Index: head/sys/dev/cxgbe/tom/t4_cpl_io.c =================================================================== --- head/sys/dev/cxgbe/tom/t4_cpl_io.c (revision 299205) +++ head/sys/dev/cxgbe/tom/t4_cpl_io.c (revision 299206) @@ -1,1749 +1,1753 @@ /*- * Copyright (c) 2012, 2015 Chelsio Communications, Inc. * All rights reserved. * Written by: Navdeep Parhar * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #ifdef TCP_OFFLOAD #include #include #include #include #include #include #include #include #include #include #include #include #include #include #define TCPSTATES #include #include #include #include #include "common/common.h" #include "common/t4_msg.h" #include "common/t4_regs.h" #include "common/t4_tcb.h" #include "tom/t4_tom_l2t.h" #include "tom/t4_tom.h" VNET_DECLARE(int, tcp_do_autosndbuf); #define V_tcp_do_autosndbuf VNET(tcp_do_autosndbuf) VNET_DECLARE(int, tcp_autosndbuf_inc); #define V_tcp_autosndbuf_inc VNET(tcp_autosndbuf_inc) VNET_DECLARE(int, tcp_autosndbuf_max); #define V_tcp_autosndbuf_max VNET(tcp_autosndbuf_max) VNET_DECLARE(int, tcp_do_autorcvbuf); #define V_tcp_do_autorcvbuf VNET(tcp_do_autorcvbuf) VNET_DECLARE(int, tcp_autorcvbuf_inc); #define V_tcp_autorcvbuf_inc VNET(tcp_autorcvbuf_inc) VNET_DECLARE(int, tcp_autorcvbuf_max); #define V_tcp_autorcvbuf_max VNET(tcp_autorcvbuf_max) void send_flowc_wr(struct toepcb *toep, struct flowc_tx_params *ftxp) { struct wrqe *wr; struct fw_flowc_wr *flowc; unsigned int nparams = ftxp ? 8 : 6, flowclen; struct vi_info *vi = toep->vi; struct port_info *pi = vi->pi; struct adapter *sc = pi->adapter; unsigned int pfvf = G_FW_VIID_PFN(vi->viid) << S_FW_VIID_PFN; struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx]; KASSERT(!(toep->flags & TPF_FLOWC_WR_SENT), ("%s: flowc for tid %u sent already", __func__, toep->tid)); flowclen = sizeof(*flowc) + nparams * sizeof(struct fw_flowc_mnemval); wr = alloc_wrqe(roundup2(flowclen, 16), toep->ofld_txq); if (wr == NULL) { /* XXX */ panic("%s: allocation failure.", __func__); } flowc = wrtod(wr); memset(flowc, 0, wr->wr_len); flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) | V_FW_FLOWC_WR_NPARAMS(nparams)); flowc->flowid_len16 = htonl(V_FW_WR_LEN16(howmany(flowclen, 16)) | V_FW_WR_FLOWID(toep->tid)); flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_PFNVFN; flowc->mnemval[0].val = htobe32(pfvf); flowc->mnemval[1].mnemonic = FW_FLOWC_MNEM_CH; flowc->mnemval[1].val = htobe32(pi->tx_chan); flowc->mnemval[2].mnemonic = FW_FLOWC_MNEM_PORT; flowc->mnemval[2].val = htobe32(pi->tx_chan); flowc->mnemval[3].mnemonic = FW_FLOWC_MNEM_IQID; flowc->mnemval[3].val = htobe32(toep->ofld_rxq->iq.abs_id); if (ftxp) { uint32_t sndbuf = min(ftxp->snd_space, sc->tt.sndbuf); flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_SNDNXT; flowc->mnemval[4].val = htobe32(ftxp->snd_nxt); flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_RCVNXT; flowc->mnemval[5].val = htobe32(ftxp->rcv_nxt); flowc->mnemval[6].mnemonic = FW_FLOWC_MNEM_SNDBUF; flowc->mnemval[6].val = htobe32(sndbuf); flowc->mnemval[7].mnemonic = FW_FLOWC_MNEM_MSS; flowc->mnemval[7].val = htobe32(ftxp->mss); CTR6(KTR_CXGBE, "%s: tid %u, mss %u, sndbuf %u, snd_nxt 0x%x, rcv_nxt 0x%x", __func__, toep->tid, ftxp->mss, sndbuf, ftxp->snd_nxt, ftxp->rcv_nxt); } else { flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_SNDBUF; flowc->mnemval[4].val = htobe32(512); flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_MSS; flowc->mnemval[5].val = htobe32(512); CTR2(KTR_CXGBE, "%s: tid %u", __func__, toep->tid); } txsd->tx_credits = howmany(flowclen, 16); txsd->plen = 0; KASSERT(toep->tx_credits >= txsd->tx_credits && toep->txsd_avail > 0, ("%s: not enough credits (%d)", __func__, toep->tx_credits)); toep->tx_credits -= txsd->tx_credits; if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) toep->txsd_pidx = 0; toep->txsd_avail--; toep->flags |= TPF_FLOWC_WR_SENT; t4_wrq_tx(sc, wr); } void send_reset(struct adapter *sc, struct toepcb *toep, uint32_t snd_nxt) { struct wrqe *wr; struct cpl_abort_req *req; int tid = toep->tid; struct inpcb *inp = toep->inp; struct tcpcb *tp = intotcpcb(inp); /* don't use if INP_DROPPED */ INP_WLOCK_ASSERT(inp); CTR6(KTR_CXGBE, "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x%s", __func__, toep->tid, inp->inp_flags & INP_DROPPED ? "inp dropped" : tcpstates[tp->t_state], toep->flags, inp->inp_flags, toep->flags & TPF_ABORT_SHUTDOWN ? " (abort already in progress)" : ""); if (toep->flags & TPF_ABORT_SHUTDOWN) return; /* abort already in progress */ toep->flags |= TPF_ABORT_SHUTDOWN; KASSERT(toep->flags & TPF_FLOWC_WR_SENT, ("%s: flowc_wr not sent for tid %d.", __func__, tid)); wr = alloc_wrqe(sizeof(*req), toep->ofld_txq); if (wr == NULL) { /* XXX */ panic("%s: allocation failure.", __func__); } req = wrtod(wr); INIT_TP_WR_MIT_CPL(req, CPL_ABORT_REQ, tid); if (inp->inp_flags & INP_DROPPED) req->rsvd0 = htobe32(snd_nxt); else req->rsvd0 = htobe32(tp->snd_nxt); req->rsvd1 = !(toep->flags & TPF_TX_DATA_SENT); req->cmd = CPL_ABORT_SEND_RST; /* * XXX: What's the correct way to tell that the inp hasn't been detached * from its socket? Should I even be flushing the snd buffer here? */ if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) { struct socket *so = inp->inp_socket; if (so != NULL) /* because I'm not sure. See comment above */ sbflush(&so->so_snd); } t4_l2t_send(sc, wr, toep->l2te); } /* * Called when a connection is established to translate the TCP options * reported by HW to FreeBSD's native format. */ static void assign_rxopt(struct tcpcb *tp, unsigned int opt) { struct toepcb *toep = tp->t_toe; struct inpcb *inp = tp->t_inpcb; struct adapter *sc = td_adapter(toep->td); int n; INP_LOCK_ASSERT(inp); if (inp->inp_inc.inc_flags & INC_ISIPV6) n = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); else n = sizeof(struct ip) + sizeof(struct tcphdr); tp->t_maxseg = sc->params.mtus[G_TCPOPT_MSS(opt)] - n; CTR4(KTR_CXGBE, "%s: tid %d, mtu_idx %u (%u)", __func__, toep->tid, G_TCPOPT_MSS(opt), sc->params.mtus[G_TCPOPT_MSS(opt)]); if (G_TCPOPT_TSTAMP(opt)) { tp->t_flags |= TF_RCVD_TSTMP; /* timestamps ok */ tp->ts_recent = 0; /* hmmm */ tp->ts_recent_age = tcp_ts_getticks(); } if (G_TCPOPT_SACK(opt)) tp->t_flags |= TF_SACK_PERMIT; /* should already be set */ else tp->t_flags &= ~TF_SACK_PERMIT; /* sack disallowed by peer */ if (G_TCPOPT_WSCALE_OK(opt)) tp->t_flags |= TF_RCVD_SCALE; /* Doing window scaling? */ if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == (TF_RCVD_SCALE | TF_REQ_SCALE)) { tp->rcv_scale = tp->request_r_scale; tp->snd_scale = G_TCPOPT_SND_WSCALE(opt); } } /* * Completes some final bits of initialization for just established connections * and changes their state to TCPS_ESTABLISHED. * * The ISNs are from after the exchange of SYNs. i.e., the true ISN + 1. */ void make_established(struct toepcb *toep, uint32_t snd_isn, uint32_t rcv_isn, uint16_t opt) { struct inpcb *inp = toep->inp; struct socket *so = inp->inp_socket; struct tcpcb *tp = intotcpcb(inp); long bufsize; uint32_t iss = be32toh(snd_isn) - 1; /* true ISS */ uint32_t irs = be32toh(rcv_isn) - 1; /* true IRS */ uint16_t tcpopt = be16toh(opt); struct flowc_tx_params ftxp; + CURVNET_SET(so->so_vnet); INP_WLOCK_ASSERT(inp); KASSERT(tp->t_state == TCPS_SYN_SENT || tp->t_state == TCPS_SYN_RECEIVED, ("%s: TCP state %s", __func__, tcpstates[tp->t_state])); CTR4(KTR_CXGBE, "%s: tid %d, toep %p, inp %p", __func__, toep->tid, toep, inp); tp->t_state = TCPS_ESTABLISHED; tp->t_starttime = ticks; TCPSTAT_INC(tcps_connects); tp->irs = irs; tcp_rcvseqinit(tp); tp->rcv_wnd = toep->rx_credits << 10; tp->rcv_adv += tp->rcv_wnd; tp->last_ack_sent = tp->rcv_nxt; /* * If we were unable to send all rx credits via opt0, save the remainder * in rx_credits so that they can be handed over with the next credit * update. */ SOCKBUF_LOCK(&so->so_rcv); bufsize = select_rcv_wnd(so); SOCKBUF_UNLOCK(&so->so_rcv); toep->rx_credits = bufsize - tp->rcv_wnd; tp->iss = iss; tcp_sendseqinit(tp); tp->snd_una = iss + 1; tp->snd_nxt = iss + 1; tp->snd_max = iss + 1; assign_rxopt(tp, tcpopt); SOCKBUF_LOCK(&so->so_snd); if (so->so_snd.sb_flags & SB_AUTOSIZE && V_tcp_do_autosndbuf) bufsize = V_tcp_autosndbuf_max; else bufsize = sbspace(&so->so_snd); SOCKBUF_UNLOCK(&so->so_snd); ftxp.snd_nxt = tp->snd_nxt; ftxp.rcv_nxt = tp->rcv_nxt; ftxp.snd_space = bufsize; ftxp.mss = tp->t_maxseg; send_flowc_wr(toep, &ftxp); soisconnected(so); + CURVNET_RESTORE(); } static int send_rx_credits(struct adapter *sc, struct toepcb *toep, int credits) { struct wrqe *wr; struct cpl_rx_data_ack *req; uint32_t dack = F_RX_DACK_CHANGE | V_RX_DACK_MODE(1); KASSERT(credits >= 0, ("%s: %d credits", __func__, credits)); wr = alloc_wrqe(sizeof(*req), toep->ctrlq); if (wr == NULL) return (0); req = wrtod(wr); INIT_TP_WR_MIT_CPL(req, CPL_RX_DATA_ACK, toep->tid); req->credit_dack = htobe32(dack | V_RX_CREDITS(credits)); t4_wrq_tx(sc, wr); return (credits); } void t4_rcvd(struct toedev *tod, struct tcpcb *tp) { struct adapter *sc = tod->tod_softc; struct inpcb *inp = tp->t_inpcb; struct socket *so = inp->inp_socket; struct sockbuf *sb = &so->so_rcv; struct toepcb *toep = tp->t_toe; int credits; INP_WLOCK_ASSERT(inp); SOCKBUF_LOCK(sb); KASSERT(toep->sb_cc >= sbused(sb), ("%s: sb %p has more data (%d) than last time (%d).", __func__, sb, sbused(sb), toep->sb_cc)); toep->rx_credits += toep->sb_cc - sbused(sb); toep->sb_cc = sbused(sb); if (toep->rx_credits > 0 && (tp->rcv_wnd <= 32 * 1024 || toep->rx_credits >= 64 * 1024 || (toep->rx_credits >= 16 * 1024 && tp->rcv_wnd <= 128 * 1024) || toep->sb_cc + tp->rcv_wnd < sb->sb_lowat)) { credits = send_rx_credits(sc, toep, toep->rx_credits); toep->rx_credits -= credits; tp->rcv_wnd += credits; tp->rcv_adv += credits; } SOCKBUF_UNLOCK(sb); } /* * Close a connection by sending a CPL_CLOSE_CON_REQ message. */ static int close_conn(struct adapter *sc, struct toepcb *toep) { struct wrqe *wr; struct cpl_close_con_req *req; unsigned int tid = toep->tid; CTR3(KTR_CXGBE, "%s: tid %u%s", __func__, toep->tid, toep->flags & TPF_FIN_SENT ? ", IGNORED" : ""); if (toep->flags & TPF_FIN_SENT) return (0); KASSERT(toep->flags & TPF_FLOWC_WR_SENT, ("%s: flowc_wr not sent for tid %u.", __func__, tid)); wr = alloc_wrqe(sizeof(*req), toep->ofld_txq); if (wr == NULL) { /* XXX */ panic("%s: allocation failure.", __func__); } req = wrtod(wr); req->wr.wr_hi = htonl(V_FW_WR_OP(FW_TP_WR) | V_FW_WR_IMMDLEN(sizeof(*req) - sizeof(req->wr))); req->wr.wr_mid = htonl(V_FW_WR_LEN16(howmany(sizeof(*req), 