1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright (c) 2012 Chelsio Communications, Inc. 5 * All rights reserved. 6 * Written by: Navdeep Parhar <np@FreeBSD.org> 7 * 8 * Redistribution and use in source and binary forms, with or without 9 * modification, are permitted provided that the following conditions 10 * are met: 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 16 * 17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 20 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 27 * SUCH DAMAGE. 28 */ 29 30 #include <sys/cdefs.h> 31 __FBSDID("$FreeBSD$"); 32 33 #include "opt_inet.h" 34 #include "opt_inet6.h" 35 #include "opt_kern_tls.h" 36 #include "opt_ratelimit.h" 37 38 #include <sys/param.h> 39 #include <sys/types.h> 40 #include <sys/systm.h> 41 #include <sys/kernel.h> 42 #include <sys/ktr.h> 43 #include <sys/lock.h> 44 #include <sys/limits.h> 45 #include <sys/module.h> 46 #include <sys/protosw.h> 47 #include <sys/domain.h> 48 #include <sys/refcount.h> 49 #include <sys/rmlock.h> 50 #include <sys/socket.h> 51 #include <sys/socketvar.h> 52 #include <sys/sysctl.h> 53 #include <sys/taskqueue.h> 54 #include <net/if.h> 55 #include <net/if_var.h> 56 #include <net/if_types.h> 57 #include <net/if_vlan_var.h> 58 #include <netinet/in.h> 59 #include <netinet/in_pcb.h> 60 #include <netinet/in_var.h> 61 #include <netinet/ip.h> 62 #include <netinet/ip6.h> 63 #include <netinet6/scope6_var.h> 64 #define TCPSTATES 65 #include <netinet/tcp_fsm.h> 66 #include <netinet/tcp_timer.h> 67 #include <netinet/tcp_var.h> 68 #include <netinet/toecore.h> 69 #include <netinet/cc/cc.h> 70 71 #ifdef TCP_OFFLOAD 72 #include "common/common.h" 73 #include "common/t4_msg.h" 74 #include "common/t4_regs.h" 75 #include "common/t4_regs_values.h" 76 #include "common/t4_tcb.h" 77 #include "t4_clip.h" 78 #include "tom/t4_tom_l2t.h" 79 #include "tom/t4_tom.h" 80 #include "tom/t4_tls.h" 81 82 static struct protosw toe_protosw; 83 static struct pr_usrreqs toe_usrreqs; 84 85 static struct protosw toe6_protosw; 86 static struct pr_usrreqs toe6_usrreqs; 87 88 /* Module ops */ 89 static int t4_tom_mod_load(void); 90 static int t4_tom_mod_unload(void); 91 static int t4_tom_modevent(module_t, int, void *); 92 93 /* ULD ops and helpers */ 94 static int t4_tom_activate(struct adapter *); 95 static int t4_tom_deactivate(struct adapter *); 96 97 static struct uld_info tom_uld_info = { 98 .uld_id = ULD_TOM, 99 .activate = t4_tom_activate, 100 .deactivate = t4_tom_deactivate, 101 }; 102 103 static void release_offload_resources(struct toepcb *); 104 static int alloc_tid_tabs(struct tid_info *); 105 static void free_tid_tabs(struct tid_info *); 106 static void free_tom_data(struct adapter *, struct tom_data *); 107 static void reclaim_wr_resources(void *, int); 108 109 struct toepcb * 110 alloc_toepcb(struct vi_info *vi, int flags) 111 { 112 struct port_info *pi = vi->pi; 113 struct adapter *sc = pi->adapter; 114 struct toepcb *toep; 115 int tx_credits, txsd_total, len; 116 117 /* 118 * The firmware counts tx work request credits in units of 16 bytes 119 * each. Reserve room for an ABORT_REQ so the driver never has to worry 120 * about tx credits if it wants to abort a connection. 121 */ 122 tx_credits = sc->params.ofldq_wr_cred; 123 tx_credits -= howmany(sizeof(struct cpl_abort_req), 16); 124 125 /* 126 * Shortest possible tx work request is a fw_ofld_tx_data_wr + 1 byte 127 * immediate payload, and firmware counts tx work request credits in 128 * units of 16 byte. Calculate the maximum work requests possible. 129 */ 130 txsd_total = tx_credits / 131 howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16); 132 133 len = offsetof(struct toepcb, txsd) + 134 txsd_total * sizeof(struct ofld_tx_sdesc); 135 136 toep = malloc(len, M_CXGBE, M_ZERO | flags); 137 if (toep == NULL) 138 return (NULL); 139 140 refcount_init(&toep->refcount, 1); 141 toep->td = sc->tom_softc; 142 toep->vi = vi; 143 toep->tid = -1; 144 toep->tx_total = tx_credits; 145 toep->tx_credits = tx_credits; 146 mbufq_init(&toep->ulp_pduq, INT_MAX); 147 mbufq_init(&toep->ulp_pdu_reclaimq, INT_MAX); 148 toep->txsd_total = txsd_total; 149 toep->txsd_avail = txsd_total; 150 toep->txsd_pidx = 0; 151 toep->txsd_cidx = 0; 152 aiotx_init_toep(toep); 153 154 return (toep); 155 } 156 157 /* 158 * Initialize a toepcb after its params have been filled out. 159 */ 160 int 161 init_toepcb(struct vi_info *vi, struct toepcb *toep) 162 { 163 struct conn_params *cp = &toep->params; 164 struct port_info *pi = vi->pi; 165 struct adapter *sc = pi->adapter; 166 struct tx_cl_rl_params *tc; 167 168 if (cp->tc_idx >= 0 && cp->tc_idx < sc->chip_params->nsched_cls) { 169 tc = &pi->sched_params->cl_rl[cp->tc_idx]; 170 mtx_lock(&sc->tc_lock); 171 if (tc->flags & CLRL_ERR) { 172 log(LOG_ERR, 173 "%s: failed to associate traffic class %u with tid %u\n", 174 device_get_nameunit(vi->dev), cp->tc_idx, 175 toep->tid); 176 cp->tc_idx = -1; 177 } else { 178 tc->refcount++; 179 } 180 mtx_unlock(&sc->tc_lock); 181 } 182 toep->ofld_txq = &sc->sge.ofld_txq[cp->txq_idx]; 183 toep->ofld_rxq = &sc->sge.ofld_rxq[cp->rxq_idx]; 184 toep->ctrlq = &sc->sge.ctrlq[pi->port_id]; 185 186 tls_init_toep(toep); 187 if (ulp_mode(toep) == ULP_MODE_TCPDDP) 188 ddp_init_toep(toep); 189 190 toep->flags |= TPF_INITIALIZED; 191 192 return (0); 193 } 194 195 struct toepcb * 196 hold_toepcb(struct toepcb *toep) 197 { 198 199 refcount_acquire(&toep->refcount); 200 return (toep); 201 } 202 203 void 204 free_toepcb(struct toepcb *toep) 205 { 206 207 if (refcount_release(&toep->refcount) == 0) 208 return; 209 210 KASSERT(!