); tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT); goto out; } 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. */ TCPSTAT_INC(tcps_keeptimeo); if (tp->t_state < TCPS_ESTABLISHED) goto dropit; - if ((tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && + 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) { 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: TCPSTAT_INC(tcps_keepdrops); tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT); return (1); } /* * 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 ((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; 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 (bbr_progress_timeout_check(bbr)) { retval = 1; tcp_set_inp_to_drop(bbr->rc_inp, ETIMEDOUT); goto out; } 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; TCPSTAT_INC(tcps_timeoutdrop); retval = 1; tcp_set_inp_to_drop(bbr->rc_inp, (tp->t_softerror ? (uint16_t) tp->t_softerror : ETIMEDOUT)); goto out; } 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; } 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. */ if (V_tcp_pmtud_blackhole_detect && (((tp->t_state == TCPS_ESTABLISHED)) || (tp->t_state == TCPS_FIN_WAIT_1))) { #ifdef INET6 int32_t isipv6; #endif /* * 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; 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) { /* 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 } 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; 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; out: 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 (bbr->rc_inp->inp_in_hpts && (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, HPTS_REMOVE_OUTPUT); 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_rt == NULL) || (bbr->rc_inp->inp_route.ro_rt->rt_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_rt->rt_ifp, rate, (RS_PACING_GEQ|RS_PACING_SUB_OK), &error); 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) * Also note that Hardware-TLS is special and does alternate * things to minimize PCI Bus Bandwidth use. */ 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; } #ifdef KERN_TLS if (bbr->rc_inp->inp_socket->so_snd.sb_flags & SB_TLS_IFNET) { tls_seg = ctf_get_opt_tls_size(bbr->rc_inp->inp_socket, bbr->rc_tp->snd_wnd); bbr->r_ctl.rc_pace_min_segs = (tls_seg + bbr->rc_last_options); } #endif 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 unless we are hardware TLS. Set up * the inital 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; #ifdef KERN_TLS if ((bbr->rc_inp->inp_socket->so_snd.sb_flags & SB_TLS_IFNET) && tls_seg) { /* * For hardware TLS we set our min to the tls_seg size. */ bbr->r_ctl.rc_pace_max_segs = tls_seg; bbr->r_ctl.rc_pace_min_segs = tls_seg + bbr->rc_last_options; } #endif 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); #ifdef KERN_TLS if ((bbr->rc_inp->inp_socket->so_snd.sb_flags & SB_TLS_IFNET) == 0) #endif 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 we are not HW-TLS */ #ifdef KERN_TLS if ((bbr->rc_inp->inp_socket->so_snd.sb_flags & SB_TLS_IFNET) == 0) #endif 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; } #ifdef KERN_TLS if (tls_seg) { /* * Lets move the output size * up to 1 or more TLS record sizes. */ uint32_t temp; temp = roundup(new_tso, tls_seg); new_tso = temp; /* Back down if needed to under a full frame */ while (new_tso > PACE_MAX_IP_BYTES) new_tso -= tls_seg; } #endif 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; } #ifdef KERN_TLS if ((bbr->rc_inp->inp_socket->so_snd.sb_flags & SB_TLS_IFNET) && tls_seg) { bbr->r_ctl.rc_pace_min_segs = tls_seg + bbr->rc_last_options; } else #endif /* We have hardware pacing and not hardware TLS! */ 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, uint8_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) 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; } 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_earlier_retran(struct tcpcb *tp, struct tcp_bbr *bbr, struct bbr_sendmap *rsm, uint32_t t, uint32_t cts, int ack_type) { /* * For this RSM, we acknowledged the data from a previous * transmission, not the last one we made. This means we did a false * retransmit. */ if (rsm->r_flags & BBR_HAS_FIN) { /* * The sending of the FIN often is multiple sent when we * have everything outstanding ack'd. We ignore this case * since its over now. */ return; } if (rsm->r_flags & BBR_TLP) { /* * We expect TLP's to have this occur often */ bbr->rc_tlp_rtx_out = 0; return; } if (ack_type != BBR_CUM_ACKED) { /* * If it was not a cum-ack we * don't really know for sure since * the timestamp could be from some * other transmission. */ return; } if (rsm->r_flags & BBR_WAS_SACKPASS) { /* * We retransmitted based on a sack and the earlier * retransmission ack'd it - re-ordering is occuring. */ BBR_STAT_INC(bbr_reorder_seen); bbr->r_ctl.rc_reorder_ts = cts; } /* Back down the loss count */ 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; rsm->r_flags &= ~BBR_MARKED_LOST; 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; } /***** RRS HERE ************************/ /* Do we need to do this??? */ /* bbr_reset_lt_bw_sampling(bbr, cts); */ /***** RRS HERE ************************/ BBR_STAT_INC(bbr_badfr); BBR_STAT_ADD(bbr_badfr_bytes, (rsm->r_end - rsm->r_start)); } 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 measurment */ 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 upper, lower, 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; } upper = (bw >> 32) & 0x00000000ffffffff; lower = bw & 0x00000000ffffffff; /* * 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 upper, lower, 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; } upper = (bw >> 32) & 0x00000000ffffffff; lower = bw & 0x00000000ffffffff; /* * 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_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 */ bbr_earlier_retran(tp, bbr, rsm, t, cts, ack_type); } 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); bbr_earlier_retran(tp, bbr, rsm, t, cts, ack_type); } 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, maxseg, changed = 0; struct bbr_sendmap *rsm, *nrsm; int32_t used_ref = 1; uint8_t went_back = 0, went_fwd = 0; maxseg = tp->t_maxseg - bbr->rc_last_options; 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 (th->th_flags & 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++; #ifdef NETFLIX_STATS } else if (SEQ_LEQ(sack.start, th_ack) && SEQ_LEQ(sack.end, th_ack)) { /* * Its a D-SACK block. */ tcp_record_dsack(sack.start, sack.end); #endif } } 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); TCPSTAT_ADD(tcps_rcvackpack, (int)nsegs); 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; 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 { 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; tp->t_flags &= ~TF_FORCEDATA; 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 (bbr->rc_inp->inp_in_hpts) { tcp_hpts_remove(bbr->rc_inp, HPTS_REMOVE_OUTPUT); 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) 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); } /* * Process segments with URG. */ if ((thflags & TH_URG) && th->th_urp && TCPS_HAVERCVDFIN(tp->t_state) == 0) { /* * This is a kludge, but if we receive and accept random * urgent pointers, we'll crash in soreceive. It's hard to * imagine someone actually wanting to send this much urgent * data. */ SOCKBUF_LOCK(&so->so_rcv); if (th->th_urp + sbavail(&so->so_rcv) > sb_max) { th->th_urp = 0; /* XXX */ thflags &= ~TH_URG; /* XXX */ SOCKBUF_UNLOCK(&so->so_rcv); /* XXX */ goto dodata; /* XXX */ } /* * If this segment advances the known urgent pointer, then * mark the data stream. This should not happen in * CLOSE_WAIT, CLOSING, LAST_ACK or TIME_WAIT STATES since a * FIN has been received from the remote side. In these * states we ignore the URG. * * 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 as the original * spec states (in one of two places). */ if (SEQ_GT(th->th_seq + th->th_urp, tp->rcv_up)) { tp->rcv_up = th->th_seq + th->th_urp; so->so_oobmark = sbavail(&so->so_rcv) + (tp->rcv_up - tp->rcv_nxt) - 1; if (so->so_oobmark == 0) so->so_rcv.sb_state |= SBS_RCVATMARK; sohasoutofband(so); tp->t_oobflags &= ~(TCPOOB_HAVEDATA | TCPOOB_HADDATA); } SOCKBUF_UNLOCK(&so->so_rcv); /* * Remove out of band data so doesn't get presented to user. * This can happen independent of advancing the URG pointer, * but if two URG's are pending at once, some out-of-band * data may creep in... ick. */ if (th->th_urp <= (uint32_t)tlen && !(so->so_options & SO_OOBINLINE)) { /* hdr drop is delayed */ tcp_pulloutofband(so, th, m, drop_hdrlen); } } else { /* * If no out of band data is expected, pull receive urgent * pointer along with the receive window. */ if (SEQ_GT(tp->rcv_nxt, tp->rcv_up)) tp->rcv_up = tp->rcv_nxt; } dodata: /* XXX */ 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) && 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; thflags = th->th_flags & TH_FIN; TCPSTAT_ADD(tcps_rcvpack, (int)nsegs); 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_SACK_PERMIT) && (save_tlen > 0)) { 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) { 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); TCPSTAT_INC(tcps_preddat); tp->rcv_nxt += tlen; /* * 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; TCPSTAT_ADD(tcps_rcvpack, (int)nsegs); 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, so, 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) { 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. */ 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 TCPSTAT_ADD(tcps_rcvackpack, (int)nsegs); 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) { 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))) { 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; 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 == 0)) { 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); 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; TCPSTAT_INC(tcps_rcvpackafterwin); 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) { 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))) { 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)) { 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)) { 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)); } TCPSTAT_INC(tcps_connects); 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; /* * Account for the ACK of our SYN prior to regular * ACK processing below. */ tp->snd_una++; } /* * 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); } /* * 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); 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) { 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)) { 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 (bbr_progress_timeout_check(bbr)) { 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) { 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 (bbr_progress_timeout_check(bbr)) { 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: tp = tcp_close(tp); 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) { 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. */ if ((so->so_state & SS_NOFDREF) && tlen) { /* * We call a new function now so we might continue and setup * to reset at all data being ack'd. */ if (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 (bbr_progress_timeout_check(bbr)) { 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) { 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. */ if ((so->so_state & SS_NOFDREF) && tlen) { /* * We call a new function now so we might continue and setup * to reset at all data being ack'd. */ if (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 (bbr_progress_timeout_check(bbr)) { 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) { 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. */ if ((so->so_state & SS_NOFDREF) && tlen) { /* * We call a new function now so we might continue and setup * to reset at all data being ack'd. */ if (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 (bbr_progress_timeout_check(bbr)) { 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) { 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. */ if ((so->so_state & SS_NOFDREF) && tlen) { /* * We call a new function now so we might continue and setup * to reset at all data being ack'd. */ if (bbr_check_data_after_close(m, bbr, tp, &tlen, th, so)) return (1); } INP_WLOCK_ASSERT(tp->t_inpcb); /* * 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 (bbr_progress_timeout_check(bbr)) { 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 recusive (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_flags = BBR_OVERMAX; 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_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); 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_gain; old_state = bbr_state_val(bbr); old_gain = bbr->r_ctl.rc_bbr_hptsi_gain; 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 * occured. * 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 measurment 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 occured 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 occured 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; int32_t in_recovery; 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 = th->th_flags; /* * 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__)); KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT", __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 /* * 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 (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); } /* * 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; } /* * 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_RCVD_TSTMP; tp->ts_recent = to.to_tsval; tp->ts_recent_age = tcp_tv_to_mssectick(&bbr->rc_tv); } 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 (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))) { ctf_do_dropwithreset_conn(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return (1); } in_recovery = IN_RECOVERY(tp->t_flags); 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); #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; (void)tp->t_fb->tfb_tcp_output(tp); } 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) || - ((tcp_always_keepalive || bbr->rc_inp->inp_socket->so_options & SO_KEEPALIVE) && + ((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) && (bbr->rc_inp->inp_in_hpts) && (bbr->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { /* * keep alive not needed if we are hptsi * output yet */ ; } else { if (bbr->rc_inp->inp_in_hpts) { tcp_hpts_remove(bbr->rc_inp, HPTS_REMOVE_OUTPUT); 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); (void)tp->t_fb->tfb_tcp_output(tp); } } } 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_log_type_hrdwtso(struct tcpcb *tp, struct tcp_bbr *bbr, int len, int mod, int what_we_can_send) { if (tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; uint32_t cts; cts = tcp_get_usecs(&tv); bbr_fill_in_logging_data(bbr, &log.u_bbr, cts); log.u_bbr.flex1 = bbr->r_ctl.rc_pace_min_segs; log.u_bbr.flex2 = what_we_can_send; log.u_bbr.flex3 = bbr->r_ctl.rc_pace_max_segs; log.u_bbr.flex4 = len; log.u_bbr.flex5 = 0; log.u_bbr.flex7 = mod; log.u_bbr.flex8 = 1; TCP_LOG_EVENTP(tp, NULL, &tp->t_inpcb->inp_socket->so_rcv, &tp->t_inpcb->inp_socket->so_snd, TCP_HDWR_TLS, 0, 0, &log, false, &tv); } } 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) { tv.tv_sec = m->m_pkthdr.rcv_tstmp /1000000000; tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000)/1000; } 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 TCPSTAT_INC(tcps_sndpack_error); 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 ((tp->t_flags & TF_FORCEDATA) && len == 1) { /* Window probe */ TCPSTAT_INC(tcps_sndprobe); #ifdef STATS stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB, len); #endif } else if (rsm) { if (rsm->r_flags & BBR_TLP) { /* * TLP should not count in retran count, but in its * own bin */ #ifdef NETFLIX_STATS tp->t_sndtlppack++; tp->t_sndtlpbyte += len; TCPSTAT_INC(tcps_tlpresends); TCPSTAT_ADD(tcps_tlpresend_bytes, len); #endif } else { /* Retransmit */ tp->t_sndrexmitpack++; TCPSTAT_INC(tcps_sndrexmitpack); 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 */ TCPSTAT_INC(tcps_sndpack); 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. */ uint32_t adv; int32_t oldwin; adv = min(recwin, TCP_MAXWIN << tp->rcv_scale); if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { oldwin = (tp->rcv_adv - tp->rcv_nxt); adv -= oldwin; } else oldwin = 0; /* * If the new window size ends up being the same as 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; #ifdef NETFLIX_TCPOUDP struct udphdr *udp = NULL; #endif u_char opt[TCP_MAXOLEN]; unsigned ipoptlen, optlen, hdrlen; #ifdef NETFLIX_TCPOUDP unsigned ulen; #endif 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 what_we_can = 0; uint32_t tot_len = 0; uint32_t rtr_cnt = 0; uint32_t maxseg, pace_max_segs, p_maxseg; int32_t csum_flags; 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; int force_tso = 0; 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; #ifdef KERN_TLS if (sb->sb_flags & SB_TLS_IFNET) hw_tls = 1; else #endif 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) && inp->inp_in_hpts) { /* * 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 = min(max(sbspace(&so->so_rcv), 0), TCP_MAXWIN << tp->rcv_scale); if ((bbr_window_update_needed(tp, so, recwin, maxseg) == 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, HPTS_REMOVE_OUTPUT); 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_flags & TF_FORCEDATA) || (tp->t_state < TCPS_ESTABLISHED)) { /* Timeouts or early states are exempt */ if (inp->inp_in_hpts) tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT); } else if (inp->inp_in_hpts) { 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, HPTS_REMOVE_OUTPUT); } else if (tp->t_state == TCPS_CLOSED) { bbr->r_ctl.rc_last_delay_val = 0; tcp_hpts_remove(inp, HPTS_REMOVE_OUTPUT); } 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) { if (bbr_process_timers(tp, bbr, cts, hpts_calling)) { counter_u64_add(bbr_out_size[TCP_MSS_ACCT_ATIMER], 1); return (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) && /* inital SYN or SYN|ACK sent */ (tp->t_rxtshift == 0)) { /* not a retransmit */ return (0); } /* * 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 } rtr_cnt++; 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) && #ifdef KERN_TLS ((sb->sb_flags & SB_TLS_IFNET) == 0) && #endif (len > maxseg)) { len = maxseg; more_to_rxt = 1; } sb_offset = rsm->r_start - tp->snd_una; if (len > 0) { sack_rxmit = 1; TCPSTAT_INC(tcps_sack_rexmits); 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; rtr_cnt++; 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 in persist timeout with window of 0, send 1 byte. Otherwise, * if window is small but nonzero and time TF_SENTFIN expired, we * will send what we can and go to transmit state. */ if (tp->t_flags & TF_FORCEDATA) { if ((sendwin == 0) || (sendwin <= (tp->snd_max - tp->snd_una))) { /* * If we still have some data to send, then clear * the FIN bit. Usually this would happen below * when it realizes that we aren't sending all the * data. However, if we have exactly 1 byte of * unsent data, then it won't clear the FIN bit * below, and if we are in persist state, we wind up * sending the packet without recording that we sent * the FIN bit. * * We can't just blindly clear the FIN bit, because * if we don't have any more data to send then the * probe will be the FIN itself. */ if (sb_offset < sbused(sb)) flags &= ~TH_FIN; sendwin = 1; } else { if ((bbr->rc_in_persist != 0) && (tp->snd_wnd >= min((bbr->r_ctl.rc_high_rwnd/2), bbr_minseg(bbr)))) { /* Exit persists if there is space */ bbr_exit_persist(tp, bbr, cts, __LINE__); } if (rsm == NULL) { /* * If we are dropping persist mode then we * need to correct sb_offset if not a * retransmit. */ sb_offset = tp->snd_max - tp->snd_una; } } } /* * 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 { what_we_can = 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 ((tp->t_flags & TF_FORCEDATA) && (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 = min(max(sbspace(&so->so_rcv), 0), 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 (tp->t_flags & TF_FORCEDATA) { /* typ. timeout case */ 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, or urgent data. ACKNOW * is also a catch-all for the retransmit timer timeout case. */ if (tp->t_flags & TF_ACKNOW) { goto send; } if (((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0)) { goto send; } if (SEQ_GT(tp->snd_up, tp->snd_una)) { 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); tp->t_flags &= ~TF_FORCEDATA; /* 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); #ifdef NETFLIX_TCPOUDP if (tp->t_port) to.to_mss -= V_tcp_udp_tunneling_overhead; #endif 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; } #ifdef NETFLIX_TCPOUDP 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); } #endif #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. */ #ifdef KERN_TLS /* force TSO for so TLS offload can get mss */ if (sb->sb_flags & SB_TLS_IFNET) { force_tso = 1; } #endif 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)) && (hw_tls == 0)) { 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; uint32_t orig_len; /* * 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); tcp_set_inp_to_drop(inp, EFAULT); return (0); } 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 orig_len = len; m->m_next = tcp_m_copym( #ifdef NETFLIX_COPY_ARGS tp, #endif 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 && !force_tso) { /* * 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) TCPSTAT_INC(tcps_sndacks); else if (flags & (TH_SYN | TH_FIN | TH_RST)) TCPSTAT_INC(tcps_sndctrl); else if (SEQ_GT(tp->snd_up, tp->snd_una)) TCPSTAT_INC(tcps_sndurg); else 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 *); #ifdef NETFLIX_TCPOUDP if (tp->t_port) { udp = (struct udphdr *)((caddr_t)ip6 + ipoptlen + 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 { #endif th = (struct tcphdr *)(ip6 + 1); #ifdef NETFLIX_TCPOUDP } #endif tcpip_fillheaders(inp, #ifdef NETFLIX_TCPOUDP tp->t_port, #endif ip6, th); } else #endif /* INET6 */ { ip = mtod(m, struct ip *); #ifdef TCPDEBUG ipov = (struct ipovly *)ip; #endif #ifdef NETFLIX_TCPOUDP if (tp->t_port) { udp = (struct udphdr *)((caddr_t)ip + ipoptlen + 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 #endif th = (struct tcphdr *)(ip + 1); tcpip_fillheaders(inp, #ifdef NETFLIX_TCPOUDP tp->t_port, #endif 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 if (flags & TH_RST) { /* * For a Reset send the last cum ack in sequence * (this like any other choice may still generate a * challenge ack, if a ack-update packet is in * flight). */ th->th_seq = htonl(tp->snd_una); bbr_seq = tp->snd_una; } 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; } th->th_flags = 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 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; if (SEQ_GT(tp->snd_up, tp->snd_max)) { th->th_urp = htons((u_short)(tp->snd_up - tp->snd_max)); th->th_flags |= TH_URG; } else /* * If no urgent pointer to send, then we pull the urgent * pointer to the left edge of the send window so that it * doesn't drift into the send window on sequence number * wraparound. */ tp->snd_up = tp->snd_una; /* drag it along */ #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. */ #ifdef NETFLIX_TCPOUDP 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 { #endif 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); #ifdef NETFLIX_TCPOUDP } #endif } #endif #if defined(INET6) && defined(INET) else #endif #ifdef INET { #ifdef NETFLIX_TCPOUDP 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 { #endif 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)); #ifdef NETFLIX_TCPOUDP } #endif /* 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 || force_tso) { KASSERT(force_tso || 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_rt != NULL) mtu = inp->inp_route6.ro_rt->rt_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_rt != NULL) mtu = inp->inp_route.ro_rt->rt_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) { 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); else if (hw_tls) { if (filled_all || (len >= bbr->r_ctl.rc_pace_max_segs)) BBR_STAT_INC(bbr_meets_tso_thresh); else { if (doing_tlp) { BBR_STAT_INC(bbr_miss_tlp); bbr_log_type_hrdwtso(tp, bbr, len, 1, what_we_can); } else if (rsm) { BBR_STAT_INC(bbr_miss_retran); bbr_log_type_hrdwtso(tp, bbr, len, 2, what_we_can); } else if ((ctf_outstanding(tp) + bbr->r_ctl.rc_pace_max_segs) > sbavail(sb)) { BBR_STAT_INC(bbr_miss_tso_app); bbr_log_type_hrdwtso(tp, bbr, len, 3, what_we_can); } else if ((ctf_flight_size(tp, (bbr->r_ctl.rc_sacked + bbr->r_ctl.rc_lost_bytes)) + bbr->r_ctl.rc_pace_max_segs) > tp->snd_cwnd) { BBR_STAT_INC(bbr_miss_tso_cwnd); bbr_log_type_hrdwtso(tp, bbr, len, 4, what_we_can); } else if ((ctf_outstanding(tp) + bbr->r_ctl.rc_pace_max_segs) > tp->snd_wnd) { BBR_STAT_INC(bbr_miss_tso_rwnd); bbr_log_type_hrdwtso(tp, bbr, len, 5, what_we_can); } else { BBR_STAT_INC(bbr_miss_unknown); bbr_log_type_hrdwtso(tp, bbr, len, 6, what_we_can); } } } /* 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 (((tp->t_flags & TF_FORCEDATA) == 0) || (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) { tp->snd_max++; } 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 */ tcp_set_inp_to_drop(inp, 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); } tp->t_flags &= ~TF_FORCEDATA; /* * 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: tp->t_flags &= ~TF_FORCEDATA; 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 } 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_rt && inp->inp_route.ro_rt->rt_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_rt->rt_ifp, rate_wanted, (RS_PACING_GEQ|RS_PACING_SUB_OK), &err); 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_rt->rt_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_rt == NULL) || (inp->inp_route.ro_rt->rt_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) && (hw_tls == 0) && (len > 0) && ((flags & TH_RST) == 0) && (IN_RECOVERY(tp->t_flags) == 0) && (bbr->rc_in_persist == 0) && ((tp->t_flags & TF_FORCEDATA) == 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 | TF_FORCEDATA); goto again; } skip_again: 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 | TF_FORCEDATA); 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; struct tcp_bbr *bbr; bbr = (struct tcp_bbr *)tp->t_fb_ptr; 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; } } /* * 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 socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp, struct tcp_bbr *bbr) { int32_t error = 0, optval; 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(so, sopt, inp, tp)); 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)) { INP_WUNLOCK(inp); return (ECONNRESET); } tp = intotcpcb(inp); 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->bbr_hdrw_pacing) { bbr->bbr_hdrw_pacing = 0; in_pcbdetach_txrtlmt(bbr->rc_inp); } #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; bbr_output(tp); } } 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(so, sopt, inp, tp)); 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 socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp, struct tcp_bbr *bbr) { int32_t error, optval; /* * 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(so, sopt, inp, tp)); 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 socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp) { int32_t error = EINVAL; struct tcp_bbr *bbr; bbr = (struct tcp_bbr *)tp->t_fb_ptr; if (bbr == NULL) { /* Huh? */ goto out; } if (sopt->sopt_dir == SOPT_SET) { return (bbr_set_sockopt(so, sopt, inp, tp, bbr)); } else if (sopt->sopt_dir == SOPT_GET) { return (bbr_get_sockopt(so, sopt, inp, tp, bbr)); } out: INP_WUNLOCK(inp); return (error); } 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 }; 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, 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); Index: head/sys/netinet/tcp_stacks/rack.c =================================================================== --- head/sys/netinet/tcp_stacks/rack.c (revision 356526) +++ head/sys/netinet/tcp_stacks/rack.c (revision 356527) @@ -1,10561 +1,10561 @@ /*- * Copyright (c) 2016-9 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 #ifdef KERN_TLS #include #endif #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 #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 #ifdef TCPDEBUG #include #endif /* TCPDEBUG */ #ifdef TCP_OFFLOAD #include #endif #ifdef INET6 #include #endif #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 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_segement * 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_reorder_thresh = 2; static int32_t rack_reorder_fade = 60000; /* 0 - never fade, def 60,000 * - 60 seconds */ /* 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_pkt_delay = 1; static int32_t rack_min_pace_time = 0; static int32_t rack_early_recovery = 1; static int32_t rack_send_a_lot_in_prr = 1; static int32_t rack_min_to = 1; /* Number of ms minimum timeout */ static int32_t rack_verbose_logging = 0; static int32_t rack_ignore_data_after_close = 1; static int32_t use_rack_cheat = 1; static int32_t rack_persist_min = 250; /* 250ms */ static int32_t rack_persist_max = 1000; /* 1 Second */ static int32_t rack_sack_not_required = 0; /* set to one to allow non-sack to use rack */ static int32_t rack_hw_tls_max_seg = 0; /* 0 means use hw-tls single segment */ /* * 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 rack_tlp_min = 10; static int32_t rack_rto_min = 30; /* 30ms same as main freebsd */ static int32_t rack_rto_max = 4000; /* 4 seconds */ 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 = 200; /* 200ms */ static int32_t rack_slot_reduction = 4; static int32_t rack_lower_cwnd_at_tlp = 0; static int32_t rack_use_proportional_reduce = 0; static int32_t rack_proportional_rate = 10; static int32_t rack_tlp_max_resend = 2; static int32_t rack_limited_retran = 0; static int32_t rack_always_send_oldest = 0; static int32_t rack_use_sack_filter = 1; static int32_t rack_tlp_threshold_use = TLP_USE_TWO_ONE; static int32_t rack_per_of_gp = 50; /* Rack specific counters */ counter_u64_t rack_badfr; counter_u64_t rack_badfr_bytes; counter_u64_t rack_rtm_prr_retran; counter_u64_t rack_rtm_prr_newdata; counter_u64_t rack_timestamp_mismatch; counter_u64_t rack_reorder_seen; counter_u64_t rack_paced_segments; counter_u64_t rack_unpaced_segments; counter_u64_t rack_calc_zero; counter_u64_t rack_calc_nonzero; counter_u64_t rack_saw_enobuf; counter_u64_t rack_saw_enetunreach; counter_u64_t rack_per_timer_hole; /* 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_tlp_retran_fail; counter_u64_t rack_to_tot; counter_u64_t rack_to_arm_rack; counter_u64_t rack_to_arm_tlp; 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_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_used_tlpmethod; counter_u64_t rack_used_tlpmethod2; counter_u64_t rack_enter_tlp_calc; counter_u64_t rack_input_idle_reduces; counter_u64_t rack_collapsed_win; counter_u64_t rack_tlp_does_nada; /* Counters for HW TLS */ counter_u64_t rack_tls_rwnd; counter_u64_t rack_tls_cwnd; counter_u64_t rack_tls_app; counter_u64_t rack_tls_other; counter_u64_t rack_tls_filled; counter_u64_t rack_tls_rxt; counter_u64_t rack_tls_tlp; /* Temp CPU counters */ counter_u64_t rack_find_high; counter_u64_t rack_progress_drops; counter_u64_t rack_out_size[TCP_MSS_ACCT_SIZE]; counter_u64_t rack_opts_arry[RACK_OPTS_SIZE]; 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, struct tcphdr *th, 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, struct tcphdr *th, uint32_t type); static void rack_counter_destroy(void); static int rack_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp); static int32_t rack_ctor(void *mem, int32_t size, void *arg, int32_t how); 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_earlier_retran(struct tcpcb *tp, struct rack_sendmap *rsm, uint32_t t, uint32_t cts); 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 socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack); 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); static void rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len, uint32_t seq_out, uint8_t th_flags, int32_t err, uint32_t ts, uint8_t pass, struct rack_sendmap *hintrsm); 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, int num); 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, struct tcphdr *th); static void rack_remxt_tmr(struct tcpcb *tp); static int rack_set_sockopt(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack); 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, uint32_t ts, int32_t * lenp); static void rack_update_rsm(struct tcpcb *tp, struct tcp_rack *rack, struct rack_sendmap *rsm, uint32_t ts); 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); 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); static void tcp_rack_partialack(struct tcpcb *tp, struct tcphdr *th); int32_t rack_clear_counter=0; 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_badfr); counter_u64_zero(rack_badfr_bytes); counter_u64_zero(rack_rtm_prr_retran); counter_u64_zero(rack_rtm_prr_newdata); counter_u64_zero(rack_timestamp_mismatch); counter_u64_zero(rack_reorder_seen); 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_tlp_retran_fail); counter_u64_zero(rack_to_tot); counter_u64_zero(rack_to_arm_rack); counter_u64_zero(rack_to_arm_tlp); counter_u64_zero(rack_paced_segments); counter_u64_zero(rack_calc_zero); counter_u64_zero(rack_calc_nonzero); counter_u64_zero(rack_unpaced_segments); counter_u64_zero(rack_saw_enobuf); counter_u64_zero(rack_saw_enetunreach); counter_u64_zero(rack_per_timer_hole); 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_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_find_high); counter_u64_zero(rack_tls_rwnd); counter_u64_zero(rack_tls_cwnd); counter_u64_zero(rack_tls_app); counter_u64_zero(rack_tls_other); counter_u64_zero(rack_tls_filled); counter_u64_zero(rack_tls_rxt); counter_u64_zero(rack_tls_tlp); 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_used_tlpmethod); counter_u64_zero(rack_used_tlpmethod2); counter_u64_zero(rack_enter_tlp_calc); counter_u64_zero(rack_progress_drops); counter_u64_zero(rack_tlp_does_nada); counter_u64_zero(rack_collapsed_win); } rack_clear_counter = 0; return (0); } static void rack_init_sysctls(void) { struct sysctl_oid *rack_counters; struct sysctl_oid *rack_attack; 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 "); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "hw_tlsmax", CTLFLAG_RW, &rack_hw_tls_max_seg , 0, "Do we have a multplier of TLS records we can send as a max (0=1 TLS record)? "); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "data_after_close", CTLFLAG_RW, &rack_ignore_data_after_close, 0, "Do we hold off sending a RST until all pending data is ack'd"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "cheat_rxt", CTLFLAG_RW, &use_rack_cheat, 1, "Do we use the rxt cheat for rack?"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "persmin", CTLFLAG_RW, &rack_persist_min, 250, "What is the minimum time in milliseconds between persists"); SYSCTL_ADD_U32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "persmax", CTLFLAG_RW, &rack_persist_max, 1000, "What is the largest delay in milliseconds between persists"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "no_sack_needed", CTLFLAG_RW, &rack_sack_not_required, 0, "Do we allow rack to run on connections not supporting SACK?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), 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_sysctl_root), OID_AUTO, "gp_percentage", CTLFLAG_RW, &rack_per_of_gp, 50, "Do we pace to percentage of goodput (0=old method)?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "min_pace_time", CTLFLAG_RW, &rack_min_pace_time, 0, "Should we enforce a minimum pace time of 1ms"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), 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_sysctl_root), OID_AUTO, "sackfiltering", CTLFLAG_RW, &rack_use_sack_filter, 1, "Do we use sack filtering?