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 #include <sys/param.h> 35 #include <sys/types.h> 36 #include <sys/systm.h> 37 #include <sys/kernel.h> 38 #include <sys/ktr.h> 39 #include <sys/lock.h> 40 #include <sys/limits.h> 41 #include <sys/module.h> 42 #include <sys/protosw.h> 43 #include <sys/domain.h> 44 #include <sys/refcount.h> 45 #include <sys/rmlock.h> 46 #include <sys/socket.h> 47 #include <sys/socketvar.h> 48 #include <sys/taskqueue.h> 49 #include <net/if.h> 50 #include <net/if_var.h> 51 #include <netinet/in.h> 52 #include <netinet/in_pcb.h> 53 #include <netinet/in_var.h> 54 #include <netinet/ip.h> 55 #include <netinet/ip6.h> 56 #include <netinet6/scope6_var.h> 57 #define TCPSTATES 58 #include <netinet/tcp_fsm.h> 59 #include <netinet/tcp_var.h> 60 #include <netinet/toecore.h> 61 62 #ifdef TCP_OFFLOAD 63 #include "common/common.h" 64 #include "common/t4_msg.h" 65 #include "common/t4_regs.h" 66 #include "common/t4_regs_values.h" 67 #include "common/t4_tcb.h" 68 #include "tom/t4_tom_l2t.h" 69 #include "tom/t4_tom.h" 70 71 static struct protosw toe_protosw; 72 static struct pr_usrreqs toe_usrreqs; 73 74 static struct protosw toe6_protosw; 75 static struct pr_usrreqs toe6_usrreqs; 76 77 /* Module ops */ 78 static int t4_tom_mod_load(void); 79 static int t4_tom_mod_unload(void); 80 static int t4_tom_modevent(module_t, int, void *); 81 82 /* ULD ops and helpers */ 83 static int t4_tom_activate(struct adapter *); 84 static int t4_tom_deactivate(struct adapter *); 85 86 static struct uld_info tom_uld_info = { 87 .uld_id = ULD_TOM, 88 .activate = t4_tom_activate, 89 .deactivate = t4_tom_deactivate, 90 }; 91 92 static void queue_tid_release(struct adapter *, int); 93 static void release_offload_resources(struct toepcb *); 94 static int alloc_tid_tabs(struct tid_info *); 95 static void free_tid_tabs(struct tid_info *); 96 static int add_lip(struct adapter *, struct in6_addr *); 97 static int delete_lip(struct adapter *, struct in6_addr *); 98 static struct clip_entry *search_lip(struct tom_data *, struct in6_addr *); 99 static void init_clip_table(struct adapter *, struct tom_data *); 100 static void update_clip(struct adapter *, void *); 101 static void t4_clip_task(void *, int); 102 static void update_clip_table(struct adapter *, struct tom_data *); 103 static void destroy_clip_table(struct adapter *, struct tom_data *); 104 static void free_tom_data(struct adapter *, struct tom_data *); 105 static void reclaim_wr_resources(void *, int); 106 107 static int in6_ifaddr_gen; 108 static eventhandler_tag ifaddr_evhandler; 109 static struct timeout_task clip_task; 110 111 struct toepcb * 112 alloc_toepcb(struct vi_info *vi, int txqid, int rxqid, int flags) 113 { 114 struct port_info *pi = vi->pi; 115 struct adapter *sc = pi->adapter; 116 struct toepcb *toep; 117 int tx_credits, txsd_total, len; 118 119 /* 120 * The firmware counts tx work request credits in units of 16 bytes 121 * each. Reserve room for an ABORT_REQ so the driver never has to worry 122 * about tx credits if it wants to abort a connection. 123 */ 124 tx_credits = sc->params.ofldq_wr_cred; 125 tx_credits -= howmany(sizeof(struct cpl_abort_req), 16); 126 127 /* 128 * Shortest possible tx work request is a fw_ofld_tx_data_wr + 1 byte 129 * immediate payload, and firmware counts tx work request credits in 130 * units of 16 byte. Calculate the maximum work requests possible. 131 */ 132 txsd_total = tx_credits / 133 howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16); 134 135 if (txqid < 0) 136 txqid = (arc4random() % vi->nofldtxq) + vi->first_ofld_txq; 137 KASSERT(txqid >= vi->first_ofld_txq && 138 txqid < vi->first_ofld_txq + vi->nofldtxq, 139 ("%s: txqid %d for vi %p (first %d, n %d)", __func__, txqid, vi, 140 vi->first_ofld_txq, vi->nofldtxq)); 141 142 if (rxqid < 0) 143 rxqid = (arc4random() % vi->nofldrxq) + vi->first_ofld_rxq; 144 KASSERT(rxqid >= vi->first_ofld_rxq && 145 rxqid < vi->first_ofld_rxq + vi->nofldrxq, 146 ("%s: rxqid %d for vi %p (first %d, n %d)", __func__, rxqid, vi, 147 vi->first_ofld_rxq, vi->nofldrxq)); 148 149 len = offsetof(struct