16)) | V_FW_WR_FLOWID(tid)); req->wr.wr_lo = cpu_to_be64(0); OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_CLOSE_CON_REQ, tid)); req->rsvd = 0; toep->flags |= TPF_FIN_SENT; toep->flags &= ~TPF_SEND_FIN; t4_l2t_send(sc, wr, toep->l2te); return (0); } #define MAX_OFLD_TX_CREDITS (SGE_MAX_WR_LEN / 16) #define MIN_OFLD_TX_CREDITS (howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16)) /* Maximum amount of immediate data we could stuff in a WR */ static inline int max_imm_payload(int tx_credits) { const int n = 2; /* Use only up to 2 desc for imm. data WR */ KASSERT(tx_credits >= 0 && tx_credits <= MAX_OFLD_TX_CREDITS, ("%s: %d credits", __func__, tx_credits)); if (tx_credits < MIN_OFLD_TX_CREDITS) return (0); if (tx_credits >= (n * EQ_ESIZE) / 16) return ((n * EQ_ESIZE) - sizeof(struct fw_ofld_tx_data_wr)); else return (tx_credits * 16 - sizeof(struct fw_ofld_tx_data_wr)); } /* Maximum number of SGL entries we could stuff in a WR */ static inline int max_dsgl_nsegs(int tx_credits) { int nseg = 1; /* ulptx_sgl has room for 1, rest ulp_tx_sge_pair */ int sge_pair_credits = tx_credits - MIN_OFLD_TX_CREDITS; KASSERT(tx_credits >= 0 && tx_credits <= MAX_OFLD_TX_CREDITS, ("%s: %d credits", __func__, tx_credits)); if (tx_credits < MIN_OFLD_TX_CREDITS) return (0); nseg += 2 * (sge_pair_credits * 16 / 24); if ((sge_pair_credits * 16) % 24 == 16) nseg++; return (nseg); } static inline void write_tx_wr(void *dst, struct toepcb *toep, unsigned int immdlen, unsigned int plen, uint8_t credits, int shove, int ulp_submode, int txalign) { struct fw_ofld_tx_data_wr *txwr = dst; txwr->op_to_immdlen = htobe32(V_WR_OP(FW_OFLD_TX_DATA_WR) | V_FW_WR_IMMDLEN(immdlen)); txwr->flowid_len16 = htobe32(V_FW_WR_FLOWID(toep->tid) | V_FW_WR_LEN16(credits)); txwr->lsodisable_to_flags = htobe32(V_TX_ULP_MODE(toep->ulp_mode) | V_TX_ULP_SUBMODE(ulp_submode) | V_TX_URG(0) | V_TX_SHOVE(shove)); txwr->plen = htobe32(plen); if (txalign > 0) { struct tcpcb *tp = intotcpcb(toep->inp); if (plen < 2 * tp->t_maxseg || is_10G_port(toep->vi->pi)) txwr->lsodisable_to_flags |= htobe32(F_FW_OFLD_TX_DATA_WR_LSODISABLE); else txwr->lsodisable_to_flags |= htobe32(F_FW_OFLD_TX_DATA_WR_ALIGNPLD | (tp->t_flags & TF_NODELAY ? 0 : F_FW_OFLD_TX_DATA_WR_ALIGNPLDSHOVE)); } } /* * Generate a DSGL from a starting mbuf. The total number of segments and the * maximum segments in any one mbuf are provided. */ static void write_tx_sgl(void *dst, struct mbuf *start, struct mbuf *stop, int nsegs, int n) { struct mbuf *m; struct ulptx_sgl *usgl = dst; int i, j, rc; struct sglist sg; struct sglist_seg segs[n]; KASSERT(nsegs > 0, ("%s: nsegs 0", __func__)); sglist_init(&sg, n, segs); usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) | V_ULPTX_NSGE(nsegs)); i = -1; for (m = start; m != stop; m = m->m_next) { rc = sglist_append(&sg, mtod(m, void *), m->m_len); if (__predict_false(rc != 0)) panic("%s: sglist_append %d", __func__, rc); for (j = 0; j < sg.sg_nseg; i++, j++) { if (i < 0) { usgl->len0 = htobe32(segs[j].ss_len); usgl->addr0 = htobe64(segs[j].ss_paddr); } else { usgl->sge[i / 2].len[i & 1] = htobe32(segs[j].ss_len); usgl->sge[i / 2].addr[i & 1] = htobe64(segs[j].ss_paddr); } #ifdef INVARIANTS nsegs--; #endif } sglist_reset(&sg); } if (i & 1) usgl->sge[i / 2].len[1] = htobe32(0); KASSERT(nsegs == 0, ("%s: nsegs %d, start %p, stop %p", __func__, nsegs, start, stop)); } /* * Max number of SGL entries an offload tx work request can have. This is 41 * (1 + 40) for a full 512B work request. * fw_ofld_tx_data_wr(16B) + ulptx_sgl(16B, 1) + ulptx_sge_pair(480B, 40) */ #define OFLD_SGL_LEN (41) /* * Send data and/or a FIN to the peer. * * The socket's so_snd buffer consists of a stream of data starting with sb_mb * and linked together with m_next. sb_sndptr, if set, is the last mbuf that * was transmitted. * * drop indicates the number of bytes that should be dropped from the head of * the send buffer. It is an optimization that lets do_fw4_ack avoid creating * contention on the send buffer lock (before this change it used to do * sowwakeup and then t4_push_frames right after that when recovering from tx * stalls). When drop is set this function MUST drop the bytes and wake up any * writers. */ void t4_push_frames(struct adapter *sc, struct toepcb *toep, int drop) { struct mbuf *sndptr, *m, *sb_sndptr; struct fw_ofld_tx_data_wr *txwr; struct wrqe *wr; u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf; struct inpcb *inp = toep->inp; struct tcpcb *tp = intotcpcb(inp); struct socket *so = inp->inp_socket; struct sockbuf *sb = &so->so_snd; int tx_credits, shove, compl, space, sowwakeup; struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx]; INP_WLOCK_ASSERT(inp); KASSERT(toep->flags & TPF_FLOWC_WR_SENT, ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid)); KASSERT(toep->ulp_mode == ULP_MODE_NONE || toep->ulp_mode == ULP_MODE_TCPDDP || toep->ulp_mode == ULP_MODE_RDMA, ("%s: ulp_mode %u for toep %p", __func__, toep->ulp_mode, toep)); if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN)) return; /* * This function doesn't resume by itself. Someone else must clear the * flag and call this function. */ if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) { KASSERT(drop == 0, ("%s: drop (%d) != 0 but tx is suspended", __func__, drop)); return; } do { tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS); max_imm = max_imm_payload(tx_credits); max_nsegs = max_dsgl_nsegs(tx_credits); SOCKBUF_LOCK(sb); sowwakeup = drop; if (drop) { sbdrop_locked(sb, drop); drop = 0; } sb_sndptr = sb->sb_sndptr; sndptr = sb_sndptr ? sb_sndptr->m_next : sb->sb_mb; plen = 0; nsegs = 0; max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */ for (m = sndptr; m != NULL; m = m->m_next) { int n = sglist_count(mtod(m, void *), m->m_len); nsegs += n; plen += m->m_len; /* This mbuf sent us _over_ the nsegs limit, back out */ if (plen > max_imm && nsegs > max_nsegs) { nsegs -= n; plen -= m->m_len; if (plen == 0) { /* Too few credits */ toep->flags |= TPF_TX_SUSPENDED; if (sowwakeup) sowwakeup_locked(so); else SOCKBUF_UNLOCK(sb); SOCKBUF_UNLOCK_ASSERT(sb); return; } break; } if (max_nsegs_1mbuf < n) max_nsegs_1mbuf = n; sb_sndptr = m; /* new sb->sb_sndptr if all goes well */ /* This mbuf put us right at the max_nsegs limit */ if (plen > max_imm && nsegs == max_nsegs) { m = m->m_next; break; } } space = sbspace(sb); if (space <= sb->sb_hiwat * 3 / 8 && toep->plen_nocompl + plen >= sb->sb_hiwat / 4) compl = 1; else compl = 0; if (sb->sb_flags & SB_AUTOSIZE && V_tcp_do_autosndbuf && sb->sb_hiwat < V_tcp_autosndbuf_max && space < sb->sb_hiwat / 8) { int newsize = min(sb->sb_hiwat + V_tcp_autosndbuf_inc, V_tcp_autosndbuf_max); if (!sbreserve_locked(sb, newsize, so, NULL)) sb->sb_flags &= ~SB_AUTOSIZE; else sowwakeup = 1; /* room available */ } if (sowwakeup) sowwakeup_locked(so); else SOCKBUF_UNLOCK(sb); SOCKBUF_UNLOCK_ASSERT(sb); /* nothing to send */ if (plen == 0) { KASSERT(m == NULL, ("%s: nothing to send, but m != NULL", __func__)); break; } if (__predict_false(toep->flags & TPF_FIN_SENT)) panic("%s: excess tx.", __func__); shove = m == NULL && !(tp->t_flags & TF_MORETOCOME); if (plen <= max_imm) { /* Immediate data tx */ wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16), toep->ofld_txq); if (wr == NULL) { /* XXX: how will we recover from this? */ toep->flags |= TPF_TX_SUSPENDED; return; } txwr = wrtod(wr); credits = howmany(wr->wr_len, 16); write_tx_wr(txwr, toep, plen, plen, credits, shove, 0, sc->tt.tx_align); m_copydata(sndptr, 0, plen, (void *)(txwr + 1)); nsegs = 0; } else { int wr_len; /* DSGL tx */ wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) + ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8; wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq); if (wr == NULL) { /* XXX: how will we recover from this? */ toep->flags |= TPF_TX_SUSPENDED; return; } txwr = wrtod(wr); credits = howmany(wr_len, 16); write_tx_wr(txwr, toep, 0, plen, credits, shove, 0, sc->tt.tx_align); write_tx_sgl(txwr + 1, sndptr, m, nsegs, max_nsegs_1mbuf); if (wr_len & 0xf) { uint64_t *pad = (uint64_t *) ((uintptr_t)txwr + wr_len); *pad = 0; } } KASSERT(toep->tx_credits >= credits, ("%s: not enough credits", __func__)); toep->tx_credits -= credits; toep->tx_nocompl += credits; toep->plen_nocompl += plen; if (toep->tx_credits <= toep->tx_total * 3 / 8 && toep->tx_nocompl >= toep->tx_total / 4) compl = 1; if (compl || toep->ulp_mode == ULP_MODE_RDMA) { txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL); toep->tx_nocompl = 0; toep->plen_nocompl = 0; } tp->snd_nxt += plen; tp->snd_max += plen; SOCKBUF_LOCK(sb); KASSERT(sb_sndptr, ("%s: sb_sndptr is NULL", __func__)); sb->sb_sndptr = sb_sndptr; SOCKBUF_UNLOCK(sb); toep->flags |= TPF_TX_DATA_SENT; if (toep->tx_credits < MIN_OFLD_TX_CREDITS) toep->flags |= TPF_TX_SUSPENDED; KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__)); txsd->plen = plen; txsd->tx_credits = credits; txsd++; if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) { toep->txsd_pidx = 0; txsd = &toep->txsd[0]; } toep->txsd_avail--; t4_l2t_send(sc, wr, toep->l2te); } while (m != NULL); /* Send a FIN if requested, but only if there's no more data to send */ if (m == NULL && toep->flags & TPF_SEND_FIN) close_conn(sc, toep); } static inline void rqdrop_locked(struct mbufq *q, int plen) { struct mbuf *m; while (plen > 0) { m = mbufq_dequeue(q); /* Too many credits. */ MPASS(m != NULL); M_ASSERTPKTHDR(m); /* Partial credits. */ MPASS(plen >= m->m_pkthdr.len); plen -= m->m_pkthdr.len; m_freem(m); } } void t4_push_pdus(struct adapter *sc, struct toepcb *toep, int drop) { struct mbuf *sndptr, *m; struct fw_ofld_tx_data_wr *txwr; struct wrqe *wr; u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf; u_int adjusted_plen, ulp_submode; struct inpcb *inp = toep->inp; struct tcpcb *tp = intotcpcb(inp); int tx_credits, shove; struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx]; struct mbufq *pduq = &toep->ulp_pduq; static const u_int ulp_extra_len[] = {0, 4, 4, 8}; INP_WLOCK_ASSERT(inp); KASSERT(toep->flags & TPF_FLOWC_WR_SENT, ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid)); KASSERT(toep->ulp_mode == ULP_MODE_ISCSI, ("%s: ulp_mode %u for toep %p", __func__, toep->ulp_mode, toep)); if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN)) return; /* * This function doesn't resume by itself. Someone else must clear the * flag and call this function. */ if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) { KASSERT(drop == 0, ("%s: drop (%d) != 0 but tx is suspended", __func__, drop)); return; } if (drop) rqdrop_locked(&toep->ulp_pdu_reclaimq, drop); while ((sndptr = mbufq_first(pduq)) != NULL) { M_ASSERTPKTHDR(sndptr); tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS); max_imm = max_imm_payload(tx_credits); max_nsegs = max_dsgl_nsegs(tx_credits); plen = 0; nsegs = 0; max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */ for (m = sndptr; m != NULL; m = m->m_next) { int n = sglist_count(mtod(m, void *), m->m_len); nsegs += n; plen += m->m_len; /* * This mbuf would send us _over_ the nsegs limit. * Suspend tx because the PDU can't be sent out. */ if (plen > max_imm && nsegs > max_nsegs) { toep->flags |= TPF_TX_SUSPENDED; return; } if (max_nsegs_1mbuf < n) max_nsegs_1mbuf = n; } if (__predict_false(toep->flags & TPF_FIN_SENT)) panic("%s: excess tx.", __func__); /* * We have a PDU to send. All of it goes out in one WR so 'm' * is NULL. A PDU's length is always a multiple of 4. */ MPASS(m == NULL); MPASS((plen & 3) == 0); MPASS(sndptr->m_pkthdr.len == plen); shove = !