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_fib.h> 53 #include <netinet/in_pcb.h> 54 #include <netinet/ip.h> 55 #include <netinet/ip6.h> 56 #include <netinet6/in6_fib.h> 57 #include <netinet6/scope6_var.h> 58 #include <netinet/tcp_timer.h> 59 #include <netinet/tcp_var.h> 60 #define TCPSTATES 61 #include <netinet/tcp_fsm.h> 62 #include <netinet/toecore.h> 63 64 #include "common/common.h" 65 #include "common/t4_msg.h" 66 #include "common/t4_regs.h" 67 #include "tom/t4_tom_l2t.h" 68 #include "tom/t4_tom.h" 69 70 /* stid services */ 71 static int alloc_stid(struct adapter *, struct listen_ctx *, int); 72 static struct listen_ctx *lookup_stid(struct adapter *, int); 73 static void free_stid(struct adapter *, struct listen_ctx *); 74 75 /* lctx services */ 76 static struct listen_ctx *alloc_lctx(struct adapter *, struct inpcb *, 77 struct vi_info *); 78 static int free_lctx(struct adapter *, struct listen_ctx *); 79 static void hold_lctx(struct listen_ctx *); 80 static void listen_hash_add(struct adapter *, struct listen_ctx *); 81 static struct listen_ctx *listen_hash_find(struct adapter *, struct inpcb *); 82 static struct listen_ctx *listen_hash_del(struct adapter *, struct inpcb *); 83 static struct inpcb *release_lctx(struct adapter *, struct listen_ctx *); 84 85 static inline void save_qids_in_mbuf(struct mbuf *, struct vi_info *); 86 static inline void get_qids_from_mbuf(struct mbuf *m, int *, int *); 87 static void send_reset_synqe(struct toedev *, struct synq_entry *); 88 89 static int 90 alloc_stid(struct adapter *sc, struct listen_ctx *lctx, int isipv6) 91 { 92 struct tid_info *t = &sc->tids; 93 u_int stid, n, f, mask; 94 struct stid_region *sr = &lctx->stid_region; 95 96 /* 97 * An IPv6 server needs 2 naturally aligned stids (1 stid = 4 cells) in 98 * the TCAM. The start of the stid region is properly aligned (the chip 99 * requires each region to be 128-cell aligned). 100 */ 101 n = isipv6 ? 2 : 1; 102 mask = n - 1; 103 KASSERT((t->stid_base & mask) == 0 && (t->nstids & mask) == 0, 104 ("%s: stid region (%u, %u) not properly aligned. n = %u", 105 __func__, t->stid_base, t->nstids, n)); 106 107 mtx_lock(&t->stid_lock); 108 if (n > t->nstids - t->stids_in_use) { 109 mtx_unlock(&t->stid_lock); 110 return (-1); 111 } 112 113 if (t->nstids_free_head >= n) { 114 /* 115 * This allocation will definitely succeed because the region 116 * starts at a good alignment and we just checked we have enough 117 * stids free. 118 */ 119 f = t->nstids_free_head & mask; 120 t->nstids_free_head -= n + f; 121 stid = t->nstids_free_head; 122 TAILQ_INSERT_HEAD(&t->stids, sr, link); 123 } else { 124 struct stid_region *s; 125 126 stid = t->nstids_free_head; 127 TAILQ_FOREACH(s, &t->stids, link) { 128 stid += s->used + s->free; 129 f = stid & mask; 130 if (s->free >= n + f) { 131 stid -= n + f; 132 s->free -= n + f; 133 TAILQ_INSERT_AFTER(&t->stids, s, sr, link); 134 goto allocated; 135 } 136 } 137 138 if (__predict_false(stid != t->nstids)) { 139 panic("%s: stids TAILQ (%p) corrupt." 140 " At %d instead of %d at the end of the queue.", 141 __func__, &t->stids, stid, t->nstids); 142 } 143 144 mtx_unlock(&t->stid_lock); 145 return (-1); 146 } 147 148 allocated: 149 sr->used = n; 150 sr->free = f; 151 t->stids_in_use += n; 152 t->stid_tab[stid] = lctx; 153 mtx_unlock(&t->stid_lock); 154 155 KASSERT(((stid + t->stid_base) & mask) == 0, 156 ("%s: EDOOFUS.", __func__)); 157 return (stid + t->stid_base); 158 } 159 160 static struct listen_ctx * 161 lookup_stid(struct adapter *sc, int stid) 162 { 163 struct tid_info *t = &sc->tids; 164 165 return (t->stid_tab[stid - t->stid_base]); 166 } 167 168 static void 169 free_stid(struct adapter *sc, struct listen_ctx *lctx) 170 { 171 struct tid_info *t = &sc->tids; 172 struct stid_region *sr = &lctx->stid_region; 173 struct stid_region *s; 174 175 KASSERT(sr->used > 0, ("%s: nonsense free (%d)", __func__, sr->used)); 176 177 mtx_lock(&t->stid_lock); 178 s = TAILQ_PREV(sr, stid_head, link); 179 if (s != NULL) 180 s->free += sr->used + sr->free; 181 else 182 t->nstids_free_head += sr->used + sr->free; 183 KASSERT(t->stids_in_use >= sr->used, 184 ("%s: stids_in_use (%u) < stids being freed (%u)", __func__, 185 t->stids_in_use, sr->used)); 186 t->stids_in_use -= sr->used; 187 TAILQ_REMOVE(&t->stids, sr, link); 188 mtx_unlock(&t->stid_lock); 189 } 190 191 static struct listen_ctx * 192 alloc_lctx(struct adapter *sc, struct inpcb *inp, struct vi_info *vi) 193 { 194 struct listen_ctx *lctx; 195 196 INP_WLOCK_ASSERT(inp); 197 198 lctx = malloc(sizeof(struct listen_ctx), M_CXGBE, M_NOWAIT | M_ZERO); 199 if (lctx == NULL) 200 return (NULL); 201 202 lctx->stid = alloc_stid(sc, lctx, inp->inp_vflag & INP_IPV6); 203 if (lctx->stid < 0) { 204 free(lctx, M_CXGBE); 205 return (NULL); 206 } 207 208 if (inp->inp_vflag & INP_IPV6 && 209 !IN6_ARE_ADDR_EQUAL(&in6addr_any, &inp->in6p_laddr)) { 210 struct tom_data *td = sc->tom_softc; 211 212 lctx->ce = hold_lip(td, &inp->in6p_laddr); 213 if (lctx->ce == NULL) { 214 free(lctx, M_CXGBE); 215 return (NULL); 216 } 217 } 218 219 lctx->ctrlq = &sc->sge.ctrlq[vi->pi->port_id]; 220 lctx->ofld_rxq = &sc->sge.ofld_rxq[vi->first_ofld_rxq]; 221 refcount_init(&lctx->refcount, 1); 222 TAILQ_INIT(&lctx->synq); 223 224 lctx->inp = inp; 225 in_pcbref(inp); 226 227 return (lctx); 228 } 229 230 /* Don't call this directly, use release_lctx instead */ 231 static int 232 free_lctx(struct adapter *sc, struct listen_ctx *lctx) 233 { 234 struct inpcb *inp = lctx->inp; 235 struct tom_data *td = sc->tom_softc; 236 237 INP_WLOCK_ASSERT(inp); 238 KASSERT(lctx->refcount == 0, 239 ("%s: refcount %d", __func__, lctx->refcount)); 240 KASSERT(TAILQ_EMPTY(&lctx->synq), 241 ("%s: synq not empty.", __func__)); 242 KASSERT(lctx->stid >= 0, ("%s: bad stid %d.", __func__, lctx->stid)); 243 244 CTR4(KTR_CXGBE, "%s: stid %u, lctx %p, inp %p", 245 __func__, lctx->stid, lctx, lctx->inp); 246 247 if (lctx->ce) 248 release_lip(td, lctx->ce); 249 free_stid(sc, lctx); 250 free(lctx, M_CXGBE); 251 252 return (in_pcbrele_wlocked(inp)); 253 } 254 255 static void 256 hold_lctx(struct listen_ctx *lctx) 257 { 258 259 refcount_acquire(&lctx->refcount); 260 } 261 262 static inline uint32_t 263 listen_hashfn(void *key, u_long mask) 264 { 265 266 return (fnv_32_buf(&key, sizeof(key), FNV1_32_INIT) & mask); 267 } 268 269 /* 270 * Add a listen_ctx entry to the listen hash table. 