toepcb, txsd) + 150 txsd_total * sizeof(struct ofld_tx_sdesc); 151 152 toep = malloc(len, M_CXGBE, M_ZERO | flags); 153 if (toep == NULL) 154 return (NULL); 155 156 refcount_init(&toep->refcount, 1); 157 toep->td = sc->tom_softc; 158 toep->vi = vi; 159 toep->tx_total = tx_credits; 160 toep->tx_credits = tx_credits; 161 toep->ofld_txq = &sc->sge.ofld_txq[txqid]; 162 toep->ofld_rxq = &sc->sge.ofld_rxq[rxqid]; 163 toep->ctrlq = &sc->sge.ctrlq[pi->port_id]; 164 mbufq_init(&toep->ulp_pduq, INT_MAX); 165 mbufq_init(&toep->ulp_pdu_reclaimq, INT_MAX); 166 toep->txsd_total = txsd_total; 167 toep->txsd_avail = txsd_total; 168 toep->txsd_pidx = 0; 169 toep->txsd_cidx = 0; 170 aiotx_init_toep(toep); 171 ddp_init_toep(toep); 172 173 return (toep); 174 } 175 176 struct toepcb * 177 hold_toepcb(struct toepcb *toep) 178 { 179 180 refcount_acquire(&toep->refcount); 181 return (toep); 182 } 183 184 void 185 free_toepcb(struct toepcb *toep) 186 { 187 188 if (refcount_release(&toep->refcount) == 0) 189 return; 190 191 KASSERT(!(toep->flags & TPF_ATTACHED), 192 ("%s: attached to an inpcb", __func__)); 193 KASSERT(!(toep->flags & TPF_CPL_PENDING), 194 ("%s: CPL pending", __func__)); 195 196 ddp_uninit_toep(toep); 197 free(toep, M_CXGBE); 198 } 199 200 /* 201 * Set up the socket for TCP offload. 202 */ 203 void 204 offload_socket(struct socket *so, struct toepcb *toep) 205 { 206 struct tom_data *td = toep->td; 207 struct inpcb *inp = sotoinpcb(so); 208 struct tcpcb *tp = intotcpcb(inp); 209 struct sockbuf *sb; 210 211 INP_WLOCK_ASSERT(inp); 212 213 /* Update socket */ 214 sb = &so->so_snd; 215 SOCKBUF_LOCK(sb); 216 sb->sb_flags |= SB_NOCOALESCE; 217 SOCKBUF_UNLOCK(sb); 218 sb = &so->so_rcv; 219 SOCKBUF_LOCK(sb); 220 sb->sb_flags |= SB_NOCOALESCE; 221 if (inp->inp_vflag & INP_IPV6) 222 so->so_proto = &toe6_protosw; 223 else 224 so->so_proto = &toe_protosw; 225 SOCKBUF_UNLOCK(sb); 226 227 /* Update TCP PCB */ 228 tp->tod = &td->tod; 229 tp->t_toe = toep; 230 tp->t_flags |= TF_TOE; 231 232 /* Install an extra hold on inp */ 233 toep->inp = inp; 234 toep->flags |= TPF_ATTACHED; 235 in_pcbref(inp); 236 237 /* Add the TOE PCB to the active list */ 238 mtx_lock(&td->toep_list_lock); 239 TAILQ_INSERT_HEAD(&td->toep_list, toep, link); 240 mtx_unlock(&td->toep_list_lock); 241 } 242 243 /* This is _not_ the normal way to "unoffload" a socket. */ 244 void 245 undo_offload_socket(struct socket *so) 246 { 247 struct inpcb *inp = sotoinpcb(so); 248 struct tcpcb *tp = intotcpcb(inp); 249 struct toepcb *toep = tp->t_toe; 250 struct tom_data *td = toep->td; 251 struct sockbuf *sb; 252 253 INP_WLOCK_ASSERT(inp); 254 255 sb = &so->so_snd; 256 SOCKBUF_LOCK(sb); 257 sb->sb_flags &= ~SB_NOCOALESCE; 258 SOCKBUF_UNLOCK(sb); 259 sb = &so->so_rcv; 260 SOCKBUF_LOCK(sb); 261 sb->sb_flags &= ~SB_NOCOALESCE; 262 SOCKBUF_UNLOCK(sb); 263 264 tp->tod = NULL; 265 tp->t_toe = NULL; 266 tp->t_flags &= ~TF_TOE; 267 268 toep->inp = NULL; 269 toep->flags &= ~TPF_ATTACHED; 270 if (in_pcbrele_wlocked(inp)) 271 panic("%s: inp freed.", __func__); 272 273 mtx_lock(&td->toep_list_lock); 274 TAILQ_REMOVE(&td->toep_list, toep, link); 275 mtx_unlock(&td->toep_list_lock); 276 277 free_toepcb(toep); 278 } 279 280 static void 281 release_offload_resources(struct toepcb *toep) 282 { 283 struct tom_data *td = toep->td; 284 struct adapter *sc = td_adapter(td); 285 int tid = toep->tid; 286 287 KASSERT(!(toep->flags & TPF_CPL_PENDING), 288 ("%s: %p has CPL pending.", __func__, toep)); 289 KASSERT(!(toep->flags & TPF_ATTACHED), 290 ("%s: %p is still attached.", __func__, toep)); 291 292 CTR5(KTR_CXGBE, "%s: toep %p (tid %d, l2te %p, ce %p)", 293 __func__, toep, tid, toep->l2te, toep->ce); 294 295 /* 296 * These queues should have been emptied at approximately the same time 297 * that a normal connection's socket's so_snd would have been purged or 298 * drained. Do _not_ clean up here. 299 */ 300 MPASS(mbufq_len(&toep->ulp_pduq) == 0); 301 MPASS(mbufq_len(&toep->ulp_pdu_reclaimq) == 0); 302 #ifdef INVARIANTS 303 ddp_assert_empty(toep); 304 #endif 305 306 if (toep->l2te) 307 t4_l2t_release(toep->l2te); 308 309 if (tid >= 0) { 310 remove_tid(sc, tid); 311 release_tid(sc, tid, toep->ctrlq); 312 } 313 314 if (toep->ce) 315 release_lip(td, toep->ce); 316 317 mtx_lock(&td->toep_list_lock); 318 TAILQ_REMOVE(&td->toep_list, toep, link); 319 mtx_unlock(&td->toep_list_lock); 320 321 free_toepcb(toep); 322 } 323 324 /* 325 * The kernel is done with the TCP PCB and this is our opportunity to unhook the 326 * toepcb hanging off of it. If the TOE driver is also done with the toepcb (no 327 * pending CPL) then it is time to release all resources tied to the toepcb. 