(toep->flags & TPF_ATTACHED), 211 ("%s: attached to an inpcb", __func__)); 212 KASSERT(!(toep->flags & TPF_CPL_PENDING), 213 ("%s: CPL pending", __func__)); 214 215 if (toep->flags & TPF_INITIALIZED) { 216 if (ulp_mode(toep) == ULP_MODE_TCPDDP) 217 ddp_uninit_toep(toep); 218 tls_uninit_toep(toep); 219 } 220 free(toep, M_CXGBE); 221 } 222 223 /* 224 * Set up the socket for TCP offload. 225 */ 226 void 227 offload_socket(struct socket *so, struct toepcb *toep) 228 { 229 struct tom_data *td = toep->td; 230 struct inpcb *inp = sotoinpcb(so); 231 struct tcpcb *tp = intotcpcb(inp); 232 struct sockbuf *sb; 233 234 INP_WLOCK_ASSERT(inp); 235 236 /* Update socket */ 237 sb = &so->so_snd; 238 SOCKBUF_LOCK(sb); 239 sb->sb_flags |= SB_NOCOALESCE; 240 SOCKBUF_UNLOCK(sb); 241 sb = &so->so_rcv; 242 SOCKBUF_LOCK(sb); 243 sb->sb_flags |= SB_NOCOALESCE; 244 if (inp->inp_vflag & INP_IPV6) 245 so->so_proto = &toe6_protosw; 246 else 247 so->so_proto = &toe_protosw; 248 SOCKBUF_UNLOCK(sb); 249 250 /* Update TCP PCB */ 251 tp->tod = &td->tod; 252 tp->t_toe = toep; 253 tp->t_flags |= TF_TOE; 254 255 /* Install an extra hold on inp */ 256 toep->inp = inp; 257 toep->flags |= TPF_ATTACHED; 258 in_pcbref(inp); 259 260 /* Add the TOE PCB to the active list */ 261 mtx_lock(&td->toep_list_lock); 262 TAILQ_INSERT_HEAD(&td->toep_list, toep, link); 263 mtx_unlock(&td->toep_list_lock); 264 } 265 266 /* This is _not_ the normal way to "unoffload" a socket. */ 267 void 268 undo_offload_socket(struct socket *so) 269 { 270 struct inpcb *inp = sotoinpcb(so); 271 struct tcpcb *tp = intotcpcb(inp); 272 struct toepcb *toep = tp->t_toe; 273 struct tom_data *td = toep->td; 274 struct sockbuf *sb; 275 276 INP_WLOCK_ASSERT(inp); 277 278 sb = &so->so_snd; 279 SOCKBUF_LOCK(sb); 280 sb->sb_flags &= ~SB_NOCOALESCE; 281 SOCKBUF_UNLOCK(sb); 282 sb = &so->so_rcv; 283 SOCKBUF_LOCK(sb); 284 sb->sb_flags &= ~SB_NOCOALESCE; 285 SOCKBUF_UNLOCK(sb); 286 287 tp->tod = NULL; 288 tp->t_toe = NULL; 289 tp->t_flags &= ~TF_TOE; 290 291 toep->inp = NULL; 292 toep->flags &= ~TPF_ATTACHED; 293 if (in_pcbrele_wlocked(inp)) 294 panic("%s: inp freed.", __func__); 295 296 mtx_lock(&td->toep_list_lock); 297 TAILQ_REMOVE(&td->toep_list, toep, link); 298 mtx_unlock(&td->toep_list_lock); 299 } 300 301 static void 302 release_offload_resources(struct toepcb *toep) 303 { 304 struct tom_data *td = toep->td; 305 struct adapter *sc = td_adapter(td); 306 int tid = toep->tid; 307 308 KASSERT(!(toep->flags & TPF_CPL_PENDING), 309 ("%s: %p has CPL pending.", __func__, toep)); 310 KASSERT(!(toep->flags & TPF_ATTACHED), 311 ("%s: %p is still attached.", __func__, toep)); 312 313 CTR5(KTR_CXGBE, "%s: toep %p (tid %d, l2te %p, ce %p)", 314 __func__, toep, tid, toep->l2te, toep->ce); 315 316 /* 317 * These queues should have been emptied at approximately the same time 318 * that a normal connection's socket's so_snd would have been purged or 319 * drained. Do _not_ clean up here. 320 */ 321 MPASS(mbufq_len(&toep->ulp_pduq) == 0); 322 MPASS(mbufq_len(&toep->ulp_pdu_reclaimq) == 0); 323 #ifdef INVARIANTS 324 if (ulp_mode(toep) == ULP_MODE_TCPDDP) 325 ddp_assert_empty(toep); 326 #endif 327 MPASS(TAILQ_EMPTY(&toep->aiotx_jobq)); 328 329 if (toep->l2te) 330 t4_l2t_release(toep->l2te); 331 332 if (tid >= 0) { 333 remove_tid(sc, tid, toep->ce ? 2 : 1); 334 release_tid(sc, tid, toep->ctrlq); 335 } 336 337 if (toep->ce) 338 t4_release_lip(sc, toep->ce); 339 340 if (toep->params.tc_idx != -1) 341 t4_release_cl_rl(sc, toep->vi->pi->port_id, toep->params.tc_idx); 342 343 mtx_lock(&td->toep_list_lock); 344 TAILQ_REMOVE(&td->toep_list, toep, link); 345 mtx_unlock(&td->toep_list_lock); 346 347 free_toepcb(toep); 348 } 349 350 /* 351 * The kernel is done with the TCP PCB and this is our opportunity to unhook the 352 * toepcb hanging off of it. If the TOE driver is also done with the toepcb (no 353 * pending CPL) then it is time to release all resources tied to the toepcb. 354 * 355 * Also gets called when an offloaded active open fails and the TOM wants the 356 * kernel to take the TCP PCB back. 357 */ 358 static void 359 t4_pcb_detach(struct toedev *tod __unused, struct tcpcb *tp) 360 { 361 #if defined(KTR) || defined(INVARIANTS) 362 struct inpcb *inp = tp->t_inpcb; 363 #endif 364 struct toepcb *toep = tp->t_toe; 365 366 INP_WLOCK_ASSERT(inp); 367 368 KASSERT(toep != NULL, ("%s: toep is NULL", __func__)); 369 KASSERT(toep->flags & TPF_ATTACHED, 370 ("%s: not attached", __func__)); 371 372 #ifdef KTR 373 if (tp->t_state == TCPS_SYN_SENT) { 374 CTR6(KTR_CXGBE, "%s: atid %d, toep %p (0x%x), inp %p (0x%x)", 375 __func__, toep->tid, toep, toep->flags, inp, 376 inp->inp_flags); 377 } else { 378 CTR6(KTR_CXGBE, 379 "t4_pcb_detach: tid %d (%s), toep %p (0x%x), inp %p (0x%x)", 380 toep->tid, tcpstates[tp->t_state], toep, toep->flags, inp, 381 inp->inp_flags); 382 } 383 #endif 384 385 tp->t_toe = NULL; 386 tp->t_flags &= ~TF_TOE; 387 toep->flags &= ~TPF_ATTACHED; 388 389 if (!(toep->flags & TPF_CPL_PENDING)) 390 release_offload_resources(toep); 391 } 392 393 /* 394 * setsockopt handler. 