1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright (c) 2012, 2015 Chelsio Communications, Inc. 5 * All rights reserved. 6 * Written by: Navdeep Parhar <np@FreeBSD.org> 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 20 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 27 * SUCH DAMAGE. 28 */ 29 30 #include <sys/cdefs.h> 31 __FBSDID("$FreeBSD$"); 32 33 #include "opt_inet.h" 34 #include "opt_inet6.h" 35 #include "opt_kern_tls.h" 36 #include "opt_ratelimit.h" 37 38 #ifdef TCP_OFFLOAD 39 #include <sys/param.h> 40 #include <sys/aio.h> 41 #include <sys/file.h> 42 #include <sys/kernel.h> 43 #include <sys/ktr.h> 44 #include <sys/module.h> 45 #include <sys/proc.h> 46 #include <sys/protosw.h> 47 #include <sys/domain.h> 48 #include <sys/socket.h> 49 #include <sys/socketvar.h> 50 #include <sys/sglist.h> 51 #include <sys/taskqueue.h> 52 #include <netinet/in.h> 53 #include <netinet/in_pcb.h> 54 #include <netinet/ip.h> 55 #include <netinet/ip6.h> 56 #define TCPSTATES 57 #include <netinet/tcp_fsm.h> 58 #include <netinet/tcp_seq.h> 59 #include <netinet/tcp_var.h> 60 #include <netinet/toecore.h> 61 62 #include <security/mac/mac_framework.h> 63 64 #include <vm/vm.h> 65 #include <vm/vm_extern.h> 66 #include <vm/pmap.h> 67 #include <vm/vm_map.h> 68 #include <vm/vm_page.h> 69 70 #include "common/common.h" 71 #include "common/t4_msg.h" 72 #include "common/t4_regs.h" 73 #include "common/t4_tcb.h" 74 #include "tom/t4_tom_l2t.h" 75 #include "tom/t4_tom.h" 76 77 static void t4_aiotx_cancel(struct kaiocb *job); 78 static void t4_aiotx_queue_toep(struct socket *so, struct toepcb *toep); 79 80 void 81 send_flowc_wr(struct toepcb *toep, struct tcpcb *tp) 82 { 83 struct wrqe *wr; 84 struct fw_flowc_wr *flowc; 85 unsigned int nparams, flowclen, paramidx; 86 struct vi_info *vi = toep->vi; 87 struct port_info *pi = vi->pi; 88 struct adapter *sc = pi->adapter; 89 unsigned int pfvf = sc->pf << S_FW_VIID_PFN; 90 struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx]; 91 92 KASSERT(!(toep->flags & TPF_FLOWC_WR_SENT), 93 ("%s: flowc for tid %u sent already", __func__, toep->tid)); 94 95 if (tp != NULL) 96 nparams = 8; 97 else 98 nparams = 6; 99 if (ulp_mode(toep) == ULP_MODE_TLS) 100 nparams++; 101 if (toep->tls.fcplenmax != 0) 102 nparams++; 103 if (toep->params.tc_idx != -1) { 104 MPASS(toep->params.tc_idx >= 0 && 105 toep->params.tc_idx < sc->chip_params->nsched_cls); 106 nparams++; 107 } 108 109 flowclen = sizeof(*flowc) + nparams * sizeof(struct fw_flowc_mnemval); 110 111 wr = alloc_wrqe(roundup2(flowclen, 16), toep->ofld_txq); 112 if (wr == NULL) { 113 /* XXX */ 114 panic("%s: allocation failure.", __func__); 115 } 116 flowc = wrtod(wr); 117 memset(flowc, 0, wr->wr_len); 118 119 flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) | 120 V_FW_FLOWC_WR_NPARAMS(nparams)); 121 flowc->flowid_len16 = htonl(V_FW_WR_LEN16(howmany(flowclen, 16)) | 122 V_FW_WR_FLOWID(toep->tid)); 123 124 #define FLOWC_PARAM(__m, __v) \ 125 do { \ 126 flowc->mnemval[paramidx].mnemonic = FW_FLOWC_MNEM_##__m; \ 127 flowc->mnemval[paramidx].val = htobe32(__v); \ 128 paramidx++; \ 129 } while (0) 130 131 paramidx = 0; 132 133 FLOWC_PARAM(PFNVFN, pfvf); 134 FLOWC_PARAM(CH, pi->tx_chan); 135 FLOWC_PARAM(PORT, pi->tx_chan); 136 FLOWC_PARAM(IQID, toep->ofld_rxq->iq.abs_id); 137 FLOWC_PARAM(SNDBUF, toep->params.sndbuf); 138 if (tp) { 139 FLOWC_PARAM(MSS, toep->params.emss); 140 FLOWC_PARAM(SNDNXT, tp->snd_nxt); 141 FLOWC_PARAM(RCVNXT, tp->rcv_nxt); 142 } else 143 FLOWC_PARAM(MSS, 512); 144 CTR6(KTR_CXGBE, 145 "%s: tid %u, mss %u, sndbuf %u, snd_nxt 0x%x, rcv_nxt 0x%x", 146 __func__, toep->tid, toep->params.emss, toep->params.sndbuf, 147 tp ? tp->snd_nxt : 0, tp ? tp->rcv_nxt : 0); 148 149 if (ulp_mode(toep) == ULP_MODE_TLS) 150 FLOWC_PARAM(ULP_MODE, ulp_mode(toep)); 151 if (toep->tls.fcplenmax != 0) 152 FLOWC_PARAM(TXDATAPLEN_MAX, toep->tls.fcplenmax); 153 if (toep->params.tc_idx != -1) 154 FLOWC_PARAM(SCHEDCLASS, toep->params.tc_idx); 155 #undef FLOWC_PARAM 156 157 KASSERT(paramidx == nparams, ("nparams mismatch")); 158 159 txsd->tx_credits = howmany(flowclen, 16); 160 txsd->plen = 0; 161 KASSERT(toep->tx_credits >= txsd->tx_credits && toep->txsd_avail > 0, 162 ("%s: not enough credits (%d)", __func__, toep->tx_credits)); 163 toep->tx_credits -= txsd->tx_credits; 164 if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) 165 toep->txsd_pidx = 0; 166 toep->txsd_avail--; 167 168 toep->flags |= TPF_FLOWC_WR_SENT; 169 t4_wrq_tx(sc, wr); 170 } 171 172 #ifdef RATELIMIT 173 /* 174 * Input is Bytes/second (so_max_pacing_rate), chip counts in Kilobits/second. 175 */ 176 static int 177 update_tx_rate_limit(struct adapter *sc, struct toepcb *toep, u_int Bps) 178 { 179 int tc_idx, rc; 180 const u_int kbps = (u_int) (uint64_t)Bps * 8ULL / 1000; 181 const int port_id = toep->vi->pi->port_id; 182 183 CTR3(KTR_CXGBE, "%s: tid %u, rate %uKbps", __func__, toep->tid, kbps); 184 185 if (kbps == 0) { 186 /* unbind */ 187 tc_idx = -1; 188 } else { 189 rc = t4_reserve_cl_rl_kbps(sc, port_id, kbps, &tc_idx); 190 if (rc != 0) 191 return (rc); 192 MPASS(tc_idx >= 0 && tc_idx < sc->chip_params->nsched_cls); 193 } 194 195 if (toep->params.tc_idx != tc_idx) { 196 struct wrqe *wr; 197 struct fw_flowc_wr *flowc; 198 int nparams = 1, flowclen, flowclen16; 199 struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx]; 200 201 flowclen = sizeof(*flowc) + nparams * sizeof(struct 202 fw_flowc_mnemval); 203 flowclen16 = howmany(flowclen, 16); 204 if (toep->tx_credits < flowclen16 || toep->txsd_avail == 0 || 205 (wr = alloc_wrqe(roundup2(flowclen, 16), toep->ofld_txq)) == NULL) { 206 if (tc_idx >= 0) 207 t4_release_cl_rl(sc, port_id, tc_idx); 208 return (ENOMEM); 209 } 210 211 flowc = wrtod(wr); 212 memset(flowc, 0, wr->wr_len); 213 214 flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) | 215 V_FW_FLOWC_WR_NPARAMS(nparams)); 216 flowc->flowid_len16 = htonl(V_FW_WR_LEN16(flowclen16) | 217 V_FW_WR_FLOWID(toep->tid)); 218 219 flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_SCHEDCLASS; 220 if (tc_idx == -1) 221 flowc->mnemval[0].val = htobe32(0xff); 222 else 223 flowc->mnemval[0].val = htobe32(tc_idx); 224 225 txsd->tx_credits = flowclen16; 226 txsd->plen = 0; 227 toep->tx_credits -= txsd->tx_credits; 228 if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) 229 toep->txsd_pidx = 0; 230 toep->txsd_avail--; 231 t4_wrq_tx(sc, wr); 232 } 233 234 if (toep->params.tc_idx >= 0) 235 t4_release_cl_rl(sc, port_id, toep->params.tc_idx); 236 toep->params.tc_idx = tc_idx; 237 238 return (0); 239 } 240 #endif 241 242 void 243 send_reset(struct adapter *sc, struct toepcb *toep, uint32_t snd_nxt) 244 { 245 struct wrqe *wr; 246 struct cpl_abort_req *req; 247 int tid = toep->tid; 248 struct inpcb *inp = toep->inp; 249 struct tcpcb *tp = intotcpcb(inp); /* don't use if INP_DROPPED */ 250 251 INP_WLOCK_ASSERT(inp); 252 253 CTR6(KTR_CXGBE, "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x%s", 254 __func__, toep->tid, 255 inp->inp_flags & INP_DROPPED ? "inp dropped" : 256 tcpstates[tp->t_state], 257 toep->flags, inp->inp_flags, 258 toep->flags & TPF_ABORT_SHUTDOWN ? 259 " (abort already in progress)" : ""); 260 261 if (toep->flags & TPF_ABORT_SHUTDOWN) 262 return; /* abort already in progress */ 263 264 toep->flags |= TPF_ABORT_SHUTDOWN; 265 266 KASSERT(toep->flags & TPF_FLOWC_WR_SENT, 267 ("%s: flowc_wr not sent for tid %d.", __func__, tid)); 268 269 wr = alloc_wrqe(sizeof(*req), toep->ofld_txq); 270 if (wr == NULL) { 271 /* XXX */ 272 panic("%s: allocation failure.", __func__); 273 } 274 req = wrtod(wr); 275 276 INIT_TP_WR_MIT_CPL(req, CPL_ABORT_REQ, tid); 277 if (inp->inp_flags & INP_DROPPED) 278 req->rsvd0 = htobe32(snd_nxt); 279 else 280 req->rsvd0 = htobe32(tp->snd_nxt); 281 req->rsvd1 = !(toep->flags & TPF_TX_DATA_SENT); 282 req->cmd = CPL_ABORT_SEND_RST; 283 284 /* 285 * XXX: What's the correct way to tell that the inp hasn't been detached 286 * from its socket? Should I even be flushing the snd buffer here? 287 */ 288 if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) { 289 struct socket *so = inp->inp_socket; 290 291 if (so != NULL) /* because I'm not sure. See comment above */ 292 sbflush(&so->so_snd); 293 } 294 295 t4_l2t_send(sc, wr, toep->l2te); 296 } 297 298 /* 299 * Called when a connection is established to translate the TCP options 300 * reported by HW to FreeBSD's native format. 