328 * 329 * Also gets called when an offloaded active open fails and the TOM wants the 330 * kernel to take the TCP PCB back. 331 */ 332 static void 333 t4_pcb_detach(struct toedev *tod __unused, struct tcpcb *tp) 334 { 335 #if defined(KTR) || defined(INVARIANTS) 336 struct inpcb *inp = tp->t_inpcb; 337 #endif 338 struct toepcb *toep = tp->t_toe; 339 340 INP_WLOCK_ASSERT(inp); 341 342 KASSERT(toep != NULL, ("%s: toep is NULL", __func__)); 343 KASSERT(toep->flags & TPF_ATTACHED, 344 ("%s: not attached", __func__)); 345 346 #ifdef KTR 347 if (tp->t_state == TCPS_SYN_SENT) { 348 CTR6(KTR_CXGBE, "%s: atid %d, toep %p (0x%x), inp %p (0x%x)", 349 __func__, toep->tid, toep, toep->flags, inp, 350 inp->inp_flags); 351 } else { 352 CTR6(KTR_CXGBE, 353 "t4_pcb_detach: tid %d (%s), toep %p (0x%x), inp %p (0x%x)", 354 toep->tid, tcpstates[tp->t_state], toep, toep->flags, inp, 355 inp->inp_flags); 356 } 357 #endif 358 359 tp->t_toe = NULL; 360 tp->t_flags &= ~TF_TOE; 361 toep->flags &= ~TPF_ATTACHED; 362 363 if (!(toep->flags & TPF_CPL_PENDING)) 364 release_offload_resources(toep); 365 } 366 367 /* 368 * setsockopt handler. 369 */ 370 static void 371 t4_ctloutput(struct toedev *tod, struct tcpcb *tp, int dir, int name) 372 { 373 struct adapter *sc = tod->tod_softc; 374 struct toepcb *toep = tp->t_toe; 375 376 if (dir == SOPT_GET) 377 return; 378 379 CTR4(KTR_CXGBE, "%s: tp %p, dir %u, name %u", __func__, tp, dir, name); 380 381 switch (name) { 382 case TCP_NODELAY: 383 t4_set_tcb_field(sc, toep->ctrlq, toep->tid, W_TCB_T_FLAGS, 384 V_TF_NAGLE(1), V_TF_NAGLE(tp->t_flags & TF_NODELAY ? 0 : 1), 385 0, 0, toep->ofld_rxq->iq.abs_id); 386 break; 387 default: 388 break; 389 } 390 } 391 392 /* 393 * The TOE driver will not receive any more CPLs for the tid associated with the 394 * toepcb; release the hold on the inpcb. 395 */ 396 void 397 final_cpl_received(struct toepcb *toep) 398 { 399 struct inpcb *inp = toep->inp; 400 401 KASSERT(inp != NULL, ("%s: inp is NULL", __func__)); 402 INP_WLOCK_ASSERT(inp); 403 KASSERT(toep->flags & TPF_CPL_PENDING, 404 ("%s: CPL not pending already?", __func__)); 405 406 CTR6(KTR_CXGBE, "%s: tid %d, toep %p (0x%x), inp %p (0x%x)", 407 __func__, toep->tid, toep, toep->flags, inp, inp->inp_flags); 408 409 if (toep->ulp_mode == ULP_MODE_TCPDDP) 410 release_ddp_resources(toep); 411 toep->inp = NULL; 412 toep->flags &= ~TPF_CPL_PENDING; 413 mbufq_drain(&toep->ulp_pdu_reclaimq); 414 415 if (!(toep->flags & TPF_ATTACHED)) 416 release_offload_resources(toep); 417 418 if (!in_pcbrele_wlocked(inp)) 419 INP_WUNLOCK(inp); 420 } 421 422 void 423 insert_tid(struct adapter *sc, int tid, void *ctx) 424 { 425 struct tid_info *t = &sc->tids; 426 427 t->tid_tab[tid] = ctx; 428 atomic_add_int(&t->tids_in_use, 1); 429 } 430 431 void * 432 lookup_tid(struct adapter *sc, int tid) 433 { 434 struct tid_info *t = &sc->tids; 435 436 return (t->tid_tab[tid]); 437 } 438 439 void 440 update_tid(struct adapter *sc, int tid, void *ctx) 441 { 442 struct tid_info *t = &sc->tids; 443 444 t->tid_tab[tid] = ctx; 445 } 446 447 void 448 remove_tid(struct adapter *sc, int tid) 449 { 450 struct tid_info *t = &sc->tids; 451 452 t->tid_tab[tid] = NULL; 453 atomic_subtract_int(&t->tids_in_use, 1); 454 } 455 456 void 457 release_tid(struct adapter *sc, int tid, struct sge_wrq *ctrlq) 458 { 459 struct wrqe *wr; 460 struct cpl_tid_release *req; 461 462 wr = alloc_wrqe(sizeof(*req), ctrlq); 463 if (wr == NULL) { 464 queue_tid_release(sc, tid); /* defer */ 465 return; 466 } 467 req = wrtod(wr); 468 469 INIT_TP_WR_MIT_CPL(req, CPL_TID_RELEASE, tid); 