1/*- 2 * Copyright (c) 2012 Chelsio Communications, Inc. 3 * All rights reserved. 4 * Written by: Navdeep Parhar <np@FreeBSD.org> 5 * 6 * Redistribution and use in source and binary forms, with or without 7 * modification, are permitted provided that the following conditions 8 * are met: 9 * 1. Redistributions of source code must retain the above copyright 10 * notice, this list of conditions and the following disclaimer. 11 * 2. Redistributions in binary form must reproduce the above copyright 12 * notice, this list of conditions and the following disclaimer in the 13 * documentation and/or other materials provided with the distribution. 14 * 15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 25 * SUCH DAMAGE. 26 */ 27 28#include <sys/cdefs.h> 29__FBSDID("$FreeBSD$"); 30 31#include "opt_inet.h" 32#include "opt_inet6.h" 33 34#ifdef TCP_OFFLOAD 35#include <sys/param.h> 36#include <sys/types.h> 37#include <sys/kernel.h> 38#include <sys/ktr.h> 39#include <sys/module.h> 40#include <sys/protosw.h> 41#include <sys/refcount.h> 42#include <sys/domain.h> 43#include <sys/fnv_hash.h> 44#include <sys/socket.h> 45#include <sys/socketvar.h> 46#include <net/ethernet.h> 47#include <net/if.h> 48#include <net/if_types.h> 49#include <net/if_vlan_var.h> 50#include <net/route.h> 51#include <netinet/in.h> 52#include <netinet/in_pcb.h> 53#include <netinet/ip.h> 54#include <netinet/ip6.h> 55#include <netinet6/scope6_var.h> 56#include <netinet/tcp_timer.h> 57#include <netinet/tcp_var.h> 58#define TCPSTATES 59#include <netinet/tcp_fsm.h> 60#include <netinet/toecore.h> 61 62#include "common/common.h" 63#include "common/t4_msg.h" 64#include "common/t4_regs.h" 65#include "tom/t4_tom_l2t.h" 66#include "tom/t4_tom.h" 67 68/* stid services */ 69static int alloc_stid(struct adapter *, struct listen_ctx *, int); 70static struct listen_ctx *lookup_stid(struct adapter *, int); 71static void free_stid(struct adapter *, struct listen_ctx *); 72 73/* lctx services */ 74static struct listen_ctx *alloc_lctx(struct adapter *, struct inpcb *, 75 struct port_info *); 76static int free_lctx(struct adapter *, struct listen_ctx *); 77static void hold_lctx(struct listen_ctx *); 78static void listen_hash_add(struct adapter *, struct listen_ctx *); 79static struct listen_ctx *listen_hash_find(struct adapter *, struct inpcb *); 80static struct listen_ctx *listen_hash_del(struct adapter *, struct inpcb *); 81static struct inpcb *release_lctx(struct adapter *, struct listen_ctx *); 82 83static inline void save_qids_in_mbuf(struct mbuf *, struct port_info *); 84static inline void get_qids_from_mbuf(struct mbuf *m, int *, int *); 85static void send_reset_synqe(struct toedev *, struct synq_entry *); 86 87static int 88alloc_stid(struct adapter *sc, struct listen_ctx *lctx, int isipv6) 89{ 90 struct tid_info *t = &sc->tids; 91 u_int stid, n, f, mask; 92 struct stid_region *sr = &lctx->stid_region; 93 94 /* 95 * An IPv6 server needs 2 naturally aligned stids (1 stid = 4 cells) in 96 * the TCAM. The start of the stid region is properly aligned (the chip 97 * requires each region to be 128-cell aligned). 98 */ 99 n = isipv6 ? 2 : 1; 100 mask = n - 1; 101 KASSERT((t->stid_base & mask) == 0 && (t->nstids & mask) == 0, 102 ("%s: stid region (%u, %u) not properly aligned. n = %u", 103 __func__, t->stid_base, t->nstids, n)); 104 105 mtx_lock(&t->stid_lock); 106 if (n > t->nstids - t->stids_in_use) { 107 mtx_unlock(&t->stid_lock); 108 return (-1); 109 } 110 111 if (t->nstids_free_head >= n) { 112 /* 113 * This allocation will definitely succeed because the region 114 * starts at a good alignment and we just checked we have enough 115 * stids free. 116 */ 117 f = t->nstids_free_head & mask; 118 t->nstids_free_head -= n + f; 119 stid = t->nstids_free_head; 120 TAILQ_INSERT_HEAD(&t->stids, sr, link); 121 } else { 122 struct stid_region *s; 123 124 stid = t->nstids_free_head; 125 TAILQ_FOREACH(s, &t->stids, link) { 126 stid += s->used + s->free; 127 f = stid & mask; 128 if (s->free >= n + f) { 129 stid -= n + f; 130 s->free -= n + f; 131 TAILQ_INSERT_AFTER(&t->stids, s, sr, link); 132 goto allocated; 133 } 134 } 135 136 if (__predict_false(stid != t->nstids)) { 137 panic("%s: stids TAILQ (%p) corrupt." 138 " At %d instead of %d at the end of the queue.", 139 __func__, &t->stids, stid, t->nstids); 140 } 141 142 mtx_unlock(&t->stid_lock); 143 return (-1); 144 } 145 146allocated: 147 sr->used = n; 148 sr->free = f; 149 t->stids_in_use += n; 150 t->stid_tab[stid] = lctx; 151 mtx_unlock(&t->stid_lock); 152 153 KASSERT(((stid + t->stid_base) & mask) == 0, 154 ("%s: EDOOFUS.", __func__)); 155 return (stid + t->stid_base); 156} 157 158static struct listen_ctx * 159lookup_stid(struct adapter *sc, int stid) 160{ 161 struct tid_info *t = &sc->tids; 162 163 return (t->stid_tab[stid - t->stid_base]); 164} 165 166static void 167free_stid(struct adapter *sc, struct listen_ctx *lctx) 168{ 169 struct tid_info *t = &sc->tids; 170 struct stid_region *sr = &lctx->stid_region; 171 struct stid_region *s; 172 173 KASSERT(sr->used > 0, ("%s: nonsense free (%d)", __func__, sr->used)); 174 175 mtx_lock(&t->stid_lock); 176 s = TAILQ_PREV(sr, stid_head, link); 177 if (s != NULL) 178 s->free += sr->used + sr->free; 179 else 180 t->nstids_free_head += sr->used + sr->free; 181 KASSERT(t->stids_in_use >= sr->used, 182 ("%s: stids_in_use (%u) < stids being freed (%u)", __func__, 183 t->stids_in_use, sr->used)); 184 t->stids_in_use -= sr->used; 185 TAILQ_REMOVE(&t->stids, sr, link); 186 mtx_unlock(&t->stid_lock); 187} 188 189static struct listen_ctx * 190alloc_lctx(struct adapter *sc, struct inpcb *inp, struct port_info *pi) 191{ 192 struct listen_ctx *lctx; 193 194 INP_WLOCK_ASSERT(inp); 195 196 lctx = malloc(sizeof(struct listen_ctx), M_CXGBE, M_NOWAIT | M_ZERO); 197 if (lctx == NULL) 198 return (NULL); 199 200 lctx->stid = alloc_stid(sc, lctx, inp->inp_vflag & INP_IPV6); 201 if (lctx->stid < 0) { 202 free(lctx, M_CXGBE); 203 return (NULL); 204 } 205 206 if (inp->inp_vflag & INP_IPV6 && 207 !IN6_ARE_ADDR_EQUAL(&in6addr_any, &inp->in6p_laddr)) { 