301 */ 302 static void 303 assign_rxopt(struct tcpcb *tp, uint16_t opt) 304 { 305 struct toepcb *toep = tp->t_toe; 306 struct inpcb *inp = tp->t_inpcb; 307 struct adapter *sc = td_adapter(toep->td); 308 309 INP_LOCK_ASSERT(inp); 310 311 toep->params.mtu_idx = G_TCPOPT_MSS(opt); 312 tp->t_maxseg = sc->params.mtus[toep->params.mtu_idx]; 313 if (inp->inp_inc.inc_flags & INC_ISIPV6) 314 tp->t_maxseg -= sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 315 else 316 tp->t_maxseg -= sizeof(struct ip) + sizeof(struct tcphdr); 317 318 toep->params.emss = tp->t_maxseg; 319 if (G_TCPOPT_TSTAMP(opt)) { 320 toep->params.tstamp = 1; 321 toep->params.emss -= TCPOLEN_TSTAMP_APPA; 322 tp->t_flags |= TF_RCVD_TSTMP; /* timestamps ok */ 323 tp->ts_recent = 0; /* hmmm */ 324 tp->ts_recent_age = tcp_ts_getticks(); 325 } else 326 toep->params.tstamp = 0; 327 328 if (G_TCPOPT_SACK(opt)) { 329 toep->params.sack = 1; 330 tp->t_flags |= TF_SACK_PERMIT; /* should already be set */ 331 } else { 332 toep->params.sack = 0; 333 tp->t_flags &= ~TF_SACK_PERMIT; /* sack disallowed by peer */ 334 } 335 336 if (G_TCPOPT_WSCALE_OK(opt)) 337 tp->t_flags |= TF_RCVD_SCALE; 338 339 /* Doing window scaling? */ 340 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 341 (TF_RCVD_SCALE | TF_REQ_SCALE)) { 342 tp->rcv_scale = tp->request_r_scale; 343 tp->snd_scale = G_TCPOPT_SND_WSCALE(opt); 344 } else 345 toep->params.wscale = 0; 346 347 CTR6(KTR_CXGBE, 348 "assign_rxopt: tid %d, mtu_idx %u, emss %u, ts %u, sack %u, wscale %u", 349 toep->tid, toep->params.mtu_idx, toep->params.emss, 350 toep->params.tstamp, toep->params.sack, toep->params.wscale); 351 } 352 353 /* 354 * Completes some final bits of initialization for just established connections 355 * and changes their state to TCPS_ESTABLISHED. 356 * 357 * The ISNs are from the exchange of SYNs. 358 */ 359 void 360 make_established(struct toepcb *toep, uint32_t iss, uint32_t irs, uint16_t opt) 361 { 362 struct inpcb *inp = toep->inp; 363 struct socket *so = inp->inp_socket; 364 struct tcpcb *tp = intotcpcb(inp); 365 uint16_t tcpopt = be16toh(opt); 366 367 INP_WLOCK_ASSERT(inp); 368 KASSERT(tp->t_state == TCPS_SYN_SENT || 369 tp->t_state == TCPS_SYN_RECEIVED, 370 ("%s: TCP state %s", __func__, tcpstates[tp->t_state])); 371 372 CTR6(KTR_CXGBE, "%s: tid %d, so %p, inp %p, tp %p, toep %p", 373 __func__, toep->tid, so, inp, tp, toep); 374 375 tcp_state_change(tp, TCPS_ESTABLISHED); 376 tp->t_starttime = ticks; 377 TCPSTAT_INC(tcps_connects); 378 379 tp->irs = irs; 380 tcp_rcvseqinit(tp); 381 tp->rcv_wnd = (u_int)toep->params.opt0_bufsize << 10; 382 tp->rcv_adv += tp->rcv_wnd; 383 tp->last_ack_sent = tp->rcv_nxt; 384 385 tp->iss = iss; 386 tcp_sendseqinit(tp); 387 tp->snd_una = iss + 1; 388 tp->snd_nxt = iss + 1; 389 tp->snd_max = iss + 1; 390 391 assign_rxopt(tp, tcpopt); 392 send_flowc_wr(toep, tp); 393 394 soisconnected(so); 395 } 396 397 int 398 send_rx_credits(struct adapter *sc, struct toepcb *toep, int credits) 399 { 400 struct wrqe *wr; 401 struct cpl_rx_data_ack *req; 402 uint32_t dack = F_RX_DACK_CHANGE | V_RX_DACK_MODE(1); 403 404 KASSERT(credits >= 0, ("%s: %d credits", __func__, credits)); 405 406 wr = alloc_wrqe(sizeof(*req), toep->ctrlq); 407 if (wr == NULL) 408 return (0); 409 req = wrtod(wr); 410 411 INIT_TP_WR_MIT_CPL(req, CPL_RX_DATA_ACK, toep->tid); 412 req->credit_dack = htobe32(dack | V_RX_CREDITS(credits)); 413 414 t4_wrq_tx(sc, wr); 415 return (credits); 416 } 417 418 void 419 send_rx_modulate(struct adapter *sc, struct toepcb *toep) 420 { 421 struct wrqe *wr; 422 struct cpl_rx_data_ack *req; 423 424 wr = alloc_wrqe(sizeof(*req), toep->ctrlq); 425 if (wr == NULL) 426 return; 427 req = wrtod(wr); 428 429 INIT_TP_WR_MIT_CPL(req, CPL_RX_DATA_ACK, toep->tid); 430 req->credit_dack = htobe32(F_RX_MODULATE_RX); 431 432 t4_wrq_tx(sc, wr); 433 } 434 435 void 436 t4_rcvd_locked(struct toedev *tod, struct tcpcb *tp) 437 { 438 struct adapter *sc = tod->tod_softc; 439 struct inpcb *inp = tp->t_inpcb; 440 struct socket *so = inp->inp_socket; 441 struct sockbuf *sb = &so->so_rcv; 442 struct toepcb *toep = tp->t_toe; 443 int rx_credits; 444 445 INP_WLOCK_ASSERT(inp); 446 SOCKBUF_LOCK_ASSERT(sb); 447 448 rx_credits = sbspace(sb) > tp->rcv_wnd ? sbspace(sb) - tp->rcv_wnd : 0; 449 if (ulp_mode(toep) == ULP_MODE_TLS) { 450 if (toep->tls.rcv_over >= rx_credits) { 451 toep->tls.rcv_over -= rx_credits; 452 rx_credits = 0; 453 } else { 454 rx_credits -= toep->tls.rcv_over; 455 toep->tls.rcv_over = 0; 456 } 457 } 458 459 if (rx_credits > 0 && 460 (tp->rcv_wnd <= 32 * 1024 || rx_credits >= 64 * 1024 || 461 (rx_credits >= 16 * 1024 && tp->rcv_wnd <= 128 * 1024) || 462 sbused(sb) + tp->rcv_wnd < sb->sb_lowat)) { 463 rx_credits = send_rx_credits(sc, toep, rx_credits); 464 tp->rcv_wnd += rx_credits; 465 tp->rcv_adv += rx_credits; 466 } else if (toep->flags & TPF_FORCE_CREDITS) 467 send_rx_modulate(sc, toep); 468 } 469 470 void 471 t4_rcvd(struct toedev *tod, struct tcpcb *tp) 472 { 473 struct inpcb *inp = tp->t_inpcb; 474 struct socket *so = inp->inp_socket; 475 struct sockbuf *sb = &so->so_rcv; 476 477 SOCKBUF_LOCK(sb); 478 t4_rcvd_locked(tod, tp); 479 SOCKBUF_UNLOCK(sb); 480 } 481 482 /* 483 * Close a connection by sending a CPL_CLOSE_CON_REQ message. 484 */ 485 int 486 t4_close_conn(struct adapter *sc, struct toepcb *toep) 487 { 488 struct wrqe *wr; 489 struct cpl_close_con_req *req; 490 unsigned int tid = toep->tid; 491 492 CTR3(KTR_CXGBE, "%s: tid %u%s", __func__, toep->tid, 493 toep->flags & TPF_FIN_SENT ? ", IGNORED" : ""); 494 495 if (toep->flags & TPF_FIN_SENT) 496 return (0); 497 498 KASSERT(toep->flags & TPF_FLOWC_WR_SENT, 499 ("%s: flowc_wr not sent for tid %u.", __func__, tid)); 500 501 wr = alloc_wrqe(sizeof(*req), toep->ofld_txq); 502 if (wr == NULL) { 503 /* XXX */ 504 panic("%s: allocation failure.", __func__); 505 } 506 req = wrtod(wr); 507 508 req->wr.wr_hi = htonl(V_FW_WR_OP(FW_TP_WR) | 509 V_FW_WR_IMMDLEN(sizeof(*req) - sizeof(req->wr))); 510 req->wr.wr_mid = htonl(V_FW_WR_LEN16(howmany(sizeof(*req), 16)) | 511 V_FW_WR_FLOWID(tid)); 512 req->wr.wr_lo = cpu_to_be64(0); 513 OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_CLOSE_CON_REQ, tid)); 514 req->rsvd = 0; 515 516 toep->flags |= TPF_FIN_SENT; 517 toep->flags &= ~TPF_SEND_FIN; 518 t4_l2t_send(sc, wr, toep->l2te); 519 520 return (0); 521 } 522 523 #define MAX_OFLD_TX_CREDITS (SGE_MAX_WR_LEN / 16) 524 #define MIN_OFLD_TX_CREDITS (howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16)) 525 526 /* Maximum amount of immediate data we could stuff in a WR */ 527 static inline int 528 max_imm_payload(int tx_credits) 529 { 530 const int n = 1; /* Use no more than one desc for imm. data WR */ 531 532 KASSERT(tx_credits >= 0 && 533 tx_credits <= MAX_OFLD_TX_CREDITS, 534 ("%s: %d credits", __func__, tx_credits)); 535 536 if (tx_credits < MIN_OFLD_TX_CREDITS) 537 return (0); 538 539 if (tx_credits >= (n * EQ_ESIZE) / 16) 540 return ((n * EQ_ESIZE) - sizeof(struct fw_ofld_tx_data_wr)); 541 else 542 return (tx_credits * 16 - sizeof(struct fw_ofld_tx_data_wr)); 543 } 544 545 /* Maximum number of SGL entries we could stuff in a WR */ 546 static inline int 547 max_dsgl_nsegs(int tx_credits) 548 { 549 int nseg = 1; /* ulptx_sgl has room for 1, rest ulp_tx_sge_pair */ 550 int sge_pair_credits = tx_credits - MIN_OFLD_TX_CREDITS; 551 552 KASSERT(tx_credits >= 0 && 553 tx_credits <= MAX_OFLD_TX_CREDITS, 554 ("%s: %d credits", __func__, tx_credits)); 555 556 if (tx_credits < MIN_OFLD_TX_CREDITS) 557 return (0); 558 559 nseg += 2 * (sge_pair_credits * 16 / 24); 560 if ((sge_pair_credits * 16) % 24 == 16) 561 nseg++; 562 563 return (nseg); 564 } 565 566 static inline void 567 write_tx_wr(void *dst, struct toepcb *toep, unsigned int immdlen, 568 unsigned int plen, uint8_t credits, int shove, int ulp_submode) 569 { 570 struct fw_ofld_tx_data_wr *txwr = dst; 571 572 txwr->op_to_immdlen = htobe32(V_WR_OP(FW_OFLD_TX_DATA_WR) | 573 V_FW_WR_IMMDLEN(immdlen)); 574 txwr->flowid_len16 = htobe32(V_FW_WR_FLOWID(toep->tid) | 575 V_FW_WR_LEN16(credits)); 576 txwr->lsodisable_to_flags = htobe32(V_TX_ULP_MODE(ulp_mode(toep)) | 577 V_TX_ULP_SUBMODE(ulp_submode) | V_TX_URG(0) | V_TX_SHOVE(shove)); 578 txwr->plen = htobe32(plen); 579 580 if (toep->params.tx_align > 0) { 581 if (plen < 2 * toep->params.emss) 582 txwr->lsodisable_to_flags |= 583 htobe32(F_FW_OFLD_TX_DATA_WR_LSODISABLE); 584 else 585 txwr->lsodisable_to_flags |= 586 htobe32(F_FW_OFLD_TX_DATA_WR_ALIGNPLD | 587 (toep->params.nagle == 0 ? 