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "delayed_ack", CTLFLAG_RW, &rack_delayed_ack_time, 200, "Delayed ack time (200ms)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "tlpminto", CTLFLAG_RW, &rack_tlp_min, 10, "TLP minimum timeout per the specification (10ms)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "send_oldest", CTLFLAG_RW, &rack_always_send_oldest, 1, "Should we always send the oldest TLP and RACK-TLP"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), 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_sysctl_root), OID_AUTO, "minrto", CTLFLAG_RW, &rack_rto_min, 0, "Minimum RTO in ms -- set with caution below 1000 due to TLP"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "maxrto", CTLFLAG_RW, &rack_rto_max, 0, "Maxiumum RTO in ms -- should be at least as large as min_rto"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "tlp_retry", CTLFLAG_RW, &rack_tlp_max_resend, 2, "How many times does TLP retry a single segment or multiple with no ACK"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "recovery_loss_prop", CTLFLAG_RW, &rack_use_proportional_reduce, 0, "Should we proportionaly reduce cwnd based on the number of losses "); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "recovery_prop", CTLFLAG_RW, &rack_proportional_rate, 10, "What percent reduction per loss"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), 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_sysctl_root), OID_AUTO, "hptsi_reduces", CTLFLAG_RW, &rack_slot_reduction, 4, "When setting a slot should we reduce by divisor"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "hptsi_every_seg", CTLFLAG_RW, &rack_pace_every_seg, 0, "Should we use the original pacing mechanism that did not pace much?"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "hptsi_seg_max", CTLFLAG_RW, &rack_hptsi_segments, 40, "Should we pace out only a limited size of segments"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), 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_sysctl_root), OID_AUTO, "minto", CTLFLAG_RW, &rack_min_to, 1, "Minimum rack timeout in milliseconds"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "earlyrecovery", CTLFLAG_RW, &rack_early_recovery, 1, "Do we do early recovery with rack"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), 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_sysctl_root), 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_sysctl_root), OID_AUTO, "reorder_fade", CTLFLAG_RW, &rack_reorder_fade, 0, "Does reorder detection fade, if so how many ms (0 means never)"); SYSCTL_ADD_S32(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "pktdelay", CTLFLAG_RW, &rack_pkt_delay, 1, "Extra RACK time (in ms) besides reordering thresh"); rack_counters = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "stats", CTLFLAG_RW, 0, "Rack Counters"); rack_badfr = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "badfr", CTLFLAG_RD, &rack_badfr, "Total number of bad FRs"); rack_badfr_bytes = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "badfr_bytes", CTLFLAG_RD, &rack_badfr_bytes, "Total number of bad FRs"); rack_rtm_prr_retran = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "prrsndret", CTLFLAG_RD, &rack_rtm_prr_retran, "Total number of prr based retransmits"); rack_rtm_prr_newdata = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "prrsndnew", CTLFLAG_RD, &rack_rtm_prr_newdata, "Total number of prr based new transmits"); rack_timestamp_mismatch = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "tsnf", CTLFLAG_RD, &rack_timestamp_mismatch, "Total number of timestamps that we could not find the reported ts"); rack_find_high = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "findhigh", CTLFLAG_RD, &rack_find_high, "Total number of FIN causing find-high"); rack_reorder_seen = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "reordering", CTLFLAG_RD, &rack_reorder_seen, "Total number of times we added delay due to reordering"); 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_tlp_retran_fail = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "tlp_retran_fail", CTLFLAG_RD, &rack_tlp_retran_fail, "Total number of tail loss probe sending retransmitted data that failed (wait for t3)"); 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_to_arm_rack = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "arm_rack", CTLFLAG_RD, &rack_to_arm_rack, "Total number of times the rack timer armed?"); rack_to_arm_tlp = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "arm_tlp", CTLFLAG_RD, &rack_to_arm_tlp, "Total number of times the tlp timer armed?"); rack_calc_zero = counter_u64_alloc(M_WAITOK); rack_calc_nonzero = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "calc_zero", CTLFLAG_RD, &rack_calc_zero, "Total number of times pacing time worked out to zero?"); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "calc_nonzero", CTLFLAG_RD, &rack_calc_nonzero, "Total number of times pacing time worked out to non-zero?"); rack_paced_segments = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "paced", CTLFLAG_RD, &rack_paced_segments, "Total number of times a segment send caused hptsi"); rack_unpaced_segments = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "unpaced", CTLFLAG_RD, &rack_unpaced_segments, "Total number of times a segment did not cause hptsi"); 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 segment did not cause hptsi"); 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 segment did not cause hptsi"); 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_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_enter_tlp_calc = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "tlp_calc_entered", CTLFLAG_RD, &rack_enter_tlp_calc, "Total times we called calc-tlp"); rack_used_tlpmethod = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "hit_tlp_method", CTLFLAG_RD, &rack_used_tlpmethod, "Total number of runt sacks"); rack_used_tlpmethod2 = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "hit_tlp_method2", CTLFLAG_RD, &rack_used_tlpmethod2, "Total number of times we hit TLP method 2"); /* Sack Attacker detection stuff */ rack_attack = SYSCTL_ADD_NODE(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_sysctl_root), OID_AUTO, "sack_attack", CTLFLAG_RW, 0, "Rack Sack Attack Counters and Controls"); 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 SACK's"); 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 postions 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 postions 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"); 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_progress_drops = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "prog_drops", CTLFLAG_RD, &rack_progress_drops, "Total number of progress drops"); 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 = 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 windows"); rack_tlp_does_nada = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "tlp_nada", CTLFLAG_RD, &rack_tlp_does_nada, "Total number of nada tlp calls"); rack_tls_rwnd = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "tls_rwnd", CTLFLAG_RD, &rack_tls_rwnd, "Total hdwr tls rwnd limited"); rack_tls_cwnd = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "tls_cwnd", CTLFLAG_RD, &rack_tls_cwnd, "Total hdwr tls cwnd limited"); rack_tls_app = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "tls_app", CTLFLAG_RD, &rack_tls_app, "Total hdwr tls app limited"); rack_tls_other = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "tls_other", CTLFLAG_RD, &rack_tls_other, "Total hdwr tls other limited"); rack_tls_filled = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "tls_filled", CTLFLAG_RD, &rack_tls_filled, "Total hdwr tls filled"); rack_tls_rxt = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "tls_rxt", CTLFLAG_RD, &rack_tls_rxt, "Total hdwr rxt"); rack_tls_tlp = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "tls_tlp", CTLFLAG_RD, &rack_tls_tlp, "Total hdwr tls tlp"); rack_per_timer_hole = counter_u64_alloc(M_WAITOK); SYSCTL_ADD_COUNTER_U64(&rack_sysctl_ctx, SYSCTL_CHILDREN(rack_counters), OID_AUTO, "timer_hole", CTLFLAG_RD, &rack_per_timer_hole, "Total persists start in timer hole"); 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 inline int32_t rack_progress_timeout_check(struct tcpcb *tp) { if (tp->t_maxunacktime && tp->t_acktime && TSTMP_GT(ticks, tp->t_acktime)) { if ((ticks - tp->t_acktime) >= tp->t_maxunacktime) { /* * There is an assumption that the caller * will drop the connection so we will * increment the counters here. */ struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; counter_u64_add(rack_progress_drops, 1); #ifdef NETFLIX_STATS TCPSTAT_INC(tcps_progdrops); #endif rack_log_progress_event(rack, tp, ticks, PROGRESS_DROP, __LINE__); return (1); } } return (0); } 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; 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 = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; 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_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 = TICKS_2_MSEC(rack->rc_tp->t_srtt >> TCP_RTT_SHIFT); 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; log.u_bbr.pkts_out = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; 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_TIMERSTAR, 0, 0, &log, false, &tv); } } static void rack_log_to_event(struct tcp_rack *rack, int32_t to_num, int no) { 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 = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; 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; log.u_bbr.flex3 = no; 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); 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_rtt_upd(struct tcpcb *tp, struct tcp_rack *rack, int32_t t, uint32_t o_srtt, uint32_t o_var) { 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 = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; log.u_bbr.flex1 = t; log.u_bbr.flex2 = o_srtt; log.u_bbr.flex3 = o_var; 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_rtt_cnt; log.u_bbr.rttProp = rack->r_ctl.rack_rs.rs_rtt_tot; log.u_bbr.flex8 = rack->r_ctl.rc_rate_sample_method; log.u_bbr.pkts_out = 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); 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 * 1000; 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.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); 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 = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; 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); 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) { 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 = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; log.u_bbr.flex1 = slot; 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 = 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_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) { 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; log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.applimited = rack->r_ctl.rc_pace_min_segs; log.u_bbr.flex7 = rack->r_wanted_output; log.u_bbr.flex8 = rack->rc_in_persist; 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_DOSEG_DONE, 0, 0, &log, false, &tv); } } static void rack_log_type_hrdwtso(struct tcpcb *tp, struct tcp_rack *rack, int len, int mod, int32_t orig_len, int frm) { if (tp->t_logstate != TCP_LOG_STATE_OFF) { union tcp_log_stackspecific log; struct timeval tv; uint32_t cts; memset(&log, 0, sizeof(log)); cts = tcp_get_usecs(&tv); 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 = len; log.u_bbr.flex5 = orig_len; log.u_bbr.flex6 = rack->r_ctl.rc_sacked; log.u_bbr.flex7 = mod; 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); TCP_LOG_EVENTP(tp, NULL, &tp->t_inpcb->inp_socket->so_rcv, &tp->t_inpcb->inp_socket->so_snd, TCP_HDWR_TLS, 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) { 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 = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; log.u_bbr.flex1 = slot; log.u_bbr.flex2 = rack->r_ctl.rc_hpts_flags; 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.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_JUSTRET, 0, tlen, &log, false, &tv); } } static void rack_log_to_cancel(struct tcp_rack *rack, int32_t hpts_removed, 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.inhpts = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; log.u_bbr.flex1 = line; log.u_bbr.flex2 = 0; log.u_bbr.flex3 = rack->r_ctl.rc_hpts_flags; log.u_bbr.flex4 = 0; log.u_bbr.flex5 = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.flex6 = rack->rc_tp->t_rxtcur; log.u_bbr.flex8 = hpts_removed; 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_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; log.u_bbr.flex6 = 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); 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) { 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; 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.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); 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_badfr); counter_u64_free(rack_badfr_bytes); counter_u64_free(rack_rtm_prr_retran); counter_u64_free(rack_rtm_prr_newdata); counter_u64_free(rack_timestamp_mismatch); counter_u64_free(rack_reorder_seen); 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_tlp_retran_fail); counter_u64_free(rack_to_tot); counter_u64_free(rack_to_arm_rack); counter_u64_free(rack_to_arm_tlp); counter_u64_free(rack_paced_segments); counter_u64_free(rack_unpaced_segments); counter_u64_free(rack_saw_enobuf); counter_u64_free(rack_saw_enetunreach); counter_u64_free(rack_to_alloc_hard); counter_u64_free(rack_to_alloc_emerg); counter_u64_free(rack_sack_proc_all); counter_u64_free(rack_sack_proc_short); counter_u64_free(rack_sack_proc_restart); counter_u64_free(rack_to_alloc); counter_u64_free(rack_to_alloc_limited); counter_u64_free(rack_alloc_limited_conns); counter_u64_free(rack_split_limited); counter_u64_free(rack_find_high); counter_u64_free(rack_enter_tlp_calc); counter_u64_free(rack_used_tlpmethod); counter_u64_free(rack_used_tlpmethod2); counter_u64_free(rack_progress_drops); counter_u64_free(rack_input_idle_reduces); counter_u64_free(rack_collapsed_win); counter_u64_free(rack_tlp_does_nada); 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; 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); } 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_limit_type) { /* currently there is only one limit type */ rack->r_ctl.rc_num_split_allocs--; } 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; if (rack->rc_free_cnt < rack_free_cache) { memset(rsm, 0, sizeof(struct rack_sendmap)); TAILQ_INSERT_TAIL(&rack->r_ctl.rc_free, rsm, r_tnext); rsm->r_limit_type = 0; rack->rc_free_cnt++; return; } rack->r_ctl.rc_num_maps_alloced--; uma_zfree(rack_zone, rsm); } /* * CC wrapper hook functions */ static void rack_ack_received(struct tcpcb *tp, struct tcp_rack *rack, struct tcphdr *th, uint16_t nsegs, uint16_t type, int32_t recovery) { #ifdef STATS int32_t gput; #endif INP_WLOCK_ASSERT(tp->t_inpcb); tp->ccv->nsegs = nsegs; tp->ccv->bytes_this_ack = BYTES_THIS_ACK(tp, th); 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; } } if ((!V_tcp_do_newcwv && (tp->snd_cwnd <= tp->snd_wnd)) || (V_tcp_do_newcwv && (tp->snd_cwnd <= tp->snd_wnd) && (tp->snd_cwnd < (ctf_flight_size(tp, rack->r_ctl.rc_sacked) * 2)))) tp->ccv->flags |= CCF_CWND_LIMITED; else tp->ccv->flags &= ~CCF_CWND_LIMITED; if (type == CC_ACK) { #ifdef STATS stats_voi_update_abs_s32(tp->t_stats, VOI_TCP_CALCFRWINDIFF, ((int32_t) tp->snd_cwnd) - tp->snd_wnd); if ((tp->t_flags & TF_GPUTINPROG) && SEQ_GEQ(th->th_ack, tp->gput_ack)) { gput = (((int64_t) (th->th_ack - tp->gput_seq)) << 3) / max(1, tcp_ts_getticks() - tp->gput_ts); /* We store it in bytes per ms (or kbytes per sec) */ rack->r_ctl.rc_gp_history[rack->r_ctl.rc_gp_hist_idx] = gput / 8; rack->r_ctl.rc_gp_hist_idx++; if (rack->r_ctl.rc_gp_hist_idx >= RACK_GP_HIST) rack->r_ctl.rc_gp_hist_filled = 1; rack->r_ctl.rc_gp_hist_idx %= RACK_GP_HIST; 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); tp->t_flags &= ~TF_GPUTINPROG; tp->t_stats_gput_prev = gput; if (tp->t_maxpeakrate) { /* * We update t_peakrate_thr. This gives us roughly * one update per round trip time. */ tcp_update_peakrate_thr(tp); } } #endif 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)); if (tp->t_bytes_acked >= tp->snd_cwnd) { tp->t_bytes_acked -= tp->snd_cwnd; tp->ccv->flags |= CCF_ABC_SENTAWND; } } else { tp->ccv->flags &= ~CCF_ABC_SENTAWND; tp->t_bytes_acked = 0; } } if (CC_ALGO(tp)->ack_received != NULL) { /* XXXLAS: Find a way to live without this */ tp->ccv->curack = th->th_ack; CC_ALGO(tp)->ack_received(tp->ccv, type); } #ifdef STATS stats_voi_update_abs_ulong(tp->t_stats, VOI_TCP_LCWIN, tp->snd_cwnd); #endif if (rack->r_ctl.rc_rack_largest_cwnd < tp->snd_cwnd) { rack->r_ctl.rc_rack_largest_cwnd = tp->snd_cwnd; } /* we enforce max peak rate if it is set. */ if (tp->t_peakrate_thr && tp->snd_cwnd > tp->t_peakrate_thr) { tp->snd_cwnd = tp->t_peakrate_thr; } } static void tcp_rack_partialack(struct tcpcb *tp, struct tcphdr *th) { struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; INP_WLOCK_ASSERT(tp->t_inpcb); if (rack->r_ctl.rc_prr_sndcnt > 0) rack->r_wanted_output++; } static void rack_post_recovery(struct tcpcb *tp, struct tcphdr *th) { struct tcp_rack *rack; INP_WLOCK_ASSERT(tp->t_inpcb); rack = (struct tcp_rack *)tp->t_fb_ptr; if (CC_ALGO(tp)->post_recovery != NULL) { tp->ccv->curack = th->th_ack; CC_ALGO(tp)->post_recovery(tp->ccv); } /* * Here we can in theory adjust cwnd to be based on the number of * losses in the window (rack->r_ctl.rc_loss_count). This is done * based on the rack_use_proportional flag. */ if (rack->r_ctl.rc_prop_reduce && rack->r_ctl.rc_prop_rate) { int32_t reduce; reduce = (rack->r_ctl.rc_loss_count * rack->r_ctl.rc_prop_rate); if (reduce > 50) { reduce = 50; } tp->snd_cwnd -= ((reduce * tp->snd_cwnd) / 100); } else { if (tp->snd_cwnd > tp->snd_ssthresh) { /* Drop us down to the ssthresh (1/2 cwnd at loss) */ tp->snd_cwnd = tp->snd_ssthresh; } } if (rack->r_ctl.rc_prr_sndcnt > 0) { /* Suck the next prr cnt back into cwnd */ tp->snd_cwnd += rack->r_ctl.rc_prr_sndcnt; rack->r_ctl.rc_prr_sndcnt = 0; rack_log_to_prr(rack, 1); } tp->snd_recover = tp->snd_una; EXIT_RECOVERY(tp->t_flags); } static void rack_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type) { struct tcp_rack *rack; INP_WLOCK_ASSERT(tp->t_inpcb); 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_tlp_rtx_out = 0; rack->r_ctl.rc_prr_delivered = 0; rack->r_ctl.rc_prr_out = 0; rack->r_ctl.rc_loss_count = 0; rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp); rack_log_to_prr(rack, 2); 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)) { TCPSTAT_INC(tcps_ecn_rcwnd); tp->snd_recover = tp->snd_max; 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, tp->snd_cwnd) / 2 / ctf_fixed_maxseg(tp)) * ctf_fixed_maxseg(tp); tp->snd_cwnd = ctf_fixed_maxseg(tp); break; case CC_RTO_ERR: 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) { if (th != NULL) tp->ccv->curack = th->th_ack; CC_ALGO(tp)->cong_signal(tp->ccv, type); } } static inline void rack_cc_after_idle(struct tcpcb *tp) { uint32_t i_cwnd; INP_WLOCK_ASSERT(tp->t_inpcb); #ifdef NETFLIX_STATS TCPSTAT_INC(tcps_idle_restarts); if (tp->t_state == TCPS_ESTABLISHED) 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 = tcp_compute_initwnd(tcp_maxseg(tp)); /* * Being idle is no differnt 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). */ counter_u64_add(rack_find_high, 1); 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; } thresh = srtt + rack->r_ctl.rc_pkt_delay; if (lro) { /* 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); } else { thresh += 1; } /* We don't let the rack timeout be above a RTO */ if (thresh > TICKS_2_MSEC(rack->rc_tp->t_rxtcur)) { thresh = TICKS_2_MSEC(rack->rc_tp->t_rxtcur); } /* And we don't want it above the RTO max either */ if (thresh > rack_rto_max) { thresh = rack_rto_max; } 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 maxseg; 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 */ maxseg = ctf_fixed_maxseg(tp); counter_u64_add(rack_enter_tlp_calc, 1); 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) <= maxseg) { uint32_t alt_thresh; /* * Compensate for delayed-ack with the d-ack time. */ counter_u64_add(rack_used_tlpmethod, 1); 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 <= maxseg)) { /* * Two packets outstanding, thresh should be (2*srtt) + * possible inter-packet delay (if any). */ uint32_t inter_gap = 0; int idx, nidx; counter_u64_add(rack_used_tlpmethod, 1); 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; counter_u64_add(rack_used_tlpmethod2, 1); 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 <= maxseg) { uint32_t alt_thresh; /* * Compensate for delayed-ack with the d-ack time. */ counter_u64_add(rack_used_tlpmethod, 1); alt_thresh = srtt + (srtt / 2) + rack_delayed_ack_time; if (alt_thresh > thresh) thresh = alt_thresh; } } /* Not above an RTO */ if (thresh > TICKS_2_MSEC(tp->t_rxtcur)) { thresh = TICKS_2_MSEC(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 (TICKS_2_MSEC(tp->t_srtt >> TCP_RTT_SHIFT)); } 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 (tsused < rsm->r_tim_lastsent[idx]) { return (NULL); } if ((tsused - rsm->r_tim_lastsent[idx]) < thresh) { return (NULL); } /* Ok if we reach here we are over-due */ rack->r_ctl.rc_rsm_start = rsm->r_start; rack->r_ctl.rc_cwnd_at = tp->snd_cwnd; rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh; rack_cong_signal(tp, NULL, CC_NDUPACK); 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 = TICKS_2_MSEC((tp->t_srtt >> TCP_RTT_SHIFT) + ((tp->t_rttvar * 4) >> TCP_RTT_SHIFT)); TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], rack_persist_min, rack_persist_max); if (tp->t_rxtshift < TCP_MAXRXTSHIFT) tp->t_rxtshift++; 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)); } 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 */ activate_rxt: time_since_sent = 0; rsm = TAILQ_FIRST(&rack->r_ctl.rc_tmap); if (rsm) { idx = rsm->r_rtr_cnt - 1; if (TSTMP_GEQ(rsm->r_tim_lastsent[idx], rack->r_ctl.rc_tlp_rxt_last_time)) tstmp_touse = rsm->r_tim_lastsent[idx]; else tstmp_touse = rack->r_ctl.rc_tlp_rxt_last_time; if (TSTMP_GT(tstmp_touse, cts)) 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 = TICKS_2_MSEC(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; 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) { 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_cheat == 0) && (IN_RECOVERY(tp->t_flags)) && (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 = 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; } } else { to = rack->r_ctl.rc_min_to; } } else { /* Ok we need to do a TLP not RACK */ activate_tlp: if ((rack->rc_tlp_in_progress != 0) || (rack->r_ctl.rc_tlp_rtx_out != 0)) { /* * The previous send was a TLP or a tlp_rtx is in * process. */ 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(rsm->r_tim_lastsent[idx], rack->r_ctl.rc_tlp_rxt_last_time)) tstmp_touse = rsm->r_tim_lastsent[idx]; else tstmp_touse = rack->r_ctl.rc_tlp_rxt_last_time; if (TSTMP_GT(tstmp_touse, cts)) time_since_sent = cts - tstmp_touse; is_tlp_timer = 1; if (tp->t_srtt) { srtt_cur = (tp->t_srtt >> TCP_RTT_SHIFT); srtt = TICKS_2_MSEC(srtt_cur); } else srtt = RACK_INITIAL_RTO; 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; if (to > TCPTV_REXMTMAX) { /* * If the TLP time works out to larger than the max * RTO lets not do TLP.. just RTO. */ goto activate_rxt; } if (rsm->r_start != rack->r_ctl.rc_last_tlp_seq) { /* * The tail is no longer the last one I did a probe * on */ rack->r_ctl.rc_tlp_seg_send_cnt = 0; rack->r_ctl.rc_last_tlp_seq = rsm->r_start; } } if (is_tlp_timer == 0) { rack->r_ctl.rc_hpts_flags |= PACE_TMR_RACK; } else { if ((rack->r_ctl.rc_tlp_send_cnt > rack_tlp_max_resend) || (rack->r_ctl.rc_tlp_seg_send_cnt > rack_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 { 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) { rack->r_ctl.rc_went_idle_time = cts; rack_timer_cancel(tp, rack, cts, __LINE__); tp->t_rxtshift = 0; rack->rc_in_persist = 1; } } static void rack_exit_persist(struct tcpcb *tp, struct tcp_rack *rack) { if (rack->rc_inp->inp_in_hpts) { tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); rack->r_ctl.rc_hpts_flags = 0; } rack->rc_in_persist = 0; rack->r_ctl.rc_went_idle_time = 0; tp->t_flags &= ~TF_FORCEDATA; tp->t_rxtshift = 0; } 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 inpcb *inp; uint32_t delayed_ack = 0; uint32_t hpts_timeout; uint8_t stopped; uint32_t left = 0; inp = tp->t_inpcb; if (inp->inp_in_hpts) { /* A previous call is already set up */ return; } if ((tp->t_state == TCPS_CLOSED) || (tp->t_state == TCPS_LISTEN)) { 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->tlp_timer_up = 0; rack->r_ctl.rc_timer_exp = 0; if (rack->rc_inp->inp_in_hpts == 0) { rack->r_ctl.rc_hpts_flags = 0; } if (slot) { /* We are hptsi too */ rack->r_ctl.rc_hpts_flags |= PACE_PKT_OUTPUT; } else if (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) { /* * We are still left on the hpts when the to goes * it will be for output. */ if (TSTMP_GT(rack->r_ctl.rc_last_output_to, cts)) slot = rack->r_ctl.rc_last_output_to - cts; else slot = 1; } hpts_timeout = rack_timer_start(tp, rack, cts, sup_rack); #ifdef NETFLIX_EXP_DETECTION if (rack->sack_attack_disable && (slot < USEC_TO_MSEC(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 = USEC_TO_MSEC(tcp_sad_pacing_interval); } #endif if (tp->t_flags & TF_DELACK) { delayed_ack = TICKS_2_MSEC(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 ((tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && + 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 = TP_KEEPIDLE(tp); } else { /* Get the initial setup keep-alive time */ hpts_timeout = TP_KEEPINIT(tp); } rack->r_ctl.rc_hpts_flags |= PACE_TMR_KEEP; } } 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; } if (slot) { rack->rc_inp->inp_flags2 |= INP_MBUF_QUEUE_READY; if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) inp->inp_flags2 |= INP_DONT_SACK_QUEUE; else inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE; rack->r_ctl.rc_last_output_to = cts + slot; if ((hpts_timeout == 0) || (hpts_timeout > slot)) { if (rack->rc_inp->inp_in_hpts == 0) tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(slot)); rack_log_to_start(rack, cts, hpts_timeout, slot, 1); } else { /* * Arrange for the hpts to kick back in after the * t-o if the t-o does not cause a send. */ if (rack->rc_inp->inp_in_hpts == 0) tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(hpts_timeout)); rack_log_to_start(rack, cts, hpts_timeout, slot, 0); } } else if (hpts_timeout) { if (rack->r_ctl.rc_hpts_flags & PACE_TMR_RACK) { /* For a rack timer, don't wake us */ rack->rc_inp->inp_flags2 |= INP_MBUF_QUEUE_READY; inp->inp_flags2 |= INP_DONT_SACK_QUEUE; } else { /* All other timers wake us up */ rack->rc_inp->inp_flags2 &= ~INP_MBUF_QUEUE_READY; inp->inp_flags2 &= ~INP_DONT_SACK_QUEUE; } if (rack->rc_inp->inp_in_hpts == 0) tcp_hpts_insert(tp->t_inpcb, HPTS_MS_TO_SLOTS(hpts_timeout)); 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, cts); } /* * 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; int32_t recovery, ll; 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); } recovery = IN_RECOVERY(tp->t_flags); counter_u64_add(rack_to_tot, 1); if (rack->r_state && (rack->r_state != tp->t_state)) rack_set_state(tp, rack); rsm = rack_check_recovery_mode(tp, cts); if (rsm) ll = rsm->r_end - rsm->r_start; else ll = 0; rack_log_to_event(rack, RACK_TO_FRM_RACK, ll); if (rsm) { uint32_t rtt; rtt = rack->rc_rack_rtt; if (rtt == 0) rtt = 1; if ((recovery == 0) && (rack->r_ctl.rc_prr_sndcnt < ctf_fixed_maxseg(tp))) { /* * The rack-timeout that enter's us into recovery * will force out one MSS and set us up so that we * can do one more send in 2*rtt (transitioning the * rack timeout into a rack-tlp). */ rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp); rack_log_to_prr(rack, 3); } else if ((rack->r_ctl.rc_prr_sndcnt < (rsm->r_end - rsm->r_start)) && rack->use_rack_cheat) { /* * When a rack timer goes, if the rack cheat is * on, arrange it so we can send a full segment. */ rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp); rack_log_to_prr(rack, 4); } } else { /* This is a case that should happen rarely if ever */ counter_u64_add(rack_tlp_does_nada, 1); #ifdef TCP_BLACKBOX tcp_log_dump_tp_logbuf(tp, "nada counter trips", M_NOWAIT, true); #endif rack->r_ctl.rc_resend = TAILQ_FIRST(&rack->r_ctl.rc_tmap); } rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RACK; return (0); } 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_rtr_bytes = 0; rsm->r_end = nrsm->r_start; for (idx = 0; idx < nrsm->r_rtr_cnt; idx++) { nrsm->r_tim_lastsent[idx] = rsm->r_tim_lastsent[idx]; } } 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). */ struct rack_sendmap *rm; 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; rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, r_rsm); #ifdef INVARIANTS 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) { /* * Tail Loss Probe. */ struct rack_sendmap *rsm = NULL; struct rack_sendmap *insret; struct socket *so; uint32_t amm, old_prr_snd = 0; 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 (rack_progress_timeout_check(tp)) { tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); return (1); } /* * 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, 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; #ifdef KERN_TLS if (rack->rc_inp->inp_socket->so_snd.sb_flags & SB_TLS_IFNET) { /* * For hardware TLS we do *not* want to send * new data, lets instead just do a retransmission. */ goto need_retran; } #endif avail = sbavail(&so->so_snd); out = tp->snd_max - tp->snd_una; rack->tlp_timer_up = 1; if (out > tp->snd_wnd) { /* special case, we need a retransmission */ collapsed_win = 1; goto need_retran; } /* * If we are in recovery we can jazz out a segment if new data is * present simply by setting rc_prr_sndcnt to a segment. */ if ((avail > out) && ((rack_always_send_oldest == 0) || (TAILQ_EMPTY(&rack->r_ctl.rc_tmap)))) { /* New data is available */ amm = avail - out; if (amm > ctf_fixed_maxseg(tp)) { amm = ctf_fixed_maxseg(tp); } else if ((amm < ctf_fixed_maxseg(tp)) && ((tp->t_flags & TF_NODELAY) == 0)) { /* not enough to fill a MTU and no-delay is off */ goto need_retran; } if (IN_RECOVERY(tp->t_flags)) { /* Unlikely */ old_prr_snd = rack->r_ctl.rc_prr_sndcnt; if (out + amm <= tp->snd_wnd) { rack->r_ctl.rc_prr_sndcnt = amm; rack_log_to_prr(rack, 4); } 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_tlp_seg_send_cnt = 0; rack->r_ctl.rc_last_tlp_seq = tp->snd_max; 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) { counter_u64_add(rack_tlp_does_nada, 1); #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) { counter_u64_add(rack_tlp_does_nada, 1); #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. */ counter_u64_add(rack_tlp_does_nada, 1); goto out; } rack_clone_rsm(rack, nrsm, rsm, (rsm->r_end - ctf_fixed_maxseg(tp))); insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); #ifdef INVARIANTS 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); rsm = nrsm; } rack->r_ctl.rc_tlpsend = rsm; rack->r_ctl.rc_tlp_rtx_out = 1; if (rsm->r_start == rack->r_ctl.rc_last_tlp_seq) { rack->r_ctl.rc_tlp_seg_send_cnt++; tp->t_rxtshift++; } else { rack->r_ctl.rc_last_tlp_seq = rsm->r_start; rack->r_ctl.rc_tlp_seg_send_cnt = 1; } send: rack->r_ctl.rc_tlp_send_cnt++; if (rack->r_ctl.rc_tlp_send_cnt > rack_tlp_max_resend) { /* * Can't [re]/transmit a segment we have not heard from the * peer in max times. We need the retransmit timer to take * over. */ restore: rack->r_ctl.rc_tlpsend = NULL; if (rsm) rsm->r_flags &= ~RACK_TLP; rack->r_ctl.rc_prr_sndcnt = old_prr_snd; rack_log_to_prr(rack, 5); counter_u64_add(rack_tlp_retran_fail, 1); goto out; } else if (rsm) { rsm->r_flags |= RACK_TLP; } if (rsm && (rsm->r_start == rack->r_ctl.rc_last_tlp_seq) && (rack->r_ctl.rc_tlp_seg_send_cnt > rack_tlp_max_resend)) { /* * We don't want to send a single segment more than the max * either. */ goto restore; } rack->r_timer_override = 1; rack->r_tlp_running = 1; rack->rc_tlp_in_progress = 1; rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_TLP; return (0); out: rack->tlp_timer_up = 0; 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, 0); tp->t_flags &= ~TF_DELACK; tp->t_flags |= TF_ACKNOW; TCPSTAT_INC(tcps_delack); rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_DELACK; return (0); } /* * Persists timer, here we simply need to setup the * FORCE-DATA flag the output routine will send * 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; struct inpcb *inp; int32_t retval = 1; inp = tp->t_inpcb; if (tp->t_timers->tt_flags & TT_STOPPED) { return (1); } if (rack->rc_in_persist == 0) return (0); if (rack_progress_timeout_check(tp)) { tcp_set_inp_to_drop(inp, ETIMEDOUT); return (1); } 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. */ 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 - tp->t_rcvtime >= TCP_REXMTVAL(tp) * tcp_totbackoff)) { TCPSTAT_INC(tcps_persistdrop); retval = 1; tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT); goto out; } if ((sbavail(&rack->rc_inp->inp_socket->so_snd) == 0) && tp->snd_una == tp->snd_max) rack_exit_persist(tp, rack); 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) { retval = 1; TCPSTAT_INC(tcps_persistdrop); tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT); goto out; } t_template = tcpip_maketemplate(rack->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++; out: rack_log_to_event(rack, RACK_TO_FRM_PERSIST, 0); 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, 0); /* * 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 ((tcp_always_keepalive || inp->inp_socket->so_options & SO_KEEPALIVE) && + 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) { 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: TCPSTAT_INC(tcps_keepdrops); tcp_set_inp_to_drop(rack->rc_inp, ETIMEDOUT); return (1); } /* * 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; int32_t cnt = 0; rack = (struct tcp_rack *)tp->t_fb_ptr; rack_timer_cancel(tp, rack, tcp_ts_getticks(), __LINE__); rack_log_to_event(rack, RACK_TO_FRM_TMR, 0); 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. */ RB_FOREACH(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) { if (rsm->r_flags & RACK_ACKED) { cnt++; rsm->r_dupack = 0; rack_log_retran_reason(rack, rsm, __LINE__, 0, 2); if (rsm->r_in_tmap == 0) { /* 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; rsm->r_flags &= ~(RACK_ACKED | RACK_SACK_PASSED | RACK_WAS_SACKPASS); } /* Clear the count (we just un-acked them) */ rack->r_ctl.rc_sacked = 0; /* Clear the tlp rtx mark */ rack->r_ctl.rc_tlp_rtx_out = 0; rack->r_ctl.rc_tlp_seg_send_cnt = 0; rack->r_ctl.rc_resend = RB_MIN(rack_rb_tree_head, &rack->r_ctl.rc_mtree); rack->r_ctl.rc_prr_sndcnt = 0; rack_log_to_prr(rack, 6); rack->r_timer_override = 1; } /* * 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; struct inpcb *inp; int32_t retval = 0; inp = tp->t_inpcb; if (tp->t_timers->tt_flags & TT_STOPPED) { return (1); } if (rack_progress_timeout_check(tp)) { tcp_set_inp_to_drop(inp, ETIMEDOUT); return (1); } rack->r_ctl.rc_hpts_flags &= ~PACE_TMR_RXT; 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. */ 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) { tp->t_rxtshift = TCP_MAXRXTSHIFT; TCPSTAT_INC(tcps_timeoutdrop); retval = 1; tcp_set_inp_to_drop(rack->rc_inp, (tp->t_softerror ? (uint16_t) tp->t_softerror : ETIMEDOUT)); goto out; } 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; tp->t_badrxtwin = ticks + (tp->t_srtt >> (TCP_RTT_SHIFT + 1)); 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 = MSEC_2_TICKS(RACK_INITIAL_RTO * tcp_backoff[tp->t_rxtshift]); else rexmt = TCP_REXMTVAL(tp) * tcp_backoff[tp->t_rxtshift]; TCPT_RANGESET(tp->t_rxtcur, rexmt, max(MSEC_2_TICKS(rack_rto_min), rexmt), MSEC_2_TICKS(rack_rto_max)); /* * 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. */ if (V_tcp_pmtud_blackhole_detect && (((tp->t_state == TCPS_ESTABLISHED)) || (tp->t_state == TCPS_FIN_WAIT_1))) { #ifdef INET6 int32_t isipv6; #endif /* * 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 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? 