271 */ 272 static void 273 listen_hash_add(struct adapter *sc, struct listen_ctx *lctx) 274 { 275 struct tom_data *td = sc->tom_softc; 276 int bucket = listen_hashfn(lctx->inp, td->listen_mask); 277 278 mtx_lock(&td->lctx_hash_lock); 279 LIST_INSERT_HEAD(&td->listen_hash[bucket], lctx, link); 280 td->lctx_count++; 281 mtx_unlock(&td->lctx_hash_lock); 282 } 283 284 /* 285 * Look for the listening socket's context entry in the hash and return it. 286 */ 287 static struct listen_ctx * 288 listen_hash_find(struct adapter *sc, struct inpcb *inp) 289 { 290 struct tom_data *td = sc->tom_softc; 291 int bucket = listen_hashfn(inp, td->listen_mask); 292 struct listen_ctx *lctx; 293 294 mtx_lock(&td->lctx_hash_lock); 295 LIST_FOREACH(lctx, &td->listen_hash[bucket], link) { 296 if (lctx->inp == inp) 297 break; 298 } 299 mtx_unlock(&td->lctx_hash_lock); 300 301 return (lctx); 302 } 303 304 /* 305 * Removes the listen_ctx structure for inp from the hash and returns it. 306 */ 307 static struct listen_ctx * 308 listen_hash_del(struct adapter *sc, struct inpcb *inp) 309 { 310 struct tom_data *td = sc->tom_softc; 311 int bucket = listen_hashfn(inp, td->listen_mask); 312 struct listen_ctx *lctx, *l; 313 314 mtx_lock(&td->lctx_hash_lock); 315 LIST_FOREACH_SAFE(lctx, &td->listen_hash[bucket], link, l) { 316 if (lctx->inp == inp) { 317 LIST_REMOVE(lctx, link); 318 td->lctx_count--; 319 break; 320 } 321 } 322 mtx_unlock(&td->lctx_hash_lock); 323 324 return (lctx); 325 } 326 327 /* 328 * Releases a hold on the lctx. Must be called with the listening socket's inp 329 * locked. The inp may be freed by this function and it returns NULL to 330 * indicate this. 331 */ 332 static struct inpcb * 333 release_lctx(struct adapter *sc, struct listen_ctx *lctx) 334 { 335 struct inpcb *inp = lctx->inp; 336 int inp_freed = 0; 337 338 INP_WLOCK_ASSERT(inp); 339 if (refcount_release(&lctx->refcount)) 340 inp_freed = free_lctx(sc, lctx); 341 342 return (inp_freed ? NULL : inp); 343 } 344 345 static void 346 send_reset_synqe(struct toedev *tod, struct synq_entry *synqe) 347 { 348 struct adapter *sc = tod->tod_softc; 349 struct mbuf *m = synqe->syn; 350 struct ifnet *ifp = m->m_pkthdr.rcvif; 351 struct vi_info *vi = ifp->if_softc; 352 struct port_info *pi = vi->pi; 353 struct l2t_entry *e = &sc->l2t->l2tab[synqe->l2e_idx]; 354 struct wrqe *wr; 355 struct fw_flowc_wr *flowc; 356 struct cpl_abort_req *req; 357 int txqid, rxqid, flowclen; 358 struct sge_wrq *ofld_txq; 359 struct sge_ofld_rxq *ofld_rxq; 360 const int nparams = 6; 361 unsigned int pfvf = G_FW_VIID_PFN(vi->viid) << S_FW_VIID_PFN; 362 363 INP_WLOCK_ASSERT(synqe->lctx->inp); 364 365 CTR5(KTR_CXGBE, "%s: synqe %p (0x%x), tid %d%s", 366 __func__, synqe, synqe->flags, synqe->tid, 367 synqe->flags & TPF_ABORT_SHUTDOWN ? 368 " (abort already in progress)" : ""); 369 if (synqe->flags & TPF_ABORT_SHUTDOWN) 370 return; /* abort already in progress */ 371 synqe->flags |= TPF_ABORT_SHUTDOWN; 372 373 get_qids_from_mbuf(m, &txqid, &rxqid); 374 ofld_txq = &sc->sge.ofld_txq[txqid]; 375 ofld_rxq = &sc->sge.ofld_rxq[rxqid]; 376 377 /* The wrqe will have two WRs - a flowc followed by an abort_req */ 378 flowclen = sizeof(*flowc) + nparams * sizeof(struct fw_flowc_mnemval); 379 380 wr = alloc_wrqe(roundup2(flowclen, EQ_ESIZE) + sizeof(*req), ofld_txq); 381 if (wr == NULL) { 382 /* XXX */ 383 panic("%s: allocation failure.", __func__); 384 } 385 flowc = wrtod(wr); 386 req = (void *)((caddr_t)flowc + roundup2(flowclen, EQ_ESIZE)); 387 388 /* First the flowc ... */ 389 memset(flowc, 0, wr->wr_len); 390 flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) | 391 V_FW_FLOWC_WR_NPARAMS(nparams)); 392 flowc->flowid_len16 = htonl(V_FW_WR_LEN16(howmany(flowclen, 16)) | 393 V_FW_WR_FLOWID(synqe->tid)); 394 flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_PFNVFN; 395 flowc->mnemval[0].val = htobe32(pfvf); 396 flowc->mnemval[1].mnemonic = FW_FLOWC_MNEM_CH; 397 flowc->mnemval[1].val = htobe32(pi->tx_chan); 398 flowc->mnemval[2].mnemonic = FW_FLOWC_MNEM_PORT; 399 flowc->mnemval[2].val = htobe32(pi->tx_chan); 400 flowc->mnemval[3].mnemonic = FW_FLOWC_MNEM_IQID; 401 flowc->mnemval[3].val = htobe32(ofld_rxq->iq.abs_id); 402 flowc->mnemval[4].mnemonic = FW_FLOWC_MNEM_SNDBUF; 403 flowc->mnemval[4].val = htobe32(512); 404 flowc->mnemval[5].mnemonic = FW_FLOWC_MNEM_MSS; 405 flowc->mnemval[5].val = htobe32(512); 406 synqe->flags |= TPF_FLOWC_WR_SENT; 407 408 /* ... then ABORT request */ 409 INIT_TP_WR_MIT_CPL(req, CPL_ABORT_REQ, synqe->tid); 410 req->rsvd0 = 0; /* don't have a snd_nxt */ 411 req->rsvd1 = 1; /* no data sent yet */ 412 req->cmd = CPL_ABORT_SEND_RST; 413 414 t4_l2t_send(sc, wr, e); 415 } 416 417 static int 418 create_server(struct adapter *sc, struct listen_ctx *lctx) 419 { 420 struct wrqe *wr; 421 struct cpl_pass_open_req *req; 422 struct inpcb *inp = lctx->inp; 423 424 wr = alloc_wrqe(sizeof(*req), lctx->ctrlq); 425 if (wr == NULL) { 426 log(LOG_ERR, "%s: allocation failure", __func__); 427 return (ENOMEM); 428 } 429 req = wrtod(wr); 430 431 INIT_TP_WR(req, 0); 432 OPCODE_TID(req) = htobe32(MK_OPCODE_TID(CPL_PASS_OPEN_REQ, lctx->stid)); 433 req->local_port = inp->inp_lport; 434 req->peer_port = 0; 435 req->local_ip = inp->inp_laddr.s_addr; 436 req->peer_ip = 0; 437 req->opt0 = htobe64(V_TX_CHAN(lctx->ctrlq->eq.tx_chan)); 438 req->opt1 = htobe64(V_CONN_POLICY(CPL_CONN_POLICY_ASK) | 439 F_SYN_RSS_ENABLE | V_SYN_RSS_QUEUE(lctx->ofld_rxq->iq.abs_id)); 440 441 t4_wrq_tx(sc, wr); 442 return (0); 443 } 444 445 static int 446 create_server6(struct adapter *sc, struct listen_ctx *lctx) 447 { 448 struct wrqe *wr; 449 struct cpl_pass_open_req6 *req; 450 struct inpcb *inp = lctx->inp; 451 452 wr = alloc_wrqe(sizeof(*req), lctx->ctrlq); 453 if (wr == NULL) { 454 log(LOG_ERR, "%s: allocation failure", __func__); 455 return (ENOMEM); 456 } 457 req = wrtod(wr); 458 459 INIT_TP_WR(req, 0); 460 OPCODE_TID(req) = htobe32(MK_OPCODE_TID(CPL_PASS_OPEN_REQ6, lctx->stid)); 461 req->local_port = inp->inp_lport; 462 req->peer_port = 0; 463 req->local_ip_hi = *(uint64_t *)&inp->in6p_laddr.s6_addr[0]; 464 req->local_ip_lo = *(uint64_t *)&inp->in6p_laddr.s6_addr[8]; 465 req->peer_ip_hi = 0; 466 req->peer_ip_lo = 0; 467 req->opt0 = htobe64(V_TX_CHAN(lctx->ctrlq->eq.tx_chan)); 468 req->opt1 = htobe64(V_CONN_POLICY(CPL_CONN_POLICY_ASK) | 469 F_SYN_RSS_ENABLE | V_SYN_RSS_QUEUE(lctx->ofld_rxq->iq.abs_id)); 470 471 t4_wrq_tx(sc, wr); 472 return (0); 473 } 474 475 static int 476 destroy_server(struct adapter *sc, struct listen_ctx *lctx) 477 { 478 struct wrqe *wr; 479 struct cpl_close_listsvr_req *req; 480 481 wr = alloc_wrqe(sizeof(*req), lctx->ctrlq); 482 if (wr == NULL) { 483 /* XXX */ 484 panic("%s: allocation failure.", __func__); 485 } 486 req = wrtod(wr); 487 488 INIT_TP_WR(req, 0); 489 OPCODE_TID(req) = htonl(MK_OPCODE_TID(CPL_CLOSE_LISTSRV_REQ, 490 lctx->stid)); 491 req->reply_ctrl = htobe16(lctx->ofld_rxq->iq.abs_id); 492 req->rsvd = htobe16(0); 493 494 t4_wrq_tx(sc, wr); 495 return (0); 496 } 497 498 /* 499 * Start a listening server by sending a passive open request to HW. 500 * 501 * Can't take adapter lock here and access to sc->flags, 502 * sc->offload_map, if_capenable are all race prone. 503 */ 504 int 505 t4_listen_start(struct toedev *tod, struct tcpcb *tp) 506 { 507 struct adapter *sc = tod->tod_softc; 508 struct vi_info *vi; 509 struct port_info *pi; 510 struct inpcb *inp = tp->t_inpcb; 511 struct listen_ctx *lctx; 512 int i, rc, v; 513 514 INP_WLOCK_ASSERT(inp); 515 516 /* Don't start a hardware listener for any loopback address. */ 517 if (inp->inp_vflag & INP_IPV6 && IN6_IS_ADDR_LOOPBACK(&inp->in6p_laddr)) 518 return (0); 519 if (!(inp->inp_vflag & INP_IPV6) && 520 IN_LOOPBACK(ntohl(inp->inp_laddr.s_addr))) 521 return (0); 522 #if 0 523 ADAPTER_LOCK(sc); 524 if (IS_BUSY(sc)) { 525 log(LOG_ERR, "%s: listen request ignored, %s is busy", 526 __func__, device_get_nameunit(sc->dev)); 527 goto done; 528 } 529 530 KASSERT(uld_active(sc, ULD_TOM), 531 ("%s: TOM not initialized", __func__)); 532 #endif 533 534 /* 535 * Find a running VI with IFCAP_TOE (4 or 6). We'll use the first 536 * such VI's queues to send the passive open and receive the reply to 537 * it. 538 * 539 * XXX: need a way to mark a port in use by offload. if_cxgbe should 540 * then reject any attempt to bring down such a port (and maybe reject 541 * attempts to disable IFCAP_TOE on that port too?). 542 */ 543 for_each_port(sc, i) { 544 pi = sc->port[i]; 545 for_each_vi(pi, v, vi) { 546 if (vi->ifp->if_drv_flags & IFF_DRV_RUNNING && 547 vi->ifp->if_capenable & IFCAP_TOE) 548 goto found; 549 } 550 } 551 goto done; /* no port that's UP with IFCAP_TOE enabled */ 552 found: 553 554 if (listen_hash_find(sc, inp) != NULL) 555 goto done; /* already setup */ 556 557 lctx = alloc_lctx(sc, inp, vi); 558 if (lctx == NULL) { 559 log(LOG_ERR, 560 "%s: listen request ignored, %s couldn't allocate lctx\n", 561 __func__, device_get_nameunit(sc->dev)); 562 goto done; 563 } 564 listen_hash_add(sc, lctx); 565 566 CTR6(KTR_CXGBE, "%s: stid %u (%s), lctx %p, inp %p vflag 0x%x", 567 __func__, lctx->stid, tcpstates[tp->t_state], lctx, inp, 568 inp->inp_vflag); 569 570 if (inp->inp_vflag & INP_IPV6) 571 rc = create_server6(sc, lctx); 572 else 573 rc = create_server(sc, lctx); 574 if (rc != 0) { 575 log(LOG_ERR, "%s: %s failed to create hw listener: %d.\n", 576 __func__, device_get_nameunit(sc->dev), rc); 577 (void) listen_hash_del(sc, inp); 578 inp = release_lctx(sc, lctx); 579 /* can't be freed, host stack has a reference */ 580 KASSERT(inp != NULL, ("%s: inp freed", __func__)); 581 goto done; 582 } 583 lctx->flags |= LCTX_RPL_PENDING; 584 done: 585 #if 0 586 ADAPTER_UNLOCK(sc); 587 #endif 588 return (0); 589 } 590 591 int 592 t4_listen_stop(struct toedev *tod, struct tcpcb *tp) 593 { 594 struct listen_ctx *lctx; 595 struct adapter *sc = tod->tod_softc; 596 struct inpcb *inp = tp->t_inpcb; 597 struct synq_entry *synqe; 598 599 INP_WLOCK_ASSERT(inp); 600 601 lctx = listen_hash_del(sc, inp); 602 if (lctx == NULL) 603 return (ENOENT); /* no hardware listener for this inp */ 604 605 CTR4(KTR_CXGBE, "%s: stid %u, lctx %p, flags %x", __func__, lctx->stid, 606 lctx, lctx->flags); 607 608 /* 609 * If the reply to the PASS_OPEN is still pending we'll wait for it to 610 * arrive and clean up when it does. 611 */ 612 if (lctx->flags & LCTX_RPL_PENDING) { 613 KASSERT(TAILQ_EMPTY(&lctx->synq), 614 ("%s: synq not empty.", __func__)); 615 return (EINPROGRESS); 616 } 617 618 /* 619 * The host stack will abort all the connections on the listening 620 * socket's so_comp. It doesn't know about the connections on the synq 621 * so we need to take care of those. 