0 : 588 F_FW_OFLD_TX_DATA_WR_ALIGNPLDSHOVE)); 589 } 590 } 591 592 /* 593 * Generate a DSGL from a starting mbuf. The total number of segments and the 594 * maximum segments in any one mbuf are provided. 595 */ 596 static void 597 write_tx_sgl(void *dst, struct mbuf *start, struct mbuf *stop, int nsegs, int n) 598 { 599 struct mbuf *m; 600 struct ulptx_sgl *usgl = dst; 601 int i, j, rc; 602 struct sglist sg; 603 struct sglist_seg segs[n]; 604 605 KASSERT(nsegs > 0, ("%s: nsegs 0", __func__)); 606 607 sglist_init(&sg, n, segs); 608 usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) | 609 V_ULPTX_NSGE(nsegs)); 610 611 i = -1; 612 for (m = start; m != stop; m = m->m_next) { 613 if (m->m_flags & M_EXTPG) 614 rc = sglist_append_mbuf_epg(&sg, m, 615 mtod(m, vm_offset_t), m->m_len); 616 else 617 rc = sglist_append(&sg, mtod(m, void *), m->m_len); 618 if (__predict_false(rc != 0)) 619 panic("%s: sglist_append %d", __func__, rc); 620 621 for (j = 0; j < sg.sg_nseg; i++, j++) { 622 if (i < 0) { 623 usgl->len0 = htobe32(segs[j].ss_len); 624 usgl->addr0 = htobe64(segs[j].ss_paddr); 625 } else { 626 usgl->sge[i / 2].len[i & 1] = 627 htobe32(segs[j].ss_len); 628 usgl->sge[i / 2].addr[i & 1] = 629 htobe64(segs[j].ss_paddr); 630 } 631 #ifdef INVARIANTS 632 nsegs--; 633 #endif 634 } 635 sglist_reset(&sg); 636 } 637 if (i & 1) 638 usgl->sge[i / 2].len[1] = htobe32(0); 639 KASSERT(nsegs == 0, ("%s: nsegs %d, start %p, stop %p", 640 __func__, nsegs, start, stop)); 641 } 642 643 /* 644 * Max number of SGL entries an offload tx work request can have. This is 41 645 * (1 + 40) for a full 512B work request. 646 * fw_ofld_tx_data_wr(16B) + ulptx_sgl(16B, 1) + ulptx_sge_pair(480B, 40) 647 */ 648 #define OFLD_SGL_LEN (41) 649 650 /* 651 * Send data and/or a FIN to the peer. 652 * 653 * The socket's so_snd buffer consists of a stream of data starting with sb_mb 654 * and linked together with m_next. sb_sndptr, if set, is the last mbuf that 655 * was transmitted. 656 * 657 * drop indicates the number of bytes that should be dropped from the head of 658 * the send buffer. It is an optimization that lets do_fw4_ack avoid creating 659 * contention on the send buffer lock (before this change it used to do 660 * sowwakeup and then t4_push_frames right after that when recovering from tx 661 * stalls). When drop is set this function MUST drop the bytes and wake up any 662 * writers. 663 */ 664 void 665 t4_push_frames(struct adapter *sc, struct toepcb *toep, int drop) 666 { 667 struct mbuf *sndptr, *m, *sb_sndptr; 668 struct fw_ofld_tx_data_wr *txwr; 669 struct wrqe *wr; 670 u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf; 671 struct inpcb *inp = toep->inp; 672 struct tcpcb *tp = intotcpcb(inp); 673 struct socket *so = inp->inp_socket; 674 struct sockbuf *sb = &so->so_snd; 675 int tx_credits, shove, compl, sowwakeup; 676 struct ofld_tx_sdesc *txsd; 677 bool nomap_mbuf_seen; 678 679 INP_WLOCK_ASSERT(inp); 680 KASSERT(toep->flags & TPF_FLOWC_WR_SENT, 681 ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid)); 682 683 KASSERT(ulp_mode(toep) == ULP_MODE_NONE || 684 ulp_mode(toep) == ULP_MODE_TCPDDP || 685 ulp_mode(toep) == ULP_MODE_TLS || 686 ulp_mode(toep) == ULP_MODE_RDMA, 687 ("%s: ulp_mode %u for toep %p", __func__, ulp_mode(toep), toep)); 688 689 #ifdef VERBOSE_TRACES 690 CTR5(KTR_CXGBE, "%s: tid %d toep flags %#x tp flags %#x drop %d", 691 __func__, toep->tid, toep->flags, tp->t_flags, drop); 692 #endif 693 if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN)) 694 return; 695 696 #ifdef RATELIMIT 697 if (__predict_false(inp->inp_flags2 & INP_RATE_LIMIT_CHANGED) && 698 (update_tx_rate_limit(sc, toep, so->so_max_pacing_rate) == 0)) { 699 inp->inp_flags2 &= ~INP_RATE_LIMIT_CHANGED; 700 } 701 #endif 702 703 /* 704 * This function doesn't resume by itself. Someone else must clear the 705 * flag and call this function. 706 */ 707 if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) { 708 KASSERT(drop == 0, 709 ("%s: drop (%d) != 0 but tx is suspended", __func__, drop)); 710 return; 711 } 712 713 txsd = &toep->txsd[toep->txsd_pidx]; 714 do { 715 tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS); 716 max_imm = max_imm_payload(tx_credits); 717 max_nsegs = max_dsgl_nsegs(tx_credits); 718 719 SOCKBUF_LOCK(sb); 720 sowwakeup = drop; 721 if (drop) { 722 sbdrop_locked(sb, drop); 723 drop = 0; 724 } 725 sb_sndptr = sb->sb_sndptr; 726 sndptr = sb_sndptr ? sb_sndptr->m_next : sb->sb_mb; 727 plen = 0; 728 nsegs = 0; 729 max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */ 730 nomap_mbuf_seen = false; 731 for (m = sndptr; m != NULL; m = m->m_next) { 732 int n; 733 734 if (m->m_flags & M_EXTPG) { 735 #ifdef KERN_TLS 736 if (m->m_epg_tls != NULL) { 737 toep->flags |= TPF_KTLS; 738 if (plen == 0) { 739 SOCKBUF_UNLOCK(sb); 740 t4_push_ktls(sc, toep, 0); 741 return; 742 } 743 break; 744 } 745 #endif 746 n = sglist_count_mbuf_epg(m, 747 mtod(m, vm_offset_t), m->m_len); 748 } else 749 n = sglist_count(mtod(m, void *), m->m_len); 750 751 nsegs += n; 752 plen += m->m_len; 753 754 /* This mbuf sent us _over_ the nsegs limit, back out */ 755 if (plen > max_imm && nsegs > max_nsegs) { 756 nsegs -= n; 757 plen -= m->m_len; 758 if (plen == 0) { 759 /* Too few credits */ 760 toep->flags |= TPF_TX_SUSPENDED; 761 if (sowwakeup) { 762 if (!TAILQ_EMPTY( 763 &toep->aiotx_jobq)) 764 t4_aiotx_queue_toep(so, 765 toep); 766 sowwakeup_locked(so); 767 } else 768 SOCKBUF_UNLOCK(sb); 769 SOCKBUF_UNLOCK_ASSERT(sb); 770 return; 771 } 772 break; 773 } 774 775 if (m->m_flags & M_EXTPG) 776 nomap_mbuf_seen = true; 777 if (max_nsegs_1mbuf < n) 778 max_nsegs_1mbuf = n; 779 sb_sndptr = m; /* new sb->sb_sndptr if all goes well */ 780 781 /* This mbuf put us right at the max_nsegs limit */ 782 if (plen > max_imm && nsegs == max_nsegs) { 783 m = m->m_next; 784 break; 785 } 786 } 787 788 if (sbused(sb) > sb->sb_hiwat * 5 / 8 && 789 toep->plen_nocompl + plen >= sb->sb_hiwat / 4) 790 compl = 1; 791 else 792 compl = 0; 793 794 if (sb->sb_flags & SB_AUTOSIZE && 795 V_tcp_do_autosndbuf && 796 sb->sb_hiwat < V_tcp_autosndbuf_max && 797 sbused(sb) >= sb->sb_hiwat * 7 / 8) { 798 int newsize = min(sb->sb_hiwat + V_tcp_autosndbuf_inc, 799 V_tcp_autosndbuf_max); 800 801 if (!sbreserve_locked(sb, newsize, so, NULL)) 802 sb->sb_flags &= ~SB_AUTOSIZE; 803 else 804 sowwakeup = 1; /* room available */ 805 } 806 if (sowwakeup) { 807 if (!TAILQ_EMPTY(&toep->aiotx_jobq)) 808 t4_aiotx_queue_toep(so, toep); 809 sowwakeup_locked(so); 810 } else 811 SOCKBUF_UNLOCK(sb); 812 SOCKBUF_UNLOCK_ASSERT(sb); 813 814 /* nothing to send */ 815 if (plen == 0) { 816 KASSERT(m == NULL, 817 ("%s: nothing to send, but m != NULL", __func__)); 818 break; 819 } 820 821 if (__predict_false(toep->flags & TPF_FIN_SENT)) 822 panic("%s: excess tx.", __func__); 823 824 shove = m == NULL && !