(tp->t_flags & TF_MORETOCOME); ulp_submode = mbuf_ulp_submode(sndptr); MPASS(ulp_submode < nitems(ulp_extra_len)); /* * plen doesn't include header and data digests, which are * generated and inserted in the right places by the TOE, but * they do occupy TCP sequence space and need to be accounted * for. */ adjusted_plen = plen + ulp_extra_len[ulp_submode]; if (plen <= max_imm) { /* Immediate data tx */ wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16), toep->ofld_txq); if (wr == NULL) { /* XXX: how will we recover from this? */ toep->flags |= TPF_TX_SUSPENDED; return; } txwr = wrtod(wr); credits = howmany(wr->wr_len, 16); write_tx_wr(txwr, toep, plen, adjusted_plen, credits, shove, ulp_submode, sc->tt.tx_align); m_copydata(sndptr, 0, plen, (void *)(txwr + 1)); nsegs = 0; } else { int wr_len; /* DSGL tx */ wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) + ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8; wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq); if (wr == NULL) { /* XXX: how will we recover from this? */ toep->flags |= TPF_TX_SUSPENDED; return; } txwr = wrtod(wr); credits = howmany(wr_len, 16); write_tx_wr(txwr, toep, 0, adjusted_plen, credits, shove, ulp_submode, sc->tt.tx_align); write_tx_sgl(txwr + 1, sndptr, m, nsegs, max_nsegs_1mbuf); if (wr_len & 0xf) { uint64_t *pad = (uint64_t *) ((uintptr_t)txwr + wr_len); *pad = 0; } } KASSERT(toep->tx_credits >= credits, ("%s: not enough credits", __func__)); m = mbufq_dequeue(pduq); MPASS(m == sndptr); mbufq_enqueue(&toep->ulp_pdu_reclaimq, m); toep->tx_credits -= credits; toep->tx_nocompl += credits; toep->plen_nocompl += plen; if (toep->tx_credits <= toep->tx_total * 3 / 8 && toep->tx_nocompl >= toep->tx_total / 4) { txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL); toep->tx_nocompl = 0; toep->plen_nocompl = 0; } tp->snd_nxt += adjusted_plen; tp->snd_max += adjusted_plen; toep->flags |= TPF_TX_DATA_SENT; if (toep->tx_credits < MIN_OFLD_TX_CREDITS) toep->flags |= TPF_TX_SUSPENDED; KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__)); txsd->plen = plen; txsd->tx_credits = credits; txsd++; if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) { toep->txsd_pidx = 0; txsd = &toep->txsd[0]; } toep->txsd_avail--; t4_l2t_send(sc, wr, toep->l2te); } /* Send a FIN if requested, but only if there are no more PDUs to send */ if (mbufq_first(pduq) == NULL && toep->flags & TPF_SEND_FIN) close_conn(sc, toep); } int t4_tod_output(struct toedev *tod, struct tcpcb *tp) { struct adapter *sc = tod->tod_softc; #ifdef INVARIANTS struct inpcb *inp = tp->t_inpcb; #endif struct toepcb *toep = tp->t_toe; INP_WLOCK_ASSERT(inp); KASSERT((inp->inp_flags & INP_DROPPED) == 0, ("%s: inp %p dropped.", __func__, inp)); KASSERT(toep != NULL, ("%s: toep is NULL", __func__)); if (toep->ulp_mode == ULP_MODE_ISCSI) t4_push_pdus(sc, toep, 0); else t4_push_frames(sc, toep, 0); return (0); } int t4_send_fin(struct toedev *tod, struct tcpcb *tp) { struct adapter *sc = tod->tod_softc; #ifdef INVARIANTS struct inpcb *inp = tp->t_inpcb; #endif struct toepcb *toep = tp->t_toe; INP_WLOCK_ASSERT(inp); KASSERT((inp->inp_flags & INP_DROPPED) == 0, ("%s: inp %p dropped.", __func__, inp)); KASSERT(toep != NULL, ("%s: toep is NULL", __func__)); toep->flags |= TPF_SEND_FIN; if (tp->t_state >= TCPS_ESTABLISHED) { if (toep->ulp_mode == ULP_MODE_ISCSI) t4_push_pdus(sc, toep, 0); else t4_push_frames(sc, toep, 0); } return (0); } int t4_send_rst(struct toedev *tod, struct tcpcb *tp) { struct adapter *sc = tod->tod_softc; #if defined(INVARIANTS) struct inpcb *inp = tp->t_inpcb; #endif struct toepcb *toep = tp->t_toe; INP_WLOCK_ASSERT(inp); KASSERT((inp->inp_flags & INP_DROPPED) == 0, ("%s: inp %p dropped.", __func__, inp)); KASSERT(toep != NULL, ("%s: toep is NULL", __func__)); /* hmmmm */ KASSERT(toep->flags & TPF_FLOWC_WR_SENT, ("%s: flowc for tid %u [%s] not sent already", __func__, toep->tid, tcpstates[tp->t_state])); send_reset(sc, toep, 0); return (0); } /* * Peer has sent us a FIN. */ static int do_peer_close(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { struct adapter *sc = iq->adapter; const struct cpl_peer_close *cpl = (const void *)(rss + 1); unsigned int tid = GET_TID(cpl); struct toepcb *toep = lookup_tid(sc, tid); struct inpcb *inp = toep->inp; struct tcpcb *tp = NULL; struct socket *so; struct sockbuf *sb; #ifdef INVARIANTS unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); #endif KASSERT(opcode == CPL_PEER_CLOSE, ("%s: unexpected opcode 0x%x", __func__, opcode)); KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); if (__predict_false(toep->flags & TPF_SYNQE)) { #ifdef INVARIANTS struct synq_entry *synqe = (void *)toep; INP_WLOCK(synqe->lctx->inp); if (synqe->flags & TPF_SYNQE_HAS_L2TE) { KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN, ("%s: listen socket closed but tid %u not aborted.", __func__, tid)); } else { /* * do_pass_accept_req is still running and will * eventually take care of this tid. */ } INP_WUNLOCK(synqe->lctx->inp); #endif CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid, toep, toep->flags); return (0); } KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__)); INP_INFO_RLOCK(&V_tcbinfo); INP_WLOCK(inp); tp = intotcpcb(inp); CTR5(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x, inp %p", __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags, inp); if (toep->flags & TPF_ABORT_SHUTDOWN) goto done; tp->rcv_nxt++; /* FIN */ so = inp->inp_socket; sb = &so->so_rcv; SOCKBUF_LOCK(sb); if (__predict_false(toep->ddp_flags & (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE))) { handle_ddp_close(toep, tp, sb, cpl->rcv_nxt); } socantrcvmore_locked(so); /* unlocks the sockbuf */ if (toep->ulp_mode != ULP_MODE_RDMA) { KASSERT(tp->rcv_nxt == be32toh(cpl->rcv_nxt), ("%s: rcv_nxt mismatch: %u %u", __func__, tp->rcv_nxt, be32toh(cpl->rcv_nxt))); } switch (tp->t_state) { case TCPS_SYN_RECEIVED: tp->t_starttime = ticks; /* FALLTHROUGH */ case TCPS_ESTABLISHED: tp->t_state = TCPS_CLOSE_WAIT; break; case TCPS_FIN_WAIT_1: tp->t_state = TCPS_CLOSING; break; case TCPS_FIN_WAIT_2: tcp_twstart(tp); INP_UNLOCK_ASSERT(inp); /* safe, we have a ref on the inp */ INP_INFO_RUNLOCK(&V_tcbinfo); INP_WLOCK(inp); final_cpl_received(toep); return (0); default: log(LOG_ERR, "%s: TID %u received CPL_PEER_CLOSE in state %d\n", __func__, tid, tp->t_state); } done: INP_WUNLOCK(inp); INP_INFO_RUNLOCK(&V_tcbinfo); return (0); } /* * Peer has ACK'd our FIN. */ static int do_close_con_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { struct adapter *sc = iq->adapter; const struct cpl_close_con_rpl *cpl = (const void *)(rss + 1); unsigned int tid = GET_TID(cpl); struct toepcb *toep = lookup_tid(sc, tid); struct inpcb *inp = toep->inp; struct tcpcb *tp = NULL; struct socket *so = NULL; #ifdef INVARIANTS unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); #endif KASSERT(opcode == CPL_CLOSE_CON_RPL, ("%s: unexpected opcode 0x%x", __func__, opcode)); KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__)); INP_INFO_RLOCK(&V_tcbinfo); INP_WLOCK(inp); tp = intotcpcb(inp); CTR4(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x", __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags); if (toep->flags & TPF_ABORT_SHUTDOWN) goto done; so = inp->inp_socket; tp->snd_una = be32toh(cpl->snd_nxt) - 1; /* exclude FIN */ switch (tp->t_state) { case TCPS_CLOSING: /* see TCPS_FIN_WAIT_2 in do_peer_close too */ tcp_twstart(tp); release: INP_UNLOCK_ASSERT(inp); /* safe, we have a ref on the inp */ INP_INFO_RUNLOCK(&V_tcbinfo); INP_WLOCK(inp); final_cpl_received(toep); /* no more CPLs expected */ return (0); case TCPS_LAST_ACK: if (tcp_close(tp)) INP_WUNLOCK(inp); goto release; case TCPS_FIN_WAIT_1: if (so->so_rcv.sb_state & SBS_CANTRCVMORE) soisdisconnected(so); tp->t_state = TCPS_FIN_WAIT_2; break; default: log(LOG_ERR, "%s: TID %u received CPL_CLOSE_CON_RPL in state %s\n", __func__, tid, tcpstates[tp->t_state]); } done: INP_WUNLOCK(inp); INP_INFO_RUNLOCK(&V_tcbinfo); return (0); } void send_abort_rpl(struct adapter *sc, struct sge_wrq *ofld_txq, int tid, int rst_status) { struct wrqe *wr; struct cpl_abort_rpl *cpl; wr = alloc_wrqe(sizeof(*cpl), ofld_txq); if (wr == NULL) { /* XXX */ panic("%s: allocation failure.", __func__); } cpl = wrtod(wr); INIT_TP_WR_MIT_CPL(cpl, CPL_ABORT_RPL, tid); cpl->cmd = rst_status; t4_wrq_tx(sc, wr); } static int abort_status_to_errno(struct tcpcb *tp, unsigned int abort_reason) { switch (abort_reason) { case CPL_ERR_BAD_SYN: case CPL_ERR_CONN_RESET: return (tp->t_state == TCPS_CLOSE_WAIT ? EPIPE : ECONNRESET); case CPL_ERR_XMIT_TIMEDOUT: case CPL_ERR_PERSIST_TIMEDOUT: case CPL_ERR_FINWAIT2_TIMEDOUT: case CPL_ERR_KEEPALIVE_TIMEDOUT: return (ETIMEDOUT); default: return (EIO); } } /* * TCP RST from the peer, timeout, or some other such critical error. */ static int do_abort_req(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { struct adapter *sc = iq->adapter; const struct cpl_abort_req_rss *cpl = (const void *)(rss + 1); unsigned int tid = GET_TID(cpl); struct toepcb *toep = lookup_tid(sc, tid); struct sge_wrq *ofld_txq = toep->ofld_txq; struct inpcb *inp; struct tcpcb *tp; #ifdef INVARIANTS unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); #endif KASSERT(opcode == CPL_ABORT_REQ_RSS, ("%s: unexpected opcode 0x%x", __func__, opcode)); KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); if (toep->flags & TPF_SYNQE) return (do_abort_req_synqe(iq, rss, m)); KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__)); if (negative_advice(cpl->status)) { CTR4(KTR_CXGBE, "%s: negative advice %d for tid %d (0x%x)", __func__, cpl->status, tid, toep->flags); return (0); /* Ignore negative advice */ } inp = toep->inp; INP_INFO_RLOCK(&V_tcbinfo); /* for tcp_close */ INP_WLOCK(inp); tp = intotcpcb(inp); CTR6(KTR_CXGBE, "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x, status %d", __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags, inp->inp_flags, cpl->status); /* * If we'd initiated an abort earlier the reply to it is responsible for * cleaning up resources. Otherwise we tear everything down right here * right now. We owe the T4 a CPL_ABORT_RPL no matter what. */ if (toep->flags & TPF_ABORT_SHUTDOWN) { INP_WUNLOCK(inp); goto done; } toep->flags |= TPF_ABORT_SHUTDOWN; if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) { struct socket *so = inp->inp_socket; if (so != NULL) so_error_set(so, abort_status_to_errno(tp, cpl->status)); tp = tcp_close(tp); if (tp == NULL) INP_WLOCK(inp); /* re-acquire */ } final_cpl_received(toep); done: INP_INFO_RUNLOCK(&V_tcbinfo); send_abort_rpl(sc, ofld_txq, tid, CPL_ABORT_NO_RST); return (0); } /* * Reply to the CPL_ABORT_REQ (send_reset) */ static int do_abort_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { struct adapter *sc = iq->adapter; const struct cpl_abort_rpl_rss *cpl = (const void *)(rss + 1); unsigned int tid = GET_TID(cpl); struct toepcb *toep = lookup_tid(sc, tid); struct inpcb *inp = toep->inp; #ifdef INVARIANTS unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); #endif KASSERT(opcode == CPL_ABORT_RPL_RSS, ("%s: unexpected opcode 0x%x", __func__, opcode)); KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); if (toep->flags & TPF_SYNQE) return (do_abort_rpl_synqe(iq, rss, m)); KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__)); CTR5(KTR_CXGBE, "%s: tid %u, toep %p, inp %p, status %d", __func__, tid, toep, inp, cpl->status); KASSERT(toep->flags & TPF_ABORT_SHUTDOWN, ("%s: wasn't expecting abort reply", __func__)); INP_WLOCK(inp); final_cpl_received(toep); return (0); } static int do_rx_data(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { struct adapter *sc = iq->adapter; const struct cpl_rx_data *cpl = mtod(m, const void *); unsigned int tid = GET_TID(cpl); struct toepcb *toep = lookup_tid(sc, tid); struct inpcb *inp = toep->inp; struct tcpcb *tp; struct socket *so; struct sockbuf *sb; int len; uint32_t ddp_placed = 0; if (__predict_false(toep->flags & TPF_SYNQE)) { #ifdef INVARIANTS struct synq_entry *synqe = (void *)toep; INP_WLOCK(synqe->lctx->inp); if (synqe->flags & TPF_SYNQE_HAS_L2TE) { KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN, ("%s: listen socket closed but tid %u not aborted.", __func__, tid)); } else { /* * do_pass_accept_req is still running and will * eventually take care of this tid. */ } INP_WUNLOCK(synqe->lctx->inp); #endif CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid, toep, toep->flags); m_freem(m); return (0); } KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__)); /* strip off CPL header */ m_adj(m, sizeof(*cpl)); len = m->m_pkthdr.len; INP_WLOCK(inp); if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) { CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x", __func__, tid, len, inp->inp_flags); INP_WUNLOCK(inp); m_freem(m); return (0); } tp = intotcpcb(inp); if (__predict_false(tp->rcv_nxt != be32toh(cpl->seq))) ddp_placed = be32toh(cpl->seq) - tp->rcv_nxt; tp->rcv_nxt += len; KASSERT(tp->rcv_wnd >= len, ("%s: negative window size", __func__)); tp->rcv_wnd -= len; tp->t_rcvtime = ticks; so = inp_inpcbtosocket(inp); sb = &so->so_rcv; SOCKBUF_LOCK(sb); if (__predict_false(sb->sb_state & SBS_CANTRCVMORE)) { CTR3(KTR_CXGBE, "%s: tid %u, excess rx (%d bytes)", __func__, tid, len); m_freem(m); SOCKBUF_UNLOCK(sb); INP_WUNLOCK(inp); INP_INFO_RLOCK(&V_tcbinfo); INP_WLOCK(inp); tp = tcp_drop(tp, ECONNRESET); if (tp) INP_WUNLOCK(inp); INP_INFO_RUNLOCK(&V_tcbinfo); return (0); } /* receive buffer autosize */ + CURVNET_SET(so->so_vnet); if (sb->sb_flags & SB_AUTOSIZE && V_tcp_do_autorcvbuf && sb->sb_hiwat < V_tcp_autorcvbuf_max && len > (sbspace(sb) / 8 * 7)) { unsigned int hiwat = sb->sb_hiwat; unsigned int newsize = min(hiwat + V_tcp_autorcvbuf_inc, V_tcp_autorcvbuf_max); if (!sbreserve_locked(sb, newsize, so, NULL)) sb->sb_flags &= ~SB_AUTOSIZE; else toep->rx_credits += newsize - hiwat; } if (toep->ulp_mode == ULP_MODE_TCPDDP) { int changed = !(toep->ddp_flags & DDP_ON) ^ cpl->ddp_off; if (changed) { if (toep->ddp_flags & DDP_SC_REQ) toep->ddp_flags ^= DDP_ON | DDP_SC_REQ; else { KASSERT(cpl->ddp_off == 1, ("%s: DDP switched on by itself.", __func__)); /* Fell out of DDP mode */ toep->ddp_flags &= ~(DDP_ON | DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE); if (ddp_placed) insert_ddp_data(toep, ddp_placed); } } if ((toep->ddp_flags & DDP_OK) == 0 && time_uptime >= toep->ddp_disabled + DDP_RETRY_WAIT) { toep->ddp_score = DDP_LOW_SCORE; toep->ddp_flags |= DDP_OK; CTR3(KTR_CXGBE, "%s: tid %u DDP_OK @ %u", __func__, tid, time_uptime); } if (toep->ddp_flags & DDP_ON) { /* * CPL_RX_DATA with DDP on can only be an indicate. Ask * soreceive to post a buffer or disable DDP. The * payload that arrived in this indicate is appended to * the socket buffer as usual. */ #if 0 CTR5(KTR_CXGBE, "%s: tid %u (0x%x) DDP indicate (seq 0x%x, len %d)", __func__, tid, toep->flags, be32toh(cpl->seq), len); #endif sb->sb_flags |= SB_DDP_INDICATE; } else if ((toep->ddp_flags & (DDP_OK|DDP_SC_REQ)) == DDP_OK && tp->rcv_wnd > DDP_RSVD_WIN && len >= sc->tt.ddp_thres) { /* * DDP allowed but isn't on (and a request to switch it * on isn't pending either), and conditions are ripe for * it to work. Switch it on. */ enable_ddp(sc, toep); } } KASSERT(toep->sb_cc >= sbused(sb), ("%s: sb %p has more data (%d) than last time (%d).", __func__, sb, sbused(sb), toep->sb_cc)); toep->rx_credits += toep->sb_cc - sbused(sb); sbappendstream_locked(sb, m, 0); toep->sb_cc = sbused(sb); if (toep->rx_credits > 0 && toep->sb_cc + tp->rcv_wnd < sb->sb_lowat) { int credits; credits = send_rx_credits(sc, toep, toep->rx_credits); toep->rx_credits -= credits; tp->rcv_wnd += credits; tp->rcv_adv += credits; } sorwakeup_locked(so); SOCKBUF_UNLOCK_ASSERT(sb); INP_WUNLOCK(inp); + CURVNET_RESTORE(); return (0); } #define S_CPL_FW4_ACK_OPCODE 24 #define M_CPL_FW4_ACK_OPCODE 0xff #define V_CPL_FW4_ACK_OPCODE(x) ((x) << S_CPL_FW4_ACK_OPCODE) #define G_CPL_FW4_ACK_OPCODE(x) \ (((x) >> S_CPL_FW4_ACK_OPCODE) & M_CPL_FW4_ACK_OPCODE) #define S_CPL_FW4_ACK_FLOWID 0 #define M_CPL_FW4_ACK_FLOWID 0xffffff #define V_CPL_FW4_ACK_FLOWID(x) ((x) << S_CPL_FW4_ACK_FLOWID) #define G_CPL_FW4_ACK_FLOWID(x) \ (((x) >> S_CPL_FW4_ACK_FLOWID) & M_CPL_FW4_ACK_FLOWID) #define S_CPL_FW4_ACK_CR 24 #define M_CPL_FW4_ACK_CR 0xff #define V_CPL_FW4_ACK_CR(x) ((x) << S_CPL_FW4_ACK_CR) #define G_CPL_FW4_ACK_CR(x) (((x) >> S_CPL_FW4_ACK_CR) & M_CPL_FW4_ACK_CR) #define S_CPL_FW4_ACK_SEQVAL 0 #define M_CPL_FW4_ACK_SEQVAL 0x1 #define V_CPL_FW4_ACK_SEQVAL(x) ((x) << S_CPL_FW4_ACK_SEQVAL) #define G_CPL_FW4_ACK_SEQVAL(x) \ (((x) >> S_CPL_FW4_ACK_SEQVAL) & M_CPL_FW4_ACK_SEQVAL) #define F_CPL_FW4_ACK_SEQVAL V_CPL_FW4_ACK_SEQVAL(1U) static int do_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { struct adapter *sc = iq->adapter; const struct cpl_fw4_ack *cpl = (const void *)(rss + 1); unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl))); struct toepcb *toep = lookup_tid(sc, tid); struct inpcb *inp; struct tcpcb *tp; struct socket *so; uint8_t credits = cpl->credits; struct ofld_tx_sdesc *txsd; int plen; #ifdef INVARIANTS unsigned int opcode = G_CPL_FW4_ACK_OPCODE(be32toh(OPCODE_TID(cpl))); #endif /* * Very unusual case: we'd sent a flowc + abort_req for a synq entry and * now this comes back carrying the credits for the flowc. */ if (__predict_false(toep->flags & TPF_SYNQE)) { KASSERT(toep->flags & TPF_ABORT_SHUTDOWN, ("%s: credits for a synq entry %p", __func__, toep)); return (0); } inp = toep->inp; KASSERT(opcode == CPL_FW4_ACK, ("%s: unexpected opcode 0x%x", __func__, opcode)); KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__)); INP_WLOCK(inp); if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN)) { INP_WUNLOCK(inp); return (0); } KASSERT((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) == 0, ("%s: inp_flags 0x%x", __func__, inp->inp_flags)); tp = intotcpcb(inp); if (cpl->flags & CPL_FW4_ACK_FLAGS_SEQVAL) { tcp_seq snd_una = be32toh(cpl->snd_una); #ifdef INVARIANTS if (__predict_false(SEQ_LT(snd_una, tp->snd_una))) { log(LOG_ERR, "%s: unexpected seq# %x for TID %u, snd_una %x\n", __func__, snd_una, toep->tid, tp->snd_una); } #endif if (tp->snd_una != snd_una) { tp->snd_una = snd_una; tp->ts_recent_age = tcp_ts_getticks(); } } so = inp->inp_socket; txsd = &toep->txsd[toep->txsd_cidx]; plen = 0; while (credits) { KASSERT(credits >= txsd->tx_credits, ("%s: too many (or partial) credits", __func__)); credits -= txsd->tx_credits; toep->tx_credits += txsd->tx_credits; plen += txsd->plen; txsd++; toep->txsd_avail++; KASSERT(toep->txsd_avail <= toep->txsd_total, ("%s: txsd avail > total", __func__)); if (__predict_false(++toep->txsd_cidx == toep->txsd_total)) { txsd = &toep->txsd[0]; toep->txsd_cidx = 0; } } if (toep->tx_credits == toep->tx_total) { toep->tx_nocompl = 0; toep->plen_nocompl = 0; } if (toep->flags & TPF_TX_SUSPENDED && toep->tx_credits >= toep->tx_total / 4) { toep->flags &= ~TPF_TX_SUSPENDED; if (toep->ulp_mode == ULP_MODE_ISCSI) t4_push_pdus(sc, toep, plen); else t4_push_frames(sc, toep, plen); } else if (plen > 0) { struct sockbuf *sb = &so->so_snd; int sbu; SOCKBUF_LOCK(sb); sbu = sbused(sb); if (toep->ulp_mode == ULP_MODE_ISCSI) { if (__predict_false(sbu > 0)) { /* * The data trasmitted before the tid's ULP mode * changed to ISCSI is still in so_snd. * Incoming credits should account for so_snd * first. */ sbdrop_locked(sb, min(sbu, plen)); plen -= min(sbu, plen); } sowwakeup_locked(so); /* unlocks so_snd */ rqdrop_locked(&toep->ulp_pdu_reclaimq, plen); } else { sbdrop_locked(sb, plen); sowwakeup_locked(so); /* unlocks so_snd */ } SOCKBUF_UNLOCK_ASSERT(sb); } INP_WUNLOCK(inp); return (0); } static int do_set_tcb_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { struct adapter *sc = iq->adapter; const struct cpl_set_tcb_rpl *cpl = (const void *)(rss + 1); unsigned int tid = GET_TID(cpl); #ifdef