1 : 0; if (isipv6 && tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) { /* 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) { /* 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 } 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; TCPSTAT_INC(tcps_pmtud_blackhole_failed); } } } /* * 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 >> TCP_RTT_SHIFT); tp->t_srtt = 0; } if (rack_use_sack_filter) 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, NULL, CC_RTO); out: return (retval); } static int rack_process_timers(struct tcpcb *tp, struct tcp_rack *rack, uint32_t cts, uint8_t hpts_calling) { int32_t ret = 0; int32_t timers = (rack->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 (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) return (0); return (1); } 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) { 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); rack->rc_last_pto_set = 0; return (1); } 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; 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); } else if (timers & PACE_TMR_RXT) { rack->r_ctl.rc_tlp_rxt_last_time = cts; 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) { uint8_t hpts_removed = 0; if ((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) && TSTMP_GEQ(cts, rack->r_ctl.rc_last_output_to)) { tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); hpts_removed = 1; } if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { rack->rc_tmr_stopped = rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK; if (rack->rc_inp->inp_in_hpts && ((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_REMOVE_OUTPUT); hpts_removed = 1; } rack_log_to_cancel(rack, hpts_removed, line); rack->r_ctl.rc_hpts_flags &= ~(PACE_TMR_MASK); } } 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, uint32_t ts) { 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) && (rack->r_tlp_running == 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; 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; 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, uint32_t ts, int32_t *lenp) { /* * We (re-)transmitted starting at rsm->r_start for some length * (possibly less than r_end. */ struct rack_sendmap *nrsm, *insret; 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); 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); insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); #ifdef INVARIANTS 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); *lenp = 0; return (0); } static void rack_log_output(struct tcpcb *tp, struct tcpopt *to, int32_t len, uint32_t seq_out, uint8_t th_flags, int32_t err, uint32_t ts, uint8_t pass, struct rack_sendmap *hintrsm) { struct tcp_rack *rack; struct rack_sendmap *rsm, *nrsm, *insret, fe; 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; 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; } 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) 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 (len == 0) { /* We don't log zero window probes */ return; } rack->r_ctl.rc_time_last_sent = ts; if (IN_RECOVERY(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; } else { rsm->r_flags = 0; } rsm->r_tim_lastsent[0] = ts; 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_start = seq_out + 1; rsm->r_end = rsm->r_start + (len - 1); } else { /* Normal case */ rsm->r_start = seq_out; rsm->r_end = rsm->r_start + len; } rsm->r_dupack = 0; rack_log_retran_reason(rack, rsm, __LINE__, 0, 2); insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); #ifdef INVARIANTS 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; 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, ts, &len); 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, ts, &len); 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, ts); 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); insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); #ifdef INVARIANTS 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, ts, &len); 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) { 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; } 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; uint32_t o_srtt, o_var; 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; rack_log_rtt_sample(rack, rtt); o_srtt = tp->t_srtt; o_var = tp->t_rttvar; rack = (struct tcp_rack *)tp->t_fb_ptr; 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 - 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 << TCP_RTT_SHIFT; tp->t_rttvar = rtt << (TCP_RTTVAR_SHIFT - 1); tp->t_rttbest = tp->t_srtt + tp->t_rttvar; } TCPSTAT_INC(tcps_rttupdated); rack_log_rtt_upd(tp, rack, rtt, o_srtt, o_var); tp->t_rttupdated++; #ifdef STATS stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RTT, imax(0, rtt)); #endif tp->t_rxtshift = 0; /* * 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(rack_rto_min), rtt + 2), MSEC_2_TICKS(rack_rto_max)); tp->t_softerror = 0; } static void rack_earlier_retran(struct tcpcb *tp, struct rack_sendmap *rsm, uint32_t t, uint32_t cts) { /* * For this RSM, we acknowledged the data from a previous * transmission, not the last one we made. This means we did a false * retransmit. */ struct tcp_rack *rack; if (rsm->r_flags & RACK_HAS_FIN) { /* * The sending of the FIN often is multiple sent when we * have everything outstanding ack'd. We ignore this case * since its over now. */ return; } if (rsm->r_flags & RACK_TLP) { /* * We expect TLP's to have this occur. */ return; } rack = (struct tcp_rack *)tp->t_fb_ptr; /* should we undo cc changes and exit recovery? */ if (IN_RECOVERY(tp->t_flags)) { if (rack->r_ctl.rc_rsm_start == rsm->r_start) { /* * Undo what we ratched down and exit recovery if * possible */ EXIT_RECOVERY(tp->t_flags); tp->snd_recover = tp->snd_una; if (rack->r_ctl.rc_cwnd_at > tp->snd_cwnd) tp->snd_cwnd = rack->r_ctl.rc_cwnd_at; if (rack->r_ctl.rc_ssthresh_at > tp->snd_ssthresh) tp->snd_ssthresh = rack->r_ctl.rc_ssthresh_at; } } if (rsm->r_flags & RACK_WAS_SACKPASS) { /* * We retransmitted based on a sack and the earlier * retransmission ack'd it - re-ordering is occuring. */ counter_u64_add(rack_reorder_seen, 1); rack->r_ctl.rc_reorder_ts = cts; } counter_u64_add(rack_badfr, 1); counter_u64_add(rack_badfr_bytes, (rsm->r_end - rsm->r_start)); } 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) { int32_t i; uint32_t t; if (rsm->r_flags & RACK_ACKED) /* Already done */ return (0); if ((rsm->r_rtr_cnt == 1) || ((ack_type == CUM_ACKED) && (to->to_flags & TOF_TS) && (to->to_tsecr) && (rsm->r_tim_lastsent[rsm->r_rtr_cnt - 1] == to->to_tsecr)) ) { /* * We will only find a matching timestamp if its cum-acked. * But if its only one retransmission its for-sure matching * :-) */ t = cts - 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; } } tcp_rack_xmit_timer(rack, t + 1); if ((rsm->r_flags & RACK_TLP) && (!IN_RECOVERY(tp->t_flags))) { /* Segment was a TLP and our retrans matched */ if (rack->r_ctl.rc_tlp_cwnd_reduce) { rack->r_ctl.rc_rsm_start = tp->snd_max; rack->r_ctl.rc_cwnd_at = tp->snd_cwnd; rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh; rack_cong_signal(tp, NULL, CC_NDUPACK); /* * When we enter recovery we need to assure * we send one packet. */ rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp); rack_log_to_prr(rack, 7); } } if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) { /* New more recent rack_tmit_time */ rack->r_ctl.rc_rack_tmit_time = 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; tp->t_softerror = 0; if ((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 (rsm->r_tim_lastsent[i] == to->to_tsecr) { t = cts - rsm->r_tim_lastsent[i]; if ((int)t <= 0) t = 1; if ((i + 1) < rsm->r_rtr_cnt) { /* Likely */ rack_earlier_retran(tp, rsm, t, cts); } 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; } } /* * Note the following calls to * tcp_rack_xmit_timer() are being commented * out for now. They give us no more accuracy * and often lead to a wrong choice. We have * enough samples that have not been * retransmitted. I leave the commented out * code in here in case in the future we * decide to add it back (though I can't forsee * doing that). That way we will easily see * where they need to be placed. */ if (SEQ_LT(rack->r_ctl.rc_rack_tmit_time, rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)])) { /* New more recent rack_tmit_time */ rack->r_ctl.rc_rack_tmit_time = rsm->r_tim_lastsent[(rsm->r_rtr_cnt - 1)]; rack->rc_rack_rtt = t; } 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 - 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 * likey did an improper retransmit as outlined in * 4.2 Step 3 point 2 in the rack-draft. */ i = rsm->r_rtr_cnt - 2; t = cts - rsm->r_tim_lastsent[i]; rack_earlier_retran(tp, rsm, t, cts); } 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, rsm->r_tim_lastsent[i])) { /* New more recent rack_tmit_time */ rack->r_ctl.rc_rack_tmit_time = 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_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 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, *insret, *prev, *next; 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) { /** * 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; /* Clear out the dup ack count of the remainder */ rsm->r_dupack = 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); 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) { counter_u64_add(rack_reorder_seen, 1); 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; } 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); insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); #ifdef INVARIANTS 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); /* 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 * repostion 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_TLP) rack->r_ctl.rc_tlp_rtx_out = 0; if ((rsm->r_flags & RACK_ACKED) == 0) { rack_update_rtt(tp, rack, rsm, to, cts, SACKED); 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; counter_u64_add(rack_reorder_seen, 1); rack->r_ctl.rc_reorder_ts = cts; } rsm->r_flags |= RACK_ACKED; rsm->r_flags &= ~RACK_TLP; if (rsm->r_in_tmap) { TAILQ_REMOVE(&rack->r_ctl.rc_tmap, rsm, r_tnext); rsm->r_in_tmap = 0; } } 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) { 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) */ 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 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); 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) { counter_u64_add(rack_reorder_seen, 1); rack->r_ctl.rc_reorder_ts = cts; } 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; } /** * 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); insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); #ifdef INVARIANTS 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); if (rsm->r_flags & RACK_TLP) rack->r_ctl.rc_tlp_rtx_out = 0; rack_update_rtt(tp, rack, rsm, to, cts, SACKED); 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; counter_u64_add(rack_reorder_seen, 1); rack->r_ctl.rc_reorder_ts = cts; } rsm->r_flags |= RACK_ACKED; rsm->r_flags &= ~RACK_TLP; 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_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_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_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. */ if (rack_use_sack_filter) sack_filter_clear(&rack->r_ctl.rack_sf, th_ack); } static void rack_do_decay(struct tcp_rack *rack) { #ifdef NETFLIX_EXP_DETECTION 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.rc_last_ack, &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. */ uint32_t pkt_delta; pkt_delta = rack->r_ctl.input_pkt - rack->r_ctl.saved_input_pkt; /* 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.rc_last_ack; /* Now do we escape without decay? */ 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_log_ack(struct tcpcb *tp, struct tcpopt *to, struct tcphdr *th) { uint32_t changed, entered_recovery = 0; struct tcp_rack *rack; struct rack_sendmap *rsm, *rm; 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, sack_changed = 0; int loop_start = 0, moved_two = 0; INP_WLOCK_ASSERT(tp->t_inpcb); if (th->th_flags & TH_RST) { /* We don't log resets */ return; } rack = (struct tcp_rack *)tp->t_fb_ptr; cts = tcp_ts_getticks(); 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) { /* * 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++; 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. */ goto proc_sack; } if (tp->t_flags & TF_SENTFIN) { /* if we send a FIN we will not hav a map */ goto proc_sack; } #ifdef INVARIANTS panic("No rack map tp:%p for th:%p state:%d rack:%p snd_una:%u snd_max:%u snd_nxt:%u chg:%d\n", tp, th, tp->t_state, rack, tp->snd_una, tp->snd_max, tp->snd_nxt, changed); #endif goto proc_sack; } 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 goto proc_sack; } rack_update_rtt(tp, rack, rsm, to, cts, CUM_ACKED); /* Now do we consume the whole thing? */ if (SEQ_GEQ(th_ack, rsm->r_end)) { /* Its all consumed. */ uint32_t left; rack->r_ctl.rc_holes_rxt -= rsm->r_rtr_bytes; rsm->r_rtr_bytes = 0; if (rsm->r_flags & RACK_TLP) rack->r_ctl.rc_tlp_rtx_out = 0; rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); #ifdef INVARIANTS 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; } 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); } 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; counter_u64_add(rack_reorder_seen, 1); rsm->r_flags |= RACK_ACKED; rack->r_ctl.rc_reorder_ts = cts; } left = th_ack - rsm->r_end; if (rsm->r_rtr_cnt > 1) { /* * Technically we should make r_rtr_cnt be * monotonicly increasing and just mod it to * the timestamp it is replacing.. that way * we would have the last 3 retransmits. Now * rc_loss_count will be wrong if we * retransmit something more than 2 times in * recovery :( */ rack->r_ctl.rc_loss_count += (rsm->r_rtr_cnt - 1); } /* Free back to zone */ rack_free(rack, rsm); if (left) { goto more; } goto proc_sack; } 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 */ rsm->r_start = th_ack; } proc_sack: /* 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->th_ack); } if ((to->to_flags & TOF_SACK) == 0) { /* We are done nothing left */ 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++; #ifdef NETFLIX_STATS } else if (SEQ_LEQ(sack.start, th_ack) && SEQ_LEQ(sack.end, th_ack)) { /* * Its a D-SACK block. */ tcp_record_dsack(sack.start, sack.end); #endif } } /* * Sort the SACK blocks so we can update the rack scoreboard with * just one pass. */ if (rack_use_sack_filter) { 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++; changed += acked; sack_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++; changed += acked; sack_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 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(tp, th)/ctf_fixed_maxseg(rack->rc_tp))); 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 if (changed) { /* Something changed cancel the rack timer */ rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); } 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 rack_sendmap *rsm; uint32_t tsused; tsused = tcp_ts_getticks(); rsm = tcp_rack_output(tp, rack, tsused); if (rsm) { /* Enter recovery */ rack->r_ctl.rc_rsm_start = rsm->r_start; rack->r_ctl.rc_cwnd_at = tp->snd_cwnd; rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh; entered_recovery = 1; rack_cong_signal(tp, NULL, CC_NDUPACK); /* * When we enter recovery we need to assure we send * one packet. */ rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp); rack_log_to_prr(rack, 8); rack->r_timer_override = 1; } } if (IN_RECOVERY(tp->t_flags) && (entered_recovery == 0)) { /* Deal with changed and PRR here (in recovery only) */ uint32_t pipe, snd_una; rack->r_ctl.rc_prr_delivered += changed; /* Compute prr_sndcnt */ if (SEQ_GT(tp->snd_una, th_ack)) { snd_una = tp->snd_una; } else { snd_una = th_ack; } pipe = ((tp->snd_max - snd_una) - rack->r_ctl.rc_sacked) + rack->r_ctl.rc_holes_rxt; 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); 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); } 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); } else { rack->r_ctl.rc_prr_sndcnt = min(0, limit); rack_log_to_prr(rack, 12); } } if (rack->r_ctl.rc_prr_sndcnt >= ctf_fixed_maxseg(tp)) { 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); if (rsm && (rsm->r_dupack < 0xff)) { rsm->r_dupack++; if (rsm->r_dupack >= DUP_ACK_THRESHOLD) { rack->r_wanted_output = 1; rack_log_retran_reason(rack, rsm, __LINE__, 1, 3); } else { rack_log_retran_reason(rack, rsm, __LINE__, 0, 3); } } } /* * 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 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_wanted_output++; return (1); } if (SEQ_GEQ(th->th_ack, tp->snd_una) || to->to_nsacks) { if (rack->rc_in_persist) tp->t_rxtshift = 0; if ((th->th_ack == tp->snd_una) && (tiwin == tp->snd_wnd)) rack_strike_dupack(rack); rack_log_ack(tp, to, th); } if (__predict_false(SEQ_LEQ(th->th_ack, tp->snd_una))) { /* * Old ack, behind (or duplicate to) the last one rcv'd * Note: Should mark reordering is occuring! We should also * look for sack blocks arriving e.g. ack 1, 4-4 then ack 1, * 3-3, 4-4 would be reording. As well as ack 1, 3-3 ack 3 */ 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); TCPSTAT_ADD(tcps_rcvackpack, nsegs); 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) rack_cong_signal(tp, th, CC_RTO_ERR); } /* * 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 (th->th_ack == tp->snd_max) { rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); rack->r_wanted_output++; } if (acked == 0) { if (ofia) *ofia = ourfinisacked; return (0); } if (rack->r_ctl.rc_early_recovery) { 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, th); } else { rack_post_recovery(tp, th); 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, 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)) { ourfinisacked = 1; } /* NB: sowwakeup_locked() does an implicit unlock. */ sowwakeup_locked(so); m_freem(mfree); if (rack->r_ctl.rc_early_recovery == 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, th); } else { rack_post_recovery(tp, th); } } } tp->snd_una = th->th_ack; 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 (tp->snd_una == tp->snd_max) { /* Nothing left outstanding */ 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++; if (rack_use_sack_filter) 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, time to * reset him. */ *ret_val = 1; tp = tcp_close(tp); ctf_do_dropwithreset(m, tp, th, BANDLIM_UNLIMITED, tlen); return (1); } } if (ofia) *ofia = ourfinisacked; return (0); } static void rack_collapsed_window(struct tcp_rack *rack) { /* * 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 rack_sendmap *rsm, *nrsm, fe, *insret; tcp_seq max_seq; uint32_t maxseg; max_seq = rack->rc_tp->snd_una + rack->rc_tp->snd_wnd; maxseg = ctf_fixed_maxseg(rack->rc_tp); memset(&fe, 0, sizeof(fe)); fe.r_start = max_seq; /* Find the first seq past or at maxseq */ rsm = RB_FIND(rack_rb_tree_head, &rack->r_ctl.rc_mtree, &fe); if (rsm == NULL) { /* Nothing to do strange */ rack->rc_has_collapsed = 0; return; } /* * Now do we need to split at * the collapse point? */ if (SEQ_GT(max_seq, rsm->r_start)) { 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 */ rack_clone_rsm(rack, nrsm, rsm, max_seq); insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm); #ifdef INVARIANTS 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; } /* * Set in the new RSM as the * collapsed starting point */ rsm = nrsm; } no_split: counter_u64_add(rack_collapsed_win, 1); RB_FOREACH_FROM(nrsm, rack_rb_tree_head, rsm) { nrsm->r_flags |= RACK_RWND_COLLAPSED; rack->rc_has_collapsed = 1; } } static void rack_un_collapse_window(struct tcp_rack *rack) { struct rack_sendmap *rsm; RB_FOREACH_REVERSE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree) { if (rsm->r_flags & RACK_RWND_COLLAPSED) rsm->r_flags &= ~RACK_RWND_COLLAPSED; else break; } rack->rc_has_collapsed = 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) 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; rack->r_wanted_output++; } 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 the window */ rack_collapsed_window(rack); else if (rack->rc_has_collapsed) rack_un_collapse_window(rack); /* 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); tp->snd_nxt = tp->snd_max; /* Make sure we output to start the timer */ rack->r_wanted_output++; } /* 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) && 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. */ tp->snd_nxt = tp->snd_una; rack_enter_persist(tp, rack, rack->r_ctl.rc_rcvtime); } if (tp->t_flags2 & TF2_DROP_AF_DATA) { m_freem(m); return (0); } /* * Process segments with URG. */ if ((thflags & TH_URG) && th->th_urp && TCPS_HAVERCVDFIN(tp->t_state) == 0) { /* * This is a kludge, but if we receive and accept random * urgent pointers, we'll crash in soreceive. It's hard to * imagine someone actually wanting to send this much urgent * data. */ SOCKBUF_LOCK(&so->so_rcv); if (th->th_urp + sbavail(&so->so_rcv) > sb_max) { th->th_urp = 0; /* XXX */ thflags &= ~TH_URG; /* XXX */ SOCKBUF_UNLOCK(&so->so_rcv); /* XXX */ goto dodata; /* XXX */ } /* * If this segment advances the known urgent pointer, then * mark the data stream. This should not happen in * CLOSE_WAIT, CLOSING, LAST_ACK or TIME_WAIT STATES since a * FIN has been received from the remote side. In these * states we ignore the URG. * * 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 as the original * spec states (in one of two places). */ if (SEQ_GT(th->th_seq + th->th_urp, tp->rcv_up)) { tp->rcv_up = th->th_seq + th->th_urp; so->so_oobmark = sbavail(&so->so_rcv) + (tp->rcv_up - tp->rcv_nxt) - 1; if (so->so_oobmark == 0) so->so_rcv.sb_state |= SBS_RCVATMARK; sohasoutofband(so); tp->t_oobflags &= ~(TCPOOB_HAVEDATA | TCPOOB_HADDATA); } SOCKBUF_UNLOCK(&so->so_rcv); /* * Remove out of band data so doesn't get presented to user. * This can happen independent of advancing the URG pointer, * but if two URG's are pending at once, some out-of-band * data may creep in... ick. */ if (th->th_urp <= (uint32_t) tlen && !(so->so_options & SO_OOBINLINE)) { /* hdr drop is delayed */ tcp_pulloutofband(so, th, m, drop_hdrlen); } } else { /* * If no out of band data is expected, pull receive urgent * pointer along with the receive window. */ if (SEQ_GT(tp->rcv_nxt, tp->rcv_up)) tp->rcv_up = tp->rcv_nxt; } dodata: /* XXX */ 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) && 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, tlen) || tfo_syn) { rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); tp->t_flags |= TF_DELACK; } else { rack->r_wanted_output++; tp->t_flags |= TF_ACKNOW; } tp->rcv_nxt += tlen; thflags = th->th_flags & TH_FIN; TCPSTAT_ADD(tcps_rcvpack, nsegs); 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_SACK_PERMIT) && (save_tlen > 0)) { 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) { 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++; } 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); TCPSTAT_INC(tcps_preddat); tp->rcv_nxt += tlen; /* * 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; TCPSTAT_ADD(tcps_rcvpack, nsegs); 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, so, 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, tlen)) { rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); tp->t_flags |= TF_DELACK; } else { tp->t_flags |= TF_ACKNOW; rack->r_wanted_output++; } if ((tp->snd_una == tp->snd_max) && rack_use_sack_filter) 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, uint8_t iptos) { 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 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 */ nsegs = max(1, m->m_pkthdr.lro_nsegs); rack_log_ack(tp, to, th); /* * We made progress, clear the tlp * out flag so we could start a TLP * again. */ rack->r_ctl.rc_tlp_rtx_out = 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 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); } /* 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) && 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. */ tp->snd_nxt = tp->snd_una; 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. */ 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) rack_cong_signal(tp, th, CC_RTO_ERR); } /* * 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 TCPSTAT_ADD(tcps_rcvackpack, nsegs); TCPSTAT_ADD(tcps_rcvackbyte, acked); sbdrop(&so->so_snd, acked); /* * Let the congestion control algorithm update congestion control * related information. This typically means increasing the * congestion window. */ rack_ack_received(tp, rack, th, nsegs, CC_ACK, 0); tp->snd_una = th->th_ack; if (tp->snd_wnd < ctf_outstanding(tp)) { /* The peer collapsed the window */ rack_collapsed_window(rack); } else if (rack->rc_has_collapsed) rack_un_collapse_window(rack); /* * 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 (tp->snd_una == tp->snd_max) { 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__); } /* Wake up the socket if we have room to write more */ sowwakeup(so); if (sbavail(&so->so_snd)) { rack->r_wanted_output++; } 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 tos) { 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, beginning with URG */ if ((thflags & TH_ACK) && (SEQ_LEQ(th->th_ack, tp->iss) || SEQ_GT(th->th_ack, tp->snd_max))) { 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; 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 == 0)) { rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); tp->t_flags |= TF_DELACK; } else { rack->r_wanted_output++; tp->t_flags |= TF_ACKNOW; } if (((thflags & (TH_CWR | TH_ECE)) == TH_ECE) && V_tcp_do_ecn) { tp->t_flags2 |= TF2_ECN_PERMIT; TCPSTAT_INC(tcps_ecn_shs); } 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); 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; TCPSTAT_INC(tcps_rcvpackafterwin); 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; t = tcp_ts_getticks() - to->to_tsecr; if (!tp->t_rttlow || tp->t_rttlow > t) tp->t_rttlow = t; tcp_rack_xmit_timer(rack, t + 1); 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))) { 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)) { 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)); /* * 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)) { 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)) { 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; /* * 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 (rack_process_data(m, th, so, tp, drop_hdrlen, tlen, tiwin, thflags, nxt_pkt)); } TCPSTAT_INC(tcps_connects); 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; /* * Account for the ACK of our SYN prior to * regular ACK processing below. */ tp->snd_una++; } 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); } /* * 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); 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; t = tcp_ts_getticks() - to->to_tsecr; if (!tp->t_rttlow || tp->t_rttlow > t) tp->t_rttlow = t; tcp_rack_xmit_timer(rack, t + 1); 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; /* * 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. */ 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)) { struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; if (tlen == 0) { if (rack_fastack(m, th, so, tp, to, drop_hdrlen, tlen, tiwin, nxt_pkt, rack->r_ctl.rc_rcvtime, iptos)) { 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)); /* * 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)) { 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++; 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 (rack_progress_timeout_check(tp)) { tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); ctf_do_dropwithreset(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; 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); } 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_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++; 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 (rack_progress_timeout_check(tp)) { tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); ctf_do_dropwithreset(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: tp = tcp_close(tp); 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; 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; 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); } 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. */ if ((so->so_state & SS_NOFDREF) && tlen) { if (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++; 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 (rack_progress_timeout_check(tp)) { tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); ctf_do_dropwithreset(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; 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); } 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. */ if ((so->so_state & SS_NOFDREF) && tlen) { if (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++; 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 (rack_progress_timeout_check(tp)) { tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); ctf_do_dropwithreset(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; 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); } 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. */ if ((so->so_state & SS_NOFDREF) && tlen) { if (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++; 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 (rack_progress_timeout_check(tp)) { tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); ctf_do_dropwithreset(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; 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); } /* * 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)) { return (ret_val); } /* * If new data are received on a connection after the user processes * are gone, then RST the other end. */ if ((so->so_state & SS_NOFDREF) && tlen) { if (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++; 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 (rack_progress_timeout_check(tp)) { tcp_set_inp_to_drop(tp->t_inpcb, ETIMEDOUT); ctf_do_dropwithreset(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 tls_seg = 0; #ifdef KERN_TLS if (rack->rc_inp->inp_socket->so_snd.sb_flags & SB_TLS_IFNET) { tls_seg = ctf_get_opt_tls_size(rack->rc_inp->inp_socket, rack->rc_tp->snd_wnd); rack->r_ctl.rc_pace_min_segs = tls_seg; } else #endif rack->r_ctl.rc_pace_min_segs = ctf_fixed_maxseg(tp); rack->r_ctl.rc_pace_max_segs = ctf_fixed_maxseg(tp) * rack->rc_pace_max_segs; if (rack->r_ctl.rc_pace_max_segs > PACE_MAX_IP_BYTES) rack->r_ctl.rc_pace_max_segs = PACE_MAX_IP_BYTES; #ifdef KERN_TLS if (tls_seg != 0) { if (rack_hw_tls_max_seg > 1) { rack->r_ctl.rc_pace_max_segs /= tls_seg; if (rack_hw_tls_max_seg < rack->r_ctl.rc_pace_max_segs) rack->r_ctl.rc_pace_max_segs = rack_hw_tls_max_seg; } else { rack->r_ctl.rc_pace_max_segs = 1; } if (rack->r_ctl.rc_pace_max_segs == 0) rack->r_ctl.rc_pace_max_segs = 1; rack->r_ctl.rc_pace_max_segs *= tls_seg; } #endif rack_log_type_hrdwtso(tp, rack, tls_seg, rack->rc_inp->inp_socket->so_snd.sb_flags, 0, 2); } static int rack_init(struct tcpcb *tp) { struct tcp_rack *rack = NULL; struct rack_sendmap *insret; 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 recusive (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; if (tp->t_inpcb) { rack->rc_inp = tp->t_inpcb; } tp->t_inpcb->inp_flags2 |= INP_SUPPORTS_MBUFQ; /* Probably not needed but lets be sure */ rack_clear_rate_sample(rack); rack->r_cpu = 0; 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->rc_pace_reduce = rack_slot_reduction; if (use_rack_cheat) rack->use_rack_cheat = 1; if (V_tcp_delack_enabled) tp->t_delayed_ack = 1; else tp->t_delayed_ack = 0; rack->rc_pace_max_segs = rack_hptsi_segments; rack->r_ctl.rc_reorder_shift = rack_reorder_thresh; rack->r_ctl.rc_pkt_delay = rack_pkt_delay; rack->r_ctl.rc_prop_reduce = rack_use_proportional_reduce; rack->r_enforce_min_pace = rack_min_pace_time; rack->r_ctl.rc_prop_rate = rack_proportional_rate; rack->r_ctl.rc_tlp_cwnd_reduce = rack_lower_cwnd_at_tlp; rack->r_ctl.rc_early_recovery = rack_early_recovery; rack->rc_always_pace = rack_pace_every_seg; rack_set_pace_segments(tp, rack); rack->r_ctl.rc_high_rwnd = tp->snd_wnd; 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; rack->rack_per_of_gp = rack_per_of_gp; microuptime(&rack->r_ctl.rc_last_ack); rack->r_ctl.rc_last_time_decay = rack->r_ctl.rc_last_ack; rack->r_ctl.rc_tlp_rxt_last_time = tcp_ts_getticks(); /* Do we force on detection? */ #ifdef NETFLIX_EXP_DETECTION if (tcp_force_detection) rack->do_detection = 1; else #endif rack->do_detection = 0; 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_flags = RACK_OVERMAX; rsm->r_tim_lastsent[0] = rack->r_ctl.rc_tlp_rxt_last_time; 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; insret = RB_INSERT(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); #ifdef INVARIANTS 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; } rack_stop_all_timers(tp); rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), 0, 0, 0); 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 you have to get to ESTAB or beyond * to tell. */ 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, *rm; if (tp->t_inpcb) { tp->t_inpcb->inp_flags2 &= ~INP_SUPPORTS_MBUFQ; tp->t_inpcb->inp_flags2 &= ~INP_MBUF_QUEUE_READY; } rack = (struct tcp_rack *)tp->t_fb_ptr; #ifdef TCP_BLACKBOX tcp_log_flowend(tp); #endif RB_FOREACH_SAFE(rsm, rack_rb_tree_head, &rack->r_ctl.rc_mtree, nrsm) { rm = RB_REMOVE(rack_rb_tree_head, &rack->r_ctl.rc_mtree, rsm); #ifdef INVARIANTS 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; } /* 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) { 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); rack->r_state = TCPS_ESTABLISHED; rack->r_substate = rack_do_established; break; case TCPS_CLOSE_WAIT: rack->r_state = TCPS_CLOSE_WAIT; rack->r_substate = rack_do_close_wait; break; case TCPS_FIN_WAIT_1: rack->r_state = TCPS_FIN_WAIT_1; rack->r_substate = rack_do_fin_wait_1; break; case TCPS_CLOSING: rack->r_state = TCPS_CLOSING; rack->r_substate = rack_do_closing; break; case TCPS_LAST_ACK: rack->r_state = TCPS_LAST_ACK; rack->r_substate = rack_do_lastack; break; case TCPS_FIN_WAIT_2: 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; }; } 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 possiblity * of nothing outstanding and the RXT up (and the hptsi timer). */ return; - } else if (((tcp_always_keepalive || + } 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. */ rack_timer_cancel(tp, rack, rack->r_ctl.rc_rcvtime, __LINE__); rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), 0, 0, 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) { int32_t thflags, retval, did_out = 0; int32_t way_out = 0; uint32_t cts; uint32_t tiwin; struct tcpopt to; struct tcp_rack *rack; struct rack_sendmap *rsm; int32_t prev_state = 0; if (m->m_flags & M_TSTMP_LRO) { tv->tv_sec = m->m_pkthdr.rcv_tstmp /1000000000; tv->tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000)/1000; } cts = tcp_tv_to_mssectick(tv); rack = (struct tcp_rack *)tp->t_fb_ptr; kern_prefetch(rack, &prev_state); prev_state = 0; thflags = th->th_flags; NET_EPOCH_ASSERT(); INP_WLOCK_ASSERT(tp->t_inpcb); KASSERT(tp->t_state > TCPS_LISTEN, ("%s: TCPS_LISTEN", __func__)); KASSERT(tp->t_state != TCPS_TIME_WAIT, ("%s: TCPS_TIME_WAIT", __func__)); 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 = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; log.u_bbr.flex1 = rack->r_ctl.rc_prr_sndcnt; log.u_bbr.flex2 = rack->r_ctl.rc_num_maps_alloced; 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->rc_tp->t_maxseg; TCP_LOG_EVENTP(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_IN, 0, tlen, &log, true, &tv); } if ((thflags & TH_SYN) && (thflags & TH_FIN) && V_drop_synfin) { way_out = 