(tp->t_flags & TF_MORETOCOME); 825 if (plen <= max_imm && !nomap_mbuf_seen) { 826 827 /* Immediate data tx */ 828 829 wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16), 830 toep->ofld_txq); 831 if (wr == NULL) { 832 /* XXX: how will we recover from this? */ 833 toep->flags |= TPF_TX_SUSPENDED; 834 return; 835 } 836 txwr = wrtod(wr); 837 credits = howmany(wr->wr_len, 16); 838 write_tx_wr(txwr, toep, plen, plen, credits, shove, 0); 839 m_copydata(sndptr, 0, plen, (void *)(txwr + 1)); 840 nsegs = 0; 841 } else { 842 int wr_len; 843 844 /* DSGL tx */ 845 846 wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) + 847 ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8; 848 wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq); 849 if (wr == NULL) { 850 /* XXX: how will we recover from this? */ 851 toep->flags |= TPF_TX_SUSPENDED; 852 return; 853 } 854 txwr = wrtod(wr); 855 credits = howmany(wr_len, 16); 856 write_tx_wr(txwr, toep, 0, plen, credits, shove, 0); 857 write_tx_sgl(txwr + 1, sndptr, m, nsegs, 858 max_nsegs_1mbuf); 859 if (wr_len & 0xf) { 860 uint64_t *pad = (uint64_t *) 861 ((uintptr_t)txwr + wr_len); 862 *pad = 0; 863 } 864 } 865 866 KASSERT(toep->tx_credits >= credits, 867 ("%s: not enough credits", __func__)); 868 869 toep->tx_credits -= credits; 870 toep->tx_nocompl += credits; 871 toep->plen_nocompl += plen; 872 if (toep->tx_credits <= toep->tx_total * 3 / 8 && 873 toep->tx_nocompl >= toep->tx_total / 4) 874 compl = 1; 875 876 if (compl || ulp_mode(toep) == ULP_MODE_RDMA) { 877 txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL); 878 toep->tx_nocompl = 0; 879 toep->plen_nocompl = 0; 880 } 881 882 tp->snd_nxt += plen; 883 tp->snd_max += plen; 884 885 SOCKBUF_LOCK(sb); 886 KASSERT(sb_sndptr, ("%s: sb_sndptr is NULL", __func__)); 887 sb->sb_sndptr = sb_sndptr; 888 SOCKBUF_UNLOCK(sb); 889 890 toep->flags |= TPF_TX_DATA_SENT; 891 if (toep->tx_credits < MIN_OFLD_TX_CREDITS) 892 toep->flags |= TPF_TX_SUSPENDED; 893 894 KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__)); 895 txsd->plen = plen; 896 txsd->tx_credits = credits; 897 txsd++; 898 if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) { 899 toep->txsd_pidx = 0; 900 txsd = &toep->txsd[0]; 901 } 902 toep->txsd_avail--; 903 904 t4_l2t_send(sc, wr, toep->l2te); 905 } while (m != NULL); 906 907 /* Send a FIN if requested, but only if there's no more data to send */ 908 if (m == NULL && toep->flags & TPF_SEND_FIN) 909 t4_close_conn(sc, toep); 910 } 911 912 static inline void 913 rqdrop_locked(struct mbufq *q, int plen) 914 { 915 struct mbuf *m; 916 917 while (plen > 0) { 918 m = mbufq_dequeue(q); 919 920 /* Too many credits. */ 921 MPASS(m != NULL); 922 M_ASSERTPKTHDR(m); 923 924 /* Partial credits. */ 925 MPASS(plen >= m->m_pkthdr.len); 926 927 plen -= m->m_pkthdr.len; 928 m_freem(m); 929 } 930 } 931 932 void 933 t4_push_pdus(struct adapter *sc, struct toepcb *toep, int drop) 934 { 935 struct mbuf *sndptr, *m; 936 struct fw_ofld_tx_data_wr *txwr; 937 struct wrqe *wr; 938 u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf; 939 u_int adjusted_plen, ulp_submode; 940 struct inpcb *inp = toep->inp; 941 struct tcpcb *tp = intotcpcb(inp); 942 int tx_credits, shove; 943 struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx]; 944 struct mbufq *pduq = &toep->ulp_pduq; 945 static const u_int ulp_extra_len[] = {0, 4, 4, 8}; 946 947 INP_WLOCK_ASSERT(inp); 948 KASSERT(toep->flags & TPF_FLOWC_WR_SENT, 949 ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid)); 950 KASSERT(ulp_mode(toep) == ULP_MODE_ISCSI, 951 ("%s: ulp_mode %u for toep %p", __func__, ulp_mode(toep), toep)); 952 953 if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN)) 954 return; 955 956 /* 957 * This function doesn't resume by itself. Someone else must clear the 958 * flag and call this function. 959 */ 960 if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) { 961 KASSERT(drop == 0, 962 ("%s: drop (%d) != 0 but tx is suspended", __func__, drop)); 963 return; 964 } 965 966 if (drop) 967 rqdrop_locked(&toep->ulp_pdu_reclaimq, drop); 968 969 while ((sndptr = mbufq_first(pduq)) != NULL) { 970 M_ASSERTPKTHDR(sndptr); 971 972 tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS); 973 max_imm = max_imm_payload(tx_credits); 974 max_nsegs = max_dsgl_nsegs(tx_credits); 975 976 plen = 0; 977 nsegs = 0; 978 max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */ 979 for (m = sndptr; m != NULL; m = m->m_next) { 980 int n = sglist_count(mtod(m, void *), m->m_len); 981 982 nsegs += n; 983 plen += m->m_len; 984 985 /* 986 * This mbuf would send us _over_ the nsegs limit. 987 * Suspend tx because the PDU can't be sent out. 988 */ 989 if (plen > max_imm && nsegs > max_nsegs) { 990 toep->flags |= TPF_TX_SUSPENDED; 991 return; 992 } 993 994 if (max_nsegs_1mbuf < n) 995 max_nsegs_1mbuf = n; 996 } 997 998 if (__predict_false(toep->flags & TPF_FIN_SENT)) 999 panic("%s: excess tx.", __func__); 1000 1001 /* 1002 * We have a PDU to send. All of it goes out in one WR so 'm' 1003 * is NULL. A PDU's length is always a multiple of 4. 1004 */ 1005 MPASS(m == NULL); 1006 MPASS((plen & 3) == 0); 1007 MPASS(sndptr->m_pkthdr.len == plen); 1008 1009 shove = !(tp->t_flags & TF_MORETOCOME); 1010 ulp_submode = mbuf_ulp_submode(sndptr); 1011 MPASS(ulp_submode < nitems(ulp_extra_len)); 1012 1013 /* 1014 * plen doesn't include header and data digests, which are 1015 * generated and inserted in the right places by the TOE, but 1016 * they do occupy TCP sequence space and need to be accounted 1017 * for. 1018 */ 1019 adjusted_plen = plen + ulp_extra_len[ulp_submode]; 1020 if (plen <= max_imm) { 1021 1022 /* Immediate data tx */ 1023 1024 wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16), 1025 toep->ofld_txq); 1026 if (wr == NULL) { 1027 /* XXX: how will we recover from this? */ 1028 toep->flags |= TPF_TX_SUSPENDED; 1029 return; 1030 } 1031 txwr = wrtod(wr); 1032 credits = howmany(wr->wr_len, 16); 1033 write_tx_wr(txwr, toep, plen, adjusted_plen, credits, 1034 shove, ulp_submode); 1035 m_copydata(sndptr, 0, plen, (void *)(txwr + 1)); 1036 nsegs = 0; 1037 } else { 1038 int wr_len; 1039 1040 /* DSGL tx */ 1041 wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) + 1042 ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8; 1043 wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq); 1044 if (wr == NULL) { 1045 /* XXX: how will we recover from this? */ 1046 toep->flags |= TPF_TX_SUSPENDED; 1047 return; 1048 } 1049 txwr = wrtod(wr); 1050 credits = howmany(wr_len, 16); 1051 write_tx_wr(txwr, toep, 0, adjusted_plen, credits, 1052 shove, ulp_submode); 1053 write_tx_sgl(txwr + 1, sndptr, m, nsegs, 1054 max_nsegs_1mbuf); 1055 if (wr_len & 0xf) { 1056 uint64_t *pad = (uint64_t *) 1057 ((uintptr_t)txwr + wr_len); 1058 *pad = 0; 1059 } 1060 } 1061 1062 KASSERT(toep->tx_credits >= credits, 1063 ("%s: not enough credits", __func__)); 1064 1065 m = mbufq_dequeue(pduq); 1066 MPASS(m == sndptr); 1067 mbufq_enqueue(&toep->ulp_pdu_reclaimq, m); 1068 1069 toep->tx_credits -= credits; 1070 toep->tx_nocompl += credits; 1071 toep->plen_nocompl += plen; 1072 if (toep->tx_credits <= toep->tx_total * 3 / 8 && 1073 toep->tx_nocompl >= toep->tx_total / 4) { 1074 txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL); 1075 toep->tx_nocompl = 0; 1076 toep->plen_nocompl = 0; 1077 } 1078 1079 tp->snd_nxt += adjusted_plen; 1080 tp->snd_max += adjusted_plen; 1081 1082 toep->flags |= TPF_TX_DATA_SENT; 1083 if (toep->tx_credits < MIN_OFLD_TX_CREDITS) 1084 toep->flags |= TPF_TX_SUSPENDED; 1085 1086 KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__)); 1087 txsd->plen = plen; 1088 txsd->tx_credits = credits; 1089 txsd++; 1090 if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) { 1091 toep->txsd_pidx = 0; 1092 txsd = &toep->txsd[0]; 1093 } 1094 toep->txsd_avail--; 1095 1096 t4_l2t_send(sc, wr, toep->l2te); 1097 } 1098 1099 /* Send a FIN if requested, but only if there are no more PDUs to send */ 1100 if (mbufq_first(pduq) == NULL && toep->flags & TPF_SEND_FIN) 1101 t4_close_conn(sc, toep); 1102 } 1103 1104 static inline void 1105 t4_push_data(struct adapter *sc, struct toepcb *toep, int drop) 1106 { 1107 1108 if (ulp_mode(toep) == ULP_MODE_ISCSI) 1109 t4_push_pdus(sc, toep, drop); 1110 else if (tls_tx_key(toep) && toep->tls.mode == TLS_MODE_TLSOM) 1111 t4_push_tls_records(sc, toep, drop); 1112 #ifdef KERN_TLS 1113 else if (toep->flags & TPF_KTLS) 1114 t4_push_ktls(sc, toep, drop); 1115 #endif 1116 else 1117 t4_push_frames(sc, toep, drop); 1118 } 1119 1120 int 1121 t4_tod_output(struct toedev *tod, struct tcpcb *tp) 1122 { 1123 struct adapter *sc = tod->tod_softc; 1124 #ifdef INVARIANTS 1125 struct inpcb *inp = tp->t_inpcb; 1126 #endif 1127 struct toepcb *toep = tp->t_toe; 1128 1129 INP_WLOCK_ASSERT(inp); 1130 KASSERT((inp->inp_flags & INP_DROPPED) == 0, 1131 ("%s: inp %p dropped.", __func__, inp)); 1132 KASSERT(toep != NULL, ("%s: toep is NULL", __func__)); 1133 1134 t4_push_data(sc, toep, 0); 1135 1136 