622 */ 623 TAILQ_FOREACH(synqe, &lctx->synq, link) { 624 if (synqe->flags & TPF_SYNQE_HAS_L2TE) 625 send_reset_synqe(tod, synqe); 626 } 627 628 destroy_server(sc, lctx); 629 return (0); 630 } 631 632 static inline void 633 hold_synqe(struct synq_entry *synqe) 634 { 635 636 refcount_acquire(&synqe->refcnt); 637 } 638 639 static inline void 640 release_synqe(struct synq_entry *synqe) 641 { 642 643 if (refcount_release(&synqe->refcnt)) { 644 int needfree = synqe->flags & TPF_SYNQE_NEEDFREE; 645 646 m_freem(synqe->syn); 647 if (needfree) 648 free(synqe, M_CXGBE); 649 } 650 } 651 652 void 653 t4_syncache_added(struct toedev *tod __unused, void *arg) 654 { 655 struct synq_entry *synqe = arg; 656 657 hold_synqe(synqe); 658 } 659 660 void 661 t4_syncache_removed(struct toedev *tod __unused, void *arg) 662 { 663 struct synq_entry *synqe = arg; 664 665 release_synqe(synqe); 666 } 667 668 /* XXX */ 669 extern void tcp_dooptions(struct tcpopt *, u_char *, int, int); 670 671 int 672 t4_syncache_respond(struct toedev *tod, void *arg, struct mbuf *m) 673 { 674 struct adapter *sc = tod->tod_softc; 675 struct synq_entry *synqe = arg; 676 struct wrqe *wr; 677 struct l2t_entry *e; 678 struct tcpopt to; 679 struct ip *ip = mtod(m, struct ip *); 680 struct tcphdr *th; 681 682 wr = (struct wrqe *)atomic_readandclear_ptr(&synqe->wr); 683 if (wr == NULL) { 684 m_freem(m); 685 return (EALREADY); 686 } 687 688 if (ip->ip_v == IPVERSION) 689 th = (void *)(ip + 1); 690 else 691 th = (void *)((struct ip6_hdr *)ip + 1); 692 bzero(&to, sizeof(to)); 693 tcp_dooptions(&to, (void *)(th + 1), (th->th_off << 2) - sizeof(*th), 694 TO_SYN); 695 696 /* save these for later */ 697 synqe->iss = be32toh(th->th_seq); 698 synqe->ts = to.to_tsval; 699 700 if (is_t5(sc)) { 701 struct cpl_t5_pass_accept_rpl *rpl5 = wrtod(wr); 702 703 rpl5->iss = th->th_seq; 704 } 705 706 e = &sc->l2t->l2tab[synqe->l2e_idx]; 707 t4_l2t_send(sc, wr, e); 708 709 m_freem(m); /* don't need this any more */ 710 return (0); 711 } 712 713 static int 714 do_pass_open_rpl(struct sge_iq *iq, const struct rss_header *rss, 715 struct mbuf *m) 716 { 717 struct adapter *sc = iq->adapter; 718 const struct cpl_pass_open_rpl *cpl = (const void *)(rss + 1); 719 int stid = GET_TID(cpl); 720 unsigned int status = cpl->status; 721 struct listen_ctx *lctx = lookup_stid(sc, stid); 722 struct inpcb *inp = lctx->inp; 723 #ifdef INVARIANTS 724 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); 725 #endif 726 727 KASSERT(opcode == CPL_PASS_OPEN_RPL, 728 ("%s: unexpected opcode 0x%x", __func__, opcode)); 729 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 730 KASSERT(lctx->stid == stid, ("%s: lctx stid mismatch", __func__)); 731 732 INP_WLOCK(inp); 733 734 CTR4(KTR_CXGBE, "%s: stid %d, status %u, flags 0x%x", 735 __func__, stid, status, lctx->flags); 736 737 lctx->flags &= ~LCTX_RPL_PENDING; 738 739 if (status != CPL_ERR_NONE) 740 log(LOG_ERR, "listener (stid %u) failed: %d\n", stid, status); 741 742 #ifdef INVARIANTS 743 /* 744 * If the inp has been dropped (listening socket closed) then 745 * listen_stop must have run and taken the inp out of the hash. 746 */ 747 if (inp->inp_flags & INP_DROPPED) { 748 KASSERT(listen_hash_del(sc, inp) == NULL, 749 ("%s: inp %p still in listen hash", __func__, inp)); 750 } 751 #endif 752 753 if (inp->inp_flags & INP_DROPPED && status != CPL_ERR_NONE) { 754 if (release_lctx(sc, lctx) != NULL) 755 INP_WUNLOCK(inp); 756 return (status); 757 } 758 759 /* 760 * Listening socket stopped listening earlier and now the chip tells us 761 * it has started the hardware listener. Stop it; the lctx will be 762 * released in do_close_server_rpl. 763 */ 764 if (inp->inp_flags & INP_DROPPED) { 765 destroy_server(sc, lctx); 766 INP_WUNLOCK(inp); 767 return (status); 768 } 769 770 /* 771 * Failed to start hardware listener. Take inp out of the hash and 772 * release our reference on it. An error message has been logged 773 * already. 774 */ 775 if (status != CPL_ERR_NONE) { 776 listen_hash_del(sc, inp); 777 if (release_lctx(sc, lctx) != NULL) 778 INP_WUNLOCK(inp); 779 return (status); 780 } 781 782 /* hardware listener open for business */ 783 784 INP_WUNLOCK(inp); 785 return (status); 786 } 787 788 static int 789 do_close_server_rpl(struct sge_iq *iq, const struct rss_header *rss, 790 struct mbuf *m) 791 { 792 struct adapter *sc = iq->adapter; 793 const struct cpl_close_listsvr_rpl *cpl = (const void *)(rss + 1); 794 int stid = GET_TID(cpl); 795 unsigned int status = cpl->status; 796 struct listen_ctx *lctx = lookup_stid(sc, stid); 797 struct inpcb *inp = lctx->inp; 798 #ifdef INVARIANTS 799 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); 800 #endif 801 802 KASSERT(opcode == CPL_CLOSE_LISTSRV_RPL, 803 ("%s: unexpected opcode 0x%x", __func__, opcode)); 804 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 805 KASSERT(lctx->stid == stid, ("%s: lctx stid mismatch", __func__)); 806 807 CTR3(KTR_CXGBE, "%s: stid %u, status %u", __func__, stid, status); 808 809 if (status != CPL_ERR_NONE) { 810 log(LOG_ERR, "%s: failed (%u) to close listener for stid %u\n", 811 __func__, status, stid); 812 return (status); 813 } 814 815 INP_WLOCK(inp); 816 inp = release_lctx(sc, lctx); 817 if (inp != NULL) 818 INP_WUNLOCK(inp); 819 820 return (status); 821 } 822 823 static void 824 done_with_synqe(struct adapter *sc, struct synq_entry *synqe) 825 { 826 struct listen_ctx *lctx = synqe->lctx; 827 struct inpcb *inp = lctx->inp; 828 struct vi_info *vi = synqe->syn->m_pkthdr.rcvif->if_softc; 829 struct l2t_entry *e = &sc->l2t->l2tab[synqe->l2e_idx]; 830 831 INP_WLOCK_ASSERT(inp); 832 833 TAILQ_REMOVE(&lctx->synq, synqe, link); 834 inp = release_lctx(sc, lctx); 835 if (inp) 836 INP_WUNLOCK(inp); 837 remove_tid(sc, synqe->tid); 838 release_tid(sc, synqe->tid, &sc->sge.ctrlq[vi->pi->port_id]); 839 t4_l2t_release(e); 840 release_synqe(synqe); /* removed from synq list */ 841 } 842 843 int 844 do_abort_req_synqe(struct sge_iq *iq, const struct rss_header *rss, 845 struct mbuf *m) 846 { 847 struct adapter *sc = iq->adapter; 848 const struct cpl_abort_req_rss *cpl = (const void *)(rss + 1); 849 unsigned int tid = GET_TID(cpl); 850 struct synq_entry *synqe = lookup_tid(sc, tid); 851 struct listen_ctx *lctx = synqe->lctx; 852 struct inpcb *inp = lctx->inp; 853 int txqid; 854 struct sge_wrq *ofld_txq; 855 #ifdef INVARIANTS 856 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); 857 #endif 858 859 KASSERT(opcode == CPL_ABORT_REQ_RSS, 860 ("%s: unexpected opcode 0x%x", __func__, opcode)); 861 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 862 KASSERT(synqe->tid == tid, ("%s: toep tid mismatch", __func__)); 863 864 CTR6(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x), lctx %p, status %d", 865 __func__, tid, synqe, synqe->flags, synqe->lctx, cpl->status); 866 867 if (negative_advice(cpl->status)) 868 return (0); /* Ignore negative advice */ 869 870 INP_WLOCK(inp); 871 872 get_qids_from_mbuf(synqe->syn, &txqid, NULL); 873 ofld_txq = &sc->sge.ofld_txq[txqid]; 874 875 /* 876 * If we'd initiated an abort earlier the reply to it is responsible for 877 * cleaning up resources. Otherwise we tear everything down right here 878 * right now. We owe the T4 a CPL_ABORT_RPL no matter what. 879 */ 880 if (synqe->flags & TPF_ABORT_SHUTDOWN) { 881 INP_WUNLOCK(inp); 882 goto done; 883 } 884 885 done_with_synqe(sc, synqe); 886 /* inp lock released by done_with_synqe */ 887 done: 888 send_abort_rpl(sc, ofld_txq, tid, CPL_ABORT_NO_RST); 889 return (0); 890 } 891 892 int 893 do_abort_rpl_synqe(struct sge_iq *iq, const struct rss_header *rss, 894 struct mbuf *m) 895 { 896 struct adapter *sc = iq->adapter; 897 const struct cpl_abort_rpl_rss *cpl = (const void *)(rss + 1); 898 unsigned int tid = GET_TID(cpl); 899 struct synq_entry *synqe = lookup_tid(sc, tid); 900 struct listen_ctx *lctx = synqe->lctx; 901 struct inpcb *inp = lctx->inp; 902 #ifdef INVARIANTS 903 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); 904 #endif 905 906 KASSERT(opcode == CPL_ABORT_RPL_RSS, 907 ("%s: unexpected opcode 0x%x", __func__, opcode)); 908 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 909 KASSERT(synqe->tid == tid, ("%s: toep tid mismatch", __func__)); 910 911 CTR6(KTR_CXGBE, "%s: tid %u, synqe %p (0x%x), lctx %p, status %d", 912 __func__, tid, synqe, synqe->flags, synqe->lctx, cpl->status); 913 914 INP_WLOCK(inp); 915 KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN, 916 ("%s: wasn't expecting abort reply for synqe %p (0x%x)", 917 __func__, synqe, synqe->flags)); 918 919 done_with_synqe(sc, synqe); 920 /* inp lock released by done_with_synqe */ 921 922 return (0); 923 } 924 925 void 926 t4_offload_socket(struct toedev *tod, void *arg, struct socket *so) 927 { 928 struct adapter *sc = tod->tod_softc; 929 struct synq_entry *synqe = arg; 930 #ifdef INVARIANTS 931 struct inpcb *inp = sotoinpcb(so); 932 #endif 933 struct cpl_pass_establish *cpl = mtod(synqe->syn, void *); 934 struct toepcb *toep = *(struct toepcb **)(cpl + 1); 935 936 INP_INFO_RLOCK_ASSERT(&V_tcbinfo); /* prevents bad race with accept() */ 937 INP_WLOCK_ASSERT(inp); 938 KASSERT(synqe->flags & TPF_SYNQE, 939 ("%s: %p not a synq_entry?", __func__, arg)); 940 941 offload_socket(so, toep); 942 make_established(toep, cpl->snd_isn, cpl->rcv_isn, cpl->tcp_opt); 943 toep->flags |= TPF_CPL_PENDING; 944 update_tid(sc, synqe->tid, toep); 945 synqe->flags |= TPF_SYNQE_EXPANDED; 946 } 947 948 static inline void 949 save_qids_in_mbuf(struct mbuf *m, struct vi_info *vi) 950 { 951 uint32_t txqid, rxqid; 952 953 txqid = (arc4random() % vi->nofldtxq) + vi->first_ofld_txq; 954 rxqid = (arc4random() % vi->nofldrxq) + vi->first_ofld_rxq; 955 956 m->m_pkthdr.flowid = (txqid << 16) | (rxqid & 0xffff); 957 } 958 959 static inline void 960 get_qids_from_mbuf(struct mbuf *m, int *txqid, int *rxqid) 961 { 962 963 if (txqid) 964 *txqid = m->m_pkthdr.flowid >> 16; 965 if (rxqid) 966 *rxqid = m->m_pkthdr.flowid & 0xffff; 967 } 968 969 /* 970 * Use the trailing space in the mbuf in which the PASS_ACCEPT_REQ arrived to 971 * store some state temporarily. 972 */ 973 static struct synq_entry * 974 mbuf_to_synqe(struct mbuf *m) 975 { 976 int len = roundup2(sizeof (struct synq_entry), 8); 977 int tspace = M_TRAILINGSPACE(m); 978 struct synq_entry *synqe = NULL; 979 980 if (tspace < len) { 981 synqe = malloc(sizeof(*synqe), M_CXGBE, M_NOWAIT); 982 if (synqe == NULL) 983 return (NULL); 984 synqe->flags = TPF_SYNQE | TPF_SYNQE_NEEDFREE; 985 } else { 986 synqe = (void *)(m->m_data + m->m_len + tspace - len); 987 synqe->flags = TPF_SYNQE; 988 } 989 990 return (synqe); 991 } 992 993 static void 994 t4opt_to_tcpopt(const struct tcp_options *t4opt, struct tcpopt *to) 995 { 996 bzero(to, sizeof(*to)); 997 998 if (t4opt->mss) { 999 to->to_flags |= TOF_MSS; 1000 to->to_mss = be16toh(t4opt->mss); 1001 } 1002 1003 if (t4opt->wsf) { 1004 to->to_flags |= TOF_SCALE; 1005 to->to_wscale = t4opt->wsf; 1006 } 1007 1008 if (t4opt->tstamp) 1009 to->to_flags |= TOF_TS; 1010 1011 if (t4opt->sack) 1012 to->to_flags |= TOF_SACKPERM; 1013 } 1014 1015 /* 1016 * Options2 for passive open. 1017 */ 1018 static uint32_t 1019 calc_opt2p(struct adapter *sc, struct port_info *pi, int rxqid, 1020 const struct tcp_options *tcpopt, struct tcphdr *th, int ulp_mode) 1021 { 1022 struct sge_ofld_rxq *ofld_rxq = &sc->sge.ofld_rxq[rxqid]; 1023 uint32_t opt2; 1024 1025 opt2 = V_TX_QUEUE(sc->params.tp.tx_modq[pi->tx_chan]) | 1026 F_RSS_QUEUE_VALID | V_RSS_QUEUE(ofld_rxq->iq.abs_id); 1027 1028 if (V_tcp_do_rfc1323) { 1029 if (tcpopt->tstamp) 1030 opt2 |= F_TSTAMPS_EN; 1031 if (tcpopt->sack) 1032 opt2 |= F_SACK_EN; 1033 if (tcpopt->wsf <= 14) 1034 opt2 |= F_WND_SCALE_EN; 1035 } 1036 1037 if (V_tcp_do_ecn && th->th_flags & (TH_ECE | TH_CWR)) 1038 opt2 |= F_CCTRL_ECN; 1039 1040 /* RX_COALESCE is always a valid value (0 or M_RX_COALESCE). */ 1041 if (is_t4(sc)) 1042 opt2 |= F_RX_COALESCE_VALID; 1043 else { 1044 opt2 |= F_T5_OPT_2_VALID; 1045 opt2 |= F_CONG_CNTRL_VALID; /* OPT_2_ISS really, for T5 */ 1046 } 1047 if (sc->tt.rx_coalesce) 1048 opt2 |= V_RX_COALESCE(M_RX_COALESCE); 1049 1050 #ifdef USE_DDP_RX_FLOW_CONTROL 1051 if (ulp_mode == ULP_MODE_TCPDDP) 1052 opt2 |= F_RX_FC_VALID | F_RX_FC_DDP; 1053 #endif 1054 1055 return htobe32(opt2); 1056 } 1057 1058 static void 1059 pass_accept_req_to_protohdrs(const struct mbuf *m, struct in_conninfo *inc, 1060 struct tcphdr *th) 1061 { 1062 const struct cpl_pass_accept_req *cpl = mtod(m, const void *); 1063 const struct ether_header *eh; 1064 unsigned int hlen = be32toh(cpl->hdr_len); 1065 uintptr_t l3hdr; 1066 const struct tcphdr *tcp; 1067 1068 eh = (const void *)(cpl + 1); 1069 l3hdr = ((uintptr_t)eh + G_ETH_HDR_LEN(hlen)); 1070 tcp = (const void *)(l3hdr + G_IP_HDR_LEN(hlen)); 1071 1072 if (inc) { 1073 bzero(inc, sizeof(*inc)); 1074 inc->inc_fport = tcp->th_sport; 1075 inc->inc_lport = tcp->th_dport; 1076 if (((struct ip *)l3hdr)->ip_v == IPVERSION) { 1077 const struct ip *ip = (const void *)l3hdr; 1078 1079 inc->inc_faddr = ip->ip_src; 1080 inc->inc_laddr = ip->ip_dst; 1081 } else { 1082 const struct ip6_hdr *ip6 = (const void *)l3hdr; 1083 1084 inc->inc_flags |= INC_ISIPV6; 1085 inc->inc6_faddr = ip6->ip6_src; 1086 inc->inc6_laddr = ip6->ip6_dst; 1087 } 1088 } 1089 1090 if (th) { 1091 bcopy(tcp, th, sizeof(*th)); 1092 tcp_fields_to_host(th); /* just like tcp_input */ 1093 } 1094 } 1095 1096 static struct l2t_entry * 1097 get_l2te_for_nexthop(struct port_info *pi, struct ifnet *ifp, 1098 struct in_conninfo *inc) 1099 { 1100 struct l2t_entry *e; 1101 struct sockaddr_in6 sin6; 1102 struct sockaddr *dst = (void *)&sin6; 1103 1104 if (inc->inc_flags & INC_ISIPV6) { 1105 struct nhop6_basic nh6; 1106 1107 bzero(dst, sizeof(struct sockaddr_in6)); 1108 dst->sa_len = sizeof(struct sockaddr_in6); 1109 dst->sa_family = AF_INET6; 1110 1111 if (IN6_IS_ADDR_LINKLOCAL(&inc->inc6_laddr)) { 1112 /* no need for route lookup */ 1113 e = t4_l2t_get(pi, ifp, dst); 1114 return (e); 1115 } 1116 1117 if (fib6_lookup_nh_basic(RT_DEFAULT_FIB, &inc->inc6_faddr, 1118 0, 0, 0, &nh6) != 0) 1119 return (NULL); 1120 if (nh6.nh_ifp != ifp) 1121 return (NULL); 1122 ((struct sockaddr_in6 *)dst)->sin6_addr = nh6.nh_addr; 1123 } else { 1124 struct nhop4_basic nh4; 1125 1126 dst->sa_len = sizeof(struct sockaddr_in); 1127 dst->sa_family = AF_INET; 1128 1129 if (fib4_lookup_nh_basic(RT_DEFAULT_FIB, inc->inc_faddr, 0, 0, 1130 &nh4) != 0) 1131 return (NULL); 1132 if (nh4.nh_ifp != ifp) 1133 return (NULL); 1134 ((struct sockaddr_in *)dst)->sin_addr = nh4.nh_addr; 1135 } 1136 1137 e = t4_l2t_get(pi, ifp, dst); 1138 return (e); 1139 } 1140 1141 #define REJECT_PASS_ACCEPT() do { \ 1142 reject_reason = __LINE__; \ 1143 goto reject; \ 1144 } while (0) 1145 1146 /* 1147 * The context associated with a tid entry via insert_tid could be a synq_entry 1148 * or a toepcb. The only way CPL handlers can tell is via a bit in these flags. 1149 */ 1150 CTASSERT(offsetof(struct toepcb, flags) == offsetof(struct synq_entry, flags)); 1151 1152 /* 1153 * Incoming SYN on a listening socket. 1154 * 1155 * XXX: Every use of ifp in this routine has a bad race with up/down, toe/-toe, 1156 * etc. 1157 */ 1158 static int 1159 do_pass_accept_req(struct sge_iq *iq, const struct rss_header *rss, 1160 struct mbuf *m) 1161 { 1162 struct adapter *sc = iq->adapter; 1163 struct toedev *tod; 1164 const struct cpl_pass_accept_req *cpl = mtod(m, const void *); 1165 struct cpl_pass_accept_rpl *rpl; 1166 struct wrqe *wr; 1167 unsigned int stid = G_PASS_OPEN_TID(be32toh(cpl->tos_stid)); 1168 unsigned int tid = GET_TID(cpl); 1169 struct listen_ctx *lctx = lookup_stid(sc, stid); 1170 struct inpcb *inp; 1171 struct socket *so; 1172 struct in_conninfo inc; 1173 struct tcphdr th; 1174 struct tcpopt to; 1175 struct port_info *pi; 1176 struct vi_info *vi; 1177 struct ifnet *hw_ifp, *ifp; 1178 struct l2t_entry *e = NULL; 1179 int rscale, mtu_idx, rx_credits, rxqid, ulp_mode; 1180 struct synq_entry *synqe = NULL; 1181 int reject_reason, v; 1182 uint16_t vid; 1183 #ifdef INVARIANTS 1184 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); 1185 #endif 1186 1187 KASSERT(opcode == CPL_PASS_ACCEPT_REQ, 1188 ("%s: unexpected opcode 0x%x", __func__, opcode)); 1189 KASSERT(lctx->stid == stid, ("%s: lctx stid mismatch", __func__)); 1190 1191 CTR4(KTR_CXGBE, "%s: stid %u, tid %u, lctx %p", __func__, stid, tid, 1192 lctx); 1193 1194 pass_accept_req_to_protohdrs(m, &inc, &th); 1195 t4opt_to_tcpopt(&cpl->tcpopt, &to); 1196 1197 pi = sc->port[G_SYN_INTF(be16toh(cpl->l2info))]; 1198 1199 /* 1200 * Use the MAC index to lookup the associated VI. If this SYN 1201 * didn't match a perfect MAC filter, punt. 1202 */ 1203 if (!(be16toh(cpl->l2info) & F_SYN_XACT_MATCH)) { 1204 m_freem(m); 1205 m = NULL; 1206 REJECT_PASS_ACCEPT(); 1207 } 1208 for_each_vi(pi, v, vi) { 1209 if (vi->xact_addr_filt == G_SYN_MAC_IDX(be16toh(cpl->l2info))) 1210 goto found; 1211 } 1212 m_freem(m); 1213 m = NULL; 1214 REJECT_PASS_ACCEPT(); 1215 1216 found: 1217 hw_ifp = vi->ifp; /* the (v)cxgbeX ifnet */ 1218 m->m_pkthdr.rcvif = hw_ifp; 1219 tod = TOEDEV(hw_ifp); 1220 1221 /* 1222 * Figure out if there is a pseudo interface (vlan, lagg, etc.) 1223 * involved. Don't offload if the SYN had a VLAN tag and the vid 1224 * doesn't match anything on this interface. 1225 * 1226 * XXX: lagg support, lagg + vlan support. 1227 */ 1228 vid = EVL_VLANOFTAG(be16toh(cpl->vlan)); 1229 if (vid != 0xfff) { 1230 ifp = VLAN_DEVAT(hw_ifp, vid); 1231 if (ifp == NULL) 1232 REJECT_PASS_ACCEPT(); 1233 } else 1234 ifp = hw_ifp; 1235 1236 /* 1237 * Don't offload if the peer requested a TCP option that's not known to 1238 * the silicon. 1239 */ 1240 if (cpl->tcpopt.unknown) 1241 REJECT_PASS_ACCEPT(); 1242 1243 if (inc.inc_flags & INC_ISIPV6) { 1244 1245 /* Don't offload if the ifcap isn't enabled */ 1246 if ((ifp->if_capenable & IFCAP_TOE6) == 0) 1247 REJECT_PASS_ACCEPT(); 1248 1249 /* 1250 * SYN must be directed to an IP6 address on this ifnet. This 1251 * is more restrictive than in6_localip. 