470 471 t4_wrq_tx(sc, wr); 472 } 473 474 static void 475 queue_tid_release(struct adapter *sc, int tid) 476 { 477 478 CXGBE_UNIMPLEMENTED("deferred tid release"); 479 } 480 481 /* 482 * What mtu_idx to use, given a 4-tuple and/or an MSS cap 483 */ 484 int 485 find_best_mtu_idx(struct adapter *sc, struct in_conninfo *inc, int pmss) 486 { 487 unsigned short *mtus = &sc->params.mtus[0]; 488 int i, mss, n; 489 490 KASSERT(inc != NULL || pmss > 0, 491 ("%s: at least one of inc/pmss must be specified", __func__)); 492 493 mss = inc ? tcp_mssopt(inc) : pmss; 494 if (pmss > 0 && mss > pmss) 495 mss = pmss; 496 497 if (inc->inc_flags & INC_ISIPV6) 498 n = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 499 else 500 n = sizeof(struct ip) + sizeof(struct tcphdr); 501 502 for (i = 0; i < NMTUS - 1 && mtus[i + 1] <= mss + n; i++) 503 continue; 504 505 return (i); 506 } 507 508 /* 509 * Determine the receive window size for a socket. 510 */ 511 u_long 512 select_rcv_wnd(struct socket *so) 513 { 514 unsigned long wnd; 515 516 SOCKBUF_LOCK_ASSERT(&so->so_rcv); 517 518 wnd = sbspace(&so->so_rcv); 519 if (wnd < MIN_RCV_WND) 520 wnd = MIN_RCV_WND; 521 522 return min(wnd, MAX_RCV_WND); 523 } 524 525 int 526 select_rcv_wscale(void) 527 { 528 int wscale = 0; 529 unsigned long space = sb_max; 530 531 if (space > MAX_RCV_WND) 532 space = MAX_RCV_WND; 533 534 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < space) 535 wscale++; 536 537 return (wscale); 538 } 539 540 extern int always_keepalive; 541 #define VIID_SMACIDX(v) (((unsigned int)(v) & 0x7f) << 1) 542 543 /* 544 * socket so could be a listening socket too. 545 */ 546 uint64_t 547 calc_opt0(struct socket *so, struct vi_info *vi, struct l2t_entry *e, 548 int mtu_idx, int rscale, int rx_credits, int ulp_mode) 549 { 550 uint64_t opt0; 551 552 KASSERT(rx_credits <= M_RCV_BUFSIZ, 553 ("%s: rcv_bufsiz too high", __func__)); 554 555 opt0 = F_TCAM_BYPASS | V_WND_SCALE(rscale) | V_MSS_IDX(mtu_idx) | 556 V_ULP_MODE(ulp_mode) | V_RCV_BUFSIZ(rx_credits); 557 558 if (so != NULL) { 559 struct inpcb *inp = sotoinpcb(so); 560 struct tcpcb *tp = intotcpcb(inp); 561 int keepalive = always_keepalive || 562 so_options_get(so) & SO_KEEPALIVE; 563 564 opt0 |= V_NAGLE((tp->t_flags & TF_NODELAY) == 0); 565 opt0 |= V_KEEP_ALIVE(keepalive != 0); 566 } 567 568 if (e != NULL) 569 opt0 |= V_L2T_IDX(e->idx); 570 571 if (vi != NULL) { 572 opt0 |= V_SMAC_SEL(VIID_SMACIDX(vi->viid)); 573 opt0 |= V_TX_CHAN(vi->pi->tx_chan); 574 } 575 576 return htobe64(opt0); 577 } 578 579 uint64_t 580 select_ntuple(struct vi_info *vi, struct l2t_entry *e) 581 { 582 struct adapter *sc = vi->pi->adapter; 583 struct tp_params *tp = &sc->params.tp; 584 uint16_t viid = vi->viid; 585 uint64_t ntuple = 0; 586 587 /* 588 * Initialize each of the fields which we care about which are present 589 * in the Compressed Filter Tuple. 590 */ 591 if (tp->vlan_shift >= 0 && e->vlan != CPL_L2T_VLAN_NONE) 592 ntuple |= (uint64_t)(F_FT_VLAN_VLD | e->vlan) << tp->vlan_shift; 593 594 if (tp->port_shift >= 0) 595 ntuple |= (uint64_t)e->lport << tp->port_shift; 596 597 if (tp->protocol_shift >= 0) 598 ntuple |= (uint64_t)IPPROTO_TCP << tp->protocol_shift; 599 600 if (tp->vnic_shift >= 0) { 601 uint32_t vf = G_FW_VIID_VIN(viid); 602 uint32_t pf = G_FW_VIID_PFN(viid); 603 uint32_t vld = G_FW_VIID_VIVLD(viid); 604 605 ntuple |= (uint64_t)(V_FT_VNID_ID_VF(vf) | V_FT_VNID_ID_PF(pf) | 606 V_FT_VNID_ID_VLD(vld)) << tp->vnic_shift; 607 } 608 609 if (is_t4(sc)) 610 return (htobe32((uint32_t)ntuple)); 611 else 612 return (htobe64(V_FILTER_TUPLE(ntuple))); 613 } 614 615 void 616 set_tcpddp_ulp_mode(struct toepcb *toep) 617 { 618 619 toep->ulp_mode = ULP_MODE_TCPDDP; 620 toep->ddp_flags = DDP_OK; 621 } 622 623 int 624 negative_advice(int status) 625 { 626 627 return (status == CPL_ERR_RTX_NEG_ADVICE || 628 status == CPL_ERR_PERSIST_NEG_ADVICE || 629 status == CPL_ERR_KEEPALV_NEG_ADVICE); 630 } 631 632 static int 633 alloc_tid_tabs(struct tid_info *t) 634 { 635 size_t size; 636 unsigned int i; 637 638 size = t->ntids * sizeof(*t->tid_tab) + 639 