395 */ 396 static void 397 t4_ctloutput(struct toedev *tod, struct tcpcb *tp, int dir, int name) 398 { 399 struct adapter *sc = tod->tod_softc; 400 struct toepcb *toep = tp->t_toe; 401 402 if (dir == SOPT_GET) 403 return; 404 405 CTR4(KTR_CXGBE, "%s: tp %p, dir %u, name %u", __func__, tp, dir, name); 406 407 switch (name) { 408 case TCP_NODELAY: 409 if (tp->t_state != TCPS_ESTABLISHED) 410 break; 411 toep->params.nagle = tp->t_flags & TF_NODELAY ? 0 : 1; 412 t4_set_tcb_field(sc, toep->ctrlq, toep, W_TCB_T_FLAGS, 413 V_TF_NAGLE(1), V_TF_NAGLE(toep->params.nagle), 0, 0); 414 break; 415 default: 416 break; 417 } 418 } 419 420 static inline uint64_t 421 get_tcb_tflags(const uint64_t *tcb) 422 { 423 424 return ((be64toh(tcb[14]) << 32) | (be64toh(tcb[15]) >> 32)); 425 } 426 427 static inline uint32_t 428 get_tcb_field(const uint64_t *tcb, u_int word, uint32_t mask, u_int shift) 429 { 430 #define LAST_WORD ((TCB_SIZE / 4) - 1) 431 uint64_t t1, t2; 432 int flit_idx; 433 434 MPASS(mask != 0); 435 MPASS(word <= LAST_WORD); 436 MPASS(shift < 32); 437 438 flit_idx = (LAST_WORD - word) / 2; 439 if (word & 0x1) 440 shift += 32; 441 t1 = be64toh(tcb[flit_idx]) >> shift; 442 t2 = 0; 443 if (fls(mask) > 64 - shift) { 444 /* 445 * Will spill over into the next logical flit, which is the flit 446 * before this one. The flit_idx before this one must be valid. 447 */ 448 MPASS(flit_idx > 0); 449 t2 = be64toh(tcb[flit_idx - 1]) << (64 - shift); 450 } 451 return ((t2 | t1) & mask); 452 #undef LAST_WORD 453 } 454 #define GET_TCB_FIELD(tcb, F) \ 455 get_tcb_field(tcb, W_TCB_##F, M_TCB_##F, S_TCB_##F) 456 457 /* 458 * Issues a CPL_GET_TCB to read the entire TCB for the tid. 459 */ 460 static int 461 send_get_tcb(struct adapter *sc, u_int tid) 462 { 463 struct cpl_get_tcb *cpl; 464 struct wrq_cookie cookie; 465 466 MPASS(tid < sc->tids.ntids); 467 468 cpl = start_wrq_wr(&sc->sge.ctrlq[0], howmany(sizeof(*cpl), 16), 469 &cookie); 470 if (__predict_false(cpl == NULL)) 471 return (ENOMEM); 472 bzero(cpl, sizeof(*cpl)); 473 INIT_TP_WR(cpl, tid); 474 OPCODE_TID(cpl) = htobe32(MK_OPCODE_TID(CPL_GET_TCB, tid)); 475 cpl->reply_ctrl = htobe16(V_REPLY_CHAN(0) | 476 V_QUEUENO(sc->sge.ofld_rxq[0].iq.cntxt_id)); 477 cpl->cookie = 0xff; 478 commit_wrq_wr(&sc->sge.ctrlq[0], cpl, &cookie); 479 480 return (0); 481 } 482 483 static struct tcb_histent * 484 alloc_tcb_histent(struct adapter *sc, u_int tid, int flags) 485 { 486 struct tcb_histent *te; 487 488 MPASS(flags == M_NOWAIT || flags == M_WAITOK); 489 490 te = malloc(sizeof(*te), M_CXGBE, M_ZERO | flags); 491 if (te == NULL) 492 return (NULL); 493 mtx_init(&te->te_lock, "TCB entry", NULL, MTX_DEF); 494 callout_init_mtx(&te->te_callout, &te->te_lock, 0); 495 te->te_adapter = sc; 496 te->te_tid = tid; 497 498 return (te); 499 } 500 501 static void 502 free_tcb_histent(struct tcb_histent *te) 503 { 504 505 mtx_destroy(&te->te_lock); 506 free(te, M_CXGBE); 507 } 508 509 /* 510 * Start tracking the tid in the TCB history. 511 */ 512 int 513 add_tid_to_history(struct adapter *sc, u_int tid) 514 { 515 struct tcb_histent *te = NULL; 516 struct tom_data *td = sc->tom_softc; 517 int rc; 518 519 MPASS(tid < sc->tids.ntids); 520 521 if (td->tcb_history == NULL) 522 return (ENXIO); 523 524 rw_wlock(&td->tcb_history_lock); 525 if (td->tcb_history[tid] != NULL) { 526 rc = EEXIST; 527 goto done; 528 } 529 te = alloc_tcb_histent(sc, tid, M_NOWAIT); 530 if (te == NULL) { 531 rc = ENOMEM; 532 goto done; 533 } 534 mtx_lock(&te->te_lock); 535 rc = send_get_tcb(sc, tid); 536 if (rc == 0) { 537 te->te_flags |= TE_RPL_PENDING; 538 td->tcb_history[tid] = te; 539 } else { 540 free(te, M_CXGBE); 541 } 542 mtx_unlock(&te->te_lock); 543 done: 544 rw_wunlock(&td->tcb_history_lock); 545 return (rc); 546 } 547 548 static void 549 remove_tcb_histent(struct tcb_histent *te) 550 { 551 struct adapter *sc = te->te_adapter; 552 struct tom_data *td = sc->tom_softc; 553 554 rw_assert(&td->tcb_history_lock, RA_WLOCKED); 555 mtx_assert(&te->te_lock, MA_OWNED); 556 MPASS(td->tcb_history[te->te_tid] == te); 557 558 td->tcb_history[te->te_tid] = NULL; 559 free_tcb_histent(te); 560 rw_wunlock(&td->tcb_history_lock); 561 } 562 563 static inline struct tcb_histent * 564 lookup_tcb_histent(struct adapter *sc, u_int tid, bool addrem) 565 { 566 struct tcb_histent *te; 567 struct tom_data *td = sc->tom_softc; 568 569 MPASS(tid < sc->tids.ntids); 570 571 if (td->tcb_history == NULL) 572 return (NULL); 573 574 if (addrem) 575 rw_wlock(&td->tcb_history_lock); 576 else 577 rw_rlock(&td->tcb_history_lock); 578 te = td->tcb_history[tid]; 579 if (te != NULL) { 580 mtx_lock(&te->te_lock); 581 return (te); /* with both locks held */ 582 } 583 if (addrem) 584 rw_wunlock(&td->tcb_history_lock); 585 else 586 rw_runlock(&td->tcb_history_lock); 587 588 return (te); 589 } 590 591 static inline void 592 release_tcb_histent(struct tcb_histent *te) 593 { 594 struct adapter *sc = te->te_adapter; 595 struct tom_data *td = sc->tom_softc; 596 597 mtx_assert(&te->te_lock, MA_OWNED); 598 mtx_unlock(&te->te_lock); 599 rw_assert(&td->tcb_history_lock, RA_RLOCKED); 600 rw_runlock(&td->tcb_history_lock); 601 } 602 603 static void 604 request_tcb(void *arg) 605 { 606 struct tcb_histent *te = arg; 607 608 mtx_assert(&te->te_lock, MA_OWNED); 609 610 /* Noone else is supposed to update the histent. */ 611 MPASS(!