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 FLOWC_PARAM(MSS, toep->params.emss); 139 if (tp) { 140 FLOWC_PARAM(SNDNXT, tp->snd_nxt); 141 FLOWC_PARAM(RCVNXT, tp->rcv_nxt); 142 } 143 CTR6(KTR_CXGBE, 144 "%s: tid %u, mss %u, sndbuf %u, snd_nxt 0x%x, rcv_nxt 0x%x", 145 __func__, toep->tid, toep->params.emss, toep->params.sndbuf, 146 tp ? tp->snd_nxt : 0, tp ? tp->rcv_nxt : 0); 147 148 if (ulp_mode(toep) == ULP_MODE_TLS) 149 FLOWC_PARAM(ULP_MODE, ulp_mode(toep)); 150 if (toep->tls.fcplenmax != 0) 151 FLOWC_PARAM(TXDATAPLEN_MAX, toep->tls.fcplenmax); 152 if (toep->params.tc_idx != -1) 153 FLOWC_PARAM(SCHEDCLASS, toep->params.tc_idx); 154 #undef FLOWC_PARAM 155 156 KASSERT(paramidx == nparams, ("nparams mismatch")); 157 158 txsd->tx_credits = howmany(flowclen, 16); 159 txsd->plen = 0; 160 KASSERT(toep->tx_credits >= txsd->tx_credits && toep->txsd_avail > 0, 161 ("%s: not enough credits (%d)", __func__, toep->tx_credits)); 162 toep->tx_credits -= txsd->tx_credits; 163 if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) 164 toep->txsd_pidx = 0; 165 toep->txsd_avail--; 166 167 toep->flags |= TPF_FLOWC_WR_SENT; 168 t4_wrq_tx(sc, wr); 169 } 170 171 #ifdef RATELIMIT 172 /* 173 * Input is Bytes/second (so_max_pacing_rate), chip counts in Kilobits/second. 174 */ 175 static int 176 update_tx_rate_limit(struct adapter *sc, struct toepcb *toep, u_int Bps) 177 { 178 int tc_idx, rc; 179 const u_int kbps = (u_int) (uint64_t)Bps * 8ULL / 1000; 180 const int port_id = toep->vi->pi->port_id; 181 182 CTR3(KTR_CXGBE, "%s: tid %u, rate %uKbps", __func__, toep->tid, kbps); 183 184 if (kbps == 0) { 185 /* unbind */ 186 tc_idx = -1; 187 } else { 188 rc = t4_reserve_cl_rl_kbps(sc, port_id, kbps, &tc_idx); 189 if (rc != 0) 190 return (rc); 191 MPASS(tc_idx >= 0 && tc_idx < sc->chip_params->nsched_cls); 192 } 193 194 if (toep->params.tc_idx != tc_idx) { 195 struct wrqe *wr; 196 struct fw_flowc_wr *flowc; 197 int nparams = 1, flowclen, flowclen16; 198 struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx]; 199 200 flowclen = sizeof(*flowc) + nparams * sizeof(struct 201 fw_flowc_mnemval); 202 flowclen16 = howmany(flowclen, 16); 203 if (toep->tx_credits < flowclen16 || toep->txsd_avail == 0 || 204 (wr = alloc_wrqe(roundup2(flowclen, 16), toep->ofld_txq)) == NULL) { 205 if (tc_idx >= 0) 206 t4_release_cl_rl(sc, port_id, tc_idx); 207 return (ENOMEM); 208 } 209 210 flowc = wrtod(wr); 211 memset(flowc, 0, wr->wr_len); 212 213 flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) | 214 V_FW_FLOWC_WR_NPARAMS(nparams)); 215 flowc->flowid_len16 = htonl(V_FW_WR_LEN16(flowclen16) | 216 V_FW_WR_FLOWID(toep->tid)); 217 218 flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_SCHEDCLASS; 219 if (tc_idx == -1) 220 flowc->mnemval[0].val = htobe32(0xff); 221 else 222 flowc->mnemval[0].val = htobe32(tc_idx); 223 224 txsd->tx_credits = flowclen16; 225 txsd->plen = 0; 226 toep->tx_credits -= txsd->tx_credits; 227 if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) 228 toep->txsd_pidx = 0; 229 toep->txsd_avail--; 230 t4_wrq_tx(sc, wr); 231 } 232 233 if (toep->params.tc_idx >= 0) 234 t4_release_cl_rl(sc, port_id, toep->params.tc_idx); 235 toep->params.tc_idx = tc_idx; 236 237 return (0); 238 } 239 #endif 240 241 void 242 send_reset(struct adapter *sc, struct toepcb *toep, uint32_t snd_nxt) 243 { 244 struct wrqe *wr; 245 struct cpl_abort_req *req; 246 int tid = toep->tid; 247 struct inpcb *inp = toep->inp; 248 struct tcpcb *tp = intotcpcb(inp); /* don't use if INP_DROPPED */ 249 250 INP_WLOCK_ASSERT(inp); 251 252 CTR6(KTR_CXGBE, "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x%s", 253 __func__, toep->tid, 254 inp->inp_flags & INP_DROPPED ? "inp dropped" : 255 tcpstates[tp->t_state], 256 toep->flags, inp->inp_flags, 257 toep->flags & TPF_ABORT_SHUTDOWN ? 258 " (abort already in progress)" : ""); 259 260 if (toep->flags & TPF_ABORT_SHUTDOWN) 261 return; /* abort already in progress */ 262 263 toep->flags |= TPF_ABORT_SHUTDOWN; 264 265 KASSERT(toep->flags & TPF_FLOWC_WR_SENT, 266 ("%s: flowc_wr not sent for tid %d.", __func__, tid)); 267 268 wr = alloc_wrqe(sizeof(*req), toep->ofld_txq); 269 if (wr == NULL) { 270 /* XXX */ 271 panic("%s: allocation failure.", __func__); 272 } 273 req = wrtod(wr); 274 275 INIT_TP_WR_MIT_CPL(req, CPL_ABORT_REQ, tid); 276 if (inp->inp_flags & INP_DROPPED) 277 req->rsvd0 = htobe32(snd_nxt); 278 else 279 req->rsvd0 = htobe32(tp->snd_nxt); 280 req->rsvd1 = !(toep->flags & TPF_TX_DATA_SENT); 281 req->cmd = CPL_ABORT_SEND_RST; 282 283 /* 284 * XXX: What's the correct way to tell that the inp hasn't been detached 285 * from its socket? Should I even be flushing the snd buffer here? 286 */ 287 if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) { 288 struct socket *so = inp->inp_socket; 289 290 if (so != NULL) /* because I'm not sure. See comment above */ 291 sbflush(&so->so_snd); 292 } 293 294 t4_l2t_send(sc, wr, toep->l2te); 295 } 296 297 /* 298 * Called when a connection is established to translate the TCP options 299 * reported by HW to FreeBSD's native format. 300 */ 301 static void 302 assign_rxopt(struct tcpcb *tp, uint16_t opt) 303 { 304 struct toepcb *toep = tp->t_toe; 305 struct inpcb *inp = tp->t_inpcb; 306 struct adapter *sc = td_adapter(toep->td); 307 308 INP_LOCK_ASSERT(inp); 309 310 toep->params.mtu_idx = G_TCPOPT_MSS(opt); 311 tp->t_maxseg = sc->params.mtus[toep->params.mtu_idx]; 312 if (inp->inp_inc.inc_flags & INC_ISIPV6) 313 tp->t_maxseg -= sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 314 else 315 tp->t_maxseg -= sizeof(struct ip) + sizeof(struct tcphdr); 316 317 toep->params.emss = tp->t_maxseg; 318 if (G_TCPOPT_TSTAMP(opt)) { 319 toep->params.tstamp = 1; 320 toep->params.emss -= TCPOLEN_TSTAMP_APPA; 321 tp->t_flags |= TF_RCVD_TSTMP; /* timestamps ok */ 322 tp->ts_recent = 0; /* hmmm */ 323 tp->ts_recent_age = tcp_ts_getticks(); 324 } else 325 toep->params.tstamp = 0; 326 327 if (G_TCPOPT_SACK(opt)) { 328 toep->params.sack = 1; 329 tp->t_flags |= TF_SACK_PERMIT; /* should already be set */ 330 } else { 331 toep->params.sack = 0; 332 tp->t_flags &= ~TF_SACK_PERMIT; /* sack disallowed by peer */ 333 } 334 335 if (G_TCPOPT_WSCALE_OK(opt)) 336 tp->t_flags |= TF_RCVD_SCALE; 337 338 /* Doing window scaling? */ 339 if ((tp->t_flags & (TF_RCVD_SCALE | TF_REQ_SCALE)) == 340 (TF_RCVD_SCALE | TF_REQ_SCALE)) { 341 tp->rcv_scale = tp->request_r_scale; 342 tp->snd_scale = G_TCPOPT_SND_WSCALE(opt); 343 } else 344 toep->params.wscale = 0; 345 346 CTR6(KTR_CXGBE, 347 "assign_rxopt: tid %d, mtu_idx %u, emss %u, ts %u, sack %u, wscale %u", 348 toep->tid, toep->params.mtu_idx, toep->params.emss, 349 toep->params.tstamp, toep->params.sack, toep->params.wscale); 350 } 351 352 /* 353 * Completes some final bits of initialization for just established connections 354 * and changes their state to TCPS_ESTABLISHED. 