t4_cpl_io.c revision 300895
1237263Snp/*- 2292736Snp * Copyright (c) 2012, 2015 Chelsio Communications, Inc. 3237263Snp * All rights reserved. 4237263Snp * Written by: Navdeep Parhar <np@FreeBSD.org> 5237263Snp * 6237263Snp * Redistribution and use in source and binary forms, with or without 7237263Snp * modification, are permitted provided that the following conditions 8237263Snp * are met: 9237263Snp * 1. Redistributions of source code must retain the above copyright 10237263Snp * notice, this list of conditions and the following disclaimer. 11237263Snp * 2. Redistributions in binary form must reproduce the above copyright 12237263Snp * notice, this list of conditions and the following disclaimer in the 13237263Snp * documentation and/or other materials provided with the distribution. 14237263Snp * 15237263Snp * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 16237263Snp * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 17237263Snp * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 18237263Snp * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 19237263Snp * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 20237263Snp * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 21237263Snp * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 22237263Snp * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 23237263Snp * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 24237263Snp * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 25237263Snp * SUCH DAMAGE. 26237263Snp */ 27237263Snp 28237263Snp#include <sys/cdefs.h> 29237263Snp__FBSDID("$FreeBSD: head/sys/dev/cxgbe/tom/t4_cpl_io.c 300895 2016-05-28 00:38:17Z np $"); 30237263Snp 31237263Snp#include "opt_inet.h" 32237263Snp 33237263Snp#ifdef TCP_OFFLOAD 34237263Snp#include <sys/param.h> 35237263Snp#include <sys/types.h> 36237263Snp#include <sys/kernel.h> 37237263Snp#include <sys/ktr.h> 38237263Snp#include <sys/module.h> 39237263Snp#include <sys/protosw.h> 40237263Snp#include <sys/domain.h> 41237263Snp#include <sys/socket.h> 42237263Snp#include <sys/socketvar.h> 43237263Snp#include <sys/sglist.h> 44237263Snp#include <netinet/in.h> 45237263Snp#include <netinet/in_pcb.h> 46237263Snp#include <netinet/ip.h> 47276574Snp#include <netinet/ip6.h> 48237263Snp#define TCPSTATES 49237263Snp#include <netinet/tcp_fsm.h> 50237263Snp#include <netinet/tcp_seq.h> 51294869Sglebius#include <netinet/tcp_var.h> 52237263Snp#include <netinet/toecore.h> 53237263Snp 54237263Snp#include "common/common.h" 55237263Snp#include "common/t4_msg.h" 56237263Snp#include "common/t4_regs.h" 57239344Snp#include "common/t4_tcb.h" 58237263Snp#include "tom/t4_tom_l2t.h" 59237263Snp#include "tom/t4_tom.h" 60237263Snp 61237263SnpVNET_DECLARE(int, tcp_do_autosndbuf); 62237263Snp#define V_tcp_do_autosndbuf VNET(tcp_do_autosndbuf) 63237263SnpVNET_DECLARE(int, tcp_autosndbuf_inc); 64237263Snp#define V_tcp_autosndbuf_inc VNET(tcp_autosndbuf_inc) 65237263SnpVNET_DECLARE(int, tcp_autosndbuf_max); 66237263Snp#define V_tcp_autosndbuf_max VNET(tcp_autosndbuf_max) 67237263SnpVNET_DECLARE(int, tcp_do_autorcvbuf); 68237263Snp#define V_tcp_do_autorcvbuf VNET(tcp_do_autorcvbuf) 69237263SnpVNET_DECLARE(int, tcp_autorcvbuf_inc); 70237263Snp#define V_tcp_autorcvbuf_inc VNET(tcp_autorcvbuf_inc) 71237263SnpVNET_DECLARE(int, tcp_autorcvbuf_max); 72237263Snp#define V_tcp_autorcvbuf_max VNET(tcp_autorcvbuf_max) 73237263Snp 74237263Snpvoid 75237263Snpsend_flowc_wr(struct toepcb *toep, struct flowc_tx_params *ftxp) 76237263Snp{ 77241626Snp struct wrqe *wr; 78241626Snp struct fw_flowc_wr *flowc; 79241642Snp unsigned int nparams = ftxp ? 8 : 6, flowclen; 80291665Sjhb struct vi_info *vi = toep->vi; 81291665Sjhb struct port_info *pi = vi->pi; 82237263Snp struct adapter *sc = pi->adapter; 83291665Sjhb unsigned int pfvf = G_FW_VIID_PFN(vi->viid) << S_FW_VIID_PFN; 84237263Snp struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx]; 85237263Snp 86239514Snp KASSERT(!(toep->flags & TPF_FLOWC_WR_SENT), 87237263Snp ("%s: flowc for tid %u sent already", __func__, toep->tid)); 88237263Snp 89237263Snp flowclen = sizeof(*flowc) + nparams * sizeof(struct fw_flowc_mnemval); 90237263Snp 91248925Snp wr = alloc_wrqe(roundup2(flowclen, 16), toep->ofld_txq); 92237263Snp if (wr == NULL) { 93237263Snp /* XXX */ 94237263Snp panic("%s: allocation failure.", __func__); 95237263Snp } 96237263Snp flowc = wrtod(wr); 97237263Snp memset(flowc, 0, wr->wr_len); 98237263Snp 99237263Snp flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) | 100237263Snp V_FW_FLOWC_WR_NPARAMS(nparams)); 101237263Snp flowc->flowid_len16 = htonl(V_FW_WR_LEN16(howmany(flowclen, 16)) | 102237263Snp V_FW_WR_FLOWID(toep->tid)); 103237263Snp 104237263Snp flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_PFNVFN; 105241626Snp flowc->mnemval[0].val = htobe32(pfvf); 106241626Snp flowc->mnemval[1].mnemonic = FW_FLOWC_MNEM_CH; 107241626Snp flowc->mnemval[1].val = htobe32(pi->tx_chan); 108241626Snp flowc->mnemval[2].mnemonic = FW_FLOWC_MNEM_PORT; 109241626Snp flowc->mnemval[2].val = htobe32(pi->tx_chan); 110241626Snp flowc->mnemval[3].mnemonic = FW_FLOWC_MNEM_IQID; 111241626Snp flowc->mnemval[3].val = htobe32(toep->ofld_rxq->iq.abs_id); 112237263Snp if (ftxp) { 113237263Snp uint32_t sndbuf = min(ftxp->snd_space, sc->tt.sndbuf); 114237263Snp 115237263Snp flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_SNDNXT; 116237263Snp flowc->mnemval[4].val = htobe32(ftxp->snd_nxt); 117237263Snp flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_RCVNXT; 118237263Snp flowc->mnemval[5].val = htobe32(ftxp->rcv_nxt); 119237263Snp flowc->mnemval[6].mnemonic = FW_FLOWC_MNEM_SNDBUF; 120237263Snp flowc->mnemval[6].val = htobe32(sndbuf); 121237263Snp flowc->mnemval[7].mnemonic = FW_FLOWC_MNEM_MSS; 122237263Snp flowc->mnemval[7].val = htobe32(ftxp->mss); 123276570Snp 124276570Snp CTR6(KTR_CXGBE, 125276570Snp "%s: tid %u, mss %u, sndbuf %u, snd_nxt 0x%x, rcv_nxt 0x%x", 126276570Snp __func__, toep->tid, ftxp->mss, sndbuf, ftxp->snd_nxt, 127276570Snp ftxp->rcv_nxt); 128241642Snp } else { 129241642Snp flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_SNDBUF; 130241642Snp flowc->mnemval[4].val = htobe32(512); 131241642Snp flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_MSS; 132241642Snp flowc->mnemval[5].val = htobe32(512); 133276570Snp 134276570Snp CTR2(KTR_CXGBE, "%s: tid %u", __func__, toep->tid); 135237263Snp } 136237263Snp 137237263Snp txsd->tx_credits = howmany(flowclen, 16); 138237263Snp txsd->plen = 0; 139237263Snp KASSERT(toep->tx_credits >= txsd->tx_credits && toep->txsd_avail > 0, 140237263Snp ("%s: not enough credits (%d)", __func__, toep->tx_credits)); 141237263Snp toep->tx_credits -= txsd->tx_credits; 142237263Snp if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) 143237263Snp toep->txsd_pidx = 0; 144237263Snp toep->txsd_avail--; 145237263Snp 146239514Snp toep->flags |= TPF_FLOWC_WR_SENT; 147237263Snp t4_wrq_tx(sc, wr); 148237263Snp} 149237263Snp 150237263Snpvoid 151237263Snpsend_reset(struct adapter *sc, struct toepcb *toep, uint32_t snd_nxt) 152237263Snp{ 153237263Snp struct wrqe *wr; 154237263Snp struct cpl_abort_req *req; 155237263Snp int tid = toep->tid; 156237263Snp struct inpcb *inp = toep->inp; 157237263Snp struct tcpcb *tp = intotcpcb(inp); /* don't use if INP_DROPPED */ 158237263Snp 159237263Snp INP_WLOCK_ASSERT(inp); 160237263Snp 161237263Snp CTR6(KTR_CXGBE, "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x%s", 162237263Snp __func__, toep->tid, 163237263Snp inp->inp_flags & INP_DROPPED ? "inp dropped" : 164237263Snp tcpstates[tp->t_state], 165237263Snp toep->flags, inp->inp_flags, 166239514Snp toep->flags & TPF_ABORT_SHUTDOWN ? 167237263Snp " (abort already in progress)" : ""); 168237263Snp 169239514Snp if (toep->flags & TPF_ABORT_SHUTDOWN) 170237263Snp return; /* abort already in progress */ 171237263Snp 172239514Snp toep->flags |= TPF_ABORT_SHUTDOWN; 173237263Snp 174239514Snp KASSERT(toep->flags & TPF_FLOWC_WR_SENT, 175237263Snp ("%s: flowc_wr not sent for tid %d.", __func__, tid)); 176237263Snp 177237263Snp wr = alloc_wrqe(sizeof(*req), toep->ofld_txq); 178237263Snp if (wr == NULL) { 179237263Snp /* XXX */ 180237263Snp panic("%s: allocation failure.", __func__); 181237263Snp } 182237263Snp req = wrtod(wr); 183237263Snp 184237263Snp INIT_TP_WR_MIT_CPL(req, CPL_ABORT_REQ, tid); 185237263Snp if (inp->inp_flags & INP_DROPPED) 186237263Snp req->rsvd0 = htobe32(snd_nxt); 187237263Snp else 188237263Snp req->rsvd0 = htobe32(tp->snd_nxt); 189239514Snp req->rsvd1 = !