(te->te_flags & TE_RPL_PENDING)); 612 if (send_get_tcb(te->te_adapter, te->te_tid) == 0) 613 te->te_flags |= TE_RPL_PENDING; 614 else 615 callout_schedule(&te->te_callout, hz / 100); 616 } 617 618 static void 619 update_tcb_histent(struct tcb_histent *te, const uint64_t *tcb) 620 { 621 struct tom_data *td = te->te_adapter->tom_softc; 622 uint64_t tflags = get_tcb_tflags(tcb); 623 uint8_t sample = 0; 624 625 if (GET_TCB_FIELD(tcb, SND_MAX_RAW) != GET_TCB_FIELD(tcb, SND_UNA_RAW)) { 626 if (GET_TCB_FIELD(tcb, T_RXTSHIFT) != 0) 627 sample |= TS_RTO; 628 if (GET_TCB_FIELD(tcb, T_DUPACKS) != 0) 629 sample |= TS_DUPACKS; 630 if (GET_TCB_FIELD(tcb, T_DUPACKS) >= td->dupack_threshold) 631 sample |= TS_FASTREXMT; 632 } 633 634 if (GET_TCB_FIELD(tcb, SND_MAX_RAW) != 0) { 635 uint32_t snd_wnd; 636 637 sample |= TS_SND_BACKLOGGED; /* for whatever reason. */ 638 639 snd_wnd = GET_TCB_FIELD(tcb, RCV_ADV); 640 if (tflags & V_TF_RECV_SCALE(1)) 641 snd_wnd <<= GET_TCB_FIELD(tcb, RCV_SCALE); 642 if (GET_TCB_FIELD(tcb, SND_CWND) < snd_wnd) 643 sample |= TS_CWND_LIMITED; /* maybe due to CWND */ 644 } 645 646 if (tflags & V_TF_CCTRL_ECN(1)) { 647 648 /* 649 * CE marker on incoming IP hdr, echoing ECE back in the TCP 650 * hdr. Indicates congestion somewhere on the way from the peer 651 * to this node. 652 */ 653 if (tflags & V_TF_CCTRL_ECE(1)) 654 sample |= TS_ECN_ECE; 655 656 /* 657 * ECE seen and CWR sent (or about to be sent). Might indicate 658 * congestion on the way to the peer. This node is reducing its 659 * congestion window in response. 660 */ 661 if (tflags & (V_TF_CCTRL_CWR(1) | V_TF_CCTRL_RFR(1))) 662 sample |= TS_ECN_CWR; 663 } 664 665 te->te_sample[te->te_pidx] = sample; 666 if (++te->te_pidx == nitems(te->te_sample)) 667 te->te_pidx = 0; 668 memcpy(te->te_tcb, tcb, TCB_SIZE); 669 te->te_flags |= TE_ACTIVE; 670 } 671 672 static int 673 do_get_tcb_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) 674 { 675 struct adapter *sc = iq->adapter; 676 const struct cpl_get_tcb_rpl *cpl = mtod(m, const void *); 677 const uint64_t *tcb = (const uint64_t *)(const void *)(cpl + 1); 678 struct tcb_histent *te; 679 const u_int tid = GET_TID(cpl); 680 bool remove; 681 682 remove = GET_TCB_FIELD(tcb, T_STATE) == TCPS_CLOSED; 683 te = lookup_tcb_histent(sc, tid, remove); 684 if (te == NULL) { 685 /* Not in the history. Who issued the GET_TCB for this? */ 686 device_printf(sc->dev, "tcb %u: flags 0x%016jx, state %u, " 687 "srtt %u, sscale %u, rscale %u, cookie 0x%x\n", tid, 688 (uintmax_t)get_tcb_tflags(tcb), GET_TCB_FIELD(tcb, T_STATE), 689 GET_TCB_FIELD(tcb, T_SRTT), GET_TCB_FIELD(tcb, SND_SCALE), 690 GET_TCB_FIELD(tcb, RCV_SCALE), cpl->cookie); 691 goto done; 692 } 693 694 MPASS(te->te_flags & TE_RPL_PENDING); 695 te->te_flags &= ~TE_RPL_PENDING; 696 if (remove) { 697 remove_tcb_histent(te); 698 } else { 699 update_tcb_histent(te, tcb); 700 callout_reset(&te->te_callout, hz / 10, request_tcb, te); 701 release_tcb_histent(te); 702 } 703 done: 704 m_freem(m); 705 return (0); 706 } 707 708 static void 709 fill_tcp_info_from_tcb(struct adapter *sc, uint64_t *tcb, struct tcp_info *ti) 710 { 711 uint32_t v; 712 713 ti->tcpi_state = GET_TCB_FIELD(tcb, T_STATE); 714 715 v = GET_TCB_FIELD(tcb, T_SRTT); 716 ti->tcpi_rtt = tcp_ticks_to_us(sc, v); 717 718 v = GET_TCB_FIELD(tcb, T_RTTVAR); 719 ti->tcpi_rttvar = tcp_ticks_to_us(sc, v); 720 721 ti->tcpi_snd_ssthresh = GET_TCB_FIELD(tcb, SND_SSTHRESH); 722 ti->tcpi_snd_cwnd = GET_TCB_FIELD(tcb, SND_CWND); 723 ti->tcpi_rcv_nxt = GET_TCB_FIELD(tcb, RCV_NXT); 724 725 v = GET_TCB_FIELD(tcb, TX_MAX); 726 ti->tcpi_snd_nxt = v - GET_TCB_FIELD(tcb, SND_NXT_RAW); 727 728 /* Receive window being advertised by us. */ 729 ti->tcpi_rcv_wscale = GET_TCB_FIELD(tcb, SND_SCALE); /* Yes, SND. */ 730 ti->tcpi_rcv_space = GET_TCB_FIELD(tcb, RCV_WND); 731 732 /* Send window */ 733 ti->tcpi_snd_wscale = GET_TCB_FIELD(tcb, RCV_SCALE); /* Yes, RCV. */ 734 ti->tcpi_snd_wnd = GET_TCB_FIELD(tcb, RCV_ADV); 735 if (get_tcb_tflags(tcb) & V_TF_RECV_SCALE(1)) 736 ti->tcpi_snd_wnd <<= ti->tcpi_snd_wscale; 737 else 738 ti->tcpi_snd_wscale = 0; 739 740 } 741 742 static void 743 fill_tcp_info_from_history(struct adapter *sc, struct tcb_histent *te, 744 struct tcp_info *ti) 745 { 746 747 fill_tcp_info_from_tcb(sc, te->te_tcb, ti); 748 } 749 750 /* 751 * Reads the TCB for the given tid using a memory window and copies it to 'buf' 752 * in the same format as CPL_GET_TCB_RPL. 753 */ 754 static void 755 read_tcb_using_memwin(struct adapter *sc, u_int tid, uint64_t *buf) 756 { 757 int i, j, k, rc; 758 uint32_t addr; 759 u_char *tcb, tmp; 760 761 MPASS(tid < sc->tids.ntids); 762 763 addr = t4_read_reg(sc, A_TP_CMM_TCB_BASE) + tid * TCB_SIZE; 764 rc = read_via_memwin(sc, 2, addr, (uint32_t *)buf, TCB_SIZE); 765 if (rc != 0) 766 return; 767 768 tcb = (u_char *)buf; 769 for (i = 0, j = TCB_SIZE - 16; i < j; i += 16, j -= 16) { 770 for (k = 0; k < 16; k++) { 771 tmp = tcb[i + k]; 772 tcb[i + k] = tcb[j + k]; 773 tcb[j + k] = tmp; 774 } 775 } 776 } 777 778 static void 779 fill_tcp_info(struct adapter *sc, u_int tid, struct tcp_info *ti) 780 { 781 uint64_t tcb[TCB_SIZE / sizeof(uint64_t)]; 782 struct tcb_histent *te; 783 784 ti->tcpi_toe_tid = tid; 785 te = lookup_tcb_histent(sc, tid, false); 786 if (te != NULL) { 787 fill_tcp_info_from_history(sc, te, ti); 788 release_tcb_histent(te); 789 } else { 790 if (!