355 * 356 * The ISNs are from the exchange of SYNs. 357 */ 358 void 359 make_established(struct toepcb *toep, uint32_t iss, uint32_t irs, uint16_t opt) 360 { 361 struct inpcb *inp = toep->inp; 362 struct socket *so = inp->inp_socket; 363 struct tcpcb *tp = intotcpcb(inp); 364 uint16_t tcpopt = be16toh(opt); 365 366 INP_WLOCK_ASSERT(inp); 367 KASSERT(tp->t_state == TCPS_SYN_SENT || 368 tp->t_state == TCPS_SYN_RECEIVED, 369 ("%s: TCP state %s", __func__, tcpstates[tp->t_state])); 370 371 CTR6(KTR_CXGBE, "%s: tid %d, so %p, inp %p, tp %p, toep %p", 372 __func__, toep->tid, so, inp, tp, toep); 373 374 tcp_state_change(tp, TCPS_ESTABLISHED); 375 tp->t_starttime = ticks; 376 TCPSTAT_INC(tcps_connects); 377 378 tp->irs = irs; 379 tcp_rcvseqinit(tp); 380 tp->rcv_wnd = (u_int)toep->params.opt0_bufsize << 10; 381 tp->rcv_adv += tp->rcv_wnd; 382 tp->last_ack_sent = tp->rcv_nxt; 383 384 tp->iss = iss; 385 tcp_sendseqinit(tp); 386 tp->snd_una = iss + 1; 387 tp->snd_nxt = iss + 1; 388 tp->snd_max = iss + 1; 389 390 assign_rxopt(tp, tcpopt); 391 send_flowc_wr(toep, tp); 392 393 soisconnected(so); 394 } 395 396 int 397 send_rx_credits(struct adapter *sc, struct toepcb *toep, int credits) 398 { 399 struct wrqe *wr; 400 struct cpl_rx_data_ack *req; 401 uint32_t dack = F_RX_DACK_CHANGE | V_RX_DACK_MODE(1); 402 403 KASSERT(credits >= 0, ("%s: %d credits", __func__, credits)); 404 405 wr = alloc_wrqe(sizeof(*req), toep->ctrlq); 406 if (wr == NULL) 407 return (0); 408 req = wrtod(wr); 409 410 INIT_TP_WR_MIT_CPL(req, CPL_RX_DATA_ACK, toep->tid); 411 req->credit_dack = htobe32(dack | V_RX_CREDITS(credits)); 412 413 t4_wrq_tx(sc, wr); 414 return (credits); 415 } 416 417 void 418 send_rx_modulate(struct adapter *sc, struct toepcb *toep) 419 { 420 struct wrqe *wr; 421 struct cpl_rx_data_ack *req; 422 423 wr = alloc_wrqe(sizeof(*req), toep->ctrlq); 424 if (wr == NULL) 425 return; 426 req = wrtod(wr); 427 428 INIT_TP_WR_MIT_CPL(req, CPL_RX_DATA_ACK, toep->tid); 429 req->credit_dack = htobe32(F_RX_MODULATE_RX); 430 431 t4_wrq_tx(sc, wr); 432 } 433 434 void 435 t4_rcvd_locked(struct toedev *tod, struct tcpcb *tp) 436 { 437 struct adapter *sc = tod->tod_softc; 438 struct inpcb *inp = tp->t_inpcb; 439 struct socket *so = inp->inp_socket; 440 struct sockbuf *sb = &so->so_rcv; 441 struct toepcb *toep = tp->t_toe; 442 int rx_credits; 443 444 INP_WLOCK_ASSERT(inp); 445 SOCKBUF_LOCK_ASSERT(sb); 446 447 rx_credits = sbspace(sb) > tp->rcv_wnd ? sbspace(sb) - tp->rcv_wnd : 0; 448 if (ulp_mode(toep) == ULP_MODE_TLS) { 449 if (toep->tls.rcv_over >= rx_credits) { 450 toep->tls.rcv_over -= rx_credits; 451 rx_credits = 0; 452 } else { 453 rx_credits -= toep->tls.rcv_over; 454 toep->tls.rcv_over = 0; 455 } 456 } 457 458 if (rx_credits > 0 && 459 (tp->rcv_wnd <= 32 * 1024 || rx_credits >= 64 * 1024 || 460 (rx_credits >= 16 * 1024 && tp->rcv_wnd <= 128 * 1024) || 461 sbused(sb) + tp->rcv_wnd < sb->sb_lowat)) { 462 rx_credits = send_rx_credits(sc, toep, rx_credits); 463 tp->rcv_wnd += rx_credits; 464 tp->rcv_adv += rx_credits; 465 } else if (toep->flags & TPF_FORCE_CREDITS) 466 send_rx_modulate(sc, toep); 467 } 468 469 void 470 t4_rcvd(struct toedev *tod, struct tcpcb *tp) 471 { 472 struct inpcb *inp = tp->t_inpcb; 473 struct socket *so = inp->inp_socket; 474 struct sockbuf *sb = &so->so_rcv; 475 476 SOCKBUF_LOCK(sb); 477 t4_rcvd_locked(tod, tp); 478 SOCKBUF_UNLOCK(sb); 479 } 480 481 /* 482 * Close a connection by sending a CPL_CLOSE_CON_REQ message. 483 */ 484 int 485 t4_close_conn(struct adapter *sc, struct toepcb *toep) 486 { 487 struct wrqe *wr; 488 struct cpl_close_con_req *req; 489 unsigned int tid = toep->tid; 490 491 CTR3(KTR_CXGBE, "%s: tid %u%s", __func__, toep->tid, 492 toep->flags & TPF_FIN_SENT ? ", IGNORED" : ""); 493 494 if (toep->flags & TPF_FIN_SENT) 495 return (0); 496 497 KASSERT(toep->flags & TPF_FLOWC_WR_SENT, 498 ("%s: flowc_wr not sent for tid %u.", __func__, tid)); 499 500 wr = alloc_wrqe(sizeof(*req), toep->ofld_txq); 501 if (wr == NULL) { 502 /* XXX */ 503 panic("%s: allocation failure.", __func__); 504 } 505 req = wrtod(wr); 506 507 req->wr.wr_hi = htonl(V_FW_WR_OP(FW_TP_WR) | 508 V_FW_WR_IMMDLEN(sizeof(*req) - sizeof(req->wr))); 509 req->wr.wr_mid = htonl(V_FW_WR_LEN16(howmany(sizeof(*req), 16)) | 510 V_FW_WR_FLOWID(tid)); 511 req->wr.wr_lo = cpu_to_be64(0); 512 OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_CLOSE_CON_REQ, tid)); 513 req->rsvd = 0; 514 515 toep->flags |= TPF_FIN_SENT; 516 toep->flags &= ~TPF_SEND_FIN; 517 t4_l2t_send(sc, wr, toep->l2te); 518 519 return (0); 520 } 521 522 #define MAX_OFLD_TX_CREDITS (SGE_MAX_WR_LEN / 16) 523 #define MIN_OFLD_TX_CREDITS (howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16)) 524 525 /* Maximum amount of immediate data we could stuff in a WR */ 526 static inline int 527 max_imm_payload(int tx_credits) 528 { 529 const int n = 1; /* Use no more than one desc for imm. data WR */ 530 531 KASSERT(tx_credits >= 0 && 532 tx_credits <= MAX_OFLD_TX_CREDITS, 533 ("%s: %d credits", __func__, tx_credits)); 534 535 if (tx_credits < MIN_OFLD_TX_CREDITS) 536 return (0); 537 538 if (tx_credits >= (n * EQ_ESIZE) / 16) 539 return ((n * EQ_ESIZE) - sizeof(struct fw_ofld_tx_data_wr)); 540 else 541 return (tx_credits * 16 - sizeof(struct fw_ofld_tx_data_wr)); 542 } 543 544 /* Maximum number of SGL entries we could stuff in a WR */ 545 static inline int 546 max_dsgl_nsegs(int tx_credits) 547 { 548 int nseg = 1; /* ulptx_sgl has room for 1, rest ulp_tx_sge_pair */ 549 int sge_pair_credits = tx_credits - MIN_OFLD_TX_CREDITS; 550 551 KASSERT(tx_credits >= 0 && 552 tx_credits <= MAX_OFLD_TX_CREDITS, 553 ("%s: %d credits", __func__, tx_credits)); 554 555 if (tx_credits < MIN_OFLD_TX_CREDITS) 556 return (0); 557 558 nseg += 2 * (sge_pair_credits * 16 / 24); 559 if ((sge_pair_credits * 16) % 24 == 16) 560 nseg++; 561 562 return (nseg); 563 } 564 565 static inline void 566 write_tx_wr(void *dst, struct toepcb *toep, unsigned int immdlen, 567 unsigned int plen, uint8_t credits, int shove, int ulp_submode) 568 { 569 struct fw_ofld_tx_data_wr *txwr = dst; 570 571 txwr->op_to_immdlen = htobe32(V_WR_OP(FW_OFLD_TX_DATA_WR) | 572 V_FW_WR_IMMDLEN(immdlen)); 573 txwr->flowid_len16 = htobe32(V_FW_WR_FLOWID(toep->tid) | 574 V_FW_WR_LEN16(credits)); 575 txwr->lsodisable_to_flags = htobe32(V_TX_ULP_MODE(ulp_mode(toep)) | 576 V_TX_ULP_SUBMODE(ulp_submode) | V_TX_URG(0) | V_TX_SHOVE(shove)); 577 txwr->plen = htobe32(plen); 578 579 if (toep->params.tx_align > 0) { 580 if (plen < 2 * toep->params.emss) 581 txwr->lsodisable_to_flags |= 582 htobe32(F_FW_OFLD_TX_DATA_WR_LSODISABLE); 583 else 584 txwr->lsodisable_to_flags |= 585 htobe32(F_FW_OFLD_TX_DATA_WR_ALIGNPLD | 586 (toep->params.nagle == 0 ? 0 : 587 F_FW_OFLD_TX_DATA_WR_ALIGNPLDSHOVE)); 588 } 589 } 590 591 /* 592 * Generate a DSGL from a starting mbuf. The total number of segments and the 593 * maximum segments in any one mbuf are provided. 594 */ 595 static void 596 write_tx_sgl(void *dst, struct mbuf *start, struct mbuf *stop, int nsegs, int n) 597 { 598 struct mbuf *m; 599 struct ulptx_sgl *usgl = dst; 600 int i, j, rc; 601 struct sglist sg; 602 struct sglist_seg segs[n]; 603 604 KASSERT(nsegs > 0, ("%s: nsegs 0", __func__)); 605 606 sglist_init(&sg, n, segs); 607 usgl->cmd_nsge = htobe32(V_ULPTX_CMD(ULP_TX_SC_DSGL) | 608 V_ULPTX_NSGE(nsegs)); 609 610 i = -1; 611 for (m = start; m != stop; m = m->m_next) { 612 if (m->m_flags & M_NOMAP) 613 rc = sglist_append_mb_ext_pgs(&sg, m); 614 else 615 rc = sglist_append(&sg, mtod(m, void *), m->m_len); 616 if (__predict_false(rc != 0)) 617 panic("%s: sglist_append %d", __func__, rc); 618 619 for (j = 0; j < sg.sg_nseg; i++, j++) { 620 if (i < 0) { 621 usgl->len0 = htobe32(segs[j].ss_len); 622 usgl->addr0 = htobe64(segs[j].ss_paddr); 623 } else { 624 usgl->sge[i / 2].len[i & 1] = 625 htobe32(segs[j].ss_len); 626 usgl->sge[i / 2].addr[i & 1] = 627 htobe64(segs[j].ss_paddr); 628 } 629 #ifdef INVARIANTS 630 nsegs--; 631 #endif 632 } 633 sglist_reset(&sg); 634 } 635 if (i & 1) 636 usgl->sge[i / 2].len[1] = htobe32(0); 637 KASSERT(nsegs == 0, ("%s: nsegs %d, start %p, stop %p", 638 __func__, nsegs, start, stop)); 639 } 640 641 /* 642 * Max number of SGL entries an offload tx work request can have. This is 41 643 * (1 + 40) for a full 512B work request. 