208 struct tom_data *td = sc->tom_softc; 209 210 lctx->ce = hold_lip(td, &inp->in6p_laddr); 211 if (lctx->ce == NULL) { 212 free(lctx, M_CXGBE); 213 return (NULL); 214 } 215 } 216 217 lctx->ctrlq = &sc->sge.ctrlq[pi->port_id]; 218 lctx->ofld_rxq = &sc->sge.ofld_rxq[pi->first_ofld_rxq]; 219 refcount_init(&lctx->refcount, 1); 220 TAILQ_INIT(&lctx->synq); 221 222 lctx->inp = inp; 223 in_pcbref(inp); 224 225 return (lctx); 226} 227 228/* Don't call this directly, use release_lctx instead */ 229static int 230free_lctx(struct adapter *sc, struct listen_ctx *lctx) 231{ 232 struct inpcb *inp = lctx->inp; 233 struct tom_data *td = sc->tom_softc; 234 235 INP_WLOCK_ASSERT(inp); 236 KASSERT(lctx->refcount == 0, 237 ("%s: refcount %d", __func__, lctx->refcount)); 238 KASSERT(TAILQ_EMPTY(&lctx->synq), 239 ("%s: synq not empty.", __func__)); 240 KASSERT(lctx->stid >= 0, ("%s: bad stid %d.", __func__, lctx->stid)); 241 242 CTR4(KTR_CXGBE, "%s: stid %u, lctx %p, inp %p", 243 __func__, lctx->stid, lctx, lctx->inp); 244 245 if (lctx->ce) 246 release_lip(td, lctx->ce); 247 free_stid(sc, lctx); 248 free(lctx, M_CXGBE); 249 250 return (in_pcbrele_wlocked(inp)); 251} 252 253static void 254hold_lctx(struct listen_ctx *lctx) 255{ 256 257 refcount_acquire(&lctx->refcount); 258} 259 260static inline uint32_t 261listen_hashfn(void *key, u_long mask) 262{ 263 264 return (fnv_32_buf(&key, sizeof(key), FNV1_32_INIT) & mask); 265} 266 267/* 268 * Add a listen_ctx entry to the listen hash table. 269 */ 270static void 271listen_hash_add(struct adapter *sc, struct listen_ctx *lctx) 272{ 273 struct tom_data *td = sc->tom_softc; 274 int bucket = listen_hashfn(lctx->inp, td->listen_mask); 275 276 mtx_lock(&td->lctx_hash_lock); 277 LIST_INSERT_HEAD(&td->listen_hash[bucket], lctx, link); 278 td->lctx_count++; 279 mtx_unlock(&td->lctx_hash_lock); 280} 281 282/* 283 * Look for the listening socket's context entry in the hash and return it. 284 */ 285static struct listen_ctx * 286listen_hash_find(struct adapter *sc, struct inpcb *inp) 287{ 288 struct tom_data *td = sc->tom_softc; 289 int bucket = listen_hashfn(inp, td->listen_mask); 290 struct listen_ctx *lctx; 291 292 mtx_lock(&td->lctx_hash_lock); 293 LIST_FOREACH(lctx, &td->listen_hash[bucket], link) { 294 if (lctx->inp == inp) 295 break; 296 } 297 mtx_unlock(&td->lctx_hash_lock); 298 299 return (lctx); 300} 301 302/* 303 * Removes the listen_ctx structure for inp from the hash and returns it. 304 */ 305static struct listen_ctx * 306listen_hash_del(struct adapter *sc, struct inpcb *inp) 307{ 308 struct tom_data *td = sc->tom_softc; 309 int bucket = listen_hashfn(inp, td->listen_mask); 310 struct listen_ctx *lctx, *l; 311 312 mtx_lock(&td->lctx_hash_lock); 313 LIST_FOREACH_SAFE(lctx, &td->listen_hash[bucket], link, l) { 314 if (lctx->inp == inp) { 315 LIST_REMOVE(lctx, link); 316 td->lctx_count--; 317 break; 318 } 319 } 320 mtx_unlock(&td->lctx_hash_lock); 321 322 return (lctx); 323} 324 325/* 326 * Releases a hold on the lctx. Must be called with the listening socket's inp 327 * locked. The inp may be freed by this function and it returns NULL to 328 * indicate this. 329 */ 330static struct inpcb * 331release_lctx(struct adapter *sc, struct listen_ctx *lctx) 332{ 333 struct inpcb *inp = lctx->inp; 334 int inp_freed = 0; 335 336 INP_WLOCK_ASSERT(inp); 337 if (refcount_release(&lctx->refcount)) 338 inp_freed = free_lctx(sc, lctx); 339 340 return (inp_freed ? NULL : inp); 341} 342 343static void 344send_reset_synqe(struct toedev *tod, struct synq_entry *synqe) 345{ 346 struct adapter *sc = tod->tod_softc; 347 struct mbuf *m = synqe->syn; 348 struct ifnet *ifp = m->m_pkthdr.rcvif; 349 struct port_info *pi = ifp->if_softc; 350 struct l2t_entry *e = &sc->l2t->l2tab[synqe->l2e_idx]; 351 struct wrqe *wr; 352 struct fw_flowc_wr *flowc; 353 struct cpl_abort_req *req; 354 int txqid, rxqid, flowclen; 355 struct sge_wrq *ofld_txq; 356 struct sge_ofld_rxq *ofld_rxq; 357 const int nparams = 6; 358 unsigned int pfvf = G_FW_VIID_PFN(pi->viid) << S_FW_VIID_PFN; 359 360 INP_WLOCK_ASSERT(synqe->lctx->inp); 361 362 CTR5(KTR_CXGBE, "%s: synqe %p (0x%x), tid %d%s", 363 __func__, synqe, synqe->flags, synqe->tid, 364 synqe->flags & TPF_ABORT_SHUTDOWN ? 365 " (abort already in progress)" : ""); 366 if (synqe->flags & TPF_ABORT_SHUTDOWN) 367 return; /* abort already in progress */ 368 synqe->flags |= TPF_ABORT_SHUTDOWN; 369 370 get_qids_from_mbuf(m, &txqid, &rxqid); 371 ofld_txq = &sc->sge.ofld_txq[txqid]; 372 ofld_rxq = &sc->sge.ofld_rxq[rxqid]; 373 374 /* The wrqe will have two WRs - a flowc followed by an abort_req */ 375 flowclen = sizeof(*flowc) + nparams * sizeof(struct fw_flowc_mnemval); 376 377 wr = alloc_wrqe(roundup2(flowclen, EQ_ESIZE) + sizeof(*req), ofld_txq); 378 if (wr == NULL) { 379 /* XXX */ 380 panic("%s: allocation failure.", __func__); 381 } 382 flowc = wrtod(wr); 383 req = (void *)((caddr_t)flowc + roundup2(flowclen, EQ_ESIZE)); 384 385 /* First the flowc ... */ 386 memset(flowc, 0, wr->wr_len); 387 flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) | 388 V_FW_FLOWC_WR_NPARAMS(nparams)); 389 flowc->flowid_len16 = htonl(V_FW_WR_LEN16(howmany(flowclen, 16)) | 390 V_FW_WR_FLOWID(synqe->tid)); 391 flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_PFNVFN; 392 flowc->mnemval[0].val = htobe32(pfvf); 393 flowc->mnemval[1].mnemonic = FW_FLOWC_MNEM_CH; 394 flowc->mnemval[1].val = htobe32(pi->tx_chan); 395 flowc->mnemval[2].mnemonic = FW_FLOWC_MNEM_PORT; 396 flowc->mnemval[2].val = htobe32(pi->tx_chan); 397 flowc->mnemval[3].mnemonic = FW_FLOWC_MNEM_IQID; 398 flowc->mnemval[3].val = htobe32(ofld_rxq->iq.abs_id); 399 flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_SNDBUF; 400 flowc->mnemval[4].val = htobe32(512); 401 flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_MSS; 402 flowc->mnemval[5].val = htobe32(512); 403 synqe->flags |= TPF_FLOWC_WR_SENT; 404 405 /* ... then ABORT request */ 406 INIT_TP_WR_MIT_CPL(req, CPL_ABORT_REQ, synqe->tid); 407 req->rsvd0 = 0; /* don't have a snd_nxt */ 408 req->rsvd1 = 