t->natids * sizeof(*t->atid_tab) + 640 t->nstids * sizeof(*t->stid_tab); 641 642 t->tid_tab = malloc(size, M_CXGBE, M_ZERO | M_NOWAIT); 643 if (t->tid_tab == NULL) 644 return (ENOMEM); 645 646 mtx_init(&t->atid_lock, "atid lock", NULL, MTX_DEF); 647 t->atid_tab = (union aopen_entry *)&t->tid_tab[t->ntids]; 648 t->afree = t->atid_tab; 649 t->atids_in_use = 0; 650 for (i = 1; i < t->natids; i++) 651 t->atid_tab[i - 1].next = &t->atid_tab[i]; 652 t->atid_tab[t->natids - 1].next = NULL; 653 654 mtx_init(&t->stid_lock, "stid lock", NULL, MTX_DEF); 655 t->stid_tab = (struct listen_ctx **)&t->atid_tab[t->natids]; 656 t->stids_in_use = 0; 657 TAILQ_INIT(&t->stids); 658 t->nstids_free_head = t->nstids; 659 660 atomic_store_rel_int(&t->tids_in_use, 0); 661 662 return (0); 663 } 664 665 static void 666 free_tid_tabs(struct tid_info *t) 667 { 668 KASSERT(t->tids_in_use == 0, 669 ("%s: %d tids still in use.", __func__, t->tids_in_use)); 670 KASSERT(t->atids_in_use == 0, 671 ("%s: %d atids still in use.", __func__, t->atids_in_use)); 672 KASSERT(t->stids_in_use == 0, 673 ("%s: %d tids still in use.", __func__, t->stids_in_use)); 674 675 free(t->tid_tab, M_CXGBE); 676 t->tid_tab = NULL; 677 678 if (mtx_initialized(&t->atid_lock)) 679 mtx_destroy(&t->atid_lock); 680 if (mtx_initialized(&t->stid_lock)) 681 mtx_destroy(&t->stid_lock); 682 } 683 684 static int 685 add_lip(struct adapter *sc, struct in6_addr *lip) 686 { 687 struct fw_clip_cmd c; 688 689 ASSERT_SYNCHRONIZED_OP(sc); 690 /* mtx_assert(&td->clip_table_lock, MA_OWNED); */ 691 692 memset(&c, 0, sizeof(c)); 693 c.op_to_write = htonl(V_FW_CMD_OP(FW_CLIP_CMD) | F_FW_CMD_REQUEST | 694 F_FW_CMD_WRITE); 695 c.alloc_to_len16 = htonl(F_FW_CLIP_CMD_ALLOC | FW_LEN16(c)); 696 c.ip_hi = *(uint64_t *)&lip->s6_addr[0]; 697 c.ip_lo = *(uint64_t *)&lip->s6_addr[8]; 698 699 return (-t4_wr_mbox_ns(sc, sc->mbox, &c, sizeof(c), &c)); 700 } 701 702 static int 703 delete_lip(struct adapter *sc, struct in6_addr *lip) 704 { 705 struct fw_clip_cmd c; 706 707 ASSERT_SYNCHRONIZED_OP(sc); 708 /* mtx_assert(&td->clip_table_lock, MA_OWNED); */ 709 710 memset(&c, 0, sizeof(c)); 711 c.op_to_write = htonl(V_FW_CMD_OP(FW_CLIP_CMD) | F_FW_CMD_REQUEST | 712 F_FW_CMD_READ); 713 c.alloc_to_len16 = htonl(F_FW_CLIP_CMD_FREE | FW_LEN16(c)); 714 c.ip_hi = *(uint64_t *)&lip->s6_addr[0]; 715 c.ip_lo = *(uint64_t *)&lip->s6_addr[8]; 716 717 return (-t4_wr_mbox_ns(sc, sc->mbox, &c, sizeof(c), &c)); 718 } 719 720 static struct clip_entry * 721 search_lip(struct tom_data *td, struct in6_addr *lip) 722 { 723 struct clip_entry *ce; 724 725 mtx_assert(&td->clip_table_lock, MA_OWNED); 726 727 TAILQ_FOREACH(ce, &td->clip_table, link) { 728 if (IN6_ARE_ADDR_EQUAL(&ce->lip, lip)) 729 return (ce); 730 } 731 732 return (NULL); 733 } 734 735 struct clip_entry * 736 hold_lip(struct tom_data *td, struct in6_addr *lip) 737 { 738 struct clip_entry *ce; 739 740 mtx_lock(&td->clip_table_lock); 741 ce = search_lip(td, lip); 742 if (ce != NULL) 743 ce->refcount++; 744 mtx_unlock(&td->clip_table_lock); 745 746 return (ce); 747 } 748 749 void 750 release_lip(struct tom_data *td, struct clip_entry *ce) 751 { 752 753 mtx_lock(&td->clip_table_lock); 754 KASSERT(search_lip(td, &ce->lip) == ce, 755 ("%s: CLIP entry %p p not in CLIP table.", __func__, ce)); 756 KASSERT(ce->refcount > 0, 757 ("%s: CLIP entry %p has refcount 0", __func__, ce)); 758 --ce->refcount; 759 mtx_unlock(&td->clip_table_lock); 760 } 761 762 static void 763 init_clip_table(struct adapter *sc, struct tom_data *td) 764 { 765 766 ASSERT_SYNCHRONIZED_OP(sc); 767 768 mtx_init(&td->clip_table_lock, "CLIP table lock", NULL, MTX_DEF); 769 TAILQ_INIT(&td->clip_table); 770 td->clip_gen = -1; 771 772 update_clip_table(sc, td); 773 } 774 775 static void 776 update_clip(struct adapter *sc, void *arg __unused) 777 { 778 779 if (begin_synchronized_op(sc, NULL, HOLD_LOCK, "t4tomuc")) 780 return; 781 782 if (uld_active(sc, ULD_TOM)) 783 update_clip_table(sc, sc->tom_softc); 784 785 end_synchronized_op(sc, LOCK_HELD); 786 } 787 788 static void 789 t4_clip_task(void *arg, int count) 790 { 791 792 