t4_listen.c revision 309560
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: stable/11/sys/dev/cxgbe/tom/t4_listen.c 309560 2016-12-05 20:43:25Z jhb $"); 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#define TCPSTATES 60#include <netinet/tcp_fsm.h> 61#include <netinet/tcp_var.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 */ 71static int alloc_stid(struct adapter *, struct listen_ctx *, int); 72static struct listen_ctx *lookup_stid(struct adapter *, int); 73static void free_stid(struct adapter *, struct listen_ctx *); 74 75/* lctx services */ 76static struct listen_ctx *alloc_lctx(struct adapter *, struct inpcb *, 77 struct vi_info *); 78static int free_lctx(struct adapter *, struct listen_ctx *); 79static void hold_lctx(struct listen_ctx *); 80static void listen_hash_add(struct adapter *, struct listen_ctx *); 81static struct listen_ctx *listen_hash_find(struct adapter *, struct inpcb *); 82static struct listen_ctx *listen_hash_del(struct adapter *, struct inpcb *); 83static struct inpcb *release_lctx(struct adapter *, struct listen_ctx *); 84 85static inline void save_qids_in_mbuf(struct mbuf *, struct vi_info *); 86static inline void get_qids_from_mbuf(struct mbuf *m, int *, int *); 87static void send_reset_synqe(struct toedev *, struct synq_entry *); 88 89static int 90alloc_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 148allocated: 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 160static struct listen_ctx * 161lookup_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 168static void 169free_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 191static struct listen_ctx * 192alloc_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 */ 231static int 232free_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 255static void 256hold_lctx(struct listen_ctx *lctx) 257{ 258 259 refcount_acquire(&lctx->refcount); 260} 261 262static inline uint32_t 263listen_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 */ 272static void 273listen_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 */ 287static struct listen_ctx * 288listen_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 */ 307static struct listen_ctx * 308listen_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 */ 332static struct inpcb * 333release_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 345static void 346send_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 417static int 418create_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 445static int 446create_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 475static int 476destroy_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 */ 504int 505t4_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 an initialized 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->flags & VI_INIT_DONE && 547 vi->ifp->if_capenable & IFCAP_TOE) 548 goto found; 549 } 550 } 551 goto done; /* no port that's UP with IFCAP_TOE enabled */ 552found: 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; 584done: 585#if 0 586 ADAPTER_UNLOCK(sc); 