1; /* no data sent yet */ 409 req->cmd = CPL_ABORT_SEND_RST; 410 411 t4_l2t_send(sc, wr, e); 412} 413 414static int 415create_server(struct adapter *sc, struct listen_ctx *lctx) 416{ 417 struct wrqe *wr; 418 struct cpl_pass_open_req *req; 419 struct inpcb *inp = lctx->inp; 420 421 wr = alloc_wrqe(sizeof(*req), lctx->ctrlq); 422 if (wr == NULL) { 423 log(LOG_ERR, "%s: allocation failure", __func__); 424 return (ENOMEM); 425 } 426 req = wrtod(wr); 427 428 INIT_TP_WR(req, 0); 429 OPCODE_TID(req) = htobe32(MK_OPCODE_TID(CPL_PASS_OPEN_REQ, lctx->stid)); 430 req->local_port = inp->inp_lport; 431 req->peer_port = 0; 432 req->local_ip = inp->inp_laddr.s_addr; 433 req->peer_ip = 0; 434 req->opt0 = htobe64(V_TX_CHAN(lctx->ctrlq->eq.tx_chan)); 435 req->opt1 = htobe64(V_CONN_POLICY(CPL_CONN_POLICY_ASK) | 436 F_SYN_RSS_ENABLE | V_SYN_RSS_QUEUE(lctx->ofld_rxq->iq.abs_id)); 437 438 t4_wrq_tx(sc, wr); 439 return (0); 440} 441 442static int 443create_server6(struct adapter *sc, struct listen_ctx *lctx) 444{ 445 struct wrqe *wr; 446 struct cpl_pass_open_req6 *req; 447 struct inpcb *inp = lctx->inp; 448 449 wr = alloc_wrqe(sizeof(*req), lctx->ctrlq); 450 if (wr == NULL) { 451 log(LOG_ERR, "%s: allocation failure", __func__); 452 return (ENOMEM); 453 } 454 req = wrtod(wr); 455 456 INIT_TP_WR(req, 0); 457 OPCODE_TID(req) = htobe32(MK_OPCODE_TID(CPL_PASS_OPEN_REQ6, lctx->stid)); 458 req->local_port = inp->inp_lport; 459 req->peer_port = 0; 460 req->local_ip_hi = *(uint64_t *)&inp->in6p_laddr.s6_addr[0]; 461 req->local_ip_lo = *(uint64_t *)&inp->in6p_laddr.s6_addr[8]; 462 req->peer_ip_hi = 0; 463 req->peer_ip_lo = 0; 464 req->opt0 = htobe64(V_TX_CHAN(lctx->ctrlq->eq.tx_chan)); 465 req->opt1 = htobe64(V_CONN_POLICY(CPL_CONN_POLICY_ASK) | 466 F_SYN_RSS_ENABLE | V_SYN_RSS_QUEUE(lctx->ofld_rxq->iq.abs_id)); 467 468 t4_wrq_tx(sc, wr); 469 return (0); 470} 471 472static int 473destroy_server(struct adapter *sc, struct listen_ctx *lctx) 474{ 475 struct wrqe *wr; 476 struct cpl_close_listsvr_req *req; 477 478 wr = alloc_wrqe(sizeof(*req), lctx->ctrlq); 479 if (wr == NULL) { 480 /* XXX */ 481 panic("%s: allocation failure.", __func__); 482 } 483 req = wrtod(wr); 484 485 INIT_TP_WR(req, 0); 486 OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_CLOSE_LISTSRV_REQ, 487 lctx->stid)); 488 req->reply_ctrl = htobe16(lctx->ofld_rxq->iq.abs_id); 489 req->rsvd = htobe16(0); 490 491 t4_wrq_tx(sc, wr); 492 return (0); 493} 494 495/* 496 * Start a listening server by sending a passive open request to HW. 497 * 498 * Can't take adapter lock here and access to sc->flags, sc->open_device_map, 499 * sc->offload_map, if_capenable are all race prone. 500 */ 501int 502t4_listen_start(struct toedev *tod, struct tcpcb *tp) 503{ 504 struct adapter *sc = tod->tod_softc; 505 struct port_info *pi; 506 struct inpcb *inp = tp->t_inpcb; 507 struct listen_ctx *lctx; 508 int i, rc; 509 510 INP_WLOCK_ASSERT(inp); 511 512 /* Don't start a hardware listener for any loopback address. */ 513 if (inp->inp_vflag & INP_IPV6 && IN6_IS_ADDR_LOOPBACK(&inp->in6p_laddr)) 514 return (0); 515 if (!(inp->inp_vflag & INP_IPV6) && 516 IN_LOOPBACK(ntohl(inp->inp_laddr.s_addr))) 517 return (0); 518#if 0 519 ADAPTER_LOCK(sc); 520 if (IS_BUSY(sc)) { 521 log(LOG_ERR, "%s: listen request ignored, %s is busy", 522 __func__, device_get_nameunit(sc->dev)); 523 goto done; 524 } 525 526 KASSERT(sc->flags & TOM_INIT_DONE, 527 ("%s: TOM not initialized", __func__)); 528#endif 529 530 if ((sc->open_device_map & sc->offload_map) == 0) 531 goto done; /* no port that's UP with IFCAP_TOE enabled */ 532 533 /* 534 * Find a running port with IFCAP_TOE (4 or 6). We'll use the first 535 * such port's queues to send the passive open and receive the reply to 536 * it. 537 * 538 * XXX: need a way to mark a port in use by offload. if_cxgbe should 539 * then reject any attempt to bring down such a port (and maybe reject 540 * attempts to disable IFCAP_TOE on that port too?). 541 */ 542 for_each_port(sc, i) { 543 if (isset(&sc->open_device_map, i) && 544 sc->port[i]->ifp->if_capenable & IFCAP_TOE) 545 break; 546 } 547 KASSERT(i < sc->params.nports, 548 ("%s: no running port with TOE capability enabled.", __func__)); 549 pi = sc->port[i]; 550 551 if (listen_hash_find(sc, inp) != NULL) 552 goto done; /* already setup */ 553 554 lctx = alloc_lctx(sc, inp, pi); 555 if (lctx == NULL) { 556 log(LOG_ERR, 557 "%s: listen request ignored, %s couldn't allocate lctx\n", 558 __func__, device_get_nameunit(sc->dev)); 559 goto done; 560 } 561 listen_hash_add(sc, lctx); 562 563 CTR6(KTR_CXGBE, "%s: stid %u (%s), lctx %p, inp %p vflag 0x%x", 564 __func__, lctx->stid, tcpstates[tp->t_state], lctx, inp, 565 inp->inp_vflag); 566 567 if (inp->inp_vflag & INP_IPV6) 568 rc = create_server6(sc, lctx); 569 else 570 rc = create_server(sc, lctx); 571 if (rc != 0) { 572 log(LOG_ERR, "%s: %s failed to create hw listener: %d.\n", 573 __func__, device_get_nameunit(sc->dev), rc); 574 (void) listen_hash_del(sc, inp); 575 inp = release_lctx(sc, lctx); 576 /* can't be freed, host stack has a reference */ 577 KASSERT(inp != NULL, ("%s: inp freed", __func__)); 578 goto done; 579 } 580 lctx->flags |= LCTX_RPL_PENDING; 581done: 582#if 0 583 ADAPTER_UNLOCK(sc); 584#endif 585 return (0); 586} 587 588int 589t4_listen_stop(struct toedev *tod, struct tcpcb *tp) 590{ 591 struct listen_ctx *lctx; 592 struct adapter *sc = tod->tod_softc; 593 struct inpcb *inp = tp->t_inpcb; 594 struct synq_entry *synqe; 595 596 INP_WLOCK_ASSERT(inp); 597 598 lctx = listen_hash_del(sc, inp); 599 if (lctx == NULL) 600 return (ENOENT); /* no hardware listener for this inp */ 601 602 CTR4(KTR_CXGBE, "%s: stid %u, lctx %p, flags %x", __func__, lctx->stid, 603 lctx, lctx->flags); 604 605 /* 606 * If the reply to the PASS_OPEN is still pending we'll wait for it to 607 * arrive and clean up when it does. 608 */ 609 if (lctx->flags & LCTX_RPL_PENDING) { 610 KASSERT(TAILQ_EMPTY(&lctx->synq), 611 ("%s: synq not empty.", __func__)); 612 return (EINPROGRESS); 613 } 614 615 /* 616 * The host stack will abort all the connections on the listening 617 * socket's so_comp. It doesn't know about the connections on the synq 618 * so we need to take care of those. 