4; retval = 0; 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))) { ctf_do_dropwithreset(m, tp, th, BANDLIM_RST_OPENPORT, tlen); return(1); } /* * 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 - tp->t_rcvtime) >= tp->t_rxtcur)) { counter_u64_add(rack_input_idle_reduces, 1); rack_cc_after_idle(tp); } 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 (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 (tp->t_flags2 & TF2_ECN_PERMIT) { if (thflags & TH_CWR) tp->t_flags2 &= ~TF2_ECN_SND_ECE; switch (iptos & IPTOS_ECN_MASK) { case IPTOS_ECN_CE: tp->t_flags2 |= TF2_ECN_SND_ECE; TCPSTAT_INC(tcps_ecn_ce); break; case IPTOS_ECN_ECT0: TCPSTAT_INC(tcps_ecn_ect0); break; case IPTOS_ECN_ECT1: TCPSTAT_INC(tcps_ecn_ect1); break; } /* Process a packet differently from RFC3168. */ cc_ecnpkt_handler(tp, th, iptos); /* Congestion experienced. */ if (thflags & TH_ECE) { rack_cong_signal(tp, th, CC_ECN); } } /* * 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 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, 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. */ rack->r_cpu = inp_to_cpuid(tp->t_inpcb); 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; } /* * 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_RCVD_TSTMP; tp->ts_recent = to.to_tsval; tp->ts_recent_age = cts; } 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) { 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 */ rack->r_is_v6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0; tcp_set_hpts(tp->t_inpcb); sack_filter_clear(&rack->r_ctl.rack_sf, th->th_ack); } /* * 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). */ memcpy(&rack->r_ctl.rc_last_ack, tv, sizeof(struct timeval)); 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; rack->r_ctl.rc_tlp_send_cnt = 0; rack_clear_rate_sample(rack); 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->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); } } tcp_rack_xmit_timer_commit(rack, tp); if ((nxt_pkt == 0) || (IN_RECOVERY(tp->t_flags))) { if (rack->r_wanted_output != 0) { did_out = 1; (void)tp->t_fb->tfb_tcp_output(tp); } rack_start_hpts_timer(rack, tp, cts, 0, 0, 0); } 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) || - ((tcp_always_keepalive || rack->rc_inp->inp_socket->so_options & SO_KEEPALIVE) && + ((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) && (rack->rc_inp->inp_in_hpts) && (rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT)) { /* keep alive not needed if we are hptsi output yet */ ; } else { if (rack->rc_inp->inp_in_hpts) { tcp_hpts_remove(rack->rc_inp, HPTS_REMOVE_OUTPUT); counter_u64_add(rack_per_timer_hole, 1); } rack_start_hpts_timer(rack, tp, tcp_ts_getticks(), 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); 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 } 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) { tv.tv_sec = m->m_pkthdr.rcv_tstmp /1000000000; tv.tv_usec = (m->m_pkthdr.rcv_tstmp % 1000000000)/1000; } 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 (rsm && ((rsm->r_flags & RACK_ACKED) == 0)) { goto check_it; } rsm = rack_find_lowest_rsm(rack); if (rsm == NULL) { return (NULL); } check_it: if (rsm->r_flags & RACK_ACKED) { return (NULL); } if ((rsm->r_flags & RACK_SACK_PASSED) == 0) { /* Its not yet ready */ return (NULL); } srtt = rack_grab_rtt(tp, rack); idx = rsm->r_rtr_cnt - 1; ts_low = 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); return (rsm); } return (NULL); } static int32_t rack_get_pacing_delay(struct tcp_rack *rack, struct tcpcb *tp, uint32_t len) { int32_t slot = 0; if ((rack->rack_per_of_gp == 0) || (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. */ uint32_t srtt, cwnd, tr_perms = 0; old_method: if (rack->r_ctl.rc_rack_min_rtt) srtt = rack->r_ctl.rc_rack_min_rtt; else srtt = TICKS_2_MSEC((tp->t_srtt >> TCP_RTT_SHIFT)); if (rack->r_ctl.rc_rack_largest_cwnd) cwnd = rack->r_ctl.rc_rack_largest_cwnd; else cwnd = tp->snd_cwnd; tr_perms = cwnd / 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->rc_pace_reduce) { int32_t reduce; reduce = (slot / rack->rc_pace_reduce); if (reduce < slot) { slot -= reduce; } else slot = 0; } } else { int cnt; uint64_t bw_est, bw_raise, res, lentim; bw_est = 0; for (cnt=0; cntr_ctl.rc_gp_hist_filled == 0) && (rack->r_ctl.rc_gp_history[cnt] == 0)) break; bw_est += rack->r_ctl.rc_gp_history[cnt]; } if (bw_est == 0) { /* * No way yet to make a b/w estimate * (no goodput est yet). */ goto old_method; } /* Covert to bytes per second */ bw_est *= MSEC_IN_SECOND; /* * Now ratchet it up by our percentage. Note * that the minimum you can do is 1 which would * get you 101% of the average last N goodput estimates. * The max you can do is 256 which would yeild you * 356% of the last N goodput estimates. */ bw_raise = bw_est * (uint64_t)rack->rack_per_of_gp; bw_est += bw_raise; /* average by the number we added */ bw_est /= cnt; /* Now calculate a rate based on this b/w */ lentim = (uint64_t) len * (uint64_t)MSEC_IN_SECOND; res = lentim / bw_est; slot = (uint32_t)res; } if (rack->r_enforce_min_pace && (slot == 0)) { /* We are enforcing a minimum pace time of 1ms */ slot = rack->r_enforce_min_pace; } if (slot) counter_u64_add(rack_calc_nonzero, 1); else counter_u64_add(rack_calc_zero, 1); return (slot); } static int rack_output(struct tcpcb *tp) { struct socket *so; uint32_t recwin, sendwin; uint32_t sb_offset; int32_t len, flags, error = 0; struct mbuf *m; struct mbuf *mb; uint32_t if_hw_tsomaxsegcount = 0; uint32_t if_hw_tsomaxsegsize = 0; int32_t maxseg; long tot_len_this_send = 0; struct ip *ip = NULL; #ifdef TCPDEBUG struct ipovly *ipov = NULL; #endif struct udphdr *udp = NULL; struct tcp_rack *rack; struct tcphdr *th; uint8_t pass = 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; uint8_t hpts_calling, new_data_tlp = 0, doing_tlp = 0; int32_t do_a_prefetch; int32_t prefetch_rsm = 0; int force_tso = 0; int32_t orig_len; int32_t prefetch_so_done = 0; struct tcp_log_buffer *lgb = NULL; struct inpcb *inp; struct sockbuf *sb; #ifdef INET6 struct ip6_hdr *ip6 = NULL; int32_t isipv6; #endif uint8_t filled_all = 0; bool hw_tls = false; /* setup and take the cache hits here */ rack = (struct tcp_rack *)tp->t_fb_ptr; inp = rack->rc_inp; so = inp->inp_socket; sb = &so->so_snd; kern_prefetch(sb, &do_a_prefetch); do_a_prefetch = 1; #ifdef KERN_TLS hw_tls = (so->so_snd.sb_flags & SB_TLS_IFNET) != 0; #endif INP_WLOCK_ASSERT(inp); #ifdef TCP_OFFLOAD if (tp->t_flags & TF_TOE) return (tcp_offload_output(tp)); #endif maxseg = ctf_fixed_maxseg(tp); /* * 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 */ return (0); #ifdef INET6 if (rack->r_state) { /* Use the cache line loaded if possible */ isipv6 = rack->r_is_v6; } else { isipv6 = (inp->inp_vflag & INP_IPV6) != 0; } #endif cts = tcp_ts_getticks(); if (((rack->r_ctl.rc_hpts_flags & PACE_PKT_OUTPUT) == 0) && inp->inp_in_hpts) { /* * We are on the hpts for some timer but not hptsi output. * Remove from the hpts unconditionally. */ rack_timer_cancel(tp, rack, cts, __LINE__); } /* Mark that we have called rack_output(). */ if ((rack->r_timer_override) || (tp->t_flags & TF_FORCEDATA) || (tp->t_state < TCPS_ESTABLISHED)) { if (tp->t_inpcb->inp_in_hpts) tcp_hpts_remove(tp->t_inpcb, HPTS_REMOVE_OUTPUT); } else if (tp->t_inpcb->inp_in_hpts) { /* * On the hpts you can't pass even if ACKNOW is on, we will * when the hpts fires. */ counter_u64_add(rack_out_size[TCP_MSS_ACCT_INPACE], 1); return (0); } hpts_calling = inp->inp_hpts_calls; inp->inp_hpts_calls = 0; if (rack->r_ctl.rc_hpts_flags & PACE_TMR_MASK) { if (rack_process_timers(tp, rack, cts, hpts_calling)) { counter_u64_add(rack_out_size[TCP_MSS_ACCT_ATIMER], 1); return (0); } } rack->r_wanted_output = 0; rack->r_timer_override = 0; /* * 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 */ return (0); /* * 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 - tp->t_rcvtime) >= tp->t_rxtcur)) rack_cc_after_idle(tp); } tp->t_flags &= ~TF_LASTIDLE; if (idle) { if (tp->t_flags & TF_MORETOCOME) { tp->t_flags |= TF_LASTIDLE; idle = 0; } } 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_ts_getticks(); tso = 0; mtu = 0; sb_offset = tp->snd_max - tp->snd_una; sendwin = min(tp->snd_wnd, 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; 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(sb); goto send; } if (rack->r_ctl.rc_tlpsend) { /* Tail loss probe */ long cwin; long tlen; doing_tlp = 1; /* * 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 = tcp_rack_output(tp, rack, cts); if (rsm == NULL) rsm = rack->r_ctl.rc_tlpsend; rack->r_ctl.rc_tlpsend = NULL; sack_rxmit = 1; tlen = rsm->r_end - rsm->r_start; if (tlen > ctf_fixed_maxseg(tp)) tlen = ctf_fixed_maxseg(tp); 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; } else 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 >= ctf_fixed_maxseg(tp)) { len = ctf_fixed_maxseg(tp); } } else if ((rack->rc_in_persist == 0) && ((rsm = tcp_rack_output(tp, rack, cts)) != NULL)) { int maxseg; maxseg = ctf_fixed_maxseg(tp); if ((!IN_RECOVERY(tp->t_flags)) && ((tp->t_flags & (TF_WASFRECOVERY | TF_WASCRECOVERY)) == 0)) { /* Enter recovery if not induced by a time-out */ rack->r_ctl.rc_rsm_start = rsm->r_start; rack->r_ctl.rc_cwnd_at = tp->snd_cwnd; rack->r_ctl.rc_ssthresh_at = tp->snd_ssthresh; rack_cong_signal(tp, NULL, CC_NDUPACK); /* * When we enter recovery we need to assure we send * one packet. */ rack->r_ctl.rc_prr_sndcnt = ctf_fixed_maxseg(tp); rack_log_to_prr(rack, 13); } #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; /* Can we send it within the PRR boundary? */ if ((rack->use_rack_cheat == 0) && (len > rack->r_ctl.rc_prr_sndcnt)) { /* It does not fit */ if ((ctf_flight_size(tp, rack->r_ctl.rc_sacked) > len) && (rack->r_ctl.rc_prr_sndcnt < maxseg)) { /* * prr is less than a segment, we * have more acks due in besides * what we need to resend. Lets not send * to avoid sending small pieces of * what we need to retransmit. */ len = 0; goto just_return_nolock; } len = rack->r_ctl.rc_prr_sndcnt; } sendalot = 0; if (len >= maxseg) { len = maxseg; } if (len > 0) { sub_from_prr = 1; sack_rxmit = 1; TCPSTAT_INC(tcps_sack_rexmits); TCPSTAT_ADD(tcps_sack_rexmit_bytes, min(len, ctf_fixed_maxseg(tp))); counter_u64_add(rack_rtm_prr_retran, 1); } } /* * 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); } 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; } } #ifdef INVARIANTS /* For debugging */ rack->r_ctl.rc_rsm_at_retran = rsm; #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 in persist timeout with window of 0, send 1 byte. Otherwise, * if window is small but nonzero and time TF_SENTFIN expired, we * will send what we can and go to transmit state. */ if (tp->t_flags & TF_FORCEDATA) { if (sendwin == 0) { /* * If we still have some data to send, then clear * the FIN bit. Usually this would happen below * when it realizes that we aren't sending all the * data. However, if we have exactly 1 byte of * unsent data, then it won't clear the FIN bit * below, and if we are in persist state, we wind up * sending the packet without recording that we sent * the FIN bit. * * We can't just blindly clear the FIN bit, because * if we don't have any more data to send then the * probe will be the FIN itself. */ if (sb_offset < sbused(sb)) flags &= ~TH_FIN; sendwin = 1; } else { 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); /* * If we are dropping persist mode then we need to * correct snd_nxt/snd_max and off. */ tp->snd_nxt = tp->snd_max; sb_offset = tp->snd_nxt - tp->snd_una; } } /* * 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_nxt, tp->snd_una) && avail) sb_offset = tp->snd_nxt - tp->snd_una; else sb_offset = 0; if (IN_RECOVERY(tp->t_flags) == 0) { 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 > tp->snd_wnd) len = tp->snd_wnd; else len = rack->r_ctl.rc_tlp_new_data; rack->r_ctl.rc_tlp_new_data = 0; new_data_tlp = doing_tlp = 1; } else { if (sendwin > avail) { /* use the available */ if (avail > sb_offset) { len = (int32_t)(avail - sb_offset); } else { len = 0; } } else { if (sendwin > sb_offset) { len = (int32_t)(sendwin - sb_offset); } else { len = 0; } } } } else { uint32_t outstanding; /* * We are inside of a SACK recovery episode and are * sending new data, having retransmitted all the * data possible so far in the scoreboard. */ 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) && (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; counter_u64_add(rack_rtm_prr_newdata, 1); } } if (len > ctf_fixed_maxseg(tp)) { /* * 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 = ctf_fixed_maxseg(tp); } else if (len < ctf_fixed_maxseg(tp)) { /* * 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; } } } } 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))) { 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++; } /* * 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))) len = 0; 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))) { tp->snd_nxt = tp->snd_una; rack_enter_persist(tp, rack, cts); } } else if ((rsm == NULL) && ((doing_tlp == 0) || (new_data_tlp == 1)) && (len < rack->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((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs)) && (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 * we send nothing, and may enter persists. */ len = 0; if (tp->snd_max == tp->snd_una) { /* * Nothing out we can * go into persists. */ rack_enter_persist(tp, rack, cts); tp->snd_nxt = tp->snd_una; } } else if ((tp->snd_cwnd >= max(rack->r_ctl.rc_pace_min_segs, (maxseg * 4))) && (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (2 * maxseg)) && (len < (int)(sbavail(sb) - sb_offset)) && (len < rack->r_ctl.rc_pace_min_segs)) { /* * 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), rack->r_ctl.rc_pace_min_segs)) && (ctf_flight_size(tp, rack->r_ctl.rc_sacked) > (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); /* * 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(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) /* * 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 > ctf_fixed_maxseg(tp) && (tp->t_port == 0) && ((tp->t_flags & TF_SIGNATURE) == 0) && tp->rcv_numsacks == 0 && sack_rxmit == 0 && ipoptlen == 0) tso = 1; { uint32_t outstanding; 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 = sbspace(&so->so_rcv); /* * 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 >= ctf_fixed_maxseg(tp)) { pass = 1; 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(&so->so_snd)) && (tp->t_flags & TF_NOPUSH) == 0) { pass = 2; goto send; } if (tp->t_flags & TF_FORCEDATA) { /* typ. timeout case */ pass = 3; goto send; } if ((tp->snd_una == tp->snd_max) && len) { /* Nothing outstanding */ 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; } } /* * 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 = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale); if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) { oldwin = (tp->rcv_adv - tp->rcv_nxt); adv -= oldwin; } else oldwin = 0; /* * If the new window size ends up being the same as 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 * ctf_fixed_maxseg(tp)) && (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) || recwin <= (int32_t)(so->so_rcv.sb_hiwat / 8) || so->so_rcv.sb_hiwat <= 8 * ctf_fixed_maxseg(tp))) { pass = 7; goto send; } if (2 * adv >= (int32_t) so->so_rcv.sb_hiwat) 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 (SEQ_GT(tp->snd_up, tp->snd_una)) { pass = 10; 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: if (tot_len_this_send == 0) counter_u64_add(rack_out_size[TCP_MSS_ACCT_JUSTRET], 1); if (slot) { /* set the rack tcb into the slot N */ counter_u64_add(rack_paced_segments, 1); } else if (tot_len_this_send) { counter_u64_add(rack_unpaced_segments, 1); } /* Check if we need to go into persists or not */ if ((rack->rc_in_persist == 0) && (tp->snd_max == tp->snd_una) && TCPS_HAVEESTABLISHED(tp->t_state) && sbavail(&tp->t_inpcb->inp_socket->so_snd) && (sbavail(&tp->t_inpcb->inp_socket->so_snd) > tp->snd_wnd) && (tp->snd_wnd < min((rack->r_ctl.rc_high_rwnd/2), rack->r_ctl.rc_pace_min_segs))) { /* 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); tp->t_flags &= ~TF_FORCEDATA; return (0); send: if ((flags & TH_FIN) && sbavail(&tp->t_inpcb->inp_socket->so_snd)) { /* * We do not transmit a FIN * with data outstanding. We * need to make it so all data * is acked first. */ flags &= ~TH_FIN; } 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. */ rack->rc_tlp_in_progress = 0; } SOCKBUF_LOCK_ASSERT(sb); if (len > 0) { if (len >= ctf_fixed_maxseg(tp)) 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); #ifdef NETFLIX_TCPOUDP if (tp->t_port) to.to_mss -= V_tcp_udp_tunneling_overhead; #endif 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 = 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 (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; } #ifdef NETFLIX_TCPOUDP 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); } #endif #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 #ifdef KERN_TLS /* force TSO for so TLS offload can get mss */ if (sb->sb_flags & SB_TLS_IFNET) { force_tso = 1; } #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; } } /* * 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)) && (hw_tls == 0)) { moff = len % (u_int)max_len; if (moff != 0) { len -= moff; sendalot = 1; } } /* * In case there are too many small fragments don't * use TSO: */ if (len <= maxseg) { len = max_len; sendalot = 1; 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 = 1; } else { 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 = 1; } } else 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. */ 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. */ if (len) { uint32_t max_val; uint32_t moff; if (rack->rc_pace_max_segs) max_val = rack->rc_pace_max_segs * ctf_fixed_maxseg(tp); else max_val = len; if (rack->r_ctl.rc_pace_max_segs < max_val) max_val = rack->r_ctl.rc_pace_max_segs; /* * We allow a limit on sending with hptsi. */ if (len > max_val) { 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); 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( #ifdef NETFLIX_COPY_ARGS tp, #endif mb, moff, &len, if_hw_tsomaxsegcount, if_hw_tsomaxsegsize, msb, ((rsm == NULL) ? hw_tls : 0) #ifdef NETFLIX_COPY_ARGS , &filled_all #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 ((tp->t_flags & TF_FORCEDATA) && len == 1) { TCPSTAT_INC(tcps_sndprobe); #ifdef STATS if (SEQ_LT(tp->snd_nxt, tp->snd_max)) stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB, len); else stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB, len); #endif } else 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++; TCPSTAT_INC(tcps_sndrexmitpack); TCPSTAT_ADD(tcps_sndrexmitbyte, len); } #ifdef STATS stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB, len); #endif } else { TCPSTAT_INC(tcps_sndpack); 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; /* * Are we doing pacing, if so we must calculate the slot. We * only do hptsi in ESTABLISHED and with no RESET being * sent where we have data to send. */ if (((tp->t_state == TCPS_ESTABLISHED) || (tp->t_state == TCPS_CLOSE_WAIT) || ((tp->t_state == TCPS_FIN_WAIT_1) && ((tp->t_flags & TF_SENTFIN) == 0) && ((flags & TH_FIN) == 0))) && ((flags & TH_RST) == 0)) { /* Get our pacing rate */ tot_len_this_send += len; slot = rack_get_pacing_delay(rack, tp, tot_len_this_send); } SOCKBUF_UNLOCK(sb); } else { SOCKBUF_UNLOCK(sb); if (tp->t_flags & TF_ACKNOW) TCPSTAT_INC(tcps_sndacks); else if (flags & (TH_SYN | TH_FIN | TH_RST)) TCPSTAT_INC(tcps_sndctrl); else if (SEQ_GT(tp->snd_up, tp->snd_una)) TCPSTAT_INC(tcps_sndurg); else 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 #ifdef INET6 if (isipv6) { ip6 = mtod(m, struct ip6_hdr *); #ifdef NETFLIX_TCPOUDP if (tp->t_port) { udp = (struct udphdr *)((caddr_t)ip6 + ipoptlen + 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 #endif th = (struct tcphdr *)(ip6 + 1); tcpip_fillheaders(inp, #ifdef NETFLIX_TCPOUDP tp->t_port, #endif ip6, th); } else #endif /* INET6 */ { ip = mtod(m, struct ip *); #ifdef TCPDEBUG ipov = (struct ipovly *)ip; #endif #ifdef NETFLIX_TCPOUDP if (tp->t_port) { udp = (struct udphdr *)((caddr_t)ip + ipoptlen + 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 #endif th = (struct tcphdr *)(ip + 1); tcpip_fillheaders(inp, #ifdef NETFLIX_TCPOUDP tp->t_port, #endif 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 == 1) { if (tp->t_rxtshift >= 1) { if (tp->t_rxtshift <= V_tcp_ecn_maxretries) flags |= TH_ECE | TH_CWR; } else flags |= TH_ECE | TH_CWR; } if (tp->t_state == TCPS_ESTABLISHED && (tp->t_flags2 & TF2_ECN_PERMIT)) { /* * If the peer has ECN, mark data packets with ECN capable * transmission (ECT). Ignore pure ack packets, * retransmissions and window probes. */ if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) && !((tp->t_flags & TF_FORCEDATA) && len == 1)) { #ifdef INET6 if (isipv6) ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20); else #endif ip->ip_tos |= IPTOS_ECN_ECT0; TCPSTAT_INC(tcps_ecn_ect0); } /* * Reply with proper ECN notifications. */ if (tp->t_flags2 & TF2_ECN_SND_CWR) { flags |= TH_CWR; tp->t_flags2 &= ~TF2_ECN_SND_CWR; } if (tp->t_flags2 & TF2_ECN_SND_ECE) flags |= TH_ECE; } /* * 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)) || rack->rc_in_persist) { th->th_seq = htonl(tp->snd_nxt); rack_seq = tp->snd_nxt; } else if (flags & TH_RST) { /* * For a Reset send the last cum ack in sequence * (this like any other choice may still generate a * challenge ack, if a ack-update packet is in * flight). */ th->th_seq = htonl(tp->snd_una); rack_seq = tp->snd_una; } 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); if (optlen) { bcopy(opt, th + 1, optlen); th->th_off = (sizeof(struct tcphdr) + optlen) >> 2; } th->th_flags = 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)ctf_fixed_maxseg(tp)) 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); if (recwin > (long)TCP_MAXWIN << tp->rcv_scale) recwin = (long)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 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; if (SEQ_GT(tp->snd_up, tp->snd_nxt)) { th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt)); th->th_flags |= TH_URG; } else /* * If no urgent pointer to send, then we pull the urgent * pointer to the left edge of the send window so that it * doesn't drift into the send window on sequence number * wraparound. */ tp->snd_up = tp->snd_una; /* drag it along */ #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 { 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 || force_tso) { KASSERT(force_tso || 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 #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 */ /* We're getting ready to send; log now. */ 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 = rack->rc_inp->inp_in_hpts; log.u_bbr.ininput = rack->rc_inp->inp_in_input; 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; if (filled_all) log.u_bbr.flex5 = 0x80000000; else log.u_bbr.flex5 = 0; if (rsm || sack_rxmit) { log.u_bbr.flex8 = 1; } else { log.u_bbr.flex8 = 0; } log.u_bbr.pkts_out = tp->t_maxseg; log.u_bbr.timeStamp = tcp_get_usecs(&tv); log.u_bbr.inflight = ctf_flight_size(rack->rc_tp, rack->r_ctl.rc_sacked); 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 && 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, tp->t_inpcb->in6p_outputopts, &inp->inp_route6, ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), NULL, NULL, inp); if (error == EMSGSIZE && inp->inp_route6.ro_rt != NULL) mtu = inp->inp_route6.ro_rt->rt_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 */ /* * 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, tp->t_inpcb->inp_options, &inp->inp_route, ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, inp); if (error == EMSGSIZE && inp->inp_route.ro_rt != NULL) mtu = inp->inp_route.ro_rt->rt_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) { if (TCPS_HAVEESTABLISHED(tp->t_state) && (tp->t_flags & TF_SACK_PERMIT) && tp->rcv_numsacks > 0) tcp_clean_dsack_blocks(tp); 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 / ctf_fixed_maxseg(tp)) + 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 (hw_tls && len > 0) { if (filled_all) { counter_u64_add(rack_tls_filled, 1); rack_log_type_hrdwtso(tp, rack, len, 0, orig_len, 1); } else { if (rsm) { counter_u64_add(rack_tls_rxt, 1); rack_log_type_hrdwtso(tp, rack, len, 2, orig_len, 1); } else if (doing_tlp) { counter_u64_add(rack_tls_tlp, 1); rack_log_type_hrdwtso(tp, rack, len, 3, orig_len, 1); } else if ( (ctf_outstanding(tp) + rack->r_ctl.rc_pace_min_segs) > sbavail(sb)) { counter_u64_add(rack_tls_app, 1); rack_log_type_hrdwtso(tp, rack, len, 4, orig_len, 1); } else if ((ctf_flight_size(tp, rack->r_ctl.rc_sacked) + rack->r_ctl.rc_pace_min_segs) > tp->snd_cwnd) { counter_u64_add(rack_tls_cwnd, 1); rack_log_type_hrdwtso(tp, rack, len, 5, orig_len, 1); } else if ((ctf_outstanding(tp) + rack->r_ctl.rc_pace_min_segs) > tp->snd_wnd) { counter_u64_add(rack_tls_rwnd, 1); rack_log_type_hrdwtso(tp, rack, len, 6, orig_len, 1); } else { rack_log_type_hrdwtso(tp, rack, len, 7, orig_len, 1); counter_u64_add(rack_tls_other, 1); } } } } if (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; rack_log_output(tp, &to, len, rack_seq, (uint8_t) flags, error, cts, pass, rsm); if ((error == 0) && (len > 0) && (tp->snd_una == tp->snd_max)) rack->r_ctl.rc_tlp_rxt_last_time = cts; if ((tp->t_flags & TF_FORCEDATA) == 0 || (rack->rc_in_persist == 0)) { tcp_seq startseq = tp->snd_nxt; /* * Advance snd_nxt over sequence space of this segment. */ if (error) /* We don't log or do anything with errors */ goto nomore; 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; TCPSTAT_INC(tcps_segstimed); } #ifdef STATS if (!(tp->t_flags & TF_GPUTINPROG) && len) { tp->t_flags |= TF_GPUTINPROG; tp->gput_seq = startseq; tp->gput_ack = startseq + ulmin(sbavail(sb) - sb_offset, sendwin); tp->gput_ts = tcp_ts_getticks(); } #endif } } else { /* * Persist case, update snd_max but since we are in persist * mode (no window) we do not update snd_nxt. */ int32_t xlen = len; if (error) goto nomore; if (flags & TH_SYN) ++xlen; if (flags & TH_FIN) { ++xlen; tp->t_flags |= TF_SENTFIN; } /* In the ENOBUFS case we do *not* update snd_max */ if (SEQ_GT(tp->snd_nxt + xlen, 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 + len; } } nomore: if (error) { 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_flags &= ~TF_FORCEDATA; tp->t_softerror = error; return (error); case ENOBUFS: if (slot == 0) { /* * Pace us right away to retry in a some * time */ slot = 1 + rack->rc_enobuf; if (rack->rc_enobuf < 255) rack->rc_enobuf++; if (slot > (rack->rc_rack_rtt / 2)) { slot = rack->rc_rack_rtt / 2; } if (slot < 10) slot = 10; } counter_u64_add(rack_saw_enobuf, 1); 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. */ if (tso) tp->t_flags &= ~TF_TSO; if (mtu != 0) { tcp_mss_update(tp, -1, mtu, NULL, NULL); goto again; } slot = 10; rack_start_hpts_timer(rack, tp, cts, slot, 0, 0); tp->t_flags &= ~TF_FORCEDATA; 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; rack_start_hpts_timer(rack, tp, cts, slot, 0, 0); tp->t_flags &= ~TF_FORCEDATA; return (error); } } else { rack->rc_enobuf = 0; } 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: rack->r_tlp_running = 0; if (flags & TH_RST) { /* * We don't send again after sending a RST. */ slot = 0; sendalot = 0; } if (rsm && (slot == 0)) { /* * Dup ack retransmission possibly, so * lets assure we have at least min rack * time, if its a rack resend then the rack * to will also be set to this. */ slot = rack->r_ctl.rc_min_to; } if (slot) { /* set the rack tcb into the slot N */ counter_u64_add(rack_paced_segments, 1); } else if (sendalot) { if (len) counter_u64_add(rack_unpaced_segments, 1); sack_rxmit = 0; tp->t_flags &= ~TF_FORCEDATA; goto again; } else if (len) { counter_u64_add(rack_unpaced_segments, 1); } tp->t_flags &= ~TF_FORCEDATA; rack_start_hpts_timer(rack, tp, cts, slot, tot_len_this_send, 0); return (error); } /* * 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 socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack) { int32_t error = 0, optval; switch (sopt->sopt_name) { case TCP_RACK_PROP_RATE: case TCP_RACK_PROP: case TCP_RACK_TLP_REDUCE: case TCP_RACK_EARLY_RECOV: case TCP_RACK_PACE_ALWAYS: case TCP_DELACK: case TCP_RACK_PACE_REDUCE: case TCP_RACK_PACE_MAX_SEG: case TCP_RACK_PRR_SENDALOT: case TCP_RACK_MIN_TO: case TCP_RACK_EARLY_SEG: case TCP_RACK_REORD_THRESH: case TCP_RACK_REORD_FADE: case TCP_RACK_TLP_THRESH: case TCP_RACK_PKT_DELAY: case TCP_RACK_TLP_USE: case TCP_RACK_TLP_INC_VAR: case TCP_RACK_IDLE_REDUCE_HIGH: case TCP_RACK_MIN_PACE: case TCP_RACK_GP_INCREASE: case TCP_BBR_RACK_RTT_USE: case TCP_BBR_USE_RACK_CHEAT: case TCP_RACK_DO_DETECTION: case TCP_DATA_AFTER_CLOSE: break; default: return (tcp_default_ctloutput(so, sopt, inp, tp)); 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)) { INP_WUNLOCK(inp); return (ECONNRESET); } tp = intotcpcb(inp); rack = (struct tcp_rack *)tp->t_fb_ptr; switch (sopt->sopt_name) { 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_PROP_RATE: if ((optval <= 0) || (optval >= 100)) { error = EINVAL; break; } RACK_OPTS_INC(tcp_rack_prop_rate); rack->r_ctl.rc_prop_rate = optval; 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_PROP: /* RACK proportional rate reduction (bool) */ RACK_OPTS_INC(tcp_rack_prop); rack->r_ctl.rc_prop_reduce = 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; case TCP_RACK_EARLY_RECOV: /* Should recovery happen early (bool) */ RACK_OPTS_INC(tcp_rack_early_recov); rack->r_ctl.rc_early_recovery = optval; break; case TCP_RACK_PACE_ALWAYS: /* Use the always pace method (bool) */ RACK_OPTS_INC(tcp_rack_pace_always); if (optval > 0) rack->rc_always_pace = 1; else rack->rc_always_pace = 0; break; case TCP_RACK_PACE_REDUCE: /* RACK Hptsi reduction factor (divisor) */ RACK_OPTS_INC(tcp_rack_pace_reduce); if (optval) /* Must be non-zero */ rack->rc_pace_reduce = optval; else error = EINVAL; break; case TCP_RACK_PACE_MAX_SEG: /* Max segments in a pace */ RACK_OPTS_INC(tcp_rack_max_seg); rack->rc_pace_max_segs = optval; rack_set_pace_segments(tp, rack); break; 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_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_CHEAT: RACK_OPTS_INC(tcp_rack_cheat); if (optval) rack->use_rack_cheat = 1; else rack->use_rack_cheat = 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_RACK_TLP_INC_VAR: /* Does TLP include rtt variance in t-o */ error = EINVAL; break; case TCP_RACK_IDLE_REDUCE_HIGH: error = EINVAL; break; case TCP_DELACK: 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; rack_output(tp); } break; case TCP_RACK_MIN_PACE: RACK_OPTS_INC(tcp_rack_min_pace); if (optval > 3) rack->r_enforce_min_pace = 3; else rack->r_enforce_min_pace = optval; break; case TCP_RACK_GP_INCREASE: if ((optval >= 0) && (optval <= 256)) rack->rack_per_of_gp = optval; else error = EINVAL; break; case TCP_BBR_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: if (optval) rack->rc_allow_data_af_clo = 1; else rack->rc_allow_data_af_clo = 0; break; default: return (tcp_default_ctloutput(so, sopt, inp, tp)); break; } #ifdef NETFLIX_STATS tcp_log_socket_option(tp, sopt->sopt_name, optval, error); #endif INP_WUNLOCK(inp); return (error); } static int rack_get_sockopt(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp, struct tcp_rack *rack) { int32_t error, optval; /* * 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; switch (sopt->sopt_name) { case TCP_RACK_DO_DETECTION: optval = rack->do_detection; break; case TCP_RACK_PROP_RATE: optval = rack->r_ctl.rc_prop_rate; break; case TCP_RACK_PROP: /* RACK proportional rate reduction (bool) */ optval = rack->r_ctl.rc_prop_reduce; break; case TCP_RACK_TLP_REDUCE: /* RACK TLP cwnd reduction (bool) */ optval = rack->r_ctl.rc_tlp_cwnd_reduce; break; case TCP_RACK_EARLY_RECOV: /* Should recovery happen early (bool) */ optval = rack->r_ctl.rc_early_recovery; break; case TCP_RACK_PACE_REDUCE: /* RACK Hptsi reduction factor (divisor) */ optval = rack->rc_pace_reduce; break; case TCP_RACK_PACE_MAX_SEG: /* Max segments in a pace */ optval = rack->rc_pace_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_CHEAT: /* Do we use the rack cheat for rxt */ optval = rack->use_rack_cheat; 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_TLP_INC_VAR: /* Does TLP include rtt variance in t-o */ error = EINVAL; break; case TCP_RACK_IDLE_REDUCE_HIGH: error = EINVAL; break; case TCP_RACK_MIN_PACE: optval = rack->r_enforce_min_pace; break; case TCP_RACK_GP_INCREASE: optval = rack->rack_per_of_gp; 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; default: return (tcp_default_ctloutput(so, sopt, inp, tp)); break; } INP_WUNLOCK(inp); if (error == 0) { error = sooptcopyout(sopt, &optval, sizeof optval); } return (error); } static int rack_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp) { int32_t error = EINVAL; struct tcp_rack *rack; rack = (struct tcp_rack *)tp->t_fb_ptr; if (rack == NULL) { /* Huh? */ goto out; } if (sopt->sopt_dir == SOPT_SET) { return (rack_set_sockopt(so, sopt, inp, tp, rack)); } else if (sopt->sopt_dir == SOPT_GET) { return (rack_get_sockopt(so, sopt, inp, tp, rack)); } out: INP_WUNLOCK(inp); return (error); } 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 }; 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, 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); Index: head/sys/netinet/tcp_subr.c =================================================================== --- head/sys/netinet/tcp_subr.c (revision 356526) +++ head/sys/netinet/tcp_subr.c (revision 356527) @@ -1,3455 +1,3455 @@ /*- * 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 #ifdef INET6 #include #endif #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 #include #include #include #include #include #include #include #include #ifdef INET6 #include #endif #include #include #ifdef TCPPCAP #include #endif #ifdef TCPDEBUG #include #endif #ifdef INET6 #include #endif #ifdef TCP_OFFLOAD #include #endif #include #include #include #include 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, 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 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, &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, &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"); 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"); 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, &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 void tcp_mtudisc(struct inpcb *, int); static char * tcp_log_addr(struct in_conninfo *inc, struct tcphdr *th, void *ip4hdr, const void *ip6hdr); static struct tcp_function_block tcp_def_funcblk = { .tfb_tcp_block_name = "freebsd", .tfb_tcp_output = tcp_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; 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); } 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 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, 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, NULL, 0, sysctl_net_inet_list_available, "A", "list available TCP Function sets"); /* * 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)); 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) /* * TCP initialization. */ static void tcp_zone_change(void *tag) { uma_zone_set_max(V_tcbinfo.ipi_zone, maxsockets); uma_zone_set_max(V_tcpcb_zone, maxsockets); tcp_tw_zone_change(); } static int tcp_inpcb_init(void *mem, int size, int flags) { struct inpcb *inp = mem; INP_LOCK_INIT(inp, "inp", "tcpinp"); return (0); } /* * 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 *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); INP_INFO_WLOCK(&V_tcbinfo); CK_LIST_FOREACH(inp, V_tcbinfo.ipi_listhead, inp_list) { INP_WLOCK(inp); if (inp->inp_flags & INP_TIMEWAIT) { INP_WUNLOCK(inp); continue; } tp = intotcpcb(inp); if (tp == NULL || tp->t_fb != blk) { INP_WUNLOCK(inp); continue; } tcp_switch_back_to_default(tp); INP_WUNLOCK(inp); } INP_INFO_WUNLOCK(&V_tcbinfo); 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); } void tcp_init(void) { const char *tcbhash_tuneable; int hashsize; tcbhash_tuneable = "net.inet.tcp.tcbhashsize"; #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 hashsize = 