t4_iterate(update_clip, NULL); 793 } 794 795 static void 796 update_clip_table(struct adapter *sc, struct tom_data *td) 797 { 798 struct rm_priotracker in6_ifa_tracker; 799 struct in6_ifaddr *ia; 800 struct in6_addr *lip, tlip; 801 struct clip_head stale; 802 struct clip_entry *ce, *ce_temp; 803 int rc, gen = atomic_load_acq_int(&in6_ifaddr_gen); 804 805 ASSERT_SYNCHRONIZED_OP(sc); 806 807 IN6_IFADDR_RLOCK(&in6_ifa_tracker); 808 mtx_lock(&td->clip_table_lock); 809 810 if (gen == td->clip_gen) 811 goto done; 812 813 TAILQ_INIT(&stale); 814 TAILQ_CONCAT(&stale, &td->clip_table, link); 815 816 TAILQ_FOREACH(ia, &V_in6_ifaddrhead, ia_link) { 817 lip = &ia->ia_addr.sin6_addr; 818 819 KASSERT(!IN6_IS_ADDR_MULTICAST(lip), 820 ("%s: mcast address in in6_ifaddr list", __func__)); 821 822 if (IN6_IS_ADDR_LOOPBACK(lip)) 823 continue; 824 if (IN6_IS_SCOPE_EMBED(lip)) { 825 /* Remove the embedded scope */ 826 tlip = *lip; 827 lip = &tlip; 828 in6_clearscope(lip); 829 } 830 /* 831 * XXX: how to weed out the link local address for the loopback 832 * interface? It's fe80::1 usually (always?). 833 */ 834 835 /* 836 * If it's in the main list then we already know it's not stale. 837 */ 838 TAILQ_FOREACH(ce, &td->clip_table, link) { 839 if (IN6_ARE_ADDR_EQUAL(&ce->lip, lip)) 840 goto next; 841 } 842 843 /* 844 * If it's in the stale list we should move it to the main list. 845 */ 846 TAILQ_FOREACH(ce, &stale, link) { 847 if (IN6_ARE_ADDR_EQUAL(&ce->lip, lip)) { 848 TAILQ_REMOVE(&stale, ce, link); 849 TAILQ_INSERT_TAIL(&td->clip_table, ce, link); 850 goto next; 851 } 852 } 853 854 /* A new IP6 address; add it to the CLIP table */ 855 ce = malloc(sizeof(*ce), M_CXGBE, M_NOWAIT); 856 memcpy(&ce->lip, lip, sizeof(ce->lip)); 857 ce->refcount = 0; 858 rc = add_lip(sc, lip); 859 if (rc == 0) 860 TAILQ_INSERT_TAIL(&td->clip_table, ce, link); 861 else { 862 char ip[INET6_ADDRSTRLEN]; 863 864 inet_ntop(AF_INET6, &ce->lip, &ip[0], sizeof(ip)); 865 log(LOG_ERR, "%s: could not add %s (%d)\n", 866 __func__, ip, rc); 867 free(ce, M_CXGBE); 868 } 869 next: 870 continue; 871 } 872 873 /* 874 * Remove stale addresses (those no longer in V_in6_ifaddrhead) that are 875 * no longer referenced by the driver. 876 */ 877 TAILQ_FOREACH_SAFE(ce, &stale, link, ce_temp) { 878 if (ce->refcount == 0) { 879 rc = delete_lip(sc, &ce->lip); 880 if (rc == 0) { 881 TAILQ_REMOVE(&stale, ce, link); 882 free(ce, M_CXGBE); 883 } else { 884 char ip[INET6_ADDRSTRLEN]; 885 886 inet_ntop(AF_INET6, &ce->lip, &ip[0], 887 sizeof(ip)); 888 log(LOG_ERR, "%s: could not delete %s (%d)\n", 889 __func__, ip, rc); 890 } 891 } 892 } 893 /* The ones that are still referenced need to stay in the CLIP table */ 894 TAILQ_CONCAT(&td->clip_table, &stale, link); 895 896 td->clip_gen = gen; 897 done: 898 mtx_unlock(&td->clip_table_lock); 899 IN6_IFADDR_RUNLOCK(&in6_ifa_tracker); 900 } 901 902 static void 903 destroy_clip_table(struct adapter *sc, struct tom_data *td) 904 { 905 struct clip_entry *ce, *ce_temp; 906 907 if (mtx_initialized(&td->clip_table_lock)) { 908 mtx_lock(&td->clip_table_lock); 909 TAILQ_FOREACH_SAFE(ce, &td->clip_table, link, ce_temp) { 910 KASSERT(ce->refcount == 0, 911 ("%s: CLIP entry %p still in use (%d)", __func__, 912 ce, ce->refcount)); 913 TAILQ_REMOVE(&td->clip_table, ce, link); 914 delete_lip(sc, &ce->lip); 915 free(ce, M_CXGBE); 916 } 917 mtx_unlock(&td->clip_table_lock); 918 mtx_destroy(&td->clip_table_lock); 919 } 920 } 921 922 static void 923 free_tom_data(struct adapter *sc, struct tom_data *td) 924 { 925 926 ASSERT_SYNCHRONIZED_OP(sc); 927 928 KASSERT(TAILQ_EMPTY(&td->toep_list), 929 ("%s: TOE PCB list is not empty.", __func__)); 930 KASSERT(td->lctx_count == 0, 931 ("%s: lctx hash table is not empty.", __func__)); 932 933 t4_uninit_ddp(sc, td); 934 destroy_clip_table(sc, td); 935 936 if (td->listen_mask != 0) 937 hashdestroy(td->listen_hash, M_CXGBE, td->listen_mask); 938 939 if (mtx_initialized(&td->unsent_wr_lock)) 940 mtx_destroy(&td->unsent_wr_lock); 941 if (mtx_initialized(&td->lctx_hash_lock)) 942 mtx_destroy(&td->lctx_hash_lock); 943 if (mtx_initialized(&td->toep_list_lock)) 944 