619 */ 620 TAILQ_FOREACH(synqe, &lctx->synq, link) { 621 if (synqe->flags & TPF_SYNQE_HAS_L2TE) 622 send_reset_synqe(tod, synqe); 623 } 624 625 destroy_server(sc, lctx); 626 return (0); 627} 628 629static inline void 630hold_synqe(struct synq_entry *synqe) 631{ 632 633 refcount_acquire(&synqe->refcnt); 634} 635 636static inline void 637release_synqe(struct synq_entry *synqe) 638{ 639 640 if (refcount_release(&synqe->refcnt)) { 641 int needfree = synqe->flags & TPF_SYNQE_NEEDFREE; 642 643 m_freem(synqe->syn); 644 if (needfree) 645 free(synqe, M_CXGBE); 646 } 647} 648 649void 650t4_syncache_added(struct toedev *tod __unused, void *arg) 651{ 652 struct synq_entry *synqe = arg; 653 654 hold_synqe(synqe); 655} 656 657void 658t4_syncache_removed(struct toedev *tod __unused, void *arg) 659{ 660 struct synq_entry *synqe = arg; 661 662 release_synqe(synqe); 663} 664 665/* XXX */ 666extern void tcp_dooptions(struct tcpopt *, u_char *, int, int); 667 668int 669t4_syncache_respond(struct toedev *tod, void *arg, struct mbuf *m) 670{ 671 struct adapter *sc = tod->tod_softc; 672 struct synq_entry *synqe = arg; 673 struct wrqe *wr; 674 struct l2t_entry *e; 675 struct tcpopt to; 676 struct ip *ip = mtod(m, struct ip *); 677 struct tcphdr *th; 678 679 wr = (struct wrqe *)atomic_readandclear_ptr(&synqe->wr); 680 if (wr == NULL) { 681 m_freem(m); 682 return (EALREADY); 683 } 684 685 if (ip->ip_v == IPVERSION) 686 th = (void *)(ip + 1); 687 else 688 th = (void *)((struct ip6_hdr *)ip + 1); 689 bzero(&to, sizeof(to)); 690 tcp_dooptions(&to, (void *)(th + 1), (th->th_off << 2) - sizeof(*th), 691 TO_SYN); 692 693 /* save these for later */ 694 synqe->iss = be32toh(th->th_seq); 695 synqe->ts = to.to_tsval; 696 697 if (is_t5(sc)) { 698 struct cpl_t5_pass_accept_rpl *rpl5 = wrtod(wr); 699 700 rpl5->iss = th->th_seq; 701 } 702 703 e = &sc->l2t->l2tab[synqe->l2e_idx]; 704 t4_l2t_send(sc, wr, e); 705 706 m_freem(m); /* don't need this any more */ 707 return (0); 708} 709 710static int 711do_pass_open_rpl(struct sge_iq *iq, const struct rss_header *rss, 712 struct mbuf *m) 713{ 714 struct adapter *sc = iq->adapter; 715 const struct cpl_pass_open_rpl *cpl = (const void *)(rss + 1); 716 int stid = GET_TID(cpl); 717 unsigned int status = cpl->status; 718 struct listen_ctx *lctx = lookup_stid(sc, stid); 719 struct inpcb *inp = lctx->inp; 720#ifdef INVARIANTS 721 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); 722#endif 723 724 KASSERT(opcode == CPL_PASS_OPEN_RPL, 725 ("%s: unexpected opcode 0x%x", __func__, opcode)); 726 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 727 KASSERT(lctx->stid == stid, ("%s: lctx stid mismatch", __func__)); 728 729 INP_WLOCK(inp); 730 731 CTR4(KTR_CXGBE, "%s: stid %d, status %u, flags 0x%x", 732 __func__, stid, status, lctx->flags); 733 734 lctx->flags &= ~LCTX_RPL_PENDING; 735 736 if (status != CPL_ERR_NONE) 737 log(LOG_ERR, "listener (stid %u) failed: %d\n", stid, status); 738 739#ifdef INVARIANTS 740 /* 741 * If the inp has been dropped (listening socket closed) then 742 * listen_stop must have run and taken the inp out of the hash. 743 */ 744 if (inp->inp_flags & INP_DROPPED) { 745 KASSERT(listen_hash_del(sc, inp) == NULL, 746 ("%s: inp %p still in listen hash", __func__, inp)); 747 } 748#endif 749 750 if (inp->inp_flags & INP_DROPPED && status != CPL_ERR_NONE) { 751 if (release_lctx(sc, lctx) != NULL) 752 INP_WUNLOCK(inp); 753 return (status); 754 } 755 756 /* 757 * Listening socket stopped listening earlier and now the chip tells us 758 * it has started the hardware listener. Stop it; the lctx will be 759 * released in do_close_server_rpl. 760 */ 761 if (inp->inp_flags & INP_DROPPED) { 762 destroy_server(sc, lctx); 763 INP_WUNLOCK(inp); 764 return (status); 765 } 766 767 /* 768 * Failed to start hardware listener. Take inp out of the hash and 769 * release our reference on it. An error message has been logged 770 * already. 771 */ 772 if (status != CPL_ERR_NONE) { 773 listen_hash_del(sc, inp); 774 if (release_lctx(sc, lctx) != NULL) 775 INP_WUNLOCK(inp); 776 return (status); 777 } 778 779 /* hardware listener open for business */ 780 781 INP_WUNLOCK(inp); 782 return (status); 783} 784 785static int 786do_close_server_rpl(struct sge_iq *iq, const struct rss_header *rss, 787 struct mbuf *m) 788{ 789 struct adapter *sc = iq->adapter; 790 const struct cpl_close_listsvr_rpl *cpl = (const void *)(rss + 1); 791 int stid = GET_TID(cpl); 792 unsigned int status = cpl->status; 793 struct listen_ctx *lctx = lookup_stid(sc, stid); 794 struct inpcb *inp = lctx->inp; 795#ifdef INVARIANTS 796 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); 797#endif 798 799 KASSERT(opcode == CPL_CLOSE_LISTSRV_RPL, 800 ("%s: unexpected opcode 0x%x", __func__, opcode)); 801 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 802 KASSERT(lctx->stid == stid, ("%s: lctx stid mismatch", __func__)); 803 804 CTR3(KTR_CXGBE, "%s: stid %u, status %u", __func__, stid, status); 805 806 if (status != CPL_ERR_NONE) { 807 log(LOG_ERR, "%s: failed (%u) to close listener for stid %u\n", 808 __func__, status, stid); 809 return (status); 810 } 811 812 INP_WLOCK(inp); 813 inp = release_lctx(sc, lctx); 814 if (inp != NULL) 815 INP_WUNLOCK(inp); 816 817 return (status); 818} 819 820static void 821done_with_synqe(struct adapter *sc, struct synq_entry *synqe) 822{ 823 struct listen_ctx *lctx = synqe->lctx; 824 struct inpcb *inp = lctx->inp; 825 struct port_info *pi = synqe->syn->m_pkthdr.rcvif->if_softc; 826 struct l2t_entry *e = &sc->l2t->l2tab[synqe->l2e_idx]; 827 828 INP_WLOCK_ASSERT(inp); 829 830 TAILQ_REMOVE(&lctx->synq, synqe, link); 831 inp = release_lctx(sc, lctx); 832 if (inp) 833 INP_WUNLOCK(inp); 834 remove_tid(sc, synqe->tid); 835 release_tid(sc, synqe->tid, &sc->sge.ctrlq[pi->port_id]); 836 t4_l2t_release(e); 837 release_synqe(synqe); /* removed from synq list */ 838} 839 840int 841do_abort_req_synqe(struct sge_iq *iq, const struct rss_header *rss, 842 struct mbuf *m) 843{ 844 struct adapter *sc = iq->adapter; 845 const struct cpl_abort_req_rss *cpl = (const void *)(rss + 1); 846 unsigned int tid = GET_TID(cpl); 847 struct synq_entry *synqe = lookup_tid(sc, tid); 848 struct listen_ctx *lctx = synqe->lctx; 849 struct inpcb *inp = lctx->inp; 850 int txqid; 851 struct sge_wrq *ofld_txq; 852#ifdef INVARIANTS 853 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); 854#endif 855 856 KASSERT(opcode == CPL_ABORT_REQ_RSS, 857 ("%s: unexpected opcode 0x%x", __func__, opcode)); 858 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 859 KASSERT(synqe->tid == tid, ("%s: toep tid mismatch", __func__)); 860 861 CTR6(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x), lctx %p, status %d", 862 __func__, tid, synqe, synqe->flags, synqe->lctx, cpl->status); 863 864 if (negative_advice(cpl->status)) 865 return (0); /* Ignore negative advice */ 866 867 INP_WLOCK(inp); 868 869 get_qids_from_mbuf(synqe->syn, &txqid, NULL); 870 ofld_txq = &sc->sge.ofld_txq[txqid]; 871 872 /* 873 * If we'd initiated an abort earlier the reply to it is responsible for 874 * cleaning up resources. Otherwise we tear everything down right here 875 * right now. We owe the T4 a CPL_ABORT_RPL no matter what. 876 */ 877 if (synqe->flags & TPF_ABORT_SHUTDOWN) { 878 INP_WUNLOCK(inp); 879 goto done; 880 } 881 882 done_with_synqe(sc, synqe); 883 /* inp lock released by done_with_synqe */ 884done: 885 send_abort_rpl(sc, ofld_txq, tid, CPL_ABORT_NO_RST); 886 return (0); 887} 888 889int 890do_abort_rpl_synqe(struct sge_iq *iq, const struct rss_header *rss, 891 struct mbuf *m) 892{ 893 struct adapter *sc = iq->adapter; 894 const struct cpl_abort_rpl_rss *cpl = (const void *)(rss + 1); 895 unsigned int tid = GET_TID(cpl); 896 struct synq_entry *synqe = lookup_tid(sc, tid); 897 struct listen_ctx *lctx = synqe->lctx; 898 struct inpcb *inp = lctx->inp; 899#ifdef INVARIANTS 900 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); 901#endif 902 903 KASSERT(opcode == CPL_ABORT_RPL_RSS, 904 ("%s: unexpected opcode 0x%x", __func__, opcode)); 905 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 906 KASSERT(synqe->tid == tid, ("%s: toep tid mismatch", __func__)); 907 908 CTR6(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x), lctx %p, status %d", 909 __func__, tid, synqe, synqe->flags, synqe->lctx, cpl->status); 910 911 INP_WLOCK(inp); 912 KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN, 913 ("%s: wasn't expecting abort reply for synqe %p (0x%x)", 914 __func__, synqe, synqe->flags)); 915 916 done_with_synqe(sc, synqe); 917 /* inp lock released by done_with_synqe */ 918 919 return (0); 920} 921 922void 923t4_offload_socket(struct toedev *tod, void *arg, struct socket *so) 924{ 925 struct adapter *sc = tod->tod_softc; 926 struct synq_entry *synqe = arg; 927#ifdef INVARIANTS 928 struct inpcb *inp = sotoinpcb(so); 929#endif 930 struct cpl_pass_establish *cpl = mtod(synqe->syn, void *); 931 struct toepcb *toep = *(struct toepcb **)(cpl + 1); 932 933 INP_INFO_LOCK_ASSERT(&V_tcbinfo); /* prevents bad race with accept() */ 934 INP_WLOCK_ASSERT(inp); 935 KASSERT(synqe->flags & TPF_SYNQE, 936 ("%s: %p not a synq_entry?", __func__, arg)); 937 938 offload_socket(so, toep); 939 make_established(toep, cpl->snd_isn, cpl->rcv_isn, cpl->tcp_opt); 940 toep->flags |= TPF_CPL_PENDING; 941 update_tid(sc, synqe->tid, toep); 942 synqe->flags |= TPF_SYNQE_EXPANDED; 943} 944 945static inline void 946save_qids_in_mbuf(struct mbuf *m, struct port_info *pi) 947{ 948 uint32_t txqid, rxqid; 949 950 txqid = (arc4random() % pi->nofldtxq) + pi->first_ofld_txq; 951 rxqid = (arc4random() % pi->nofldrxq) + pi->first_ofld_rxq; 952 953 m->m_pkthdr.flowid = (txqid << 16) | (rxqid & 0xffff); 954} 955 956static inline void 957get_qids_from_mbuf(struct mbuf *m, int *txqid, int *rxqid) 958{ 959 960 if (txqid) 961 *txqid = m->m_pkthdr.flowid >> 16; 962 if (rxqid) 963 *rxqid = m->m_pkthdr.flowid & 0xffff; 964} 965 966/* 967 * Use the trailing space in the mbuf in which the PASS_ACCEPT_REQ arrived to 968 * store some state temporarily. 969 */ 970static struct synq_entry * 971mbuf_to_synqe(struct mbuf *m) 972{ 973 int len = roundup2(sizeof (struct synq_entry), 8); 974 int tspace = M_TRAILINGSPACE(m); 975 struct synq_entry *synqe = NULL; 976 977 if (tspace < len) { 978 synqe = malloc(sizeof(*synqe), M_CXGBE, M_NOWAIT); 979 if (synqe == NULL) 980 return (NULL); 981 synqe->flags = TPF_SYNQE | TPF_SYNQE_NEEDFREE; 982 } else { 983 synqe = (void *)(m->m_data + m->m_len + tspace - len); 984 synqe->flags = TPF_SYNQE; 985 } 986 987 return (synqe); 988} 989 990static void 991t4opt_to_tcpopt(const struct tcp_options *t4opt, struct tcpopt *to) 992{ 993 bzero(to, sizeof(*to)); 994 995 if (t4opt->mss) { 996 to->to_flags |= TOF_MSS; 997 to->to_mss = be16toh(t4opt->mss); 998 } 999 1000 if (t4opt->wsf) { 1001 to->to_flags |= TOF_SCALE; 1002 to->to_wscale = t4opt->wsf; 1003 } 1004 1005 if (t4opt->tstamp) 1006 to->to_flags |= TOF_TS; 1007 1008 if (t4opt->sack) 1009 to->to_flags |= TOF_SACKPERM; 1010} 1011 1012/* 1013 * Options2 for passive open. 1014 */ 1015static uint32_t 1016calc_opt2p(struct adapter *sc, struct port_info *pi, int rxqid, 1017 const struct tcp_options *tcpopt, struct tcphdr *th, int ulp_mode) 1018{ 1019 struct sge_ofld_rxq *ofld_rxq = &sc->sge.ofld_rxq[rxqid]; 1020 uint32_t opt2; 1021 1022 opt2 = V_TX_QUEUE(sc->params.tp.tx_modq[pi->tx_chan]) | 1023 F_RSS_QUEUE_VALID | V_RSS_QUEUE(ofld_rxq->iq.abs_id); 1024 1025 if (V_tcp_do_rfc1323) { 1026 if (tcpopt->tstamp) 1027 opt2 |= F_TSTAMPS_EN; 1028 if (tcpopt->sack) 1029 opt2 |= F_SACK_EN; 1030 if (tcpopt->wsf <= 14) 1031 opt2 |= F_WND_SCALE_EN; 1032 } 1033 1034 if (V_tcp_do_ecn && th->th_flags & (TH_ECE | TH_CWR)) 1035 opt2 |= F_CCTRL_ECN; 1036 1037 /* RX_COALESCE is always a valid value (0 or M_RX_COALESCE). */ 1038 if (is_t4(sc)) 1039 opt2 |= F_RX_COALESCE_VALID; 1040 else { 1041 opt2 |= F_T5_OPT_2_VALID; 1042 opt2 |= F_CONG_CNTRL_VALID; /* OPT_2_ISS really, for T5 */ 1043 } 1044 