(sc->debug_flags & DF_DISABLE_TCB_CACHE)) { 791 /* XXX: tell firmware to flush TCB cache. */ 792 } 793 read_tcb_using_memwin(sc, tid, tcb); 794 fill_tcp_info_from_tcb(sc, tcb, ti); 795 } 796 } 797 798 /* 799 * Called by the kernel to allow the TOE driver to "refine" values filled up in 800 * the tcp_info for an offloaded connection. 801 */ 802 static void 803 t4_tcp_info(struct toedev *tod, struct tcpcb *tp, struct tcp_info *ti) 804 { 805 struct adapter *sc = tod->tod_softc; 806 struct toepcb *toep = tp->t_toe; 807 808 INP_WLOCK_ASSERT(tp->t_inpcb); 809 MPASS(ti != NULL); 810 811 fill_tcp_info(sc, toep->tid, ti); 812 } 813 814 #ifdef KERN_TLS 815 static int 816 t4_alloc_tls_session(struct toedev *tod, struct tcpcb *tp, 817 struct ktls_session *tls, int direction) 818 { 819 struct toepcb *toep = tp->t_toe; 820 821 INP_WLOCK_ASSERT(tp->t_inpcb); 822 MPASS(tls != NULL); 823 824 return (tls_alloc_ktls(toep, tls, direction)); 825 } 826 #endif 827 828 /* 829 * The TOE driver will not receive any more CPLs for the tid associated with the 830 * toepcb; release the hold on the inpcb. 831 */ 832 void 833 final_cpl_received(struct toepcb *toep) 834 { 835 struct inpcb *inp = toep->inp; 836 837 KASSERT(inp != NULL, ("%s: inp is NULL", __func__)); 838 INP_WLOCK_ASSERT(inp); 839 KASSERT(toep->flags & TPF_CPL_PENDING, 840 ("%s: CPL not pending already?", __func__)); 841 842 CTR6(KTR_CXGBE, "%s: tid %d, toep %p (0x%x), inp %p (0x%x)", 843 __func__, toep->tid, toep, toep->flags, inp, inp->inp_flags); 844 845 if (ulp_mode(toep) == ULP_MODE_TCPDDP) 846 release_ddp_resources(toep); 847 toep->inp = NULL; 848 toep->flags &= ~TPF_CPL_PENDING; 849 mbufq_drain(&toep->ulp_pdu_reclaimq); 850 851 if (!(toep->flags & TPF_ATTACHED)) 852 release_offload_resources(toep); 853 854 if (!in_pcbrele_wlocked(inp)) 855 INP_WUNLOCK(inp); 856 } 857 858 void 859 insert_tid(struct adapter *sc, int tid, void *ctx, int ntids) 860 { 861 struct tid_info *t = &sc->tids; 862 863 MPASS(tid >= t->tid_base); 864 MPASS(tid - t->tid_base < t->ntids); 865 866 t->tid_tab[tid - t->tid_base] = ctx; 867 atomic_add_int(&t->tids_in_use, ntids); 868 } 869 870 void * 871 lookup_tid(struct adapter *sc, int tid) 872 { 873 struct tid_info *t = &sc->tids; 874 875 return (t->tid_tab[tid - t->tid_base]); 876 } 877 878 void 879 update_tid(struct adapter *sc, int tid, void *ctx) 880 { 881 struct tid_info *t = &sc->tids; 882 883 t->tid_tab[tid - t->tid_base] = ctx; 884 } 885 886 void 887 remove_tid(struct adapter *sc, int tid, int ntids) 888 { 889 struct tid_info *t = &sc->tids; 890 891 t->tid_tab[tid - t->tid_base] = NULL; 892 atomic_subtract_int(&t->tids_in_use, ntids); 893 } 894 895 /* 896 * What mtu_idx to use, given a 4-tuple. Note that both s->mss and tcp_mssopt 897 * have the MSS that we should advertise in our SYN. Advertised MSS doesn't 898 * account for any TCP options so the effective MSS (only payload, no headers or 899 * options) could be different. 900 */ 901 static int 902 find_best_mtu_idx(struct adapter *sc, struct in_conninfo *inc, 903 struct offload_settings *s) 904 { 905 unsigned short *mtus = &sc->params.mtus[0]; 906 int i, mss, mtu; 907 908 MPASS(inc != NULL); 909 910 mss = s->mss > 0 ? s->mss : tcp_mssopt(inc); 911 if (inc->inc_flags & INC_ISIPV6) 912 mtu = mss + sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 913 else 914 mtu = mss + sizeof(struct ip) + sizeof(struct tcphdr); 915 916 for (i = 0; i < NMTUS - 1 && mtus[i + 1] <= mtu; i++) 917 continue; 918 919 return (i); 920 } 921 922 /* 923 * Determine the receive window size for a socket. 924 */ 925 u_long 926 select_rcv_wnd(struct socket *so) 927 { 928 unsigned long wnd; 929 930 SOCKBUF_LOCK_ASSERT(&so->so_rcv); 931 932 wnd = sbspace(&so->so_rcv); 933 if (wnd < MIN_RCV_WND) 934 wnd = MIN_RCV_WND; 935 936 return min(wnd, MAX_RCV_WND); 937 } 938 939 int 940 select_rcv_wscale(void) 941 { 942 int wscale = 0; 943 unsigned long space = sb_max; 944 945 if (space > MAX_RCV_WND) 946 space = MAX_RCV_WND; 947 948 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < space) 949 wscale++; 950 951 return (wscale); 952 } 953 954 __be64 955 calc_options0(struct vi_info *vi, struct conn_params *cp) 956 { 957 uint64_t opt0 = 0; 958 959 opt0 |= F_TCAM_BYPASS; 960 961 MPASS(cp->wscale >= 0 && cp->wscale <= M_WND_SCALE); 962 opt0 |= V_WND_SCALE(cp->wscale); 963 964 MPASS(cp->mtu_idx >= 0 && cp->mtu_idx < NMTUS); 965 opt0 |= V_MSS_IDX(cp->mtu_idx); 966 967 MPASS(cp->ulp_mode >= 0 && cp->ulp_mode <= M_ULP_MODE); 968 opt0 |= V_ULP_MODE(cp->ulp_mode); 969 970 MPASS(cp->opt0_bufsize >= 0 && cp->opt0_bufsize <= M_RCV_BUFSIZ); 971 opt0 |= V_RCV_BUFSIZ(cp->opt0_bufsize); 972 973 MPASS(cp->l2t_idx >= 0 && cp->l2t_idx < vi->pi->adapter->vres.l2t.size); 974 opt0 |= V_L2T_IDX(cp->l2t_idx); 975 976 opt0 |= V_SMAC_SEL(vi->smt_idx); 977 opt0 |= V_TX_CHAN(vi->pi->tx_chan); 978 979 MPASS(cp->keepalive == 0 || cp->keepalive == 1); 980 opt0 |= V_KEEP_ALIVE(cp->keepalive); 981 982 MPASS(cp->nagle == 0 || cp->nagle == 1); 983 opt0 |= V_NAGLE(cp->nagle); 984 985 return (htobe64(opt0)); 986 } 987 988 __be32 989 calc_options2(struct vi_info *vi, struct conn_params *cp) 990 { 991 uint32_t opt2 = 0; 992 struct port_info *pi = vi->pi; 993 struct adapter *sc = pi->adapter; 994 995 /* 996 * rx flow control, rx coalesce, congestion control, and tx pace are all 997 * explicitly set by the driver. On T5+ the ISS is also set by the 998 * driver to the value picked by the kernel. 