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 && IS_DEFAULT_VNET(curvnet)) 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); } in_pcbinfo_init(&V_tcbinfo, "tcp", &V_tcb, hashsize, hashsize, "tcp_inpcb", tcp_inpcb_init, IPI_HASHFIELDS_4TUPLE); /* * 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"); tcp_tw_init(); 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(); /* Skip initialization of globals for non-default instances. */ if (!IS_DEFAULT_VNET(curvnet)) return; tcp_reass_global_init(); /* XXX virtualize those bellow? */ tcp_delacktime = TCPTV_DELACK; tcp_keepinit = TCPTV_KEEP_INIT; tcp_keepidle = TCPTV_KEEP_IDLE; tcp_keepintvl = TCPTV_KEEPINTVL; tcp_maxpersistidle = TCPTV_KEEP_IDLE; tcp_msl = TCPTV_MSL; 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; tcp_tcbhashsize = hashsize; /* 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_usrreqs.pru_soreceive = soreceive_stream; #endif #ifdef INET6 tcp6_usrreqs.pru_soreceive = soreceive_stream; #endif /* INET6 */ } #ifdef INET6 #define TCP_MINPROTOHDR (sizeof(struct ip6_hdr) + sizeof(struct tcphdr)) #else /* INET6 */ #define TCP_MINPROTOHDR (sizeof(struct tcpiphdr)) #endif /* INET6 */ if (max_protohdr < TCP_MINPROTOHDR) max_protohdr = TCP_MINPROTOHDR; if (max_linkhdr + TCP_MINPROTOHDR > MHLEN) panic("tcp_init"); #undef TCP_MINPROTOHDR ISN_LOCK_INIT(); EVENTHANDLER_REGISTER(shutdown_pre_sync, tcp_fini, NULL, SHUTDOWN_PRI_DEFAULT); EVENTHANDLER_REGISTER(maxsockets_change, tcp_zone_change, NULL, EVENTHANDLER_PRI_ANY); 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_single_push = counter_u64_alloc(M_WAITOK); tcp_inp_lro_locks_taken = counter_u64_alloc(M_WAITOK); tcp_inp_lro_sack_wake = counter_u64_alloc(M_WAITOK); #ifdef TCPPCAP tcp_pcap_init(); #endif } #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_LIST_RLOCK(&V_tcbinfo); n = V_tcbinfo.ipi_count; INP_LIST_RUNLOCK(&V_tcbinfo); if (n == 0) break; pause("tcpdes", hz / 10); } tcp_hc_destroy(); syncache_destroy(); tcp_tw_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(); #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, 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); ip6->ip6_nxt = IPPROTO_TCP; 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; ip->ip_p = IPPROTO_TCP; 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_x2 = 0; th->th_off = 5; th->th_flags = 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, (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 tcphdr *nth; u_char *optp; #ifdef INET6 struct ip6_hdr *ip6; int isipv6; #endif /* INET6 */ int optlen, tlen, win; bool incl_opts; KASSERT(tp != NULL || m != NULL, ("tcp_respond: tp and m both NULL")); #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_WLOCK_ASSERT(inp); } else inp = NULL; 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); } else #endif /* INET6 */ { bcopy((caddr_t)ip, mtod(m, caddr_t), sizeof(struct ip)); ip = mtod(m, struct ip *); nth = (struct tcphdr *)(ip + 1); } bcopy((caddr_t)th, (caddr_t)nth, sizeof(struct tcphdr)); flags = TH_ACK; } else if (!M_WRITABLE(m)) { 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); } 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); } 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 #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) { ip6->ip6_flow = 0; ip6->ip6_vfc = IPV6_VERSION; ip6->ip6_nxt = IPPROTO_TCP; ip6->ip6_plen = htons(tlen - sizeof(*ip6)); } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { ip->ip_len = htons(tlen); ip->ip_ttl = V_ip_defttl; 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_WLOCK_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_x2 = 0; nth->th_off = (sizeof (struct tcphdr) + optlen) >> 2; nth->th_flags = 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 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum); #ifdef INET6 if (isipv6) { m->m_pkthdr.csum_flags = CSUM_TCP_IPV6; 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 { m->m_pkthdr.csum_flags = CSUM_TCP; 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); #ifdef INET6 if (isipv6) { TCP_PROBE5(send, NULL, tp, ip6, tp, nth); (void)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); (void)ip_output(m, NULL, NULL, 0, NULL, inp); } #endif } /* * 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_LIST_RLOCK(); KASSERT(!STAILQ_EMPTY(&cc_list), ("cc_list is empty!")); CC_ALGO(tp) = CC_DEFAULT(); CC_LIST_RUNLOCK(); if (CC_ALGO(tp)->cb_init != NULL) if (CC_ALGO(tp)->cb_init(tp->ccv) > 0) { if (tp->t_fb->tfb_tcp_fb_fini) (*tp->t_fb->tfb_tcp_fb_fini)(tp, 1); 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); 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); if (V_tcp_do_rfc1323) tp->t_flags = (TF_REQ_SCALE|TF_REQ_TSTMP); if (V_tcp_do_sack) tp->t_flags |= TF_SACK_PERMIT; TAILQ_INIT(&tp->snd_holes); /* * The tcpcb will hold a reference on its inpcb until tcp_discardcb() * is called. */ in_pcbref(inp); /* Reference for tcpcb */ tp->t_inpcb = inp; /* * 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 if (tp->t_fb->tfb_tcp_fb_init) { (*tp->t_fb->tfb_tcp_fb_init)(tp); } #ifdef STATS if (V_tcp_perconn_stats_enable == 1) tp->t_stats = stats_blob_alloc(V_tcp_perconn_stats_dflt_tpl, 0); #endif return (tp); /* XXX */ } /* * Switch the congestion control algorithm back to NewReno for any active * control blocks using an algorithm which is about to go away. * This ensures the CC framework can allow the unload to proceed without leaving * any dangling pointers which would trigger a panic. * Returning non-zero would inform the CC framework that something went wrong * and it would be unsafe to allow the unload to proceed. However, there is no * way for this to occur with this implementation so we always return zero. */ int tcp_ccalgounload(struct cc_algo *unload_algo) { struct cc_algo *tmpalgo; struct inpcb *inp; struct tcpcb *tp; VNET_ITERATOR_DECL(vnet_iter); /* * Check all active control blocks across all network stacks and change * any that are using "unload_algo" back to NewReno. If "unload_algo" * requires cleanup code to be run, call it. */ VNET_LIST_RLOCK(); VNET_FOREACH(vnet_iter) { CURVNET_SET(vnet_iter); INP_INFO_WLOCK(&V_tcbinfo); /* * New connections already part way through being initialised * with the CC algo we're removing will not race with this code * because the INP_INFO_WLOCK is held during initialisation. We * therefore don't enter the loop below until the connection * list has stabilised. */ CK_LIST_FOREACH(inp, &V_tcb, inp_list) { INP_WLOCK(inp); /* Important to skip tcptw structs. */ if (!(inp->inp_flags & INP_TIMEWAIT) && (tp = intotcpcb(inp)) != NULL) { /* * By holding INP_WLOCK here, we are assured * that the connection is not currently * executing inside the CC module's functions * i.e. it is safe to make the switch back to * NewReno. */ if (CC_ALGO(tp) == unload_algo) { tmpalgo = CC_ALGO(tp); if (tmpalgo->cb_destroy != NULL) tmpalgo->cb_destroy(tp->ccv); CC_DATA(tp) = NULL; /* * NewReno may allocate memory on * demand for certain stateful * configuration as needed, but is * coded to never fail on memory * allocation failure so it is a safe * fallback. */ CC_ALGO(tp) = &newreno_cc_algo; } } INP_WUNLOCK(inp); } INP_INFO_WUNLOCK(&V_tcbinfo); CURVNET_RESTORE(); } VNET_LIST_RUNLOCK(); return (0); } /* * 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; INP_INFO_LOCK_ASSERT(&V_tcbinfo); INP_WLOCK_ASSERT(tp->t_inpcb); if (TCPS_HAVERCVDSYN(tp->t_state)) { tcp_state_change(tp, TCPS_CLOSED); (void) tp->t_fb->tfb_tcp_output(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; struct socket *so = inp->inp_socket; #ifdef INET6 int isipv6 = (inp->inp_vflag & INP_IPV6) != 0; #endif /* INET6 */ int released __unused; 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); } /* * 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! */ if (tp->t_rttupdated >= 4) { 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); } /* 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; #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) { /* 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); refcount_release(&tp->t_fb->tfb_refcnt); tp->t_inpcb = NULL; uma_zfree(V_tcpcb_zone, tp); released = in_pcbrele_wlocked(inp); KASSERT(!released, ("%s: inp %p should not have been released " "here", __func__, inp)); } } 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__)); tp->t_timers->tt_draincnt--; if (tp->t_timers->tt_draincnt == 0) { /* We own the last reference on this 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); refcount_release(&tp->t_fb->tfb_refcnt); tp->t_inpcb = NULL; uma_zfree(V_tcpcb_zone, tp); if (in_pcbrele_wlocked(inp)) { NET_EPOCH_EXIT(et); CURVNET_RESTORE(); return; } } INP_WUNLOCK(inp); NET_EPOCH_EXIT(et); CURVNET_RESTORE(); } /* * 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_INFO_LOCK_ASSERT(&V_tcbinfo); 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; } 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) { KASSERT(so->so_state & SS_PROTOREF, ("tcp_close: !SS_PROTOREF")); inp->inp_flags &= ~INP_SOCKREF; INP_WUNLOCK(inp); SOCK_LOCK(so); so->so_state &= ~SS_PROTOREF; sofree(so); return (NULL); } return (tp); } void tcp_drain(void) { VNET_ITERATOR_DECL(vnet_iter); if (!do_tcpdrain) return; VNET_LIST_RLOCK_NOSLEEP(); VNET_FOREACH(vnet_iter) { CURVNET_SET(vnet_iter); 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. */ INP_INFO_WLOCK(&V_tcbinfo); CK_LIST_FOREACH(inpb, V_tcbinfo.ipi_listhead, inp_list) { INP_WLOCK(inpb); if (inpb->inp_flags & INP_TIMEWAIT) { INP_WUNLOCK(inpb); 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 } INP_WUNLOCK(inpb); } INP_INFO_WUNLOCK(&V_tcbinfo); CURVNET_RESTORE(); } VNET_LIST_RUNLOCK_NOSLEEP(); } /* * 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_INFO_LOCK_ASSERT(&V_tcbinfo); INP_WLOCK_ASSERT(inp); if ((inp->inp_flags & INP_TIMEWAIT) || (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_rt) { RTFREE(inp->inp_route.ro_rt); inp->inp_route.ro_rt = (struct rtentry *)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 epoch_tracker et; struct inpcb *inp; struct xinpgen xig; 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); NET_EPOCH_ENTER(et); for (inp = CK_LIST_FIRST(V_tcbinfo.ipi_listhead); inp != NULL; inp = CK_LIST_NEXT(inp, inp_list)) { INP_RLOCK(inp); if (inp->inp_gencnt <= xig.xig_gen) { /* * XXX: This use of cr_cansee(), introduced with * TCP state changes, is not quite right, but for * now, better than nothing. */ if (inp->inp_flags & INP_TIMEWAIT) { if (intotw(inp) != NULL) error = cr_cansee(req->td->td_ucred, intotw(inp)->tw_cred); else error = EINVAL; /* Skip this inp. */ } else error = cr_canseeinpcb(req->td->td_ucred, inp); if (error == 0) { struct xtcpcb xt; tcp_inptoxtp(inp, &xt); INP_RUNLOCK(inp); error = SYSCTL_OUT(req, &xt, sizeof xt); if (error) break; else continue; } } INP_RUNLOCK(inp); } NET_EPOCH_EXIT(et); 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, 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, 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, 0, 0, tcp6_getcred, "S,xucred", "Get the xucred of a TCP6 connection"); #endif /* INET6 */ #ifdef INET void tcp_ctlinput(int cmd, struct sockaddr *sa, void *vip) { struct ip *ip = vip; struct tcphdr *th; struct in_addr faddr; struct inpcb *inp; struct tcpcb *tp; struct inpcb *(*notify)(struct inpcb *, int) = tcp_notify; struct icmp *icp; struct in_conninfo inc; tcp_seq icmp_tcp_seq; int mtu; faddr = ((struct sockaddr_in *)sa)->sin_addr; if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY) return; if (cmd == PRC_MSGSIZE) notify = tcp_mtudisc_notify; else if (V_icmp_may_rst && (cmd == PRC_UNREACH_ADMIN_PROHIB || cmd == PRC_UNREACH_PORT || cmd == PRC_UNREACH_PROTOCOL || cmd == PRC_TIMXCEED_INTRANS) && ip) notify = tcp_drop_syn_sent; /* * Hostdead is ugly because it goes linearly through all PCBs. * XXX: We never get this from ICMP, otherwise it makes an * excellent DoS attack on machines with many connections. */ else if (cmd == PRC_HOSTDEAD) ip = NULL; else if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) return; if (ip == NULL) { in_pcbnotifyall(&V_tcbinfo, faddr, inetctlerrmap[cmd], notify); return; } icp = (struct icmp *)((caddr_t)ip - offsetof(struct icmp, icmp_ip)); th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2)); inp = in_pcblookup(&V_tcbinfo, faddr, th->th_dport, ip->ip_src, th->th_sport, INPLOOKUP_WLOCKPCB, NULL); if (inp != NULL && PRC_IS_REDIRECT(cmd)) { /* signal EHOSTDOWN, as it flushes the cached route */ inp = (*notify)(inp, EHOSTDOWN); goto out; } icmp_tcp_seq = th->th_seq; if (inp != NULL) { if (!(inp->inp_flags & INP_TIMEWAIT) && !(inp->inp_flags & INP_DROPPED) && !(inp->inp_socket == NULL)) { tp = intotcpcb(inp); if (SEQ_GEQ(ntohl(icmp_tcp_seq), tp->snd_una) && SEQ_LT(ntohl(icmp_tcp_seq), tp->snd_max)) { if (cmd == PRC_MSGSIZE) { /* * MTU discovery: * If we got a needfrag set the MTU * in the route to the suggested new * value (if given) and then notify. */ 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); /* * 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 = faddr; inc.inc_fibnum = inp->inp_inc.inc_fibnum; tcp_hc_updatemtu(&inc, mtu); tcp_mtudisc(inp, mtu); } } else inp = (*notify)(inp, inetctlerrmap[cmd]); } } } else { bzero(&inc, sizeof(inc)); inc.inc_fport = th->th_dport; inc.inc_lport = th->th_sport; inc.inc_faddr = faddr; inc.inc_laddr = ip->ip_src; syncache_unreach(&inc, icmp_tcp_seq); } out: if (inp != NULL) INP_WUNLOCK(inp); } #endif /* INET */ #ifdef INET6 void tcp6_ctlinput(int cmd, struct sockaddr *sa, void *d) { struct in6_addr *dst; struct inpcb *(*notify)(struct inpcb *, int) = tcp_notify; struct ip6_hdr *ip6; struct mbuf *m; struct inpcb *inp; struct tcpcb *tp; struct icmp6_hdr *icmp6; struct ip6ctlparam *ip6cp = NULL; const struct sockaddr_in6 *sa6_src = NULL; 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; if (sa->sa_family != AF_INET6 || sa->sa_len != sizeof(struct sockaddr_in6)) return; /* if the parameter is from icmp6, decode it. */ if (d != NULL) { ip6cp = (struct ip6ctlparam *)d; icmp6 = ip6cp->ip6c_icmp6; m = ip6cp->ip6c_m; ip6 = ip6cp->ip6c_ip6; off = ip6cp->ip6c_off; sa6_src = ip6cp->ip6c_src; dst = ip6cp->ip6c_finaldst; } else { m = NULL; ip6 = NULL; off = 0; /* fool gcc */ sa6_src = &sa6_any; dst = NULL; } if (cmd == PRC_MSGSIZE) notify = tcp_mtudisc_notify; else if (V_icmp_may_rst && (cmd == PRC_UNREACH_ADMIN_PROHIB || cmd == PRC_UNREACH_PORT || cmd == PRC_UNREACH_PROTOCOL || cmd == PRC_TIMXCEED_INTRANS) && ip6 != NULL) notify = tcp_drop_syn_sent; /* * Hostdead is ugly because it goes linearly through all PCBs. * XXX: We never get this from ICMP, otherwise it makes an * excellent DoS attack on machines with many connections. */ else if (cmd == PRC_HOSTDEAD) ip6 = NULL; else if ((unsigned)cmd >= PRC_NCMDS || inet6ctlerrmap[cmd] == 0) return; if (ip6 == NULL) { in6_pcbnotify(&V_tcbinfo, sa, 0, (const struct sockaddr *)sa6_src, 0, cmd, NULL, notify); return; } /* 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); if (inp != NULL && PRC_IS_REDIRECT(cmd)) { /* signal EHOSTDOWN, as it flushes the cached route */ inp = (*notify)(inp, EHOSTDOWN); goto out; } 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) && !(inp->inp_socket == NULL)) { tp = intotcpcb(inp); if (SEQ_GEQ(ntohl(icmp_tcp_seq), tp->snd_una) && SEQ_LT(ntohl(icmp_tcp_seq), tp->snd_max)) { if (cmd == PRC_MSGSIZE) { /* * MTU discovery: * If we got a needfrag set the MTU * in the route to the suggested new * value (if given) and then notify. */ 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; 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, inet6ctlerrmap[cmd]); } } } 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); } out: if (inp != NULL) INP_WUNLOCK(inp); } #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. */ 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)) 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) { tcp_mtudisc(inp, -1); return (inp); } static void 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)) return; 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); tp->t_fb->tfb_tcp_output(tp); } #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 nhop4_extended nh4; 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) { if (fib4_lookup_nh_ext(inc->inc_fibnum, inc->inc_faddr, NHR_REF, 0, &nh4) != 0) return (0); ifp = nh4.nh_ifp; maxmtu = nh4.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; } } fib4_free_nh_ext(inc->inc_fibnum, &nh4); } return (maxmtu); } #endif /* INET */ #ifdef INET6 uint32_t tcp_maxmtu6(struct in_conninfo *inc, struct tcp_ifcap *cap) { struct nhop6_extended nh6; 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); if (fib6_lookup_nh_ext(inc->inc_fibnum, &dst6, scopeid, 0, 0, &nh6) != 0) return (0); ifp = nh6.nh_ifp; maxmtu = nh6.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; } } fib6_free_nh_ext(inc->inc_fibnum, &nh6); } return (maxmtu); } #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. */ #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 if ((tp->t_flags & TF_SACK_PERMIT) && tp->rcv_numsacks > 0) { optlen += TCPOLEN_SACKHDR; optlen += tp->rcv_numsacks * TCPOLEN_SACK; optlen = PAD(optlen); } } 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; struct tcptw *tw; struct sockaddr_in *fin, *lin; struct epoch_tracker et; #ifdef INET6 struct sockaddr_in6 *fin6, *lin6; #endif int error; inp = NULL; fin = lin = 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]); fin = (struct sockaddr_in *)&addrs[0]; lin = (struct sockaddr_in *)&addrs[1]; 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_TIMEWAIT) { /* * XXXRW: There currently exists a state where an * inpcb is present, but its timewait state has been * discarded. For now, don't allow dropping of this * type of inpcb. */ tw = intotw(inp); if (tw != NULL) tcp_twclose(tw, 0); else INP_WUNLOCK(inp); } else if (!(inp->inp_flags & INP_DROPPED) && !(inp->inp_socket->so_options & SO_ACCEPTCONN)) { 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, NULL, 0, sysctl_drop, "", "Drop TCP connection"); #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; struct sockaddr_in *fin, *lin; struct epoch_tracker et; #ifdef INET6 struct sockaddr_in6 *fin6, *lin6; #endif int error; inp = NULL; fin = lin = 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]); fin = (struct sockaddr_in *)&addrs[0]; lin = (struct sockaddr_in *)&addrs[1]; 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 || 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); SOCK_LOCK(so); 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, 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, 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, void *ip4hdr, const void *ip6hdr) { /* Is logging enabled? */ - if (tcp_log_in_vain == 0) + 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, 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, void *ip4hdr, const void *ip6hdr) { char *s, *sp; size_t size; struct ip *ip; #ifdef INET6 const struct ip6_hdr *ip6; ip6 = (const struct ip6_hdr *)ip6hdr; #endif /* INET6 */ ip = (struct ip *)ip4hdr; /* * 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", th->th_flags, 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)); if (inp->inp_flags & INP_TIMEWAIT) { xt->t_state = TCPS_TIME_WAIT; } else { 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; 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; bcopy(tp->t_fb->tfb_tcp_block_name, xt->xt_stack, TCP_FUNCTION_NAME_LEN_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; } Index: head/sys/netinet/tcp_timer.c =================================================================== --- head/sys/netinet/tcp_timer.c (revision 356526) +++ head/sys/netinet/tcp_timer.c (revision 356527) @@ -1,1066 +1,1067 @@ /*- * 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 #ifdef TCPDEBUG #include #endif int tcp_persmin; SYSCTL_PROC(_net_inet_tcp, OID_AUTO, persmin, CTLTYPE_INT|CTLFLAG_RW, &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, &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, &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, &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, &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, &tcp_delacktime, 0, sysctl_msec_to_ticks, "I", "Time before a delayed ACK is sent"); int tcp_msl; SYSCTL_PROC(_net_inet_tcp, OID_AUTO, msl, CTLTYPE_INT|CTLFLAG_RW, &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, &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, &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, &tcp_rexmit_slop, 0, sysctl_msec_to_ticks, "I", "Retransmission Timer Slop"); -int tcp_always_keepalive = 1; -SYSCTL_INT(_net_inet_tcp, OID_AUTO, always_keepalive, CTLFLAG_RW, - &tcp_always_keepalive , 0, "Assume SO_KEEPALIVE on all TCP connections"); +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, &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"); 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; #ifdef RSS if (per_cpu_timers) { cpuid = rss_hash2cpuid(inp->inp_flowid, inp->inp_flowtype); if (cpuid == NETISR_CPUID_NONE) return (curcpu); /* XXX */ else return (cpuid); } #else /* Legacy, pre-RSS behaviour */ if (per_cpu_timers) { /* * 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); } #endif /* Default for RSS and non-RSS - cpuid 0 */ else { return (0); } } /* * Tcp protocol timeout routine called every 500 ms. * Updates timestamps used for TCP * causes finite state machine actions if timers expire. */ void tcp_slowtimo(void) { VNET_ITERATOR_DECL(vnet_iter); VNET_LIST_RLOCK_NOSLEEP(); VNET_FOREACH(vnet_iter) { CURVNET_SET(vnet_iter); (void) tcp_tw_2msl_scan(0); CURVNET_RESTORE(); } VNET_LIST_RUNLOCK_NOSLEEP(); } 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 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); (void) tp->t_fb->tfb_tcp_output(tp); INP_WUNLOCK(inp); CURVNET_RESTORE(); } void tcp_inpinfo_lock_del(struct inpcb *inp, struct tcpcb *tp) { if (inp && tp != NULL) INP_WUNLOCK(inp); } void tcp_timer_2msl(void *xtp) { struct tcpcb *tp = xtp; 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); 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_DROPPED) != 0) { 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 in TIME_WAIT state just ignore as this timeout is handled in * tcp_tw_2msl_scan(). * * 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 ((inp->inp_flags & INP_TIMEWAIT) != 0) { INP_WUNLOCK(inp); CURVNET_RESTORE(); return; } if (tcp_fast_finwait2_recycle && tp->t_state == TCPS_FIN_WAIT_2 && tp->t_inpcb && tp->t_inpcb->inp_socket && (tp->t_inpcb->inp_socket->so_rcv.sb_state & SBS_CANTRCVMORE)) { TCPSTAT_INC(tcps_finwait2_drops); if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { tcp_inpinfo_lock_del(inp, tp); goto out; } NET_EPOCH_ENTER(et); tp = tcp_close(tp); NET_EPOCH_EXIT(et); tcp_inpinfo_lock_del(inp, tp); goto out; } else { if (ticks - tp->t_rcvtime <= TP_MAXIDLE(tp)) { callout_reset(&tp->t_timers->tt_2msl, TP_KEEPINTVL(tp), tcp_timer_2msl, tp); } else { if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { tcp_inpinfo_lock_del(inp, tp); goto out; } NET_EPOCH_ENTER(et); tp = tcp_close(tp); NET_EPOCH_EXIT(et); tcp_inpinfo_lock_del(inp, tp); goto out; } } #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); if (tp != NULL) INP_WUNLOCK(inp); out: 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_DROPPED) != 0) { 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 ((tcp_always_keepalive || + 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) { 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); } 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); if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { tcp_inpinfo_lock_del(inp, tp); goto out; } NET_EPOCH_ENTER(et); 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); tcp_inpinfo_lock_del(inp, tp); out: CURVNET_RESTORE(); } void tcp_timer_persist(void *xtp) { struct tcpcb *tp = xtp; 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_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_DROPPED) != 0) { 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. */ if (tp->t_rxtshift == TCP_MAXRXTSHIFT && (ticks - tp->t_rcvtime >= tcp_maxpersistidle || ticks - tp->t_rcvtime >= TCP_REXMTVAL(tp) * tcp_totbackoff)) { TCPSTAT_INC(tcps_persistdrop); if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { tcp_inpinfo_lock_del(inp, tp); goto out; } NET_EPOCH_ENTER(et); tp = tcp_drop(tp, ETIMEDOUT); NET_EPOCH_EXIT(et); tcp_inpinfo_lock_del(inp, tp); 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) { TCPSTAT_INC(tcps_persistdrop); if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { tcp_inpinfo_lock_del(inp, tp); goto out; } NET_EPOCH_ENTER(et); tp = tcp_drop(tp, ETIMEDOUT); NET_EPOCH_EXIT(et); tcp_inpinfo_lock_del(inp, tp); goto out; } tcp_setpersist(tp); tp->t_flags |= TF_FORCEDATA; (void) tp->t_fb->tfb_tcp_output(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); INP_WUNLOCK(inp); out: CURVNET_RESTORE(); } void tcp_timer_rexmt(void * xtp) { struct tcpcb *tp = xtp; CURVNET_SET(tp->t_vnet); int rexmt; struct inpcb *inp; struct epoch_tracker et; #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_DROPPED) != 0) { 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 (++tp->t_rxtshift > TCP_MAXRXTSHIFT) { tp->t_rxtshift = TCP_MAXRXTSHIFT; TCPSTAT_INC(tcps_timeoutdrop); if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { tcp_inpinfo_lock_del(inp, tp); goto out; } NET_EPOCH_ENTER(et); tp = tcp_drop(tp, ETIMEDOUT); NET_EPOCH_EXIT(et); tcp_inpinfo_lock_del(inp, tp); goto out; } 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. */ if (V_tcp_pmtud_blackhole_detect && (((tp->t_state == TCPS_ESTABLISHED)) || (tp->t_state == TCPS_FIN_WAIT_1))) { #ifdef INET6 int isipv6; #endif /* * 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 = (tp->t_inpcb->inp_vflag & INP_IPV6) ? 1 : 0; if (isipv6 && tp->t_maxseg > V_tcp_v6pmtud_blackhole_mss) { /* 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) { /* 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. * 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; 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); (void) tp->t_fb->tfb_tcp_output(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); INP_WUNLOCK(inp); out: 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); } } 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++; } } Index: head/sys/netinet/tcp_timer.h =================================================================== --- head/sys/netinet/tcp_timer.h (revision 356526) +++ head/sys/netinet/tcp_timer.h (revision 356527) @@ -1,231 +1,232 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1982, 1986, 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. * * @(#)tcp_timer.h 8.1 (Berkeley) 6/10/93 * $FreeBSD$ */ #ifndef _NETINET_TCP_TIMER_H_ #define _NETINET_TCP_TIMER_H_ /* * The TCPT_REXMT timer is used to force retransmissions. * The TCP has the TCPT_REXMT timer set whenever segments * have been sent for which ACKs are expected but not yet * received. If an ACK is received which advances tp->snd_una, * then the retransmit timer is cleared (if there are no more * outstanding segments) or reset to the base value (if there * are more ACKs expected). Whenever the retransmit timer goes off, * we retransmit one unacknowledged segment, and do a backoff * on the retransmit timer. * * The TCPT_PERSIST timer is used to keep window size information * flowing even if the window goes shut. If all previous transmissions * have been acknowledged (so that there are no retransmissions in progress), * and the window is too small to bother sending anything, then we start * the TCPT_PERSIST timer. When it expires, if the window is nonzero, * we go to transmit state. Otherwise, at intervals send a single byte * into the peer's window to force him to update our window information. * We do this at most as often as TCPT_PERSMIN time intervals, * but no more frequently than the current estimate of round-trip * packet time. The TCPT_PERSIST timer is cleared whenever we receive * a window update from the peer. * * The TCPT_KEEP timer is used to keep connections alive. If an * connection is idle (no segments received) for TCPTV_KEEP_INIT amount of time, * but not yet established, then we drop the connection. Once the connection * is established, if the connection is idle for TCPTV_KEEP_IDLE time * (and keepalives have been enabled on the socket), we begin to probe * the connection. We force the peer to send us a segment by sending: * * This segment is (deliberately) outside the window, and should elicit * an ack segment in response from the peer. If, despite the TCPT_KEEP * initiated segments we cannot elicit a response from a peer in TCPT_MAXIDLE * amount of time probing, then we drop the connection. */ /* * Time constants. */ #define TCPTV_MSL ( 30*hz) /* max seg lifetime (hah!) */ #define TCPTV_SRTTBASE 0 /* base roundtrip time; if 0, no idea yet */ #define TCPTV_RTOBASE ( 1*hz) /* assumed RTO if no info */ #define TCPTV_PERSMIN ( 5*hz) /* minimum persist interval */ #define TCPTV_PERSMAX ( 60*hz) /* maximum persist interval */ #define TCPTV_KEEP_INIT ( 75*hz) /* initial connect keepalive */ #define TCPTV_KEEP_IDLE (120*60*hz) /* dflt time before probing */ #define TCPTV_KEEPINTVL ( 75*hz) /* default probe interval */ #define TCPTV_KEEPCNT 8 /* max probes before drop */ #define TCPTV_FINWAIT2_TIMEOUT (60*hz) /* FIN_WAIT_2 timeout if no receiver */ /* * Minimum retransmit timer is 3 ticks, for algorithmic stability. * TCPT_RANGESET() will add another TCPTV_CPU_VAR to deal with * the expected worst-case processing variances by the kernels * representing the end points. Such variances do not always show * up in the srtt because the timestamp is often calculated at * the interface rather then at the TCP layer. This value is * typically 50ms. However, it is also possible that delayed * acks (typically 100ms) could create issues so we set the slop * to 200ms to try to cover it. Note that, properly speaking, * delayed-acks should not create a major issue for interactive * environments which 'P'ush the last segment, at least as * long as implementations do the required 'at least one ack * for every two packets' for the non-interactive streaming case. * (maybe the RTO calculation should use 2*RTT instead of RTT * to handle the ack-every-other-packet case). * * The prior minimum of 1*hz (1 second) badly breaks throughput on any * networks faster then a modem that has minor (e.g. 1%) packet loss. */ #define TCPTV_MIN ( hz/33 ) /* minimum allowable value */ #define TCPTV_CPU_VAR ( hz/5 ) /* cpu variance allowed (200ms) */ #define TCPTV_REXMTMAX ( 64*hz) /* max allowable REXMT value */ #define TCPTV_TWTRUNC 8 /* RTO factor to truncate TW */ #define TCP_LINGERTIME 120 /* linger at most 2 minutes */ #define TCP_MAXRXTSHIFT 12 /* maximum retransmits */ #define TCPTV_DELACK ( hz/10 ) /* 100ms timeout */ /* * If we exceed this number of retransmits for a single segment, we'll consider * the current srtt measurement no longer valid and will recalculate from * scratch starting with the next ACK. */ #define TCP_RTT_INVALIDATE (TCP_MAXRXTSHIFT / 4) #ifdef TCPTIMERS static const char *tcptimers[] = { "REXMT", "PERSIST", "KEEP", "2MSL", "DELACK" }; #endif /* * Force a time value to be in a certain range. */ #define TCPT_RANGESET(tv, value, tvmin, tvmax) do { \ (tv) = (value) + tcp_rexmit_slop; \ if ((u_long)(tv) < (u_long)(tvmin)) \ (tv) = (tvmin); \ if ((u_long)(tv) > (u_long)(tvmax)) \ (tv) = (tvmax); \ } while(0) #ifdef _KERNEL struct xtcp_timer; struct tcp_timer { struct callout tt_rexmt; /* retransmit timer */ struct callout tt_persist; /* retransmit persistence */ struct callout tt_keep; /* keepalive */ struct callout tt_2msl; /* 2*msl TIME_WAIT timer */ struct callout tt_delack; /* delayed ACK timer */ uint32_t tt_flags; /* Timers flags */ uint32_t tt_draincnt; /* Count being drained */ }; /* * Flags for the tt_flags field. */ #define TT_DELACK 0x0001 #define TT_REXMT 0x0002 #define TT_PERSIST 0x0004 #define TT_KEEP 0x0008 #define TT_2MSL 0x0010 #define TT_MASK (TT_DELACK|TT_REXMT|TT_PERSIST|TT_KEEP|TT_2MSL) /* * Suspend flags - used when suspending a timer * from ever running again. */ #define TT_DELACK_SUS 0x0100 #define TT_REXMT_SUS 0x0200 #define TT_PERSIST_SUS 0x0400 #define TT_KEEP_SUS 0x0800 #define TT_2MSL_SUS 0x1000 #define TT_STOPPED 0x00010000 #define TP_KEEPINIT(tp) ((tp)->t_keepinit ? (tp)->t_keepinit : tcp_keepinit) #define TP_KEEPIDLE(tp) ((tp)->t_keepidle ? (tp)->t_keepidle : tcp_keepidle) #define TP_KEEPINTVL(tp) ((tp)->t_keepintvl ? (tp)->t_keepintvl : tcp_keepintvl) #define TP_KEEPCNT(tp) ((tp)->t_keepcnt ? (tp)->t_keepcnt : tcp_keepcnt) #define TP_MAXIDLE(tp) (TP_KEEPCNT(tp) * TP_KEEPINTVL(tp)) extern int tcp_persmin; /* minimum persist interval */ extern int tcp_persmax; /* maximum persist interval */ extern int tcp_keepinit; /* time to establish connection */ extern int tcp_keepidle; /* time before keepalive probes begin */ extern int tcp_keepintvl; /* time between keepalive probes */ extern int tcp_keepcnt; /* number of keepalives */ extern int tcp_delacktime; /* time before sending a delayed ACK */ extern int tcp_maxpersistidle; extern int tcp_rexmit_initial; extern int tcp_rexmit_min; extern int tcp_rexmit_slop; extern int tcp_msl; extern int tcp_ttl; /* time to live for TCP segs */ extern int tcp_backoff[]; extern int tcp_totbackoff; extern int tcp_rexmit_drop_options; -extern int tcp_always_keepalive; extern int tcp_finwait2_timeout; extern int tcp_fast_finwait2_recycle; +VNET_DECLARE(int, tcp_always_keepalive); +#define V_tcp_always_keepalive VNET(tcp_always_keepalive) VNET_DECLARE(int, tcp_pmtud_blackhole_detect); #define V_tcp_pmtud_blackhole_detect VNET(tcp_pmtud_blackhole_detect) VNET_DECLARE(int, tcp_pmtud_blackhole_mss); #define V_tcp_pmtud_blackhole_mss VNET(tcp_pmtud_blackhole_mss) VNET_DECLARE(int, tcp_v6pmtud_blackhole_mss); #define V_tcp_v6pmtud_blackhole_mss VNET(tcp_v6pmtud_blackhole_mss) void tcp_inpinfo_lock_del(struct inpcb *inp, struct tcpcb *tp); void tcp_timer_init(void); void tcp_timer_2msl(void *xtp); void tcp_timer_discard(void *); struct tcptw * tcp_tw_2msl_scan(int reuse); /* XXX temporary? */ void tcp_timer_keep(void *xtp); void tcp_timer_persist(void *xtp); void tcp_timer_rexmt(void *xtp); void tcp_timer_delack(void *xtp); #endif /* _KERNEL */ #endif /* !