(toep->flags & TPF_TX_DATA_SENT); 190237263Snp req->cmd = CPL_ABORT_SEND_RST; 191237263Snp 192237263Snp /* 193237263Snp * XXX: What's the correct way to tell that the inp hasn't been detached 194237263Snp * from its socket? Should I even be flushing the snd buffer here? 195237263Snp */ 196237263Snp if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) { 197237263Snp struct socket *so = inp->inp_socket; 198237263Snp 199237263Snp if (so != NULL) /* because I'm not sure. See comment above */ 200237263Snp sbflush(&so->so_snd); 201237263Snp } 202237263Snp 203237263Snp t4_l2t_send(sc, wr, toep->l2te); 204237263Snp} 205237263Snp 206237263Snp/* 207237263Snp * Called when a connection is established to translate the TCP options 208237263Snp * reported by HW to FreeBSD's native format. 209237263Snp */ 210237263Snpstatic void 211237263Snpassign_rxopt(struct tcpcb *tp, unsigned int opt) 212237263Snp{ 213237263Snp struct toepcb *toep = tp->t_toe; 214276574Snp struct inpcb *inp = tp->t_inpcb; 215237263Snp struct adapter *sc = td_adapter(toep->td); 216276574Snp int n; 217237263Snp 218276574Snp INP_LOCK_ASSERT(inp); 219237263Snp 220276574Snp if (inp->inp_inc.inc_flags & INC_ISIPV6) 221276574Snp n = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 222276574Snp else 223276574Snp n = sizeof(struct ip) + sizeof(struct tcphdr); 224293284Sglebius tp->t_maxseg = sc->params.mtus[G_TCPOPT_MSS(opt)] - n; 225237263Snp 226276574Snp CTR4(KTR_CXGBE, "%s: tid %d, mtu_idx %u (%u)", __func__, toep->tid, 227276574Snp G_TCPOPT_MSS(opt), sc->params.mtus[G_TCPOPT_MSS(opt)]); 228276574Snp 229237263Snp if (G_TCPOPT_TSTAMP(opt)) { 230237263Snp tp->t_flags |= TF_RCVD_TSTMP; /* timestamps ok */ 231237263Snp tp->ts_recent = 0; /* hmmm */ 232237263Snp tp->ts_recent_age = tcp_ts_getticks(); 233237263Snp } 234237263Snp 235237263Snp if (G_TCPOPT_SACK(opt)) 236237263Snp tp->t_flags |= TF_SACK_PERMIT; /* should already be set */ 237237263Snp else 238237263Snp tp->t_flags &= ~TF_SACK_PERMIT; /* sack disallowed by peer */ 239237263Snp 240237263Snp if (G_TCPOPT_WSCALE_OK(opt)) 241237263Snp tp->t_flags |= TF_RCVD_SCALE; 242237263Snp 243237263Snp /* Doing window scaling? */ 244237263Snp if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 245237263Snp (TF_RCVD_SCALE | TF_REQ_SCALE)) { 246237263Snp tp->rcv_scale = tp->request_r_scale; 247237263Snp tp->snd_scale = G_TCPOPT_SND_WSCALE(opt); 248237263Snp } 249237263Snp} 250237263Snp 251237263Snp/* 252237263Snp * Completes some final bits of initialization for just established connections 253237263Snp * and changes their state to TCPS_ESTABLISHED. 254237263Snp * 255237263Snp * The ISNs are from after the exchange of SYNs. i.e., the true ISN + 1. 256237263Snp */ 257237263Snpvoid 258237263Snpmake_established(struct toepcb *toep, uint32_t snd_isn, uint32_t rcv_isn, 259237263Snp uint16_t opt) 260237263Snp{ 261237263Snp struct inpcb *inp = toep->inp; 262237263Snp struct socket *so = inp->inp_socket; 263237263Snp struct tcpcb *tp = intotcpcb(inp); 264237263Snp long bufsize; 265237263Snp uint32_t iss = be32toh(snd_isn) - 1; /* true ISS */ 266237263Snp uint32_t irs = be32toh(rcv_isn) - 1; /* true IRS */ 267237263Snp uint16_t tcpopt = be16toh(opt); 268237263Snp struct flowc_tx_params ftxp; 269237263Snp 270299206Sjhb CURVNET_SET(so->so_vnet); 271237263Snp INP_WLOCK_ASSERT(inp); 272237263Snp KASSERT(tp->t_state == TCPS_SYN_SENT || 273237263Snp tp->t_state == TCPS_SYN_RECEIVED, 274237263Snp ("%s: TCP state %s", __func__, tcpstates[tp->t_state])); 275237263Snp 276237263Snp CTR4(KTR_CXGBE, "%s: tid %d, toep %p, inp %p", 277237263Snp __func__, toep->tid, toep, inp); 278237263Snp 279237263Snp tp->t_state = TCPS_ESTABLISHED; 280237263Snp tp->t_starttime = ticks; 281237263Snp TCPSTAT_INC(tcps_connects); 282237263Snp 283237263Snp tp->irs = irs; 284237263Snp tcp_rcvseqinit(tp); 285237263Snp tp->rcv_wnd = toep->rx_credits << 10; 286237263Snp tp->rcv_adv += tp->rcv_wnd; 287237263Snp tp->last_ack_sent = tp->rcv_nxt; 288237263Snp 289237263Snp /* 290237263Snp * If we were unable to send all rx credits via opt0, save the remainder 291237263Snp * in rx_credits so that they can be handed over with the next credit 292237263Snp * update. 293237263Snp */ 294237263Snp SOCKBUF_LOCK(&so->so_rcv); 295237263Snp bufsize = select_rcv_wnd(so); 296237263Snp SOCKBUF_UNLOCK(&so->so_rcv); 297237263Snp toep->rx_credits = bufsize - tp->rcv_wnd; 298237263Snp 299237263Snp tp->iss = iss; 300237263Snp tcp_sendseqinit(tp); 301237263Snp tp->snd_una = iss + 1; 302237263Snp tp->snd_nxt = iss + 1; 303237263Snp tp->snd_max = iss + 1; 304237263Snp 305237263Snp assign_rxopt(tp, tcpopt); 306237263Snp 307237263Snp SOCKBUF_LOCK(&so->so_snd); 308237263Snp if (so->so_snd.sb_flags & SB_AUTOSIZE && V_tcp_do_autosndbuf) 309237263Snp bufsize = V_tcp_autosndbuf_max; 310237263Snp else 311237263Snp bufsize = sbspace(&so->so_snd); 312237263Snp SOCKBUF_UNLOCK(&so->so_snd); 313237263Snp 314237263Snp ftxp.snd_nxt = tp->snd_nxt; 315237263Snp ftxp.rcv_nxt = tp->rcv_nxt; 316237263Snp ftxp.snd_space = bufsize; 317237263Snp ftxp.mss = tp->t_maxseg; 318237263Snp send_flowc_wr(toep, &ftxp); 319237263Snp 320237263Snp soisconnected(so); 321299206Sjhb CURVNET_RESTORE(); 322237263Snp} 323237263Snp 324237263Snpstatic int 325239344Snpsend_rx_credits(struct adapter *sc, struct toepcb *toep, int credits) 326237263Snp{ 327237263Snp struct wrqe *wr; 328237263Snp struct cpl_rx_data_ack *req; 329237263Snp uint32_t dack = F_RX_DACK_CHANGE | V_RX_DACK_MODE(1); 330237263Snp 331239344Snp KASSERT(credits >= 0, ("%s: %d credits", __func__, credits)); 332239344Snp 333237263Snp wr = alloc_wrqe(sizeof(*req), toep->ctrlq); 334237263Snp if (wr == NULL) 335237263Snp return (0); 336237263Snp req = wrtod(wr); 337237263Snp 338237263Snp INIT_TP_WR_MIT_CPL(req, CPL_RX_DATA_ACK, toep->tid); 339237263Snp req->credit_dack = htobe32(dack | V_RX_CREDITS(credits)); 340237263Snp 341237263Snp t4_wrq_tx(sc, wr); 342237263Snp return (credits); 343237263Snp} 344237263Snp 345237263Snpvoid 346299210Sjhbt4_rcvd_locked(struct toedev *tod, struct tcpcb *tp) 347237263Snp{ 348237263Snp struct adapter *sc = tod->tod_softc; 349237263Snp struct inpcb *inp = tp->t_inpcb; 350237263Snp struct socket *so = inp->inp_socket; 351239344Snp struct sockbuf *sb = &so->so_rcv; 352237263Snp struct toepcb *toep = tp->t_toe; 353239344Snp int credits; 354237263Snp 355237263Snp INP_WLOCK_ASSERT(inp); 356237263Snp 357299210Sjhb SOCKBUF_LOCK_ASSERT(sb); 358274421Sglebius KASSERT(toep->sb_cc >= sbused(sb), 359239344Snp ("%s: sb %p has more data (%d) than last time (%d).", 360274421Sglebius __func__, sb, sbused(sb), toep->sb_cc)); 361292736Snp 362292736Snp toep->rx_credits += toep->sb_cc - sbused(sb); 363292736Snp toep->sb_cc = sbused(sb); 364292736Snp 365280878Snp if (toep->rx_credits > 0 && 366280878Snp (tp->rcv_wnd <= 32 * 1024 || toep->rx_credits >= 64 * 1024 || 367280878Snp (toep->rx_credits >= 16 * 1024 && tp->rcv_wnd <= 128 * 1024) || 368280878Snp toep->sb_cc + tp->rcv_wnd < sb->sb_lowat)) { 369237263Snp 370280878Snp credits = send_rx_credits(sc, toep, toep->rx_credits); 371237263Snp toep->rx_credits -= credits; 372237263Snp tp->rcv_wnd += credits; 373237263Snp tp->rcv_adv += credits; 374237263Snp } 375299210Sjhb} 376299210Sjhb 377299210Sjhbvoid 378299210Sjhbt4_rcvd(struct toedev *tod, struct tcpcb *tp) 379299210Sjhb{ 380299210Sjhb struct inpcb *inp = tp->t_inpcb; 381299210Sjhb struct socket *so = inp->inp_socket; 382299210Sjhb struct sockbuf *sb = &so->so_rcv; 383299210Sjhb 384299210Sjhb SOCKBUF_LOCK(sb); 385299210Sjhb t4_rcvd_locked(tod, tp); 386280878Snp SOCKBUF_UNLOCK(sb); 387237263Snp} 388237263Snp 389237263Snp/* 390237263Snp * Close a connection by sending a CPL_CLOSE_CON_REQ message. 