999 */ 1000 if (is_t4(sc)) { 1001 opt2 |= F_RX_FC_VALID | F_RX_COALESCE_VALID; 1002 opt2 |= F_CONG_CNTRL_VALID | F_PACE_VALID; 1003 } else { 1004 opt2 |= F_T5_OPT_2_VALID; /* all 4 valid */ 1005 opt2 |= F_T5_ISS; /* ISS provided in CPL */ 1006 } 1007 1008 MPASS(cp->sack == 0 || cp->sack == 1); 1009 opt2 |= V_SACK_EN(cp->sack); 1010 1011 MPASS(cp->tstamp == 0 || cp->tstamp == 1); 1012 opt2 |= V_TSTAMPS_EN(cp->tstamp); 1013 1014 if (cp->wscale > 0) 1015 opt2 |= F_WND_SCALE_EN; 1016 1017 MPASS(cp->ecn == 0 || cp->ecn == 1); 1018 opt2 |= V_CCTRL_ECN(cp->ecn); 1019 1020 /* XXX: F_RX_CHANNEL for multiple rx c-chan support goes here. */ 1021 1022 opt2 |= V_TX_QUEUE(sc->params.tp.tx_modq[pi->tx_chan]); 1023 opt2 |= V_PACE(0); 1024 opt2 |= F_RSS_QUEUE_VALID; 1025 opt2 |= V_RSS_QUEUE(sc->sge.ofld_rxq[cp->rxq_idx].iq.abs_id); 1026 1027 MPASS(cp->cong_algo >= 0 && cp->cong_algo <= M_CONG_CNTRL); 1028 opt2 |= V_CONG_CNTRL(cp->cong_algo); 1029 1030 MPASS(cp->rx_coalesce == 0 || cp->rx_coalesce == 1); 1031 if (cp->rx_coalesce == 1) 1032 opt2 |= V_RX_COALESCE(M_RX_COALESCE); 1033 1034 opt2 |= V_RX_FC_DDP(0) | V_RX_FC_DISABLE(0); 1035 #ifdef USE_DDP_RX_FLOW_CONTROL 1036 if (cp->ulp_mode == ULP_MODE_TCPDDP) 1037 opt2 |= F_RX_FC_DDP; 1038 #endif 1039 if (cp->ulp_mode == ULP_MODE_TLS) 1040 opt2 |= F_RX_FC_DISABLE; 1041 1042 return (htobe32(opt2)); 1043 } 1044 1045 uint64_t 1046 select_ntuple(struct vi_info *vi, struct l2t_entry *e) 1047 { 1048 struct adapter *sc = vi->pi->adapter; 1049 struct tp_params *tp = &sc->params.tp; 1050 uint64_t ntuple = 0; 1051 1052 /* 1053 * Initialize each of the fields which we care about which are present 1054 * in the Compressed Filter Tuple. 1055 */ 1056 if (tp->vlan_shift >= 0 && EVL_VLANOFTAG(e->vlan) != CPL_L2T_VLAN_NONE) 1057 ntuple |= (uint64_t)(F_FT_VLAN_VLD | e->vlan) << tp->vlan_shift; 1058 1059 if (tp->port_shift >= 0) 1060 ntuple |= (uint64_t)e->lport << tp->port_shift; 1061 1062 if (tp->protocol_shift >= 0) 1063 ntuple |= (uint64_t)IPPROTO_TCP << tp->protocol_shift; 1064 1065 if (tp->vnic_shift >= 0 && tp->ingress_config & F_VNIC) { 1066 ntuple |= (uint64_t)(V_FT_VNID_ID_VF(vi->vin) | 1067 V_FT_VNID_ID_PF(sc->pf) | V_FT_VNID_ID_VLD(vi->vfvld)) << 1068 tp->vnic_shift; 1069 } 1070 1071 if (is_t4(sc)) 1072 return (htobe32((uint32_t)ntuple)); 1073 else 1074 return (htobe64(V_FILTER_TUPLE(ntuple))); 1075 } 1076 1077 static int 1078 is_tls_sock(struct socket *so, struct adapter *sc) 1079 { 1080 struct inpcb *inp = sotoinpcb(so); 1081 int i, rc; 1082 1083 /* XXX: Eventually add a SO_WANT_TLS socket option perhaps? */ 1084 rc = 0; 1085 ADAPTER_LOCK(sc); 1086 for (i = 0; i < sc->tt.num_tls_rx_ports; i++) { 1087 if (inp->inp_lport == htons(sc->tt.tls_rx_ports[i]) || 1088 inp->inp_fport == htons(sc->tt.tls_rx_ports[i])) { 1089 rc = 1; 1090 break; 1091 } 1092 } 1093 ADAPTER_UNLOCK(sc); 1094 return (rc); 1095 } 1096 1097 /* 1098 * Initialize various connection parameters. 1099 */ 1100 void 1101 init_conn_params(struct vi_info *vi , struct offload_settings *s, 1102 struct in_conninfo *inc, struct socket *so, 1103 const struct tcp_options *tcpopt, int16_t l2t_idx, struct conn_params *cp) 1104 { 1105 struct port_info *pi = vi->pi; 1106 struct adapter *sc = pi->adapter; 1107 struct tom_tunables *tt = &sc->tt; 1108 struct inpcb *inp = sotoinpcb(so); 1109 struct tcpcb *tp = intotcpcb(inp); 1110 u_long wnd; 1111 1112 MPASS(s->offload != 0); 1113 1114 /* Congestion control algorithm */ 1115 if (s->cong_algo >= 0) 1116 cp->cong_algo = s->cong_algo & M_CONG_CNTRL; 1117 else if (sc->tt.cong_algorithm >= 0) 1118 cp->cong_algo = tt->cong_algorithm & M_CONG_CNTRL; 1119 else { 1120 struct cc_algo *cc = CC_ALGO(tp); 1121 1122 if (strcasecmp(cc->name, "reno") == 0) 1123 cp->cong_algo = CONG_ALG_RENO; 1124 else if (strcasecmp(cc->name, "tahoe") == 0) 1125 cp->cong_algo = CONG_ALG_TAHOE; 1126 if (strcasecmp(cc->name, "newreno") == 0) 1127 cp->cong_algo = CONG_ALG_NEWRENO; 1128 if (strcasecmp(cc->name, "highspeed") == 0) 1129 cp->cong_algo = CONG_ALG_HIGHSPEED; 1130 else { 1131 /* 1132 * Use newreno in case the algorithm selected by the 1133 * host stack is not supported by the hardware. 1134 */ 1135 cp->cong_algo = CONG_ALG_NEWRENO; 1136 } 1137 } 1138 1139 /* Tx traffic scheduling class. */ 1140 if (s->sched_class >= 0 && 1141 s->sched_class < sc->chip_params->nsched_cls) { 1142 cp->tc_idx = s->sched_class; 1143 } else 1144 cp->tc_idx = -1; 1145 1146 /* Nagle's algorithm. */ 1147 if (s->nagle >= 0) 1148 cp->nagle = s->nagle > 0 ? 1 : 0; 1149 else 1150 cp->nagle = tp->t_flags & TF_NODELAY ? 0 : 1; 1151 1152 /* TCP Keepalive. */ 1153 if (V_tcp_always_keepalive || so_options_get(so) & SO_KEEPALIVE) 1154 cp->keepalive = 1; 1155 else 1156 cp->keepalive = 0; 1157 1158 /* Optimization that's specific to T5 @ 40G. */ 1159 if (tt->tx_align >= 0) 1160 cp->tx_align = tt->tx_align > 0 ? 