587#endif 588 return (0); 589} 590 591int 592t4_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 632static inline void 633hold_synqe(struct synq_entry *synqe) 634{ 635 636 refcount_acquire(&synqe->refcnt); 637} 638 639static inline void 640release_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 652void 653t4_syncache_added(struct toedev *tod __unused, void *arg) 654{ 655 struct synq_entry *synqe = arg; 656 657 hold_synqe(synqe); 658} 659 660void 661t4_syncache_removed(struct toedev *tod __unused, void *arg) 662{ 663 struct synq_entry *synqe = arg; 664 665 release_synqe(synqe); 666} 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 (chip_id(sc) >= CHELSIO_T5) { 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 vi_info *vi = 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[vi->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_RLOCK_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 vi_info *vi) 947{ 948 uint32_t txqid, rxqid; 949 950 txqid = (arc4random() % vi->nofldtxq) + vi->first_ofld_txq; 951 rxqid = (arc4random() % vi->nofldrxq) + vi->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(struct adapter *sc, const struct mbuf *m, 1057 struct in_conninfo *inc, 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 if (chip_id(sc) >= CHELSIO_T6) { 1067 l3hdr = ((uintptr_t)eh + G_T6_ETH_HDR_LEN(hlen)); 1068 tcp = (const void *)(l3hdr + G_T6_IP_HDR_LEN(hlen)); 1069 } else { 1070 l3hdr = ((uintptr_t)eh + G_ETH_HDR_LEN(hlen)); 1071 tcp = (const void *)(l3hdr + G_IP_HDR_LEN(hlen)); 1072 } 1073 1074 if (inc) { 1075 bzero(inc, sizeof(*inc)); 1076 inc->inc_fport = tcp->th_sport; 1077 inc->inc_lport = tcp->th_dport; 1078 if (((struct ip *)l3hdr)->ip_v == IPVERSION) { 1079 const struct ip *ip = (const void *)l3hdr; 1080 1081 inc->inc_faddr = ip->ip_src; 1082 inc->inc_laddr = ip->ip_dst; 1083 } else { 1084 const struct ip6_hdr *ip6 = (const void *)l3hdr; 1085 1086 inc->inc_flags |= INC_ISIPV6; 1087 inc->inc6_faddr = ip6->ip6_src; 1088 inc->inc6_laddr = ip6->ip6_dst; 1089 } 1090 } 1091 1092 if (th) { 1093 bcopy(tcp, th, sizeof(*th)); 1094 tcp_fields_to_host(th); /* just like tcp_input */ 1095 } 1096} 1097 1098static struct l2t_entry * 1099get_l2te_for_nexthop(struct port_info *pi, struct ifnet *ifp, 1100 struct in_conninfo *inc) 1101{ 1102 struct l2t_entry *e; 1103 struct sockaddr_in6 sin6; 1104 struct sockaddr *dst = (void *)&sin6; 1105 1106 if (inc->inc_flags & INC_ISIPV6) { 1107 struct nhop6_basic nh6; 1108 1109 bzero(dst, sizeof(struct sockaddr_in6)); 1110 dst->sa_len = sizeof(struct sockaddr_in6); 1111 dst->sa_family = AF_INET6; 1112 1113 if (IN6_IS_ADDR_LINKLOCAL(&inc->inc6_laddr)) { 1114 /* no need for route lookup */ 1115 e = t4_l2t_get(pi, ifp, dst); 1116 return (e); 1117 } 1118 1119 if (fib6_lookup_nh_basic(RT_DEFAULT_FIB, &inc->inc6_faddr, 1120 0, 0, 0, &nh6) != 0) 1121 return (NULL); 1122 if (nh6.nh_ifp != ifp) 1123 return (NULL); 1124 ((struct sockaddr_in6 *)dst)->sin6_addr = nh6.nh_addr; 1125 } else { 1126 struct nhop4_basic nh4; 1127 1128 dst->sa_len = sizeof(struct sockaddr_in); 1129 