INVARIANTS unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); #endif KASSERT(opcode == CPL_SET_TCB_RPL, ("%s: unexpected opcode 0x%x", __func__, opcode)); KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); if (is_ftid(sc, tid)) return (t4_filter_rpl(iq, rss, m)); /* TCB is a filter */ /* * TOM and/or other ULPs don't request replies for CPL_SET_TCB or * CPL_SET_TCB_FIELD requests. This can easily change and when it does * the dispatch code will go here. */ #ifdef INVARIANTS panic("%s: Unexpected CPL_SET_TCB_RPL for tid %u on iq %p", __func__, tid, iq); #else log(LOG_ERR, "%s: Unexpected CPL_SET_TCB_RPL for tid %u on iq %p\n", __func__, tid, iq); #endif return (0); } void t4_set_tcb_field(struct adapter *sc, struct toepcb *toep, int ctrl, uint16_t word, uint64_t mask, uint64_t val) { struct wrqe *wr; struct cpl_set_tcb_field *req; wr = alloc_wrqe(sizeof(*req), ctrl ? toep->ctrlq : toep->ofld_txq); if (wr == NULL) { /* XXX */ panic("%s: allocation failure.", __func__); } req = wrtod(wr); INIT_TP_WR_MIT_CPL(req, CPL_SET_TCB_FIELD, toep->tid); req->reply_ctrl = htobe16(V_NO_REPLY(1) | V_QUEUENO(toep->ofld_rxq->iq.abs_id)); req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(0)); req->mask = htobe64(mask); req->val = htobe64(val); t4_wrq_tx(sc, wr); } void t4_init_cpl_io_handlers(struct adapter *sc) { t4_register_cpl_handler(sc, CPL_PEER_CLOSE, do_peer_close); t4_register_cpl_handler(sc, CPL_CLOSE_CON_RPL, do_close_con_rpl); t4_register_cpl_handler(sc, CPL_ABORT_REQ_RSS, do_abort_req); t4_register_cpl_handler(sc, CPL_ABORT_RPL_RSS, do_abort_rpl); t4_register_cpl_handler(sc, CPL_RX_DATA, do_rx_data); t4_register_cpl_handler(sc, CPL_FW4_ACK, do_fw4_ack); t4_register_cpl_handler(sc, CPL_SET_TCB_RPL, do_set_tcb_rpl); } void t4_uninit_cpl_io_handlers(struct adapter *sc) { t4_register_cpl_handler(sc, CPL_SET_TCB_RPL, t4_filter_rpl); } #endif Index: head/sys/dev/cxgbe/tom/t4_listen.c =================================================================== --- head/sys/dev/cxgbe/tom/t4_listen.c (revision 299205) +++ head/sys/dev/cxgbe/tom/t4_listen.c (revision 299206) @@ -1,1598 +1,1600 @@ /*- * 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" #ifdef TCP_OFFLOAD #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #define TCPSTATES #include #include #include #include "common/common.h" #include "common/t4_msg.h" #include "common/t4_regs.h" #include "tom/t4_tom_l2t.h" #include "tom/t4_tom.h" /* stid services */ static int alloc_stid(struct adapter *, struct listen_ctx *, int); static struct listen_ctx *lookup_stid(struct adapter *, int); static void free_stid(struct adapter *, struct listen_ctx *); /* lctx services */ static struct listen_ctx *alloc_lctx(struct adapter *, struct inpcb *, struct vi_info *); static int free_lctx(struct adapter *, struct listen_ctx *); static void hold_lctx(struct listen_ctx *); static void listen_hash_add(struct adapter *, struct listen_ctx *); static struct listen_ctx *listen_hash_find(struct adapter *, struct inpcb *); static struct listen_ctx *listen_hash_del(struct adapter *, struct inpcb *); static struct inpcb *release_lctx(struct adapter *, struct listen_ctx *); static inline void save_qids_in_mbuf(struct mbuf *, struct vi_info *); static inline void get_qids_from_mbuf(struct mbuf *m, int *, int *); static void send_reset_synqe(struct toedev *, struct synq_entry *); static int alloc_stid(struct adapter *sc, struct listen_ctx *lctx, int isipv6) { struct tid_info *t = &sc->tids; u_int stid, n, f, mask; struct stid_region *sr = &lctx->stid_region; /* * An IPv6 server needs 2 naturally aligned stids (1 stid = 4 cells) in * the TCAM. The start of the stid region is properly aligned (the chip * requires each region to be 128-cell aligned). */ n = isipv6 ? 2 : 1; mask = n - 1; KASSERT((t->stid_base & mask) == 0 && (t->nstids & mask) == 0, ("%s: stid region (%u, %u) not properly aligned. n = %u", __func__, t->stid_base, t->nstids, n)); mtx_lock(&t->stid_lock); if (n > t->nstids - t->stids_in_use) { mtx_unlock(&t->stid_lock); return (-1); } if (t->nstids_free_head >= n) { /* * This allocation will definitely succeed because the region * starts at a good alignment and we just checked we have enough * stids free. */ f = t->nstids_free_head & mask; t->nstids_free_head -= n + f; stid = t->nstids_free_head; TAILQ_INSERT_HEAD(&t->stids, sr, link); } else { struct stid_region *s; stid = t->nstids_free_head; TAILQ_FOREACH(s, &t->stids, link) { stid += s->used + s->free; f = stid & mask; if (s->free >= n + f) { stid -= n + f; s->free -= n + f; TAILQ_INSERT_AFTER(&t->stids, s, sr, link); goto allocated; } } if (__predict_false(stid != t->nstids)) { panic("%s: stids TAILQ (%p) corrupt." " At %d instead of %d at the end of the queue.", __func__, &t->stids, stid, t->nstids); } mtx_unlock(&t->stid_lock); return (-1); } allocated: sr->used = n; sr->free = f; t->stids_in_use += n; t->stid_tab[stid] = lctx; mtx_unlock(&t->stid_lock); KASSERT(((stid + t->stid_base) & mask) == 0, ("%s: EDOOFUS.", __func__)); return (stid + t->stid_base); } static struct listen_ctx * lookup_stid(struct adapter *sc, int stid) { struct tid_info *t = &sc->tids; return (t->stid_tab[stid - t->stid_base]); } static void free_stid(struct adapter *sc, struct listen_ctx *lctx) { struct tid_info *t = &sc->tids; struct stid_region *sr = &lctx->stid_region; struct stid_region *s; KASSERT(sr->used > 0, ("%s: nonsense free (%d)", __func__, sr->used)); mtx_lock(&t->stid_lock); s = TAILQ_PREV(sr, stid_head, link); if (s != NULL) s->free += sr->used + sr->free; else t->nstids_free_head += sr->used + sr->free; KASSERT(t->stids_in_use >= sr->used, ("%s: stids_in_use (%u) < stids being freed (%u)", __func__, t->stids_in_use, sr->used)); t->stids_in_use -= sr->used; TAILQ_REMOVE(&t->stids, sr, link); mtx_unlock(&t->stid_lock); } static struct listen_ctx * alloc_lctx(struct adapter *sc, struct inpcb *inp, struct vi_info *vi) { struct listen_ctx *lctx; INP_WLOCK_ASSERT(inp); lctx = malloc(sizeof(struct listen_ctx), M_CXGBE, M_NOWAIT | M_ZERO); if (lctx == NULL) return (NULL); lctx->stid = alloc_stid(sc, lctx, inp->inp_vflag & INP_IPV6); if (lctx->stid < 0) { free(lctx, M_CXGBE); return (NULL); } if (inp->inp_vflag & INP_IPV6 && !IN6_ARE_ADDR_EQUAL(&in6addr_any, &inp->in6p_laddr)) { struct tom_data *td = sc->tom_softc; lctx->ce = hold_lip(td, &inp->in6p_laddr); if (lctx->ce == NULL) { free(lctx, M_CXGBE); return (NULL); } } lctx->ctrlq = &sc->sge.ctrlq[vi->pi->port_id]; lctx->ofld_rxq = &sc->sge.ofld_rxq[vi->first_ofld_rxq]; refcount_init(&lctx->refcount, 1); TAILQ_INIT(&lctx->synq); lctx->inp = inp; in_pcbref(inp); return (lctx); } /* Don't call this directly, use release_lctx instead */ static int free_lctx(struct adapter *sc, struct listen_ctx *lctx) { struct inpcb *inp = lctx->inp; struct tom_data *td = sc->tom_softc; INP_WLOCK_ASSERT(inp); KASSERT(lctx->refcount == 0, ("%s: refcount %d", __func__, lctx->refcount)); KASSERT(TAILQ_EMPTY(&lctx->synq), ("%s: synq not empty.", __func__)); KASSERT(lctx->stid >= 0, ("%s: bad stid %d.", __func__, lctx->stid)); CTR4(KTR_CXGBE, "%s: stid %u, lctx %p, inp %p", __func__, lctx->stid, lctx, lctx->inp); if (lctx->ce) release_lip(td, lctx->ce); free_stid(sc, lctx); free(lctx, M_CXGBE); return (in_pcbrele_wlocked(inp)); } static void hold_lctx(struct listen_ctx *lctx) { refcount_acquire(&lctx->refcount); } static inline uint32_t listen_hashfn(void *key, u_long mask) { return (fnv_32_buf(&key, sizeof(key), FNV1_32_INIT) & mask); } /* * Add a listen_ctx entry to the listen hash table. */ static void listen_hash_add(struct adapter *sc, struct listen_ctx *lctx) { struct tom_data *td = sc->tom_softc; int bucket = listen_hashfn(lctx->inp, td->listen_mask); mtx_lock(&td->lctx_hash_lock); LIST_INSERT_HEAD(&td->listen_hash[bucket], lctx, link); td->lctx_count++; mtx_unlock(&td->lctx_hash_lock); } /* * Look for the listening socket's context entry in the hash and return it. */ static struct listen_ctx * listen_hash_find(struct adapter *sc, struct inpcb *inp) { struct tom_data *td = sc->tom_softc; int bucket = listen_hashfn(inp, td->listen_mask); struct listen_ctx *lctx; mtx_lock(&td->lctx_hash_lock); LIST_FOREACH(lctx, &td->listen_hash[bucket], link) { if (lctx->inp == inp) break; } mtx_unlock(&td->lctx_hash_lock); return (lctx); } /* * Removes the listen_ctx structure for inp from the hash and returns it. */ static struct listen_ctx * listen_hash_del(struct adapter *sc, struct inpcb *inp) { struct tom_data *td = sc->tom_softc; int bucket = listen_hashfn(inp, td->listen_mask); struct listen_ctx *lctx, *l; mtx_lock(&td->lctx_hash_lock); LIST_FOREACH_SAFE(lctx, &td->listen_hash[bucket], link, l) { if (lctx->inp == inp) { LIST_REMOVE(lctx, link); td->lctx_count--; break; } } mtx_unlock(&td->lctx_hash_lock); return (lctx); } /* * Releases a hold on the lctx. Must be called with the listening socket's inp * locked. The inp may be freed by this function and it returns NULL to * indicate this. */ static struct inpcb * release_lctx(struct adapter *sc, struct listen_ctx *lctx) { struct inpcb *inp = lctx->inp; int inp_freed = 0; INP_WLOCK_ASSERT(inp); if (refcount_release(&lctx->refcount)) inp_freed = free_lctx(sc, lctx); return (inp_freed ? NULL : inp); } static void send_reset_synqe(struct toedev *tod, struct synq_entry *synqe) { struct adapter *sc = tod->tod_softc; struct mbuf *m = synqe->syn; struct ifnet *ifp = m->m_pkthdr.rcvif; struct vi_info *vi = ifp->if_softc; struct port_info *pi = vi->pi; struct l2t_entry *e = &sc->l2t->l2tab[synqe->l2e_idx]; struct wrqe *wr; struct fw_flowc_wr *flowc; struct cpl_abort_req *req; int txqid, rxqid, flowclen; struct sge_wrq *ofld_txq; struct sge_ofld_rxq *ofld_rxq; const int nparams = 6; unsigned int pfvf = G_FW_VIID_PFN(vi->viid) << S_FW_VIID_PFN; INP_WLOCK_ASSERT(synqe->lctx->inp); CTR5(KTR_CXGBE, "%s: synqe %p (0x%x), tid %d%s", __func__, synqe, synqe->flags, synqe->tid, synqe->flags & TPF_ABORT_SHUTDOWN ? " (abort already in progress)" : ""); if (synqe->flags & TPF_ABORT_SHUTDOWN) return; /* abort already in progress */ synqe->flags |= TPF_ABORT_SHUTDOWN; get_qids_from_mbuf(m, &txqid, &rxqid); ofld_txq = &sc->sge.ofld_txq[txqid]; ofld_rxq = &sc->sge.ofld_rxq[rxqid]; /* The wrqe will have two WRs - a flowc followed by an abort_req */ flowclen = sizeof(*flowc) + nparams * sizeof(struct fw_flowc_mnemval); wr = alloc_wrqe(roundup2(flowclen, EQ_ESIZE) + sizeof(*req), ofld_txq); if (wr == NULL) { /* XXX */ panic("%s: allocation failure.", __func__); } flowc = wrtod(wr); req = (void *)((caddr_t)flowc + roundup2(flowclen, EQ_ESIZE)); /* First the flowc ... */ memset(flowc, 0, wr->wr_len); flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) | V_FW_FLOWC_WR_NPARAMS(nparams)); flowc->flowid_len16 = htonl(V_FW_WR_LEN16(howmany(flowclen, 16)) | V_FW_WR_FLOWID(synqe->tid)); flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_PFNVFN; flowc->mnemval[0].val = htobe32(pfvf); flowc->mnemval[1].mnemonic = FW_FLOWC_MNEM_CH; flowc->mnemval[1].val = htobe32(pi->tx_chan); flowc->mnemval[2].mnemonic = FW_FLOWC_MNEM_PORT; flowc->mnemval[2].val = htobe32(pi->tx_chan); flowc->mnemval[3].mnemonic = FW_FLOWC_MNEM_IQID; flowc->mnemval[3].val = htobe32(ofld_rxq->iq.abs_id); flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_SNDBUF; flowc->mnemval[4].val = htobe32(512); flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_MSS; flowc->mnemval[5].val = htobe32(512); synqe->flags |= TPF_FLOWC_WR_SENT; /* ... then ABORT request */ INIT_TP_WR_MIT_CPL(req, CPL_ABORT_REQ, synqe->tid); req->rsvd0 = 0; /* don't have a snd_nxt */ req->rsvd1 = 1; /* no data sent yet */ req->cmd = CPL_ABORT_SEND_RST; t4_l2t_send(sc, wr, e); } static int create_server(struct adapter *sc, struct listen_ctx *lctx) { struct wrqe *wr; struct cpl_pass_open_req *req; struct inpcb *inp = lctx->inp; wr = alloc_wrqe(sizeof(*req), lctx->ctrlq); if (wr == NULL) { log(LOG_ERR, "%s: allocation failure", __func__); return (ENOMEM); } req = wrtod(wr); INIT_TP_WR(req, 0); OPCODE_TID(req) = htobe32(MK_OPCODE_TID(CPL_PASS_OPEN_REQ, lctx->stid)); req->local_port = inp->inp_lport; req->peer_port = 0; req->local_ip = inp->inp_laddr.s_addr; req->peer_ip = 0; req->opt0 = htobe64(V_TX_CHAN(lctx->ctrlq->eq.tx_chan)); req->opt1 = htobe64(V_CONN_POLICY(CPL_CONN_POLICY_ASK) | F_SYN_RSS_ENABLE | V_SYN_RSS_QUEUE(lctx->ofld_rxq->iq.abs_id)); t4_wrq_tx(sc, wr); return (0); } static int create_server6(struct adapter *sc, struct listen_ctx *lctx) { struct wrqe *wr; struct cpl_pass_open_req6 *req; struct inpcb *inp = lctx->inp; wr = alloc_wrqe(sizeof(*req), lctx->ctrlq); if (wr == NULL) { log(LOG_ERR, "%s: allocation failure", __func__); return (ENOMEM); } req = wrtod(wr); INIT_TP_WR(req, 0); OPCODE_TID(req) = htobe32(MK_OPCODE_TID(CPL_PASS_OPEN_REQ6, lctx->stid)); req->local_port = inp->inp_lport; req->peer_port = 0; req->local_ip_hi = *(uint64_t *)&inp->in6p_laddr.s6_addr[0]; req->local_ip_lo = *(uint64_t *)&inp->in6p_laddr.s6_addr[8]; req->peer_ip_hi = 0; req->peer_ip_lo = 0; req->opt0 = htobe64(V_TX_CHAN(lctx->ctrlq->eq.tx_chan)); req->opt1 = htobe64(V_CONN_POLICY(CPL_CONN_POLICY_ASK) | F_SYN_RSS_ENABLE | V_SYN_RSS_QUEUE(lctx->ofld_rxq->iq.abs_id)); t4_wrq_tx(sc, wr); return (0); } static int destroy_server(struct adapter *sc, struct listen_ctx *lctx) { struct wrqe *wr; struct cpl_close_listsvr_req *req; wr = alloc_wrqe(sizeof(*req), lctx->ctrlq); if (wr == NULL) { /* XXX */ panic("%s: allocation failure.", __func__); } req = wrtod(wr); INIT_TP_WR(req, 0); OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_CLOSE_LISTSRV_REQ, lctx->stid)); req->reply_ctrl = htobe16(lctx->ofld_rxq->iq.abs_id); req->rsvd = htobe16(0); t4_wrq_tx(sc, wr); return (0); } /* * Start a listening server by sending a passive open request to HW. * * Can't take adapter lock here and access to sc->flags, * sc->offload_map, if_capenable are all race prone. */ int t4_listen_start(struct toedev *tod, struct tcpcb *tp) { struct adapter *sc = tod->tod_softc; struct vi_info *vi; struct port_info *pi; struct inpcb *inp = tp->t_inpcb; struct listen_ctx *lctx; int i, rc, v; INP_WLOCK_ASSERT(inp); /* Don't start a hardware listener for any loopback address. */ if (inp->inp_vflag & INP_IPV6 && IN6_IS_ADDR_LOOPBACK(&inp->in6p_laddr)) return (0); if (!(inp->inp_vflag & INP_IPV6) && IN_LOOPBACK(ntohl(inp->inp_laddr.s_addr))) return (0); #if 0 ADAPTER_LOCK(sc); if (IS_BUSY(sc)) { log(LOG_ERR, "%s: listen request ignored, %s is busy", __func__, device_get_nameunit(sc->dev)); goto done; } KASSERT(uld_active(sc, ULD_TOM), ("%s: TOM not initialized", __func__)); #endif /* * Find a running VI with IFCAP_TOE (4 or 6). We'll use the first * such VI's queues to send the passive open and receive the reply to * it. * * XXX: need a way to mark a port in use by offload. if_cxgbe should * then reject any attempt to bring down such a port (and maybe reject * attempts to disable IFCAP_TOE on that port too?). */ for_each_port(sc, i) { pi = sc->port[i]; for_each_vi(pi, v, vi) { if (vi->ifp->if_drv_flags & IFF_DRV_RUNNING && vi->ifp->if_capenable & IFCAP_TOE) goto found; } } goto done; /* no port that's UP with IFCAP_TOE enabled */ found: if (listen_hash_find(sc, inp) != NULL) goto done; /* already setup */ lctx = alloc_lctx(sc, inp, vi); if (lctx == NULL) { log(LOG_ERR, "%s: listen request ignored, %s couldn't allocate lctx\n", __func__, device_get_nameunit(sc->dev)); goto done; } listen_hash_add(sc, lctx); CTR6(KTR_CXGBE, "%s: stid %u (%s), lctx %p, inp %p vflag 0x%x", __func__, lctx->stid, tcpstates[tp->t_state], lctx, inp, inp->inp_vflag); if (inp->inp_vflag & INP_IPV6) rc = create_server6(sc, lctx); else rc = create_server(sc, lctx); if (rc != 0) { log(LOG_ERR, "%s: %s failed to create hw listener: %d.\n", __func__, device_get_nameunit(sc->dev), rc); (void) listen_hash_del(sc, inp); inp = release_lctx(sc, lctx); /* can't be freed, host stack has a reference */ KASSERT(inp != NULL, ("%s: inp freed", __func__)); goto done; } lctx->flags |= LCTX_RPL_PENDING; done: #if 0 ADAPTER_UNLOCK(sc); #endif return (0); } int t4_listen_stop(struct toedev *tod, struct tcpcb *tp) { struct listen_ctx *lctx; struct adapter *sc = tod->tod_softc; struct inpcb *inp = tp->t_inpcb; struct synq_entry *synqe; INP_WLOCK_ASSERT(inp); lctx = listen_hash_del(sc, inp); if (lctx == NULL) return (ENOENT); /* no hardware listener for this inp */ CTR4(KTR_CXGBE, "%s: stid %u, lctx %p, flags %x", __func__, lctx->stid, lctx, lctx->flags); /* * If the reply to the PASS_OPEN is still pending we'll wait for it to * arrive and clean up when it does. */ if (lctx->flags & LCTX_RPL_PENDING) { KASSERT(TAILQ_EMPTY(&lctx->synq), ("%s: synq not empty.", __func__)); return (EINPROGRESS); } /* * The host stack will abort all the connections on the listening * socket's so_comp. It doesn't know about the connections on the synq * so we need to take care of those. */ TAILQ_FOREACH(synqe, &lctx->synq, link) { if (synqe->flags & TPF_SYNQE_HAS_L2TE) send_reset_synqe(tod, synqe); } destroy_server(sc, lctx); return (0); } static inline void hold_synqe(struct synq_entry *synqe) { refcount_acquire(&synqe->refcnt); } static inline void release_synqe(struct synq_entry *synqe) { if (refcount_release(&synqe->refcnt)) { int needfree = synqe->flags & TPF_SYNQE_NEEDFREE; m_freem(synqe->syn); if (needfree) free(synqe, M_CXGBE); } } void t4_syncache_added(struct toedev *tod __unused, void *arg) { struct synq_entry *synqe = arg; hold_synqe(synqe); } void t4_syncache_removed(struct toedev *tod __unused, void *arg) { struct synq_entry *synqe = arg; release_synqe(synqe); } /* XXX */ extern void tcp_dooptions(struct tcpopt *, u_char *, int, int); int t4_syncache_respond(struct toedev *tod, void *arg, struct mbuf *m) { struct adapter *sc = tod->tod_softc; struct synq_entry *synqe = arg; struct wrqe *wr; struct l2t_entry *e; struct tcpopt to; struct ip *ip = mtod(m, struct ip *); struct tcphdr *th; wr = (struct wrqe *)atomic_readandclear_ptr(&synqe->wr); if (wr == NULL) { m_freem(m); return (EALREADY); } if (ip->ip_v == IPVERSION) th = (void *)(ip + 1); else th = (void *)((struct ip6_hdr *)ip + 1); bzero(&to, sizeof(to)); tcp_dooptions(&to, (void *)(th + 1), (th->th_off << 2) - sizeof(*th), TO_SYN); /* save these for later */ synqe->iss = be32toh(th->th_seq); synqe->ts = to.to_tsval; if (is_t5(sc)) { struct cpl_t5_pass_accept_rpl *rpl5 = wrtod(wr); rpl5->iss = th->th_seq; } e = &sc->l2t->l2tab[synqe->l2e_idx]; t4_l2t_send(sc, wr, e); m_freem(m); /* don't need this any more */ return (0); } static int do_pass_open_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { struct adapter *sc = iq->adapter; const struct cpl_pass_open_rpl *cpl = (const void *)(rss + 1); int stid = GET_TID(cpl); unsigned int status = cpl->status; struct listen_ctx *lctx = lookup_stid(sc, stid); struct inpcb *inp = lctx->inp; #ifdef INVARIANTS unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); #endif KASSERT(opcode == CPL_PASS_OPEN_RPL, ("%s: unexpected opcode 0x%x", __func__, opcode)); KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); KASSERT(lctx->stid == stid, ("%s: lctx stid mismatch", __func__)); INP_WLOCK(inp); CTR4(KTR_CXGBE, "%s: stid %d, status %u, flags 0x%x", __func__, stid, status, lctx->flags); lctx->flags &= ~LCTX_RPL_PENDING; if (status != CPL_ERR_NONE) log(LOG_ERR, "listener (stid %u) failed: %d\n", stid, status); #ifdef INVARIANTS /* * If the inp has been dropped (listening socket closed) then * listen_stop must have run and taken the inp out of the hash. */ if (inp->inp_flags & INP_DROPPED) { KASSERT(listen_hash_del(sc, inp) == NULL, ("%s: inp %p still in listen hash", __func__, inp)); } #endif if (inp->inp_flags & INP_DROPPED && status != CPL_ERR_NONE) { if (release_lctx(sc, lctx) != NULL) INP_WUNLOCK(inp); return (status); } /* * Listening socket stopped listening earlier and now the chip tells us * it has started the hardware listener. Stop it; the lctx will be * released in do_close_server_rpl. */ if (inp->inp_flags & INP_DROPPED) { destroy_server(sc, lctx); INP_WUNLOCK(inp); return (status); } /* * Failed to start hardware listener. Take inp out of the hash and * release our reference on it. An error message has been logged * already. */ if (status != CPL_ERR_NONE) { listen_hash_del(sc, inp); if (release_lctx(sc, lctx) != NULL) INP_WUNLOCK(inp); return (status); } /* hardware listener open for business */ INP_WUNLOCK(inp); return (status); } static int do_close_server_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { struct adapter *sc = iq->adapter; const struct cpl_close_listsvr_rpl *cpl = (const void *)(rss + 1); int stid = GET_TID(cpl); unsigned int status = cpl->status; struct listen_ctx *lctx = lookup_stid(sc, stid); struct inpcb *inp = lctx->inp; #ifdef INVARIANTS unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); #endif KASSERT(opcode == CPL_CLOSE_LISTSRV_RPL, ("%s: unexpected opcode 0x%x", __func__, opcode)); KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); KASSERT(lctx->stid == stid, ("%s: lctx