1 : 0; 1161 else if (chip_id(sc) == CHELSIO_T5 && 1162 (port_top_speed(pi) > 10 || sc->params.nports > 2)) 1163 cp->tx_align = 1; 1164 else 1165 cp->tx_align = 0; 1166 1167 /* ULP mode. */ 1168 if (can_tls_offload(sc) && 1169 (s->tls > 0 || (s->tls < 0 && is_tls_sock(so, sc)))) 1170 cp->ulp_mode = ULP_MODE_TLS; 1171 else if (s->ddp > 0 || 1172 (s->ddp < 0 && sc->tt.ddp && (so_options_get(so) & SO_NO_DDP) == 0)) 1173 cp->ulp_mode = ULP_MODE_TCPDDP; 1174 else 1175 cp->ulp_mode = ULP_MODE_NONE; 1176 1177 /* Rx coalescing. */ 1178 if (s->rx_coalesce >= 0) 1179 cp->rx_coalesce = s->rx_coalesce > 0 ? 1 : 0; 1180 else if (cp->ulp_mode == ULP_MODE_TLS) 1181 cp->rx_coalesce = 0; 1182 else if (tt->rx_coalesce >= 0) 1183 cp->rx_coalesce = tt->rx_coalesce > 0 ? 1 : 0; 1184 else 1185 cp->rx_coalesce = 1; /* default */ 1186 1187 /* 1188 * Index in the PMTU table. This controls the MSS that we announce in 1189 * our SYN initially, but after ESTABLISHED it controls the MSS that we 1190 * use to send data. 1191 */ 1192 cp->mtu_idx = find_best_mtu_idx(sc, inc, s); 1193 1194 /* Tx queue for this connection. */ 1195 if (s->txq >= 0 && s->txq < vi->nofldtxq) 1196 cp->txq_idx = s->txq; 1197 else 1198 cp->txq_idx = arc4random() % vi->nofldtxq; 1199 cp->txq_idx += vi->first_ofld_txq; 1200 1201 /* Rx queue for this connection. */ 1202 if (s->rxq >= 0 && s->rxq < vi->nofldrxq) 1203 cp->rxq_idx = s->rxq; 1204 else 1205 cp->rxq_idx = arc4random() % vi->nofldrxq; 1206 cp->rxq_idx += vi->first_ofld_rxq; 1207 1208 if (SOLISTENING(so)) { 1209 /* Passive open */ 1210 MPASS(tcpopt != NULL); 1211 1212 /* TCP timestamp option */ 1213 if (tcpopt->tstamp && 1214 (s->tstamp > 0 || (s->tstamp < 0 && V_tcp_do_rfc1323))) 1215 cp->tstamp = 1; 1216 else 1217 cp->tstamp = 0; 1218 1219 /* SACK */ 1220 if (tcpopt->sack && 1221 (s->sack > 0 || (s->sack < 0 && V_tcp_do_sack))) 1222 cp->sack = 1; 1223 else 1224 cp->sack = 0; 1225 1226 /* Receive window scaling. */ 1227 if (tcpopt->wsf > 0 && tcpopt->wsf < 15 && V_tcp_do_rfc1323) 1228 cp->wscale = select_rcv_wscale(); 1229 else 1230 cp->wscale = 0; 1231 1232 /* ECN */ 1233 if (tcpopt->ecn && /* XXX: review. */ 1234 (s->ecn > 0 || (s->ecn < 0 && V_tcp_do_ecn))) 1235 cp->ecn = 1; 1236 else 1237 cp->ecn = 0; 1238 1239 wnd = max(so->sol_sbrcv_hiwat, MIN_RCV_WND); 1240 cp->opt0_bufsize = min(wnd >> 10, M_RCV_BUFSIZ); 1241 1242 if (tt->sndbuf > 0) 1243 cp->sndbuf = tt->sndbuf; 1244 else if (so->sol_sbsnd_flags & SB_AUTOSIZE && 1245 V_tcp_do_autosndbuf) 1246 cp->sndbuf = 256 * 1024; 1247 else 1248 cp->sndbuf = so->sol_sbsnd_hiwat; 1249 } else { 1250 /* Active open */ 1251 1252 /* TCP timestamp option */ 1253 if (s->tstamp > 0 || 1254 (s->tstamp < 0 && (tp->t_flags & TF_REQ_TSTMP))) 1255 cp->tstamp = 1; 1256 else 1257 cp->tstamp = 0; 1258 1259 /* SACK */ 1260 if (s->sack > 0 || 1261 (s->sack < 0 && (tp->t_flags & TF_SACK_PERMIT))) 1262 cp->sack = 1; 1263 else 1264 cp->sack = 0; 1265 1266 /* Receive window scaling */ 1267 if (tp->t_flags & TF_REQ_SCALE) 1268 cp->wscale = select_rcv_wscale(); 1269 else 1270 cp->wscale = 0; 1271 1272 /* ECN */ 1273 if (s->ecn > 0 || (s->ecn < 0 && V_tcp_do_ecn == 1)) 1274 cp->ecn = 1; 1275 else 1276 cp->ecn = 0; 1277 1278 SOCKBUF_LOCK(&so->so_rcv); 1279 wnd = max(select_rcv_wnd(so), MIN_RCV_WND); 1280 SOCKBUF_UNLOCK(&so->so_rcv); 1281 cp->opt0_bufsize = min(wnd >> 10, M_RCV_BUFSIZ); 1282 1283 if (tt->sndbuf > 0) 1284 cp->sndbuf = tt->sndbuf; 1285 else { 1286 SOCKBUF_LOCK(&so->so_snd); 1287 if (so->so_snd.sb_flags & SB_AUTOSIZE && 1288 V_tcp_do_autosndbuf) 1289 cp->sndbuf = 256 * 1024; 1290 else 1291 cp->sndbuf = so->so_snd.sb_hiwat; 1292 SOCKBUF_UNLOCK(&so->so_snd); 1293 } 1294 } 1295 1296 cp->l2t_idx = l2t_idx; 1297 1298 /* This will be initialized on ESTABLISHED. */ 1299 cp->emss = 0; 1300 } 1301 1302 int 1303 negative_advice(int status) 1304 { 1305 1306 return (status == CPL_ERR_RTX_NEG_ADVICE || 1307 status == CPL_ERR_PERSIST_NEG_ADVICE || 1308 status == CPL_ERR_KEEPALV_NEG_ADVICE); 1309 } 1310 1311 static int 1312 alloc_tid_tab(struct tid_info *t, int flags) 1313 { 1314 1315 MPASS(t->ntids > 0); 1316 MPASS(t->tid_tab == NULL); 1317 1318 t->tid_tab = malloc(t->ntids * sizeof(*t->tid_tab), M_CXGBE, 1319 M_ZERO | flags); 1320 if (t->tid_tab == NULL) 1321 return (ENOMEM); 1322 atomic_store_rel_int(&t->tids_in_use, 0); 1323 1324 return (0); 1325 } 1326 1327 static void 1328 free_tid_tab(struct tid_info *t) 1329 { 1330 1331 KASSERT(t->tids_in_use == 0, 1332 ("%s: %d tids still in use.", __func__, t->tids_in_use)); 1333 1334 free(t->tid_tab, M_CXGBE); 1335 t->tid_tab = NULL; 1336 } 1337 1338 static int 1339 alloc_stid_tab(struct tid_info *t, int flags) 1340 { 1341 1342 MPASS(t->nstids > 0); 1343 MPASS(t->stid_tab == NULL); 1344 1345 t->stid_tab = malloc(t->nstids * sizeof(*t->stid_tab), M_CXGBE, 1346 M_ZERO | flags); 1347 if (t->stid_tab == NULL) 1348 return (ENOMEM); 1349 mtx_init(&t->stid_lock, "stid lock", NULL, MTX_DEF); 1350 t->stids_in_use = 0; 1351 TAILQ_INIT(&t->stids); 1352 t->nstids_free_head = t->nstids; 1353 1354 return (0); 1355 } 1356 1357 static void 1358 free_stid_tab(struct tid_info *t) 1359 { 1360 1361 KASSERT(t->stids_in_use == 0, 1362 ("%s: %d tids still in use.", __func__, t->stids_in_use)); 1363 1364 if (mtx_initialized(&t->stid_lock)) 1365 mtx_destroy(&t->stid_lock); 1366 free(t->stid_tab, M_CXGBE); 1367 t->stid_tab = NULL; 1368 } 1369 1370 static void 1371 free_tid_tabs(struct tid_info *t) 1372 { 1373 1374 free_tid_tab(t); 1375 free_stid_tab(t); 1376 } 1377 1378 static int 1379 alloc_tid_tabs(struct