return (0); 1137 } 1138 1139 int 1140 t4_send_fin(struct toedev *tod, struct tcpcb *tp) 1141 { 1142 struct adapter *sc = tod->tod_softc; 1143 #ifdef INVARIANTS 1144 struct inpcb *inp = tp->t_inpcb; 1145 #endif 1146 struct toepcb *toep = tp->t_toe; 1147 1148 INP_WLOCK_ASSERT(inp); 1149 KASSERT((inp->inp_flags & INP_DROPPED) == 0, 1150 ("%s: inp %p dropped.", __func__, inp)); 1151 KASSERT(toep != NULL, ("%s: toep is NULL", __func__)); 1152 1153 toep->flags |= TPF_SEND_FIN; 1154 if (tp->t_state >= TCPS_ESTABLISHED) 1155 t4_push_data(sc, toep, 0); 1156 1157 return (0); 1158 } 1159 1160 int 1161 t4_send_rst(struct toedev *tod, struct tcpcb *tp) 1162 { 1163 struct adapter *sc = tod->tod_softc; 1164 #if defined(INVARIANTS) 1165 struct inpcb *inp = tp->t_inpcb; 1166 #endif 1167 struct toepcb *toep = tp->t_toe; 1168 1169 INP_WLOCK_ASSERT(inp); 1170 KASSERT((inp->inp_flags & INP_DROPPED) == 0, 1171 ("%s: inp %p dropped.", __func__, inp)); 1172 KASSERT(toep != NULL, ("%s: toep is NULL", __func__)); 1173 1174 /* hmmmm */ 1175 KASSERT(toep->flags & TPF_FLOWC_WR_SENT, 1176 ("%s: flowc for tid %u [%s] not sent already", 1177 __func__, toep->tid, tcpstates[tp->t_state])); 1178 1179 send_reset(sc, toep, 0); 1180 return (0); 1181 } 1182 1183 /* 1184 * Peer has sent us a FIN. 1185 */ 1186 static int 1187 do_peer_close(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) 1188 { 1189 struct adapter *sc = iq->adapter; 1190 const struct cpl_peer_close *cpl = (const void *)(rss + 1); 1191 unsigned int tid = GET_TID(cpl); 1192 struct toepcb *toep = lookup_tid(sc, tid); 1193 struct inpcb *inp = toep->inp; 1194 struct tcpcb *tp = NULL; 1195 struct socket *so; 1196 struct epoch_tracker et; 1197 #ifdef INVARIANTS 1198 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); 1199 #endif 1200 1201 KASSERT(opcode == CPL_PEER_CLOSE, 1202 ("%s: unexpected opcode 0x%x", __func__, opcode)); 1203 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 1204 1205 if (__predict_false(toep->flags & TPF_SYNQE)) { 1206 /* 1207 * do_pass_establish must have run before do_peer_close and if 1208 * this is still a synqe instead of a toepcb then the connection 1209 * must be getting aborted. 1210 */ 1211 MPASS(toep->flags & TPF_ABORT_SHUTDOWN); 1212 CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid, 1213 toep, toep->flags); 1214 return (0); 1215 } 1216 1217 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__)); 1218 1219 CURVNET_SET(toep->vnet); 1220 NET_EPOCH_ENTER(et); 1221 INP_WLOCK(inp); 1222 tp = intotcpcb(inp); 1223 1224 CTR6(KTR_CXGBE, 1225 "%s: tid %u (%s), toep_flags 0x%x, ddp_flags 0x%x, inp %p", 1226 __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags, 1227 toep->ddp.flags, inp); 1228 1229 if (toep->flags & TPF_ABORT_SHUTDOWN) 1230 goto done; 1231 1232 tp->rcv_nxt++; /* FIN */ 1233 1234 so = inp->inp_socket; 1235 socantrcvmore(so); 1236 if (ulp_mode(toep) == ULP_MODE_TCPDDP) { 1237 DDP_LOCK(toep); 1238 if (__predict_false(toep->ddp.flags & 1239 (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE))) 1240 handle_ddp_close(toep, tp, cpl->rcv_nxt); 1241 DDP_UNLOCK(toep); 1242 } 1243 1244 if (ulp_mode(toep) != ULP_MODE_RDMA) { 1245 KASSERT(tp->rcv_nxt == be32toh(cpl->rcv_nxt), 1246 ("%s: rcv_nxt mismatch: %u %u", __func__, tp->rcv_nxt, 1247 be32toh(cpl->rcv_nxt))); 1248 } 1249 1250 switch (tp->t_state) { 1251 case TCPS_SYN_RECEIVED: 1252 tp->t_starttime = ticks; 1253 /* FALLTHROUGH */ 1254 1255 case TCPS_ESTABLISHED: 1256 tcp_state_change(tp, TCPS_CLOSE_WAIT); 1257 break; 1258 1259 case TCPS_FIN_WAIT_1: 1260 tcp_state_change(tp, TCPS_CLOSING); 1261 break; 1262 1263 case TCPS_FIN_WAIT_2: 1264 tcp_twstart(tp); 1265 INP_UNLOCK_ASSERT(inp); /* safe, we have a ref on the inp */ 1266 NET_EPOCH_EXIT(et); 1267 CURVNET_RESTORE(); 1268 1269 INP_WLOCK(inp); 1270 final_cpl_received(toep); 1271 return (0); 1272 1273 default: 1274 log(LOG_ERR, "%s: TID %u received CPL_PEER_CLOSE in state %d\n", 1275 __func__, tid, tp->t_state); 1276 } 1277 done: 1278 INP_WUNLOCK(inp); 1279 NET_EPOCH_EXIT(et); 1280 CURVNET_RESTORE(); 1281 return (0); 1282 } 1283 1284 /* 1285 * Peer has ACK'd our FIN. 1286 */ 1287 static int 1288 do_close_con_rpl(struct sge_iq *iq, const struct rss_header *rss, 1289 struct mbuf *m) 1290 { 1291 struct adapter *sc = iq->adapter; 1292 const struct cpl_close_con_rpl *cpl = (const void *)(rss + 1); 1293 unsigned int tid = GET_TID(cpl); 1294 struct toepcb *toep = lookup_tid(sc, tid); 1295 struct inpcb *inp = toep->inp; 1296 struct tcpcb *tp = NULL; 1297 struct socket *so = NULL; 1298 struct epoch_tracker et; 1299 #ifdef INVARIANTS 1300 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); 1301 #endif 1302 1303 KASSERT(opcode == CPL_CLOSE_CON_RPL, 1304 ("%s: unexpected opcode 0x%x", __func__, opcode)); 1305 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 1306 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__)); 1307 1308 CURVNET_SET(toep->vnet); 1309 NET_EPOCH_ENTER(et); 1310 INP_WLOCK(inp); 1311 tp = intotcpcb(inp); 1312 1313 CTR4(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x", 1314 __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags); 1315 1316 if (toep->flags & TPF_ABORT_SHUTDOWN) 1317 goto done; 1318 1319 so = inp->inp_socket; 1320 tp->snd_una = be32toh(cpl->snd_nxt) - 1; /* exclude FIN */ 1321 1322 switch (tp->t_state) { 1323 case TCPS_CLOSING: /* see TCPS_FIN_WAIT_2 in do_peer_close too */ 1324 tcp_twstart(tp); 1325 release: 1326 INP_UNLOCK_ASSERT(inp); /* safe, we have a ref on the inp */ 1327 NET_EPOCH_EXIT(et); 1328 CURVNET_RESTORE(); 1329 1330 INP_WLOCK(inp); 1331 final_cpl_received(toep); /* no more CPLs expected */ 1332 1333 return (0); 1334 case TCPS_LAST_ACK: 1335 if (tcp_close(tp)) 1336 INP_WUNLOCK(inp); 1337 goto release; 1338 1339 case TCPS_FIN_WAIT_1: 1340 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) 1341 soisdisconnected(so); 1342 tcp_state_change(tp, TCPS_FIN_WAIT_2); 1343 break; 1344 1345 default: 1346 log(LOG_ERR, 1347 "%s: TID %u received CPL_CLOSE_CON_RPL in state %s\n", 1348 __func__, tid, tcpstates[tp->t_state]); 1349 } 1350 done: 1351 INP_WUNLOCK(inp); 1352 NET_EPOCH_EXIT(et); 1353 CURVNET_RESTORE(); 1354 return (0); 1355 } 1356 1357 void 1358 send_abort_rpl(struct adapter *sc, struct sge_wrq *ofld_txq, int tid, 1359 int rst_status) 1360 { 1361 struct wrqe *wr; 1362 struct cpl_abort_rpl *cpl; 1363 1364 wr = alloc_wrqe(sizeof(*cpl), ofld_txq); 1365 if (wr == NULL) { 1366 /* XXX */ 1367 panic("%s: allocation failure.", __func__); 1368 } 1369 cpl = wrtod(wr); 1370 1371 INIT_TP_WR_MIT_CPL(cpl, CPL_ABORT_RPL, tid); 1372 cpl->cmd = rst_status; 1373 1374 t4_wrq_tx(sc, wr); 1375 } 1376 1377 static int 1378 abort_status_to_errno(struct tcpcb *tp, unsigned int abort_reason) 1379 { 1380 switch (abort_reason) { 1381 case CPL_ERR_BAD_SYN: 1382 case CPL_ERR_CONN_RESET: 1383 return (tp->t_state == TCPS_CLOSE_WAIT ? EPIPE : ECONNRESET); 1384 case CPL_ERR_XMIT_TIMEDOUT: 1385 case CPL_ERR_PERSIST_TIMEDOUT: 1386 case CPL_ERR_FINWAIT2_TIMEDOUT: 1387 case CPL_ERR_KEEPALIVE_TIMEDOUT: 1388 return (ETIMEDOUT); 1389 default: 1390 return (EIO); 1391 } 1392 } 1393 1394 /* 1395 * TCP RST from the peer, timeout, or some other such critical error. 