1252 */ 1253 if (!in6_ifhasaddr(ifp, &inc.inc6_laddr)) 1254 REJECT_PASS_ACCEPT(); 1255 } else { 1256 1257 /* Don't offload if the ifcap isn't enabled */ 1258 if ((ifp->if_capenable & IFCAP_TOE4) == 0) 1259 REJECT_PASS_ACCEPT(); 1260 1261 /* 1262 * SYN must be directed to an IP address on this ifnet. This 1263 * is more restrictive than in_localip. 1264 */ 1265 if (!in_ifhasaddr(ifp, inc.inc_laddr)) 1266 REJECT_PASS_ACCEPT(); 1267 } 1268 1269 e = get_l2te_for_nexthop(pi, ifp, &inc); 1270 if (e == NULL) 1271 REJECT_PASS_ACCEPT(); 1272 1273 synqe = mbuf_to_synqe(m); 1274 if (synqe == NULL) 1275 REJECT_PASS_ACCEPT(); 1276 1277 wr = alloc_wrqe(is_t4(sc) ? sizeof(struct cpl_pass_accept_rpl) : 1278 sizeof(struct cpl_t5_pass_accept_rpl), &sc->sge.ctrlq[pi->port_id]); 1279 if (wr == NULL) 1280 REJECT_PASS_ACCEPT(); 1281 rpl = wrtod(wr); 1282 1283 INP_INFO_RLOCK(&V_tcbinfo); /* for 4-tuple check */ 1284 1285 /* Don't offload if the 4-tuple is already in use */ 1286 if (toe_4tuple_check(&inc, &th, ifp) != 0) { 1287 INP_INFO_RUNLOCK(&V_tcbinfo); 1288 free(wr, M_CXGBE); 1289 REJECT_PASS_ACCEPT(); 1290 } 1291 INP_INFO_RUNLOCK(&V_tcbinfo); 1292 1293 inp = lctx->inp; /* listening socket, not owned by TOE */ 1294 INP_WLOCK(inp); 1295 1296 /* Don't offload if the listening socket has closed */ 1297 if (__predict_false(inp->inp_flags & INP_DROPPED)) { 1298 /* 1299 * The listening socket has closed. The reply from the TOE to 1300 * our CPL_CLOSE_LISTSRV_REQ will ultimately release all 1301 * resources tied to this listen context. 1302 */ 1303 INP_WUNLOCK(inp); 1304 free(wr, M_CXGBE); 1305 REJECT_PASS_ACCEPT(); 1306 } 1307 so = inp->inp_socket; 1308 1309 mtu_idx = find_best_mtu_idx(sc, &inc, be16toh(cpl->tcpopt.mss)); 1310 rscale = cpl->tcpopt.wsf && V_tcp_do_rfc1323 ? select_rcv_wscale() : 0; 1311 SOCKBUF_LOCK(&so->so_rcv); 1312 /* opt0 rcv_bufsiz initially, assumes its normal meaning later */ 1313 rx_credits = min(select_rcv_wnd(so) >> 10, M_RCV_BUFSIZ); 1314 SOCKBUF_UNLOCK(&so->so_rcv); 1315 1316 save_qids_in_mbuf(m, vi); 1317 get_qids_from_mbuf(m, NULL, &rxqid); 1318 1319 if (is_t4(sc)) 1320 INIT_TP_WR_MIT_CPL(rpl, CPL_PASS_ACCEPT_RPL, tid); 1321 else { 1322 struct cpl_t5_pass_accept_rpl *rpl5 = (void *)rpl; 1323 1324 INIT_TP_WR_MIT_CPL(rpl5, CPL_PASS_ACCEPT_RPL, tid); 1325 } 1326 if (sc->tt.ddp && (so->so_options & SO_NO_DDP) == 0) { 1327 ulp_mode = ULP_MODE_TCPDDP; 1328 synqe->flags |= TPF_SYNQE_TCPDDP; 1329 } else 1330 ulp_mode = ULP_MODE_NONE; 1331 rpl->opt0 = calc_opt0(so, vi, e, mtu_idx, rscale, rx_credits, ulp_mode); 1332 rpl->opt2 = calc_opt2p(sc, pi, rxqid, &cpl->tcpopt, &th, ulp_mode); 1333 1334 synqe->tid = tid; 1335 synqe->lctx = lctx; 1336 synqe->syn = m; 1337 m = NULL; 1338 refcount_init(&synqe->refcnt, 1); /* 1 means extra hold */ 1339 synqe->l2e_idx = e->idx; 1340 synqe->rcv_bufsize = rx_credits; 1341 atomic_store_rel_ptr(&synqe->wr, (uintptr_t)wr); 1342 1343 insert_tid(sc, tid, synqe); 1344 TAILQ_INSERT_TAIL(&lctx->synq, synqe, link); 1345 hold_synqe(synqe); /* hold for the duration it's in the synq */ 1346 hold_lctx(lctx); /* A synqe on the list has a ref on its lctx */ 1347 1348 /* 1349 * If all goes well t4_syncache_respond will get called during 1350 * syncache_add. Note that syncache_add releases the pcb lock. 1351 */ 1352 toe_syncache_add(&inc, &to, &th, inp, tod, synqe); 1353 INP_UNLOCK_ASSERT(inp); /* ok to assert, we have a ref on the inp */ 1354 1355 /* 1356 * If we replied during syncache_add (synqe->wr has been consumed), 1357 * good. Otherwise, set it to 0 so that further syncache_respond 1358 * attempts by the kernel will be ignored. 1359 */ 1360 if (atomic_cmpset_ptr(&synqe->wr, (uintptr_t)wr, 0)) { 1361 1362 /* 1363 * syncache may or may not have a hold on the synqe, which may 1364 * or may not be stashed in the original SYN mbuf passed to us. 1365 * Just copy it over instead of dealing with all possibilities. 1366 */ 1367 m = m_dup(synqe->syn, M_NOWAIT); 1368 if (m) 1369 m->m_pkthdr.rcvif = hw_ifp; 1370 1371 remove_tid(sc, synqe->tid); 1372 free(wr, M_CXGBE); 1373 1374 /* Yank the synqe out of the lctx synq. */ 1375 INP_WLOCK(inp); 1376 TAILQ_REMOVE(&lctx->synq, synqe, link); 1377 release_synqe(synqe); /* removed from synq list */ 1378 inp = release_lctx(sc, lctx); 1379 if (inp) 1380 INP_WUNLOCK(inp); 1381 1382 release_synqe(synqe); /* extra hold */ 1383 REJECT_PASS_ACCEPT(); 1384 } 1385 1386 CTR5(KTR_CXGBE, "%s: stid %u, tid %u, lctx %p, synqe %p, SYNACK", 1387 __func__, stid, tid, lctx, synqe); 1388 1389 INP_WLOCK(inp); 1390 synqe->flags |= TPF_SYNQE_HAS_L2TE; 1391 if (__predict_false(inp->inp_flags & INP_DROPPED)) { 1392 /* 1393 * Listening socket closed but tod_listen_stop did not abort 1394 * this tid because there was no L2T entry for the tid at that 1395 * time. Abort it now. The reply to the abort will clean up. 1396 */ 1397 CTR6(KTR_CXGBE, 1398 "%s: stid %u, tid %u, lctx %p, synqe %p (0x%x), ABORT", 1399 __func__, stid, tid, lctx, synqe, synqe->flags); 1400 if (!