dst->sa_family = AF_INET; 1130 1131 if (fib4_lookup_nh_basic(RT_DEFAULT_FIB, inc->inc_faddr, 0, 0, 1132 &nh4) != 0) 1133 return (NULL); 1134 if (nh4.nh_ifp != ifp) 1135 return (NULL); 1136 ((struct sockaddr_in *)dst)->sin_addr = nh4.nh_addr; 1137 } 1138 1139 e = t4_l2t_get(pi, ifp, dst); 1140 return (e); 1141} 1142 1143#define REJECT_PASS_ACCEPT() do { \ 1144 reject_reason = __LINE__; \ 1145 goto reject; \ 1146} while (0) 1147 1148/* 1149 * The context associated with a tid entry via insert_tid could be a synq_entry 1150 * or a toepcb. The only way CPL handlers can tell is via a bit in these flags. 1151 */ 1152CTASSERT(offsetof(struct toepcb, flags) == offsetof(struct synq_entry, flags)); 1153 1154/* 1155 * Incoming SYN on a listening socket. 1156 * 1157 * XXX: Every use of ifp in this routine has a bad race with up/down, toe/-toe, 1158 * etc. 1159 */ 1160static int 1161do_pass_accept_req(struct sge_iq *iq, const struct rss_header *rss, 1162 struct mbuf *m) 1163{ 1164 struct adapter *sc = iq->adapter; 1165 struct toedev *tod; 1166 const struct cpl_pass_accept_req *cpl = mtod(m, const void *); 1167 struct cpl_pass_accept_rpl *rpl; 1168 struct wrqe *wr; 1169 unsigned int stid = G_PASS_OPEN_TID(be32toh(cpl->tos_stid)); 1170 unsigned int tid = GET_TID(cpl); 1171 struct listen_ctx *lctx = lookup_stid(sc, stid); 1172 struct inpcb *inp; 1173 struct socket *so; 1174 struct in_conninfo inc; 1175 struct tcphdr th; 1176 struct tcpopt to; 1177 struct port_info *pi; 1178 struct vi_info *vi; 1179 struct ifnet *hw_ifp, *ifp; 1180 struct l2t_entry *e = NULL; 1181 int rscale, mtu_idx, rx_credits, rxqid, ulp_mode; 1182 struct synq_entry *synqe = NULL; 1183 int reject_reason, v; 1184 uint16_t vid; 1185#ifdef INVARIANTS 1186 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); 1187#endif 1188 1189 KASSERT(opcode == CPL_PASS_ACCEPT_REQ, 1190 ("%s: unexpected opcode 0x%x", __func__, opcode)); 1191 KASSERT(lctx->stid == stid, ("%s: lctx stid mismatch", __func__)); 1192 1193 CTR4(KTR_CXGBE, "%s: stid %u, tid %u, lctx %p", __func__, stid, tid, 1194 lctx); 1195 1196 pass_accept_req_to_protohdrs(sc, m, &inc, &th); 1197 t4opt_to_tcpopt(&cpl->tcpopt, &to); 1198 1199 pi = sc->port[G_SYN_INTF(be16toh(cpl->l2info))]; 1200 1201 /* 1202 * Use the MAC index to lookup the associated VI. If this SYN 1203 * didn't match a perfect MAC filter, punt. 1204 */ 1205 if (!(be16toh(cpl->l2info) & F_SYN_XACT_MATCH)) { 1206 m_freem(m); 1207 m = NULL; 1208 REJECT_PASS_ACCEPT(); 1209 } 1210 for_each_vi(pi, v, vi) { 1211 if (vi->xact_addr_filt == G_SYN_MAC_IDX(be16toh(cpl->l2info))) 1212 goto found; 1213 } 1214 m_freem(m); 1215 m = NULL; 1216 REJECT_PASS_ACCEPT(); 1217 1218found: 1219 hw_ifp = vi->ifp; /* the (v)cxgbeX ifnet */ 1220 m->m_pkthdr.rcvif = hw_ifp; 1221 tod = TOEDEV(hw_ifp); 1222 1223 /* 1224 * Figure out if there is a pseudo interface (vlan, lagg, etc.) 1225 * involved. Don't offload if the SYN had a VLAN tag and the vid 1226 * doesn't match anything on this interface. 1227 * 1228 * XXX: lagg support, lagg + vlan support. 