if (sc->tt.rx_coalesce) 1045 opt2 |= V_RX_COALESCE(M_RX_COALESCE); 1046 1047#ifdef USE_DDP_RX_FLOW_CONTROL 1048 if (ulp_mode == ULP_MODE_TCPDDP) 1049 opt2 |= F_RX_FC_VALID | F_RX_FC_DDP; 1050#endif 1051 1052 return htobe32(opt2); 1053} 1054 1055static void 1056pass_accept_req_to_protohdrs(const struct mbuf *m, struct in_conninfo *inc, 1057 struct tcphdr *th) 1058{ 1059 const struct cpl_pass_accept_req *cpl = mtod(m, const void *); 1060 const struct ether_header *eh; 1061 unsigned int hlen = be32toh(cpl->hdr_len); 1062 uintptr_t l3hdr; 1063 const struct tcphdr *tcp; 1064 1065 eh = (const void *)(cpl + 1); 1066 l3hdr = ((uintptr_t)eh + G_ETH_HDR_LEN(hlen)); 1067 tcp = (const void *)(l3hdr + G_IP_HDR_LEN(hlen)); 1068 1069 if (inc) { 1070 bzero(inc, sizeof(*inc)); 1071 inc->inc_fport = tcp->th_sport; 1072 inc->inc_lport = tcp->th_dport; 1073 if (((struct ip *)l3hdr)->ip_v == IPVERSION) { 1074 const struct ip *ip = (const void *)l3hdr; 1075 1076 inc->inc_faddr = ip->ip_src; 1077 inc->inc_laddr = ip->ip_dst; 1078 } else { 1079 const struct ip6_hdr *ip6 = (const void *)l3hdr; 1080 1081 inc->inc_flags |= INC_ISIPV6; 1082 inc->inc6_faddr = ip6->ip6_src; 1083 inc->inc6_laddr = ip6->ip6_dst; 1084 } 1085 } 1086 1087 if (th) { 1088 bcopy(tcp, th, sizeof(*th)); 1089 tcp_fields_to_host(th); /* just like tcp_input */ 1090 } 1091} 1092 1093static int 1094ifnet_has_ip6(struct ifnet *ifp, struct in6_addr *ip6) 1095{ 1096 struct ifaddr *ifa; 1097 struct sockaddr_in6 *sin6; 1098 int found = 0; 1099 struct in6_addr in6 = *ip6; 1100 1101 /* Just as in ip6_input */ 1102 if (in6_clearscope(&in6) || in6_clearscope(&in6)) 1103 return (0); 1104 in6_setscope(&in6, ifp, NULL); 1105 1106 if_addr_rlock(ifp); 1107 TAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) { 1108 sin6 = (void *)ifa->ifa_addr; 1109 if (sin6->sin6_family != AF_INET6) 1110 continue; 1111 1112 if (IN6_ARE_ADDR_EQUAL(&sin6->sin6_addr, &in6)) { 1113 found = 1; 1114 break; 1115 } 1116 } 1117 if_addr_runlock(ifp); 1118 1119 return (found); 1120} 1121 1122static struct l2t_entry * 1123get_l2te_for_nexthop(struct port_info *pi, struct ifnet *ifp, 1124 struct in_conninfo *inc) 1125{ 1126 struct rtentry *rt; 1127 struct l2t_entry *e; 1128 struct sockaddr_in6 sin6; 1129 struct sockaddr *dst = (void *)&sin6; 1130 1131 if (inc->inc_flags & INC_ISIPV6) { 1132 dst->sa_len = sizeof(struct sockaddr_in6); 1133 dst->sa_family = AF_INET6; 1134 ((struct sockaddr_in6 *)dst)->sin6_addr = inc->inc6_faddr; 1135 1136 if (IN6_IS_ADDR_LINKLOCAL(&inc->inc6_laddr)) { 1137 /* no need for route lookup */ 1138 e = t4_l2t_get(pi, ifp, dst); 1139 return (e); 1140 } 1141 } else { 1142 dst->sa_len = sizeof(struct sockaddr_in); 1143 dst->sa_family = AF_INET; 1144 ((struct sockaddr_in *)dst)->sin_addr = inc->inc_faddr; 1145 } 1146 1147 rt = rtalloc1(dst, 0, 0); 1148 if (rt == NULL) 1149 return (NULL); 1150 else { 1151 struct sockaddr *nexthop; 1152 1153 RT_UNLOCK(rt); 1154 if (rt->rt_ifp != ifp) 1155 e = NULL; 1156 else { 1157 if (rt->rt_flags & RTF_GATEWAY) 1158 nexthop = rt->rt_gateway; 1159 else 1160 nexthop = dst; 1161 e = t4_l2t_get(pi, ifp, nexthop); 1162 } 1163 RTFREE(rt); 1164 } 1165 1166 return (e); 1167} 1168 1169static int 1170ifnet_has_ip(struct ifnet *ifp, struct in_addr in) 1171{ 1172 struct ifaddr *ifa; 1173 struct sockaddr_in *sin; 1174 int found = 0; 1175 1176 if_addr_rlock(ifp); 1177 TAILQ_FOREACH(ifa, &ifp->if_addrhead, ifa_link) { 1178 sin = (void *)ifa->ifa_addr; 1179 if (sin->sin_family != AF_INET) 1180 continue; 1181 1182 if (sin->sin_addr.s_addr == in.s_addr) { 1183 found = 1; 1184 break; 1185 } 1186 } 1187 if_addr_runlock(ifp); 1188 1189 return (found); 1190} 1191 1192#define REJECT_PASS_ACCEPT() do { \ 1193 reject_reason = __LINE__; \ 1194 goto reject; \ 1195} while (0) 1196 1197/* 1198 * The context associated with a tid entry via insert_tid could be a synq_entry 1199 * or a toepcb. The only way CPL handlers can tell is via a bit in these flags. 1200 */ 1201CTASSERT(offsetof(struct toepcb, flags) == offsetof(struct synq_entry, flags)); 1202 1203/* 1204 * Incoming SYN on a listening socket. 1205 * 1206 * XXX: Every use of ifp in this routine has a bad race with up/down, toe/-toe, 1207 * etc. 1208 */ 1209static int 1210do_pass_accept_req(struct sge_iq *iq, const struct rss_header *rss, 1211 struct mbuf *m) 1212{ 1213 struct adapter *sc = iq->adapter; 1214 struct toedev *tod; 1215 const struct cpl_pass_accept_req *cpl = mtod(m, const void *); 1216 struct cpl_pass_accept_rpl *rpl; 1217 struct wrqe *wr; 1218 unsigned int stid = G_PASS_OPEN_TID(be32toh(cpl->tos_stid)); 1219 unsigned int tid = GET_TID(cpl); 1220 struct listen_ctx *lctx = lookup_stid(sc, stid); 1221 struct inpcb *inp; 1222 struct socket *so; 1223 struct in_conninfo inc; 1224 struct tcphdr th; 1225 struct tcpopt to; 1226 struct port_info *pi; 1227 struct ifnet *hw_ifp, *ifp; 1228 struct l2t_entry *e = NULL; 1229 int rscale, mtu_idx, rx_credits, rxqid, ulp_mode; 1230 struct synq_entry *synqe = NULL; 1231 int reject_reason; 1232 uint16_t vid; 1233#ifdef INVARIANTS 1234 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); 1235#endif 1236 1237 KASSERT(opcode == CPL_PASS_ACCEPT_REQ, 1238 ("%s: unexpected opcode 0x%x", __func__, opcode)); 1239 KASSERT(lctx->stid == stid, ("%s: lctx stid mismatch", __func__)); 1240 1241 CTR4(KTR_CXGBE, "%s: stid %u, tid %u, lctx %p", __func__, stid, tid, 1242 lctx); 1243 1244 pass_accept_req_to_protohdrs(m, &inc, &th); 1245 t4opt_to_tcpopt(&cpl->tcpopt, &to); 1246 1247 pi = sc->port[G_SYN_INTF(be16toh(cpl->l2info))]; 1248 hw_ifp = pi->ifp; /* the cxgbeX ifnet */ 1249 m->m_pkthdr.rcvif = hw_ifp; 1250 tod = TOEDEV(hw_ifp); 1251 1252 /* 1253 * Figure out if there is a pseudo interface (vlan, lagg, etc.) 1254 * involved. Don't offload if the SYN had a VLAN tag and the vid 1255 * doesn't match anything on this interface. 1256 * 1257 * XXX: lagg support, lagg + vlan support. 