mtx_destroy(&td->toep_list_lock); 945 946 free_tid_tabs(&sc->tids); 947 free(td, M_CXGBE); 948 } 949 950 static void 951 reclaim_wr_resources(void *arg, int count) 952 { 953 struct tom_data *td = arg; 954 STAILQ_HEAD(, wrqe) twr_list = STAILQ_HEAD_INITIALIZER(twr_list); 955 struct cpl_act_open_req *cpl; 956 u_int opcode, atid; 957 struct wrqe *wr; 958 struct adapter *sc; 959 960 mtx_lock(&td->unsent_wr_lock); 961 STAILQ_SWAP(&td->unsent_wr_list, &twr_list, wrqe); 962 mtx_unlock(&td->unsent_wr_lock); 963 964 while ((wr = STAILQ_FIRST(&twr_list)) != NULL) { 965 STAILQ_REMOVE_HEAD(&twr_list, link); 966 967 cpl = wrtod(wr); 968 opcode = GET_OPCODE(cpl); 969 970 switch (opcode) { 971 case CPL_ACT_OPEN_REQ: 972 case CPL_ACT_OPEN_REQ6: 973 atid = G_TID_TID(be32toh(OPCODE_TID(cpl))); 974 sc = td_adapter(td); 975 976 CTR2(KTR_CXGBE, "%s: atid %u ", __func__, atid); 977 act_open_failure_cleanup(sc, atid, EHOSTUNREACH); 978 free(wr, M_CXGBE); 979 break; 980 default: 981 log(LOG_ERR, "%s: leaked work request %p, wr_len %d, " 982 "opcode %x\n", __func__, wr, wr->wr_len, opcode); 983 /* WR not freed here; go look at it with a debugger. */ 984 } 985 } 986 } 987 988 /* 989 * Ground control to Major TOM 990 * Commencing countdown, engines on 991 */ 992 static int 993 t4_tom_activate(struct adapter *sc) 994 { 995 struct tom_data *td; 996 struct toedev *tod; 997 struct vi_info *vi; 998 struct sge_ofld_rxq *ofld_rxq; 999 int i, j, rc, v; 1000 1001 ASSERT_SYNCHRONIZED_OP(sc); 1002 1003 /* per-adapter softc for TOM */ 1004 td = malloc(sizeof(*td), M_CXGBE, M_ZERO | M_NOWAIT); 1005 if (td == NULL) 1006 return (ENOMEM); 1007 1008 /* List of TOE PCBs and associated lock */ 1009 mtx_init(&td->toep_list_lock, "PCB list lock", NULL, MTX_DEF); 1010 TAILQ_INIT(&td->toep_list); 1011 1012 /* Listen context */ 1013 mtx_init(&td->lctx_hash_lock, "lctx hash lock", NULL, MTX_DEF); 1014 td->listen_hash = hashinit_flags(LISTEN_HASH_SIZE, M_CXGBE, 1015 &td->listen_mask, HASH_NOWAIT); 1016 1017 /* List of WRs for which L2 resolution failed */ 1018 mtx_init(&td->unsent_wr_lock, "Unsent WR list lock", NULL, MTX_DEF); 1019 STAILQ_INIT(&td->unsent_wr_list); 1020 TASK_INIT(&td->reclaim_wr_resources, 0, reclaim_wr_resources, td); 1021 1022 /* TID tables */ 1023 rc = alloc_tid_tabs(&sc->tids); 1024 if (rc != 0) 1025 goto done; 1026 1027 /* DDP page pods and CPL handlers */ 1028 t4_init_ddp(sc, td); 1029 1030 /* CLIP table for IPv6 offload */ 1031 init_clip_table(sc, td); 1032 1033 /* toedev ops */ 1034 tod = &td->tod; 1035 init_toedev(tod); 1036 tod->tod_softc = sc; 1037 tod->tod_connect = t4_connect; 1038 tod->tod_listen_start = t4_listen_start; 1039 tod->tod_listen_stop = t4_listen_stop; 1040 tod->tod_rcvd = t4_rcvd; 1041 tod->tod_output = t4_tod_output; 1042 tod->tod_send_rst = t4_send_rst; 1043 tod->tod_send_fin = t4_send_fin; 1044 tod->tod_pcb_detach = t4_pcb_detach; 1045 tod->tod_l2_update = t4_l2_update; 1046 tod->tod_syncache_added = t4_syncache_added; 1047 tod->tod_syncache_removed = t4_syncache_removed; 1048 tod->tod_syncache_respond = t4_syncache_respond; 1049 tod->tod_offload_socket = t4_offload_socket; 1050 tod->tod_ctloutput = t4_ctloutput; 1051 1052 for_each_port(sc, i) { 1053 for_each_vi(sc->port[i], v, vi) { 1054 TOEDEV(vi->ifp) = &td->tod; 1055 for_each_ofld_rxq(vi, j, ofld_rxq) { 1056 ofld_rxq->iq.set_tcb_rpl = do_set_tcb_rpl; 1057 ofld_rxq->iq.l2t_write_rpl = do_l2t_write_rpl2; 1058 } 1059 } 1060 } 1061 1062 sc->tom_softc = td; 1063 register_toedev(sc->tom_softc); 1064 1065 done: 1066 if (rc != 0) 1067 free_tom_data(sc, td); 1068 return (rc); 1069 } 1070 1071 static int 1072 t4_tom_deactivate(struct adapter *sc) 1073 { 1074 int rc = 0; 1075 struct tom_data *td = sc->tom_softc; 1076 1077 ASSERT_SYNCHRONIZED_OP(sc); 1078 1079 if (td == NULL) 1080 return (0); /* XXX. KASSERT? */ 1081 1082 if (sc->offload_map != 0) 1083 return (EBUSY); /* at least one port has IFCAP_TOE enabled */ 1084 1085 if (uld_active(sc, ULD_IWARP) || uld_active(sc, ULD_ISCSI)) 1086 return (EBUSY); /* both iWARP and iSCSI rely on the TOE. */ 1087 1088 mtx_lock(&td->toep_list_lock); 1089 if (!TAILQ_EMPTY(&td->toep_list)) 1090 rc = EBUSY; 1091 mtx_unlock(&td->toep_list_lock); 1092 1093 mtx_lock(&td->lctx_hash_lock); 1094 if (td->lctx_count > 0) 1095 rc = EBUSY; 1096 mtx_unlock(&td->lctx_hash_lock); 1097 1098 taskqueue_drain(taskqueue_thread, &td->reclaim_wr_resources); 1099 mtx_lock(&td->unsent_wr_lock); 1100 if (!STAILQ_EMPTY(&td->unsent_wr_list)) 1101 rc = EBUSY; 1102 mtx_unlock(&td->unsent_wr_lock); 1103 1104 if (rc == 0) { 1105 unregister_toedev(sc->tom_softc); 1106 free_tom_data(sc, td); 1107 sc->tom_softc = NULL; 1108 } 1109 1110 return (rc); 1111 } 1112 1113 static void 1114 t4_tom_ifaddr_event(void *arg __unused, struct ifnet *ifp) 1115 { 1116 1117 atomic_add_rel_int(&in6_ifaddr_gen, 1); 1118 taskqueue_enqueue_timeout(taskqueue_thread, &clip_task, -hz / 4); 1119 } 1120 1121 static int 1122 t4_aio_queue_tom(struct socket *so, struct kaiocb *job) 1123 { 1124 struct tcpcb *tp = so_sototcpcb(so); 1125 struct toepcb *toep = tp->t_toe; 1126 int error; 1127 1128 if (toep->ulp_mode == ULP_MODE_TCPDDP) { 1129 error = t4_aio_queue_ddp(so, job); 1130 if (error != EOPNOTSUPP) 1131 return (error); 1132 } 1133 1134 return (t4_aio_queue_aiotx(so, job)); 1135 } 1136 1137 static int 1138 t4_tom_mod_load(void) 1139 { 1140 int rc; 1141 struct protosw *tcp_protosw, *tcp6_protosw; 1142 1143 /* CPL handlers */ 1144 t4_init_connect_cpl_handlers(); 1145 t4_init_listen_cpl_handlers(); 1146 t4_init_cpl_io_handlers(); 1147 1148 rc = t4_ddp_mod_load(); 1149 if (rc != 0) 1150 return (rc); 1151 1152 tcp_protosw = pffindproto(PF_INET, IPPROTO_TCP, SOCK_STREAM); 1153 if (tcp_protosw == NULL) 1154 return (ENOPROTOOPT); 1155 bcopy(tcp_protosw, &toe_protosw, sizeof(toe_protosw)); 1156 bcopy(tcp_protosw->pr_usrreqs, &toe_usrreqs, sizeof(toe_usrreqs)); 1157 toe_usrreqs.pru_aio_queue = t4_aio_queue_tom; 1158 toe_protosw.pr_usrreqs = &toe_usrreqs; 1159 1160 tcp6_protosw = pffindproto(PF_INET6, IPPROTO_TCP, SOCK_STREAM); 1161 if (tcp6_protosw == NULL) 1162 return (ENOPROTOOPT); 1163 bcopy(tcp6_protosw, &toe6_protosw, sizeof(toe6_protosw)); 1164 bcopy(tcp6_protosw->pr_usrreqs, &toe6_usrreqs, sizeof(toe6_usrreqs)); 1165 toe6_usrreqs.pru_aio_queue = t4_aio_queue_tom; 1166 toe6_protosw.pr_usrreqs = &toe6_usrreqs; 1167 1168 TIMEOUT_TASK_INIT(taskqueue_thread, &clip_task, 0, t4_clip_task, NULL); 1169 ifaddr_evhandler = EVENTHANDLER_REGISTER(ifaddr_event, 1170 t4_tom_ifaddr_event, NULL, EVENTHANDLER_PRI_ANY); 1171 1172 rc = t4_register_uld(&tom_uld_info); 1173 if (rc != 0) 1174 t4_tom_mod_unload(); 1175 1176 return (rc); 1177 } 1178 1179 static void 1180 tom_uninit(struct adapter *sc, void *arg __unused) 1181 { 1182 if (begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4tomun")) 1183 return; 1184 1185 /* Try to free resources (works only if no port has IFCAP_TOE) */ 1186 if (uld_active(sc, ULD_TOM)) 1187 t4_deactivate_uld(sc, ULD_TOM); 1188 1189 end_synchronized_op(sc, 0); 1190 } 1191 1192 static int 1193 t4_tom_mod_unload(void) 1194 { 1195 t4_iterate(tom_uninit, NULL); 1196 1197 if (t4_unregister_uld(&tom_uld_info) == EBUSY) 1198 return (EBUSY); 1199 1200 if (ifaddr_evhandler) { 1201 EVENTHANDLER_DEREGISTER(ifaddr_event, ifaddr_evhandler); 1202 taskqueue_cancel_timeout(taskqueue_thread, &clip_task, NULL); 1203 } 1204 1205 t4_ddp_mod_unload(); 1206 1207 return (0); 1208 } 1209 #endif /* TCP_OFFLOAD */ 1210 1211 static int 1212 t4_tom_modevent(module_t mod, int cmd, void *arg) 1213 { 1214 int rc = 0; 1215 1216 #ifdef TCP_OFFLOAD 1217 switch (cmd) { 1218 case MOD_LOAD: 1219 rc = t4_tom_mod_load(); 1220 break; 1221 1222 case MOD_UNLOAD: 1223 rc = t4_tom_mod_unload(); 1224 break; 1225 1226 default: 1227 rc = EINVAL; 1228 } 1229 #else 1230 printf("t4_tom: compiled without TCP_OFFLOAD support.\n"); 1231 rc = EOPNOTSUPP; 1232 #endif 1233 return (rc); 1234 } 1235 1236 static moduledata_t t4_tom_moddata= { 1237 "t4_tom", 1238 t4_tom_modevent, 1239 0 1240 }; 1241 1242 MODULE_VERSION(t4_tom, 1); 1243 MODULE_DEPEND(t4_tom, toecore, 1, 1, 1); 1244 MODULE_DEPEND(t4_tom, t4nex, 1, 1, 1); 1245 DECLARE_MODULE(t4_tom, t4_tom_moddata, SI_SUB_EXEC, SI_ORDER_ANY); 1246