391237263Snp */ 392237263Snpstatic int 393237263Snpclose_conn(struct adapter *sc, struct toepcb *toep) 394237263Snp{ 395237263Snp struct wrqe *wr; 396237263Snp struct cpl_close_con_req *req; 397237263Snp unsigned int tid = toep->tid; 398237263Snp 399237263Snp CTR3(KTR_CXGBE, "%s: tid %u%s", __func__, toep->tid, 400239514Snp toep->flags & TPF_FIN_SENT ? ", IGNORED" : ""); 401237263Snp 402239514Snp if (toep->flags & TPF_FIN_SENT) 403237263Snp return (0); 404237263Snp 405239514Snp KASSERT(toep->flags & TPF_FLOWC_WR_SENT, 406237263Snp ("%s: flowc_wr not sent for tid %u.", __func__, tid)); 407237263Snp 408237263Snp wr = alloc_wrqe(sizeof(*req), toep->ofld_txq); 409237263Snp if (wr == NULL) { 410237263Snp /* XXX */ 411237263Snp panic("%s: allocation failure.", __func__); 412237263Snp } 413237263Snp req = wrtod(wr); 414237263Snp 415237263Snp req->wr.wr_hi = htonl(V_FW_WR_OP(FW_TP_WR) | 416237263Snp V_FW_WR_IMMDLEN(sizeof(*req) - sizeof(req->wr))); 417237263Snp req->wr.wr_mid = htonl(V_FW_WR_LEN16(howmany(sizeof(*req), 16)) | 418237263Snp V_FW_WR_FLOWID(tid)); 419237263Snp req->wr.wr_lo = cpu_to_be64(0); 420237263Snp OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_CLOSE_CON_REQ, tid)); 421237263Snp req->rsvd = 0; 422237263Snp 423239514Snp toep->flags |= TPF_FIN_SENT; 424239514Snp toep->flags &= ~TPF_SEND_FIN; 425237263Snp t4_l2t_send(sc, wr, toep->l2te); 426237263Snp 427237263Snp return (0); 428237263Snp} 429237263Snp 430237263Snp#define MAX_OFLD_TX_CREDITS (SGE_MAX_WR_LEN / 16) 431237263Snp#define MIN_OFLD_TX_CREDITS (howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16)) 432237263Snp 433237263Snp/* Maximum amount of immediate data we could stuff in a WR */ 434237263Snpstatic inline int 435237263Snpmax_imm_payload(int tx_credits) 436237263Snp{ 437237263Snp const int n = 2; /* Use only up to 2 desc for imm. data WR */ 438237263Snp 439237263Snp KASSERT(tx_credits >= 0 && 440237263Snp tx_credits <= MAX_OFLD_TX_CREDITS, 441237263Snp ("%s: %d credits", __func__, tx_credits)); 442237263Snp 443237263Snp if (tx_credits < MIN_OFLD_TX_CREDITS) 444237263Snp return (0); 445237263Snp 446237263Snp if (tx_credits >= (n * EQ_ESIZE) / 16) 447237263Snp return ((n * EQ_ESIZE) - sizeof(struct fw_ofld_tx_data_wr)); 448237263Snp else 449237263Snp return (tx_credits * 16 - sizeof(struct fw_ofld_tx_data_wr)); 450237263Snp} 451237263Snp 452237263Snp/* Maximum number of SGL entries we could stuff in a WR */ 453237263Snpstatic inline int 454237263Snpmax_dsgl_nsegs(int tx_credits) 455237263Snp{ 456237263Snp int nseg = 1; /* ulptx_sgl has room for 1, rest ulp_tx_sge_pair */ 457237263Snp int sge_pair_credits = tx_credits - MIN_OFLD_TX_CREDITS; 458237263Snp 459237263Snp KASSERT(tx_credits >= 0 && 460237263Snp tx_credits <= MAX_OFLD_TX_CREDITS, 461237263Snp ("%s: %d credits", __func__, tx_credits)); 462237263Snp 463237263Snp if (tx_credits < MIN_OFLD_TX_CREDITS) 464237263Snp return (0); 465237263Snp 466237263Snp nseg += 2 * (sge_pair_credits * 16 / 24); 467237263Snp if ((sge_pair_credits * 16) % 24 == 16) 468237263Snp nseg++; 469237263Snp 470237263Snp return (nseg); 471237263Snp} 472237263Snp 473237263Snpstatic inline void 474237263Snpwrite_tx_wr(void *dst, struct toepcb *toep, unsigned int immdlen, 475292736Snp unsigned int plen, uint8_t credits, int shove, int ulp_submode, int txalign) 476237263Snp{ 477237263Snp struct fw_ofld_tx_data_wr *txwr = dst; 478237263Snp 479237263Snp txwr->op_to_immdlen = htobe32(V_WR_OP(FW_OFLD_TX_DATA_WR) | 480255411Snp V_FW_WR_IMMDLEN(immdlen)); 481237263Snp txwr->flowid_len16 = htobe32(V_FW_WR_FLOWID(toep->tid) | 482237263Snp V_FW_WR_LEN16(credits)); 483292736Snp txwr->lsodisable_to_flags = htobe32(V_TX_ULP_MODE(toep->ulp_mode) | 484292736Snp V_TX_ULP_SUBMODE(ulp_submode) | V_TX_URG(0) | V_TX_SHOVE(shove)); 485237263Snp txwr->plen = htobe32(plen); 486276597Snp 487276597Snp if (txalign > 0) { 488276597Snp struct tcpcb *tp = intotcpcb(toep->inp); 489276597Snp 490291665Sjhb if (plen < 2 * tp->t_maxseg || is_10G_port(toep->vi->pi)) 491285527Snp txwr->lsodisable_to_flags |= 492276597Snp htobe32(F_FW_OFLD_TX_DATA_WR_LSODISABLE); 493276597Snp else 494285527Snp txwr->lsodisable_to_flags |= 495276597Snp htobe32(F_FW_OFLD_TX_DATA_WR_ALIGNPLD | 496276597Snp (tp->t_flags & TF_NODELAY ? 0 : 497276597Snp F_FW_OFLD_TX_DATA_WR_ALIGNPLDSHOVE)); 498276597Snp } 499237263Snp} 500237263Snp 501237263Snp/* 502237263Snp * Generate a DSGL from a starting mbuf. The total number of segments and the 503237263Snp * maximum segments in any one mbuf are provided. 504237263Snp */ 505237263Snpstatic void 506237263Snpwrite_tx_sgl(void *dst, struct mbuf *start, struct mbuf *stop, int nsegs, int n) 507237263Snp{ 508237263Snp struct mbuf *m; 509237263Snp struct ulptx_sgl *usgl = dst; 510237263Snp int i, j, rc; 511237263Snp struct sglist sg; 512237263Snp struct sglist_seg segs[n]; 513237263Snp 514237263Snp KASSERT(nsegs > 0, ("%s: nsegs 0", __func__)); 515237263Snp 516237263Snp sglist_init(&sg, n, segs); 517237263Snp usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) | 518237263Snp V_ULPTX_NSGE(nsegs)); 519237263Snp 520237263Snp i = -1; 521237263Snp for (m = start; m != stop; m = m->m_next) { 522237263Snp rc = sglist_append(&sg, mtod(m, void *), m->m_len); 523237263Snp if (__predict_false(rc != 0)) 524237263Snp panic("%s: sglist_append %d", __func__, rc); 525237263Snp 526237263Snp for (j = 0; j < sg.sg_nseg; i++, j++) { 527237263Snp if (i < 0) { 528237263Snp usgl->len0 = htobe32(segs[j].ss_len); 529237263Snp usgl->addr0 = htobe64(segs[j].ss_paddr); 530237263Snp } else { 531237263Snp usgl->sge[i / 2].len[i & 1] = 532237263Snp htobe32(segs[j].ss_len); 533237263Snp usgl->sge[i / 2].addr[i & 1] = 534237263Snp htobe64(segs[j].ss_paddr); 535237263Snp } 536237263Snp#ifdef INVARIANTS 537237263Snp nsegs--; 538237263Snp#endif 539237263Snp } 540237263Snp sglist_reset(&sg); 541237263Snp } 542237263Snp if (i & 1) 543237263Snp usgl->sge[i / 2].len[1] = htobe32(0); 544237263Snp KASSERT(nsegs == 0, ("%s: nsegs %d, start %p, stop %p", 545237263Snp __func__, nsegs, start, stop)); 546237263Snp} 547237263Snp 548237263Snp/* 549237263Snp * Max number of SGL entries an offload tx work request can have. This is 41 550237263Snp * (1 + 40) for a full 512B work request. 551237263Snp * fw_ofld_tx_data_wr(16B) + ulptx_sgl(16B, 1) + ulptx_sge_pair(480B, 40) 552237263Snp */ 553237263Snp#define OFLD_SGL_LEN (41) 554237263Snp 555237263Snp/* 556237263Snp * Send data and/or a FIN to the peer. 557237263Snp * 558237263Snp * The socket's so_snd buffer consists of a stream of data starting with sb_mb 559237263Snp * and linked together with m_next. sb_sndptr, if set, is the last mbuf that 560237263Snp * was transmitted. 561255411Snp * 562255411Snp * drop indicates the number of bytes that should be dropped from the head of 563255411Snp * the send buffer. It is an optimization that lets do_fw4_ack avoid creating 564255411Snp * contention on the send buffer lock (before this change it used to do 565255411Snp * sowwakeup and then t4_push_frames right after that when recovering from tx 566255411Snp * stalls). When drop is set this function MUST drop the bytes and wake up any 567255411Snp * writers. 568237263Snp */ 569269076Snpvoid 570255411Snpt4_push_frames(struct adapter *sc, struct toepcb *toep, int drop) 571237263Snp{ 572237263Snp struct mbuf *sndptr, *m, *sb_sndptr; 573237263Snp struct fw_ofld_tx_data_wr *txwr; 574237263Snp struct wrqe *wr; 575255411Snp u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf; 576237263Snp struct inpcb *inp = toep->inp; 577237263Snp struct tcpcb *tp = intotcpcb(inp); 578237263Snp struct socket *so = inp->inp_socket; 579237263Snp struct sockbuf *sb = &so->so_snd; 580255411Snp int tx_credits, shove, compl, space, sowwakeup; 581237263Snp struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx]; 582237263Snp 583237263Snp INP_WLOCK_ASSERT(inp); 584239514Snp KASSERT(toep->flags & TPF_FLOWC_WR_SENT, 585237263Snp ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid)); 586237263Snp 587255005Snp KASSERT(toep->ulp_mode == ULP_MODE_NONE || 588255005Snp toep->ulp_mode == ULP_MODE_TCPDDP || 589255005Snp toep->ulp_mode == ULP_MODE_RDMA, 590255005Snp ("%s: ulp_mode %u for toep %p", __func__, toep->ulp_mode, toep)); 591237263Snp 592292736Snp if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN)) 593292736Snp return; 594292736Snp 595237263Snp /* 596237263Snp * This function doesn't resume by itself. Someone else must clear the 597237263Snp * flag and call this function. 598237263Snp */ 599255411Snp if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) { 600255411Snp KASSERT(drop == 0, 601255411Snp ("%s: drop (%d) != 0 but tx is suspended", __func__, drop)); 602237263Snp return; 603255411Snp } 604237263Snp 605237263Snp do { 606237263Snp tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS); 607237263Snp max_imm = max_imm_payload(tx_credits); 608237263Snp max_nsegs = max_dsgl_nsegs(tx_credits); 609237263Snp 610237263Snp SOCKBUF_LOCK(sb); 611255411Snp sowwakeup = drop; 612255411Snp if (drop) { 613255411Snp sbdrop_locked(sb, drop); 614255411Snp drop = 0; 615255411Snp } 616237263Snp sb_sndptr = sb->sb_sndptr; 617237263Snp sndptr = sb_sndptr ? sb_sndptr->m_next : sb->sb_mb; 618237263Snp plen = 0; 619237263Snp nsegs = 0; 620237263Snp max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */ 621237263Snp for (m = sndptr; m != NULL; m = m->m_next) { 622237263Snp int n = sglist_count(mtod(m, void *), m->m_len); 623237263Snp 624237263Snp nsegs += n; 625237263Snp plen += m->m_len; 626237263Snp 627237263Snp /* This mbuf sent us _over_ the nsegs limit, back out */ 628237263Snp if (plen > max_imm && nsegs > max_nsegs) { 629237263Snp nsegs -= n; 630237263Snp plen -= m->m_len; 631237263Snp if (plen == 0) { 632237263Snp /* Too few credits */ 633239514Snp toep->flags |= TPF_TX_SUSPENDED; 634255411Snp if (sowwakeup) 635255411Snp sowwakeup_locked(so); 636255411Snp else 637255411Snp SOCKBUF_UNLOCK(sb); 638255411Snp SOCKBUF_UNLOCK_ASSERT(sb); 639237263Snp return; 640237263Snp } 641237263Snp