1258 */ 1259 vid = EVL_VLANOFTAG(be16toh(cpl->vlan)); 1260 if (vid != 0xfff) { 1261 ifp = VLAN_DEVAT(hw_ifp, vid); 1262 if (ifp == NULL) 1263 REJECT_PASS_ACCEPT(); 1264 } else 1265 ifp = hw_ifp; 1266 1267 /* 1268 * Don't offload if the peer requested a TCP option that's not known to 1269 * the silicon. 1270 */ 1271 if (cpl->tcpopt.unknown) 1272 REJECT_PASS_ACCEPT(); 1273 1274 if (inc.inc_flags & INC_ISIPV6) { 1275 1276 /* Don't offload if the ifcap isn't enabled */ 1277 if ((ifp->if_capenable & IFCAP_TOE6) == 0) 1278 REJECT_PASS_ACCEPT(); 1279 1280 /* 1281 * SYN must be directed to an IP6 address on this ifnet. This 1282 * is more restrictive than in6_localip. 1283 */ 1284 if (!ifnet_has_ip6(ifp, &inc.inc6_laddr)) 1285 REJECT_PASS_ACCEPT(); 1286 } else { 1287 1288 /* Don't offload if the ifcap isn't enabled */ 1289 if ((ifp->if_capenable & IFCAP_TOE4) == 0) 1290 REJECT_PASS_ACCEPT(); 1291 1292 /* 1293 * SYN must be directed to an IP address on this ifnet. This 1294 * is more restrictive than in_localip. 1295 */ 1296 if (!ifnet_has_ip(ifp, inc.inc_laddr)) 1297 REJECT_PASS_ACCEPT(); 1298 } 1299 1300 e = get_l2te_for_nexthop(pi, ifp, &inc); 1301 if (e == NULL) 1302 REJECT_PASS_ACCEPT(); 1303 1304 synqe = mbuf_to_synqe(m); 1305 if (synqe == NULL) 1306 REJECT_PASS_ACCEPT(); 1307 1308 wr = alloc_wrqe(is_t4(sc) ? sizeof(struct cpl_pass_accept_rpl) : 1309 sizeof(struct cpl_t5_pass_accept_rpl), &sc->sge.ctrlq[pi->port_id]); 1310 if (wr == NULL) 1311 REJECT_PASS_ACCEPT(); 1312 rpl = wrtod(wr); 1313 1314 INP_INFO_WLOCK(&V_tcbinfo); /* for 4-tuple check, syncache_add */ 1315 1316 /* Don't offload if the 4-tuple is already in use */ 1317 if (toe_4tuple_check(&inc, &th, ifp) != 0) { 1318 INP_INFO_WUNLOCK(&V_tcbinfo); 1319 free(wr, M_CXGBE); 1320 REJECT_PASS_ACCEPT(); 1321 } 1322 1323 inp = lctx->inp; /* listening socket, not owned by TOE */ 1324 INP_WLOCK(inp); 1325 1326 /* Don't offload if the listening socket has closed */ 1327 if (__predict_false(inp->inp_flags & INP_DROPPED)) { 1328 /* 1329 * The listening socket has closed. The reply from the TOE to 1330 * our CPL_CLOSE_LISTSRV_REQ will ultimately release all 1331 * resources tied to this listen context. 1332 */ 1333 INP_WUNLOCK(inp); 1334 INP_INFO_WUNLOCK(&V_tcbinfo); 1335 free(wr, M_CXGBE); 1336 REJECT_PASS_ACCEPT(); 1337 } 1338 so = inp->inp_socket; 1339 1340 mtu_idx = find_best_mtu_idx(sc, &inc, be16toh(cpl->tcpopt.mss)); 1341 rscale = cpl->tcpopt.wsf && V_tcp_do_rfc1323 ? select_rcv_wscale() : 0; 1342 SOCKBUF_LOCK(&so->so_rcv); 1343 /* opt0 rcv_bufsiz initially, assumes its normal meaning later */ 1344 rx_credits = min(select_rcv_wnd(so) >> 10, M_RCV_BUFSIZ); 1345 SOCKBUF_UNLOCK(&so->so_rcv); 1346 1347 save_qids_in_mbuf(m, pi); 1348 get_qids_from_mbuf(m, NULL, &rxqid); 1349 1350 if (is_t4(sc)) 1351 INIT_TP_WR_MIT_CPL(rpl, CPL_PASS_ACCEPT_RPL, tid); 1352 else { 1353 struct cpl_t5_pass_accept_rpl *rpl5 = (void *)rpl; 1354 1355 INIT_TP_WR_MIT_CPL(rpl5, CPL_PASS_ACCEPT_RPL, tid); 1356 } 1357 if (sc->tt.ddp && (so->so_options & SO_NO_DDP) == 0) { 1358 ulp_mode = ULP_MODE_TCPDDP; 1359 synqe->flags |= TPF_SYNQE_TCPDDP; 1360 } else 1361 ulp_mode = ULP_MODE_NONE; 1362 rpl->opt0 = calc_opt0(so, pi, e, mtu_idx, rscale, rx_credits, ulp_mode); 1363 rpl->opt2 = calc_opt2p(sc, pi, rxqid, &cpl->tcpopt, &th, ulp_mode); 1364 1365 synqe->tid = tid; 1366 synqe->lctx = lctx; 1367 synqe->syn = m; 1368 m = NULL; 1369 refcount_init(&synqe->refcnt, 1); /* 1 means extra hold */ 1370 synqe->l2e_idx = e->idx; 1371 synqe->rcv_bufsize = rx_credits; 1372 atomic_store_rel_ptr(&synqe->wr, (uintptr_t)wr); 1373 1374 insert_tid(sc, tid, synqe); 1375 TAILQ_INSERT_TAIL(&lctx->synq, synqe, link); 1376 hold_synqe(synqe); /* hold for the duration it's in the synq */ 1377 hold_lctx(lctx); /* A synqe on the list has a ref on its lctx */ 1378 1379 /* 1380 * If all goes well t4_syncache_respond will get called during 1381 * syncache_add. Also note that syncache_add releases both pcbinfo and 1382 * pcb locks. 1383 */ 1384 toe_syncache_add(&inc, &to, &th, inp, tod, synqe); 1385 INP_UNLOCK_ASSERT(inp); /* ok to assert, we have a ref on the inp */ 1386 INP_INFO_UNLOCK_ASSERT(&V_tcbinfo); 1387 1388 /* 1389 * If we replied during syncache_add (synqe->wr has been consumed), 1390 * good. Otherwise, set it to 0 so that further syncache_respond 1391 * attempts by the kernel will be ignored. 1392 */ 1393 if (atomic_cmpset_ptr(&synqe->wr, (uintptr_t)wr, 0)) { 1394 1395 /* 1396 * syncache may or may not have a hold on the synqe, which may 1397 * or may not be stashed in the original SYN mbuf passed to us. 1398 * Just copy it over instead of dealing with all possibilities. 1399 */ 1400 m = m_dup(synqe->syn, M_NOWAIT); 1401 if (m) 1402 m->m_pkthdr.rcvif = hw_ifp; 1403 1404 remove_tid(sc, synqe->tid); 1405 free(wr, M_CXGBE); 1406 1407 /* Yank the synqe out of the lctx synq. */ 1408 INP_WLOCK(inp); 1409 TAILQ_REMOVE(&lctx->synq, synqe, link); 1410 release_synqe(synqe); /* removed from synq list */ 1411 inp = release_lctx(sc, lctx); 1412 if (inp) 1413 INP_WUNLOCK(inp); 1414 1415 release_synqe(synqe); /* extra hold */ 1416 REJECT_PASS_ACCEPT(); 1417 } 1418 1419 CTR5(KTR_CXGBE, "%s: stid %u, tid %u, lctx %p, synqe %p, SYNACK", 1420 __func__, stid, tid, lctx, synqe); 1421 1422 INP_WLOCK(inp); 1423 synqe->flags |= TPF_SYNQE_HAS_L2TE; 1424 if (__predict_false(inp->inp_flags & INP_DROPPED)) { 1425 /* 1426 * Listening socket closed but tod_listen_stop did not abort 1427 * this tid because there was no L2T entry for the tid at that 1428 * time. Abort it now. The reply to the abort will clean up. 1429 */ 1430 CTR6(KTR_CXGBE, 1431 "%s: stid %u, tid %u, lctx %p, synqe %p (0x%x), ABORT", 1432 __func__, stid, tid, lctx, synqe, synqe->flags); 1433 if (!