1229 */ 1230 vid = EVL_VLANOFTAG(be16toh(cpl->vlan)); 1231 if (vid != 0xfff) { 1232 ifp = VLAN_DEVAT(hw_ifp, vid); 1233 if (ifp == NULL) 1234 REJECT_PASS_ACCEPT(); 1235 } else 1236 ifp = hw_ifp; 1237 1238 /* 1239 * Don't offload if the peer requested a TCP option that's not known to 1240 * the silicon. 1241 */ 1242 if (cpl->tcpopt.unknown) 1243 REJECT_PASS_ACCEPT(); 1244 1245 if (inc.inc_flags & INC_ISIPV6) { 1246 1247 /* Don't offload if the ifcap isn't enabled */ 1248 if ((ifp->if_capenable & IFCAP_TOE6) == 0) 1249 REJECT_PASS_ACCEPT(); 1250 1251 /* 1252 * SYN must be directed to an IP6 address on this ifnet. This 1253 * is more restrictive than in6_localip. 1254 */ 1255 if (!in6_ifhasaddr(ifp, &inc.inc6_laddr)) 1256 REJECT_PASS_ACCEPT(); 1257 } else { 1258 1259 /* Don't offload if the ifcap isn't enabled */ 1260 if ((ifp->if_capenable & IFCAP_TOE4) == 0) 1261 REJECT_PASS_ACCEPT(); 1262 1263 /* 1264 * SYN must be directed to an IP address on this ifnet. This 1265 * is more restrictive than in_localip. 1266 */ 1267 if (!in_ifhasaddr(ifp, inc.inc_laddr)) 1268 REJECT_PASS_ACCEPT(); 1269 } 1270 1271 e = get_l2te_for_nexthop(pi, ifp, &inc); 1272 if (e == NULL) 1273 REJECT_PASS_ACCEPT(); 1274 1275 synqe = mbuf_to_synqe(m); 1276 if (synqe == NULL) 1277 REJECT_PASS_ACCEPT(); 1278 1279 wr = alloc_wrqe(is_t4(sc) ? sizeof(struct cpl_pass_accept_rpl) : 1280 sizeof(struct cpl_t5_pass_accept_rpl), &sc->sge.ctrlq[pi->port_id]); 1281 if (wr == NULL) 1282 REJECT_PASS_ACCEPT(); 1283 rpl = wrtod(wr); 1284 1285 INP_INFO_RLOCK(&V_tcbinfo); /* for 4-tuple check */ 1286 1287 /* Don't offload if the 4-tuple is already in use */ 1288 if (toe_4tuple_check(&inc, &th, ifp) != 0) { 1289 INP_INFO_RUNLOCK(&V_tcbinfo); 1290 free(wr, M_CXGBE); 1291 REJECT_PASS_ACCEPT(); 1292 } 1293 INP_INFO_RUNLOCK(&V_tcbinfo); 1294 1295 inp = lctx->inp; /* listening socket, not owned by TOE */ 1296 INP_WLOCK(inp); 1297 1298 /* Don't offload if the listening socket has closed */ 1299 if (__predict_false(inp->inp_flags & INP_DROPPED)) { 1300 /* 1301 * The listening socket has closed. The reply from the TOE to 1302 * our CPL_CLOSE_LISTSRV_REQ will ultimately release all 1303 * resources tied to this listen context. 1304 */ 1305 INP_WUNLOCK(inp); 1306 free(wr, M_CXGBE); 1307 REJECT_PASS_ACCEPT(); 1308 } 1309 so = inp->inp_socket; 1310 CURVNET_SET(so->so_vnet); 1311 1312 mtu_idx = find_best_mtu_idx(sc, &inc, be16toh(cpl->tcpopt.mss)); 1313 rscale = cpl->tcpopt.wsf && V_tcp_do_rfc1323 ? select_rcv_wscale() : 0; 1314 SOCKBUF_LOCK(&so->so_rcv); 1315 /* opt0 rcv_bufsiz initially, assumes its normal meaning later */ 1316 rx_credits = min(select_rcv_wnd(so) >> 10, M_RCV_BUFSIZ); 1317 SOCKBUF_UNLOCK(&so->so_rcv); 1318 1319 save_qids_in_mbuf(m, vi); 1320 get_qids_from_mbuf(m, NULL, &rxqid); 1321 1322 if (is_t4(sc)) 1323 INIT_TP_WR_MIT_CPL(rpl, CPL_PASS_ACCEPT_RPL, tid); 1324 else { 1325 struct cpl_t5_pass_accept_rpl *rpl5 = (void *)rpl; 1326 1327 INIT_TP_WR_MIT_CPL(rpl5, CPL_PASS_ACCEPT_RPL, tid); 1328 } 1329 if (sc->tt.ddp && (so->so_options & SO_NO_DDP) == 0) { 1330 ulp_mode = ULP_MODE_TCPDDP; 1331 