644 * fw_ofld_tx_data_wr(16B) + ulptx_sgl(16B, 1) + ulptx_sge_pair(480B, 40) 645 */ 646 #define OFLD_SGL_LEN (41) 647 648 /* 649 * Send data and/or a FIN to the peer. 650 * 651 * The socket's so_snd buffer consists of a stream of data starting with sb_mb 652 * and linked together with m_next. sb_sndptr, if set, is the last mbuf that 653 * was transmitted. 654 * 655 * drop indicates the number of bytes that should be dropped from the head of 656 * the send buffer. It is an optimization that lets do_fw4_ack avoid creating 657 * contention on the send buffer lock (before this change it used to do 658 * sowwakeup and then t4_push_frames right after that when recovering from tx 659 * stalls). When drop is set this function MUST drop the bytes and wake up any 660 * writers. 661 */ 662 void 663 t4_push_frames(struct adapter *sc, struct toepcb *toep, int drop) 664 { 665 struct mbuf *sndptr, *m, *sb_sndptr; 666 struct fw_ofld_tx_data_wr *txwr; 667 struct wrqe *wr; 668 u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf; 669 struct inpcb *inp = toep->inp; 670 struct tcpcb *tp = intotcpcb(inp); 671 struct socket *so = inp->inp_socket; 672 struct sockbuf *sb = &so->so_snd; 673 int tx_credits, shove, compl, sowwakeup; 674 struct ofld_tx_sdesc *txsd; 675 bool nomap_mbuf_seen; 676 677 INP_WLOCK_ASSERT(inp); 678 KASSERT(toep->flags & TPF_FLOWC_WR_SENT, 679 ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid)); 680 681 KASSERT(ulp_mode(toep) == ULP_MODE_NONE || 682 ulp_mode(toep) == ULP_MODE_TCPDDP || 683 ulp_mode(toep) == ULP_MODE_TLS || 684 ulp_mode(toep) == ULP_MODE_RDMA, 685 ("%s: ulp_mode %u for toep %p", __func__, ulp_mode(toep), toep)); 686 687 #ifdef VERBOSE_TRACES 688 CTR5(KTR_CXGBE, "%s: tid %d toep flags %#x tp flags %#x drop %d", 689 __func__, toep->tid, toep->flags, tp->t_flags, drop); 690 #endif 691 if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN)) 692 return; 693 694 #ifdef RATELIMIT 695 if (__predict_false(inp->inp_flags2 & INP_RATE_LIMIT_CHANGED) && 696 (update_tx_rate_limit(sc, toep, so->so_max_pacing_rate) == 0)) { 697 inp->inp_flags2 &= ~INP_RATE_LIMIT_CHANGED; 698 } 699 #endif 700 701 /* 702 * This function doesn't resume by itself. Someone else must clear the 703 * flag and call this function. 704 */ 705 if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) { 706 KASSERT(drop == 0, 707 ("%s: drop (%d) != 0 but tx is suspended", __func__, drop)); 708 return; 709 } 710 711 txsd = &toep->txsd[toep->txsd_pidx]; 712 do { 713 tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS); 714 max_imm = max_imm_payload(tx_credits); 715 max_nsegs = max_dsgl_nsegs(tx_credits); 716 717 SOCKBUF_LOCK(sb); 718 sowwakeup = drop; 719 if (drop) { 720 sbdrop_locked(sb, drop); 721 drop = 0; 722 } 723 sb_sndptr = sb->sb_sndptr; 724 sndptr = sb_sndptr ? sb_sndptr->m_next : sb->sb_mb; 725 plen = 0; 726 nsegs = 0; 727 max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */ 728 nomap_mbuf_seen = false; 729 for (m = sndptr; m != NULL; m = m->m_next) { 730 int n; 731 732 if (m->m_flags & M_NOMAP) { 733 #ifdef KERN_TLS 734 if (m->m_ext.ext_pgs->tls != NULL) { 735 toep->flags |= TPF_KTLS; 736 if (plen == 0) { 737 SOCKBUF_UNLOCK(sb); 738 t4_push_ktls(sc, toep, 0); 739 return; 740 } 741 break; 742 } 743 #endif 744 n = sglist_count_mb_ext_pgs(m); 745 } else 746 n = sglist_count(mtod(m, void *), m->m_len); 747 748 nsegs += n; 749 plen += m->m_len; 750 751 /* This mbuf sent us _over_ the nsegs limit, back out */ 752 if (plen > max_imm && nsegs > max_nsegs) { 753 nsegs -= n; 754 plen -= m->m_len; 755 if (plen == 0) { 756 /* Too few credits */ 757 toep->flags |= TPF_TX_SUSPENDED; 758 if (sowwakeup) { 759 if (!TAILQ_EMPTY( 760 &toep->aiotx_jobq)) 761 t4_aiotx_queue_toep(so, 762 toep); 763 sowwakeup_locked(so); 764 } else 765 SOCKBUF_UNLOCK(sb); 766 SOCKBUF_UNLOCK_ASSERT(sb); 767 return; 768 } 769 break; 770 } 771 772 if (m->m_flags & M_NOMAP) 773 nomap_mbuf_seen = true; 774 if (max_nsegs_1mbuf < n) 775 max_nsegs_1mbuf = n; 776 sb_sndptr = m; /* new sb->sb_sndptr if all goes well */ 777 778 /* This mbuf put us right at the max_nsegs limit */ 779 if (plen > max_imm && nsegs == max_nsegs) { 780 m = m->m_next; 781 break; 782 } 783 } 784 785 if (sbused(sb) > sb->sb_hiwat * 5 / 8 && 786 toep->plen_nocompl + plen >= sb->sb_hiwat / 4) 787 compl = 1; 788 else 789 compl = 0; 790 791 if (sb->sb_flags & SB_AUTOSIZE && 792 V_tcp_do_autosndbuf && 793 sb->sb_hiwat < V_tcp_autosndbuf_max && 794 sbused(sb) >= sb->sb_hiwat * 7 / 8) { 795 int newsize = min(sb->sb_hiwat + V_tcp_autosndbuf_inc, 796 V_tcp_autosndbuf_max); 797 798 if (!sbreserve_locked(sb, newsize, so, NULL)) 799 sb->sb_flags &= ~SB_AUTOSIZE; 800 else 801 sowwakeup = 1; /* room available */ 802 } 803 if (sowwakeup) { 804 if (!TAILQ_EMPTY(&toep->aiotx_jobq)) 805 t4_aiotx_queue_toep(so, toep); 806 sowwakeup_locked(so); 807 } else 808 SOCKBUF_UNLOCK(sb); 809 SOCKBUF_UNLOCK_ASSERT(sb); 810 811 /* nothing to send */ 812 if (plen == 0) { 813 KASSERT(m == NULL, 814 ("%s: nothing to send, but m != NULL", __func__)); 815 break; 816 } 817 818 if (__predict_false(toep->flags & TPF_FIN_SENT)) 819 panic("%s: excess tx.", __func__); 820 821 shove = m == NULL && !(tp->t_flags & TF_MORETOCOME); 822 if (plen <= max_imm && !nomap_mbuf_seen) { 823 824 /* Immediate data tx */ 825 826 wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16), 827 toep->ofld_txq); 828 if (wr == NULL) { 829 /* XXX: how will we recover from this? */ 830 toep->flags |= TPF_TX_SUSPENDED; 831 return; 832 } 833 txwr = wrtod(wr); 834 credits = howmany(wr->wr_len, 16); 835 write_tx_wr(txwr, toep, plen, plen, credits, shove, 0); 836 m_copydata(sndptr, 0, plen, (void *)(txwr + 1)); 837 nsegs = 0; 838 } else { 839 int wr_len; 840 841 /* DSGL tx */ 842 843 wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) + 844 ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8; 845 wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq); 846 if (wr == NULL) { 847 /* XXX: how will we recover from this? */ 848 toep->flags |= TPF_TX_SUSPENDED; 849 return; 850 } 851 txwr = wrtod(wr); 852 credits = howmany(wr_len, 16); 853 write_tx_wr(txwr, toep, 0, plen, credits, shove, 0); 854 write_tx_sgl(txwr + 1, sndptr, m, nsegs, 855 max_nsegs_1mbuf); 856 if (wr_len & 0xf) { 857 uint64_t *pad = (uint64_t *) 858 ((uintptr_t)txwr + wr_len); 859 *pad = 0; 860 } 861 } 862 863 KASSERT(toep->tx_credits >= credits, 864 ("%s: not enough credits", __func__)); 865 866 toep->tx_credits -= credits; 867 toep->tx_nocompl += credits; 868 toep->plen_nocompl += plen; 869 if (toep->tx_credits <= toep->tx_total * 3 / 8 && 870 toep->tx_nocompl >= toep->tx_total / 4) 871 compl = 1; 872 873 if (compl || ulp_mode(toep) == ULP_MODE_RDMA) { 874 txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL); 875 toep->tx_nocompl = 0; 876 toep->plen_nocompl = 0; 877 } 878 879 tp->snd_nxt += plen; 880 tp->snd_max += plen; 881 882 SOCKBUF_LOCK(sb); 883 KASSERT(sb_sndptr, ("%s: sb_sndptr is NULL", __func__)); 884 sb->sb_sndptr = sb_sndptr; 885 SOCKBUF_UNLOCK(sb); 886 887 toep->flags |= TPF_TX_DATA_SENT; 888 if (toep->tx_credits < MIN_OFLD_TX_CREDITS) 889 toep->flags |= TPF_TX_SUSPENDED; 890 891 KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__)); 892 txsd->plen = plen; 893 txsd->tx_credits = credits; 894 txsd++; 895 if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) { 896 toep->txsd_pidx = 0; 897 txsd = &toep->txsd[0]; 898 } 899 toep->txsd_avail--; 900 901 t4_l2t_send(sc, wr, toep->l2te); 902 } while (m != NULL); 903 904 /* Send a FIN if requested, but only if there's no more data to send */ 905 if (m == NULL && toep->flags & TPF_SEND_FIN) 906 t4_close_conn(sc, toep); 907 } 908 909 static inline void 910 rqdrop_locked(struct mbufq *q, int plen) 911 { 912 struct mbuf *m; 913 914 while (plen > 0) { 915 m = mbufq_dequeue(q); 916 917 /* Too many credits. */ 918 MPASS(m != NULL); 919 M_ASSERTPKTHDR(m); 920 921 /* Partial credits. */ 922 MPASS(plen >= m->m_pkthdr.len); 923 924 plen -= m->m_pkthdr.len; 925 m_freem(m); 926 } 927 } 928 929 void 930 t4_push_pdus(struct adapter *sc, struct toepcb *toep, int drop) 931 { 932 struct mbuf *sndptr, *m; 933 struct fw_ofld_tx_data_wr *txwr; 934 struct wrqe *wr; 935 u_int plen, nsegs, credits, max_imm, max_nsegs, max_nsegs_1mbuf; 936 u_int adjusted_plen, ulp_submode; 937 struct inpcb *inp = toep->inp; 938 struct tcpcb *tp = intotcpcb(inp); 939 int tx_credits, shove; 940 struct ofld_tx_sdesc *txsd = &toep->txsd[toep->txsd_pidx]; 941 struct mbufq *pduq = &toep->ulp_pduq; 942 static const u_int ulp_extra_len[] = {0, 4, 4, 8}; 943 944 INP_WLOCK_ASSERT(inp); 945 KASSERT(toep->flags & TPF_FLOWC_WR_SENT, 946 ("%s: flowc_wr not sent for tid %u.", __func__, toep->tid)); 947 KASSERT(ulp_mode(toep) == ULP_MODE_ISCSI, 948 ("%s: ulp_mode %u for toep %p", __func__, ulp_mode(toep), toep)); 949 950 if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN)) 951 return; 952 953 /* 954 * This function doesn't resume by itself. Someone else must clear the 955 * flag and call this function. 956 */ 957 if (__predict_false(toep->flags & TPF_TX_SUSPENDED)) { 958 KASSERT(drop == 0, 959 ("%s: drop (%d) != 0 but tx is suspended", __func__, drop)); 960 return; 961 } 962 963 if (drop) 964 rqdrop_locked(&toep->ulp_pdu_reclaimq, drop); 965 966 while ((sndptr = mbufq_first(pduq)) != NULL) { 967 M_ASSERTPKTHDR(sndptr); 968 969 tx_credits = min(toep->tx_credits, MAX_OFLD_TX_CREDITS); 970 max_imm = max_imm_payload(tx_credits); 971 max_nsegs = max_dsgl_nsegs(tx_credits); 972 973 plen = 0; 974 nsegs = 0; 975 max_nsegs_1mbuf = 0; /* max # of SGL segments in any one mbuf */ 976 for (m = sndptr; m != NULL; m = m->m_next) { 977 int n = sglist_count(mtod(m, void *), m->m_len); 978 979 nsegs += n; 980 plen += m->m_len; 981 982 /* 983 * This mbuf would send us _over_ the nsegs limit. 984 * Suspend tx because the PDU can't be sent out. 985 */ 986 if (plen > max_imm && nsegs > max_nsegs) { 987 toep->flags |= TPF_TX_SUSPENDED; 988 return; 989 } 990 991 if (max_nsegs_1mbuf < n) 992 max_nsegs_1mbuf = n; 993 } 994 995 if (__predict_false(toep->flags & TPF_FIN_SENT)) 996 panic("%s: excess tx.", __func__); 997 998 /* 999 * We have a PDU to send. All of it goes out in one WR so 'm' 1000 * is NULL. A PDU's length is always a multiple of 4. 1001 */ 1002 MPASS(m == NULL); 1003 MPASS((plen & 3) == 0); 1004 MPASS(sndptr->m_pkthdr.len == plen); 1005 1006 shove = !(tp->t_flags & TF_MORETOCOME); 1007 ulp_submode = mbuf_ulp_submode(sndptr); 1008 MPASS(ulp_submode < nitems(ulp_extra_len)); 1009 1010 /* 1011 * plen doesn't include header and data digests, which are 1012 * generated and inserted in the right places by the TOE, but 1013 * they do occupy TCP sequence space and need to be accounted 1014 * for. 1015 */ 1016 adjusted_plen = plen + ulp_extra_len[ulp_submode]; 1017 if (plen <= max_imm) { 1018 1019 /* Immediate data tx */ 1020 1021 wr = alloc_wrqe(roundup2(sizeof(*txwr) + plen, 16), 1022 toep->ofld_txq); 1023 if (wr == NULL) { 1024 /* XXX: how will we recover from this? */ 1025 toep->flags |= TPF_TX_SUSPENDED; 1026 return; 1027 } 1028 txwr = wrtod(wr); 1029 credits = howmany(wr->wr_len, 16); 1030 write_tx_wr(txwr, toep, plen, adjusted_plen, credits, 1031 shove, ulp_submode); 1032 m_copydata(sndptr, 0, plen, (void *)(txwr + 1)); 1033 nsegs = 0; 1034 } else { 1035 int wr_len; 1036 1037 /* DSGL tx */ 1038 wr_len = sizeof(*txwr) + sizeof(struct ulptx_sgl) + 1039 ((3 * (nsegs - 1)) / 2 + ((nsegs - 1) & 1)) * 8; 1040 wr = alloc_wrqe(roundup2(wr_len, 16), toep->ofld_txq); 1041 if (wr == NULL) { 1042 /* XXX: how will we recover from this? */ 1043 toep->flags |= TPF_TX_SUSPENDED; 1044 return; 1045 } 1046 txwr = wrtod(wr); 1047 credits = howmany(wr_len, 16); 1048 write_tx_wr(txwr, toep, 0, adjusted_plen, credits, 1049 shove, ulp_submode); 1050 write_tx_sgl(txwr + 1, sndptr, m, nsegs, 1051 max_nsegs_1mbuf); 1052 if (wr_len & 0xf) { 1053 uint64_t *pad = (uint64_t *) 1054 ((uintptr_t)txwr + wr_len); 1055 *pad = 0; 1056 } 1057 } 1058 1059 KASSERT(toep->tx_credits >= credits, 1060 ("%s: not enough credits", __func__)); 1061 1062 m = mbufq_dequeue(pduq); 1063 MPASS(m == sndptr); 1064 mbufq_enqueue(&toep->ulp_pdu_reclaimq, m); 1065 1066 toep->tx_credits -= credits; 1067 toep->tx_nocompl += credits; 1068 toep->plen_nocompl += plen; 1069 if (toep->tx_credits <= toep->tx_total * 3 / 8 && 1070 toep->tx_nocompl >= toep->tx_total / 4) { 1071 txwr->op_to_immdlen |= htobe32(F_FW_WR_COMPL); 1072 toep->tx_nocompl = 0; 1073 toep->plen_nocompl = 0; 1074 } 1075 1076 tp->snd_nxt += adjusted_plen; 1077 tp->snd_max += adjusted_plen; 1078 1079 toep->flags |= TPF_TX_DATA_SENT; 1080 if (toep->tx_credits < MIN_OFLD_TX_CREDITS) 1081 toep->flags |= TPF_TX_SUSPENDED; 1082 1083 KASSERT(toep->txsd_avail > 0, ("%s: no txsd", __func__)); 1084 txsd->plen = plen; 1085 txsd->tx_credits = credits; 1086 txsd++; 1087 if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) { 1088 toep->txsd_pidx = 0; 1089 txsd = &toep->txsd[0]; 1090 } 1091 toep->txsd_avail--; 1092 1093 t4_l2t_send(sc, wr, toep->l2te); 1094 } 1095 1096 /* Send a FIN if requested, but only if there are no more PDUs to send */ 1097 if (mbufq_first(pduq) == NULL && toep->flags & TPF_SEND_FIN) 1098 t4_close_conn(sc, toep); 1099 } 1100 1101 static inline void 1102 t4_push_data(struct adapter *sc, struct toepcb *toep, int drop) 1103 { 1104 1105 if (ulp_mode(toep) == ULP_MODE_ISCSI) 1106 t4_push_pdus(sc, toep, drop); 1107 else if (tls_tx_key(toep) && toep->tls.mode == TLS_MODE_TLSOM) 1108 t4_push_tls_records(sc, toep, drop); 1109 #ifdef KERN_TLS 1110 else if (toep->flags & TPF_KTLS) 1111 t4_push_ktls(sc, toep, drop); 1112 #endif 1113 else 1114 t4_push_frames(sc, toep, drop); 1115 } 1116 1117 int 1118 t4_tod_output(struct toedev *tod, struct tcpcb *tp) 1119 { 1120 struct adapter *sc = tod->tod_softc; 1121 #ifdef INVARIANTS 1122 struct inpcb *inp = tp->t_inpcb; 1123 #endif 1124 struct toepcb *toep = tp->t_toe; 1125 1126 INP_WLOCK_ASSERT(inp); 1127 KASSERT((inp->inp_flags & INP_DROPPED) == 0, 1128 ("%s: inp %p dropped.", __func__, inp)); 1129 KASSERT(toep != NULL, ("%s: toep is NULL", __func__)); 1130 1131 t4_push_data(sc, toep, 0); 1132 1133 return (0); 1134 } 1135 1136 int 1137 t4_send_fin(struct toedev *tod, struct tcpcb *tp) 1138 { 1139 struct adapter *sc = tod->tod_softc; 1140 #ifdef INVARIANTS 1141 struct inpcb *inp = tp->t_inpcb; 1142 #endif 1143 struct toepcb *toep = tp->t_toe; 1144 1145 INP_WLOCK_ASSERT(inp); 1146 KASSERT((inp->inp_flags & INP_DROPPED) == 0, 1147 ("%s: inp %p dropped.", __func__, inp)); 1148 KASSERT(toep != NULL, ("%s: toep is NULL", __func__)); 1149 1150 toep->flags |= TPF_SEND_FIN; 1151 if (tp->t_state >= TCPS_ESTABLISHED) 1152 t4_push_data(sc, toep, 0); 1153 1154 return (0); 1155 } 1156 1157 int 1158 t4_send_rst(struct toedev *tod, struct tcpcb *tp) 1159 { 1160 struct adapter *sc = tod->tod_softc; 1161 #if defined(INVARIANTS) 1162 struct inpcb *inp = tp->t_inpcb; 1163 #endif 1164 struct toepcb *toep = tp->t_toe; 1165 1166 INP_WLOCK_ASSERT(inp); 1167 KASSERT((inp->inp_flags & INP_DROPPED) == 0, 1168 ("%s: inp %p dropped.", __func__, inp)); 1169 KASSERT(toep != NULL, ("%s: toep is NULL", __func__)); 1170 1171 /* hmmmm */ 1172 KASSERT(toep->flags & TPF_FLOWC_WR_SENT, 1173 ("%s: flowc for tid %u [%s] not sent already", 1174 __func__, toep->tid, tcpstates[tp->t_state])); 1175 1176 send_reset(sc, toep, 0); 1177 return (0); 1178 } 1179 1180 /* 1181 * Peer has sent us a FIN. 1182 */ 1183 static int 1184 do_peer_close(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) 1185 { 1186 struct adapter *sc = iq->adapter; 1187 const struct cpl_peer_close *cpl = (const void *)(rss + 1); 1188 unsigned int tid = GET_TID(cpl); 1189 struct toepcb *toep = lookup_tid(sc, tid); 1190 struct inpcb *inp = toep->inp; 1191 struct tcpcb *tp = NULL; 1192 struct socket *so; 1193 struct epoch_tracker et; 1194 #ifdef INVARIANTS 1195 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); 1196 #endif 1197 1198 KASSERT(opcode == CPL_PEER_CLOSE, 1199 ("%s: unexpected opcode 0x%x", __func__, opcode)); 1200 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 1201 1202 if (__predict_false(toep->flags & TPF_SYNQE)) { 1203 /* 1204 * do_pass_establish must have run before do_peer_close and if 1205 * this is still a synqe instead of a toepcb then the connection 1206 * must be getting aborted. 