(synqe->flags & TPF_SYNQE_EXPANDED)) 1401 send_reset_synqe(tod, synqe); 1402 INP_WUNLOCK(inp); 1403 1404 release_synqe(synqe); /* extra hold */ 1405 return (__LINE__); 1406 } 1407 INP_WUNLOCK(inp); 1408 1409 release_synqe(synqe); /* extra hold */ 1410 return (0); 1411 reject: 1412 CTR4(KTR_CXGBE, "%s: stid %u, tid %u, REJECT (%d)", __func__, stid, tid, 1413 reject_reason); 1414 1415 if (e) 1416 t4_l2t_release(e); 1417 release_tid(sc, tid, lctx->ctrlq); 1418 1419 if (__predict_true(m != NULL)) { 1420 m_adj(m, sizeof(*cpl)); 1421 m->m_pkthdr.csum_flags |= (CSUM_IP_CHECKED | CSUM_IP_VALID | 1422 CSUM_DATA_VALID | CSUM_PSEUDO_HDR); 1423 m->m_pkthdr.csum_data = 0xffff; 1424 hw_ifp->if_input(hw_ifp, m); 1425 } 1426 1427 return (reject_reason); 1428 } 1429 1430 static void 1431 synqe_to_protohdrs(struct synq_entry *synqe, 1432 const struct cpl_pass_establish *cpl, struct in_conninfo *inc, 1433 struct tcphdr *th, struct tcpopt *to) 1434 { 1435 uint16_t tcp_opt = be16toh(cpl->tcp_opt); 1436 1437 /* start off with the original SYN */ 1438 pass_accept_req_to_protohdrs(synqe->syn, inc, th); 1439 1440 /* modify parts to make it look like the ACK to our SYN|ACK */ 1441 th->th_flags = TH_ACK; 1442 th->th_ack = synqe->iss + 1; 1443 th->th_seq = be32toh(cpl->rcv_isn); 1444 bzero(to, sizeof(*to)); 1445 if (G_TCPOPT_TSTAMP(tcp_opt)) { 1446 to->to_flags |= TOF_TS; 1447 to->to_tsecr = synqe->ts; 1448 } 1449 } 1450 1451 static int 1452 do_pass_establish(struct sge_iq *iq, const struct rss_header *rss, 1453 struct mbuf *m) 1454 { 1455 struct adapter *sc = iq->adapter; 1456 struct vi_info *vi; 1457 struct ifnet *ifp; 1458 const struct cpl_pass_establish *cpl = (const void *)(rss + 1); 1459 #if defined(KTR) || defined(INVARIANTS) 1460 unsigned int stid = G_PASS_OPEN_TID(be32toh(cpl->tos_stid)); 1461 #endif 1462 unsigned int tid = GET_TID(cpl); 1463 struct synq_entry *synqe = lookup_tid(sc, tid); 1464 struct listen_ctx *lctx = synqe->lctx; 1465 struct inpcb *inp = lctx->inp, *new_inp; 1466 struct socket *so; 1467 struct tcphdr th; 1468 struct tcpopt to; 1469 struct in_conninfo inc; 1470 struct toepcb *toep; 1471 u_int txqid, rxqid; 1472 #ifdef INVARIANTS 1473 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); 1474 #endif 1475 1476 KASSERT(opcode == CPL_PASS_ESTABLISH, 1477 ("%s: unexpected opcode 0x%x", __func__, opcode)); 1478 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 1479 KASSERT(lctx->stid == stid, ("%s: lctx stid mismatch", __func__)); 1480 KASSERT(synqe->flags & TPF_SYNQE, 1481 ("%s: tid %u (ctx %p) not a synqe", __func__, tid, synqe)); 1482 1483 INP_INFO_RLOCK(&V_tcbinfo); /* for syncache_expand */ 1484 INP_WLOCK(inp); 1485 1486 CTR6(KTR_CXGBE, 1487 "%s: stid %u, tid %u, synqe %p (0x%x), inp_flags 0x%x", 1488 __func__, stid, tid, synqe, synqe->flags, inp->inp_flags); 1489 1490 if (__predict_false(inp->inp_flags & INP_DROPPED)) { 1491 1492 if (synqe->flags & TPF_SYNQE_HAS_L2TE) { 1493 KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN, 1494 ("%s: listen socket closed but tid %u not aborted.", 1495 __func__, tid)); 1496 } 1497 1498 INP_WUNLOCK(inp); 1499 INP_INFO_RUNLOCK(&V_tcbinfo); 1500 return (0); 1501 } 1502 1503 ifp = synqe->syn->m_pkthdr.rcvif; 1504 vi = ifp->if_softc; 1505 KASSERT(vi->pi->adapter == sc, 1506 ("%s: vi %p, sc %p mismatch", __func__, vi, sc)); 1507 1508 get_qids_from_mbuf(synqe->syn, &txqid, &rxqid); 1509 KASSERT(rxqid == iq_to_ofld_rxq(iq) - &sc->sge.ofld_rxq[0], 1510 ("%s: CPL arrived on unexpected rxq. %d %d", __func__, rxqid, 1511 (int)(iq_to_ofld_rxq(iq) - &sc->sge.ofld_rxq[0]))); 1512 1513 toep = alloc_toepcb(vi, txqid, rxqid, M_NOWAIT); 1514 if (toep == NULL) { 1515 reset: 1516 /* 1517 * The reply to this abort will perform final cleanup. There is 1518 * no need to check for HAS_L2TE here. We can be here only if 1519 * we responded to the PASS_ACCEPT_REQ, and our response had the 1520 * L2T idx. 1521 */ 1522 send_reset_synqe(TOEDEV(ifp), synqe); 1523 INP_WUNLOCK(inp); 1524 INP_INFO_RUNLOCK(&V_tcbinfo); 1525 return (0); 1526 } 1527 toep->tid = tid; 1528 toep->l2te = &sc->l2t->l2tab[synqe->l2e_idx]; 1529 if (synqe->flags & TPF_SYNQE_TCPDDP) 1530 set_tcpddp_ulp_mode(toep); 1531 else 1532 toep->ulp_mode = ULP_MODE_NONE; 1533 /* opt0 rcv_bufsiz initially, assumes its normal meaning later */ 1534 toep->rx_credits = synqe->rcv_bufsize; 1535 1536 so = inp->inp_socket; 1537 KASSERT(so != NULL, ("%s: socket is NULL", __func__)); 1538 1539 /* Come up with something that syncache_expand should be ok with. */ 1540 synqe_to_protohdrs(synqe, cpl, &inc, &th, &to); 1541 1542 /* 1543 * No more need for anything in the mbuf that carried the 1544 * CPL_PASS_ACCEPT_REQ. Drop the CPL_PASS_ESTABLISH and toep pointer 1545 * there. XXX: bad form but I don't want to increase the size of synqe. 1546 */ 1547 m = synqe->syn; 1548 KASSERT(sizeof(*cpl) + sizeof(toep) <= m->m_len, 1549 ("%s: no room in mbuf %p (m_len %d)", __func__, m, m->m_len)); 1550 bcopy(cpl, mtod(m, void *), sizeof(*cpl)); 1551 *(struct toepcb **)(mtod(m, struct cpl_pass_establish *) + 1) = toep; 1552 1553 if (!toe_syncache_expand(&inc, &to, &th, &so) || so == NULL) { 1554 free_toepcb(toep); 1555 goto reset; 1556 } 1557 1558 /* New connection inpcb is already locked by syncache_expand(). */ 1559 new_inp = sotoinpcb(so); 1560 INP_WLOCK_ASSERT(new_inp); 1561 1562 /* 1563 * This is for the unlikely case where the syncache entry that we added 1564 * has been evicted from the syncache, but the syncache_expand above 1565 * works because of syncookies. 1566 * 1567 * XXX: we've held the tcbinfo lock throughout so there's no risk of 1568 * anyone accept'ing a connection before we've installed our hooks, but 1569 * this somewhat defeats the purpose of having a tod_offload_socket :-( 1570 */ 1571 if (__predict_false(!(synqe->flags & TPF_SYNQE_EXPANDED))) { 1572 tcp_timer_activate(intotcpcb(new_inp), TT_KEEP, 0); 1573 t4_offload_socket(TOEDEV(ifp), synqe, so); 1574 } 1575 1576 INP_WUNLOCK(new_inp); 1577 1578 /* Done with the synqe */ 1579 TAILQ_REMOVE(&lctx->synq, synqe, link); 1580 inp = release_lctx(sc, lctx); 1581 if (inp != NULL) 1582 INP_WUNLOCK(inp); 1583 INP_INFO_RUNLOCK(&V_tcbinfo); 1584 release_synqe(synqe); 1585 1586 return (0); 1587 } 1588 1589 void 1590 t4_init_listen_cpl_handlers(struct adapter *sc) 1591 { 1592 1593 t4_register_cpl_handler(sc, CPL_PASS_OPEN_RPL, do_pass_open_rpl); 1594 t4_register_cpl_handler(sc, CPL_CLOSE_LISTSRV_RPL, do_close_server_rpl); 1595 t4_register_cpl_handler(sc, CPL_PASS_ACCEPT_REQ, do_pass_accept_req); 1596 t4_register_cpl_handler(sc, CPL_PASS_ESTABLISH, do_pass_establish); 1597 } 1598 #endif 1599