synqe->flags |= TPF_SYNQE_TCPDDP; 1332 } else 1333 ulp_mode = ULP_MODE_NONE; 1334 rpl->opt0 = calc_opt0(so, vi, e, mtu_idx, rscale, rx_credits, ulp_mode); 1335 rpl->opt2 = calc_opt2p(sc, pi, rxqid, &cpl->tcpopt, &th, ulp_mode); 1336 1337 synqe->tid = tid; 1338 synqe->lctx = lctx; 1339 synqe->syn = m; 1340 m = NULL; 1341 refcount_init(&synqe->refcnt, 1); /* 1 means extra hold */ 1342 synqe->l2e_idx = e->idx; 1343 synqe->rcv_bufsize = rx_credits; 1344 atomic_store_rel_ptr(&synqe->wr, (uintptr_t)wr); 1345 1346 insert_tid(sc, tid, synqe); 1347 TAILQ_INSERT_TAIL(&lctx->synq, synqe, link); 1348 hold_synqe(synqe); /* hold for the duration it's in the synq */ 1349 hold_lctx(lctx); /* A synqe on the list has a ref on its lctx */ 1350 1351 /* 1352 * If all goes well t4_syncache_respond will get called during 1353 * syncache_add. Note that syncache_add releases the pcb lock. 1354 */ 1355 toe_syncache_add(&inc, &to, &th, inp, tod, synqe); 1356 INP_UNLOCK_ASSERT(inp); /* ok to assert, we have a ref on the inp */ 1357 CURVNET_RESTORE(); 1358 1359 /* 1360 * If we replied during syncache_add (synqe->wr has been consumed), 1361 * good. Otherwise, set it to 0 so that further syncache_respond 1362 * attempts by the kernel will be ignored. 1363 */ 1364 if (atomic_cmpset_ptr(&synqe->wr, (uintptr_t)wr, 0)) { 1365 1366 /* 1367 * syncache may or may not have a hold on the synqe, which may 1368 * or may not be stashed in the original SYN mbuf passed to us. 1369 * Just copy it over instead of dealing with all possibilities. 1370 */ 1371 m = m_dup(synqe->syn, M_NOWAIT); 1372 if (m) 1373 m->m_pkthdr.rcvif = hw_ifp; 1374 1375 remove_tid(sc, synqe->tid); 1376 free(wr, M_CXGBE); 1377 1378 /* Yank the synqe out of the lctx synq. */ 1379 INP_WLOCK(inp); 1380 TAILQ_REMOVE(&lctx->synq, synqe, link); 1381 release_synqe(synqe); /* removed from synq list */ 1382 inp = release_lctx(sc, lctx); 1383 if (inp) 1384 INP_WUNLOCK(inp); 1385 1386 release_synqe(synqe); /* extra hold */ 1387 REJECT_PASS_ACCEPT(); 1388 } 1389 1390 CTR5(KTR_CXGBE, "%s: stid %u, tid %u, lctx %p, synqe %p, SYNACK", 1391 __func__, stid, tid, lctx, synqe); 1392 1393 INP_WLOCK(inp); 1394 synqe->flags |= TPF_SYNQE_HAS_L2TE; 1395 if (__predict_false(inp->inp_flags & INP_DROPPED)) { 1396 /* 1397 * Listening socket closed but tod_listen_stop did not abort 1398 * this tid because there was no L2T entry for the tid at that 1399 * time. Abort it now. The reply to the abort will clean up. 1400 */ 1401 CTR6(KTR_CXGBE, 1402 "%s: stid %u, tid %u, lctx %p, synqe %p (0x%x), ABORT", 1403 __func__, stid, tid, lctx, synqe, synqe->flags); 1404 if (!(synqe->flags & TPF_SYNQE_EXPANDED)) 1405 send_reset_synqe(tod, synqe); 1406 INP_WUNLOCK(inp); 1407 1408 release_synqe(synqe); /* extra hold */ 1409 return (__LINE__); 1410 } 1411 INP_WUNLOCK(inp); 1412 1413 release_synqe(synqe); /* extra hold */ 1414 return (0); 1415reject: 1416 CTR4(KTR_CXGBE, "%s: stid %u, tid %u, REJECT (%d)", __func__, stid, tid, 1417 reject_reason); 1418 1419 if (e) 1420 t4_l2t_release(e); 1421 release_tid(sc, tid, lctx->ctrlq); 1422 