break; 642237263Snp } 643237263Snp 644237263Snp if (max_nsegs_1mbuf < n) 645237263Snp max_nsegs_1mbuf = n; 646237263Snp sb_sndptr = m; /* new sb->sb_sndptr if all goes well */ 647237263Snp 648237263Snp /* This mbuf put us right at the max_nsegs limit */ 649237263Snp if (plen > max_imm && nsegs == max_nsegs) { 650237263Snp m = m->m_next; 651237263Snp break; 652237263Snp } 653237263Snp } 654237263Snp 655255411Snp space = sbspace(sb); 656255411Snp 657255411Snp if (space <= sb->sb_hiwat * 3 / 8 && 658255411Snp toep->plen_nocompl + plen >= sb->sb_hiwat / 4) 659255411Snp compl = 1; 660255411Snp else 661255411Snp compl = 0; 662255411Snp 663237263Snp if (sb->sb_flags & SB_AUTOSIZE && 664237263Snp V_tcp_do_autosndbuf && 665237263Snp sb->sb_hiwat < V_tcp_autosndbuf_max && 666255411Snp space < sb->sb_hiwat / 8) { 667237263Snp int newsize = min(sb->sb_hiwat + V_tcp_autosndbuf_inc, 668237263Snp V_tcp_autosndbuf_max); 669237263Snp 670237263Snp if (!sbreserve_locked(sb, newsize, so, NULL)) 671237263Snp sb->sb_flags &= ~SB_AUTOSIZE; 672255411Snp else 673255411Snp sowwakeup = 1; /* room available */ 674237263Snp } 675255411Snp if (sowwakeup) 676255411Snp sowwakeup_locked(so); 677255411Snp else 678255411Snp SOCKBUF_UNLOCK(sb); 679255411Snp SOCKBUF_UNLOCK_ASSERT(sb); 680237263Snp 681237263Snp /* nothing to send */ 682237263Snp if (plen == 0) { 683237263Snp KASSERT(m == NULL, 684237263Snp ("%s: nothing to send, but m != NULL", __func__)); 685237263Snp break; 686237263Snp } 687237263Snp 688239514Snp if (__predict_false(toep->flags & TPF_FIN_SENT)) 689237263Snp panic("%s: excess tx.", __func__); 690237263Snp 691290175Snp shove = m == NULL && !(tp->t_flags & TF_MORETOCOME); 692237263Snp if (plen <= max_imm) { 693237263Snp 694237263Snp /* Immediate data tx */ 695237263Snp 696248925Snp wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16), 697237263Snp toep->ofld_txq); 698237263Snp if (wr == NULL) { 699237263Snp /* XXX: how will we recover from this? */ 700239514Snp toep->flags |= TPF_TX_SUSPENDED; 701237263Snp return; 702237263Snp } 703237263Snp txwr = wrtod(wr); 704237263Snp credits = howmany(wr->wr_len, 16); 705276597Snp write_tx_wr(txwr, toep, plen, plen, credits, shove, 0, 706276597Snp sc->tt.tx_align); 707237263Snp m_copydata(sndptr, 0, plen, (void *)(txwr + 1)); 708255411Snp nsegs = 0; 709237263Snp } else { 710237263Snp int wr_len; 711237263Snp 712237263Snp /* DSGL tx */ 713237263Snp 714237263Snp wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) + 715237263Snp ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8; 716248925Snp wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq); 717237263Snp if (wr == NULL) { 718237263Snp /* XXX: how will we recover from this? */ 719239514Snp toep->flags |= TPF_TX_SUSPENDED; 720237263Snp return; 721237263Snp } 722237263Snp txwr = wrtod(wr); 723237263Snp credits = howmany(wr_len, 16); 724276597Snp write_tx_wr(txwr, toep, 0, plen, credits, shove, 0, 725276597Snp sc->tt.tx_align); 726237263Snp write_tx_sgl(txwr + 1, sndptr, m, nsegs, 727237263Snp max_nsegs_1mbuf); 728237263Snp if (wr_len & 0xf) { 729237263Snp uint64_t *pad = (uint64_t *) 730237263Snp ((uintptr_t)txwr + wr_len); 731237263Snp *pad = 0; 732237263Snp } 733237263Snp } 734237263Snp 735237263Snp KASSERT(toep->tx_credits >= credits, 736237263Snp ("%s: not enough credits", __func__)); 737237263Snp 738237263Snp toep->tx_credits -= credits; 739255411Snp toep->tx_nocompl += credits; 740255411Snp toep->plen_nocompl += plen; 741255411Snp if (toep->tx_credits <= toep->tx_total * 3 / 8 && 742255411Snp toep->tx_nocompl >= toep->tx_total / 4) 743255411Snp compl = 1; 744237263Snp 745273797Snp if (compl || toep->ulp_mode == ULP_MODE_RDMA) { 746255411Snp txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL); 747255411Snp toep->tx_nocompl = 0; 748255411Snp toep->plen_nocompl = 0; 749255411Snp } 750255411Snp 751237263Snp tp->snd_nxt += plen; 752237263Snp tp->snd_max += plen; 753237263Snp 754237263Snp SOCKBUF_LOCK(sb); 755237263Snp KASSERT(sb_sndptr, ("%s: sb_sndptr is NULL", __func__)); 756237263Snp sb->sb_sndptr = sb_sndptr; 757237263Snp SOCKBUF_UNLOCK(sb); 758237263Snp 759239514Snp toep->flags |= TPF_TX_DATA_SENT; 760255411Snp if (toep->tx_credits < MIN_OFLD_TX_CREDITS) 761255411Snp toep->flags |= TPF_TX_SUSPENDED; 762237263Snp 763237263Snp KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__)); 764237263Snp txsd->plen = plen; 765237263Snp txsd->tx_credits = credits; 766237263Snp txsd++; 767237263Snp if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) { 768237263Snp toep->txsd_pidx = 0; 769237263Snp txsd = &toep->txsd[0]; 770237263Snp } 771237263Snp toep->txsd_avail--; 772237263Snp 773237263Snp t4_l2t_send(sc, wr, toep->l2te); 774237263Snp } while (m != NULL); 775237263Snp 776237263Snp /* Send a FIN if requested, but only if there's no more data to send */ 777239514Snp if (m == NULL && toep->flags & TPF_SEND_FIN) 778237263Snp close_conn(sc, toep); 779237263Snp} 780237263Snp 781292736Snpstatic inline void 782292736Snprqdrop_locked(struct mbufq *q, int plen) 783292736Snp{ 784292736Snp struct mbuf *m; 785292736Snp 786292736Snp while (plen > 0) { 787292736Snp m = mbufq_dequeue(q); 788292736Snp 789292736Snp /* Too many credits. */ 790292736Snp MPASS(m != NULL); 791292736Snp M_ASSERTPKTHDR(m); 792292736Snp 793292736Snp /* Partial credits. */ 794292736Snp MPASS(plen >= m->m_pkthdr.len); 795292736Snp 796292736Snp plen -= m->m_pkthdr.len; 797292736Snp m_freem(m); 798292736Snp } 799292736Snp} 800292736Snp 801269076Snpvoid 802292736Snpt4_push_pdus(struct adapter *sc, struct toepcb *toep, int drop) 803269076Snp{ 804292736Snp struct mbuf *sndptr, *m; 805269076Snp struct fw_ofld_tx_data_wr *txwr; 806269076Snp struct wrqe *wr; 807292736Snp u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf; 808292736Snp u_int adjusted_plen, ulp_submode; 809269076Snp struct inpcb *inp = toep->inp; 810292736Snp struct tcpcb *tp = intotcpcb(inp); 811292736Snp int tx_credits, shove; 812292736Snp struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx]; 813292736Snp struct mbufq *pduq = &toep->ulp_pduq; 814292736Snp static const u_int ulp_extra_len[] = {0, 4, 4, 8}; 815269076Snp 816269076Snp INP_WLOCK_ASSERT(inp); 817269076Snp KASSERT(toep->flags & TPF_FLOWC_WR_SENT, 818269076Snp ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid)); 819292736Snp KASSERT(toep->ulp_mode == ULP_MODE_ISCSI, 820292736Snp ("%s: ulp_mode %u for toep %p", __func__, toep->ulp_mode, toep)); 821269076Snp 822292736Snp if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN)) 823292736Snp return; 824292736Snp 825269076Snp /* 826269076Snp * This function doesn't resume by itself. Someone else must clear the 827269076Snp * flag and call this function. 828269076Snp */ 829292736Snp if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) { 830292736Snp KASSERT(drop == 0, 831292736Snp ("%s: drop (%d) != 0 but tx is suspended", __func__, drop)); 832269076Snp return; 833292736Snp } 834269076Snp 835292736Snp if (drop) 836292736Snp rqdrop_locked(&toep->ulp_pdu_reclaimq, drop); 837269076Snp 838292736Snp while ((sndptr = mbufq_first(pduq)) != NULL) { 839292736Snp M_ASSERTPKTHDR(sndptr); 840292736Snp 841269076Snp tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS); 842269076Snp max_imm = max_imm_payload(tx_credits); 843269076Snp max_nsegs = max_dsgl_nsegs(tx_credits); 844269076Snp 845269076Snp plen = 0; 846269076Snp nsegs = 0; 847269076Snp max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */ 848269076Snp for (m = sndptr; m != NULL; m = m->m_next) { 849269076Snp int n = sglist_count(mtod(m, void *), m->m_len); 850269076Snp 851269076Snp nsegs += n; 852269076Snp plen += m->m_len; 853269076Snp 854292736Snp /* 855292736Snp * This mbuf would send us _over_ the nsegs limit. 856292736Snp * Suspend tx because the PDU can't be sent out. 857292736Snp */ 858269076Snp if (plen > max_imm && nsegs > max_nsegs) { 859269076Snp toep->flags |= TPF_TX_SUSPENDED; 860269076Snp return; 861269076Snp } 862269076Snp 863269076Snp if (max_nsegs_1mbuf < n) 864269076Snp max_nsegs_1mbuf = n; 865269076Snp } 866269076Snp 867269076Snp if (__predict_false(toep->flags & TPF_FIN_SENT)) 868269076Snp panic("%s: excess tx.", __func__); 869269076Snp 870292736Snp /* 871292736Snp * We have a PDU to send. All of it goes out in one WR so 'm' 872292736Snp * is NULL. A PDU's length is always a multiple of 4. 873292736Snp */ 874292736Snp MPASS(m == NULL); 875292736Snp MPASS((plen & 3) == 0); 876292736Snp MPASS(sndptr->m_pkthdr.len == plen); 877292736Snp 878292736Snp shove = !