1207 */ 1208 MPASS(toep->flags & TPF_ABORT_SHUTDOWN); 1209 CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid, 1210 toep, toep->flags); 1211 return (0); 1212 } 1213 1214 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__)); 1215 1216 CURVNET_SET(toep->vnet); 1217 NET_EPOCH_ENTER(et); 1218 INP_WLOCK(inp); 1219 tp = intotcpcb(inp); 1220 1221 CTR6(KTR_CXGBE, 1222 "%s: tid %u (%s), toep_flags 0x%x, ddp_flags 0x%x, inp %p", 1223 __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags, 1224 toep->ddp.flags, inp); 1225 1226 if (toep->flags & TPF_ABORT_SHUTDOWN) 1227 goto done; 1228 1229 tp->rcv_nxt++; /* FIN */ 1230 1231 so = inp->inp_socket; 1232 socantrcvmore(so); 1233 if (ulp_mode(toep) == ULP_MODE_TCPDDP) { 1234 DDP_LOCK(toep); 1235 if (__predict_false(toep->ddp.flags & 1236 (DDP_BUF0_ACTIVE | DDP_BUF1_ACTIVE))) 1237 handle_ddp_close(toep, tp, cpl->rcv_nxt); 1238 DDP_UNLOCK(toep); 1239 } 1240 1241 if (ulp_mode(toep) != ULP_MODE_RDMA) { 1242 KASSERT(tp->rcv_nxt == be32toh(cpl->rcv_nxt), 1243 ("%s: rcv_nxt mismatch: %u %u", __func__, tp->rcv_nxt, 1244 be32toh(cpl->rcv_nxt))); 1245 } 1246 1247 switch (tp->t_state) { 1248 case TCPS_SYN_RECEIVED: 1249 tp->t_starttime = ticks; 1250 /* FALLTHROUGH */ 1251 1252 case TCPS_ESTABLISHED: 1253 tcp_state_change(tp, TCPS_CLOSE_WAIT); 1254 break; 1255 1256 case TCPS_FIN_WAIT_1: 1257 tcp_state_change(tp, TCPS_CLOSING); 1258 break; 1259 1260 case TCPS_FIN_WAIT_2: 1261 tcp_twstart(tp); 1262 INP_UNLOCK_ASSERT(inp); /* safe, we have a ref on the inp */ 1263 NET_EPOCH_EXIT(et); 1264 CURVNET_RESTORE(); 1265 1266 INP_WLOCK(inp); 1267 final_cpl_received(toep); 1268 return (0); 1269 1270 default: 1271 log(LOG_ERR, "%s: TID %u received CPL_PEER_CLOSE in state %d\n", 1272 __func__, tid, tp->t_state); 1273 } 1274 done: 1275 INP_WUNLOCK(inp); 1276 NET_EPOCH_EXIT(et); 1277 CURVNET_RESTORE(); 1278 return (0); 1279 } 1280 1281 /* 1282 * Peer has ACK'd our FIN. 1283 */ 1284 static int 1285 do_close_con_rpl(struct sge_iq *iq, const struct rss_header *rss, 1286 struct mbuf *m) 1287 { 1288 struct adapter *sc = iq->adapter; 1289 const struct cpl_close_con_rpl *cpl = (const void *)(rss + 1); 1290 unsigned int tid = GET_TID(cpl); 1291 struct toepcb *toep = lookup_tid(sc, tid); 1292 struct inpcb *inp = toep->inp; 1293 struct tcpcb *tp = NULL; 1294 struct socket *so = NULL; 1295 struct epoch_tracker et; 1296 #ifdef INVARIANTS 1297 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); 1298 #endif 1299 1300 KASSERT(opcode == CPL_CLOSE_CON_RPL, 1301 ("%s: unexpected opcode 0x%x", __func__, opcode)); 1302 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 1303 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__)); 1304 1305 CURVNET_SET(toep->vnet); 1306 NET_EPOCH_ENTER(et); 1307 INP_WLOCK(inp); 1308 tp = intotcpcb(inp); 1309 1310 CTR4(KTR_CXGBE, "%s: tid %u (%s), toep_flags 0x%x", 1311 __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags); 1312 1313 if (toep->flags & TPF_ABORT_SHUTDOWN) 1314 goto done; 1315 1316 so = inp->inp_socket; 1317 tp->snd_una = be32toh(cpl->snd_nxt) - 1; /* exclude FIN */ 1318 1319 switch (tp->t_state) { 1320 case TCPS_CLOSING: /* see TCPS_FIN_WAIT_2 in do_peer_close too */ 1321 tcp_twstart(tp); 1322 release: 1323 INP_UNLOCK_ASSERT(inp); /* safe, we have a ref on the inp */ 1324 NET_EPOCH_EXIT(et); 1325 CURVNET_RESTORE(); 1326 1327 INP_WLOCK(inp); 1328 final_cpl_received(toep); /* no more CPLs expected */ 1329 1330 return (0); 1331 case TCPS_LAST_ACK: 1332 if (tcp_close(tp)) 1333 INP_WUNLOCK(inp); 1334 goto release; 1335 1336 case TCPS_FIN_WAIT_1: 1337 if (so->so_rcv.sb_state & SBS_CANTRCVMORE) 1338 soisdisconnected(so); 1339 tcp_state_change(tp, TCPS_FIN_WAIT_2); 1340 break; 1341 1342 default: 1343 log(LOG_ERR, 1344 "%s: TID %u received CPL_CLOSE_CON_RPL in state %s\n", 1345 __func__, tid, tcpstates[tp->t_state]); 1346 } 1347 done: 1348 INP_WUNLOCK(inp); 1349 NET_EPOCH_EXIT(et); 1350 CURVNET_RESTORE(); 1351 return (0); 1352 } 1353 1354 void 1355 send_abort_rpl(struct adapter *sc, struct sge_wrq *ofld_txq, int tid, 1356 int rst_status) 1357 { 1358 struct wrqe *wr; 1359 struct cpl_abort_rpl *cpl; 1360 1361 wr = alloc_wrqe(sizeof(*cpl), ofld_txq); 1362 if (wr == NULL) { 1363 /* XXX */ 1364 panic("%s: allocation failure.", __func__); 1365 } 1366 cpl = wrtod(wr); 1367 1368 INIT_TP_WR_MIT_CPL(cpl, CPL_ABORT_RPL, tid); 1369 cpl->cmd = rst_status; 1370 1371 t4_wrq_tx(sc, wr); 1372 } 1373 1374 static int 1375 abort_status_to_errno(struct tcpcb *tp, unsigned int abort_reason) 1376 { 1377 switch (abort_reason) { 1378 case CPL_ERR_BAD_SYN: 1379 case CPL_ERR_CONN_RESET: 1380 return (tp->t_state == TCPS_CLOSE_WAIT ? EPIPE : ECONNRESET); 1381 case CPL_ERR_XMIT_TIMEDOUT: 1382 case CPL_ERR_PERSIST_TIMEDOUT: 1383 case CPL_ERR_FINWAIT2_TIMEDOUT: 1384 case CPL_ERR_KEEPALIVE_TIMEDOUT: 1385 return (ETIMEDOUT); 1386 default: 1387 return (EIO); 1388 } 1389 } 1390 1391 /* 1392 * TCP RST from the peer, timeout, or some other such critical error. 1393 */ 1394 static int 1395 do_abort_req(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) 1396 { 1397 struct adapter *sc = iq->adapter; 1398 const struct cpl_abort_req_rss *cpl = (const void *)(rss + 1); 1399 unsigned int tid = GET_TID(cpl); 1400 struct toepcb *toep = lookup_tid(sc, tid); 1401 struct sge_wrq *ofld_txq = toep->ofld_txq; 1402 struct inpcb *inp; 1403 struct tcpcb *tp; 1404 struct epoch_tracker et; 1405 #ifdef INVARIANTS 1406 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); 1407 #endif 1408 1409 KASSERT(opcode == CPL_ABORT_REQ_RSS, 1410 ("%s: unexpected opcode 0x%x", __func__, opcode)); 1411 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 1412 1413 if (toep->flags & TPF_SYNQE) 1414 return (do_abort_req_synqe(iq, rss, m)); 1415 1416 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__)); 1417 1418 if (negative_advice(cpl->status)) { 1419 CTR4(KTR_CXGBE, "%s: negative advice %d for tid %d (0x%x)", 1420 __func__, cpl->status, tid, toep->flags); 1421 return (0); /* Ignore negative advice */ 1422 } 1423 1424 inp = toep->inp; 1425 CURVNET_SET(toep->vnet); 1426 NET_EPOCH_ENTER(et); /* for tcp_close */ 1427 INP_WLOCK(inp); 1428 1429 tp = intotcpcb(inp); 1430 1431 CTR6(KTR_CXGBE, 1432 "%s: tid %d (%s), toep_flags 0x%x, inp_flags 0x%x, status %d", 1433 __func__, tid, tp ? tcpstates[tp->t_state] : "no tp", toep->flags, 1434 inp->inp_flags, cpl->status); 1435 1436 /* 1437 * If we'd initiated an abort earlier the reply to it is responsible for 1438 * cleaning up resources. Otherwise we tear everything down right here 1439 * right now. We owe the T4 a CPL_ABORT_RPL no matter what. 1440 */ 1441 if (toep->flags & TPF_ABORT_SHUTDOWN) { 1442 INP_WUNLOCK(inp); 1443 goto done; 1444 } 1445 toep->flags |= TPF_ABORT_SHUTDOWN; 1446 1447 if ((inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) == 0) { 1448 struct socket *so = inp->inp_socket; 1449 1450 if (so != NULL) 1451 so_error_set(so, abort_status_to_errno(tp, 1452 cpl->status)); 1453 tp = tcp_close(tp); 1454 if (tp == NULL) 1455 INP_WLOCK(inp); /* re-acquire */ 1456 } 1457 1458 final_cpl_received(toep); 1459 done: 1460 NET_EPOCH_EXIT(et); 1461 CURVNET_RESTORE(); 1462 send_abort_rpl(sc, ofld_txq, tid, CPL_ABORT_NO_RST); 1463 return (0); 1464 } 1465 1466 /* 1467 * Reply to the CPL_ABORT_REQ (send_reset) 1468 */ 1469 static int 1470 do_abort_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) 1471 { 1472 struct adapter *sc = iq->adapter; 1473 const struct cpl_abort_rpl_rss *cpl = (const void *)(rss + 1); 1474 unsigned int tid = GET_TID(cpl); 1475 struct toepcb *toep = lookup_tid(sc, tid); 1476 struct inpcb *inp = toep->inp; 1477 #ifdef INVARIANTS 1478 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); 1479 #endif 1480 1481 KASSERT(opcode == CPL_ABORT_RPL_RSS, 1482 ("%s: unexpected opcode 0x%x", __func__, opcode)); 1483 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 1484 1485 if (toep->flags & TPF_SYNQE) 1486 return (do_abort_rpl_synqe(iq, rss, m)); 1487 1488 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__)); 1489 1490 CTR5(KTR_CXGBE, "%s: tid %u, toep %p, inp %p, status %d", 1491 __func__, tid, toep, inp, cpl->status); 1492 1493 KASSERT(toep->flags & TPF_ABORT_SHUTDOWN, 1494 ("%s: wasn't expecting abort reply", __func__)); 1495 1496 INP_WLOCK(inp); 1497 final_cpl_received(toep); 1498 1499 return (0); 1500 } 1501 1502 static int 1503 do_rx_data(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) 1504 { 1505 struct adapter *sc = iq->adapter; 1506 const struct cpl_rx_data *cpl = mtod(m, const void *); 1507 unsigned int tid = GET_TID(cpl); 1508 struct toepcb *toep = lookup_tid(sc, tid); 1509 struct inpcb *inp = toep->inp; 1510 struct tcpcb *tp; 1511 struct socket *so; 1512 struct sockbuf *sb; 1513 struct epoch_tracker et; 1514 int len, rx_credits; 1515 uint32_t ddp_placed = 0; 1516 1517 if (__predict_false(toep->flags & TPF_SYNQE)) { 1518 /* 1519 * do_pass_establish must have run before do_rx_data and if this 1520 * is still a synqe instead of a toepcb then the connection must 1521 * be getting aborted. 1522 */ 1523 MPASS(toep->flags & TPF_ABORT_SHUTDOWN); 1524 CTR4(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x)", __func__, tid, 1525 toep, toep->flags); 1526 m_freem(m); 1527 return (0); 1528 } 1529 1530 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__)); 1531 1532 /* strip off CPL header */ 1533 m_adj(m, sizeof(*cpl)); 1534 len = m->m_pkthdr.len; 1535 1536 INP_WLOCK(inp); 1537 if (inp->inp_flags & (INP_DROPPED | INP_TIMEWAIT)) { 1538 CTR4(KTR_CXGBE, "%s: tid %u, rx (%d bytes), inp_flags 0x%x", 1539 __func__, tid, len, inp->inp_flags); 1540 INP_WUNLOCK(inp); 1541 m_freem(m); 1542 return (0); 1543 } 1544 1545 tp = intotcpcb(inp); 1546 1547 if (__predict_false(tp->rcv_nxt != be32toh(cpl->seq))) 1548 ddp_placed = be32toh(cpl->seq) - tp->rcv_nxt; 1549 1550 tp->rcv_nxt += len; 1551 if (tp->rcv_wnd < len) { 1552 KASSERT(ulp_mode(toep) == ULP_MODE_RDMA, 1553 ("%s: negative window size", __func__)); 1554 } 1555 1556 tp->rcv_wnd -= len; 1557 tp->t_rcvtime = ticks; 1558 1559 if (ulp_mode(toep) == ULP_MODE_TCPDDP) 1560 DDP_LOCK(toep); 1561 so = inp_inpcbtosocket(inp); 1562 sb = &so->so_rcv; 1563 SOCKBUF_LOCK(sb); 1564 1565 if (__predict_false(sb->sb_state & SBS_CANTRCVMORE)) { 1566 CTR3(KTR_CXGBE, "%s: tid %u, excess rx (%d bytes)", 1567 __func__, tid, len); 1568 m_freem(m); 1569 SOCKBUF_UNLOCK(sb); 1570 if (ulp_mode(toep) == ULP_MODE_TCPDDP) 1571 DDP_UNLOCK(toep); 1572 INP_WUNLOCK(inp); 1573 1574 CURVNET_SET(toep->vnet); 1575 NET_EPOCH_ENTER(et); 1576 INP_WLOCK(inp); 1577 tp = tcp_drop(tp, ECONNRESET); 1578 if (tp) 1579 INP_WUNLOCK(inp); 1580 NET_EPOCH_EXIT(et); 1581 CURVNET_RESTORE(); 1582 1583 return (0); 1584 } 1585 1586 /* receive buffer autosize */ 1587 MPASS(toep->vnet == so->so_vnet); 1588 CURVNET_SET(toep->vnet); 1589 if (sb->sb_flags & SB_AUTOSIZE && 1590 V_tcp_do_autorcvbuf && 1591 sb->sb_hiwat < V_tcp_autorcvbuf_max && 1592 len > (sbspace(sb) / 8 * 7)) { 1593 unsigned int hiwat = sb->sb_hiwat; 1594 unsigned int newsize = min(hiwat + sc->tt.autorcvbuf_inc, 1595 V_tcp_autorcvbuf_max); 1596 1597 if (!sbreserve_locked(sb, newsize, so, NULL)) 1598 sb->sb_flags &= ~SB_AUTOSIZE; 1599 } 1600 1601 if (ulp_mode(toep) == ULP_MODE_TCPDDP) { 1602 int changed = !(toep->ddp.flags & DDP_ON) ^ cpl->ddp_off; 1603 1604 if (toep->ddp.waiting_count != 0 || toep->ddp.active_count != 0) 1605 CTR3(KTR_CXGBE, "%s: tid %u, non-ddp rx (%d bytes)", 1606 __func__, tid, len); 1607 1608 if (changed) { 1609 if (toep->ddp.flags & DDP_SC_REQ) 1610 toep->ddp.flags ^= DDP_ON | DDP_SC_REQ; 1611 else { 1612 KASSERT(cpl->ddp_off == 1, 1613 ("%s: DDP switched on by itself.", 1614 __func__)); 1615 1616 /* Fell out of DDP mode */ 1617 toep->ddp.flags &= ~DDP_ON; 1618 CTR1(KTR_CXGBE, "%s: fell out of DDP mode", 1619 __func__); 1620 1621 insert_ddp_data(toep, ddp_placed); 1622 } 1623 } 1624 1625 if (toep->ddp.flags & DDP_ON) { 1626 /* 1627 * CPL_RX_DATA with DDP on can only be an indicate. 1628 * Start posting queued AIO requests via DDP. The 1629 * payload that arrived in this indicate is appended 1630 * to the socket buffer as usual. 1631 */ 1632 handle_ddp_indicate(toep); 1633 } 1634 } 1635 1636 sbappendstream_locked(sb, m, 0); 1637 rx_credits = sbspace(sb) > tp->rcv_wnd ? sbspace(sb) - tp->rcv_wnd : 0; 1638 if (rx_credits > 0 && sbused(sb) + tp->rcv_wnd < sb->sb_lowat) { 1639 rx_credits = send_rx_credits(sc, toep, rx_credits); 1640 tp->rcv_wnd += rx_credits; 1641 tp->rcv_adv += rx_credits; 1642 } 1643 1644 if (ulp_mode(toep) == ULP_MODE_TCPDDP && toep->ddp.waiting_count > 0 && 1645 sbavail(sb) != 0) { 1646 CTR2(KTR_CXGBE, "%s: tid %u queueing AIO task", __func__, 1647 tid); 1648 ddp_queue_toep(toep); 1649 } 1650 sorwakeup_locked(so); 1651 SOCKBUF_UNLOCK_ASSERT(sb); 1652 if (ulp_mode(toep) == ULP_MODE_TCPDDP) 1653 DDP_UNLOCK(toep); 1654 1655 INP_WUNLOCK(inp); 1656 CURVNET_RESTORE(); 1657 return (0); 1658 } 1659 1660 static int 1661 do_fw4_ack(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) 1662 { 1663 struct adapter *sc = iq->adapter; 1664 const struct cpl_fw4_ack *cpl = (const void *)(rss + 1); 1665 unsigned int tid = G_CPL_FW4_ACK_FLOWID(be32toh(OPCODE_TID(cpl))); 1666 struct toepcb *toep = lookup_tid(sc, tid); 1667 struct inpcb *inp; 1668 struct tcpcb *tp; 1669 struct socket *so; 1670 uint8_t credits = cpl->credits; 1671 struct ofld_tx_sdesc *txsd; 1672 int plen; 1673 #ifdef INVARIANTS 1674 unsigned int opcode = G_CPL_FW4_ACK_OPCODE(be32toh(OPCODE_TID(cpl))); 1675 #endif 1676 1677 /* 1678 * Very unusual case: we'd sent a flowc + abort_req for a synq entry and 1679 * now this comes back carrying the credits for the flowc. 