stid mismatch", __func__)); CTR3(KTR_CXGBE, "%s: stid %u, status %u", __func__, stid, status); if (status != CPL_ERR_NONE) { log(LOG_ERR, "%s: failed (%u) to close listener for stid %u\n", __func__, status, stid); return (status); } INP_WLOCK(inp); inp = release_lctx(sc, lctx); if (inp != NULL) INP_WUNLOCK(inp); return (status); } static void done_with_synqe(struct adapter *sc, struct synq_entry *synqe) { struct listen_ctx *lctx = synqe->lctx; struct inpcb *inp = lctx->inp; struct vi_info *vi = synqe->syn->m_pkthdr.rcvif->if_softc; struct l2t_entry *e = &sc->l2t->l2tab[synqe->l2e_idx]; INP_WLOCK_ASSERT(inp); TAILQ_REMOVE(&lctx->synq, synqe, link); inp = release_lctx(sc, lctx); if (inp) INP_WUNLOCK(inp); remove_tid(sc, synqe->tid); release_tid(sc, synqe->tid, &sc->sge.ctrlq[vi->pi->port_id]); t4_l2t_release(e); release_synqe(synqe); /* removed from synq list */ } int do_abort_req_synqe(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { struct adapter *sc = iq->adapter; const struct cpl_abort_req_rss *cpl = (const void *)(rss + 1); unsigned int tid = GET_TID(cpl); struct synq_entry *synqe = lookup_tid(sc, tid); struct listen_ctx *lctx = synqe->lctx; struct inpcb *inp = lctx->inp; int txqid; struct sge_wrq *ofld_txq; #ifdef INVARIANTS unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); #endif KASSERT(opcode == CPL_ABORT_REQ_RSS, ("%s: unexpected opcode 0x%x", __func__, opcode)); KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); KASSERT(synqe->tid == tid, ("%s: toep tid mismatch", __func__)); CTR6(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x), lctx %p, status %d", __func__, tid, synqe, synqe->flags, synqe->lctx, cpl->status); if (negative_advice(cpl->status)) return (0); /* Ignore negative advice */ INP_WLOCK(inp); get_qids_from_mbuf(synqe->syn, &txqid, NULL); ofld_txq = &sc->sge.ofld_txq[txqid]; /* * If we'd initiated an abort earlier the reply to it is responsible for * cleaning up resources. Otherwise we tear everything down right here * right now. We owe the T4 a CPL_ABORT_RPL no matter what. */ if (synqe->flags & TPF_ABORT_SHUTDOWN) { INP_WUNLOCK(inp); goto done; } done_with_synqe(sc, synqe); /* inp lock released by done_with_synqe */ done: send_abort_rpl(sc, ofld_txq, tid, CPL_ABORT_NO_RST); return (0); } int do_abort_rpl_synqe(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { struct adapter *sc = iq->adapter; const struct cpl_abort_rpl_rss *cpl = (const void *)(rss + 1); unsigned int tid = GET_TID(cpl); struct synq_entry *synqe = lookup_tid(sc, tid); struct listen_ctx *lctx = synqe->lctx; struct inpcb *inp = lctx->inp; #ifdef INVARIANTS unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); #endif KASSERT(opcode == CPL_ABORT_RPL_RSS, ("%s: unexpected opcode 0x%x", __func__, opcode)); KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); KASSERT(synqe->tid == tid, ("%s: toep tid mismatch", __func__)); CTR6(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x), lctx %p, status %d", __func__, tid, synqe, synqe->flags, synqe->lctx, cpl->status); INP_WLOCK(inp); KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN, ("%s: wasn't expecting abort reply for synqe %p (0x%x)", __func__, synqe, synqe->flags)); done_with_synqe(sc, synqe); /* inp lock released by done_with_synqe */ return (0); } void t4_offload_socket(struct toedev *tod, void *arg, struct socket *so) { struct adapter *sc = tod->tod_softc; struct synq_entry *synqe = arg; #ifdef INVARIANTS struct inpcb *inp = sotoinpcb(so); #endif struct cpl_pass_establish *cpl = mtod(synqe->syn, void *); struct toepcb *toep = *(struct toepcb **)(cpl + 1); INP_INFO_RLOCK_ASSERT(&V_tcbinfo); /* prevents bad race with accept() */ INP_WLOCK_ASSERT(inp); KASSERT(synqe->flags & TPF_SYNQE, ("%s: %p not a synq_entry?", __func__, arg)); offload_socket(so, toep); make_established(toep, cpl->snd_isn, cpl->rcv_isn, cpl->tcp_opt); toep->flags |= TPF_CPL_PENDING; update_tid(sc, synqe->tid, toep); synqe->flags |= TPF_SYNQE_EXPANDED; } static inline void save_qids_in_mbuf(struct mbuf *m, struct vi_info *vi) { uint32_t txqid, rxqid; txqid = (arc4random() % vi->nofldtxq) + vi->first_ofld_txq; rxqid = (arc4random() % vi->nofldrxq) + vi->first_ofld_rxq; m->m_pkthdr.flowid = (txqid << 16) | (rxqid & 0xffff); } static inline void get_qids_from_mbuf(struct mbuf *m, int *txqid, int *rxqid) { if (txqid) *txqid = m->m_pkthdr.flowid >> 16; if (rxqid) *rxqid = m->m_pkthdr.flowid & 0xffff; } /* * Use the trailing space in the mbuf in which the PASS_ACCEPT_REQ arrived to * store some state temporarily. */ static struct synq_entry * mbuf_to_synqe(struct mbuf *m) { int len = roundup2(sizeof (struct synq_entry), 8); int tspace = M_TRAILINGSPACE(m); struct synq_entry *synqe = NULL; if (tspace < len) { synqe = malloc(sizeof(*synqe), M_CXGBE, M_NOWAIT); if (synqe == NULL) return (NULL); synqe->flags = TPF_SYNQE | TPF_SYNQE_NEEDFREE; } else { synqe = (void *)(m->m_data + m->m_len + tspace - len); synqe->flags = TPF_SYNQE; } return (synqe); } static void t4opt_to_tcpopt(const struct tcp_options *t4opt, struct tcpopt *to) { bzero(to, sizeof(*to)); if (t4opt->mss) { to->to_flags |= TOF_MSS; to->to_mss = be16toh(t4opt->mss); } if (t4opt->wsf) { to->to_flags |= TOF_SCALE; to->to_wscale = t4opt->wsf; } if (t4opt->tstamp) to->to_flags |= TOF_TS; if (t4opt->sack) to->to_flags |= TOF_SACKPERM; } /* * Options2 for passive open. */ static uint32_t calc_opt2p(struct adapter *sc, struct port_info *pi, int rxqid, const struct tcp_options *tcpopt, struct tcphdr *th, int ulp_mode) { struct sge_ofld_rxq *ofld_rxq = &sc->sge.ofld_rxq[rxqid]; uint32_t opt2; opt2 = V_TX_QUEUE(sc->params.tp.tx_modq[pi->tx_chan]) | F_RSS_QUEUE_VALID | V_RSS_QUEUE(ofld_rxq->iq.abs_id); if (V_tcp_do_rfc1323) { if (tcpopt->tstamp) opt2 |= F_TSTAMPS_EN; if (tcpopt->sack) opt2 |= F_SACK_EN; if (tcpopt->wsf <= 14) opt2 |= F_WND_SCALE_EN; } if (V_tcp_do_ecn && th->th_flags & (TH_ECE | TH_CWR)) opt2 |= F_CCTRL_ECN; /* RX_COALESCE is always a valid value (0 or M_RX_COALESCE). */ if (is_t4(sc)) opt2 |= F_RX_COALESCE_VALID; else { opt2 |= F_T5_OPT_2_VALID; opt2 |= F_CONG_CNTRL_VALID; /* OPT_2_ISS really, for T5 */ } if (sc->tt.rx_coalesce) opt2 |= V_RX_COALESCE(M_RX_COALESCE); #ifdef USE_DDP_RX_FLOW_CONTROL if (ulp_mode == ULP_MODE_TCPDDP) opt2 |= F_RX_FC_VALID | F_RX_FC_DDP; #endif return htobe32(opt2); } static void pass_accept_req_to_protohdrs(const struct mbuf *m, struct in_conninfo *inc, struct tcphdr *th) { const struct cpl_pass_accept_req *cpl = mtod(m, const void *); const struct ether_header *eh; unsigned int hlen = be32toh(cpl->hdr_len); uintptr_t l3hdr; const struct tcphdr *tcp; eh = (const void *)(cpl + 1); l3hdr = ((uintptr_t)eh + G_ETH_HDR_LEN(hlen)); tcp = (const void *)(l3hdr + G_IP_HDR_LEN(hlen)); if (inc) { bzero(inc, sizeof(*inc)); inc->inc_fport = tcp->th_sport; inc->inc_lport = tcp->th_dport; if (((struct ip *)l3hdr)->ip_v == IPVERSION) { const struct ip *ip = (const void *)l3hdr; inc->inc_faddr = ip->ip_src; inc->inc_laddr = ip->ip_dst; } else { const struct ip6_hdr *ip6 = (const void *)l3hdr; inc->inc_flags |= INC_ISIPV6; inc->inc6_faddr = ip6->ip6_src; inc->inc6_laddr = ip6->ip6_dst; } } if (th) { bcopy(tcp, th, sizeof(*th)); tcp_fields_to_host(th); /* just like tcp_input */ } } static struct l2t_entry * get_l2te_for_nexthop(struct port_info *pi, struct ifnet *ifp, struct in_conninfo *inc) { struct l2t_entry *e; struct sockaddr_in6 sin6; struct sockaddr *dst = (void *)&sin6; if (inc->inc_flags & INC_ISIPV6) { struct nhop6_basic nh6; bzero(dst, sizeof(struct sockaddr_in6)); dst->sa_len = sizeof(struct sockaddr_in6); dst->sa_family = AF_INET6; if (IN6_IS_ADDR_LINKLOCAL(&inc->inc6_laddr)) { /* no need for route lookup */ e = t4_l2t_get(pi, ifp, dst); return (e); } if (fib6_lookup_nh_basic(RT_DEFAULT_FIB, &inc->inc6_faddr, 0, 0, 0, &nh6) != 0) return (NULL); if (nh6.nh_ifp != ifp) return (NULL); ((struct sockaddr_in6 *)dst)->sin6_addr = nh6.nh_addr; } else { struct nhop4_basic nh4; dst->sa_len = sizeof(struct sockaddr_in); dst->sa_family = AF_INET; if (fib4_lookup_nh_basic(RT_DEFAULT_FIB, inc->inc_faddr, 0, 0, &nh4) != 0) return (NULL); if (nh4.nh_ifp != ifp) return (NULL); ((struct sockaddr_in *)dst)->sin_addr = nh4.nh_addr; } e = t4_l2t_get(pi, ifp, dst); return (e); } #define REJECT_PASS_ACCEPT() do { \ reject_reason = __LINE__; \ goto reject; \ } while (0) /* * The context associated with a tid entry via insert_tid could be a synq_entry * or a toepcb. The only way CPL handlers can tell is via a bit in these flags. */ CTASSERT(offsetof(struct toepcb, flags) == offsetof(struct synq_entry, flags)); /* * Incoming SYN on a listening socket. * * XXX: Every use of ifp in this routine has a bad race with up/down, toe/-toe, * etc. */ static int do_pass_accept_req(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { struct adapter *sc = iq->adapter; struct toedev *tod; const struct cpl_pass_accept_req *cpl = mtod(m, const void *); struct cpl_pass_accept_rpl *rpl; struct wrqe *wr; unsigned int stid = G_PASS_OPEN_TID(be32toh(cpl->tos_stid)); unsigned int tid = GET_TID(cpl); struct listen_ctx *lctx = lookup_stid(sc, stid); struct inpcb *inp; struct socket *so; struct in_conninfo inc; struct tcphdr th; struct tcpopt to; struct port_info *pi; struct vi_info *vi; struct ifnet *hw_ifp, *ifp; struct l2t_entry *e = NULL; int rscale, mtu_idx, rx_credits, rxqid, ulp_mode; struct synq_entry *synqe = NULL; int reject_reason, v; uint16_t vid; #ifdef INVARIANTS unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); #endif KASSERT(opcode == CPL_PASS_ACCEPT_REQ, ("%s: unexpected opcode 0x%x", __func__, opcode)); KASSERT(lctx->stid == stid, ("%s: lctx stid mismatch", __func__)); CTR4(KTR_CXGBE, "%s: stid %u, tid %u, lctx %p", __func__, stid, tid, lctx); pass_accept_req_to_protohdrs(m, &inc, &th); t4opt_to_tcpopt(&cpl->tcpopt, &to); pi = sc->port[G_SYN_INTF(be16toh(cpl->l2info))]; /* * Use the MAC index to lookup the associated VI. If this SYN * didn't match a perfect MAC filter, punt. */ if (!