(tp->t_flags & TF_MORETOCOME); 879292736Snp ulp_submode = mbuf_ulp_submode(sndptr); 880292736Snp MPASS(ulp_submode < nitems(ulp_extra_len)); 881292736Snp 882292736Snp /* 883292736Snp * plen doesn't include header and data digests, which are 884292736Snp * generated and inserted in the right places by the TOE, but 885292736Snp * they do occupy TCP sequence space and need to be accounted 886292736Snp * for. 887292736Snp */ 888292736Snp adjusted_plen = plen + ulp_extra_len[ulp_submode]; 889269076Snp if (plen <= max_imm) { 890269076Snp 891269076Snp /* Immediate data tx */ 892292736Snp 893292736Snp wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16), 894269076Snp toep->ofld_txq); 895269076Snp if (wr == NULL) { 896269076Snp /* XXX: how will we recover from this? */ 897269076Snp toep->flags |= TPF_TX_SUSPENDED; 898269076Snp return; 899269076Snp } 900269076Snp txwr = wrtod(wr); 901269076Snp credits = howmany(wr->wr_len, 16); 902292736Snp write_tx_wr(txwr, toep, plen, adjusted_plen, credits, 903292736Snp shove, ulp_submode, sc->tt.tx_align); 904269076Snp m_copydata(sndptr, 0, plen, (void *)(txwr + 1)); 905292736Snp nsegs = 0; 906269076Snp } else { 907269076Snp int wr_len; 908269076Snp 909269076Snp /* DSGL tx */ 910269076Snp wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) + 911269076Snp ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8; 912292736Snp wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq); 913269076Snp if (wr == NULL) { 914269076Snp /* XXX: how will we recover from this? */ 915269076Snp toep->flags |= TPF_TX_SUSPENDED; 916269076Snp return; 917269076Snp } 918269076Snp txwr = wrtod(wr); 919269076Snp credits = howmany(wr_len, 16); 920292736Snp write_tx_wr(txwr, toep, 0, adjusted_plen, credits, 921292736Snp shove, ulp_submode, sc->tt.tx_align); 922269076Snp write_tx_sgl(txwr + 1, sndptr, m, nsegs, 923269076Snp max_nsegs_1mbuf); 924269076Snp if (wr_len & 0xf) { 925269076Snp uint64_t *pad = (uint64_t *) 926269076Snp ((uintptr_t)txwr + wr_len); 927269076Snp *pad = 0; 928269076Snp } 929269076Snp } 930269076Snp 931269076Snp KASSERT(toep->tx_credits >= credits, 932269076Snp ("%s: not enough credits", __func__)); 933269076Snp 934292736Snp m = mbufq_dequeue(pduq); 935292736Snp MPASS(m == sndptr); 936292736Snp mbufq_enqueue(&toep->ulp_pdu_reclaimq, m); 937292736Snp 938269076Snp toep->tx_credits -= credits; 939269076Snp toep->tx_nocompl += credits; 940269076Snp toep->plen_nocompl += plen; 941269076Snp if (toep->tx_credits <= toep->tx_total * 3 / 8 && 942292736Snp toep->tx_nocompl >= toep->tx_total / 4) { 943269076Snp txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL); 944269076Snp toep->tx_nocompl = 0; 945269076Snp toep->plen_nocompl = 0; 946269076Snp } 947269076Snp 948292736Snp tp->snd_nxt += adjusted_plen; 949292736Snp tp->snd_max += adjusted_plen; 950269076Snp 951269076Snp toep->flags |= TPF_TX_DATA_SENT; 952292736Snp if (toep->tx_credits < MIN_OFLD_TX_CREDITS) 953269076Snp toep->flags |= TPF_TX_SUSPENDED; 954269076Snp 955269076Snp KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__)); 956269076Snp txsd->plen = plen; 957269076Snp txsd->tx_credits = credits; 958269076Snp txsd++; 959269076Snp if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) { 960269076Snp toep->txsd_pidx = 0; 961269076Snp txsd = &toep->txsd[0]; 962269076Snp } 963269076Snp toep->txsd_avail--; 964269076Snp 965269076Snp t4_l2t_send(sc, wr, toep->l2te); 966292736Snp } 967269076Snp 968292736Snp /* Send a FIN if requested, but only if there are no more PDUs to send */ 969292736Snp if (mbufq_first(pduq) == NULL && toep->flags & TPF_SEND_FIN) 970269076Snp close_conn(sc, toep); 971269076Snp} 972269076Snp 973237263Snpint 974237263Snpt4_tod_output(struct toedev *tod, struct tcpcb *tp) 975237263Snp{ 976237263Snp struct adapter *sc = tod->tod_softc; 977237263Snp#ifdef INVARIANTS 978237263Snp struct inpcb *inp = tp->t_inpcb; 979237263Snp#endif 980237263Snp struct toepcb *toep = tp->t_toe; 981237263Snp 982237263Snp INP_WLOCK_ASSERT(inp); 983237263Snp KASSERT((inp->inp_flags & INP_DROPPED) == 0, 984237263Snp ("%s: inp %p dropped.", __func__, inp)); 985237263Snp KASSERT(toep != NULL, ("%s: toep is NULL", __func__)); 986237263Snp 987292736Snp if (toep->ulp_mode == ULP_MODE_ISCSI) 988292736Snp t4_push_pdus(sc, toep, 0); 989292736Snp else 990292736Snp t4_push_frames(sc, toep, 0); 991237263Snp 992237263Snp return (0); 993237263Snp} 994237263Snp 995237263Snpint 996237263Snpt4_send_fin(struct toedev *tod, struct tcpcb *tp) 997237263Snp{ 998237263Snp struct adapter *sc = tod->tod_softc; 999237263Snp#ifdef INVARIANTS 1000237263Snp struct inpcb *inp = tp->t_inpcb; 1001237263Snp#endif 1002237263Snp struct toepcb *toep = tp->t_toe; 1003237263Snp 1004237263Snp INP_WLOCK_ASSERT(inp); 1005237263Snp KASSERT((inp->inp_flags & INP_DROPPED) == 0, 1006237263Snp ("%s: inp %p dropped.", __func__, inp)); 1007237263Snp KASSERT(toep != NULL, ("%s: toep is NULL", __func__)); 1008237263Snp 1009239514Snp toep->flags |= TPF_SEND_FIN; 1010269076Snp if (tp->t_state >= TCPS_ESTABLISHED) { 1011269076Snp if (toep->ulp_mode == ULP_MODE_ISCSI) 1012292736Snp t4_push_pdus(sc, toep, 0); 1013269076Snp else 1014269076Snp t4_push_frames(sc, toep, 0); 1015269076Snp } 1016237263Snp 1017237263Snp return (0); 1018237263Snp} 1019237263Snp 1020237263Snpint 1021237263Snpt4_send_rst(struct toedev *tod, struct tcpcb *tp) 1022237263Snp{ 1023237263Snp struct adapter *sc = tod->tod_softc; 1024237263Snp#if defined(INVARIANTS) 1025237263Snp struct inpcb *inp = tp->t_inpcb; 1026237263Snp#endif 1027237263Snp struct toepcb *toep = tp->t_toe; 1028237263Snp 1029237263Snp INP_WLOCK_ASSERT(inp); 1030237263Snp KASSERT((inp->inp_flags & INP_DROPPED) == 0, 1031237263Snp ("%s: inp %p dropped.", __func__, inp)); 1032237263Snp KASSERT(toep != NULL, ("%s: toep is NULL", __func__)); 1033237263Snp 1034237263Snp /* hmmmm */ 1035239514Snp KASSERT(toep->flags & TPF_FLOWC_WR_SENT, 1036237263Snp ("%s: flowc for tid %u [%s] not sent already", 1037237263Snp __func__, toep->tid, tcpstates[tp->t_state])); 1038237263Snp 1039237263Snp send_reset(sc, toep, 0); 1040237263Snp return (0); 1041237263Snp} 1042237263Snp 1043237263Snp/* 1044237263Snp * Peer has sent us a FIN. 1045237263Snp */ 1046237263Snpstatic int 1047237263Snpdo_peer_close(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) 1048237263Snp{ 1049237263Snp struct adapter *sc = iq->adapter; 1050237263Snp const struct cpl_peer_close *cpl = (const void *)(rss + 1); 1051237263Snp unsigned int tid = GET_TID(cpl); 1052237263Snp struct toepcb *toep = lookup_tid(sc, tid); 1053237263Snp struct inpcb *inp = toep->inp; 1054237263Snp struct tcpcb *tp = NULL; 1055239344Snp struct socket *so; 1056237263Snp#ifdef INVARIANTS 1057237263Snp unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); 1058237263Snp#endif 1059237263Snp 1060237263Snp KASSERT(opcode == CPL_PEER_CLOSE, 1061237263Snp ("%s: unexpected opcode 0x%x", __func__, opcode)); 1062237263Snp KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 1063243680Snp 1064243680Snp if (__predict_false(toep->flags & TPF_SYNQE)) { 1065243680Snp#ifdef INVARIANTS 1066243680Snp struct synq_entry *synqe = (void *)toep; 1067243680Snp 1068243680Snp INP_WLOCK(synqe->lctx->inp); 1069243680Snp if (synqe->flags & TPF_SYNQE_HAS_L2TE) { 1070243680Snp KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN, 1071243680Snp ("%s: listen socket closed but tid %u not aborted.", 1072243680Snp __func__, tid)); 1073243680Snp } else { 1074243680Snp /* 1075243680Snp * do_pass_accept_req is still running and will 1076243680Snp * eventually take care of this tid. 1077243680Snp */ 1078243680Snp } 1079243680Snp INP_WUNLOCK(synqe->lctx->inp); 1080243680Snp#endif 1081243680Snp CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid, 1082243680Snp toep, toep->flags); 1083243680Snp return (0); 1084243680Snp } 1085243680Snp 1086237263Snp KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__)); 1087237263Snp 1088286227Sjch INP_INFO_RLOCK(&V_tcbinfo); 1089237263Snp INP_WLOCK(inp); 1090237263Snp tp = intotcpcb(inp); 1091237263Snp 1092237263Snp CTR5(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x, inp %p", __func__, 1093237263Snp tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags, inp); 1094237263Snp 1095239514Snp if (toep->flags & TPF_ABORT_SHUTDOWN) 1096237263Snp goto done; 1097237263Snp 1098239344Snp tp->rcv_nxt++; /* FIN */ 1099239344Snp 1100237263Snp so = inp->inp_socket; 1101299210Sjhb if (toep->ulp_mode == ULP_MODE_TCPDDP) { 1102299210Sjhb DDP_LOCK(toep); 1103299210Sjhb if (__predict_false(toep->ddp_flags & 1104299210Sjhb (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE))) 1105299210Sjhb handle_ddp_close(toep, tp, cpl->rcv_nxt); 1106299210Sjhb DDP_UNLOCK(toep); 1107239344Snp } 1108299210Sjhb socantrcvmore(so); 1109239344Snp 1110255005Snp if (toep->ulp_mode != ULP_MODE_RDMA) { 1111255005Snp KASSERT(tp->rcv_nxt == be32toh(cpl->rcv_nxt), 1112255005Snp ("%s: rcv_nxt mismatch: %u %u", __func__, tp->rcv_nxt, 