1396 */ 1397 static int 1398 do_abort_req(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) 1399 { 1400 struct adapter *sc = iq->adapter; 1401 const struct cpl_abort_req_rss *cpl = (const void *)(rss + 1); 1402 unsigned int tid = GET_TID(cpl); 1403 struct toepcb *toep = lookup_tid(sc, tid); 1404 struct sge_wrq *ofld_txq = toep->ofld_txq; 1405 struct inpcb *inp; 1406 struct tcpcb *tp; 1407 struct epoch_tracker et; 1408 #ifdef INVARIANTS 1409 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); 1410 #endif 1411 1412 KASSERT(opcode == CPL_ABORT_REQ_RSS, 1413 ("%s: unexpected opcode 0x%x", __func__, opcode)); 1414 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 1415 1416 if (toep->flags & TPF_SYNQE) 1417 return (do_abort_req_synqe(iq, rss, m)); 1418 1419 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__)); 1420 1421 if (negative_advice(cpl->status)) { 1422 CTR4(KTR_CXGBE, "%s: negative advice %d for tid %d (0x%x)", 1423 __func__, cpl->status, tid, toep->flags); 1424 return (0); /* Ignore negative advice */ 1425 } 1426 1427 inp = toep->inp; 1428 CURVNET_SET(toep->vnet); 1429 NET_EPOCH_ENTER(et); /* for tcp_close */ 1430 INP_WLOCK(inp); 1431 1432 tp = intotcpcb(inp); 1433 1434 CTR6(KTR_CXGBE, 1435 "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x, status %d", 1436 __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags, 1437 inp->inp_flags, cpl->status); 1438 1439 /* 1440 * If we'd initiated an abort earlier the reply to it is responsible for 1441 * cleaning up resources. Otherwise we tear everything down right here 1442 * right now. We owe the T4 a CPL_ABORT_RPL no matter what. 1443 */ 1444 if (toep->flags & TPF_ABORT_SHUTDOWN) { 1445 INP_WUNLOCK(inp); 1446 goto done; 1447 } 1448 toep->flags |= TPF_ABORT_SHUTDOWN; 1449 1450 if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) { 1451 struct socket *so = inp->inp_socket; 1452 1453 if (so != NULL) 1454 so_error_set(so, abort_status_to_errno(tp, 1455 cpl->status)); 1456 tp = tcp_close(tp); 1457 if (tp == NULL) 1458 INP_WLOCK(inp); /* re-acquire */ 1459 } 1460 1461 final_cpl_received(toep); 1462 done: 1463 NET_EPOCH_EXIT(et); 1464 CURVNET_RESTORE(); 1465 send_abort_rpl(sc, ofld_txq, tid, CPL_ABORT_NO_RST); 1466 return (0); 1467 } 1468 1469 /* 1470 * Reply to the CPL_ABORT_REQ (send_reset) 1471 */ 1472 static int 1473 do_abort_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) 1474 { 1475 struct adapter *sc = iq->adapter; 1476 const struct cpl_abort_rpl_rss *cpl = (const void *)(rss + 1); 1477 unsigned int tid = GET_TID(cpl); 1478 struct toepcb *toep = lookup_tid(sc, tid); 1479 struct inpcb *inp = toep->inp; 1480 #ifdef INVARIANTS 1481 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); 1482 #endif 1483 1484 KASSERT(opcode == CPL_ABORT_RPL_RSS, 1485 ("%s: unexpected opcode 0x%x", __func__, opcode)); 1486 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 1487 1488 if (toep->flags & TPF_SYNQE) 1489 return (do_abort_rpl_synqe(iq, rss, m)); 1490 1491 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__)); 1492 1493 CTR5(KTR_CXGBE, "%s: tid %u, toep %p, inp %p, status %d", 1494 __func__, tid, toep, inp, cpl->status); 1495 1496 KASSERT(toep->flags & TPF_ABORT_SHUTDOWN, 1497 ("%s: wasn't expecting abort reply", __func__)); 1498 1499 INP_WLOCK(inp); 1500 final_cpl_received(toep); 1501 1502 return (0); 1503 } 1504 1505 static int 1506 do_rx_data(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) 1507 { 1508 struct adapter *sc = iq->adapter; 1509 const struct cpl_rx_data *cpl = mtod(m, const void *); 1510 unsigned int tid = GET_TID(cpl); 1511 struct toepcb *toep = lookup_tid(sc, tid); 1512 struct inpcb *inp = toep->inp; 1513 struct tcpcb *tp; 1514 struct socket *so; 1515 struct sockbuf *sb; 1516 struct epoch_tracker et; 1517 int len, rx_credits; 1518 uint32_t ddp_placed = 0; 1519 1520 if (__predict_false(toep->flags & TPF_SYNQE)) { 1521 /* 1522 * do_pass_establish must have run before do_rx_data and if this 1523 * is still a synqe instead of a toepcb then the connection must 1524 * be getting aborted. 1525 */ 1526 MPASS(toep->flags & TPF_ABORT_SHUTDOWN); 1527 CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid, 1528 toep, toep->flags); 1529 m_freem(m); 1530 return (0); 1531 } 1532 1533 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__)); 1534 1535 /* strip off CPL header */ 1536 m_adj(m, sizeof(*cpl)); 1537 len = m->m_pkthdr.len; 1538 1539 INP_WLOCK(inp); 1540 if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) { 1541 CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x", 1542 __func__, tid, len, inp->inp_flags); 1543 INP_WUNLOCK(inp); 1544 m_freem(m); 1545 return (0); 1546 } 1547 1548 tp = intotcpcb(inp); 1549 1550 if (__predict_false(tp->rcv_nxt != be32toh(cpl->seq))) 1551 ddp_placed = be32toh(cpl->seq) - tp->rcv_nxt; 1552 1553 tp->rcv_nxt += len; 1554 if (tp->rcv_wnd < len) { 1555 KASSERT(ulp_mode(toep) == ULP_MODE_RDMA, 1556 ("%s: negative window size", __func__)); 1557 } 1558 1559 tp->rcv_wnd -= len; 1560 tp->t_rcvtime = ticks; 1561 1562 if (ulp_mode(toep) == ULP_MODE_TCPDDP) 1563 DDP_LOCK(toep); 1564 so = inp_inpcbtosocket(inp); 1565 sb = &so->so_rcv; 1566 SOCKBUF_LOCK(sb); 1567 1568 if (__predict_false(sb->sb_state & SBS_CANTRCVMORE)) { 1569 CTR3(KTR_CXGBE, "%s: tid %u, excess rx (%d bytes)", 1570 __func__, tid, len); 1571 m_freem(m); 1572 SOCKBUF_UNLOCK(sb); 1573 if (ulp_mode(toep) == ULP_MODE_TCPDDP) 1574 DDP_UNLOCK(toep); 1575 INP_WUNLOCK(inp); 1576 1577 CURVNET_SET(toep->vnet); 1578 NET_EPOCH_ENTER(et); 1579 INP_WLOCK(inp); 1580 tp = tcp_drop(tp, ECONNRESET); 1581 if (tp) 1582 INP_WUNLOCK(inp); 1583 NET_EPOCH_EXIT(et); 1584 CURVNET_RESTORE(); 1585 1586 return (0); 1587 } 1588 1589 /* receive buffer autosize */ 1590 MPASS(toep->vnet == so->so_vnet); 1591 CURVNET_SET(toep->vnet); 1592 if (sb->sb_flags & SB_AUTOSIZE && 1593 V_tcp_do_autorcvbuf && 1594 sb->sb_hiwat < V_tcp_autorcvbuf_max && 1595 len > (sbspace(sb) / 8 * 7)) { 1596 unsigned int hiwat = sb->sb_hiwat; 1597 unsigned int newsize = min(hiwat + sc->tt.autorcvbuf_inc, 1598 V_tcp_autorcvbuf_max); 1599 1600 if (!sbreserve_locked(sb, newsize, so, NULL)) 1601 sb->sb_flags &= ~SB_AUTOSIZE; 1602 } 1603 1604 if (ulp_mode(toep) == ULP_MODE_TCPDDP) { 1605 int changed = !(toep->ddp.flags & DDP_ON) ^ cpl->ddp_off; 1606 1607 if (toep->ddp.waiting_count != 0 || toep->ddp.active_count != 0) 1608 CTR3(KTR_CXGBE, "%s: tid %u, non-ddp rx (%d bytes)", 1609 __func__, tid, len); 1610 1611 if (changed) { 1612 if (toep->ddp.flags & DDP_SC_REQ) 1613 toep->ddp.flags ^= DDP_ON | DDP_SC_REQ; 1614 else { 1615 KASSERT(cpl->ddp_off == 1, 1616 ("%s: DDP switched on by itself.", 1617 __func__)); 1618 1619 /* Fell out of DDP mode */ 1620 toep->ddp.flags &= ~DDP_ON; 1621 CTR1(KTR_CXGBE, "%s: fell out of DDP mode", 1622 __func__); 1623 1624 insert_ddp_data(toep, ddp_placed); 1625 } 1626 } 1627 1628 if (toep->ddp.flags & DDP_ON) { 1629 /* 1630 * CPL_RX_DATA with DDP on can only be an indicate. 1631 * Start posting queued AIO requests via DDP. The 1632 * payload that arrived in this indicate is appended 1633 * to the socket buffer as usual. 1634 */ 1635 handle_ddp_indicate(toep); 1636 } 1637 } 1638 1639 sbappendstream_locked(sb, m, 0); 1640 rx_credits = sbspace(sb) > tp->rcv_wnd ? sbspace(sb) - tp->rcv_wnd : 0; 1641 if (rx_credits > 0 && sbused(sb) + tp->rcv_wnd < sb->sb_lowat) { 1642 rx_credits = send_rx_credits(sc, toep, rx_credits); 1643 tp->rcv_wnd += rx_credits; 1644 tp->rcv_adv += rx_credits; 1645 } 1646 1647 if (ulp_mode(toep) == ULP_MODE_TCPDDP && toep->ddp.waiting_count > 0 && 1648 sbavail(sb) != 0) { 1649 CTR2(KTR_CXGBE, "%s: tid %u queueing AIO task", __func__, 1650 tid); 1651 ddp_queue_toep(toep); 1652 } 1653 sorwakeup_locked(so); 1654 SOCKBUF_UNLOCK_ASSERT(sb); 1655 if (ulp_mode(toep) == ULP_MODE_TCPDDP) 1656 DDP_UNLOCK(toep); 1657 1658 INP_WUNLOCK(inp); 1659 CURVNET_RESTORE(); 1660 return (0); 1661 } 1662 1663 static int 1664 do_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) 1665 { 1666 struct adapter *sc = iq->adapter; 1667 const struct cpl_fw4_ack *cpl = (const void *)(rss + 1); 1668 unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl))); 1669 struct toepcb *toep = lookup_tid(sc, tid); 1670 struct inpcb *inp; 1671 struct tcpcb *tp; 1672 struct socket *so; 1673 uint8_t credits = cpl->credits; 1674 struct ofld_tx_sdesc *txsd; 1675 int plen; 1676 #ifdef INVARIANTS 1677 unsigned int opcode = G_CPL_FW4_ACK_OPCODE(be32toh(OPCODE_TID(cpl))); 1678 #endif 1679 1680 /* 1681 * Very unusual case: we'd sent a flowc + abort_req for a synq entry and 1682 * now this comes back carrying the credits for the flowc. 