(be16toh(cpl->l2info) & F_SYN_XACT_MATCH)) { m_freem(m); m = NULL; REJECT_PASS_ACCEPT(); } for_each_vi(pi, v, vi) { if (vi->xact_addr_filt == G_SYN_MAC_IDX(be16toh(cpl->l2info))) goto found; } m_freem(m); m = NULL; REJECT_PASS_ACCEPT(); found: hw_ifp = vi->ifp; /* the (v)cxgbeX ifnet */ m->m_pkthdr.rcvif = hw_ifp; tod = TOEDEV(hw_ifp); /* * Figure out if there is a pseudo interface (vlan, lagg, etc.) * involved. Don't offload if the SYN had a VLAN tag and the vid * doesn't match anything on this interface. * * XXX: lagg support, lagg + vlan support. */ vid = EVL_VLANOFTAG(be16toh(cpl->vlan)); if (vid != 0xfff) { ifp = VLAN_DEVAT(hw_ifp, vid); if (ifp == NULL) REJECT_PASS_ACCEPT(); } else ifp = hw_ifp; /* * Don't offload if the peer requested a TCP option that's not known to * the silicon. */ if (cpl->tcpopt.unknown) REJECT_PASS_ACCEPT(); if (inc.inc_flags & INC_ISIPV6) { /* Don't offload if the ifcap isn't enabled */ if ((ifp->if_capenable & IFCAP_TOE6) == 0) REJECT_PASS_ACCEPT(); /* * SYN must be directed to an IP6 address on this ifnet. This * is more restrictive than in6_localip. */ if (!in6_ifhasaddr(ifp, &inc.inc6_laddr)) REJECT_PASS_ACCEPT(); } else { /* Don't offload if the ifcap isn't enabled */ if ((ifp->if_capenable & IFCAP_TOE4) == 0) REJECT_PASS_ACCEPT(); /* * SYN must be directed to an IP address on this ifnet. This * is more restrictive than in_localip. */ if (!in_ifhasaddr(ifp, inc.inc_laddr)) REJECT_PASS_ACCEPT(); } e = get_l2te_for_nexthop(pi, ifp, &inc); if (e == NULL) REJECT_PASS_ACCEPT(); synqe = mbuf_to_synqe(m); if (synqe == NULL) REJECT_PASS_ACCEPT(); wr = alloc_wrqe(is_t4(sc) ? sizeof(struct cpl_pass_accept_rpl) : sizeof(struct cpl_t5_pass_accept_rpl), &sc->sge.ctrlq[pi->port_id]); if (wr == NULL) REJECT_PASS_ACCEPT(); rpl = wrtod(wr); INP_INFO_RLOCK(&V_tcbinfo); /* for 4-tuple check */ /* Don't offload if the 4-tuple is already in use */ if (toe_4tuple_check(&inc, &th, ifp) != 0) { INP_INFO_RUNLOCK(&V_tcbinfo); free(wr, M_CXGBE); REJECT_PASS_ACCEPT(); } INP_INFO_RUNLOCK(&V_tcbinfo); inp = lctx->inp; /* listening socket, not owned by TOE */ INP_WLOCK(inp); /* Don't offload if the listening socket has closed */ if (__predict_false(inp->inp_flags & INP_DROPPED)) { /* * The listening socket has closed. The reply from the TOE to * our CPL_CLOSE_LISTSRV_REQ will ultimately release all * resources tied to this listen context. */ INP_WUNLOCK(inp); free(wr, M_CXGBE); REJECT_PASS_ACCEPT(); } so = inp->inp_socket; + CURVNET_SET(so->so_vnet); mtu_idx = find_best_mtu_idx(sc, &inc, be16toh(cpl->tcpopt.mss)); rscale = cpl->tcpopt.wsf && V_tcp_do_rfc1323 ? select_rcv_wscale() : 0; SOCKBUF_LOCK(&so->so_rcv); /* opt0 rcv_bufsiz initially, assumes its normal meaning later */ rx_credits = min(select_rcv_wnd(so) >> 10, M_RCV_BUFSIZ); SOCKBUF_UNLOCK(&so->so_rcv); save_qids_in_mbuf(m, vi); get_qids_from_mbuf(m, NULL, &rxqid); if (is_t4(sc)) INIT_TP_WR_MIT_CPL(rpl, CPL_PASS_ACCEPT_RPL, tid); else { struct cpl_t5_pass_accept_rpl *rpl5 = (void *)rpl; INIT_TP_WR_MIT_CPL(rpl5, CPL_PASS_ACCEPT_RPL, tid); } if (sc->tt.ddp && (so->so_options & SO_NO_DDP) == 0) { ulp_mode = ULP_MODE_TCPDDP; synqe->flags |= TPF_SYNQE_TCPDDP; } else ulp_mode = ULP_MODE_NONE; rpl->opt0 = calc_opt0(so, vi, e, mtu_idx, rscale, rx_credits, ulp_mode); rpl->opt2 = calc_opt2p(sc, pi, rxqid, &cpl->tcpopt, &th, ulp_mode); synqe->tid = tid; synqe->lctx = lctx; synqe->syn = m; m = NULL; refcount_init(&synqe->refcnt, 1); /* 1 means extra hold */ synqe->l2e_idx = e->idx; synqe->rcv_bufsize = rx_credits; atomic_store_rel_ptr(&synqe->wr, (uintptr_t)wr); insert_tid(sc, tid, synqe); TAILQ_INSERT_TAIL(&lctx->synq, synqe, link); hold_synqe(synqe); /* hold for the duration it's in the synq */ hold_lctx(lctx); /* A synqe on the list has a ref on its lctx */ /* * If all goes well t4_syncache_respond will get called during * syncache_add. Note that syncache_add releases the pcb lock. */ toe_syncache_add(&inc, &to, &th, inp, tod, synqe); INP_UNLOCK_ASSERT(inp); /* ok to assert, we have a ref on the inp */ + CURVNET_RESTORE(); /* * If we replied during syncache_add (synqe->wr has been consumed), * good. Otherwise, set it to 0 so that further syncache_respond * attempts by the kernel will be ignored. */ if (atomic_cmpset_ptr(&synqe->wr, (uintptr_t)wr, 0)) { /* * syncache may or may not have a hold on the synqe, which may * or may not be stashed in the original SYN mbuf passed to us. * Just copy it over instead of dealing with all possibilities. */ m = m_dup(synqe->syn, M_NOWAIT); if (m) m->m_pkthdr.rcvif = hw_ifp; remove_tid(sc, synqe->tid); free(wr, M_CXGBE); /* Yank the synqe out of the lctx synq. */ INP_WLOCK(inp); TAILQ_REMOVE(&lctx->synq, synqe, link); release_synqe(synqe); /* removed from synq list */ inp = release_lctx(sc, lctx); if (inp) INP_WUNLOCK(inp); release_synqe(synqe); /* extra hold */ REJECT_PASS_ACCEPT(); } CTR5(KTR_CXGBE, "%s: stid %u, tid %u, lctx %p, synqe %p, SYNACK", __func__, stid, tid, lctx, synqe); INP_WLOCK(inp); synqe->flags |= TPF_SYNQE_HAS_L2TE; if (__predict_false(inp->inp_flags & INP_DROPPED)) { /* * Listening socket closed but tod_listen_stop did not abort * this tid because there was no L2T entry for the tid at that * time. Abort it now. The reply to the abort will clean up. */ CTR6(KTR_CXGBE, "%s: stid %u, tid %u, lctx %p, synqe %p (0x%x), ABORT", __func__, stid, tid, lctx, synqe, synqe->flags); if (!(synqe->flags & TPF_SYNQE_EXPANDED)) send_reset_synqe(tod, synqe); INP_WUNLOCK(inp); release_synqe(synqe); /* extra hold */ return (__LINE__); } INP_WUNLOCK(inp); release_synqe(synqe); /* extra hold */ return (0); reject: CTR4(KTR_CXGBE, "%s: stid %u, tid %u, REJECT (%d)", __func__, stid, tid, reject_reason); if (e) t4_l2t_release(e); release_tid(sc, tid, lctx->ctrlq); if (__predict_true(m != NULL)) { m_adj(m, sizeof(*cpl)); m->m_pkthdr.csum_flags |= (CSUM_IP_CHECKED | CSUM_IP_VALID | CSUM_DATA_VALID | CSUM_PSEUDO_HDR); m->m_pkthdr.csum_data = 0xffff; hw_ifp->if_input(hw_ifp, m); } return (reject_reason); } static void synqe_to_protohdrs(struct synq_entry *synqe, const struct cpl_pass_establish *cpl, struct in_conninfo *inc, struct tcphdr *th, struct tcpopt *to) { uint16_t tcp_opt = be16toh(cpl->tcp_opt); /* start off with the original SYN */ pass_accept_req_to_protohdrs(synqe->syn, inc, th); /* modify parts to make it look like the ACK to our SYN|ACK */ th->th_flags = TH_ACK; th->th_ack = synqe->iss + 1; th->th_seq = be32toh(cpl->rcv_isn); bzero(to, sizeof(*to)); if (G_TCPOPT_TSTAMP(tcp_opt)) { to->to_flags |= TOF_TS; to->to_tsecr = synqe->ts; } } static int do_pass_establish(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) { struct adapter *sc = iq->adapter; struct vi_info *vi; struct ifnet *ifp; const struct cpl_pass_establish *cpl = (const void *)(rss + 1); #if defined(KTR) || defined(INVARIANTS) unsigned int stid = G_PASS_OPEN_TID(be32toh(cpl->tos_stid)); #endif unsigned int tid = GET_TID(cpl); struct synq_entry *synqe = lookup_tid(sc, tid); struct listen_ctx *lctx = synqe->lctx; struct inpcb *inp = lctx->inp, *new_inp; struct socket *so; struct tcphdr th; struct tcpopt to; struct in_conninfo inc; struct toepcb *toep; u_int txqid, rxqid; #ifdef INVARIANTS unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); #endif KASSERT(opcode == CPL_PASS_ESTABLISH, ("%s: unexpected opcode 0x%x", __func__, opcode)); KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); KASSERT(lctx->stid == stid, ("%s: lctx stid mismatch", __func__)); KASSERT(synqe->flags & TPF_SYNQE, ("%s: tid %u (ctx %p) not a synqe", __func__, tid, synqe)); INP_INFO_RLOCK(&V_tcbinfo); /* for syncache_expand */ INP_WLOCK(inp); CTR6(KTR_CXGBE, "%s: stid %u, tid %u, synqe %p (0x%x), inp_flags 0x%x", __func__, stid, tid, synqe, synqe->flags, inp->inp_flags); if (__predict_false(inp->inp_flags & INP_DROPPED)) { if (synqe->flags & TPF_SYNQE_HAS_L2TE) { KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN, ("%s: listen socket closed but tid %u not aborted.", __func__, tid)); } INP_WUNLOCK(inp); INP_INFO_RUNLOCK(&V_tcbinfo); return (0); } ifp = synqe->syn->m_pkthdr.rcvif; vi = ifp->if_softc; KASSERT(vi->pi->adapter == sc, ("%s: vi %p, sc %p mismatch", __func__, vi, sc)); get_qids_from_mbuf(synqe->syn, &txqid, &rxqid); KASSERT(rxqid == iq_to_ofld_rxq(iq) - &sc->sge.ofld_rxq[0], ("%s: CPL arrived on unexpected rxq. %d %d", __func__, rxqid, (int)(iq_to_ofld_rxq(iq) - &sc->sge.ofld_rxq[0]))); toep = alloc_toepcb(vi, txqid, rxqid, M_NOWAIT); if (toep == NULL) { reset: /* * The reply to this abort will perform final cleanup. There is * no need to check for HAS_L2TE here. We can be here only if * we responded to the PASS_ACCEPT_REQ, and our response had the * L2T idx. */ send_reset_synqe(TOEDEV(ifp), synqe); INP_WUNLOCK(inp); INP_INFO_RUNLOCK(&V_tcbinfo); return (0); } toep->tid = tid; toep->l2te = &sc->l2t->l2tab[synqe->l2e_idx]; if (synqe->flags & TPF_SYNQE_TCPDDP) set_tcpddp_ulp_mode(toep); else toep->ulp_mode = ULP_MODE_NONE; /* opt0 rcv_bufsiz initially, assumes its normal meaning later */ toep->rx_credits = synqe->rcv_bufsize; so = inp->inp_socket; KASSERT(so != NULL, ("%s: socket is NULL", __func__)); /* Come up with something that syncache_expand should be ok with. */ synqe_to_protohdrs(synqe, cpl, &inc, &th, &to); /* * No more need for anything in the mbuf that carried the * CPL_PASS_ACCEPT_REQ. Drop the CPL_PASS_ESTABLISH and toep pointer * there. XXX: bad form but I don't want to increase the size of synqe. */ m = synqe->syn; KASSERT(sizeof(*cpl) + sizeof(toep) <= m->m_len, ("%s: no room in mbuf %p (m_len %d)", __func__, m, m->m_len)); bcopy(cpl, mtod(m, void *), sizeof(*cpl)); *(struct toepcb **)(mtod(m, struct cpl_pass_establish *) + 1) = toep; if (!toe_syncache_expand(&inc, &to, &th, &so) || so == NULL) { free_toepcb(toep); goto reset; } /* New connection inpcb is already locked by syncache_expand(). */ new_inp = sotoinpcb(so); INP_WLOCK_ASSERT(new_inp); /* * This is for the unlikely case where the syncache entry that we added * has been evicted from the syncache, but the syncache_expand above * works because of syncookies. * * XXX: we've held the tcbinfo lock throughout so there's no risk of * anyone accept'ing a connection before we've installed our hooks, but * this somewhat defeats the purpose of having a tod_offload_socket :-( */ if (__predict_false(!(synqe->flags & TPF_SYNQE_EXPANDED))) { tcp_timer_activate(intotcpcb(new_inp), TT_KEEP, 0); t4_offload_socket(TOEDEV(ifp), synqe, so); } INP_WUNLOCK(new_inp); /* Done with the synqe */ TAILQ_REMOVE(&lctx->synq, synqe, link); inp = release_lctx(sc, lctx); if (inp != NULL) INP_WUNLOCK(inp); INP_INFO_RUNLOCK(&V_tcbinfo); release_synqe(synqe); return (0); } void t4_init_listen_cpl_handlers(struct adapter *sc) { t4_register_cpl_handler(sc, CPL_PASS_OPEN_RPL, do_pass_open_rpl); t4_register_cpl_handler(sc, CPL_CLOSE_LISTSRV_RPL, do_close_server_rpl); t4_register_cpl_handler(sc, CPL_PASS_ACCEPT_REQ, do_pass_accept_req); t4_register_cpl_handler(sc, CPL_PASS_ESTABLISH, do_pass_establish); } #endif