1113255005Snp be32toh(cpl->rcv_nxt))); 1114255005Snp } 1115237263Snp 1116237263Snp switch (tp->t_state) { 1117237263Snp case TCPS_SYN_RECEIVED: 1118237263Snp tp->t_starttime = ticks; 1119237263Snp /* FALLTHROUGH */ 1120237263Snp 1121237263Snp case TCPS_ESTABLISHED: 1122237263Snp tp->t_state = TCPS_CLOSE_WAIT; 1123237263Snp break; 1124237263Snp 1125237263Snp case TCPS_FIN_WAIT_1: 1126237263Snp tp->t_state = TCPS_CLOSING; 1127237263Snp break; 1128237263Snp 1129237263Snp case TCPS_FIN_WAIT_2: 1130237263Snp tcp_twstart(tp); 1131237263Snp INP_UNLOCK_ASSERT(inp); /* safe, we have a ref on the inp */ 1132286227Sjch INP_INFO_RUNLOCK(&V_tcbinfo); 1133237263Snp 1134237263Snp INP_WLOCK(inp); 1135237263Snp final_cpl_received(toep); 1136237263Snp return (0); 1137237263Snp 1138237263Snp default: 1139237263Snp log(LOG_ERR, "%s: TID %u received CPL_PEER_CLOSE in state %d\n", 1140237263Snp __func__, tid, tp->t_state); 1141237263Snp } 1142237263Snpdone: 1143237263Snp INP_WUNLOCK(inp); 1144286227Sjch INP_INFO_RUNLOCK(&V_tcbinfo); 1145237263Snp return (0); 1146237263Snp} 1147237263Snp 1148237263Snp/* 1149237263Snp * Peer has ACK'd our FIN. 1150237263Snp */ 1151237263Snpstatic int 1152237263Snpdo_close_con_rpl(struct sge_iq *iq, const struct rss_header *rss, 1153237263Snp struct mbuf *m) 1154237263Snp{ 1155237263Snp struct adapter *sc = iq->adapter; 1156237263Snp const struct cpl_close_con_rpl *cpl = (const void *)(rss + 1); 1157237263Snp unsigned int tid = GET_TID(cpl); 1158237263Snp struct toepcb *toep = lookup_tid(sc, tid); 1159237263Snp struct inpcb *inp = toep->inp; 1160237263Snp struct tcpcb *tp = NULL; 1161237263Snp struct socket *so = NULL; 1162237263Snp#ifdef INVARIANTS 1163237263Snp unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); 1164237263Snp#endif 1165237263Snp 1166237263Snp KASSERT(opcode == CPL_CLOSE_CON_RPL, 1167237263Snp ("%s: unexpected opcode 0x%x", __func__, opcode)); 1168237263Snp KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 1169237263Snp KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__)); 1170237263Snp 1171286227Sjch INP_INFO_RLOCK(&V_tcbinfo); 1172237263Snp INP_WLOCK(inp); 1173237263Snp tp = intotcpcb(inp); 1174237263Snp 1175237263Snp CTR4(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x", 1176237263Snp __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags); 1177237263Snp 1178239514Snp if (toep->flags & TPF_ABORT_SHUTDOWN) 1179237263Snp goto done; 1180237263Snp 1181237263Snp so = inp->inp_socket; 1182237263Snp tp->snd_una = be32toh(cpl->snd_nxt) - 1; /* exclude FIN */ 1183237263Snp 1184237263Snp switch (tp->t_state) { 1185237263Snp case TCPS_CLOSING: /* see TCPS_FIN_WAIT_2 in do_peer_close too */ 1186237263Snp tcp_twstart(tp); 1187237263Snprelease: 1188237263Snp INP_UNLOCK_ASSERT(inp); /* safe, we have a ref on the inp */ 1189286227Sjch INP_INFO_RUNLOCK(&V_tcbinfo); 1190237263Snp 1191237263Snp INP_WLOCK(inp); 1192237263Snp final_cpl_received(toep); /* no more CPLs expected */ 1193237263Snp 1194237263Snp return (0); 1195237263Snp case TCPS_LAST_ACK: 1196237263Snp if (tcp_close(tp)) 1197237263Snp INP_WUNLOCK(inp); 1198237263Snp goto release; 1199237263Snp 1200237263Snp case TCPS_FIN_WAIT_1: 1201237263Snp if (so->so_rcv.sb_state & SBS_CANTRCVMORE) 1202237263Snp soisdisconnected(so); 1203237263Snp tp->t_state = TCPS_FIN_WAIT_2; 1204237263Snp break; 1205237263Snp 1206237263Snp default: 1207237263Snp log(LOG_ERR, 1208237263Snp "%s: TID %u received CPL_CLOSE_CON_RPL in state %s\n", 1209237263Snp __func__, tid, tcpstates[tp->t_state]); 1210237263Snp } 1211237263Snpdone: 1212237263Snp INP_WUNLOCK(inp); 1213286227Sjch INP_INFO_RUNLOCK(&V_tcbinfo); 1214237263Snp return (0); 1215237263Snp} 1216237263Snp 1217237263Snpvoid 1218237263Snpsend_abort_rpl(struct adapter *sc, struct sge_wrq *ofld_txq, int tid, 1219237263Snp int rst_status) 1220237263Snp{ 1221237263Snp struct wrqe *wr; 1222237263Snp struct cpl_abort_rpl *cpl; 1223237263Snp 1224237263Snp wr = alloc_wrqe(sizeof(*cpl), ofld_txq); 1225237263Snp if (wr == NULL) { 1226237263Snp /* XXX */ 1227237263Snp panic("%s: allocation failure.", __func__); 1228237263Snp } 1229237263Snp cpl = wrtod(wr); 1230237263Snp 1231237263Snp INIT_TP_WR_MIT_CPL(cpl, CPL_ABORT_RPL, tid); 1232237263Snp cpl->cmd = rst_status; 1233237263Snp 1234237263Snp t4_wrq_tx(sc, wr); 1235237263Snp} 1236237263Snp 1237237263Snpstatic int 1238237263Snpabort_status_to_errno(struct tcpcb *tp, unsigned int abort_reason) 1239237263Snp{ 1240237263Snp switch (abort_reason) { 1241237263Snp case CPL_ERR_BAD_SYN: 1242237263Snp case CPL_ERR_CONN_RESET: 1243237263Snp return (tp->t_state == TCPS_CLOSE_WAIT ? EPIPE : ECONNRESET); 1244237263Snp case CPL_ERR_XMIT_TIMEDOUT: 1245237263Snp case CPL_ERR_PERSIST_TIMEDOUT: 1246237263Snp case CPL_ERR_FINWAIT2_TIMEDOUT: 1247237263Snp case CPL_ERR_KEEPALIVE_TIMEDOUT: 1248237263Snp return (ETIMEDOUT); 1249237263Snp default: 1250237263Snp return (EIO); 1251237263Snp } 1252237263Snp} 1253237263Snp 1254237263Snp/* 1255237263Snp * TCP RST from the peer, timeout, or some other such critical error. 1256237263Snp */ 1257237263Snpstatic int 1258237263Snpdo_abort_req(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) 1259237263Snp{ 1260237263Snp struct adapter *sc = iq->adapter; 1261237263Snp const struct cpl_abort_req_rss *cpl = (const void *)(rss + 1); 1262237263Snp unsigned int tid = GET_TID(cpl); 1263237263Snp struct toepcb *toep = lookup_tid(sc, tid); 1264237263Snp struct sge_wrq *ofld_txq = toep->ofld_txq; 1265237263Snp struct inpcb *inp; 1266237263Snp struct tcpcb *tp; 1267237263Snp#ifdef INVARIANTS 1268237263Snp unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); 1269237263Snp#endif 1270237263Snp 1271237263Snp KASSERT(opcode == CPL_ABORT_REQ_RSS, 1272237263Snp ("%s: unexpected opcode 0x%x", __func__, opcode)); 1273237263Snp KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 1274237263Snp 1275239514Snp if (toep->flags & TPF_SYNQE) 1276237263Snp return (do_abort_req_synqe(iq, rss, m)); 1277237263Snp 1278237263Snp KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__)); 1279237263Snp 1280245935Snp if (negative_advice(cpl->status)) { 1281237263Snp CTR4(KTR_CXGBE, "%s: negative advice %d for tid %d (0x%x)", 1282237263Snp __func__, cpl->status, tid, toep->flags); 1283237263Snp return (0); /* Ignore negative advice */ 1284237263Snp } 1285237263Snp 1286237263Snp inp = toep->inp; 1287286227Sjch INP_INFO_RLOCK(&V_tcbinfo); /* for tcp_close */ 1288237263Snp INP_WLOCK(inp); 1289237263Snp 1290237263Snp tp = intotcpcb(inp); 1291237263Snp 1292237263Snp CTR6(KTR_CXGBE, 1293237263Snp "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x, status %d", 1294239528Snp __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags, 1295239528Snp inp->inp_flags, cpl->status); 1296237263Snp 1297237263Snp /* 1298237263Snp * If we'd initiated an abort earlier the reply to it is responsible for 1299237263Snp * cleaning up resources. Otherwise we tear everything down right here 1300237263Snp * right now. We owe the T4 a CPL_ABORT_RPL no matter what. 1301237263Snp */ 1302239514Snp if (toep->flags & TPF_ABORT_SHUTDOWN) { 1303237263Snp INP_WUNLOCK(inp); 1304237263Snp goto done; 1305237263Snp } 1306239514Snp toep->flags |= TPF_ABORT_SHUTDOWN; 1307237263Snp 1308242671Snp if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) { 1309242671Snp struct socket *so = inp->inp_socket; 1310237263Snp 1311242671Snp if (so != NULL) 1312242671Snp so_error_set(so, abort_status_to_errno(tp, 1313242671Snp cpl->status)); 1314242671Snp tp = tcp_close(tp); 1315242671Snp if (tp == NULL) 1316242671Snp INP_WLOCK(inp); /* re-acquire */ 1317242671Snp } 1318242671Snp 1319237263Snp final_cpl_received(toep); 1320237263Snpdone: 1321286227Sjch INP_INFO_RUNLOCK(&V_tcbinfo); 1322237263Snp send_abort_rpl(sc, ofld_txq, tid, CPL_ABORT_NO_RST); 1323237263Snp return (0); 1324237263Snp} 1325237263Snp 1326237263Snp/* 1327237263Snp * Reply to the CPL_ABORT_REQ (send_reset) 1328237263Snp */ 1329237263Snpstatic int 1330237263Snpdo_abort_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) 1331237263Snp{ 1332237263Snp struct adapter *sc = iq->adapter; 1333237263Snp const struct cpl_abort_rpl_rss *cpl = (const void *)(rss + 1); 1334237263Snp unsigned int tid = GET_TID(cpl); 1335237263Snp struct toepcb *toep = lookup_tid(sc, tid); 1336237263Snp struct inpcb *inp = toep->inp; 1337237263Snp#ifdef INVARIANTS 1338237263Snp unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); 1339237263Snp#endif 1340237263Snp 1341237263Snp KASSERT(opcode == CPL_ABORT_RPL_RSS, 1342237263Snp ("%s: unexpected opcode 0x%x", __func__, opcode)); 1343237263Snp KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 1344237263Snp 1345239514Snp if (toep->flags & TPF_SYNQE) 1346237263Snp