1680 */ 1681 if (__predict_false(toep->flags & TPF_SYNQE)) { 1682 KASSERT(toep->flags & TPF_ABORT_SHUTDOWN, 1683 ("%s: credits for a synq entry %p", __func__, toep)); 1684 return (0); 1685 } 1686 1687 inp = toep->inp; 1688 1689 KASSERT(opcode == CPL_FW4_ACK, 1690 ("%s: unexpected opcode 0x%x", __func__, opcode)); 1691 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 1692 KASSERT(toep->tid == tid, ("%s: toep tid mismatch", __func__)); 1693 1694 INP_WLOCK(inp); 1695 1696 if (__predict_false(toep->flags & TPF_ABORT_SHUTDOWN)) { 1697 INP_WUNLOCK(inp); 1698 return (0); 1699 } 1700 1701 KASSERT((inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) == 0, 1702 ("%s: inp_flags 0x%x", __func__, inp->inp_flags)); 1703 1704 tp = intotcpcb(inp); 1705 1706 if (cpl->flags & CPL_FW4_ACK_FLAGS_SEQVAL) { 1707 tcp_seq snd_una = be32toh(cpl->snd_una); 1708 1709 #ifdef INVARIANTS 1710 if (__predict_false(SEQ_LT(snd_una, tp->snd_una))) { 1711 log(LOG_ERR, 1712 "%s: unexpected seq# %x for TID %u, snd_una %x\n", 1713 __func__, snd_una, toep->tid, tp->snd_una); 1714 } 1715 #endif 1716 1717 if (tp->snd_una != snd_una) { 1718 tp->snd_una = snd_una; 1719 tp->ts_recent_age = tcp_ts_getticks(); 1720 } 1721 } 1722 1723 #ifdef VERBOSE_TRACES 1724 CTR3(KTR_CXGBE, "%s: tid %d credits %u", __func__, tid, credits); 1725 #endif 1726 so = inp->inp_socket; 1727 txsd = &toep->txsd[toep->txsd_cidx]; 1728 plen = 0; 1729 while (credits) { 1730 KASSERT(credits >= txsd->tx_credits, 1731 ("%s: too many (or partial) credits", __func__)); 1732 credits -= txsd->tx_credits; 1733 toep->tx_credits += txsd->tx_credits; 1734 plen += txsd->plen; 1735 if (txsd->iv_buffer) { 1736 free(txsd->iv_buffer, M_CXGBE); 1737 txsd->iv_buffer = NULL; 1738 } 1739 txsd++; 1740 toep->txsd_avail++; 1741 KASSERT(toep->txsd_avail <= toep->txsd_total, 1742 ("%s: txsd avail > total", __func__)); 1743 if (__predict_false(++toep->txsd_cidx == toep->txsd_total)) { 1744 txsd = &toep->txsd[0]; 1745 toep->txsd_cidx = 0; 1746 } 1747 } 1748 1749 if (toep->tx_credits == toep->tx_total) { 1750 toep->tx_nocompl = 0; 1751 toep->plen_nocompl = 0; 1752 } 1753 1754 if (toep->flags & TPF_TX_SUSPENDED && 1755 toep->tx_credits >= toep->tx_total / 4) { 1756 #ifdef VERBOSE_TRACES 1757 CTR2(KTR_CXGBE, "%s: tid %d calling t4_push_frames", __func__, 1758 tid); 1759 #endif 1760 toep->flags &= ~TPF_TX_SUSPENDED; 1761 CURVNET_SET(toep->vnet); 1762 t4_push_data(sc, toep, plen); 1763 CURVNET_RESTORE(); 1764 } else if (plen > 0) { 1765 struct sockbuf *sb = &so->so_snd; 1766 int sbu; 1767 1768 SOCKBUF_LOCK(sb); 1769 sbu = sbused(sb); 1770 if (ulp_mode(toep) == ULP_MODE_ISCSI) { 1771 1772 if (__predict_false(sbu > 0)) { 1773 /* 1774 * The data trasmitted before the tid's ULP mode 1775 * changed to ISCSI is still in so_snd. 1776 * Incoming credits should account for so_snd 1777 * first. 1778 */ 1779 sbdrop_locked(sb, min(sbu, plen)); 1780 plen -= min(sbu, plen); 1781 } 1782 sowwakeup_locked(so); /* unlocks so_snd */ 1783 rqdrop_locked(&toep->ulp_pdu_reclaimq, plen); 1784 } else { 1785 #ifdef VERBOSE_TRACES 1786 CTR3(KTR_CXGBE, "%s: tid %d dropped %d bytes", __func__, 1787 tid, plen); 1788 #endif 1789 sbdrop_locked(sb, plen); 1790 if (tls_tx_key(toep) && 1791 toep->tls.mode == TLS_MODE_TLSOM) { 1792 struct tls_ofld_info *tls_ofld = &toep->tls; 1793 1794 MPASS(tls_ofld->sb_off >= plen); 1795 tls_ofld->sb_off -= plen; 1796 } 1797 if (!TAILQ_EMPTY(&toep->aiotx_jobq)) 1798 t4_aiotx_queue_toep(so, toep); 1799 sowwakeup_locked(so); /* unlocks so_snd */ 1800 } 1801 SOCKBUF_UNLOCK_ASSERT(sb); 1802 } 1803 1804 INP_WUNLOCK(inp); 1805 1806 return (0); 1807 } 1808 1809 void 1810 t4_set_tcb_field(struct adapter *sc, struct sge_wrq *wrq, struct toepcb *toep, 1811 uint16_t word, uint64_t mask, uint64_t val, int reply, int cookie) 1812 { 1813 struct wrqe *wr; 1814 struct cpl_set_tcb_field *req; 1815 struct ofld_tx_sdesc *txsd; 1816 1817 MPASS((cookie & ~M_COOKIE) == 0); 1818 if (reply) { 1819 MPASS(cookie != CPL_COOKIE_RESERVED); 1820 } 1821 1822 wr = alloc_wrqe(sizeof(*req), wrq); 1823 if (wr == NULL) { 1824 /* XXX */ 1825 panic("%s: allocation failure.", __func__); 1826 } 1827 req = wrtod(wr); 1828 1829 INIT_TP_WR_MIT_CPL(req, CPL_SET_TCB_FIELD, toep->tid); 1830 req->reply_ctrl = htobe16(V_QUEUENO(toep->ofld_rxq->iq.abs_id)); 1831 if (reply == 0) 1832 req->reply_ctrl |= htobe16(F_NO_REPLY); 1833 req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(cookie)); 1834 req->mask = htobe64(mask); 1835 req->val = htobe64(val); 1836 if ((wrq->eq.flags & EQ_TYPEMASK) == EQ_OFLD) { 1837 txsd = &toep->txsd[toep->txsd_pidx]; 1838 txsd->tx_credits = howmany(sizeof(*req), 16); 1839 txsd->plen = 0; 1840 KASSERT(toep->tx_credits >= txsd->tx_credits && 1841 toep->txsd_avail > 0, 1842 ("%s: not enough credits (%d)", __func__, 1843 toep->tx_credits)); 1844 toep->tx_credits -= txsd->tx_credits; 1845 if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) 1846 toep->txsd_pidx = 0; 1847 toep->txsd_avail--; 1848 } 1849 1850 t4_wrq_tx(sc, wr); 1851 } 1852 1853 void 1854 t4_init_cpl_io_handlers(void) 1855 { 1856 1857 t4_register_cpl_handler(CPL_PEER_CLOSE, do_peer_close); 1858 t4_register_cpl_handler(CPL_CLOSE_CON_RPL, do_close_con_rpl); 1859 t4_register_cpl_handler(CPL_ABORT_REQ_RSS, do_abort_req); 1860 t4_register_shared_cpl_handler(CPL_ABORT_RPL_RSS, do_abort_rpl, 1861 CPL_COOKIE_TOM); 1862 t4_register_cpl_handler(CPL_RX_DATA, do_rx_data); 1863 t4_register_shared_cpl_handler(CPL_FW4_ACK, do_fw4_ack, CPL_COOKIE_TOM); 1864 } 1865 1866 void 1867 t4_uninit_cpl_io_handlers(void) 1868 { 1869 1870 t4_register_cpl_handler(CPL_PEER_CLOSE, NULL); 1871 t4_register_cpl_handler(CPL_CLOSE_CON_RPL, NULL); 1872 t4_register_cpl_handler(CPL_ABORT_REQ_RSS, NULL); 1873 t4_register_shared_cpl_handler(CPL_ABORT_RPL_RSS, NULL, CPL_COOKIE_TOM); 1874 t4_register_cpl_handler(CPL_RX_DATA, NULL); 1875 t4_register_shared_cpl_handler(CPL_FW4_ACK, NULL, CPL_COOKIE_TOM); 1876 } 1877 1878 /* 1879 * Use the 'backend1' field in AIO jobs to hold an error that should 1880 * be reported when the job is completed, the 'backend3' field to 1881 * store the amount of data sent by the AIO job so far, and the 1882 * 'backend4' field to hold a reference count on the job. 1883 * 1884 * Each unmapped mbuf holds a reference on the job as does the queue 1885 * so long as the job is queued. 1886 */ 1887 #define aio_error backend1 1888 #define aio_sent backend3 1889 #define aio_refs backend4 1890 1891 #define jobtotid(job) \ 1892 (((struct toepcb *)(so_sototcpcb((job)->fd_file->f_data)->t_toe))->tid) 1893 1894 static void 1895 aiotx_free_job(struct kaiocb *job) 1896 { 1897 long status; 1898 int error; 1899 1900 if (refcount_release(&job->aio_refs) == 0) 1901 return; 1902 1903 error = (intptr_t)job->aio_error; 1904 status = job->aio_sent; 1905 #ifdef VERBOSE_TRACES 1906 CTR5(KTR_CXGBE, "%s: tid %d completed %p len %ld, error %d", __func__, 1907 jobtotid(job), job, status, error); 1908 #endif 1909 if (error != 0 && status != 0) 1910 error = 0; 1911 if (error == ECANCELED) 1912 aio_cancel(job); 1913 else if (error) 1914 aio_complete(job, -1, error); 1915 else { 1916 job->msgsnd = 1; 1917 aio_complete(job, status, 0); 1918 } 1919 } 1920 1921 static void 1922 aiotx_free_pgs(struct mbuf *m) 1923 { 1924 struct mbuf_ext_pgs *ext_pgs; 1925 struct kaiocb *job; 1926 vm_page_t pg; 1927 1928 MBUF_EXT_PGS_ASSERT(m); 1929 ext_pgs = m->m_ext.ext_pgs; 1930 job = m->m_ext.ext_arg1; 1931 #ifdef VERBOSE_TRACES 1932 CTR3(KTR_CXGBE, "%s: completed %d bytes for tid %d", __func__, 1933 m->m_len, jobtotid(job)); 1934 #endif 1935 1936 for (int i = 0; i < ext_pgs->npgs; i++) { 1937 pg = PHYS_TO_VM_PAGE(ext_pgs->pa[i]); 1938 vm_page_unwire(pg, PQ_ACTIVE); 1939 } 1940 1941 aiotx_free_job(job); 1942 } 1943 1944 /* 1945 * Allocate a chain of unmapped mbufs describing the next 'len' bytes 1946 * of an AIO job. 1947 */ 1948 static struct mbuf * 1949 alloc_aiotx_mbuf(struct kaiocb *job, int len) 1950 { 1951 struct vmspace *vm; 1952 vm_page_t pgs[MBUF_PEXT_MAX_PGS]; 1953 struct mbuf *m, *top, *last; 1954 struct mbuf_ext_pgs *ext_pgs; 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, false, aiotx_free_pgs); 1987 if (m == NULL) { 1988 vm_page_unhold_pages(pgs, npages); 1989 break; 1990 } 1991 1992 ext_pgs = m->m_ext.ext_pgs; 1993 ext_pgs->first_pg_off = pgoff; 1994 ext_pgs->npgs = npages; 1995 if (npages == 1) { 1996 KASSERT(mlen + pgoff <= PAGE_SIZE, 1997 ("%s: single page is too large (off %d len %d)", 1998 __func__, pgoff, mlen)); 1999 ext_pgs->last_pg_len = mlen; 2000 } else { 2001 ext_pgs->last_pg_len = mlen - (PAGE_SIZE - pgoff) - 2002 (npages - 2) * PAGE_SIZE; 2003 } 2004 for (i = 0; i < npages; i++) 2005 ext_pgs->pa[i] = VM_PAGE_TO_PHYS(pgs[i]); 2006 2007 m->m_len = mlen; 2008 m->m_ext.ext_size = npages * PAGE_SIZE; 2009 m->m_ext.ext_arg1 = job; 2010 refcount_acquire(&job->aio_refs); 2011 2012 #ifdef VERBOSE_TRACES 2013 CTR5(KTR_CXGBE, "%s: tid %d, new mbuf %p for job %p, npages %d", 2014 __func__, jobtotid(job), m, job, npages); 2015 #endif 2016 2017 if (top == NULL) 2018 top = m; 2019 else 2020 last->m_next = m; 2021 last = m; 2022 2023 len -= mlen; 2024 start += mlen; 2025 pgoff = 0; 2026 } 2027 2028 return (top); 2029 } 2030 2031 static void 2032 t4_aiotx_process_job(struct toepcb *toep, struct socket *so, struct kaiocb *job) 2033 { 2034 struct sockbuf *sb; 2035 struct file *fp; 2036 struct inpcb *inp; 2037 struct tcpcb *tp; 2038 struct mbuf *m; 2039 int error, len; 2040 bool moretocome, sendmore; 2041 2042 sb = &so->so_snd; 2043 SOCKBUF_UNLOCK(sb); 2044 fp = job->fd_file; 2045 m = NULL; 2046 2047 #ifdef MAC 2048 error = mac_socket_check_send(fp->f_cred, so); 2049 if (error != 0) 2050 goto out; 2051 #endif 2052 2053 /* Inline sosend_generic(). */ 2054 2055 error = sblock(sb, SBL_WAIT); 2056 MPASS(error == 0); 2057 2058 sendanother: 2059 SOCKBUF_LOCK(sb); 2060 if (so->so_snd.sb_state & SBS_CANTSENDMORE) { 2061 SOCKBUF_UNLOCK(sb); 2062 sbunlock(sb); 2063 if ((so->so_options & SO_NOSIGPIPE) == 0) { 2064 PROC_LOCK(job->userproc); 2065 kern_psignal(job->userproc, SIGPIPE); 2066 PROC_UNLOCK(job->userproc); 2067 } 2068 error = EPIPE; 2069 goto out; 2070 } 2071 if (so->so_error) { 2072 error = so->so_error; 2073 so->so_error = 0; 2074 SOCKBUF_UNLOCK(sb); 2075 sbunlock(sb); 2076 goto out; 2077 } 2078 if ((so->so_state & SS_ISCONNECTED) == 0) { 2079 SOCKBUF_UNLOCK(sb); 2080 sbunlock(sb); 2081 error = ENOTCONN; 2082 goto out; 2083 } 2084 if (sbspace(sb) < sb->sb_lowat) { 2085 MPASS(job->aio_sent == 0 || !(so->so_state & SS_NBIO)); 2086 2087 /* 2088 * Don't block if there is too little room in the socket 2089 * buffer. Instead, requeue the request. 2090 */ 2091 if (!aio_set_cancel_function(job, t4_aiotx_cancel)) { 2092 SOCKBUF_UNLOCK(sb); 2093 sbunlock(sb); 2094 error = ECANCELED; 2095 goto out; 2096 } 2097 TAILQ_INSERT_HEAD(&toep->aiotx_jobq, job, list); 2098 SOCKBUF_UNLOCK(sb); 2099 sbunlock(sb); 2100 goto out; 2101 } 2102 2103 /* 2104 * Write as much data as the socket permits, but no more than a 2105 * a single sndbuf at a time. 2106 */ 2107 len = sbspace(sb); 2108 if (len > job->uaiocb.aio_nbytes - job->aio_sent) { 2109 len = job->uaiocb.aio_nbytes - job->aio_sent; 2110 moretocome = false; 2111 } else 2112 moretocome = true; 2113 if (len > toep->params.sndbuf) { 2114 len = toep->params.sndbuf; 2115 sendmore = true; 2116 } else 2117 sendmore = false; 2118 2119 if (!TAILQ_EMPTY(&toep->aiotx_jobq)) 2120 moretocome = true; 2121 SOCKBUF_UNLOCK(sb); 2122 MPASS(len != 0); 2123 2124 m = alloc_aiotx_mbuf(job, len); 2125 if (m == NULL) { 2126 sbunlock(sb); 2127 error = EFAULT; 2128 goto out; 2129 } 2130 2131 /* Inlined tcp_usr_send(). */ 2132 2133 inp = toep->inp; 2134 INP_WLOCK(inp); 2135 if (inp->inp_flags & (INP_TIMEWAIT | INP_DROPPED)) { 2136 INP_WUNLOCK(inp); 2137 sbunlock(sb); 2138 error = ECONNRESET; 2139 goto out; 2140 } 2141 2142 job->aio_sent += m_length(m, NULL); 2143 2144 sbappendstream(sb, m, 0); 2145 m = NULL; 2146 2147 if (!(inp->inp_flags & INP_DROPPED)) { 2148 tp = intotcpcb(inp); 2149 if (moretocome) 2150 tp->t_flags |= TF_MORETOCOME; 2151 error = tp->t_fb->tfb_tcp_output(tp); 2152 if (moretocome) 2153 tp->t_flags &= ~TF_MORETOCOME; 2154 } 2155 2156 INP_WUNLOCK(inp); 2157 if (sendmore) 2158 goto sendanother; 2159 sbunlock(sb); 2160 2161 if (error) 2162 goto out; 2163 2164 /* 2165 * If this is a blocking socket and the request has not been 2166 * fully completed, requeue it until the socket is ready 2167 * again. 2168 */ 2169 if (job->aio_sent < job->uaiocb.aio_nbytes && 2170 !(so->so_state & SS_NBIO)) { 2171 SOCKBUF_LOCK(sb); 2172 if (!aio_set_cancel_function(job, t4_aiotx_cancel)) { 2173 SOCKBUF_UNLOCK(sb); 2174 error = ECANCELED; 2175 goto out; 2176 } 2177 TAILQ_INSERT_HEAD(&toep->aiotx_jobq, job, list); 2178 return; 2179 } 2180 2181 /* 2182 * If the request will not be requeued, drop the queue's 2183 * reference to the job. Any mbufs in flight should still 2184 * hold a reference, but this drops the reference that the 2185 * queue owns while it is waiting to queue mbufs to the 2186 * socket. 2187 */ 2188 aiotx_free_job(job); 2189 2190 out: 2191 if (error) { 2192 job->aio_error = (void *)(intptr_t)error; 2193 aiotx_free_job(job); 2194 } 2195 if (m != NULL) 2196 m_free(m); 2197 SOCKBUF_LOCK(sb); 2198 } 2199 2200 static void 2201 t4_aiotx_task(void *context, int pending) 2202 { 2203 struct toepcb *toep = context; 2204 struct socket *so; 2205 struct kaiocb *job; 2206 2207 so = toep->aiotx_so; 2208 CURVNET_SET(toep->vnet); 2209 SOCKBUF_LOCK(&so->so_snd); 2210 while (!TAILQ_EMPTY(&toep->aiotx_jobq) && sowriteable(so)) { 2211 job = TAILQ_FIRST(&toep->aiotx_jobq); 2212 TAILQ_REMOVE(&toep->aiotx_jobq, job, list); 2213 if (!aio_clear_cancel_function(job)) 2214 continue; 2215 2216 t4_aiotx_process_job(toep, so, job); 2217 } 2218 toep->aiotx_so = NULL; 2219 SOCKBUF_UNLOCK(&so->so_snd); 2220 CURVNET_RESTORE(); 2221 2222 free_toepcb(toep); 2223 SOCK_LOCK(so); 2224 sorele(so); 2225 } 2226 2227 static void 2228 t4_aiotx_queue_toep(struct socket *so, struct toepcb *toep) 2229 { 2230 2231 SOCKBUF_LOCK_ASSERT(&toep->inp->inp_socket->so_snd); 2232 #ifdef VERBOSE_TRACES 2233 CTR3(KTR_CXGBE, "%s: queueing aiotx task for tid %d, active = %s", 2234 __func__, toep->tid, toep->aiotx_so != NULL ? "true" : "false"); 2235 #endif 2236 if (toep->aiotx_so != NULL) 2237 return; 2238 soref(so); 2239 toep->aiotx_so = so; 2240 hold_toepcb(toep); 2241 soaio_enqueue(&toep->aiotx_task); 2242 } 2243 2244 static void 2245 t4_aiotx_cancel(struct kaiocb *job) 2246 { 2247 struct socket *so; 2248 struct sockbuf *sb; 2249 struct tcpcb *tp; 2250 struct toepcb *toep; 2251 2252 so = job->fd_file->f_data; 2253 tp = so_sototcpcb(so); 2254 toep = tp->t_toe; 2255 MPASS(job->uaiocb.aio_lio_opcode == LIO_WRITE); 2256 sb = &so->so_snd; 2257 2258 SOCKBUF_LOCK(sb); 2259 if (!aio_cancel_cleared(job)) 2260 TAILQ_REMOVE(&toep->aiotx_jobq, job, list); 2261 SOCKBUF_UNLOCK(sb); 2262 2263 job->aio_error = (void *)(intptr_t)ECANCELED; 2264 aiotx_free_job(job); 2265 } 2266 2267 int 2268 t4_aio_queue_aiotx(struct socket *so, struct kaiocb *job) 2269 { 2270 struct tcpcb *tp = so_sototcpcb(so); 2271 struct toepcb *toep = tp->t_toe; 2272 struct adapter *sc = td_adapter(toep->td); 2273 2274 /* This only handles writes. */ 2275 if (job->uaiocb.aio_lio_opcode != LIO_WRITE) 2276 return (EOPNOTSUPP); 2277 2278 if (!sc->tt.tx_zcopy) 2279 return (EOPNOTSUPP); 2280 2281 if (tls_tx_key(toep)) 2282 return (EOPNOTSUPP); 2283 2284 SOCKBUF_LOCK(&so->so_snd); 2285 #ifdef VERBOSE_TRACES 2286 CTR3(KTR_CXGBE, "%s: queueing %p for tid %u", __func__, job, toep->tid); 2287 #endif 2288 if (!aio_set_cancel_function(job, t4_aiotx_cancel)) 2289 panic("new job was cancelled"); 2290 refcount_init(&job->aio_refs, 1); 2291 TAILQ_INSERT_TAIL(&toep->aiotx_jobq, job, list); 2292 if (sowriteable(so)) 2293 t4_aiotx_queue_toep(so, toep); 2294 SOCKBUF_UNLOCK(&so->so_snd); 2295 return (0); 2296 } 2297 2298 void 2299 aiotx_init_toep(struct toepcb *toep) 2300 { 2301 2302 TAILQ_INIT(&toep->aiotx_jobq); 2303 TASK_INIT(&toep->aiotx_task, 0, t4_aiotx_task, toep); 2304 } 2305 #endif 2306