_NETINET_TCP_TIMER_H_ */ Index: head/sys/netinet/tcp_var.h =================================================================== --- head/sys/netinet/tcp_var.h (revision 356526) +++ head/sys/netinet/tcp_var.h (revision 356527) @@ -1,1033 +1,1034 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1982, 1986, 1993, 1994, 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_var.h 8.4 (Berkeley) 5/24/95 * $FreeBSD$ */ #ifndef _NETINET_TCP_VAR_H_ #define _NETINET_TCP_VAR_H_ #include #include #ifdef _KERNEL #include #include #endif #if defined(_KERNEL) || defined(_WANT_TCPCB) /* TCP segment queue entry */ struct tseg_qent { TAILQ_ENTRY(tseg_qent) tqe_q; struct mbuf *tqe_m; /* mbuf contains packet */ struct mbuf *tqe_last; /* last mbuf in chain */ tcp_seq tqe_start; /* TCP Sequence number start */ int tqe_len; /* TCP segment data length */ uint32_t tqe_flags; /* The flags from the th->th_flags */ uint32_t tqe_mbuf_cnt; /* Count of mbuf overhead */ }; TAILQ_HEAD(tsegqe_head, tseg_qent); struct sackblk { tcp_seq start; /* start seq no. of sack block */ tcp_seq end; /* end seq no. */ }; struct sackhole { tcp_seq start; /* start seq no. of hole */ tcp_seq end; /* end seq no. */ tcp_seq rxmit; /* next seq. no in hole to be retransmitted */ TAILQ_ENTRY(sackhole) scblink; /* scoreboard linkage */ }; struct sackhint { struct sackhole *nexthole; int sack_bytes_rexmit; tcp_seq last_sack_ack; /* Most recent/largest sacked ack */ int ispare; /* explicit pad for 64bit alignment */ int sacked_bytes; /* * Total sacked bytes reported by the * receiver via sack option */ uint32_t _pad1[1]; /* TBD */ uint64_t _pad[1]; /* TBD */ }; #define SEGQ_EMPTY(tp) TAILQ_EMPTY(&(tp)->t_segq) STAILQ_HEAD(tcp_log_stailq, tcp_log_mem); /* * Tcp control block, one per tcp; fields: * Organized for 64 byte cacheline efficiency based * on common tcp_input/tcp_output processing. */ struct tcpcb { /* Cache line 1 */ struct inpcb *t_inpcb; /* back pointer to internet pcb */ struct tcp_function_block *t_fb;/* TCP function call block */ void *t_fb_ptr; /* Pointer to t_fb specific data */ uint32_t t_maxseg:24, /* maximum segment size */ t_logstate:8; /* State of "black box" logging */ uint32_t t_port:16, /* Tunneling (over udp) port */ t_state:4, /* state of this connection */ t_idle_reduce : 1, t_delayed_ack: 7, /* Delayed ack variable */ t_fin_is_rst: 1, /* Are fin's treated as resets */ t_log_state_set: 1, bits_spare : 2; u_int t_flags; tcp_seq snd_una; /* sent but unacknowledged */ tcp_seq snd_max; /* highest sequence number sent; * used to recognize retransmits */ tcp_seq snd_nxt; /* send next */ tcp_seq snd_up; /* send urgent pointer */ uint32_t snd_wnd; /* send window */ uint32_t snd_cwnd; /* congestion-controlled window */ uint32_t t_peakrate_thr; /* pre-calculated peak rate threshold */ /* Cache line 2 */ u_int32_t ts_offset; /* our timestamp offset */ u_int32_t rfbuf_ts; /* recv buffer autoscaling timestamp */ int rcv_numsacks; /* # distinct sack blks present */ u_int t_tsomax; /* TSO total burst length limit in bytes */ u_int t_tsomaxsegcount; /* TSO maximum segment count */ u_int t_tsomaxsegsize; /* TSO maximum segment size in bytes */ tcp_seq rcv_nxt; /* receive next */ tcp_seq rcv_adv; /* advertised window */ uint32_t rcv_wnd; /* receive window */ u_int t_flags2; /* More tcpcb flags storage */ int t_srtt; /* smoothed round-trip time */ int t_rttvar; /* variance in round-trip time */ u_int32_t ts_recent; /* timestamp echo data */ u_char snd_scale; /* window scaling for send window */ u_char rcv_scale; /* window scaling for recv window */ u_char snd_limited; /* segments limited transmitted */ u_char request_r_scale; /* pending window scaling */ tcp_seq last_ack_sent; u_int t_rcvtime; /* inactivity time */ /* Cache line 3 */ tcp_seq rcv_up; /* receive urgent pointer */ int t_segqlen; /* segment reassembly queue length */ uint32_t t_segqmbuflen; /* Count of bytes mbufs on all entries */ struct tsegqe_head t_segq; /* segment reassembly queue */ struct mbuf *t_in_pkt; struct mbuf *t_tail_pkt; struct tcp_timer *t_timers; /* All the TCP timers in one struct */ struct vnet *t_vnet; /* back pointer to parent vnet */ uint32_t snd_ssthresh; /* snd_cwnd size threshold for * for slow start exponential to * linear switch */ tcp_seq snd_wl1; /* window update seg seq number */ /* Cache line 4 */ tcp_seq snd_wl2; /* window update seg ack number */ tcp_seq irs; /* initial receive sequence number */ tcp_seq iss; /* initial send sequence number */ u_int t_acktime; u_int ts_recent_age; /* when last updated */ tcp_seq snd_recover; /* for use in NewReno Fast Recovery */ uint16_t cl4_spare; /* Spare to adjust CL 4 */ char t_oobflags; /* have some */ char t_iobc; /* input character */ int t_rxtcur; /* current retransmit value (ticks) */ int t_rxtshift; /* log(2) of rexmt exp. backoff */ u_int t_rtttime; /* RTT measurement start time */ tcp_seq t_rtseq; /* sequence number being timed */ u_int t_starttime; /* time connection was established */ u_int t_pmtud_saved_maxseg; /* pre-blackhole MSS */ u_int t_rttmin; /* minimum rtt allowed */ u_int t_rttbest; /* best rtt we've seen */ int t_softerror; /* possible error not yet reported */ uint32_t max_sndwnd; /* largest window peer has offered */ /* Cache line 5 */ uint32_t snd_cwnd_prev; /* cwnd prior to retransmit */ uint32_t snd_ssthresh_prev; /* ssthresh prior to retransmit */ tcp_seq snd_recover_prev; /* snd_recover prior to retransmit */ int t_sndzerowin; /* zero-window updates sent */ u_long t_rttupdated; /* number of times rtt sampled */ int snd_numholes; /* number of holes seen by sender */ u_int t_badrxtwin; /* window for retransmit recovery */ TAILQ_HEAD(sackhole_head, sackhole) snd_holes; /* SACK scoreboard (sorted) */ tcp_seq snd_fack; /* last seq number(+1) sack'd by rcv'r*/ tcp_seq sack_newdata; /* New data xmitted in this recovery episode starts at this seq number */ struct sackblk sackblks[MAX_SACK_BLKS]; /* seq nos. of sack blocks */ struct sackhint sackhint; /* SACK scoreboard hint */ int t_rttlow; /* smallest observerved RTT */ int rfbuf_cnt; /* recv buffer autoscaling byte count */ struct toedev *tod; /* toedev handling this connection */ int t_sndrexmitpack; /* retransmit packets sent */ int t_rcvoopack; /* out-of-order packets received */ void *t_toe; /* TOE pcb pointer */ struct cc_algo *cc_algo; /* congestion control algorithm */ struct cc_var *ccv; /* congestion control specific vars */ struct osd *osd; /* storage for Khelp module data */ int t_bytes_acked; /* # bytes acked during current RTT */ u_int t_maxunacktime; u_int t_keepinit; /* time to establish connection */ u_int t_keepidle; /* time before keepalive probes begin */ u_int t_keepintvl; /* interval between keepalives */ u_int t_keepcnt; /* number of keepalives before close */ int t_dupacks; /* consecutive dup acks recd */ int t_lognum; /* Number of log entries */ int t_loglimit; /* Maximum number of log entries */ struct tcp_log_stailq t_logs; /* Log buffer */ struct tcp_log_id_node *t_lin; struct tcp_log_id_bucket *t_lib; const char *t_output_caller; /* Function that called tcp_output */ struct statsblob *t_stats; /* Per-connection stats */ uint32_t t_logsn; /* Log "serial number" */ uint32_t gput_ts; /* Time goodput measurement started */ tcp_seq gput_seq; /* Outbound measurement seq */ tcp_seq gput_ack; /* Inbound measurement ack */ int32_t t_stats_gput_prev; /* XXXLAS: Prev gput measurement */ uint8_t t_tfo_client_cookie_len; /* TCP Fast Open client cookie length */ unsigned int *t_tfo_pending; /* TCP Fast Open server pending counter */ union { uint8_t client[TCP_FASTOPEN_MAX_COOKIE_LEN]; uint64_t server; } t_tfo_cookie; /* TCP Fast Open cookie to send */ #ifdef TCPPCAP struct mbufq t_inpkts; /* List of saved input packets. */ struct mbufq t_outpkts; /* List of saved output packets. */ #endif }; #endif /* _KERNEL || _WANT_TCPCB */ #ifdef _KERNEL struct tcptemp { u_char tt_ipgen[40]; /* the size must be of max ip header, now IPv6 */ struct tcphdr tt_t; }; /* Minimum map entries limit value, if set */ #define TCP_MIN_MAP_ENTRIES_LIMIT 128 /* * TODO: We yet need to brave plowing in * to tcp_input() and the pru_usrreq() block. * Right now these go to the old standards which * are somewhat ok, but in the long term may * need to be changed. If we do tackle tcp_input() * then we need to get rid of the tcp_do_segment() * function below. */ /* Flags for tcp functions */ #define TCP_FUNC_BEING_REMOVED 0x01 /* Can no longer be referenced */ /* * If defining the optional tcp_timers, in the * tfb_tcp_timer_stop call you must use the * callout_async_drain() function with the * tcp_timer_discard callback. You should check * the return of callout_async_drain() and if 0 * increment tt_draincnt. Since the timer sub-system * does not know your callbacks you must provide a * stop_all function that loops through and calls * tcp_timer_stop() with each of your defined timers. * Adding a tfb_tcp_handoff_ok function allows the socket * option to change stacks to query you even if the * connection is in a later stage. You return 0 to * say you can take over and run your stack, you return * non-zero (an error number) to say no you can't. * If the function is undefined you can only change * in the early states (before connect or listen). * tfb_tcp_fb_fini is changed to add a flag to tell * the old stack if the tcb is being destroyed or * not. A one in the flag means the TCB is being * destroyed, a zero indicates its transitioning to * another stack (via socket option). */ struct tcp_function_block { char tfb_tcp_block_name[TCP_FUNCTION_NAME_LEN_MAX]; int (*tfb_tcp_output)(struct tcpcb *); int (*tfb_tcp_output_wtime)(struct tcpcb *, const struct timeval *); void (*tfb_tcp_do_segment)(struct mbuf *, struct tcphdr *, struct socket *, struct tcpcb *, int, int, uint8_t); int (*tfb_do_queued_segments)(struct socket *, struct tcpcb *, int); int (*tfb_do_segment_nounlock)(struct mbuf *, struct tcphdr *, struct socket *, struct tcpcb *, int, int, uint8_t, int, struct timeval *); void (*tfb_tcp_hpts_do_segment)(struct mbuf *, struct tcphdr *, struct socket *, struct tcpcb *, int, int, uint8_t, int, struct timeval *); int (*tfb_tcp_ctloutput)(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp); /* Optional memory allocation/free routine */ int (*tfb_tcp_fb_init)(struct tcpcb *); void (*tfb_tcp_fb_fini)(struct tcpcb *, int); /* Optional timers, must define all if you define one */ int (*tfb_tcp_timer_stop_all)(struct tcpcb *); void (*tfb_tcp_timer_activate)(struct tcpcb *, uint32_t, u_int); int (*tfb_tcp_timer_active)(struct tcpcb *, uint32_t); void (*tfb_tcp_timer_stop)(struct tcpcb *, uint32_t); void (*tfb_tcp_rexmit_tmr)(struct tcpcb *); int (*tfb_tcp_handoff_ok)(struct tcpcb *); void (*tfb_tcp_mtu_chg)(struct tcpcb *); volatile uint32_t tfb_refcnt; uint32_t tfb_flags; uint8_t tfb_id; }; struct tcp_function { TAILQ_ENTRY(tcp_function) tf_next; char tf_name[TCP_FUNCTION_NAME_LEN_MAX]; struct tcp_function_block *tf_fb; }; TAILQ_HEAD(tcp_funchead, tcp_function); #endif /* _KERNEL */ /* * Flags and utility macros for the t_flags field. */ #define TF_ACKNOW 0x00000001 /* ack peer immediately */ #define TF_DELACK 0x00000002 /* ack, but try to delay it */ #define TF_NODELAY 0x00000004 /* don't delay packets to coalesce */ #define TF_NOOPT 0x00000008 /* don't use tcp options */ #define TF_SENTFIN 0x00000010 /* have sent FIN */ #define TF_REQ_SCALE 0x00000020 /* have/will request window scaling */ #define TF_RCVD_SCALE 0x00000040 /* other side has requested scaling */ #define TF_REQ_TSTMP 0x00000080 /* have/will request timestamps */ #define TF_RCVD_TSTMP 0x00000100 /* a timestamp was received in SYN */ #define TF_SACK_PERMIT 0x00000200 /* other side said I could SACK */ #define TF_NEEDSYN 0x00000400 /* send SYN (implicit state) */ #define TF_NEEDFIN 0x00000800 /* send FIN (implicit state) */ #define TF_NOPUSH 0x00001000 /* don't push */ #define TF_PREVVALID 0x00002000 /* saved values for bad rxmit valid */ #define TF_UNUSED1 0x00004000 /* unused */ #define TF_GPUTINPROG 0x00008000 /* Goodput measurement in progress */ #define TF_MORETOCOME 0x00010000 /* More data to be appended to sock */ #define TF_LQ_OVERFLOW 0x00020000 /* listen queue overflow */ #define TF_LASTIDLE 0x00040000 /* connection was previously idle */ #define TF_RXWIN0SENT 0x00080000 /* sent a receiver win 0 in response */ #define TF_FASTRECOVERY 0x00100000 /* in NewReno Fast Recovery */ #define TF_WASFRECOVERY 0x00200000 /* was in NewReno Fast Recovery */ #define TF_SIGNATURE 0x00400000 /* require MD5 digests (RFC2385) */ #define TF_FORCEDATA 0x00800000 /* force out a byte */ #define TF_TSO 0x01000000 /* TSO enabled on this connection */ #define TF_TOE 0x02000000 /* this connection is offloaded */ #define TF_UNUSED3 0x04000000 /* unused */ #define TF_UNUSED4 0x08000000 /* unused */ #define TF_UNUSED5 0x10000000 /* unused */ #define TF_CONGRECOVERY 0x20000000 /* congestion recovery mode */ #define TF_WASCRECOVERY 0x40000000 /* was in congestion recovery */ #define TF_FASTOPEN 0x80000000 /* TCP Fast Open indication */ #define IN_FASTRECOVERY(t_flags) (t_flags & TF_FASTRECOVERY) #define ENTER_FASTRECOVERY(t_flags) t_flags |= TF_FASTRECOVERY #define EXIT_FASTRECOVERY(t_flags) t_flags &= ~TF_FASTRECOVERY #define IN_CONGRECOVERY(t_flags) (t_flags & TF_CONGRECOVERY) #define ENTER_CONGRECOVERY(t_flags) t_flags |= TF_CONGRECOVERY #define EXIT_CONGRECOVERY(t_flags) t_flags &= ~TF_CONGRECOVERY #define IN_RECOVERY(t_flags) (t_flags & (TF_CONGRECOVERY | TF_FASTRECOVERY)) #define ENTER_RECOVERY(t_flags) t_flags |= (TF_CONGRECOVERY | TF_FASTRECOVERY) #define EXIT_RECOVERY(t_flags) t_flags &= ~(TF_CONGRECOVERY | TF_FASTRECOVERY) #if defined(_KERNEL) && !defined(TCP_RFC7413) #define IS_FASTOPEN(t_flags) (false) #else #define IS_FASTOPEN(t_flags) (t_flags & TF_FASTOPEN) #endif #define BYTES_THIS_ACK(tp, th) (th->th_ack - tp->snd_una) /* * Flags for the t_oobflags field. */ #define TCPOOB_HAVEDATA 0x01 #define TCPOOB_HADDATA 0x02 /* * Flags for the extended TCP flags field, t_flags2 */ #define TF2_PLPMTU_BLACKHOLE 0x00000001 /* Possible PLPMTUD Black Hole. */ #define TF2_PLPMTU_PMTUD 0x00000002 /* Allowed to attempt PLPMTUD. */ #define TF2_PLPMTU_MAXSEGSNT 0x00000004 /* Last seg sent was full seg. */ #define TF2_LOG_AUTO 0x00000008 /* Session is auto-logging. */ #define TF2_DROP_AF_DATA 0x00000010 /* Drop after all data ack'd */ #define TF2_ECN_PERMIT 0x00000020 /* connection ECN-ready */ #define TF2_ECN_SND_CWR 0x00000040 /* ECN CWR in queue */ #define TF2_ECN_SND_ECE 0x00000080 /* ECN ECE in queue */ #define TF2_ACE_PERMIT 0x00000100 /* Accurate ECN mode */ /* * Structure to hold TCP options that are only used during segment * processing (in tcp_input), but not held in the tcpcb. * It's basically used to reduce the number of parameters * to tcp_dooptions and tcp_addoptions. * The binary order of the to_flags is relevant for packing of the * options in tcp_addoptions. */ struct tcpopt { u_int32_t to_flags; /* which options are present */ #define TOF_MSS 0x0001 /* maximum segment size */ #define TOF_SCALE 0x0002 /* window scaling */ #define TOF_SACKPERM 0x0004 /* SACK permitted */ #define TOF_TS 0x0010 /* timestamp */ #define TOF_SIGNATURE 0x0040 /* TCP-MD5 signature option (RFC2385) */ #define TOF_SACK 0x0080 /* Peer sent SACK option */ #define TOF_FASTOPEN 0x0100 /* TCP Fast Open (TFO) cookie */ #define TOF_MAXOPT 0x0200 u_int32_t to_tsval; /* new timestamp */ u_int32_t to_tsecr; /* reflected timestamp */ u_char *to_sacks; /* pointer to the first SACK blocks */ u_char *to_signature; /* pointer to the TCP-MD5 signature */ u_int8_t *to_tfo_cookie; /* pointer to the TFO cookie */ u_int16_t to_mss; /* maximum segment size */ u_int8_t to_wscale; /* window scaling */ u_int8_t to_nsacks; /* number of SACK blocks */ u_int8_t to_tfo_len; /* TFO cookie length */ u_int32_t to_spare; /* UTO */ }; /* * Flags for tcp_dooptions. */ #define TO_SYN 0x01 /* parse SYN-only options */ struct hc_metrics_lite { /* must stay in sync with hc_metrics */ uint32_t rmx_mtu; /* MTU for this path */ uint32_t rmx_ssthresh; /* outbound gateway buffer limit */ uint32_t rmx_rtt; /* estimated round trip time */ uint32_t rmx_rttvar; /* estimated rtt variance */ uint32_t rmx_cwnd; /* congestion window */ uint32_t rmx_sendpipe; /* outbound delay-bandwidth product */ uint32_t rmx_recvpipe; /* inbound delay-bandwidth product */ }; /* * Used by tcp_maxmtu() to communicate interface specific features * and limits at the time of connection setup. */ struct tcp_ifcap { int ifcap; u_int tsomax; u_int tsomaxsegcount; u_int tsomaxsegsize; }; #ifndef _NETINET_IN_PCB_H_ struct in_conninfo; #endif /* _NETINET_IN_PCB_H_ */ struct tcptw { struct inpcb *tw_inpcb; /* XXX back pointer to internet pcb */ tcp_seq snd_nxt; tcp_seq rcv_nxt; tcp_seq iss; tcp_seq irs; u_short last_win; /* cached window value */ short tw_so_options; /* copy of so_options */ struct ucred *tw_cred; /* user credentials */ u_int32_t t_recent; u_int32_t ts_offset; /* our timestamp offset */ u_int t_starttime; int tw_time; TAILQ_ENTRY(tcptw) tw_2msl; void *tw_pspare; /* TCP_SIGNATURE */ u_int *tw_spare; /* TCP_SIGNATURE */ }; #define intotcpcb(ip) ((struct tcpcb *)(ip)->inp_ppcb) #define intotw(ip) ((struct tcptw *)(ip)->inp_ppcb) #define sototcpcb(so) (intotcpcb(sotoinpcb(so))) /* * The smoothed round-trip time and estimated variance * are stored as fixed point numbers scaled by the values below. * For convenience, these scales are also used in smoothing the average * (smoothed = (1/scale)sample + ((scale-1)/scale)smoothed). * With these scales, srtt has 3 bits to the right of the binary point, * and thus an "ALPHA" of 0.875. rttvar has 2 bits to the right of the * binary point, and is smoothed with an ALPHA of 0.75. */ #define TCP_RTT_SCALE 32 /* multiplier for srtt; 3 bits frac. */ #define TCP_RTT_SHIFT 5 /* shift for srtt; 3 bits frac. */ #define TCP_RTTVAR_SCALE 16 /* multiplier for rttvar; 2 bits */ #define TCP_RTTVAR_SHIFT 4 /* shift for rttvar; 2 bits */ #define TCP_DELTA_SHIFT 2 /* see tcp_input.c */ /* * The initial retransmission 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). * This version of the macro adapted from a paper by Lawrence * Brakmo and Larry Peterson which outlines a problem caused * by insufficient precision in the original implementation, * which results in inappropriately large RTO values for very * fast networks. */ #define TCP_REXMTVAL(tp) \ max((tp)->t_rttmin, (((tp)->t_srtt >> (TCP_RTT_SHIFT - TCP_DELTA_SHIFT)) \ + (tp)->t_rttvar) >> TCP_DELTA_SHIFT) /* * TCP statistics. * Many of these should be kept per connection, * but that's inconvenient at the moment. */ struct tcpstat { uint64_t tcps_connattempt; /* connections initiated */ uint64_t tcps_accepts; /* connections accepted */ uint64_t tcps_connects; /* connections established */ uint64_t tcps_drops; /* connections dropped */ uint64_t tcps_conndrops; /* embryonic connections dropped */ uint64_t tcps_minmssdrops; /* average minmss too low drops */ uint64_t tcps_closed; /* conn. closed (includes drops) */ uint64_t tcps_segstimed; /* segs where we tried to get rtt */ uint64_t tcps_rttupdated; /* times we succeeded */ uint64_t tcps_delack; /* delayed acks sent */ uint64_t tcps_timeoutdrop; /* conn. dropped in rxmt timeout */ uint64_t tcps_rexmttimeo; /* retransmit timeouts */ uint64_t tcps_persisttimeo; /* persist timeouts */ uint64_t tcps_keeptimeo; /* keepalive timeouts */ uint64_t tcps_keepprobe; /* keepalive probes sent */ uint64_t tcps_keepdrops; /* connections dropped in keepalive */ uint64_t tcps_sndtotal; /* total packets sent */ uint64_t tcps_sndpack; /* data packets sent */ uint64_t tcps_sndbyte; /* data bytes sent */ uint64_t tcps_sndrexmitpack; /* data packets retransmitted */ uint64_t tcps_sndrexmitbyte; /* data bytes retransmitted */ uint64_t tcps_sndrexmitbad; /* unnecessary packet retransmissions */ uint64_t tcps_sndacks; /* ack-only packets sent */ uint64_t tcps_sndprobe; /* window probes sent */ uint64_t tcps_sndurg; /* packets sent with URG only */ uint64_t tcps_sndwinup; /* window update-only packets sent */ uint64_t tcps_sndctrl; /* control (SYN|FIN|RST) packets sent */ uint64_t tcps_rcvtotal; /* total packets received */ uint64_t tcps_rcvpack; /* packets received in sequence */ uint64_t tcps_rcvbyte; /* bytes received in sequence */ uint64_t tcps_rcvbadsum; /* packets received with ccksum errs */ uint64_t tcps_rcvbadoff; /* packets received with bad offset */ uint64_t tcps_rcvreassfull; /* packets dropped for no reass space */ uint64_t tcps_rcvshort; /* packets received too short */ uint64_t tcps_rcvduppack; /* duplicate-only packets received */ uint64_t tcps_rcvdupbyte; /* duplicate-only bytes received */ uint64_t tcps_rcvpartduppack; /* packets with some duplicate data */ uint64_t tcps_rcvpartdupbyte; /* dup. bytes in part-dup. packets */ uint64_t tcps_rcvoopack; /* out-of-order packets received */ uint64_t tcps_rcvoobyte; /* out-of-order bytes received */ uint64_t tcps_rcvpackafterwin; /* packets with data after window */ uint64_t tcps_rcvbyteafterwin; /* bytes rcvd after window */ uint64_t tcps_rcvafterclose; /* packets rcvd after "close" */ uint64_t tcps_rcvwinprobe; /* rcvd window probe packets */ uint64_t tcps_rcvdupack; /* rcvd duplicate acks */ uint64_t tcps_rcvacktoomuch; /* rcvd acks for unsent data */ uint64_t tcps_rcvackpack; /* rcvd ack packets */ uint64_t tcps_rcvackbyte; /* bytes acked by rcvd acks */ uint64_t tcps_rcvwinupd; /* rcvd window update packets */ uint64_t tcps_pawsdrop; /* segments dropped due to PAWS */ uint64_t tcps_predack; /* times hdr predict ok for acks */ uint64_t tcps_preddat; /* times hdr predict ok for data pkts */ uint64_t tcps_pcbcachemiss; uint64_t tcps_cachedrtt; /* times cached RTT in route updated */ uint64_t tcps_cachedrttvar; /* times cached rttvar updated */ uint64_t tcps_cachedssthresh; /* times cached ssthresh updated */ uint64_t tcps_usedrtt; /* times RTT initialized from route */ uint64_t tcps_usedrttvar; /* times RTTVAR initialized from rt */ uint64_t tcps_usedssthresh; /* times ssthresh initialized from rt*/ uint64_t tcps_persistdrop; /* timeout in persist state */ uint64_t tcps_badsyn; /* bogus SYN, e.g. premature ACK */ uint64_t tcps_mturesent; /* resends due to MTU discovery */ uint64_t tcps_listendrop; /* listen queue overflows */ uint64_t tcps_badrst; /* ignored RSTs in the window */ uint64_t tcps_sc_added; /* entry added to syncache */ uint64_t tcps_sc_retransmitted; /* syncache entry was retransmitted */ uint64_t tcps_sc_dupsyn; /* duplicate SYN packet */ uint64_t tcps_sc_dropped; /* could not reply to packet */ uint64_t tcps_sc_completed; /* successful extraction of entry */ uint64_t tcps_sc_bucketoverflow;/* syncache per-bucket limit hit */ uint64_t tcps_sc_cacheoverflow; /* syncache cache limit hit */ uint64_t tcps_sc_reset; /* RST removed entry from syncache */ uint64_t tcps_sc_stale; /* timed out or listen socket gone */ uint64_t tcps_sc_aborted; /* syncache entry aborted */ uint64_t tcps_sc_badack; /* removed due to bad ACK */ uint64_t tcps_sc_unreach; /* ICMP unreachable received */ uint64_t tcps_sc_zonefail; /* zalloc() failed */ uint64_t tcps_sc_sendcookie; /* SYN cookie sent */ uint64_t tcps_sc_recvcookie; /* SYN cookie received */ uint64_t tcps_hc_added; /* entry added to hostcache */ uint64_t tcps_hc_bucketoverflow;/* hostcache per bucket limit hit */ uint64_t tcps_finwait2_drops; /* Drop FIN_WAIT_2 connection after time limit */ /* SACK related stats */ uint64_t tcps_sack_recovery_episode; /* SACK recovery episodes */ uint64_t tcps_sack_rexmits; /* SACK rexmit segments */ uint64_t tcps_sack_rexmit_bytes; /* SACK rexmit bytes */ uint64_t tcps_sack_rcv_blocks; /* SACK blocks (options) received */ uint64_t tcps_sack_send_blocks; /* SACK blocks (options) sent */ uint64_t tcps_sack_sboverflow; /* times scoreboard overflowed */ /* ECN related stats */ uint64_t tcps_ecn_ce; /* ECN Congestion Experienced */ uint64_t tcps_ecn_ect0; /* ECN Capable Transport */ uint64_t tcps_ecn_ect1; /* ECN Capable Transport */ uint64_t tcps_ecn_shs; /* ECN successful handshakes */ uint64_t tcps_ecn_rcwnd; /* # times ECN reduced the cwnd */ /* TCP_SIGNATURE related stats */ uint64_t tcps_sig_rcvgoodsig; /* Total matching signature received */ uint64_t tcps_sig_rcvbadsig; /* Total bad signature received */ uint64_t tcps_sig_err_buildsig; /* Failed to make signature */ uint64_t tcps_sig_err_sigopt; /* No signature expected by socket */ uint64_t tcps_sig_err_nosigopt; /* No signature provided by segment */ /* Path MTU Discovery Black Hole Detection related stats */ uint64_t tcps_pmtud_blackhole_activated; /* Black Hole Count */ uint64_t tcps_pmtud_blackhole_activated_min_mss; /* BH at min MSS Count */ uint64_t tcps_pmtud_blackhole_failed; /* Black Hole Failure Count */ uint64_t _pad[12]; /* 6 UTO, 6 TBD */ }; #define tcps_rcvmemdrop tcps_rcvreassfull /* compat */ #ifdef _KERNEL #define TI_UNLOCKED 1 #define TI_RLOCKED 2 #include VNET_PCPUSTAT_DECLARE(struct tcpstat, tcpstat); /* tcp statistics */ /* * In-kernel consumers can use these accessor macros directly to update * stats. */ #define TCPSTAT_ADD(name, val) \ VNET_PCPUSTAT_ADD(struct tcpstat, tcpstat, name, (val)) #define TCPSTAT_INC(name) TCPSTAT_ADD(name, 1) /* * Kernel module consumers must use this accessor macro. */ void kmod_tcpstat_inc(int statnum); #define KMOD_TCPSTAT_INC(name) \ kmod_tcpstat_inc(offsetof(struct tcpstat, name) / sizeof(uint64_t)) /* * Running TCP connection count by state. */ VNET_DECLARE(counter_u64_t, tcps_states[TCP_NSTATES]); #define V_tcps_states VNET(tcps_states) #define TCPSTATES_INC(state) counter_u64_add(V_tcps_states[state], 1) #define TCPSTATES_DEC(state) counter_u64_add(V_tcps_states[state], -1) /* * TCP specific helper hook point identifiers. */ #define HHOOK_TCP_EST_IN 0 #define HHOOK_TCP_EST_OUT 1 #define HHOOK_TCP_LAST HHOOK_TCP_EST_OUT struct tcp_hhook_data { struct tcpcb *tp; struct tcphdr *th; struct tcpopt *to; uint32_t len; int tso; tcp_seq curack; }; #ifdef TCP_HHOOK void hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th, struct tcpopt *to, uint32_t len, int tso); #endif #endif /* * TCB structure exported to user-land via sysctl(3). * * Fields prefixed with "xt_" are unique to the export structure, and fields * with "t_" or other prefixes match corresponding fields of 'struct tcpcb'. * * 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 * * Evil hack: declare only if in_pcb.h and sys/socketvar.h have been * included. Not all of our clients do. */ #if defined(_NETINET_IN_PCB_H_) && defined(_SYS_SOCKETVAR_H_) struct xtcpcb { ksize_t xt_len; /* length of this structure */ struct xinpcb xt_inp; char xt_stack[TCP_FUNCTION_NAME_LEN_MAX]; /* (s) */ char xt_logid[TCP_LOG_ID_LEN]; /* (s) */ int64_t spare64[8]; int32_t t_state; /* (s,p) */ uint32_t t_flags; /* (s,p) */ int32_t t_sndzerowin; /* (s) */ int32_t t_sndrexmitpack; /* (s) */ int32_t t_rcvoopack; /* (s) */ int32_t t_rcvtime; /* (s) */ int32_t tt_rexmt; /* (s) */ int32_t tt_persist; /* (s) */ int32_t tt_keep; /* (s) */ int32_t tt_2msl; /* (s) */ int32_t tt_delack; /* (s) */ int32_t t_logstate; /* (3) */ int32_t spare32[32]; } __aligned(8); #ifdef _KERNEL void tcp_inptoxtp(const struct inpcb *, struct xtcpcb *); #endif #endif /* * TCP function information (name-to-id mapping, aliases, and refcnt) * exported to user-land via sysctl(3). */ struct tcp_function_info { uint32_t tfi_refcnt; uint8_t tfi_id; char tfi_name[TCP_FUNCTION_NAME_LEN_MAX]; char tfi_alias[TCP_FUNCTION_NAME_LEN_MAX]; }; /* * Identifiers for TCP sysctl nodes */ #define TCPCTL_DO_RFC1323 1 /* use RFC-1323 extensions */ #define TCPCTL_MSSDFLT 3 /* MSS default */ #define TCPCTL_STATS 4 /* statistics */ #define TCPCTL_RTTDFLT 5 /* default RTT estimate */ #define TCPCTL_KEEPIDLE 6 /* keepalive idle timer */ #define TCPCTL_KEEPINTVL 7 /* interval to send keepalives */ #define TCPCTL_SENDSPACE 8 /* send buffer space */ #define TCPCTL_RECVSPACE 9 /* receive buffer space */ #define TCPCTL_KEEPINIT 10 /* timeout for establishing syn */ #define TCPCTL_PCBLIST 11 /* list of all outstanding PCBs */ #define TCPCTL_DELACKTIME 12 /* time before sending delayed ACK */ #define TCPCTL_V6MSSDFLT 13 /* MSS default for IPv6 */ #define TCPCTL_SACK 14 /* Selective Acknowledgement,rfc 2018 */ #define TCPCTL_DROP 15 /* drop tcp connection */ #define TCPCTL_STATES 16 /* connection counts by TCP state */ #ifdef _KERNEL #ifdef SYSCTL_DECL SYSCTL_DECL(_net_inet_tcp); SYSCTL_DECL(_net_inet_tcp_sack); MALLOC_DECLARE(M_TCPLOG); #endif -extern int tcp_log_in_vain; +VNET_DECLARE(int, tcp_log_in_vain); +#define V_tcp_log_in_vain VNET(tcp_log_in_vain) /* * Global TCP tunables shared between different stacks. * Please keep the list sorted. */ VNET_DECLARE(int, drop_synfin); VNET_DECLARE(int, path_mtu_discovery); VNET_DECLARE(int, tcp_abc_l_var); VNET_DECLARE(int, tcp_autorcvbuf_max); VNET_DECLARE(int, tcp_autosndbuf_inc); VNET_DECLARE(int, tcp_autosndbuf_max); VNET_DECLARE(int, tcp_delack_enabled); VNET_DECLARE(int, tcp_do_autorcvbuf); VNET_DECLARE(int, tcp_do_autosndbuf); VNET_DECLARE(int, tcp_do_ecn); VNET_DECLARE(int, tcp_do_newcwv); VNET_DECLARE(int, tcp_do_rfc1323); VNET_DECLARE(int, tcp_do_rfc3042); VNET_DECLARE(int, tcp_do_rfc3390); VNET_DECLARE(int, tcp_do_rfc3465); VNET_DECLARE(int, tcp_do_rfc6675_pipe); VNET_DECLARE(int, tcp_do_sack); VNET_DECLARE(int, tcp_do_tso); VNET_DECLARE(int, tcp_ecn_maxretries); VNET_DECLARE(int, tcp_initcwnd_segments); VNET_DECLARE(int, tcp_insecure_rst); VNET_DECLARE(int, tcp_insecure_syn); VNET_DECLARE(uint32_t, tcp_map_entries_limit); VNET_DECLARE(uint32_t, tcp_map_split_limit); VNET_DECLARE(int, tcp_minmss); VNET_DECLARE(int, tcp_mssdflt); #ifdef STATS VNET_DECLARE(int, tcp_perconn_stats_dflt_tpl); VNET_DECLARE(int, tcp_perconn_stats_enable); #endif /* STATS */ VNET_DECLARE(int, tcp_recvspace); VNET_DECLARE(int, tcp_sack_globalholes); VNET_DECLARE(int, tcp_sack_globalmaxholes); VNET_DECLARE(int, tcp_sack_maxholes); VNET_DECLARE(int, tcp_sc_rst_sock_fail); VNET_DECLARE(int, tcp_sendspace); VNET_DECLARE(struct inpcbhead, tcb); VNET_DECLARE(struct inpcbinfo, tcbinfo); #define V_tcp_do_newcwv VNET(tcp_do_newcwv) #define V_drop_synfin VNET(drop_synfin) #define V_path_mtu_discovery VNET(path_mtu_discovery) #define V_tcb VNET(tcb) #define V_tcbinfo VNET(tcbinfo) #define V_tcp_abc_l_var VNET(tcp_abc_l_var) #define V_tcp_autorcvbuf_max VNET(tcp_autorcvbuf_max) #define V_tcp_autosndbuf_inc VNET(tcp_autosndbuf_inc) #define V_tcp_autosndbuf_max VNET(tcp_autosndbuf_max) #define V_tcp_delack_enabled VNET(tcp_delack_enabled) #define V_tcp_do_autorcvbuf VNET(tcp_do_autorcvbuf) #define V_tcp_do_autosndbuf VNET(tcp_do_autosndbuf) #define V_tcp_do_ecn VNET(tcp_do_ecn) #define V_tcp_do_rfc1323 VNET(tcp_do_rfc1323) #define V_tcp_ts_offset_per_conn VNET(tcp_ts_offset_per_conn) #define V_tcp_do_rfc3042 VNET(tcp_do_rfc3042) #define V_tcp_do_rfc3390 VNET(tcp_do_rfc3390) #define V_tcp_do_rfc3465 VNET(tcp_do_rfc3465) #define V_tcp_do_rfc6675_pipe VNET(tcp_do_rfc6675_pipe) #define V_tcp_do_sack VNET(tcp_do_sack) #define V_tcp_do_tso VNET(tcp_do_tso) #define V_tcp_ecn_maxretries VNET(tcp_ecn_maxretries) #define V_tcp_initcwnd_segments VNET(tcp_initcwnd_segments) #define V_tcp_insecure_rst VNET(tcp_insecure_rst) #define V_tcp_insecure_syn VNET(tcp_insecure_syn) #define V_tcp_map_entries_limit VNET(tcp_map_entries_limit) #define V_tcp_map_split_limit VNET(tcp_map_split_limit) #define V_tcp_minmss VNET(tcp_minmss) #define V_tcp_mssdflt VNET(tcp_mssdflt) #ifdef STATS #define V_tcp_perconn_stats_dflt_tpl VNET(tcp_perconn_stats_dflt_tpl) #define V_tcp_perconn_stats_enable VNET(tcp_perconn_stats_enable) #endif /* STATS */ #define V_tcp_recvspace VNET(tcp_recvspace) #define V_tcp_sack_globalholes VNET(tcp_sack_globalholes) #define V_tcp_sack_globalmaxholes VNET(tcp_sack_globalmaxholes) #define V_tcp_sack_maxholes VNET(tcp_sack_maxholes) #define V_tcp_sc_rst_sock_fail VNET(tcp_sc_rst_sock_fail) #define V_tcp_sendspace VNET(tcp_sendspace) #define V_tcp_udp_tunneling_overhead VNET(tcp_udp_tunneling_overhead) #define V_tcp_udp_tunneling_port VNET(tcp_udp_tunneling_port) #ifdef TCP_HHOOK VNET_DECLARE(struct hhook_head *, tcp_hhh[HHOOK_TCP_LAST + 1]); #define V_tcp_hhh VNET(tcp_hhh) #endif int tcp_addoptions(struct tcpopt *, u_char *); int tcp_ccalgounload(struct cc_algo *unload_algo); struct tcpcb * tcp_close(struct tcpcb *); void tcp_discardcb(struct tcpcb *); void tcp_twstart(struct tcpcb *); void tcp_twclose(struct tcptw *, int); void tcp_ctlinput(int, struct sockaddr *, void *); int tcp_ctloutput(struct socket *, struct sockopt *); struct tcpcb * tcp_drop(struct tcpcb *, int); void tcp_drain(void); void tcp_init(void); void tcp_fini(void *); char *tcp_log_addrs(struct in_conninfo *, struct tcphdr *, void *, const void *); char *tcp_log_vain(struct in_conninfo *, struct tcphdr *, void *, const void *); int tcp_reass(struct tcpcb *, struct tcphdr *, tcp_seq *, int *, struct mbuf *); void tcp_reass_global_init(void); void tcp_reass_flush(struct tcpcb *); void tcp_dooptions(struct tcpopt *, u_char *, int, int); void tcp_dropwithreset(struct mbuf *, struct tcphdr *, struct tcpcb *, int, int); void tcp_pulloutofband(struct socket *, struct tcphdr *, struct mbuf *, int); void tcp_xmit_timer(struct tcpcb *, int); void tcp_newreno_partial_ack(struct tcpcb *, struct tcphdr *); void cc_ack_received(struct tcpcb *tp, struct tcphdr *th, uint16_t nsegs, uint16_t type); void cc_conn_init(struct tcpcb *tp); void cc_post_recovery(struct tcpcb *tp, struct tcphdr *th); void cc_ecnpkt_handler(struct tcpcb *tp, struct tcphdr *th, uint8_t iptos); void cc_cong_signal(struct tcpcb *tp, struct tcphdr *th, uint32_t type); #ifdef TCP_HHOOK void hhook_run_tcp_est_in(struct tcpcb *tp, struct tcphdr *th, struct tcpopt *to); #endif int tcp_input(struct mbuf **, int *, int); int tcp_autorcvbuf(struct mbuf *, struct tcphdr *, struct socket *, struct tcpcb *, int); void tcp_do_segment(struct mbuf *, struct tcphdr *, struct socket *, struct tcpcb *, int, int, uint8_t); int register_tcp_functions(struct tcp_function_block *blk, int wait); int register_tcp_functions_as_names(struct tcp_function_block *blk, int wait, const char *names[], int *num_names); int register_tcp_functions_as_name(struct tcp_function_block *blk, const char *name, int wait); int deregister_tcp_functions(struct tcp_function_block *blk, bool quiesce, bool force); struct tcp_function_block *find_and_ref_tcp_functions(struct tcp_function_set *fs); void tcp_switch_back_to_default(struct tcpcb *tp); struct tcp_function_block * find_and_ref_tcp_fb(struct tcp_function_block *fs); int tcp_default_ctloutput(struct socket *so, struct sockopt *sopt, struct inpcb *inp, struct tcpcb *tp); extern counter_u64_t tcp_inp_lro_direct_queue; extern counter_u64_t tcp_inp_lro_wokeup_queue; extern counter_u64_t tcp_inp_lro_compressed; extern counter_u64_t tcp_inp_lro_single_push; extern counter_u64_t tcp_inp_lro_locks_taken; extern counter_u64_t tcp_inp_lro_sack_wake; #ifdef NETFLIX_EXP_DETECTION /* Various SACK attack thresholds */ extern int32_t tcp_force_detection; extern int32_t tcp_sack_to_ack_thresh; extern int32_t tcp_sack_to_move_thresh; extern int32_t tcp_restoral_thresh; extern int32_t tcp_sad_decay_val; extern int32_t tcp_sad_pacing_interval; extern int32_t tcp_sad_low_pps; extern int32_t tcp_map_minimum; extern int32_t tcp_attack_on_turns_on_logging; #endif uint32_t tcp_maxmtu(struct in_conninfo *, struct tcp_ifcap *); uint32_t tcp_maxmtu6(struct in_conninfo *, struct tcp_ifcap *); u_int tcp_maxseg(const struct tcpcb *); void tcp_mss_update(struct tcpcb *, int, int, struct hc_metrics_lite *, struct tcp_ifcap *); void tcp_mss(struct tcpcb *, int); int tcp_mssopt(struct in_conninfo *); struct inpcb * tcp_drop_syn_sent(struct inpcb *, int); struct tcpcb * tcp_newtcpcb(struct inpcb *); int tcp_output(struct tcpcb *); void tcp_state_change(struct tcpcb *, int); void tcp_respond(struct tcpcb *, void *, struct tcphdr *, struct mbuf *, tcp_seq, tcp_seq, int); void tcp_tw_init(void); #ifdef VIMAGE void tcp_tw_destroy(void); #endif void tcp_tw_zone_change(void); int tcp_twcheck(struct inpcb *, struct tcpopt *, struct tcphdr *, struct mbuf *, int); void tcp_setpersist(struct tcpcb *); void tcp_slowtimo(void); struct tcptemp * tcpip_maketemplate(struct inpcb *); void tcpip_fillheaders(struct inpcb *, void *, void *); void tcp_timer_activate(struct tcpcb *, uint32_t, u_int); int tcp_timer_suspend(struct tcpcb *, uint32_t); void tcp_timers_unsuspend(struct tcpcb *, uint32_t); int tcp_timer_active(struct tcpcb *, uint32_t); void tcp_timer_stop(struct tcpcb *, uint32_t); void tcp_trace(short, short, struct tcpcb *, void *, struct tcphdr *, int); int inp_to_cpuid(struct inpcb *inp); /* * All tcp_hc_* functions are IPv4 and IPv6 (via in_conninfo) */ void tcp_hc_init(void); #ifdef VIMAGE void tcp_hc_destroy(void); #endif void tcp_hc_get(struct in_conninfo *, struct hc_metrics_lite *); uint32_t tcp_hc_getmtu(struct in_conninfo *); void tcp_hc_updatemtu(struct in_conninfo *, uint32_t); void tcp_hc_update(struct in_conninfo *, struct hc_metrics_lite *); extern struct pr_usrreqs tcp_usrreqs; uint32_t tcp_new_ts_offset(struct in_conninfo *); tcp_seq tcp_new_isn(struct in_conninfo *); int tcp_sack_doack(struct tcpcb *, struct tcpopt *, tcp_seq); void tcp_update_dsack_list(struct tcpcb *, tcp_seq, tcp_seq); void tcp_update_sack_list(struct tcpcb *tp, tcp_seq rcv_laststart, tcp_seq rcv_lastend); void tcp_clean_dsack_blocks(struct tcpcb *tp); void tcp_clean_sackreport(struct tcpcb *tp); void tcp_sack_adjust(struct tcpcb *tp); struct sackhole *tcp_sack_output(struct tcpcb *tp, int *sack_bytes_rexmt); void tcp_sack_partialack(struct tcpcb *, struct tcphdr *); void tcp_free_sackholes(struct tcpcb *tp); int tcp_newreno(struct tcpcb *, struct tcphdr *); int tcp_compute_pipe(struct tcpcb *); uint32_t tcp_compute_initwnd(uint32_t); void tcp_sndbuf_autoscale(struct tcpcb *, struct socket *, uint32_t); int tcp_stats_sample_rollthedice(struct tcpcb *tp, void *seed_bytes, size_t seed_len); struct mbuf * tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *plen, int32_t seglimit, int32_t segsize, struct sockbuf *sb, bool hw_tls); int tcp_stats_init(void); static inline void tcp_fields_to_host(struct tcphdr *th) { th->th_seq = ntohl(th->th_seq); th->th_ack = ntohl(th->th_ack); th->th_win = ntohs(th->th_win); th->th_urp = ntohs(th->th_urp); } static inline void tcp_fields_to_net(struct tcphdr *th) { th->th_seq = htonl(th->th_seq); th->th_ack = htonl(th->th_ack); th->th_win = htons(th->th_win); th->th_urp = htons(th->th_urp); } #endif /* _KERNEL */ #endif /* _NETINET_TCP_VAR_H_ */ Index: head/sys/netinet/udp_usrreq.c =================================================================== --- head/sys/netinet/udp_usrreq.c (revision 356526) +++ head/sys/netinet/udp_usrreq.c (revision 356527) @@ -1,1728 +1,1728 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995 * The Regents of the University of California. * Copyright (c) 2008 Robert N. M. Watson * Copyright (c) 2010-2011 Juniper Networks, Inc. * Copyright (c) 2014 Kevin Lo * 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. * * @(#)udp_usrreq.c 8.6 (Berkeley) 5/23/95 */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include "opt_ipsec.