1423 if (__predict_true(m != NULL)) { 1424 m_adj(m, sizeof(*cpl)); 1425 m->m_pkthdr.csum_flags |= (CSUM_IP_CHECKED | CSUM_IP_VALID | 1426 CSUM_DATA_VALID | CSUM_PSEUDO_HDR); 1427 m->m_pkthdr.csum_data = 0xffff; 1428 hw_ifp->if_input(hw_ifp, m); 1429 } 1430 1431 return (reject_reason); 1432} 1433 1434static void 1435synqe_to_protohdrs(struct adapter *sc, struct synq_entry *synqe, 1436 const struct cpl_pass_establish *cpl, struct in_conninfo *inc, 1437 struct tcphdr *th, struct tcpopt *to) 1438{ 1439 uint16_t tcp_opt = be16toh(cpl->tcp_opt); 1440 1441 /* start off with the original SYN */ 1442 pass_accept_req_to_protohdrs(sc, synqe->syn, inc, th); 1443 1444 /* modify parts to make it look like the ACK to our SYN|ACK */ 1445 th->th_flags = TH_ACK; 1446 th->th_ack = synqe->iss + 1; 1447 th->th_seq = be32toh(cpl->rcv_isn); 1448 bzero(to, sizeof(*to)); 1449 if (G_TCPOPT_TSTAMP(tcp_opt)) { 1450 to->to_flags |= TOF_TS; 1451 to->to_tsecr = synqe->ts; 1452 } 1453} 1454 1455static int 1456do_pass_establish(struct sge_iq *iq, const struct rss_header *rss, 1457 struct mbuf *m) 1458{ 1459 struct adapter *sc = iq->adapter; 1460 struct vi_info *vi; 1461 struct ifnet *ifp; 1462 const struct cpl_pass_establish *cpl = (const void *)(rss + 1); 1463#if defined(KTR) || defined(INVARIANTS) 1464 unsigned int stid = G_PASS_OPEN_TID(be32toh(cpl->tos_stid)); 1465#endif 1466 unsigned int tid = GET_TID(cpl); 1467 struct synq_entry *synqe = lookup_tid(sc, tid); 1468 struct listen_ctx *lctx = synqe->lctx; 1469 struct inpcb *inp = lctx->inp, *new_inp; 1470 struct socket *so; 1471 struct tcphdr th; 1472 struct tcpopt to; 1473 struct in_conninfo inc; 1474 struct toepcb *toep; 1475 u_int txqid, rxqid; 1476#ifdef INVARIANTS 1477 unsigned int opcode = G_CPL_OPCODE(be32toh(OPCODE_TID(cpl))); 1478#endif 1479 1480 KASSERT(opcode == CPL_PASS_ESTABLISH, 1481 ("%s: unexpected opcode 0x%x", __func__, opcode)); 1482 KASSERT(m == NULL, ("%s: wasn't expecting payload", __func__)); 1483 KASSERT(lctx->stid == stid, ("%s: lctx stid mismatch", __func__)); 1484 KASSERT(synqe->flags & TPF_SYNQE, 1485 ("%s: tid %u (ctx %p) not a synqe", __func__, tid, synqe)); 1486 1487 INP_INFO_RLOCK(&V_tcbinfo); /* for syncache_expand */ 1488 INP_WLOCK(inp); 1489 1490 CTR6(KTR_CXGBE, 1491 "%s: stid %u, tid %u, synqe %p (0x%x), inp_flags 0x%x", 1492 __func__, stid, tid, synqe, synqe->flags, inp->inp_flags); 1493 1494 if (__predict_false(inp->inp_flags & INP_DROPPED)) { 1495 1496 if (synqe->flags & TPF_SYNQE_HAS_L2TE) { 1497 KASSERT(synqe->flags & TPF_ABORT_SHUTDOWN, 1498 ("%s: listen socket closed but tid %u not aborted.", 1499 __func__, tid)); 1500 } 1501 1502 INP_WUNLOCK(inp); 1503 INP_INFO_RUNLOCK(&V_tcbinfo); 1504 return (0); 1505 } 1506 1507 ifp = synqe->syn->m_pkthdr.rcvif; 1508 vi = ifp->if_softc; 1509 KASSERT(vi->pi->adapter == sc, 1510 ("%s: vi %p, sc %p mismatch", __func__, vi, sc)); 1511 1512 get_qids_from_mbuf(synqe->syn, &txqid, &rxqid); 1513 KASSERT(rxqid == iq_to_ofld_rxq(iq) - &sc->sge.ofld_rxq[0], 1514 ("%s: CPL arrived on unexpected rxq. %d %d", __func__, rxqid, 