1683 */ 1684 if (__predict_false(toep->flags & TPF_SYNQE)) { 1685 KASSERT(toep->flags & TPF_ABORT_SHUTDOWN, 1686 ("%s: credits for a synq entry %p", __func__, toep)); 1687 return (0); 1688 } 1689 1690 inp = toep->inp; 1691 1692 KASSERT(opcode == CPL_FW4_ACK, 1693 ("%s: unexpected opcode 0x%x", __func__, opcode)); 1694 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 1695 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__)); 1696 1697 INP_WLOCK(inp); 1698 1699 if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN)) { 1700 INP_WUNLOCK(inp); 1701 return (0); 1702 } 1703 1704 KASSERT((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) == 0, 1705 ("%s: inp_flags 0x%x", __func__, inp->inp_flags)); 1706 1707 tp = intotcpcb(inp); 1708 1709 if (cpl->flags & CPL_FW4_ACK_FLAGS_SEQVAL) { 1710 tcp_seq snd_una = be32toh(cpl->snd_una); 1711 1712 #ifdef INVARIANTS 1713 if (__predict_false(SEQ_LT(snd_una, tp->snd_una))) { 1714 log(LOG_ERR, 1715 "%s: unexpected seq# %x for TID %u, snd_una %x\n", 1716 __func__, snd_una, toep->tid, tp->snd_una); 1717 } 1718 #endif 1719 1720 if (tp->snd_una != snd_una) { 1721 tp->snd_una = snd_una; 1722 tp->ts_recent_age = tcp_ts_getticks(); 1723 } 1724 } 1725 1726 #ifdef VERBOSE_TRACES 1727 CTR3(KTR_CXGBE, "%s: tid %d credits %u", __func__, tid, credits); 1728 #endif 1729 so = inp->inp_socket; 1730 txsd = &toep->txsd[toep->txsd_cidx]; 1731 plen = 0; 1732 while (credits) { 1733 KASSERT(credits >= txsd->tx_credits, 1734 ("%s: too many (or partial) credits", __func__)); 1735 credits -= txsd->tx_credits; 1736 toep->tx_credits += txsd->tx_credits; 1737 plen += txsd->plen; 1738 if (txsd->iv_buffer) { 1739 free(txsd->iv_buffer, M_CXGBE); 1740 txsd->iv_buffer = NULL; 1741 } 1742 txsd++; 1743 toep->txsd_avail++; 1744 KASSERT(toep->txsd_avail <= toep->txsd_total, 1745 ("%s: txsd avail > total", __func__)); 1746 if (__predict_false(++toep->txsd_cidx == toep->txsd_total)) { 1747 txsd = &toep->txsd[0]; 1748 toep->txsd_cidx = 0; 1749 } 1750 } 1751 1752 if (toep->tx_credits == toep->tx_total) { 1753 toep->tx_nocompl = 0; 1754 toep->plen_nocompl = 0; 1755 } 1756 1757 if (toep->flags & TPF_TX_SUSPENDED && 1758 toep->tx_credits >= toep->tx_total / 4) { 1759 #ifdef VERBOSE_TRACES 1760 CTR2(KTR_CXGBE, "%s: tid %d calling t4_push_frames", __func__, 1761 tid); 1762 #endif 1763 toep->flags &= ~TPF_TX_SUSPENDED; 1764 CURVNET_SET(toep->vnet); 1765 t4_push_data(sc, toep, plen); 1766 CURVNET_RESTORE(); 1767 } else if (plen > 0) { 1768 struct sockbuf *sb = &so->so_snd; 1769 int sbu; 1770 1771 SOCKBUF_LOCK(sb); 1772 sbu = sbused(sb); 1773 if (ulp_mode(toep) == ULP_MODE_ISCSI) { 1774 1775 if (__predict_false(sbu > 0)) { 1776 /* 1777 * The data trasmitted before the tid's ULP mode 1778 * changed to ISCSI is still in so_snd. 1779 * Incoming credits should account for so_snd 1780 * first. 1781 */ 1782 sbdrop_locked(sb, min(sbu, plen)); 1783 plen -= min(sbu, plen); 1784 } 1785 sowwakeup_locked(so); /* unlocks so_snd */ 1786 rqdrop_locked(&toep->ulp_pdu_reclaimq, plen); 1787 } else { 1788 #ifdef VERBOSE_TRACES 1789 CTR3(KTR_CXGBE, "%s: tid %d dropped %d bytes", __func__, 1790 tid, plen); 1791 #endif 1792 sbdrop_locked(sb, plen); 1793 if (tls_tx_key(toep) && 1794 toep->tls.mode == TLS_MODE_TLSOM) { 1795 struct tls_ofld_info *tls_ofld = &toep->tls; 1796 1797 MPASS(tls_ofld->sb_off >= plen); 1798 tls_ofld->sb_off -= plen; 1799 } 1800 if (!TAILQ_EMPTY(&toep->aiotx_jobq)) 1801 t4_aiotx_queue_toep(so, toep); 1802 sowwakeup_locked(so); /* unlocks so_snd */ 1803 } 1804 SOCKBUF_UNLOCK_ASSERT(sb); 1805 } 1806 1807 INP_WUNLOCK(inp); 1808 1809 return (0); 1810 } 1811 1812 void 1813 t4_set_tcb_field(struct adapter *sc, struct sge_wrq *wrq, struct toepcb *toep, 1814 uint16_t word, uint64_t mask, uint64_t val, int reply, int cookie) 1815 { 1816 struct wrqe *wr; 1817 struct cpl_set_tcb_field *req; 1818 struct ofld_tx_sdesc *txsd; 1819 1820 MPASS((cookie & ~M_COOKIE) == 0); 1821 if (reply) { 1822 MPASS(cookie != CPL_COOKIE_RESERVED); 1823 } 1824 1825 wr = alloc_wrqe(sizeof(*req), wrq); 1826 if (wr == NULL) { 1827 /* XXX */ 1828 panic("%s: allocation failure.", __func__); 1829 } 1830 req = wrtod(wr); 1831 1832 INIT_TP_WR_MIT_CPL(req, CPL_SET_TCB_FIELD, toep->tid); 1833 req->reply_ctrl = htobe16(V_QUEUENO(toep->ofld_rxq->iq.abs_id)); 1834 if (reply == 0) 1835 req->reply_ctrl |= htobe16(F_NO_REPLY); 1836 req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(cookie)); 1837 req->mask = htobe64(mask); 1838 req->val = htobe64(val); 1839 if ((wrq->eq.flags & EQ_TYPEMASK) == EQ_OFLD) { 1840 txsd = &toep->txsd[toep->txsd_pidx]; 1841 txsd->tx_credits = howmany(sizeof(*req), 16); 1842 txsd->plen = 0; 1843 KASSERT(toep->tx_credits >= txsd->tx_credits && 1844 toep->txsd_avail > 0, 1845 ("%s: not enough credits (%d)", __func__, 1846 toep->tx_credits)); 1847 toep->tx_credits -= txsd->tx_credits; 1848 if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) 1849 toep->txsd_pidx = 0; 1850 toep->txsd_avail--; 1851 } 1852 1853 t4_wrq_tx(sc, wr); 1854 } 1855 1856 void 1857 t4_init_cpl_io_handlers(void) 1858 { 1859 1860 t4_register_cpl_handler(CPL_PEER_CLOSE, do_peer_close); 1861 t4_register_cpl_handler(CPL_CLOSE_CON_RPL, do_close_con_rpl); 1862 t4_register_cpl_handler(CPL_ABORT_REQ_RSS, do_abort_req); 1863 t4_register_shared_cpl_handler(CPL_ABORT_RPL_RSS, do_abort_rpl, 1864 CPL_COOKIE_TOM); 1865 t4_register_cpl_handler(CPL_RX_DATA, do_rx_data); 1866 t4_register_shared_cpl_handler(CPL_FW4_ACK, do_fw4_ack, CPL_COOKIE_TOM); 1867 } 1868 1869 void 1870 t4_uninit_cpl_io_handlers(void) 1871 { 1872 1873 t4_register_cpl_handler(CPL_PEER_CLOSE, NULL); 1874 t4_register_cpl_handler(CPL_CLOSE_CON_RPL, NULL); 1875 t4_register_cpl_handler(CPL_ABORT_REQ_RSS, NULL); 1876 t4_register_shared_cpl_handler(CPL_ABORT_RPL_RSS, NULL, CPL_COOKIE_TOM); 1877 t4_register_cpl_handler(CPL_RX_DATA, NULL); 1878 t4_register_shared_cpl_handler(CPL_FW4_ACK, NULL, CPL_COOKIE_TOM); 1879 } 1880 1881 /* 1882 * Use the 'backend1' field in AIO jobs to hold an error that should 1883 * be reported when the job is completed, the 'backend3' field to 1884 * store the amount of data sent by the AIO job so far, and the 1885 * 'backend4' field to hold a reference count on the job. 1886 * 1887 * Each unmapped mbuf holds a reference on the job as does the queue 1888 * so long as the job is queued. 1889 */ 1890 #define aio_error backend1 1891 #define aio_sent backend3 1892 #define aio_refs backend4 1893 1894 #define jobtotid(job) \ 1895 (((struct toepcb *)(so_sototcpcb((job)->fd_file->f_data)->t_toe))->tid) 1896 1897 static void 1898 aiotx_free_job(struct kaiocb *job) 1899 { 1900 long status; 1901 int error; 1902 1903 if (refcount_release(&job->aio_refs) == 0) 1904 return; 1905 1906 error = (intptr_t)job->aio_error; 1907 status = job->aio_sent; 1908 #ifdef VERBOSE_TRACES 1909 CTR5(KTR_CXGBE, "%s: tid %d completed %p len %ld, error %d", __func__, 1910 jobtotid(job), job, status, error); 1911 #endif 1912 if (error != 0 && status != 0) 1913 error = 0; 1914 if (error == ECANCELED) 1915 aio_cancel(job); 1916 else if (error) 1917 aio_complete(job, -1, error); 1918 else { 1919 job->msgsnd = 1; 1920 aio_complete(job, status, 0); 1921 } 1922 } 1923 1924 static void 1925 aiotx_free_pgs(struct mbuf *m) 1926 { 1927 struct kaiocb *job; 1928 vm_page_t pg; 1929 1930 M_ASSERTEXTPG(m); 1931 job = m->m_ext.ext_arg1; 1932 #ifdef VERBOSE_TRACES 1933 CTR3(KTR_CXGBE, "%s: completed %d bytes for tid %d", __func__, 1934 m->m_len, jobtotid(job)); 1935 #endif 1936 1937 for (int i = 0; i < m->m_epg_npgs; i++) { 1938 pg = PHYS_TO_VM_PAGE(m->m_epg_pa[i]); 1939 vm_page_unwire(pg, PQ_ACTIVE); 1940 } 1941 1942 aiotx_free_job(job); 1943 } 1944 1945 /* 1946 * Allocate a chain of unmapped mbufs describing the next 'len' bytes 1947 * of an AIO job. 1948 */ 1949 static struct mbuf * 1950 alloc_aiotx_mbuf(struct kaiocb *job, int len) 1951 { 1952 struct vmspace *vm; 1953 vm_page_t pgs[MBUF_PEXT_MAX_PGS]; 1954 struct mbuf *m, *top, *last; 1955 vm_map_t map; 1956 vm_offset_t start; 1957 int i, mlen, npages, pgoff; 1958 1959 KASSERT(job->aio_sent + len <= job->uaiocb.aio_nbytes, 1960 ("%s(%p, %d): request to send beyond end of buffer", __func__, 1961 job, len)); 1962 1963 /* 1964 * The AIO subsystem will cancel and drain all requests before 1965 * permitting a process to exit or exec, so p_vmspace should 1966 * be stable here. 1967 */ 1968 vm = job->userproc->p_vmspace; 1969 map = &vm->vm_map; 1970 start = (uintptr_t)job->uaiocb.aio_buf + job->aio_sent; 1971 pgoff = start & PAGE_MASK; 1972 1973 top = NULL; 1974 last = NULL; 1975 while (len > 0) { 1976 mlen = imin(len, MBUF_PEXT_MAX_PGS * PAGE_SIZE - pgoff); 1977 KASSERT(mlen == len || ((start + mlen) & PAGE_MASK) == 0, 1978 ("%s: next start (%#jx + %#x) is not page aligned", 1979 __func__, (uintmax_t)start, mlen)); 1980 1981 npages = vm_fault_quick_hold_pages(map, start, mlen, 1982 VM_PROT_WRITE, pgs, nitems(pgs)); 1983 if (npages < 0) 1984 break; 1985 1986 m = mb_alloc_ext_pgs(M_WAITOK, aiotx_free_pgs); 1987 if (m == NULL) { 1988 vm_page_unhold_pages(pgs, npages); 1989 break; 1990 } 1991 1992 m->m_epg_1st_off = pgoff; 1993 m->m_epg_npgs = npages; 1994 if (npages == 1) { 1995 KASSERT(mlen + pgoff <= PAGE_SIZE, 1996 ("%s: single page is too large (off %d len %d)", 1997 __func__, pgoff, mlen)); 1998 m->m_epg_last_len = mlen; 1999 } else { 2000 m->m_epg_last_len = mlen - (PAGE_SIZE - pgoff) - 2001 (npages - 2) * PAGE_SIZE; 2002 } 2003 for (i = 0; i < npages; i++) 2004 m->m_epg_pa[i] = VM_PAGE_TO_PHYS(pgs[i]); 2005 2006 m->m_len = mlen; 2007 m->m_ext.ext_size = npages * PAGE_SIZE; 2008 m->m_ext.ext_arg1 = job; 2009 refcount_acquire(&job->aio_refs); 2010 2011 #ifdef VERBOSE_TRACES 2012 CTR5(KTR_CXGBE, "%s: tid %d, new mbuf %p for job %p, npages %d", 2013 __func__, jobtotid(job), m, job, npages); 2014 #endif 2015 2016 if (top == NULL) 2017 top = m; 2018 else 2019 last->m_next = m; 2020 last = m; 2021 2022 len -= mlen; 2023 start += mlen; 2024 pgoff = 0; 2025 } 2026 2027 return (top); 2028 } 2029 2030 static void 2031 t4_aiotx_process_job(struct toepcb *toep, struct socket *so, struct kaiocb *job) 2032 { 2033 struct sockbuf *sb; 2034 struct file *fp; 2035 struct inpcb *inp; 2036 struct tcpcb *tp; 2037 struct mbuf *m; 2038 int error, len; 2039 bool moretocome, sendmore; 2040 2041 sb = &so->so_snd; 2042 SOCKBUF_UNLOCK(sb); 2043 fp = job->fd_file; 2044 m = NULL; 2045 2046 #ifdef MAC 2047 error = mac_socket_check_send(fp->f_cred, so); 2048 if (error != 0) 2049 goto out; 2050 #endif 2051 2052 /* Inline sosend_generic(). */ 2053 2054 error = sblock(sb, SBL_WAIT); 2055 MPASS(error == 0); 2056 2057 sendanother: 2058 SOCKBUF_LOCK(sb); 2059 if (so->so_snd.sb_state & SBS_CANTSENDMORE) { 2060 SOCKBUF_UNLOCK(sb); 2061 sbunlock(sb); 2062 if ((so->so_options & SO_NOSIGPIPE) == 0) { 2063 PROC_LOCK(job->userproc); 2064 kern_psignal(job->userproc, SIGPIPE); 2065 PROC_UNLOCK(job->userproc); 2066 } 2067 error = EPIPE; 2068 goto out; 2069 } 2070 if (so->so_error) { 2071 error = so->so_error; 2072 so->so_error = 0; 2073 SOCKBUF_UNLOCK(sb); 2074 sbunlock(sb); 2075 goto out; 2076 } 2077 if ((so->so_state & SS_ISCONNECTED) == 0) { 2078 SOCKBUF_UNLOCK(sb); 2079 sbunlock(sb); 2080 error = ENOTCONN; 2081 goto out; 2082 } 2083 if (sbspace(sb) < sb->sb_lowat) { 2084 MPASS(job->aio_sent == 0 || !(so->so_state & SS_NBIO)); 2085 2086 /* 2087 * Don't block if there is too little room in the socket 2088 * buffer. Instead, requeue the request. 2089 */ 2090 if (!aio_set_cancel_function(job, t4_aiotx_cancel)) { 2091 SOCKBUF_UNLOCK(sb); 2092 sbunlock(sb); 2093 error = ECANCELED; 2094 goto out; 2095 } 2096 TAILQ_INSERT_HEAD(&toep->aiotx_jobq, job, list); 2097 SOCKBUF_UNLOCK(sb); 2098 sbunlock(sb); 2099 goto out; 2100 } 2101 2102 /* 2103 * Write as much data as the socket permits, but no more than a 2104 * a single sndbuf at a time. 2105 */ 2106 len = sbspace(sb); 2107 if (len > job->uaiocb.aio_nbytes - job->aio_sent) { 2108 len = job->uaiocb.aio_nbytes - job->aio_sent; 2109 moretocome = false; 2110 } else 2111 moretocome = true; 2112 if (len > toep->params.sndbuf) { 2113 len = toep->params.sndbuf; 2114 sendmore = true; 2115 } else 2116 sendmore = false; 2117 2118 if (!TAILQ_EMPTY(&toep->aiotx_jobq)) 2119 moretocome = true; 2120 SOCKBUF_UNLOCK(sb); 2121 MPASS(len != 0); 2122 2123 m = alloc_aiotx_mbuf(job, len); 2124 if (m == NULL) { 2125 sbunlock(sb); 2126 error = EFAULT; 2127 goto out; 2128 } 2129 2130 /* Inlined tcp_usr_send(). */ 2131 2132 inp = toep->inp; 2133 INP_WLOCK(inp); 2134 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { 2135 INP_WUNLOCK(inp); 2136 sbunlock(sb); 2137 error = ECONNRESET; 2138 goto out; 2139 } 2140 2141 job->aio_sent += m_length(m, NULL); 2142 2143 sbappendstream(sb, m, 0); 2144 m = NULL; 2145 2146 if (!(inp->inp_flags & INP_DROPPED)) { 2147 tp = intotcpcb(inp); 2148 if (moretocome) 2149 tp->t_flags |= TF_MORETOCOME; 2150 error = tp->t_fb->tfb_tcp_output(tp); 2151 if (moretocome) 2152 tp->t_flags &= ~TF_MORETOCOME; 2153 } 2154 2155 INP_WUNLOCK(inp); 2156 if (sendmore) 2157 goto sendanother; 2158 sbunlock(sb); 2159 2160 if (error) 2161 goto out; 2162 2163 /* 2164 * If this is a blocking socket and the request has not been 2165 * fully completed, requeue it until the socket is ready 2166 * again. 2167 */ 2168 if (job->aio_sent < job->uaiocb.aio_nbytes && 2169 !(so->so_state & SS_NBIO)) { 2170 SOCKBUF_LOCK(sb); 2171 if (!aio_set_cancel_function(job, t4_aiotx_cancel)) { 2172 SOCKBUF_UNLOCK(sb); 2173 error = ECANCELED; 2174 goto out; 2175 } 2176 TAILQ_INSERT_HEAD(&toep->aiotx_jobq, job, list); 2177 return; 2178 } 2179 2180 /* 2181 * If the request will not be requeued, drop the queue's 2182 * reference to the job. Any mbufs in flight should still 2183 * hold a reference, but this drops the reference that the 2184 * queue owns while it is waiting to queue mbufs to the 2185 * socket. 2186 */ 2187 aiotx_free_job(job); 2188 2189 out: 2190 if (error) { 2191 job->aio_error = (void *)(intptr_t)error; 2192 aiotx_free_job(job); 2193 } 2194 if (m != NULL) 2195 m_free(m); 2196 SOCKBUF_LOCK(sb); 2197 } 2198 2199 static void 2200 t4_aiotx_task(void *context, int pending) 2201 { 2202 struct toepcb *toep = context; 2203 struct socket *so; 2204 struct kaiocb *job; 2205 2206 so = toep->aiotx_so; 2207 CURVNET_SET(toep->vnet); 2208 SOCKBUF_LOCK(&so->so_snd); 2209 while (!TAILQ_EMPTY(&toep->aiotx_jobq) && sowriteable(so)) { 2210 job = TAILQ_FIRST(&toep->aiotx_jobq); 2211 TAILQ_REMOVE(&toep->aiotx_jobq, job, list); 2212 if (!aio_clear_cancel_function(job)) 2213 continue; 2214 2215 t4_aiotx_process_job(toep, so, job); 2216 } 2217 toep->aiotx_so = NULL; 2218 SOCKBUF_UNLOCK(&so->so_snd); 2219 CURVNET_RESTORE(); 2220 2221 free_toepcb(toep); 2222 SOCK_LOCK(so); 2223 sorele(so); 2224 } 2225 2226 static void 2227 t4_aiotx_queue_toep(struct socket *so, struct toepcb *toep) 2228 { 2229 2230 SOCKBUF_LOCK_ASSERT(&toep->inp->inp_socket->so_snd); 2231 #ifdef VERBOSE_TRACES 2232 CTR3(KTR_CXGBE, "%s: queueing aiotx task for tid %d, active = %s", 2233 __func__, toep->tid, toep->aiotx_so != NULL ? "true" : "false"); 2234 #endif 2235 if (toep->aiotx_so != NULL) 2236 return; 2237 soref(so); 2238 toep->aiotx_so = so; 2239 hold_toepcb(toep); 2240 soaio_enqueue(&toep->aiotx_task); 2241 } 2242 2243 static void 2244 t4_aiotx_cancel(struct kaiocb *job) 2245 { 2246 struct socket *so; 2247 struct sockbuf *sb; 2248 struct tcpcb *tp; 2249 struct toepcb *toep; 2250 2251 so = job->fd_file->f_data; 2252 tp = so_sototcpcb(so); 2253 toep = tp->t_toe; 2254 MPASS(job->uaiocb.aio_lio_opcode == LIO_WRITE); 2255 sb = &so->so_snd; 2256 2257 SOCKBUF_LOCK(sb); 2258 if (!aio_cancel_cleared(job)) 2259 TAILQ_REMOVE(&toep->aiotx_jobq, job, list); 2260 SOCKBUF_UNLOCK(sb); 2261 2262 job->aio_error = (void *)(intptr_t)ECANCELED; 2263 aiotx_free_job(job); 2264 } 2265 2266 int 2267 t4_aio_queue_aiotx(struct socket *so, struct kaiocb *job) 2268 { 2269 struct tcpcb *tp = so_sototcpcb(so); 2270 struct toepcb *toep = tp->t_toe; 2271 struct adapter *sc = td_adapter(toep->td); 2272 2273 /* This only handles writes. */ 2274 if (job->uaiocb.aio_lio_opcode != LIO_WRITE) 2275 return (EOPNOTSUPP); 2276 2277 if (!sc->tt.tx_zcopy) 2278 return (EOPNOTSUPP); 2279 2280 if (tls_tx_key(toep)) 2281 return (EOPNOTSUPP); 2282 2283 SOCKBUF_LOCK(&so->so_snd); 2284 #ifdef VERBOSE_TRACES 2285 CTR3(KTR_CXGBE, "%s: queueing %p for tid %u", __func__, job, toep->tid); 2286 #endif 2287 if (!aio_set_cancel_function(job, t4_aiotx_cancel)) 2288 panic("new job was cancelled"); 2289 refcount_init(&job->aio_refs, 1); 2290 TAILQ_INSERT_TAIL(&toep->aiotx_jobq, job, list); 2291 if (sowriteable(so)) 2292 t4_aiotx_queue_toep(so, toep); 2293 SOCKBUF_UNLOCK(&so->so_snd); 2294 return (0); 2295 } 2296 2297 void 2298 aiotx_init_toep(struct toepcb *toep) 2299 { 2300 2301 TAILQ_INIT(&toep->aiotx_jobq); 2302 TASK_INIT(&toep->aiotx_task, 0, t4_aiotx_task, toep); 2303 } 2304 #endif 2305