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 #ifdef INET6 #include #endif #include #include #include #include #ifdef INET6 #include #endif #include #include #include #include #include #include #include /* * UDP and UDP-Lite protocols implementation. * Per RFC 768, August, 1980. * Per RFC 3828, July, 2004. */ /* * BSD 4.2 defaulted the udp checksum to be off. Turning off udp checksums * removes the only data integrity mechanism for packets and malformed * packets that would otherwise be discarded due to bad checksums, and may * cause problems (especially for NFS data blocks). */ VNET_DEFINE(int, udp_cksum) = 1; SYSCTL_INT(_net_inet_udp, UDPCTL_CHECKSUM, checksum, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(udp_cksum), 0, "compute udp checksum"); -int udp_log_in_vain = 0; -SYSCTL_INT(_net_inet_udp, OID_AUTO, log_in_vain, CTLFLAG_RW, - &udp_log_in_vain, 0, "Log all incoming UDP packets"); +VNET_DEFINE(int, udp_log_in_vain) = 0; +SYSCTL_INT(_net_inet_udp, OID_AUTO, log_in_vain, CTLFLAG_VNET | CTLFLAG_RW, + &VNET_NAME(udp_log_in_vain), 0, "Log all incoming UDP packets"); VNET_DEFINE(int, udp_blackhole) = 0; SYSCTL_INT(_net_inet_udp, OID_AUTO, blackhole, CTLFLAG_VNET | CTLFLAG_RW, &VNET_NAME(udp_blackhole), 0, "Do not send port unreachables for refused connects"); u_long udp_sendspace = 9216; /* really max datagram size */ SYSCTL_ULONG(_net_inet_udp, UDPCTL_MAXDGRAM, maxdgram, CTLFLAG_RW, &udp_sendspace, 0, "Maximum outgoing UDP datagram size"); u_long udp_recvspace = 40 * (1024 + #ifdef INET6 sizeof(struct sockaddr_in6) #else sizeof(struct sockaddr_in) #endif ); /* 40 1K datagrams */ SYSCTL_ULONG(_net_inet_udp, UDPCTL_RECVSPACE, recvspace, CTLFLAG_RW, &udp_recvspace, 0, "Maximum space for incoming UDP datagrams"); VNET_DEFINE(struct inpcbhead, udb); /* from udp_var.h */ VNET_DEFINE(struct inpcbinfo, udbinfo); VNET_DEFINE(struct inpcbhead, ulitecb); VNET_DEFINE(struct inpcbinfo, ulitecbinfo); VNET_DEFINE_STATIC(uma_zone_t, udpcb_zone); #define V_udpcb_zone VNET(udpcb_zone) #ifndef UDBHASHSIZE #define UDBHASHSIZE 128 #endif VNET_PCPUSTAT_DEFINE(struct udpstat, udpstat); /* from udp_var.h */ VNET_PCPUSTAT_SYSINIT(udpstat); SYSCTL_VNET_PCPUSTAT(_net_inet_udp, UDPCTL_STATS, stats, struct udpstat, udpstat, "UDP statistics (struct udpstat, netinet/udp_var.h)"); #ifdef VIMAGE VNET_PCPUSTAT_SYSUNINIT(udpstat); #endif /* VIMAGE */ #ifdef INET static void udp_detach(struct socket *so); static int udp_output(struct inpcb *, struct mbuf *, struct sockaddr *, struct mbuf *, struct thread *); #endif static void udp_zone_change(void *tag) { uma_zone_set_max(V_udbinfo.ipi_zone, maxsockets); uma_zone_set_max(V_udpcb_zone, maxsockets); } static int udp_inpcb_init(void *mem, int size, int flags) { struct inpcb *inp; inp = mem; INP_LOCK_INIT(inp, "inp", "udpinp"); return (0); } static int udplite_inpcb_init(void *mem, int size, int flags) { struct inpcb *inp; inp = mem; INP_LOCK_INIT(inp, "inp", "udpliteinp"); return (0); } void udp_init(void) { /* * For now default to 2-tuple UDP hashing - until the fragment * reassembly code can also update the flowid. * * Once we can calculate the flowid that way and re-establish * a 4-tuple, flip this to 4-tuple. */ in_pcbinfo_init(&V_udbinfo, "udp", &V_udb, UDBHASHSIZE, UDBHASHSIZE, "udp_inpcb", udp_inpcb_init, IPI_HASHFIELDS_2TUPLE); V_udpcb_zone = uma_zcreate("udpcb", sizeof(struct udpcb), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); uma_zone_set_max(V_udpcb_zone, maxsockets); uma_zone_set_warning(V_udpcb_zone, "kern.ipc.maxsockets limit reached"); EVENTHANDLER_REGISTER(maxsockets_change, udp_zone_change, NULL, EVENTHANDLER_PRI_ANY); } void udplite_init(void) { in_pcbinfo_init(&V_ulitecbinfo, "udplite", &V_ulitecb, UDBHASHSIZE, UDBHASHSIZE, "udplite_inpcb", udplite_inpcb_init, IPI_HASHFIELDS_2TUPLE); } /* * Kernel module interface for updating udpstat. The argument is an index * into udpstat treated as an array of u_long. While this encodes the * general layout of udpstat into the caller, it doesn't encode its location, * so that future changes to add, for example, per-CPU stats support won't * cause binary compatibility problems for kernel modules. */ void kmod_udpstat_inc(int statnum) { counter_u64_add(VNET(udpstat)[statnum], 1); } int udp_newudpcb(struct inpcb *inp) { struct udpcb *up; up = uma_zalloc(V_udpcb_zone, M_NOWAIT | M_ZERO); if (up == NULL) return (ENOBUFS); inp->inp_ppcb = up; return (0); } void udp_discardcb(struct udpcb *up) { uma_zfree(V_udpcb_zone, up); } #ifdef VIMAGE static void udp_destroy(void *unused __unused) { in_pcbinfo_destroy(&V_udbinfo); uma_zdestroy(V_udpcb_zone); } VNET_SYSUNINIT(udp, SI_SUB_PROTO_DOMAIN, SI_ORDER_FOURTH, udp_destroy, NULL); static void udplite_destroy(void *unused __unused) { in_pcbinfo_destroy(&V_ulitecbinfo); } VNET_SYSUNINIT(udplite, SI_SUB_PROTO_DOMAIN, SI_ORDER_FOURTH, udplite_destroy, NULL); #endif #ifdef INET /* * Subroutine of udp_input(), which appends the provided mbuf chain to the * passed pcb/socket. The caller must provide a sockaddr_in via udp_in that * contains the source address. If the socket ends up being an IPv6 socket, * udp_append() will convert to a sockaddr_in6 before passing the address * into the socket code. * * In the normal case udp_append() will return 0, indicating that you * must unlock the inp. However if a tunneling protocol is in place we increment * the inpcb refcnt and unlock the inp, on return from the tunneling protocol we * then decrement the reference count. If the inp_rele returns 1, indicating the * inp is gone, we return that to the caller to tell them *not* to unlock * the inp. In the case of multi-cast this will cause the distribution * to stop (though most tunneling protocols known currently do *not* use * multicast). */ static int udp_append(struct inpcb *inp, struct ip *ip, struct mbuf *n, int off, struct sockaddr_in *udp_in) { struct sockaddr *append_sa; struct socket *so; struct mbuf *tmpopts, *opts = NULL; #ifdef INET6 struct sockaddr_in6 udp_in6; #endif struct udpcb *up; INP_LOCK_ASSERT(inp); /* * Engage the tunneling protocol. */ up = intoudpcb(inp); if (up->u_tun_func != NULL) { in_pcbref(inp); INP_RUNLOCK(inp); (*up->u_tun_func)(n, off, inp, (struct sockaddr *)&udp_in[0], up->u_tun_ctx); INP_RLOCK(inp); return (in_pcbrele_rlocked(inp)); } off += sizeof(struct udphdr); #if defined(IPSEC) || defined(IPSEC_SUPPORT) /* Check AH/ESP integrity. */ if (IPSEC_ENABLED(ipv4) && IPSEC_CHECK_POLICY(ipv4, n, inp) != 0) { m_freem(n); return (0); } if (up->u_flags & UF_ESPINUDP) {/* IPSec UDP encaps. */ if (IPSEC_ENABLED(ipv4) && UDPENCAP_INPUT(n, off, AF_INET) != 0) return (0); /* Consumed. */ } #endif /* IPSEC */ #ifdef MAC if (mac_inpcb_check_deliver(inp, n) != 0) { m_freem(n); return (0); } #endif /* MAC */ if (inp->inp_flags & INP_CONTROLOPTS || inp->inp_socket->so_options & (SO_TIMESTAMP | SO_BINTIME)) { #ifdef INET6 if (inp->inp_vflag & INP_IPV6) (void)ip6_savecontrol_v4(inp, n, &opts, NULL); else #endif /* INET6 */ ip_savecontrol(inp, &opts, ip, n); } if ((inp->inp_vflag & INP_IPV4) && (inp->inp_flags2 & INP_ORIGDSTADDR)) { tmpopts = sbcreatecontrol((caddr_t)&udp_in[1], sizeof(struct sockaddr_in), IP_ORIGDSTADDR, IPPROTO_IP); if (tmpopts) { if (opts) { tmpopts->m_next = opts; opts = tmpopts; } else opts = tmpopts; } } #ifdef INET6 if (inp->inp_vflag & INP_IPV6) { bzero(&udp_in6, sizeof(udp_in6)); udp_in6.sin6_len = sizeof(udp_in6); udp_in6.sin6_family = AF_INET6; in6_sin_2_v4mapsin6(&udp_in[0], &udp_in6); append_sa = (struct sockaddr *)&udp_in6; } else #endif /* INET6 */ append_sa = (struct sockaddr *)&udp_in[0]; m_adj(n, off); so = inp->inp_socket; SOCKBUF_LOCK(&so->so_rcv); if (sbappendaddr_locked(&so->so_rcv, append_sa, n, opts) == 0) { SOCKBUF_UNLOCK(&so->so_rcv); m_freem(n); if (opts) m_freem(opts); UDPSTAT_INC(udps_fullsock); } else sorwakeup_locked(so); return (0); } int udp_input(struct mbuf **mp, int *offp, int proto) { struct ip *ip; struct udphdr *uh; struct ifnet *ifp; struct inpcb *inp; uint16_t len, ip_len; struct inpcbinfo *pcbinfo; struct ip save_ip; struct sockaddr_in udp_in[2]; struct mbuf *m; struct m_tag *fwd_tag; int cscov_partial, iphlen; m = *mp; iphlen = *offp; ifp = m->m_pkthdr.rcvif; *mp = NULL; UDPSTAT_INC(udps_ipackets); /* * Strip IP options, if any; should skip this, make available to * user, and use on returned packets, but we don't yet have a way to * check the checksum with options still present. */ if (iphlen > sizeof (struct ip)) { ip_stripoptions(m); iphlen = sizeof(struct ip); } /* * Get IP and UDP header together in first mbuf. */ if (m->m_len < iphlen + sizeof(struct udphdr)) { if ((m = m_pullup(m, iphlen + sizeof(struct udphdr))) == NULL) { UDPSTAT_INC(udps_hdrops); return (IPPROTO_DONE); } } ip = mtod(m, struct ip *); uh = (struct udphdr *)((caddr_t)ip + iphlen); cscov_partial = (proto == IPPROTO_UDPLITE) ? 1 : 0; /* * Destination port of 0 is illegal, based on RFC768. */ if (uh->uh_dport == 0) goto badunlocked; /* * Construct sockaddr format source address. Stuff source address * and datagram in user buffer. */ bzero(&udp_in[0], sizeof(struct sockaddr_in) * 2); udp_in[0].sin_len = sizeof(struct sockaddr_in); udp_in[0].sin_family = AF_INET; udp_in[0].sin_port = uh->uh_sport; udp_in[0].sin_addr = ip->ip_src; udp_in[1].sin_len = sizeof(struct sockaddr_in); udp_in[1].sin_family = AF_INET; udp_in[1].sin_port = uh->uh_dport; udp_in[1].sin_addr = ip->ip_dst; /* * Make mbuf data length reflect UDP length. If not enough data to * reflect UDP length, drop. */ len = ntohs((u_short)uh->uh_ulen); ip_len = ntohs(ip->ip_len) - iphlen; if (proto == IPPROTO_UDPLITE && (len == 0 || len == ip_len)) { /* Zero means checksum over the complete packet. */ if (len == 0) len = ip_len; cscov_partial = 0; } if (ip_len != len) { if (len > ip_len || len < sizeof(struct udphdr)) { UDPSTAT_INC(udps_badlen); goto badunlocked; } if (proto == IPPROTO_UDP) m_adj(m, len - ip_len); } /* * Save a copy of the IP header in case we want restore it for * sending an ICMP error message in response. */ if (!V_udp_blackhole) save_ip = *ip; else memset(&save_ip, 0, sizeof(save_ip)); /* * Checksum extended UDP header and data. */ if (uh->uh_sum) { u_short uh_sum; if ((m->m_pkthdr.csum_flags & CSUM_DATA_VALID) && !cscov_partial) { if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR) uh_sum = m->m_pkthdr.csum_data; else uh_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, htonl((u_short)len + m->m_pkthdr.csum_data + proto)); uh_sum ^= 0xffff; } else { char b[9]; bcopy(((struct ipovly *)ip)->ih_x1, b, 9); bzero(((struct ipovly *)ip)->ih_x1, 9); ((struct ipovly *)ip)->ih_len = (proto == IPPROTO_UDP) ? uh->uh_ulen : htons(ip_len); uh_sum = in_cksum(m, len + sizeof (struct ip)); bcopy(b, ((struct ipovly *)ip)->ih_x1, 9); } if (uh_sum) { UDPSTAT_INC(udps_badsum); m_freem(m); return (IPPROTO_DONE); } } else { if (proto == IPPROTO_UDP) { UDPSTAT_INC(udps_nosum); } else { /* UDPLite requires a checksum */ /* XXX: What is the right UDPLite MIB counter here? */ m_freem(m); return (IPPROTO_DONE); } } pcbinfo = udp_get_inpcbinfo(proto); if (IN_MULTICAST(ntohl(ip->ip_dst.s_addr)) || in_broadcast(ip->ip_dst, ifp)) { struct inpcb *last; struct inpcbhead *pcblist; NET_EPOCH_ASSERT(); pcblist = udp_get_pcblist(proto); last = NULL; CK_LIST_FOREACH(inp, pcblist, inp_list) { if (inp->inp_lport != uh->uh_dport) continue; #ifdef INET6 if ((inp->inp_vflag & INP_IPV4) == 0) continue; #endif if (inp->inp_laddr.s_addr != INADDR_ANY && inp->inp_laddr.s_addr != ip->ip_dst.s_addr) continue; if (inp->inp_faddr.s_addr != INADDR_ANY && inp->inp_faddr.s_addr != ip->ip_src.s_addr) continue; if (inp->inp_fport != 0 && inp->inp_fport != uh->uh_sport) continue; INP_RLOCK(inp); if (__predict_false(inp->inp_flags2 & INP_FREED)) { INP_RUNLOCK(inp); continue; } /* * XXXRW: Because we weren't holding either the inpcb * or the hash lock when we checked for a match * before, we should probably recheck now that the * inpcb lock is held. */ /* * Handle socket delivery policy for any-source * and source-specific multicast. [RFC3678] */ if (IN_MULTICAST(ntohl(ip->ip_dst.s_addr))) { struct ip_moptions *imo; struct sockaddr_in group; int blocked; imo = inp->inp_moptions; if (imo == NULL) { INP_RUNLOCK(inp); continue; } bzero(&group, sizeof(struct sockaddr_in)); group.sin_len = sizeof(struct sockaddr_in); group.sin_family = AF_INET; group.sin_addr = ip->ip_dst; blocked = imo_multi_filter(imo, ifp, (struct sockaddr *)&group, (struct sockaddr *)&udp_in[0]); if (blocked != MCAST_PASS) { if (blocked == MCAST_NOTGMEMBER) IPSTAT_INC(ips_notmember); if (blocked == MCAST_NOTSMEMBER || blocked == MCAST_MUTED) UDPSTAT_INC(udps_filtermcast); INP_RUNLOCK(inp); continue; } } if (last != NULL) { struct mbuf *n; if ((n = m_copym(m, 0, M_COPYALL, M_NOWAIT)) != NULL) { if (proto == IPPROTO_UDPLITE) UDPLITE_PROBE(receive, NULL, last, ip, last, uh); else UDP_PROBE(receive, NULL, last, ip, last, uh); if (udp_append(last, ip, n, iphlen, udp_in)) { goto inp_lost; } } INP_RUNLOCK(last); } last = inp; /* * Don't look for additional matches if this one does * not have either the SO_REUSEPORT or SO_REUSEADDR * socket options set. This heuristic avoids * searching through all pcbs in the common case of a * non-shared port. It assumes that an application * will never clear these options after setting them. */ if ((last->inp_socket->so_options & (SO_REUSEPORT|SO_REUSEPORT_LB|SO_REUSEADDR)) == 0) break; } if (last == NULL) { /* * No matching pcb found; discard datagram. (No need * to send an ICMP Port Unreachable for a broadcast * or multicast datgram.) */ UDPSTAT_INC(udps_noportbcast); if (inp) INP_RUNLOCK(inp); goto badunlocked; } if (proto == IPPROTO_UDPLITE) UDPLITE_PROBE(receive, NULL, last, ip, last, uh); else UDP_PROBE(receive, NULL, last, ip, last, uh); if (udp_append(last, ip, m, iphlen, udp_in) == 0) INP_RUNLOCK(last); inp_lost: return (IPPROTO_DONE); } /* * Locate pcb for datagram. */ /* * Grab info from PACKET_TAG_IPFORWARD tag prepended to the chain. */ if ((m->m_flags & M_IP_NEXTHOP) && (fwd_tag = m_tag_find(m, PACKET_TAG_IPFORWARD, NULL)) != NULL) { struct sockaddr_in *next_hop; next_hop = (struct sockaddr_in *)(fwd_tag + 1); /* * Transparently forwarded. Pretend to be the destination. * Already got one like this? */ inp = in_pcblookup_mbuf(pcbinfo, ip->ip_src, uh->uh_sport, ip->ip_dst, uh->uh_dport, INPLOOKUP_RLOCKPCB, ifp, m); if (!inp) { /* * It's new. Try to find the ambushing socket. * Because we've rewritten the destination address, * any hardware-generated hash is ignored. */ inp = in_pcblookup(pcbinfo, ip->ip_src, uh->uh_sport, next_hop->sin_addr, next_hop->sin_port ? htons(next_hop->sin_port) : uh->uh_dport, INPLOOKUP_WILDCARD | INPLOOKUP_RLOCKPCB, ifp); } /* Remove the tag from the packet. We don't need it anymore. */ m_tag_delete(m, fwd_tag); m->m_flags &= ~M_IP_NEXTHOP; } else inp = in_pcblookup_mbuf(pcbinfo, ip->ip_src, uh->uh_sport, ip->ip_dst, uh->uh_dport, INPLOOKUP_WILDCARD | INPLOOKUP_RLOCKPCB, ifp, m); if (inp == NULL) { - if (udp_log_in_vain) { + if (V_udp_log_in_vain) { char src[INET_ADDRSTRLEN]; char dst[INET_ADDRSTRLEN]; log(LOG_INFO, "Connection attempt to UDP %s:%d from %s:%d\n", inet_ntoa_r(ip->ip_dst, dst), ntohs(uh->uh_dport), inet_ntoa_r(ip->ip_src, src), ntohs(uh->uh_sport)); } if (proto == IPPROTO_UDPLITE) UDPLITE_PROBE(receive, NULL, NULL, ip, NULL, uh); else UDP_PROBE(receive, NULL, NULL, ip, NULL, uh); UDPSTAT_INC(udps_noport); if (m->m_flags & (M_BCAST | M_MCAST)) { UDPSTAT_INC(udps_noportbcast); goto badunlocked; } if (V_udp_blackhole) goto badunlocked; if (badport_bandlim(BANDLIM_ICMP_UNREACH) < 0) goto badunlocked; *ip = save_ip; icmp_error(m, ICMP_UNREACH, ICMP_UNREACH_PORT, 0, 0); return (IPPROTO_DONE); } /* * Check the minimum TTL for socket. */ INP_RLOCK_ASSERT(inp); if (inp->inp_ip_minttl && inp->inp_ip_minttl > ip->ip_ttl) { if (proto == IPPROTO_UDPLITE) UDPLITE_PROBE(receive, NULL, inp, ip, inp, uh); else UDP_PROBE(receive, NULL, inp, ip, inp, uh); INP_RUNLOCK(inp); m_freem(m); return (IPPROTO_DONE); } if (cscov_partial) { struct udpcb *up; up = intoudpcb(inp); if (up->u_rxcslen == 0 || up->u_rxcslen > len) { INP_RUNLOCK(inp); m_freem(m); return (IPPROTO_DONE); } } if (proto == IPPROTO_UDPLITE) UDPLITE_PROBE(receive, NULL, inp, ip, inp, uh); else UDP_PROBE(receive, NULL, inp, ip, inp, uh); if (udp_append(inp, ip, m, iphlen, udp_in) == 0) INP_RUNLOCK(inp); return (IPPROTO_DONE); badunlocked: m_freem(m); return (IPPROTO_DONE); } #endif /* INET */ /* * Notify a udp user of an asynchronous error; just wake up so that they can * collect error status. */ struct inpcb * udp_notify(struct inpcb *inp, int errno) { INP_WLOCK_ASSERT(inp); if ((errno == EHOSTUNREACH || errno == ENETUNREACH || errno == EHOSTDOWN) && inp->inp_route.ro_rt) { RTFREE(inp->inp_route.ro_rt); inp->inp_route.ro_rt = (struct rtentry *)NULL; } inp->inp_socket->so_error = errno; sorwakeup(inp->inp_socket); sowwakeup(inp->inp_socket); return (inp); } #ifdef INET static void udp_common_ctlinput(int cmd, struct sockaddr *sa, void *vip, struct inpcbinfo *pcbinfo) { struct ip *ip = vip; struct udphdr *uh; struct in_addr faddr; struct inpcb *inp; faddr = ((struct sockaddr_in *)sa)->sin_addr; if (sa->sa_family != AF_INET || faddr.s_addr == INADDR_ANY) return; if (PRC_IS_REDIRECT(cmd)) { /* signal EHOSTDOWN, as it flushes the cached route */ in_pcbnotifyall(&V_udbinfo, faddr, EHOSTDOWN, udp_notify); return; } /* * Hostdead is ugly because it goes linearly through all PCBs. * * XXX: We never get this from ICMP, otherwise it makes an excellent * DoS attack on machines with many connections. */ if (cmd == PRC_HOSTDEAD) ip = NULL; else if ((unsigned)cmd >= PRC_NCMDS || inetctlerrmap[cmd] == 0) return; if (ip != NULL) { uh = (struct udphdr *)((caddr_t)ip + (ip->ip_hl << 2)); inp = in_pcblookup(pcbinfo, faddr, uh->uh_dport, ip->ip_src, uh->uh_sport, INPLOOKUP_WLOCKPCB, NULL); if (inp != NULL) { INP_WLOCK_ASSERT(inp); if (inp->inp_socket != NULL) { udp_notify(inp, inetctlerrmap[cmd]); } INP_WUNLOCK(inp); } else { inp = in_pcblookup(pcbinfo, faddr, uh->uh_dport, ip->ip_src, uh->uh_sport, INPLOOKUP_WILDCARD | INPLOOKUP_RLOCKPCB, NULL); if (inp != NULL) { struct udpcb *up; void *ctx; udp_tun_icmp_t func; up = intoudpcb(inp); ctx = up->u_tun_ctx; func = up->u_icmp_func; INP_RUNLOCK(inp); if (func != NULL) (*func)(cmd, sa, vip, ctx); } } } else in_pcbnotifyall(pcbinfo, faddr, inetctlerrmap[cmd], udp_notify); } void udp_ctlinput(int cmd, struct sockaddr *sa, void *vip) { return (udp_common_ctlinput(cmd, sa, vip, &V_udbinfo)); } void udplite_ctlinput(int cmd, struct sockaddr *sa, void *vip) { return (udp_common_ctlinput(cmd, sa, vip, &V_ulitecbinfo)); } #endif /* INET */ static int udp_pcblist(SYSCTL_HANDLER_ARGS) { struct xinpgen xig; struct epoch_tracker et; struct inpcb *inp; int error; if (req->newptr != 0) return (EPERM); if (req->oldptr == 0) { int n; n = V_udbinfo.ipi_count; n += imax(n / 8, 10); req->oldidx = 2 * (sizeof xig) + n * sizeof(struct xinpcb); 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_udbinfo.ipi_count; xig.xig_gen = V_udbinfo.ipi_gencnt; xig.xig_sogen = so_gencnt; error = SYSCTL_OUT(req, &xig, sizeof xig); if (error) return (error); NET_EPOCH_ENTER(et); for (inp = CK_LIST_FIRST(V_udbinfo.ipi_listhead); inp != NULL; inp = CK_LIST_NEXT(inp, inp_list)) { INP_RLOCK(inp); if (inp->inp_gencnt <= xig.xig_gen && cr_canseeinpcb(req->td->td_ucred, inp) == 0) { struct xinpcb xi; in_pcbtoxinpcb(inp, &xi); INP_RUNLOCK(inp); error = SYSCTL_OUT(req, &xi, sizeof xi); if (error) break; } else INP_RUNLOCK(inp); } NET_EPOCH_EXIT(et); 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_udbinfo.ipi_gencnt; xig.xig_sogen = so_gencnt; xig.xig_count = V_udbinfo.ipi_count; error = SYSCTL_OUT(req, &xig, sizeof xig); } return (error); } SYSCTL_PROC(_net_inet_udp, UDPCTL_PCBLIST, pcblist, CTLTYPE_OPAQUE | CTLFLAG_RD, NULL, 0, udp_pcblist, "S,xinpcb", "List of active UDP sockets"); #ifdef INET static int udp_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_udbinfo, addrs[1].sin_addr, addrs[1].sin_port, addrs[0].sin_addr, addrs[0].sin_port, INPLOOKUP_WILDCARD | INPLOOKUP_RLOCKPCB, NULL); NET_EPOCH_EXIT(et); if (inp != NULL) { INP_RLOCK_ASSERT(inp); 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_udp, OID_AUTO, getcred, CTLTYPE_OPAQUE|CTLFLAG_RW|CTLFLAG_PRISON, 0, 0, udp_getcred, "S,xucred", "Get the xucred of a UDP connection"); #endif /* INET */ int udp_ctloutput(struct socket *so, struct sockopt *sopt) { struct inpcb *inp; struct udpcb *up; int isudplite, error, optval; error = 0; isudplite = (so->so_proto->pr_protocol == IPPROTO_UDPLITE) ? 1 : 0; inp = sotoinpcb(so); KASSERT(inp != NULL, ("%s: inp == NULL", __func__)); INP_WLOCK(inp); if (sopt->sopt_level != so->so_proto->pr_protocol) { #ifdef INET6 if (INP_CHECK_SOCKAF(so, AF_INET6)) { INP_WUNLOCK(inp); error = ip6_ctloutput(so, sopt); } #endif #if defined(INET) && defined(INET6) else #endif #ifdef INET { INP_WUNLOCK(inp); error = ip_ctloutput(so, sopt); } #endif return (error); } switch (sopt->sopt_dir) { case SOPT_SET: switch (sopt->sopt_name) { #if defined(IPSEC) || defined(IPSEC_SUPPORT) #ifdef INET case UDP_ENCAP: if (!IPSEC_ENABLED(ipv4)) { INP_WUNLOCK(inp); return (ENOPROTOOPT); } error = UDPENCAP_PCBCTL(inp, sopt); break; #endif /* INET */ #endif /* IPSEC */ case UDPLITE_SEND_CSCOV: case UDPLITE_RECV_CSCOV: if (!isudplite) { INP_WUNLOCK(inp); error = ENOPROTOOPT; break; } INP_WUNLOCK(inp); error = sooptcopyin(sopt, &optval, sizeof(optval), sizeof(optval)); if (error != 0) break; inp = sotoinpcb(so); KASSERT(inp != NULL, ("%s: inp == NULL", __func__)); INP_WLOCK(inp); up = intoudpcb(inp); KASSERT(up != NULL, ("%s: up == NULL", __func__)); if ((optval != 0 && optval < 8) || (optval > 65535)) { INP_WUNLOCK(inp); error = EINVAL; break; } if (sopt->sopt_name == UDPLITE_SEND_CSCOV) up->u_txcslen = optval; else up->u_rxcslen = optval; INP_WUNLOCK(inp); break; default: INP_WUNLOCK(inp); error = ENOPROTOOPT; break; } break; case SOPT_GET: switch (sopt->sopt_name) { #if defined(IPSEC) || defined(IPSEC_SUPPORT) #ifdef INET case UDP_ENCAP: if (!IPSEC_ENABLED(ipv4)) { INP_WUNLOCK(inp); return (ENOPROTOOPT); } error = UDPENCAP_PCBCTL(inp, sopt); break; #endif /* INET */ #endif /* IPSEC */ case UDPLITE_SEND_CSCOV: case UDPLITE_RECV_CSCOV: if (!isudplite) { INP_WUNLOCK(inp); error = ENOPROTOOPT; break; } up = intoudpcb(inp); KASSERT(up != NULL, ("%s: up == NULL", __func__)); if (sopt->sopt_name == UDPLITE_SEND_CSCOV) optval = up->u_txcslen; else optval = up->u_rxcslen; INP_WUNLOCK(inp); error = sooptcopyout(sopt, &optval, sizeof(optval)); break; default: INP_WUNLOCK(inp); error = ENOPROTOOPT; break; } break; } return (error); } #ifdef INET static int udp_output(struct inpcb *inp, struct mbuf *m, struct sockaddr *addr, struct mbuf *control, struct thread *td) { struct udpiphdr *ui; int len = m->m_pkthdr.len; struct in_addr faddr, laddr; struct cmsghdr *cm; struct inpcbinfo *pcbinfo; struct sockaddr_in *sin, src; struct epoch_tracker et; int cscov_partial = 0; int error = 0; int ipflags; u_short fport, lport; u_char tos; uint8_t pr; uint16_t cscov = 0; uint32_t flowid = 0; uint8_t flowtype = M_HASHTYPE_NONE; if (len + sizeof(struct udpiphdr) > IP_MAXPACKET) { if (control) m_freem(control); m_freem(m); return (EMSGSIZE); } src.sin_family = 0; sin = (struct sockaddr_in *)addr; /* * udp_output() may need to temporarily bind or connect the current * inpcb. As such, we don't know up front whether we will need the * pcbinfo lock or not. Do any work to decide what is needed up * front before acquiring any locks. * * We will need network epoch in either case, to safely lookup into * pcb hash. */ if (sin == NULL || (inp->inp_laddr.s_addr == INADDR_ANY && inp->inp_lport == 0)) INP_WLOCK(inp); else INP_RLOCK(inp); NET_EPOCH_ENTER(et); tos = inp->inp_ip_tos; if (control != NULL) { /* * XXX: Currently, we assume all the optional information is * stored in a single mbuf. */ if (control->m_next) { m_freem(control); error = EINVAL; goto release; } for (; control->m_len > 0; control->m_data += CMSG_ALIGN(cm->cmsg_len), control->m_len -= CMSG_ALIGN(cm->cmsg_len)) { cm = mtod(control, struct cmsghdr *); if (control->m_len < sizeof(*cm) || cm->cmsg_len == 0 || cm->cmsg_len > control->m_len) { error = EINVAL; break; } if (cm->cmsg_level != IPPROTO_IP) continue; switch (cm->cmsg_type) { case IP_SENDSRCADDR: if (cm->cmsg_len != CMSG_LEN(sizeof(struct in_addr))) { error = EINVAL; break; } bzero(&src, sizeof(src)); src.sin_family = AF_INET; src.sin_len = sizeof(src); src.sin_port = inp->inp_lport; src.sin_addr = *(struct in_addr *)CMSG_DATA(cm); break; case IP_TOS: if (cm->cmsg_len != CMSG_LEN(sizeof(u_char))) { error = EINVAL; break; } tos = *(u_char *)CMSG_DATA(cm); break; case IP_FLOWID: if (cm->cmsg_len != CMSG_LEN(sizeof(uint32_t))) { error = EINVAL; break; } flowid = *(uint32_t *) CMSG_DATA(cm); break; case IP_FLOWTYPE: if (cm->cmsg_len != CMSG_LEN(sizeof(uint32_t))) { error = EINVAL; break; } flowtype = *(uint32_t *) CMSG_DATA(cm); break; #ifdef RSS case IP_RSSBUCKETID: if (cm->cmsg_len != CMSG_LEN(sizeof(uint32_t))) { error = EINVAL; break; } /* This is just a placeholder for now */ break; #endif /* RSS */ default: error = ENOPROTOOPT; break; } if (error) break; } m_freem(control); } if (error) goto release; pr = inp->inp_socket->so_proto->pr_protocol; pcbinfo = udp_get_inpcbinfo(pr); /* * If the IP_SENDSRCADDR control message was specified, override the * source address for this datagram. Its use is invalidated if the * address thus specified is incomplete or clobbers other inpcbs. */ laddr = inp->inp_laddr; lport = inp->inp_lport; if (src.sin_family == AF_INET) { INP_HASH_LOCK_ASSERT(pcbinfo); if ((lport == 0) || (laddr.s_addr == INADDR_ANY && src.sin_addr.s_addr == INADDR_ANY)) { error = EINVAL; goto release; } error = in_pcbbind_setup(inp, (struct sockaddr *)&src, &laddr.s_addr, &lport, td->td_ucred); if (error) goto release; } /* * If a UDP socket has been connected, then a local address/port will * have been selected and bound. * * If a UDP socket has not been connected to, then an explicit * destination address must be used, in which case a local * address/port may not have been selected and bound. */ if (sin != NULL) { INP_LOCK_ASSERT(inp); if (inp->inp_faddr.s_addr != INADDR_ANY) { error = EISCONN; goto release; } /* * Jail may rewrite the destination address, so let it do * that before we use it. */ error = prison_remote_ip4(td->td_ucred, &sin->sin_addr); if (error) goto release; /* * If a local address or port hasn't yet been selected, or if * the destination address needs to be rewritten due to using * a special INADDR_ constant, invoke in_pcbconnect_setup() * to do the heavy lifting. Once a port is selected, we * commit the binding back to the socket; we also commit the * binding of the address if in jail. * * If we already have a valid binding and we're not * requesting a destination address rewrite, use a fast path. */ if (inp->inp_laddr.s_addr == INADDR_ANY || inp->inp_lport == 0 || sin->sin_addr.s_addr == INADDR_ANY || sin->sin_addr.s_addr == INADDR_BROADCAST) { INP_HASH_LOCK_ASSERT(pcbinfo); error = in_pcbconnect_setup(inp, addr, &laddr.s_addr, &lport, &faddr.s_addr, &fport, NULL, td->td_ucred); if (error) goto release; /* * XXXRW: Why not commit the port if the address is * !INADDR_ANY? */ /* Commit the local port if newly assigned. */ if (inp->inp_laddr.s_addr == INADDR_ANY && inp->inp_lport == 0) { INP_WLOCK_ASSERT(inp); /* * Remember addr if jailed, to prevent * rebinding. */ if (prison_flag(td->td_ucred, PR_IP4)) inp->inp_laddr = laddr; inp->inp_lport = lport; INP_HASH_WLOCK(pcbinfo); error = in_pcbinshash(inp); INP_HASH_WUNLOCK(pcbinfo); if (error != 0) { inp->inp_lport = 0; error = EAGAIN; goto release; } inp->inp_flags |= INP_ANONPORT; } } else { faddr = sin->sin_addr; fport = sin->sin_port; } } else { INP_LOCK_ASSERT(inp); faddr = inp->inp_faddr; fport = inp->inp_fport; if (faddr.s_addr == INADDR_ANY) { error = ENOTCONN; goto release; } } /* * Calculate data length and get a mbuf for UDP, IP, and possible * link-layer headers. Immediate slide the data pointer back forward * since we won't use that space at this layer. */ M_PREPEND(m, sizeof(struct udpiphdr) + max_linkhdr, M_NOWAIT); if (m == NULL) { error = ENOBUFS; goto release; } m->m_data += max_linkhdr; m->m_len -= max_linkhdr; m->m_pkthdr.len -= max_linkhdr; /* * Fill in mbuf with extended UDP header and addresses and length put * into network format. */ ui = mtod(m, struct udpiphdr *); bzero(ui->ui_x1, sizeof(ui->ui_x1)); /* XXX still needed? */ ui->ui_v = IPVERSION << 4; ui->ui_pr = pr; ui->ui_src = laddr; ui->ui_dst = faddr; ui->ui_sport = lport; ui->ui_dport = fport; ui->ui_ulen = htons((u_short)len + sizeof(struct udphdr)); if (pr == IPPROTO_UDPLITE) { struct udpcb *up; uint16_t plen; up = intoudpcb(inp); cscov = up->u_txcslen; plen = (u_short)len + sizeof(struct udphdr); if (cscov >= plen) cscov = 0; ui->ui_len = htons(plen); ui->ui_ulen = htons(cscov); /* * For UDP-Lite, checksum coverage length of zero means * the entire UDPLite packet is covered by the checksum. */ cscov_partial = (cscov == 0) ? 0 : 1; } /* * Set the Don't Fragment bit in the IP header. */ if (inp->inp_flags & INP_DONTFRAG) { struct ip *ip; ip = (struct ip *)&ui->ui_i; ip->ip_off |= htons(IP_DF); } ipflags = 0; if (inp->inp_socket->so_options & SO_DONTROUTE) ipflags |= IP_ROUTETOIF; if (inp->inp_socket->so_options & SO_BROADCAST) ipflags |= IP_ALLOWBROADCAST; if (inp->inp_flags & INP_ONESBCAST) ipflags |= IP_SENDONES; #ifdef MAC mac_inpcb_create_mbuf(inp, m); #endif /* * Set up checksum and output datagram. */ ui->ui_sum = 0; if (pr == IPPROTO_UDPLITE) { if (inp->inp_flags & INP_ONESBCAST) faddr.s_addr = INADDR_BROADCAST; if (cscov_partial) { if ((ui->ui_sum = in_cksum(m, sizeof(struct ip) + cscov)) == 0) ui->ui_sum = 0xffff; } else { if ((ui->ui_sum = in_cksum(m, sizeof(struct udpiphdr) + len)) == 0) ui->ui_sum = 0xffff; } } else if (V_udp_cksum) { if (inp->inp_flags & INP_ONESBCAST) faddr.s_addr = INADDR_BROADCAST; ui->ui_sum = in_pseudo(ui->ui_src.s_addr, faddr.s_addr, htons((u_short)len + sizeof(struct udphdr) + pr)); m->m_pkthdr.csum_flags = CSUM_UDP; m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); } ((struct ip *)ui)->ip_len = htons(sizeof(struct udpiphdr) + len); ((struct ip *)ui)->ip_ttl = inp->inp_ip_ttl; /* XXX */ ((struct ip *)ui)->ip_tos = tos; /* XXX */ UDPSTAT_INC(udps_opackets); /* * Setup flowid / RSS information for outbound socket. * * Once the UDP code decides to set a flowid some other way, * this allows the flowid to be overridden by userland. */ if (flowtype != M_HASHTYPE_NONE) { m->m_pkthdr.flowid = flowid; M_HASHTYPE_SET(m, flowtype); } #ifdef RSS else { uint32_t hash_val, hash_type; /* * Calculate an appropriate RSS hash for UDP and * UDP Lite. * * The called function will take care of figuring out * whether a 2-tuple or 4-tuple hash is required based * on the currently configured scheme. * * Later later on connected socket values should be * cached in the inpcb and reused, rather than constantly * re-calculating it. * * UDP Lite is a different protocol number and will * likely end up being hashed as a 2-tuple until * RSS / NICs grow UDP Lite protocol awareness. */ if (rss_proto_software_hash_v4(faddr, laddr, fport, lport, pr, &hash_val, &hash_type) == 0) { m->m_pkthdr.flowid = hash_val; M_HASHTYPE_SET(m, hash_type); } } /* * Don't override with the inp cached flowid value. * * Depending upon the kind of send being done, the inp * flowid/flowtype values may actually not be appropriate * for this particular socket send. * * We should either leave the flowid at zero (which is what is * currently done) or set it to some software generated * hash value based on the packet contents. */ ipflags |= IP_NODEFAULTFLOWID; #endif /* RSS */ if (pr == IPPROTO_UDPLITE) UDPLITE_PROBE(send, NULL, inp, &ui->ui_i, inp, &ui->ui_u); else UDP_PROBE(send, NULL, inp, &ui->ui_i, inp, &ui->ui_u); error = ip_output(m, inp->inp_options, INP_WLOCKED(inp) ? &inp->inp_route : NULL, ipflags, inp->inp_moptions, inp); INP_UNLOCK(inp); NET_EPOCH_EXIT(et); return (error); release: INP_UNLOCK(inp); NET_EPOCH_EXIT(et); m_freem(m); return (error); } static void udp_abort(struct socket *so) { struct inpcb *inp; struct inpcbinfo *pcbinfo; pcbinfo = udp_get_inpcbinfo(so->so_proto->pr_protocol); inp = sotoinpcb(so); KASSERT(inp != NULL, ("udp_abort: inp == NULL")); INP_WLOCK(inp); if (inp->inp_faddr.s_addr != INADDR_ANY) { INP_HASH_WLOCK(pcbinfo); in_pcbdisconnect(inp); inp->inp_laddr.s_addr = INADDR_ANY; INP_HASH_WUNLOCK(pcbinfo); soisdisconnected(so); } INP_WUNLOCK(inp); } static int udp_attach(struct socket *so, int proto, struct thread *td) { static uint32_t udp_flowid; struct inpcb *inp; struct inpcbinfo *pcbinfo; int error; pcbinfo = udp_get_inpcbinfo(so->so_proto->pr_protocol); inp = sotoinpcb(so); KASSERT(inp == NULL, ("udp_attach: inp != NULL")); error = soreserve(so, udp_sendspace, udp_recvspace); if (error) return (error); INP_INFO_WLOCK(pcbinfo); error = in_pcballoc(so, pcbinfo); if (error) { INP_INFO_WUNLOCK(pcbinfo); return (error); } inp = sotoinpcb(so); inp->inp_vflag |= INP_IPV4; inp->inp_ip_ttl = V_ip_defttl; inp->inp_flowid = atomic_fetchadd_int(&udp_flowid, 1); inp->inp_flowtype = M_HASHTYPE_OPAQUE; error = udp_newudpcb(inp); if (error) { in_pcbdetach(inp); in_pcbfree(inp); INP_INFO_WUNLOCK(pcbinfo); return (error); } INP_WUNLOCK(inp); INP_INFO_WUNLOCK(pcbinfo); return (0); } #endif /* INET */ int udp_set_kernel_tunneling(struct socket *so, udp_tun_func_t f, udp_tun_icmp_t i, void *ctx) { struct inpcb *inp; struct udpcb *up; KASSERT(so->so_type == SOCK_DGRAM, ("udp_set_kernel_tunneling: !dgram")); inp = sotoinpcb(so); KASSERT(inp != NULL, ("udp_set_kernel_tunneling: inp == NULL")); INP_WLOCK(inp); up = intoudpcb(inp); if ((up->u_tun_func != NULL) || (up->u_icmp_func != NULL)) { INP_WUNLOCK(inp); return (EBUSY); } up->u_tun_func = f; up->u_icmp_func = i; up->u_tun_ctx = ctx; INP_WUNLOCK(inp); return (0); } #ifdef INET static int udp_bind(struct socket *so, struct sockaddr *nam, struct thread *td) { struct inpcb *inp; struct inpcbinfo *pcbinfo; int error; pcbinfo = udp_get_inpcbinfo(so->so_proto->pr_protocol); inp = sotoinpcb(so); KASSERT(inp != NULL, ("udp_bind: inp == NULL")); INP_WLOCK(inp); INP_HASH_WLOCK(pcbinfo); error = in_pcbbind(inp, nam, td->td_ucred); INP_HASH_WUNLOCK(pcbinfo); INP_WUNLOCK(inp); return (error); } static void udp_close(struct socket *so) { struct inpcb *inp; struct inpcbinfo *pcbinfo; pcbinfo = udp_get_inpcbinfo(so->so_proto->pr_protocol); inp = sotoinpcb(so); KASSERT(inp != NULL, ("udp_close: inp == NULL")); INP_WLOCK(inp); if (inp->inp_faddr.s_addr != INADDR_ANY) { INP_HASH_WLOCK(pcbinfo); in_pcbdisconnect(inp); inp->inp_laddr.s_addr = INADDR_ANY; INP_HASH_WUNLOCK(pcbinfo); soisdisconnected(so); } INP_WUNLOCK(inp); } static int udp_connect(struct socket *so, struct sockaddr *nam, struct thread *td) { struct inpcb *inp; struct inpcbinfo *pcbinfo; struct sockaddr_in *sin; int error; pcbinfo = udp_get_inpcbinfo(so->so_proto->pr_protocol); inp = sotoinpcb(so); KASSERT(inp != NULL, ("udp_connect: inp == NULL")); INP_WLOCK(inp); if (inp->inp_faddr.s_addr != INADDR_ANY) { INP_WUNLOCK(inp); return (EISCONN); } sin = (struct sockaddr_in *)nam; error = prison_remote_ip4(td->td_ucred, &sin->sin_addr); if (error != 0) { INP_WUNLOCK(inp); return (error); } INP_HASH_WLOCK(pcbinfo); error = in_pcbconnect(inp, nam, td->td_ucred); INP_HASH_WUNLOCK(pcbinfo); if (error == 0) soisconnected(so); INP_WUNLOCK(inp); return (error); } static void udp_detach(struct socket *so) { struct inpcb *inp; struct inpcbinfo *pcbinfo; struct udpcb *up; pcbinfo = udp_get_inpcbinfo(so->so_proto->pr_protocol); inp = sotoinpcb(so); KASSERT(inp != NULL, ("udp_detach: inp == NULL")); KASSERT(inp->inp_faddr.s_addr == INADDR_ANY, ("udp_detach: not disconnected")); INP_INFO_WLOCK(pcbinfo); INP_WLOCK(inp); up = intoudpcb(inp); KASSERT(up != NULL, ("%s: up == NULL", __func__)); inp->inp_ppcb = NULL; in_pcbdetach(inp); in_pcbfree(inp); INP_INFO_WUNLOCK(pcbinfo); udp_discardcb(up); } static int udp_disconnect(struct socket *so) { struct inpcb *inp; struct inpcbinfo *pcbinfo; pcbinfo = udp_get_inpcbinfo(so->so_proto->pr_protocol); inp = sotoinpcb(so); KASSERT(inp != NULL, ("udp_disconnect: inp == NULL")); INP_WLOCK(inp); if (inp->inp_faddr.s_addr == INADDR_ANY) { INP_WUNLOCK(inp); return (ENOTCONN); } INP_HASH_WLOCK(pcbinfo); in_pcbdisconnect(inp); inp->inp_laddr.s_addr = INADDR_ANY; INP_HASH_WUNLOCK(pcbinfo); SOCK_LOCK(so); so->so_state &= ~SS_ISCONNECTED; /* XXX */ SOCK_UNLOCK(so); INP_WUNLOCK(inp); return (0); } static int udp_send(struct socket *so, int flags, struct mbuf *m, struct sockaddr *addr, struct mbuf *control, struct thread *td) { struct inpcb *inp; inp = sotoinpcb(so); KASSERT(inp != NULL, ("udp_send: inp == NULL")); return (udp_output(inp, m, addr, control, td)); } #endif /* INET */ int udp_shutdown(struct socket *so) { struct inpcb *inp; inp = sotoinpcb(so); KASSERT(inp != NULL, ("udp_shutdown: inp == NULL")); INP_WLOCK(inp); socantsendmore(so); INP_WUNLOCK(inp); return (0); } #ifdef INET struct pr_usrreqs udp_usrreqs = { .pru_abort = udp_abort, .pru_attach = udp_attach, .pru_bind = udp_bind, .pru_connect = udp_connect, .pru_control = in_control, .pru_detach = udp_detach, .pru_disconnect = udp_disconnect, .pru_peeraddr = in_getpeeraddr, .pru_send = udp_send, .pru_soreceive = soreceive_dgram, .pru_sosend = sosend_dgram, .pru_shutdown = udp_shutdown, .pru_sockaddr = in_getsockaddr, .pru_sosetlabel = in_pcbsosetlabel, .pru_close = udp_close, }; #endif /* INET */ Index: head/sys/netinet/udp_var.h =================================================================== --- head/sys/netinet/udp_var.h (revision 356526) +++ head/sys/netinet/udp_var.h (revision 356527) @@ -1,186 +1,187 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1982, 1986, 1989, 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. * * @(#)udp_var.h 8.1 (Berkeley) 6/10/93 * $FreeBSD$ */ #ifndef _NETINET_UDP_VAR_H_ #define _NETINET_UDP_VAR_H_ /* * UDP kernel structures and variables. */ struct udpiphdr { struct ipovly ui_i; /* overlaid ip structure */ struct udphdr ui_u; /* udp header */ }; #define ui_x1 ui_i.ih_x1 #define ui_v ui_i.ih_x1[0] #define ui_pr ui_i.ih_pr #define ui_len ui_i.ih_len #define ui_src ui_i.ih_src #define ui_dst ui_i.ih_dst #define ui_sport ui_u.uh_sport #define ui_dport ui_u.uh_dport #define ui_ulen ui_u.uh_ulen #define ui_sum ui_u.uh_sum struct inpcb; struct mbuf; typedef void(*udp_tun_func_t)(struct mbuf *, int, struct inpcb *, const struct sockaddr *, void *); typedef void(*udp_tun_icmp_t)(int, struct sockaddr *, void *, void *); /* * UDP control block; one per udp. */ struct udpcb { udp_tun_func_t u_tun_func; /* UDP kernel tunneling callback. */ udp_tun_icmp_t u_icmp_func; /* UDP kernel tunneling icmp callback */ u_int u_flags; /* Generic UDP flags. */ uint16_t u_rxcslen; /* Coverage for incoming datagrams. */ uint16_t u_txcslen; /* Coverage for outgoing datagrams. */ void *u_tun_ctx; /* Tunneling callback context. */ }; #define intoudpcb(ip) ((struct udpcb *)(ip)->inp_ppcb) #define sotoudpcb(so) (intoudpcb(sotoinpcb(so))) /* IPsec: ESP in UDP tunneling: */ #define UF_ESPINUDP_NON_IKE 0x00000001 /* w/ non-IKE marker .. */ /* .. per draft-ietf-ipsec-nat-t-ike-0[01], * and draft-ietf-ipsec-udp-encaps-(00/)01.txt */ #define UF_ESPINUDP 0x00000002 /* w/ non-ESP marker. */ struct udpstat { /* input statistics: */ uint64_t udps_ipackets; /* total input packets */ uint64_t udps_hdrops; /* packet shorter than header */ uint64_t udps_badsum; /* checksum error */ uint64_t udps_nosum; /* no checksum */ uint64_t udps_badlen; /* data length larger than packet */ uint64_t udps_noport; /* no socket on port */ uint64_t udps_noportbcast; /* of above, arrived as broadcast */ uint64_t udps_fullsock; /* not delivered, input socket full */ uint64_t udpps_pcbcachemiss; /* input packets missing pcb cache */ uint64_t udpps_pcbhashmiss; /* input packets not for hashed pcb */ /* output statistics: */ uint64_t udps_opackets; /* total output packets */ uint64_t udps_fastout; /* output packets on fast path */ /* of no socket on port, arrived as multicast */ uint64_t udps_noportmcast; uint64_t udps_filtermcast; /* blocked by multicast filter */ }; #ifdef _KERNEL #include VNET_PCPUSTAT_DECLARE(struct udpstat, udpstat); /* * In-kernel consumers can use these accessor macros directly to update * stats. */ #define UDPSTAT_ADD(name, val) \ VNET_PCPUSTAT_ADD(struct udpstat, udpstat, name, (val)) #define UDPSTAT_INC(name) UDPSTAT_ADD(name, 1) /* * Kernel module consumers must use this accessor macro. */ void kmod_udpstat_inc(int statnum); #define KMOD_UDPSTAT_INC(name) \ kmod_udpstat_inc(offsetof(struct udpstat, name) / sizeof(uint64_t)) #endif /* * Identifiers for UDP sysctl nodes. */ #define UDPCTL_CHECKSUM 1 /* checksum UDP packets */ #define UDPCTL_STATS 2 /* statistics (read-only) */ #define UDPCTL_MAXDGRAM 3 /* max datagram size */ #define UDPCTL_RECVSPACE 4 /* default receive buffer space */ #define UDPCTL_PCBLIST 5 /* list of PCBs for UDP sockets */ #ifdef _KERNEL #include SYSCTL_DECL(_net_inet_udp); extern struct pr_usrreqs udp_usrreqs; VNET_DECLARE(struct inpcbhead, udb); VNET_DECLARE(struct inpcbinfo, udbinfo); VNET_DECLARE(struct inpcbhead, ulitecb); VNET_DECLARE(struct inpcbinfo, ulitecbinfo); #define V_udb VNET(udb) #define V_udbinfo VNET(udbinfo) #define V_ulitecb VNET(ulitecb) #define V_ulitecbinfo VNET(ulitecbinfo) extern u_long udp_sendspace; extern u_long udp_recvspace; VNET_DECLARE(int, udp_cksum); VNET_DECLARE(int, udp_blackhole); +VNET_DECLARE(int, udp_log_in_vain); #define V_udp_cksum VNET(udp_cksum) #define V_udp_blackhole VNET(udp_blackhole) -extern int udp_log_in_vain; +#define V_udp_log_in_vain VNET(udp_log_in_vain) static __inline struct inpcbinfo * udp_get_inpcbinfo(int protocol) { return (protocol == IPPROTO_UDP) ? &V_udbinfo : &V_ulitecbinfo; } static __inline struct inpcbhead * udp_get_pcblist(int protocol) { return (protocol == IPPROTO_UDP) ? &V_udb : &V_ulitecb; } int udp_newudpcb(struct inpcb *); void udp_discardcb(struct udpcb *); void udp_ctlinput(int, struct sockaddr *, void *); void udplite_ctlinput(int, struct sockaddr *, void *); int udp_ctloutput(struct socket *, struct sockopt *); void udp_init(void); void udplite_init(void); int udp_input(struct mbuf **, int *, int); void udplite_input(struct mbuf *, int); struct inpcb *udp_notify(struct inpcb *inp, int errno); int udp_shutdown(struct socket *so); int udp_set_kernel_tunneling(struct socket *so, udp_tun_func_t f, udp_tun_icmp_t i, void *ctx); #endif /* _KERNEL */ #endif /* _NETINET_UDP_VAR_H_ */ Index: head/sys/netinet6/udp6_usrreq.c =================================================================== --- head/sys/netinet6/udp6_usrreq.c (revision 356526) +++ head/sys/netinet6/udp6_usrreq.c (revision 356527) @@ -1,1372 +1,1372 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (C) 1995, 1996, 1997, and 1998 WIDE Project. * Copyright (c) 2010-2011 Juniper Networks, Inc. * Copyright (c) 2014 Kevin Lo * 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: udp6_usrreq.c,v 1.27 2001/05/21 05:45:10 jinmei Exp $ * $KAME: udp6_output.c,v 1.31 2001/05/21 16:39:15 jinmei Exp $ */ /*- * 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. * * @(#)udp_usrreq.c 8.6 (Berkeley) 5/23/95 */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include "opt_ipsec.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 #include #include #include #include #include #include /* * UDP protocol implementation. * Per RFC 768, August, 1980. */ extern struct protosw inetsw[]; static void udp6_detach(struct socket *so); static int udp6_append(struct inpcb *inp, struct mbuf *n, int off, struct sockaddr_in6 *fromsa) { struct socket *so; struct mbuf *opts = NULL, *tmp_opts; struct udpcb *up; INP_LOCK_ASSERT(inp); /* * Engage the tunneling protocol. */ up = intoudpcb(inp); if (up->u_tun_func != NULL) { in_pcbref(inp); INP_RUNLOCK(inp); (*up->u_tun_func)(n, off, inp, (struct sockaddr *)&fromsa[0], up->u_tun_ctx); INP_RLOCK(inp); return (in_pcbrele_rlocked(inp)); } #if defined(IPSEC) || defined(IPSEC_SUPPORT) /* Check AH/ESP integrity. */ if (IPSEC_ENABLED(ipv6)) { if (IPSEC_CHECK_POLICY(ipv6, n, inp) != 0) { m_freem(n); return (0); } } #endif /* IPSEC */ #ifdef MAC if (mac_inpcb_check_deliver(inp, n) != 0) { m_freem(n); return (0); } #endif opts = NULL; if (inp->inp_flags & INP_CONTROLOPTS || inp->inp_socket->so_options & SO_TIMESTAMP) ip6_savecontrol(inp, n, &opts); if ((inp->inp_vflag & INP_IPV6) && (inp->inp_flags2 & INP_ORIGDSTADDR)) { tmp_opts = sbcreatecontrol((caddr_t)&fromsa[1], sizeof(struct sockaddr_in6), IPV6_ORIGDSTADDR, IPPROTO_IPV6); if (tmp_opts) { if (opts) { tmp_opts->m_next = opts; opts = tmp_opts; } else opts = tmp_opts; } } m_adj(n, off + sizeof(struct udphdr)); so = inp->inp_socket; SOCKBUF_LOCK(&so->so_rcv); if (sbappendaddr_locked(&so->so_rcv, (struct sockaddr *)&fromsa[0], n, opts) == 0) { SOCKBUF_UNLOCK(&so->so_rcv); m_freem(n); if (opts) m_freem(opts); UDPSTAT_INC(udps_fullsock); } else sorwakeup_locked(so); return (0); } int udp6_input(struct mbuf **mp, int *offp, int proto) { struct mbuf *m = *mp; struct ifnet *ifp; struct ip6_hdr *ip6; struct udphdr *uh; struct inpcb *inp; struct inpcbinfo *pcbinfo; struct udpcb *up; int off = *offp; int cscov_partial; int plen, ulen; struct sockaddr_in6 fromsa[2]; struct m_tag *fwd_tag; uint16_t uh_sum; uint8_t nxt; NET_EPOCH_ASSERT(); ifp = m->m_pkthdr.rcvif; if (m->m_len < off + sizeof(struct udphdr)) { m = m_pullup(m, off + sizeof(struct udphdr)); if (m == NULL) { IP6STAT_INC(ip6s_exthdrtoolong); *mp = NULL; return (IPPROTO_DONE); } } ip6 = mtod(m, struct ip6_hdr *); uh = (struct udphdr *)((caddr_t)ip6 + off); UDPSTAT_INC(udps_ipackets); /* * Destination port of 0 is illegal, based on RFC768. */ if (uh->uh_dport == 0) goto badunlocked; plen = ntohs(ip6->ip6_plen) - off + sizeof(*ip6); ulen = ntohs((u_short)uh->uh_ulen); nxt = proto; cscov_partial = (nxt == IPPROTO_UDPLITE) ? 1 : 0; if (nxt == IPPROTO_UDPLITE) { /* Zero means checksum over the complete packet. */ if (ulen == 0) ulen = plen; if (ulen == plen) cscov_partial = 0; if ((ulen < sizeof(struct udphdr)) || (ulen > plen)) { /* XXX: What is the right UDPLite MIB counter? */ goto badunlocked; } if (uh->uh_sum == 0) { /* XXX: What is the right UDPLite MIB counter? */ goto badunlocked; } } else { if ((ulen < sizeof(struct udphdr)) || (plen != ulen)) { UDPSTAT_INC(udps_badlen); goto badunlocked; } if (uh->uh_sum == 0) { UDPSTAT_INC(udps_nosum); goto badunlocked; } } if ((m->m_pkthdr.csum_flags & CSUM_DATA_VALID_IPV6) && !cscov_partial) { if (m->m_pkthdr.csum_flags & CSUM_PSEUDO_HDR) uh_sum = m->m_pkthdr.csum_data; else uh_sum = in6_cksum_pseudo(ip6, ulen, nxt, m->m_pkthdr.csum_data); uh_sum ^= 0xffff; } else uh_sum = in6_cksum_partial(m, nxt, off, plen, ulen); if (uh_sum != 0) { UDPSTAT_INC(udps_badsum); goto badunlocked; } /* * Construct sockaddr format source address. */ init_sin6(&fromsa[0], m, 0); fromsa[0].sin6_port = uh->uh_sport; init_sin6(&fromsa[1], m, 1); fromsa[1].sin6_port = uh->uh_dport; pcbinfo = udp_get_inpcbinfo(nxt); if (IN6_IS_ADDR_MULTICAST(&ip6->ip6_dst)) { struct inpcb *last; struct inpcbhead *pcblist; struct ip6_moptions *imo; /* * In the event that laddr should be set to the link-local * address (this happens in RIPng), the multicast address * specified in the received packet will not match laddr. To * handle this situation, matching is relaxed if the * receiving interface is the same as one specified in the * socket and if the destination multicast address matches * one of the multicast groups specified in the socket. */ /* * KAME note: traditionally we dropped udpiphdr from mbuf * here. We need udphdr for IPsec processing so we do that * later. */ pcblist = udp_get_pcblist(nxt); last = NULL; CK_LIST_FOREACH(inp, pcblist, inp_list) { if ((inp->inp_vflag & INP_IPV6) == 0) continue; if (inp->inp_lport != uh->uh_dport) continue; if (inp->inp_fport != 0 && inp->inp_fport != uh->uh_sport) continue; if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr)) { if (!IN6_ARE_ADDR_EQUAL(&inp->in6p_laddr, &ip6->ip6_dst)) continue; } if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr)) { if (!IN6_ARE_ADDR_EQUAL(&inp->in6p_faddr, &ip6->ip6_src) || inp->inp_fport != uh->uh_sport) continue; } /* * XXXRW: Because we weren't holding either the inpcb * or the hash lock when we checked for a match * before, we should probably recheck now that the * inpcb lock is (supposed to be) held. */ /* * Handle socket delivery policy for any-source * and source-specific multicast. [RFC3678] */ imo = inp->in6p_moptions; if (imo && IN6_IS_ADDR_MULTICAST(&ip6->ip6_dst)) { struct sockaddr_in6 mcaddr; int blocked; INP_RLOCK(inp); if (__predict_false(inp->inp_flags2 & INP_FREED)) { INP_RUNLOCK(inp); continue; } bzero(&mcaddr, sizeof(struct sockaddr_in6)); mcaddr.sin6_len = sizeof(struct sockaddr_in6); mcaddr.sin6_family = AF_INET6; mcaddr.sin6_addr = ip6->ip6_dst; blocked = im6o_mc_filter(imo, ifp, (struct sockaddr *)&mcaddr, (struct sockaddr *)&fromsa[0]); if (blocked != MCAST_PASS) { if (blocked == MCAST_NOTGMEMBER) IP6STAT_INC(ip6s_notmember); if (blocked == MCAST_NOTSMEMBER || blocked == MCAST_MUTED) UDPSTAT_INC(udps_filtermcast); INP_RUNLOCK(inp); /* XXX */ continue; } INP_RUNLOCK(inp); } if (last != NULL) { struct mbuf *n; if ((n = m_copym(m, 0, M_COPYALL, M_NOWAIT)) != NULL) { INP_RLOCK(last); if (__predict_true(last->inp_flags2 & INP_FREED) == 0) { if (nxt == IPPROTO_UDPLITE) UDPLITE_PROBE(receive, NULL, last, ip6, last, uh); else UDP_PROBE(receive, NULL, last, ip6, last, uh); if (udp6_append(last, n, off, fromsa)) { /* XXX-BZ do we leak m here? */ *mp = NULL; return (IPPROTO_DONE); } } INP_RUNLOCK(last); } } last = inp; /* * Don't look for additional matches if this one does * not have either the SO_REUSEPORT or SO_REUSEADDR * socket options set. This heuristic avoids * searching through all pcbs in the common case of a * non-shared port. It assumes that an application * will never clear these options after setting them. */ if ((last->inp_socket->so_options & (SO_REUSEPORT|SO_REUSEPORT_LB|SO_REUSEADDR)) == 0) break; } if (last == NULL) { /* * No matching pcb found; discard datagram. (No need * to send an ICMP Port Unreachable for a broadcast * or multicast datgram.) */ UDPSTAT_INC(udps_noport); UDPSTAT_INC(udps_noportmcast); goto badunlocked; } INP_RLOCK(last); if (__predict_true(last->inp_flags2 & INP_FREED) == 0) { if (nxt == IPPROTO_UDPLITE) UDPLITE_PROBE(receive, NULL, last, ip6, last, uh); else UDP_PROBE(receive, NULL, last, ip6, last, uh); if (udp6_append(last, m, off, fromsa) == 0) INP_RUNLOCK(last); } else INP_RUNLOCK(last); *mp = NULL; return (IPPROTO_DONE); } /* * Locate pcb for datagram. */ /* * Grab info from PACKET_TAG_IPFORWARD tag prepended to the chain. */ if ((m->m_flags & M_IP6_NEXTHOP) && (fwd_tag = m_tag_find(m, PACKET_TAG_IPFORWARD, NULL)) != NULL) { struct sockaddr_in6 *next_hop6; next_hop6 = (struct sockaddr_in6 *)(fwd_tag + 1); /* * Transparently forwarded. Pretend to be the destination. * Already got one like this? */ inp = in6_pcblookup_mbuf(pcbinfo, &ip6->ip6_src, uh->uh_sport, &ip6->ip6_dst, uh->uh_dport, INPLOOKUP_RLOCKPCB, m->m_pkthdr.rcvif, m); if (!inp) { /* * It's new. Try to find the ambushing socket. * Because we've rewritten the destination address, * any hardware-generated hash is ignored. */ inp = in6_pcblookup(pcbinfo, &ip6->ip6_src, uh->uh_sport, &next_hop6->sin6_addr, next_hop6->sin6_port ? htons(next_hop6->sin6_port) : uh->uh_dport, INPLOOKUP_WILDCARD | INPLOOKUP_RLOCKPCB, m->m_pkthdr.rcvif); } /* Remove the tag from the packet. We don't need it anymore. */ m_tag_delete(m, fwd_tag); m->m_flags &= ~M_IP6_NEXTHOP; } else inp = in6_pcblookup_mbuf(pcbinfo, &ip6->ip6_src, uh->uh_sport, &ip6->ip6_dst, uh->uh_dport, INPLOOKUP_WILDCARD | INPLOOKUP_RLOCKPCB, m->m_pkthdr.rcvif, m); if (inp == NULL) { - if (udp_log_in_vain) { + if (V_udp_log_in_vain) { char ip6bufs[INET6_ADDRSTRLEN]; char ip6bufd[INET6_ADDRSTRLEN]; log(LOG_INFO, "Connection attempt to UDP [%s]:%d from [%s]:%d\n", ip6_sprintf(ip6bufd, &ip6->ip6_dst), ntohs(uh->uh_dport), ip6_sprintf(ip6bufs, &ip6->ip6_src), ntohs(uh->uh_sport)); } if (nxt == IPPROTO_UDPLITE) UDPLITE_PROBE(receive, NULL, NULL, ip6, NULL, uh); else UDP_PROBE(receive, NULL, NULL, ip6, NULL, uh); UDPSTAT_INC(udps_noport); if (m->m_flags & M_MCAST) { printf("UDP6: M_MCAST is set in a unicast packet.\n"); UDPSTAT_INC(udps_noportmcast); goto badunlocked; } if (V_udp_blackhole) goto badunlocked; icmp6_error(m, ICMP6_DST_UNREACH, ICMP6_DST_UNREACH_NOPORT, 0); *mp = NULL; return (IPPROTO_DONE); } INP_RLOCK_ASSERT(inp); up = intoudpcb(inp); if (cscov_partial) { if (up->u_rxcslen == 0 || up->u_rxcslen > ulen) { INP_RUNLOCK(inp); m_freem(m); *mp = NULL; return (IPPROTO_DONE); } } if (nxt == IPPROTO_UDPLITE) UDPLITE_PROBE(receive, NULL, inp, ip6, inp, uh); else UDP_PROBE(receive, NULL, inp, ip6, inp, uh); if (udp6_append(inp, m, off, fromsa) == 0) INP_RUNLOCK(inp); *mp = NULL; return (IPPROTO_DONE); badunlocked: m_freem(m); *mp = NULL; return (IPPROTO_DONE); } static void udp6_common_ctlinput(int cmd, struct sockaddr *sa, void *d, struct inpcbinfo *pcbinfo) { struct udphdr uh; struct ip6_hdr *ip6; struct mbuf *m; int off = 0; struct ip6ctlparam *ip6cp = NULL; const struct sockaddr_in6 *sa6_src = NULL; void *cmdarg; struct inpcb *(*notify)(struct inpcb *, int) = udp_notify; struct udp_portonly { u_int16_t uh_sport; u_int16_t uh_dport; } *uhp; if (sa->sa_family != AF_INET6 || sa->sa_len != sizeof(struct sockaddr_in6)) return; if ((unsigned)cmd >= PRC_NCMDS) return; if (PRC_IS_REDIRECT(cmd)) notify = in6_rtchange, d = NULL; else if (cmd == PRC_HOSTDEAD) d = NULL; else if (inet6ctlerrmap[cmd] == 0) return; /* if the parameter is from icmp6, decode it. */ if (d != NULL) { ip6cp = (struct ip6ctlparam *)d; m = ip6cp->ip6c_m; ip6 = ip6cp->ip6c_ip6; off = ip6cp->ip6c_off; cmdarg = ip6cp->ip6c_cmdarg; sa6_src = ip6cp->ip6c_src; } else { m = NULL; ip6 = NULL; cmdarg = NULL; sa6_src = &sa6_any; } if (ip6) { /* * XXX: We assume that when IPV6 is non NULL, * M and OFF are valid. */ /* Check if we can safely examine src and dst ports. */ if (m->m_pkthdr.len < off + sizeof(*uhp)) return; bzero(&uh, sizeof(uh)); m_copydata(m, off, sizeof(*uhp), (caddr_t)&uh); if (!PRC_IS_REDIRECT(cmd)) { /* Check to see if its tunneled */ struct inpcb *inp; inp = in6_pcblookup_mbuf(pcbinfo, &ip6->ip6_dst, uh.uh_dport, &ip6->ip6_src, uh.uh_sport, INPLOOKUP_WILDCARD | INPLOOKUP_RLOCKPCB, m->m_pkthdr.rcvif, m); if (inp != NULL) { struct udpcb *up; up = intoudpcb(inp); if (up->u_icmp_func) { /* Yes it is. */ INP_RUNLOCK(inp); (*up->u_icmp_func)(cmd, (struct sockaddr *)ip6cp->ip6c_src, d, up->u_tun_ctx); return; } else { /* Can't find it. */ INP_RUNLOCK(inp); } } } (void)in6_pcbnotify(pcbinfo, sa, uh.uh_dport, (struct sockaddr *)ip6cp->ip6c_src, uh.uh_sport, cmd, cmdarg, notify); } else (void)in6_pcbnotify(pcbinfo, sa, 0, (const struct sockaddr *)sa6_src, 0, cmd, cmdarg, notify); } void udp6_ctlinput(int cmd, struct sockaddr *sa, void *d) { return (udp6_common_ctlinput(cmd, sa, d, &V_udbinfo)); } void udplite6_ctlinput(int cmd, struct sockaddr *sa, void *d) { return (udp6_common_ctlinput(cmd, sa, d, &V_ulitecbinfo)); } static int udp6_getcred(SYSCTL_HANDLER_ARGS) { struct xucred xuc; struct sockaddr_in6 addrs[2]; struct epoch_tracker et; struct inpcb *inp; int error; error = priv_check(req->td, PRIV_NETINET_GETCRED); if (error) return (error); if (req->newlen != sizeof(addrs)) return (EINVAL); if (req->oldlen != sizeof(struct xucred)) return (EINVAL); 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); } NET_EPOCH_ENTER(et); inp = in6_pcblookup(&V_udbinfo, &addrs[1].sin6_addr, addrs[1].sin6_port, &addrs[0].sin6_addr, addrs[0].sin6_port, INPLOOKUP_WILDCARD | INPLOOKUP_RLOCKPCB, NULL); NET_EPOCH_EXIT(et); if (inp != NULL) { INP_RLOCK_ASSERT(inp); if (inp->inp_socket == NULL) error = ENOENT; if (error == 0) error = cr_canseesocket(req->td->td_ucred, inp->inp_socket); 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_udp6, OID_AUTO, getcred, CTLTYPE_OPAQUE|CTLFLAG_RW, 0, 0, udp6_getcred, "S,xucred", "Get the xucred of a UDP6 connection"); static int udp6_output(struct socket *so, int flags_arg, struct mbuf *m, struct sockaddr *addr6, struct mbuf *control, struct thread *td) { struct inpcb *inp; struct ip6_hdr *ip6; struct udphdr *udp6; struct in6_addr *laddr, *faddr, in6a; struct ip6_pktopts *optp, opt; struct sockaddr_in6 *sin6, tmp; struct epoch_tracker et; int cscov_partial, error, flags, hlen, scope_ambiguous; u_int32_t ulen, plen; uint16_t cscov; u_short fport; uint8_t nxt; /* addr6 has been validated in udp6_send(). */ sin6 = (struct sockaddr_in6 *)addr6; /* * In contrast to to IPv4 we do not validate the max. packet length * here due to IPv6 Jumbograms (RFC2675). */ scope_ambiguous = 0; if (sin6) { /* Protect *addr6 from overwrites. */ tmp = *sin6; sin6 = &tmp; /* * Application should provide a proper zone ID or the use of * default zone IDs should be enabled. Unfortunately, some * applications do not behave as it should, so we need a * workaround. Even if an appropriate ID is not determined, * we'll see if we can determine the outgoing interface. If we * can, determine the zone ID based on the interface below. */ if (sin6->sin6_scope_id == 0 && !V_ip6_use_defzone) scope_ambiguous = 1; if ((error = sa6_embedscope(sin6, V_ip6_use_defzone)) != 0) { if (control) m_freem(control); m_freem(m); return (error); } } inp = sotoinpcb(so); KASSERT(inp != NULL, ("%s: inp == NULL", __func__)); /* * In the following cases we want a write lock on the inp for either * local operations or for possible route cache updates in the IPv6 * output path: * - on connected sockets (sin6 is NULL) for route cache updates, * - when we are not bound to an address and source port (it is * in6_pcbsetport() which will require the write lock). * * We check the inp fields before actually locking the inp, so * here exists a race, and we may WLOCK the inp and end with already * bound one by other thread. This is fine. */ if (sin6 == NULL || (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr) && inp->inp_lport == 0)) INP_WLOCK(inp); else INP_RLOCK(inp); nxt = (inp->inp_socket->so_proto->pr_protocol == IPPROTO_UDP) ? IPPROTO_UDP : IPPROTO_UDPLITE; #ifdef INET if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0) { int hasv4addr; if (sin6 == NULL) hasv4addr = (inp->inp_vflag & INP_IPV4); else hasv4addr = IN6_IS_ADDR_V4MAPPED(&sin6->sin6_addr) ? 1 : 0; if (hasv4addr) { struct pr_usrreqs *pru; /* * XXXRW: We release UDP-layer locks before calling * udp_send() in order to avoid recursion. However, * this does mean there is a short window where inp's * fields are unstable. Could this lead to a * potential race in which the factors causing us to * select the UDPv4 output routine are invalidated? */ INP_UNLOCK(inp); if (sin6) in6_sin6_2_sin_in_sock((struct sockaddr *)sin6); pru = inetsw[ip_protox[nxt]].pr_usrreqs; /* addr will just be freed in sendit(). */ return ((*pru->pru_send)(so, flags_arg, m, (struct sockaddr *)sin6, control, td)); } } else #endif if (sin6 && IN6_IS_ADDR_V4MAPPED(&sin6->sin6_addr)) { /* * Given this is either an IPv6-only socket or no INET is * supported we will fail the send if the given destination * address is a v4mapped address. * * XXXGL: do we leak m and control? */ INP_UNLOCK(inp); return (EINVAL); } if (control) { if ((error = ip6_setpktopts(control, &opt, inp->in6p_outputopts, td->td_ucred, nxt)) != 0) { INP_UNLOCK(inp); ip6_clearpktopts(&opt, -1); if (control) m_freem(control); m_freem(m); return (error); } optp = &opt; } else optp = inp->in6p_outputopts; NET_EPOCH_ENTER(et); if (sin6) { /* * Since we saw no essential reason for calling in_pcbconnect, * we get rid of such kind of logic, and call in6_selectsrc * and in6_pcbsetport in order to fill in the local address * and the local port. */ if (sin6->sin6_port == 0) { error = EADDRNOTAVAIL; goto release; } if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr)) { /* how about ::ffff:0.0.0.0 case? */ error = EISCONN; goto release; } /* * Given we handle the v4mapped case in the INET block above * assert here that it must not happen anymore. */ KASSERT(!IN6_IS_ADDR_V4MAPPED(&sin6->sin6_addr), ("%s: sin6(%p)->sin6_addr is v4mapped which we " "should have handled.", __func__, sin6)); /* This only requires read-locking. */ error = in6_selectsrc_socket(sin6, optp, inp, td->td_ucred, scope_ambiguous, &in6a, NULL); if (error) goto release; laddr = &in6a; if (inp->inp_lport == 0) { struct inpcbinfo *pcbinfo; INP_WLOCK_ASSERT(inp); pcbinfo = udp_get_inpcbinfo(so->so_proto->pr_protocol); INP_HASH_WLOCK(pcbinfo); error = in6_pcbsetport(laddr, inp, td->td_ucred); INP_HASH_WUNLOCK(pcbinfo); if (error != 0) { /* Undo an address bind that may have occurred. */ inp->in6p_laddr = in6addr_any; goto release; } } faddr = &sin6->sin6_addr; fport = sin6->sin6_port; /* allow 0 port */ } else { if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr)) { error = ENOTCONN; goto release; } laddr = &inp->in6p_laddr; faddr = &inp->in6p_faddr; fport = inp->inp_fport; } ulen = m->m_pkthdr.len; plen = sizeof(struct udphdr) + ulen; hlen = sizeof(struct ip6_hdr); /* * Calculate data length and get a mbuf * for UDP and IP6 headers. */ M_PREPEND(m, hlen + sizeof(struct udphdr), M_NOWAIT); if (m == NULL) { error = ENOBUFS; goto release; } /* * Stuff checksum and output datagram. */ cscov = cscov_partial = 0; udp6 = (struct udphdr *)(mtod(m, caddr_t) + hlen); udp6->uh_sport = inp->inp_lport; /* lport is always set in the PCB */ udp6->uh_dport = fport; if (nxt == IPPROTO_UDPLITE) { struct udpcb *up; up = intoudpcb(inp); cscov = up->u_txcslen; if (cscov >= plen) cscov = 0; udp6->uh_ulen = htons(cscov); /* * For UDP-Lite, checksum coverage length of zero means * the entire UDPLite packet is covered by the checksum. */ cscov_partial = (cscov == 0) ? 0 : 1; } else if (plen <= 0xffff) udp6->uh_ulen = htons((u_short)plen); else udp6->uh_ulen = 0; udp6->uh_sum = 0; ip6 = mtod(m, struct ip6_hdr *); ip6->ip6_flow = inp->inp_flow & IPV6_FLOWINFO_MASK; ip6->ip6_vfc &= ~IPV6_VERSION_MASK; ip6->ip6_vfc |= IPV6_VERSION; ip6->ip6_plen = htons((u_short)plen); ip6->ip6_nxt = nxt; ip6->ip6_hlim = in6_selecthlim(inp, NULL); ip6->ip6_src = *laddr; ip6->ip6_dst = *faddr; #ifdef MAC mac_inpcb_create_mbuf(inp, m); #endif if (cscov_partial) { if ((udp6->uh_sum = in6_cksum_partial(m, nxt, sizeof(struct ip6_hdr), plen, cscov)) == 0) udp6->uh_sum = 0xffff; } else { udp6->uh_sum = in6_cksum_pseudo(ip6, plen, nxt, 0); m->m_pkthdr.csum_flags = CSUM_UDP_IPV6; m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum); } flags = 0; #ifdef RSS { uint32_t hash_val, hash_type; uint8_t pr; pr = inp->inp_socket->so_proto->pr_protocol; /* * Calculate an appropriate RSS hash for UDP and * UDP Lite. * * The called function will take care of figuring out * whether a 2-tuple or 4-tuple hash is required based * on the currently configured scheme. * * Later later on connected socket values should be * cached in the inpcb and reused, rather than constantly * re-calculating it. * * UDP Lite is a different protocol number and will * likely end up being hashed as a 2-tuple until * RSS / NICs grow UDP Lite protocol awareness. */ if (rss_proto_software_hash_v6(faddr, laddr, fport, inp->inp_lport, pr, &hash_val, &hash_type) == 0) { m->m_pkthdr.flowid = hash_val; M_HASHTYPE_SET(m, hash_type); } /* * Don't override with the inp cached flowid. * * Until the whole UDP path is vetted, it may actually * be incorrect. */ flags |= IP_NODEFAULTFLOWID; } #endif UDPSTAT_INC(udps_opackets); if (nxt == IPPROTO_UDPLITE) UDPLITE_PROBE(send, NULL, inp, ip6, inp, udp6); else UDP_PROBE(send, NULL, inp, ip6, inp, udp6); error = ip6_output(m, optp, INP_WLOCKED(inp) ? &inp->inp_route6 : NULL, flags, inp->in6p_moptions, NULL, inp); INP_UNLOCK(inp); NET_EPOCH_EXIT(et); if (control) { ip6_clearpktopts(&opt, -1); m_freem(control); } return (error); release: INP_UNLOCK(inp); NET_EPOCH_EXIT(et); if (control) { ip6_clearpktopts(&opt, -1); m_freem(control); } m_freem(m); return (error); } static void udp6_abort(struct socket *so) { struct inpcb *inp; struct inpcbinfo *pcbinfo; pcbinfo = udp_get_inpcbinfo(so->so_proto->pr_protocol); inp = sotoinpcb(so); KASSERT(inp != NULL, ("udp6_abort: inp == NULL")); INP_WLOCK(inp); #ifdef INET if (inp->inp_vflag & INP_IPV4) { struct pr_usrreqs *pru; uint8_t nxt; nxt = (inp->inp_socket->so_proto->pr_protocol == IPPROTO_UDP) ? IPPROTO_UDP : IPPROTO_UDPLITE; INP_WUNLOCK(inp); pru = inetsw[ip_protox[nxt]].pr_usrreqs; (*pru->pru_abort)(so); return; } #endif if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr)) { INP_HASH_WLOCK(pcbinfo); in6_pcbdisconnect(inp); inp->in6p_laddr = in6addr_any; INP_HASH_WUNLOCK(pcbinfo); soisdisconnected(so); } INP_WUNLOCK(inp); } static int udp6_attach(struct socket *so, int proto, struct thread *td) { struct inpcb *inp; struct inpcbinfo *pcbinfo; int error; pcbinfo = udp_get_inpcbinfo(so->so_proto->pr_protocol); inp = sotoinpcb(so); KASSERT(inp == NULL, ("udp6_attach: inp != NULL")); if (so->so_snd.sb_hiwat == 0 || so->so_rcv.sb_hiwat == 0) { error = soreserve(so, udp_sendspace, udp_recvspace); if (error) return (error); } INP_INFO_WLOCK(pcbinfo); error = in_pcballoc(so, pcbinfo); if (error) { INP_INFO_WUNLOCK(pcbinfo); return (error); } inp = (struct inpcb *)so->so_pcb; inp->inp_vflag |= INP_IPV6; if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0) inp->inp_vflag |= INP_IPV4; inp->in6p_hops = -1; /* use kernel default */ inp->in6p_cksum = -1; /* just to be sure */ /* * XXX: ugly!! * 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; error = udp_newudpcb(inp); if (error) { in_pcbdetach(inp); in_pcbfree(inp); INP_INFO_WUNLOCK(pcbinfo); return (error); } INP_WUNLOCK(inp); INP_INFO_WUNLOCK(pcbinfo); return (0); } static int udp6_bind(struct socket *so, struct sockaddr *nam, struct thread *td) { struct inpcb *inp; struct inpcbinfo *pcbinfo; int error; u_char vflagsav; pcbinfo = udp_get_inpcbinfo(so->so_proto->pr_protocol); inp = sotoinpcb(so); KASSERT(inp != NULL, ("udp6_bind: inp == NULL")); INP_WLOCK(inp); INP_HASH_WLOCK(pcbinfo); vflagsav = inp->inp_vflag; inp->inp_vflag &= ~INP_IPV4; inp->inp_vflag |= INP_IPV6; if ((inp->inp_flags & IN6P_IPV6_V6ONLY) == 0) { struct sockaddr_in6 *sin6_p; sin6_p = (struct sockaddr_in6 *)nam; if (IN6_IS_ADDR_UNSPECIFIED(&sin6_p->sin6_addr)) inp->inp_vflag |= INP_IPV4; #ifdef INET else if (IN6_IS_ADDR_V4MAPPED(&sin6_p->sin6_addr)) { struct sockaddr_in sin; in6_sin6_2_sin(&sin, sin6_p); inp->inp_vflag |= INP_IPV4; inp->inp_vflag &= ~INP_IPV6; error = in_pcbbind(inp, (struct sockaddr *)&sin, td->td_ucred); goto out; } #endif } error = in6_pcbbind(inp, nam, td->td_ucred); #ifdef INET out: #endif if (error != 0) inp->inp_vflag = vflagsav; INP_HASH_WUNLOCK(pcbinfo); INP_WUNLOCK(inp); return (error); } static void udp6_close(struct socket *so) { struct inpcb *inp; struct inpcbinfo *pcbinfo; pcbinfo = udp_get_inpcbinfo(so->so_proto->pr_protocol); inp = sotoinpcb(so); KASSERT(inp != NULL, ("udp6_close: inp == NULL")); INP_WLOCK(inp); #ifdef INET if (inp->inp_vflag & INP_IPV4) { struct pr_usrreqs *pru; uint8_t nxt; nxt = (inp->inp_socket->so_proto->pr_protocol == IPPROTO_UDP) ? IPPROTO_UDP : IPPROTO_UDPLITE; INP_WUNLOCK(inp); pru = inetsw[ip_protox[nxt]].pr_usrreqs; (*pru->pru_disconnect)(so); return; } #endif if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr)) { INP_HASH_WLOCK(pcbinfo); in6_pcbdisconnect(inp); inp->in6p_laddr = in6addr_any; INP_HASH_WUNLOCK(pcbinfo); soisdisconnected(so); } INP_WUNLOCK(inp); } static int udp6_connect(struct socket *so, struct sockaddr *nam, struct thread *td) { struct inpcb *inp; struct inpcbinfo *pcbinfo; struct sockaddr_in6 *sin6; int error; u_char vflagsav; pcbinfo = udp_get_inpcbinfo(so->so_proto->pr_protocol); inp = sotoinpcb(so); sin6 = (struct sockaddr_in6 *)nam; KASSERT(inp != NULL, ("udp6_connect: inp == NULL")); /* * XXXRW: Need to clarify locking of v4/v6 flags. */ INP_WLOCK(inp); #ifdef INET 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; } if (inp->inp_faddr.s_addr != INADDR_ANY) { error = EISCONN; goto out; } in6_sin6_2_sin(&sin, sin6); error = prison_remote_ip4(td->td_ucred, &sin.sin_addr); if (error != 0) goto out; vflagsav = inp->inp_vflag; inp->inp_vflag |= INP_IPV4; inp->inp_vflag &= ~INP_IPV6; INP_HASH_WLOCK(pcbinfo); error = in_pcbconnect(inp, (struct sockaddr *)&sin, td->td_ucred); INP_HASH_WUNLOCK(pcbinfo); /* * If connect succeeds, mark socket as connected. If * connect fails and socket is unbound, reset inp_vflag * field. */ if (error == 0) soisconnected(so); else if (inp->inp_laddr.s_addr == INADDR_ANY && inp->inp_lport == 0) inp->inp_vflag = vflagsav; goto out; } else { if ((inp->inp_vflag & INP_IPV6) == 0) { error = EAFNOSUPPORT; goto out; } } #endif if (!IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr)) { error = EISCONN; goto out; } error = prison_remote_ip6(td->td_ucred, &sin6->sin6_addr); if (error != 0) goto out; vflagsav = inp->inp_vflag; inp->inp_vflag &= ~INP_IPV4; inp->inp_vflag |= INP_IPV6; INP_HASH_WLOCK(pcbinfo); error = in6_pcbconnect(inp, nam, td->td_ucred); INP_HASH_WUNLOCK(pcbinfo); /* * If connect succeeds, mark socket as connected. If * connect fails and socket is unbound, reset inp_vflag * field. */ if (error == 0) soisconnected(so); else if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_laddr) && inp->inp_lport == 0) inp->inp_vflag = vflagsav; out: INP_WUNLOCK(inp); return (error); } static void udp6_detach(struct socket *so) { struct inpcb *inp; struct inpcbinfo *pcbinfo; struct udpcb *up; pcbinfo = udp_get_inpcbinfo(so->so_proto->pr_protocol); inp = sotoinpcb(so); KASSERT(inp != NULL, ("udp6_detach: inp == NULL")); INP_INFO_WLOCK(pcbinfo); INP_WLOCK(inp); up = intoudpcb(inp); KASSERT(up != NULL, ("%s: up == NULL", __func__)); in_pcbdetach(inp); in_pcbfree(inp); INP_INFO_WUNLOCK(pcbinfo); udp_discardcb(up); } static int udp6_disconnect(struct socket *so) { struct inpcb *inp; struct inpcbinfo *pcbinfo; pcbinfo = udp_get_inpcbinfo(so->so_proto->pr_protocol); inp = sotoinpcb(so); KASSERT(inp != NULL, ("udp6_disconnect: inp == NULL")); INP_WLOCK(inp); #ifdef INET if (inp->inp_vflag & INP_IPV4) { struct pr_usrreqs *pru; uint8_t nxt; nxt = (inp->inp_socket->so_proto->pr_protocol == IPPROTO_UDP) ? IPPROTO_UDP : IPPROTO_UDPLITE; INP_WUNLOCK(inp); pru = inetsw[ip_protox[nxt]].pr_usrreqs; (void)(*pru->pru_disconnect)(so); return (0); } #endif if (IN6_IS_ADDR_UNSPECIFIED(&inp->in6p_faddr)) { INP_WUNLOCK(inp); return (ENOTCONN); } INP_HASH_WLOCK(pcbinfo); in6_pcbdisconnect(inp); inp->in6p_laddr = in6addr_any; INP_HASH_WUNLOCK(pcbinfo); SOCK_LOCK(so); so->so_state &= ~SS_ISCONNECTED; /* XXX */ SOCK_UNLOCK(so); INP_WUNLOCK(inp); return (0); } static int udp6_send(struct socket *so, int flags, struct mbuf *m, struct sockaddr *addr, struct mbuf *control, struct thread *td) { int error; if (addr) { if (addr->sa_len != sizeof(struct sockaddr_in6)) { error = EINVAL; goto bad; } if (addr->sa_family != AF_INET6) { error = EAFNOSUPPORT; goto bad; } } return (udp6_output(so, flags, m, addr, control, td)); bad: if (control) m_freem(control); m_freem(m); return (error); } struct pr_usrreqs udp6_usrreqs = { .pru_abort = udp6_abort, .pru_attach = udp6_attach, .pru_bind = udp6_bind, .pru_connect = udp6_connect, .pru_control = in6_control, .pru_detach = udp6_detach, .pru_disconnect = udp6_disconnect, .pru_peeraddr = in6_mapped_peeraddr, .pru_send = udp6_send, .pru_shutdown = udp_shutdown, .pru_sockaddr = in6_mapped_sockaddr, .pru_soreceive = soreceive_dgram, .pru_sosend = sosend_dgram, .pru_sosetlabel = in_pcbsosetlabel, .pru_close = udp6_close };