return (do_abort_rpl_synqe(iq, rss, m)); 1347237263Snp 1348237263Snp KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__)); 1349237263Snp 1350237263Snp CTR5(KTR_CXGBE, "%s: tid %u, toep %p, inp %p, status %d", 1351237263Snp __func__, tid, toep, inp, cpl->status); 1352237263Snp 1353239514Snp KASSERT(toep->flags & TPF_ABORT_SHUTDOWN, 1354237263Snp ("%s: wasn't expecting abort reply", __func__)); 1355237263Snp 1356237263Snp INP_WLOCK(inp); 1357237263Snp final_cpl_received(toep); 1358237263Snp 1359237263Snp return (0); 1360237263Snp} 1361237263Snp 1362237263Snpstatic int 1363237263Snpdo_rx_data(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) 1364237263Snp{ 1365237263Snp struct adapter *sc = iq->adapter; 1366237263Snp const struct cpl_rx_data *cpl = mtod(m, const void *); 1367237263Snp unsigned int tid = GET_TID(cpl); 1368237263Snp struct toepcb *toep = lookup_tid(sc, tid); 1369237263Snp struct inpcb *inp = toep->inp; 1370237263Snp struct tcpcb *tp; 1371237263Snp struct socket *so; 1372239344Snp struct sockbuf *sb; 1373239344Snp int len; 1374243681Snp uint32_t ddp_placed = 0; 1375237263Snp 1376239514Snp if (__predict_false(toep->flags & TPF_SYNQE)) { 1377243680Snp#ifdef INVARIANTS 1378243680Snp struct synq_entry *synqe = (void *)toep; 1379243680Snp 1380243680Snp INP_WLOCK(synqe->lctx->inp); 1381243680Snp if (synqe->flags & TPF_SYNQE_HAS_L2TE) { 1382243680Snp KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN, 1383243680Snp ("%s: listen socket closed but tid %u not aborted.", 1384243680Snp __func__, tid)); 1385243680Snp } else { 1386243680Snp /* 1387243680Snp * do_pass_accept_req is still running and will 1388243680Snp * eventually take care of this tid. 1389243680Snp */ 1390243680Snp } 1391243680Snp INP_WUNLOCK(synqe->lctx->inp); 1392243680Snp#endif 1393243680Snp CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid, 1394243680Snp toep, toep->flags); 1395237263Snp m_freem(m); 1396237263Snp return (0); 1397237263Snp } 1398237263Snp 1399237263Snp KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__)); 1400237263Snp 1401237263Snp /* strip off CPL header */ 1402237263Snp m_adj(m, sizeof(*cpl)); 1403239344Snp len = m->m_pkthdr.len; 1404237263Snp 1405237263Snp INP_WLOCK(inp); 1406237263Snp if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) { 1407237263Snp CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x", 1408239344Snp __func__, tid, len, inp->inp_flags); 1409237263Snp INP_WUNLOCK(inp); 1410237263Snp m_freem(m); 1411237263Snp return (0); 1412237263Snp } 1413237263Snp 1414237263Snp tp = intotcpcb(inp); 1415237263Snp 1416243681Snp if (__predict_false(tp->rcv_nxt != be32toh(cpl->seq))) 1417243681Snp ddp_placed = be32toh(cpl->seq) - tp->rcv_nxt; 1418237263Snp 1419239344Snp tp->rcv_nxt += len; 1420300895Snp if (tp->rcv_wnd < len) { 1421300895Snp KASSERT(toep->ulp_mode != ULP_MODE_RDMA, 1422300895Snp ("%s: negative window size", __func__)); 1423300895Snp } 1424300895Snp 1425239344Snp tp->rcv_wnd -= len; 1426237263Snp tp->t_rcvtime = ticks; 1427237263Snp 1428299210Sjhb if (toep->ulp_mode == ULP_MODE_TCPDDP) 1429299210Sjhb DDP_LOCK(toep); 1430237263Snp so = inp_inpcbtosocket(inp); 1431239344Snp sb = &so->so_rcv; 1432239344Snp SOCKBUF_LOCK(sb); 1433237263Snp 1434239344Snp if (__predict_false(sb->sb_state & SBS_CANTRCVMORE)) { 1435237263Snp CTR3(KTR_CXGBE, "%s: tid %u, excess rx (%d bytes)", 1436239344Snp __func__, tid, len); 1437237263Snp m_freem(m); 1438239344Snp SOCKBUF_UNLOCK(sb); 1439299210Sjhb if (toep->ulp_mode == ULP_MODE_TCPDDP) 1440299210Sjhb DDP_UNLOCK(toep); 1441237263Snp INP_WUNLOCK(inp); 1442237263Snp 1443286227Sjch INP_INFO_RLOCK(&V_tcbinfo); 1444237263Snp INP_WLOCK(inp); 1445237263Snp tp = tcp_drop(tp, ECONNRESET); 1446237263Snp if (tp) 1447237263Snp INP_WUNLOCK(inp); 1448286227Sjch INP_INFO_RUNLOCK(&V_tcbinfo); 1449237263Snp 1450237263Snp return (0); 1451237263Snp } 1452237263Snp 1453237263Snp /* receive buffer autosize */ 1454299206Sjhb CURVNET_SET(so->so_vnet); 1455239344Snp if (sb->sb_flags & SB_AUTOSIZE && 1456237263Snp V_tcp_do_autorcvbuf && 1457239344Snp sb->sb_hiwat < V_tcp_autorcvbuf_max && 1458239344Snp len > (sbspace(sb) / 8 * 7)) { 1459239344Snp unsigned int hiwat = sb->sb_hiwat; 1460237263Snp unsigned int newsize = min(hiwat + V_tcp_autorcvbuf_inc, 1461237263Snp V_tcp_autorcvbuf_max); 1462237263Snp 1463239344Snp if (!sbreserve_locked(sb, newsize, so, NULL)) 1464239344Snp sb->sb_flags &= ~SB_AUTOSIZE; 1465237263Snp else 1466237263Snp toep->rx_credits += newsize - hiwat; 1467237263Snp } 1468239344Snp 1469299210Sjhb if (toep->ddp_waiting_count != 0 || toep->ddp_active_count != 0) 1470299210Sjhb CTR3(KTR_CXGBE, "%s: tid %u, non-ddp rx (%d bytes)", __func__, 1471299210Sjhb tid, len); 1472299210Sjhb 1473239344Snp if (toep->ulp_mode == ULP_MODE_TCPDDP) { 1474239344Snp int changed = !(toep->ddp_flags & DDP_ON) ^ cpl->ddp_off; 1475239344Snp 1476239344Snp if (changed) { 1477243681Snp if (toep->ddp_flags & DDP_SC_REQ) 1478243681Snp toep->ddp_flags ^= DDP_ON | DDP_SC_REQ; 1479243681Snp else { 1480243681Snp KASSERT(cpl->ddp_off == 1, 1481243681Snp ("%s: DDP switched on by itself.", 1482243681Snp __func__)); 1483243681Snp 1484243681Snp /* Fell out of DDP mode */ 1485299210Sjhb toep->ddp_flags &= ~DDP_ON; 1486299210Sjhb CTR1(KTR_CXGBE, "%s: fell out of DDP mode", 1487299210Sjhb __func__); 1488243681Snp 1489299210Sjhb insert_ddp_data(toep, ddp_placed); 1490239344Snp } 1491239344Snp } 1492239344Snp 1493239344Snp if (toep->ddp_flags & DDP_ON) { 1494239344Snp /* 1495299210Sjhb * CPL_RX_DATA with DDP on can only be an indicate. 1496299210Sjhb * Start posting queued AIO requests via DDP. The 1497299210Sjhb * payload that arrived in this indicate is appended 1498299210Sjhb * to the socket buffer as usual. 1499239344Snp */ 1500299210Sjhb handle_ddp_indicate(toep); 1501239344Snp } 1502239344Snp } 1503239344Snp 1504274421Sglebius KASSERT(toep->sb_cc >= sbused(sb), 1505239344Snp ("%s: sb %p has more data (%d) than last time (%d).", 1506274421Sglebius __func__, sb, sbused(sb), toep->sb_cc)); 1507274421Sglebius toep->rx_credits += toep->sb_cc - sbused(sb); 1508275329Sglebius sbappendstream_locked(sb, m, 0); 1509274421Sglebius toep->sb_cc = sbused(sb); 1510280878Snp if (toep->rx_credits > 0 && toep->sb_cc + tp->rcv_wnd < sb->sb_lowat) { 1511280878Snp int credits; 1512280878Snp 1513280878Snp credits = send_rx_credits(sc, toep, toep->rx_credits); 1514280878Snp toep->rx_credits -= credits; 1515280878Snp tp->rcv_wnd += credits; 1516280878Snp tp->rcv_adv += credits; 1517280878Snp } 1518299210Sjhb 1519299210Sjhb if (toep->ddp_waiting_count > 0 && sbavail(sb) != 0) { 1520299210Sjhb CTR2(KTR_CXGBE, "%s: tid %u queueing AIO task", __func__, 1521299210Sjhb tid); 1522299210Sjhb ddp_queue_toep(toep); 1523299210Sjhb } 1524237263Snp sorwakeup_locked(so); 1525239344Snp SOCKBUF_UNLOCK_ASSERT(sb); 1526299210Sjhb if (toep->ulp_mode == ULP_MODE_TCPDDP) 1527299210Sjhb DDP_UNLOCK(toep); 1528237263Snp 1529237263Snp INP_WUNLOCK(inp); 1530299206Sjhb CURVNET_RESTORE(); 1531237263Snp return (0); 1532237263Snp} 1533237263Snp 1534237263Snp#define S_CPL_FW4_ACK_OPCODE 24 1535237263Snp#define M_CPL_FW4_ACK_OPCODE 0xff 1536237263Snp#define V_CPL_FW4_ACK_OPCODE(x) ((x) << S_CPL_FW4_ACK_OPCODE) 1537237263Snp#define G_CPL_FW4_ACK_OPCODE(x) \ 1538237263Snp (((x) >> S_CPL_FW4_ACK_OPCODE) & M_CPL_FW4_ACK_OPCODE) 1539259382Snp 1540237263Snp#define S_CPL_FW4_ACK_FLOWID 0 1541237263Snp#define M_CPL_FW4_ACK_FLOWID 0xffffff 1542237263Snp#define V_CPL_FW4_ACK_FLOWID(x) ((x) << S_CPL_FW4_ACK_FLOWID) 1543237263Snp#define G_CPL_FW4_ACK_FLOWID(x) \ 1544237263Snp (((x) >> S_CPL_FW4_ACK_FLOWID) & M_CPL_FW4_ACK_FLOWID) 1545259382Snp 1546237263Snp#define S_CPL_FW4_ACK_CR 24 1547237263Snp#define M_CPL_FW4_ACK_CR 0xff 1548237263Snp#define V_CPL_FW4_ACK_CR(x) ((x) << S_CPL_FW4_ACK_CR) 1549237263Snp#define G_CPL_FW4_ACK_CR(x) (((x) >> S_CPL_FW4_ACK_CR) & M_CPL_FW4_ACK_CR) 1550259382Snp 1551237263Snp#define S_CPL_FW4_ACK_SEQVAL 0 1552237263Snp#define M_CPL_FW4_ACK_SEQVAL 0x1 1553237263Snp#define V_CPL_FW4_ACK_SEQVAL(x) ((x) << S_CPL_FW4_ACK_SEQVAL) 1554237263Snp#define G_CPL_FW4_ACK_SEQVAL(x) \ 1555237263Snp (((x) >> S_CPL_FW4_ACK_SEQVAL) & M_CPL_FW4_ACK_SEQVAL) 1556237263Snp#define F_CPL_FW4_ACK_SEQVAL V_CPL_FW4_ACK_SEQVAL(1U) 1557237263Snp 1558237263Snpstatic int 1559237263Snpdo_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) 1560237263Snp{ 1561237263Snp struct adapter *sc = iq->adapter; 1562237263Snp const struct cpl_fw4_ack *cpl = (const void *)(rss + 1); 1563237263Snp unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl))); 1564237263Snp struct toepcb *toep = lookup_tid(sc, tid); 1565237263Snp struct inpcb *inp; 1566237263Snp struct tcpcb *tp; 1567237263Snp struct socket *so; 1568237263Snp uint8_t credits = cpl->credits; 1569237263Snp struct ofld_tx_sdesc *txsd; 1570237263Snp int plen; 1571237263Snp#ifdef INVARIANTS 1572237263Snp unsigned int opcode = G_CPL_FW4_ACK_OPCODE(be32toh(OPCODE_TID(cpl))); 1573237263Snp#endif 1574237263Snp 1575237263Snp /* 1576237263Snp * Very unusual case: we'd sent a flowc + abort_req for a synq entry and 1577237263Snp * now this comes back carrying the credits for the flowc. 