(synqe->flags & TPF_SYNQE_EXPANDED)) 1434 send_reset_synqe(tod, synqe); 1435 INP_WUNLOCK(inp); 1436 1437 release_synqe(synqe); /* extra hold */ 1438 return (__LINE__); 1439 } 1440 INP_WUNLOCK(inp); 1441 1442 release_synqe(synqe); /* extra hold */ 1443 return (0); 1444reject: 1445 CTR4(KTR_CXGBE, "%s: stid %u, tid %u, REJECT (%d)", __func__, stid, tid, 1446 reject_reason); 1447 1448 if (e) 1449 t4_l2t_release(e); 1450 release_tid(sc, tid, lctx->ctrlq); 1451 1452 if (__predict_true(m != NULL)) { 1453 m_adj(m, sizeof(*cpl)); 1454 m->m_pkthdr.csum_flags |= (CSUM_IP_CHECKED | CSUM_IP_VALID | 1455 CSUM_DATA_VALID | CSUM_PSEUDO_HDR); 1456 m->m_pkthdr.csum_data = 0xffff; 1457 hw_ifp->if_input(hw_ifp, m); 1458 } 1459 1460 return (reject_reason); 1461} 1462 1463static void 1464synqe_to_protohdrs(struct synq_entry *synqe, 1465 const struct cpl_pass_establish *cpl, struct in_conninfo *inc, 1466 struct tcphdr *th, struct tcpopt *to) 1467{ 1468 uint16_t tcp_opt = be16toh(cpl->tcp_opt); 1469 1470 /* start off with the original SYN */ 1471 pass_accept_req_to_protohdrs(synqe->syn, inc, th); 1472 1473 /* modify parts to make it look like the ACK to our SYN|ACK */ 1474 th->th_flags = TH_ACK; 1475 th->th_ack = synqe->iss + 1; 1476 th->th_seq = be32toh(cpl->rcv_isn); 1477 bzero(to, sizeof(*to)); 1478 if (G_TCPOPT_TSTAMP(tcp_opt)) { 1479 to->to_flags |= TOF_TS; 1480 to->to_tsecr = synqe->ts; 1481 } 1482} 1483 1484static int 1485do_pass_establish(struct sge_iq *iq, const struct rss_header *rss, 1486 struct mbuf *m) 1487{ 1488 struct adapter *sc = iq->adapter; 1489 struct port_info *pi; 1490 struct ifnet *ifp; 1491 const struct cpl_pass_establish *cpl = (const void *)(rss + 1); 1492#if defined(KTR) || defined(INVARIANTS) 1493 unsigned int stid = G_PASS_OPEN_TID(be32toh(cpl->tos_stid)); 1494#endif 1495 unsigned int tid = GET_TID(cpl); 1496 struct synq_entry *synqe = lookup_tid(sc, tid); 1497 struct listen_ctx *lctx = synqe->lctx; 1498 struct inpcb *inp = lctx->inp; 1499 struct socket *so; 1500 struct tcphdr th; 1501 struct tcpopt to; 1502 struct in_conninfo inc; 1503 struct toepcb *toep; 1504 u_int txqid, rxqid; 1505#ifdef INVARIANTS 1506 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); 1507#endif 1508 1509 KASSERT(opcode == CPL_PASS_ESTABLISH, 1510 ("%s: unexpected opcode 0x%x", __func__, opcode)); 1511 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 1512 KASSERT(lctx->stid == stid, ("%s: lctx stid mismatch", __func__)); 1513 KASSERT(synqe->flags & TPF_SYNQE, 1514 ("%s: tid %u (ctx %p) not a synqe", __func__, tid, synqe)); 1515 1516 INP_INFO_WLOCK(&V_tcbinfo); /* for syncache_expand */ 1517 INP_WLOCK(inp); 1518 1519 CTR6(KTR_CXGBE, 1520 "%s: stid %u, tid %u, synqe %p (0x%x), inp_flags 0x%x", 1521 __func__, stid, tid, synqe, synqe->flags, inp->inp_flags); 1522 1523 if (__predict_false(inp->inp_flags & INP_DROPPED)) { 1524 1525 if (synqe->flags & TPF_SYNQE_HAS_L2TE) { 1526 KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN, 1527 ("%s: listen socket closed but tid %u not aborted.", 1528 __func__, tid)); 1529 } 1530 1531 INP_WUNLOCK(inp); 1532 INP_INFO_WUNLOCK(&V_tcbinfo); 1533 return (0); 1534 } 1535 1536 ifp = synqe->syn->m_pkthdr.rcvif; 1537 pi = ifp->if_softc; 1538 KASSERT(pi->adapter == sc, 1539 ("%s: pi %p, sc %p mismatch", __func__, pi, sc)); 1540 1541 get_qids_from_mbuf(synqe->syn, &txqid, &rxqid); 1542 KASSERT(rxqid == iq_to_ofld_rxq(iq) - &sc->sge.ofld_rxq[0], 1543 ("%s: CPL arrived on unexpected rxq. %d %d", __func__, rxqid, 1544 (int)(iq_to_ofld_rxq(iq) - &sc->sge.ofld_rxq[0]))); 1545 1546 toep = alloc_toepcb(pi, txqid, rxqid, M_NOWAIT); 1547 if (toep == NULL) { 1548reset: 1549 /* 1550 * The reply to this abort will perform final cleanup. There is 1551 * no need to check for HAS_L2TE here. We can be here only if 1552 * we responded to the PASS_ACCEPT_REQ, and our response had the 1553 * L2T idx. 1554 */ 1555 send_reset_synqe(TOEDEV(ifp), synqe); 1556 INP_WUNLOCK(inp); 1557 INP_INFO_WUNLOCK(&V_tcbinfo); 1558 return (0); 1559 } 1560 toep->tid = tid; 1561 toep->l2te = &sc->l2t->l2tab[synqe->l2e_idx]; 1562 if (synqe->flags & TPF_SYNQE_TCPDDP) 1563 set_tcpddp_ulp_mode(toep); 1564 else 1565 toep->ulp_mode = ULP_MODE_NONE; 1566 /* opt0 rcv_bufsiz initially, assumes its normal meaning later */ 1567 toep->rx_credits = synqe->rcv_bufsize; 1568 1569 so = inp->inp_socket; 1570 KASSERT(so != NULL, ("%s: socket is NULL", __func__)); 1571 1572 /* Come up with something that syncache_expand should be ok with. */ 1573 synqe_to_protohdrs(synqe, cpl, &inc, &th, &to); 1574 1575 /* 1576 * No more need for anything in the mbuf that carried the 1577 * CPL_PASS_ACCEPT_REQ. Drop the CPL_PASS_ESTABLISH and toep pointer 1578 * there. XXX: bad form but I don't want to increase the size of synqe. 1579 */ 1580 m = synqe->syn; 1581 KASSERT(sizeof(*cpl) + sizeof(toep) <= m->m_len, 1582 ("%s: no room in mbuf %p (m_len %d)", __func__, m, m->m_len)); 1583 bcopy(cpl, mtod(m, void *), sizeof(*cpl)); 1584 *(struct toepcb **)(mtod(m, struct cpl_pass_establish *) + 1) = toep; 1585 1586 if (!toe_syncache_expand(&inc, &to, &th, &so) || so == NULL) { 1587 free_toepcb(toep); 1588 goto reset; 1589 } 1590 1591 /* 1592 * This is for the unlikely case where the syncache entry that we added 1593 * has been evicted from the syncache, but the syncache_expand above 1594 * works because of syncookies. 1595 * 1596 * XXX: we've held the tcbinfo lock throughout so there's no risk of 1597 * anyone accept'ing a connection before we've installed our hooks, but 1598 * this somewhat defeats the purpose of having a tod_offload_socket :-( 1599 */ 1600 if (__predict_false(!(synqe->flags & TPF_SYNQE_EXPANDED))) { 1601 struct inpcb *new_inp = sotoinpcb(so); 1602 1603 INP_WLOCK(new_inp); 1604 tcp_timer_activate(intotcpcb(new_inp), TT_KEEP, 0); 1605 t4_offload_socket(TOEDEV(ifp), synqe, so); 1606 INP_WUNLOCK(new_inp); 1607 } 1608 1609 /* Done with the synqe */ 1610 TAILQ_REMOVE(&lctx->synq, synqe, link); 1611 inp = release_lctx(sc, lctx); 1612 if (inp != NULL) 1613 INP_WUNLOCK(inp); 1614 INP_INFO_WUNLOCK(&V_tcbinfo); 1615 release_synqe(synqe); 1616 1617 return (0); 1618} 1619 1620void 1621t4_init_listen_cpl_handlers(struct adapter *sc) 1622{ 1623 1624 t4_register_cpl_handler(sc, CPL_PASS_OPEN_RPL, do_pass_open_rpl); 1625 t4_register_cpl_handler(sc, CPL_CLOSE_LISTSRV_RPL, do_close_server_rpl); 1626 t4_register_cpl_handler(sc, CPL_PASS_ACCEPT_REQ, do_pass_accept_req); 1627 t4_register_cpl_handler(sc, CPL_PASS_ESTABLISH, do_pass_establish); 1628} 1629#endif 1630