1515 (int)(iq_to_ofld_rxq(iq) - &sc->sge.ofld_rxq[0]))); 1516 1517 toep = alloc_toepcb(vi, txqid, rxqid, M_NOWAIT); 1518 if (toep == NULL) { 1519reset: 1520 /* 1521 * The reply to this abort will perform final cleanup. There is 1522 * no need to check for HAS_L2TE here. We can be here only if 1523 * we responded to the PASS_ACCEPT_REQ, and our response had the 1524 * L2T idx. 1525 */ 1526 send_reset_synqe(TOEDEV(ifp), synqe); 1527 INP_WUNLOCK(inp); 1528 INP_INFO_RUNLOCK(&V_tcbinfo); 1529 return (0); 1530 } 1531 toep->tid = tid; 1532 toep->l2te = &sc->l2t->l2tab[synqe->l2e_idx]; 1533 if (synqe->flags & TPF_SYNQE_TCPDDP) 1534 set_tcpddp_ulp_mode(toep); 1535 else 1536 toep->ulp_mode = ULP_MODE_NONE; 1537 /* opt0 rcv_bufsiz initially, assumes its normal meaning later */ 1538 toep->rx_credits = synqe->rcv_bufsize; 1539 1540 so = inp->inp_socket; 1541 KASSERT(so != NULL, ("%s: socket is NULL", __func__)); 1542 1543 /* Come up with something that syncache_expand should be ok with. */ 1544 synqe_to_protohdrs(sc, synqe, cpl, &inc, &th, &to); 1545 1546 /* 1547 * No more need for anything in the mbuf that carried the 1548 * CPL_PASS_ACCEPT_REQ. Drop the CPL_PASS_ESTABLISH and toep pointer 1549 * there. XXX: bad form but I don't want to increase the size of synqe. 1550 */ 1551 m = synqe->syn; 1552 KASSERT(sizeof(*cpl) + sizeof(toep) <= m->m_len, 1553 ("%s: no room in mbuf %p (m_len %d)", __func__, m, m->m_len)); 1554 bcopy(cpl, mtod(m, void *), sizeof(*cpl)); 1555 *(struct toepcb **)(mtod(m, struct cpl_pass_establish *) + 1) = toep; 1556 1557 if (!toe_syncache_expand(&inc, &to, &th, &so) || so == NULL) { 1558 free_toepcb(toep); 1559 goto reset; 1560 } 1561 1562 /* New connection inpcb is already locked by syncache_expand(). */ 1563 new_inp = sotoinpcb(so); 1564 INP_WLOCK_ASSERT(new_inp); 1565 1566 /* 1567 * This is for the unlikely case where the syncache entry that we added 1568 * has been evicted from the syncache, but the syncache_expand above 1569 * works because of syncookies. 1570 * 1571 * XXX: we've held the tcbinfo lock throughout so there's no risk of 1572 * anyone accept'ing a connection before we've installed our hooks, but 1573 * this somewhat defeats the purpose of having a tod_offload_socket :-( 1574 */ 1575 if (__predict_false(!(synqe->flags & TPF_SYNQE_EXPANDED))) { 1576 tcp_timer_activate(intotcpcb(new_inp), TT_KEEP, 0); 1577 t4_offload_socket(TOEDEV(ifp), synqe, so); 1578 } 1579 1580 INP_WUNLOCK(new_inp); 1581 1582 /* Done with the synqe */ 1583 TAILQ_REMOVE(&lctx->synq, synqe, link); 1584 inp = release_lctx(sc, lctx); 1585 if (inp != NULL) 1586 INP_WUNLOCK(inp); 1587 INP_INFO_RUNLOCK(&V_tcbinfo); 1588 release_synqe(synqe); 1589 1590 return (0); 1591} 1592 1593void 1594t4_init_listen_cpl_handlers(void) 1595{ 1596 1597 t4_register_cpl_handler(CPL_PASS_OPEN_RPL, do_pass_open_rpl); 1598 t4_register_cpl_handler(CPL_CLOSE_LISTSRV_RPL, do_close_server_rpl); 1599 t4_register_cpl_handler(CPL_PASS_ACCEPT_REQ, do_pass_accept_req); 1600 t4_register_cpl_handler(CPL_PASS_ESTABLISH, do_pass_establish); 1601} 1602#endif 1603