1578237263Snp */ 1579239514Snp if (__predict_false(toep->flags & TPF_SYNQE)) { 1580239514Snp KASSERT(toep->flags & TPF_ABORT_SHUTDOWN, 1581237263Snp ("%s: credits for a synq entry %p", __func__, toep)); 1582237263Snp return (0); 1583237263Snp } 1584237263Snp 1585237263Snp inp = toep->inp; 1586237263Snp 1587237263Snp KASSERT(opcode == CPL_FW4_ACK, 1588237263Snp ("%s: unexpected opcode 0x%x", __func__, opcode)); 1589237263Snp KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 1590237263Snp KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__)); 1591237263Snp 1592237263Snp INP_WLOCK(inp); 1593237263Snp 1594239514Snp if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN)) { 1595237263Snp INP_WUNLOCK(inp); 1596237263Snp return (0); 1597237263Snp } 1598237263Snp 1599237263Snp KASSERT((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) == 0, 1600237263Snp ("%s: inp_flags 0x%x", __func__, inp->inp_flags)); 1601237263Snp 1602237263Snp tp = intotcpcb(inp); 1603237263Snp 1604237436Snp if (cpl->flags & CPL_FW4_ACK_FLAGS_SEQVAL) { 1605237263Snp tcp_seq snd_una = be32toh(cpl->snd_una); 1606237263Snp 1607237263Snp#ifdef INVARIANTS 1608237263Snp if (__predict_false(SEQ_LT(snd_una, tp->snd_una))) { 1609237263Snp log(LOG_ERR, 1610237263Snp "%s: unexpected seq# %x for TID %u, snd_una %x\n", 1611237263Snp __func__, snd_una, toep->tid, tp->snd_una); 1612237263Snp } 1613237263Snp#endif 1614237263Snp 1615237263Snp if (tp->snd_una != snd_una) { 1616237263Snp tp->snd_una = snd_una; 1617237263Snp tp->ts_recent_age = tcp_ts_getticks(); 1618237263Snp } 1619237263Snp } 1620237263Snp 1621237263Snp so = inp->inp_socket; 1622237263Snp txsd = &toep->txsd[toep->txsd_cidx]; 1623237263Snp plen = 0; 1624237263Snp while (credits) { 1625237263Snp KASSERT(credits >= txsd->tx_credits, 1626237263Snp ("%s: too many (or partial) credits", __func__)); 1627237263Snp credits -= txsd->tx_credits; 1628237263Snp toep->tx_credits += txsd->tx_credits; 1629237263Snp plen += txsd->plen; 1630237263Snp txsd++; 1631237263Snp toep->txsd_avail++; 1632237263Snp KASSERT(toep->txsd_avail <= toep->txsd_total, 1633237263Snp ("%s: txsd avail > total", __func__)); 1634237263Snp if (__predict_false(++toep->txsd_cidx == toep->txsd_total)) { 1635237263Snp txsd = &toep->txsd[0]; 1636237263Snp toep->txsd_cidx = 0; 1637237263Snp } 1638237263Snp } 1639237263Snp 1640255411Snp if (toep->tx_credits == toep->tx_total) { 1641255411Snp toep->tx_nocompl = 0; 1642255411Snp toep->plen_nocompl = 0; 1643255411Snp } 1644255411Snp 1645255411Snp if (toep->flags & TPF_TX_SUSPENDED && 1646255411Snp toep->tx_credits >= toep->tx_total / 4) { 1647255411Snp toep->flags &= ~TPF_TX_SUSPENDED; 1648269076Snp if (toep->ulp_mode == ULP_MODE_ISCSI) 1649292736Snp t4_push_pdus(sc, toep, plen); 1650269076Snp else 1651269076Snp t4_push_frames(sc, toep, plen); 1652255411Snp } else if (plen > 0) { 1653237263Snp struct sockbuf *sb = &so->so_snd; 1654292736Snp int sbu; 1655237263Snp 1656292736Snp SOCKBUF_LOCK(sb); 1657292736Snp sbu = sbused(sb); 1658292736Snp if (toep->ulp_mode == ULP_MODE_ISCSI) { 1659292736Snp 1660292736Snp if (__predict_false(sbu > 0)) { 1661292736Snp /* 1662292736Snp * The data trasmitted before the tid's ULP mode 1663292736Snp * changed to ISCSI is still in so_snd. 1664292736Snp * Incoming credits should account for so_snd 1665292736Snp * first. 1666292736Snp */ 1667292736Snp sbdrop_locked(sb, min(sbu, plen)); 1668292736Snp plen -= min(sbu, plen); 1669292736Snp } 1670292736Snp sowwakeup_locked(so); /* unlocks so_snd */ 1671292736Snp rqdrop_locked(&toep->ulp_pdu_reclaimq, plen); 1672292736Snp } else { 1673269076Snp sbdrop_locked(sb, plen); 1674292736Snp sowwakeup_locked(so); /* unlocks so_snd */ 1675269076Snp } 1676292736Snp SOCKBUF_UNLOCK_ASSERT(sb); 1677237263Snp } 1678237263Snp 1679237263Snp INP_WUNLOCK(inp); 1680237263Snp 1681237263Snp return (0); 1682237263Snp} 1683237263Snp 1684239338Snpstatic int 1685239338Snpdo_set_tcb_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) 1686239338Snp{ 1687239338Snp struct adapter *sc = iq->adapter; 1688239338Snp const struct cpl_set_tcb_rpl *cpl = (const void *)(rss + 1); 1689239338Snp unsigned int tid = GET_TID(cpl); 1690299210Sjhb struct toepcb *toep; 1691239338Snp#ifdef INVARIANTS 1692239338Snp unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); 1693239338Snp#endif 1694239338Snp 1695239338Snp KASSERT(opcode == CPL_SET_TCB_RPL, 1696239338Snp ("%s: unexpected opcode 0x%x", __func__, opcode)); 1697239338Snp KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 1698239338Snp 1699259382Snp if (is_ftid(sc, tid)) 1700239338Snp return (t4_filter_rpl(iq, rss, m)); /* TCB is a filter */ 1701239338Snp 1702299210Sjhb toep = lookup_tid(sc, tid); 1703299210Sjhb if (toep->ulp_mode == ULP_MODE_TCPDDP) { 1704299210Sjhb handle_ddp_tcb_rpl(toep, cpl); 1705299210Sjhb return (0); 1706299210Sjhb } 1707299210Sjhb 1708292736Snp /* 1709292736Snp * TOM and/or other ULPs don't request replies for CPL_SET_TCB or 1710292736Snp * CPL_SET_TCB_FIELD requests. This can easily change and when it does 1711292736Snp * the dispatch code will go here. 1712292736Snp */ 1713292736Snp#ifdef INVARIANTS 1714292736Snp panic("%s: Unexpected CPL_SET_TCB_RPL for tid %u on iq %p", __func__, 1715292736Snp tid, iq); 1716292736Snp#else 1717292736Snp log(LOG_ERR, "%s: Unexpected CPL_SET_TCB_RPL for tid %u on iq %p\n", 1718292736Snp __func__, tid, iq); 1719292736Snp#endif 1720269076Snp 1721292736Snp return (0); 1722239338Snp} 1723239338Snp 1724237263Snpvoid 1725251638Snpt4_set_tcb_field(struct adapter *sc, struct toepcb *toep, int ctrl, 1726251638Snp uint16_t word, uint64_t mask, uint64_t val) 1727239338Snp{ 1728239338Snp struct wrqe *wr; 1729239338Snp struct cpl_set_tcb_field *req; 1730239338Snp 1731251638Snp wr = alloc_wrqe(sizeof(*req), ctrl ? toep->ctrlq : toep->ofld_txq); 1732239338Snp if (wr == NULL) { 1733239338Snp /* XXX */ 1734239338Snp panic("%s: allocation failure.", __func__); 1735239338Snp } 1736239338Snp req = wrtod(wr); 1737239338Snp 1738239338Snp INIT_TP_WR_MIT_CPL(req, CPL_SET_TCB_FIELD, toep->tid); 1739239338Snp req->reply_ctrl = htobe16(V_NO_REPLY(1) | 1740239338Snp V_QUEUENO(toep->ofld_rxq->iq.abs_id)); 1741239338Snp req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(0)); 1742239338Snp req->mask = htobe64(mask); 1743239338Snp req->val = htobe64(val); 1744239338Snp 1745239338Snp t4_wrq_tx(sc, wr); 1746239338Snp} 1747239338Snp 1748239338Snpvoid 1749299210Sjhbt4_set_tcb_field_rpl(struct adapter *sc, struct toepcb *toep, int ctrl, 1750299210Sjhb uint16_t word, uint64_t mask, uint64_t val, uint8_t cookie) 1751299210Sjhb{ 1752299210Sjhb struct wrqe *wr; 1753299210Sjhb struct cpl_set_tcb_field *req; 1754299210Sjhb 1755299210Sjhb KASSERT((cookie & ~M_COOKIE) == 0, ("%s: invalid cookie %#x", __func__, 1756299210Sjhb cookie)); 1757299210Sjhb wr = alloc_wrqe(sizeof(*req), ctrl ? toep->ctrlq : toep->ofld_txq); 1758299210Sjhb if (wr == NULL) { 1759299210Sjhb /* XXX */ 1760299210Sjhb panic("%s: allocation failure.", __func__); 1761299210Sjhb } 1762299210Sjhb req = wrtod(wr); 1763299210Sjhb 1764299210Sjhb INIT_TP_WR_MIT_CPL(req, CPL_SET_TCB_FIELD, toep->tid); 1765299210Sjhb req->reply_ctrl = htobe16(V_QUEUENO(toep->ofld_rxq->iq.abs_id)); 1766299210Sjhb req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(cookie)); 1767299210Sjhb req->mask = htobe64(mask); 1768299210Sjhb req->val = htobe64(val); 1769299210Sjhb 1770299210Sjhb t4_wrq_tx(sc, wr); 1771299210Sjhb} 1772299210Sjhb 1773299210Sjhbvoid 1774237263Snpt4_init_cpl_io_handlers(struct adapter *sc) 1775237263Snp{ 1776237263Snp 1777237263Snp t4_register_cpl_handler(sc, CPL_PEER_CLOSE, do_peer_close); 1778237263Snp t4_register_cpl_handler(sc, CPL_CLOSE_CON_RPL, do_close_con_rpl); 1779237263Snp t4_register_cpl_handler(sc, CPL_ABORT_REQ_RSS, do_abort_req); 1780237263Snp t4_register_cpl_handler(sc, CPL_ABORT_RPL_RSS, do_abort_rpl); 1781237263Snp t4_register_cpl_handler(sc, CPL_RX_DATA, do_rx_data); 1782237263Snp t4_register_cpl_handler(sc, CPL_FW4_ACK, do_fw4_ack); 1783239338Snp t4_register_cpl_handler(sc, CPL_SET_TCB_RPL, do_set_tcb_rpl); 1784237263Snp} 1785239338Snp 1786239338Snpvoid 1787239338Snpt4_uninit_cpl_io_handlers(struct adapter *sc) 1788239338Snp{ 1789239338Snp 1790239338Snp t4_register_cpl_handler(sc, CPL_SET_TCB_RPL, t4_filter_rpl); 1791239338Snp} 1792237263Snp#endif 1793