tid_info *t) 1380 { 1381 int rc; 1382 1383 rc = alloc_tid_tab(t, M_NOWAIT); 1384 if (rc != 0) 1385 goto failed; 1386 1387 rc = alloc_stid_tab(t, M_NOWAIT); 1388 if (rc != 0) 1389 goto failed; 1390 1391 return (0); 1392 failed: 1393 free_tid_tabs(t); 1394 return (rc); 1395 } 1396 1397 static inline void 1398 alloc_tcb_history(struct adapter *sc, struct tom_data *td) 1399 { 1400 1401 if (sc->tids.ntids == 0 || sc->tids.ntids > 1024) 1402 return; 1403 rw_init(&td->tcb_history_lock, "TCB history"); 1404 td->tcb_history = malloc(sc->tids.ntids * sizeof(*td->tcb_history), 1405 M_CXGBE, M_ZERO | M_NOWAIT); 1406 td->dupack_threshold = G_DUPACKTHRESH(t4_read_reg(sc, A_TP_PARA_REG0)); 1407 } 1408 1409 static inline void 1410 free_tcb_history(struct adapter *sc, struct tom_data *td) 1411 { 1412 #ifdef INVARIANTS 1413 int i; 1414 1415 if (td->tcb_history != NULL) { 1416 for (i = 0; i < sc->tids.ntids; i++) { 1417 MPASS(td->tcb_history[i] == NULL); 1418 } 1419 } 1420 #endif 1421 free(td->tcb_history, M_CXGBE); 1422 if (rw_initialized(&td->tcb_history_lock)) 1423 rw_destroy(&td->tcb_history_lock); 1424 } 1425 1426 static void 1427 free_tom_data(struct adapter *sc, struct tom_data *td) 1428 { 1429 1430 ASSERT_SYNCHRONIZED_OP(sc); 1431 1432 KASSERT(TAILQ_EMPTY(&td->toep_list), 1433 ("%s: TOE PCB list is not empty.", __func__)); 1434 KASSERT(td->lctx_count == 0, 1435 ("%s: lctx hash table is not empty.", __func__)); 1436 1437 t4_free_ppod_region(&td->pr); 1438 1439 if (td->listen_mask != 0) 1440 hashdestroy(td->listen_hash, M_CXGBE, td->listen_mask); 1441 1442 if (mtx_initialized(&td->unsent_wr_lock)) 1443 mtx_destroy(&td->unsent_wr_lock); 1444 if (mtx_initialized(&td->lctx_hash_lock)) 1445 mtx_destroy(&td->lctx_hash_lock); 1446 if (mtx_initialized(&td->toep_list_lock)) 1447 mtx_destroy(&td->toep_list_lock); 1448 1449 free_tcb_history(sc, td); 1450 free_tid_tabs(&sc->tids); 1451 free(td, M_CXGBE); 1452 } 1453 1454 static char * 1455 prepare_pkt(int open_type, uint16_t vtag, struct inpcb *inp, int *pktlen, 1456 int *buflen) 1457 { 1458 char *pkt; 1459 struct tcphdr *th; 1460 int ipv6, len; 1461 const int maxlen = 1462 max(sizeof(struct ether_header), sizeof(struct ether_vlan_header)) + 1463 max(sizeof(struct ip), sizeof(struct ip6_hdr)) + 1464 sizeof(struct tcphdr); 1465 1466 MPASS(open_type == OPEN_TYPE_ACTIVE || open_type == OPEN_TYPE_LISTEN); 1467 1468 pkt = malloc(maxlen, M_CXGBE, M_ZERO | M_NOWAIT); 1469 if (pkt == NULL) 1470 return (NULL); 1471 1472 ipv6 = inp->inp_vflag & INP_IPV6; 1473 len = 0; 1474 1475 if (EVL_VLANOFTAG(vtag) == 0xfff) { 1476 struct ether_header *eh = (void *)pkt; 1477 1478 if (ipv6) 1479 eh->ether_type = htons(ETHERTYPE_IPV6); 1480 else 1481 eh->ether_type = htons(ETHERTYPE_IP); 1482 1483 len += sizeof(*eh); 1484 } else { 1485 struct ether_vlan_header *evh = (void *)pkt; 1486 1487 evh->evl_encap_proto = htons(ETHERTYPE_VLAN); 1488 evh->evl_tag = htons(vtag); 1489 if (ipv6) 1490 evh->evl_proto = htons(ETHERTYPE_IPV6); 1491 else 1492 evh->evl_proto = htons(ETHERTYPE_IP); 1493 1494 len += sizeof(*evh); 1495 } 1496 1497 if (ipv6) { 1498 struct ip6_hdr *ip6 = (void *)&pkt[len]; 1499 1500 ip6->ip6_vfc = IPV6_VERSION; 1501 ip6->ip6_plen = htons(sizeof(struct tcphdr)); 1502 ip6->ip6_nxt = IPPROTO_TCP; 1503 if (open_type == OPEN_TYPE_ACTIVE) { 1504 ip6->ip6_src = inp->in6p_laddr; 1505 ip6->ip6_dst = inp->in6p_faddr; 1506 } else if (open_type == OPEN_TYPE_LISTEN) { 1507 ip6->ip6_src = inp->in6p_laddr; 1508 ip6->ip6_dst = ip6->ip6_src; 1509 } 1510 1511 len += sizeof(*ip6); 1512 } else { 1513 struct ip *ip = (void *)&pkt[len]; 1514 1515 ip->ip_v = IPVERSION; 1516 ip->ip_hl = sizeof(*ip) >> 2; 1517 ip->ip_tos = inp->inp_ip_tos; 1518 ip->ip_len = htons(sizeof(struct ip) + sizeof(struct tcphdr)); 1519 ip->ip_ttl = inp->inp_ip_ttl; 1520 ip->ip_p = IPPROTO_TCP; 1521 if (open_type == OPEN_TYPE_ACTIVE) { 1522 ip->ip_src = inp->inp_laddr; 1523 ip->ip_dst = inp->inp_faddr; 1524 } else if (open_type == OPEN_TYPE_LISTEN) { 1525 ip->ip_src = inp->inp_laddr; 1526 ip->ip_dst = ip->ip_src; 1527 } 1528 1529 len += sizeof(*ip); 1530 } 1531 1532 th = (void *)&pkt[len]; 1533 if (open_type == OPEN_TYPE_ACTIVE) { 1534 th->th_sport = inp->inp_lport; /* network byte order already */ 1535 th->th_dport = inp->inp_fport; /* ditto */ 1536 } else if (open_type == OPEN_TYPE_LISTEN) { 1537 th->th_sport = inp->inp_lport; /* network byte order already */ 1538 th->th_dport = th->th_sport; 1539 } 1540 len += sizeof(th); 1541 1542 *pktlen = *buflen = len; 1543 return (pkt); 1544 } 1545 1546 const struct offload_settings * 1547 lookup_offload_policy(struct adapter *sc, int open_type, struct mbuf *m, 1548 uint16_t vtag, struct inpcb *inp) 1549 { 1550 const struct t4_offload_policy *op; 1551 char *pkt; 1552 struct offload_rule *r; 1553 int i, matched, pktlen, buflen; 1554 static const struct offload_settings allow_offloading_settings = { 1555 .offload = 1, 1556 .rx_coalesce = -1, 1557 .cong_algo = -1, 1558 .sched_class = -1, 1559 .tstamp = -1, 1560 .sack = -1, 1561 .nagle = -1, 1562 .ecn = -1, 1563 .ddp = -1, 1564 .tls = -1, 1565 .txq = -1, 1566 .rxq = -1, 1567 .mss = -1, 1568 }; 1569 static const struct offload_settings disallow_offloading_settings = { 1570 .offload = 0, 1571 /* rest is irrelevant when offload is off. */ 1572 }; 1573 1574 rw_assert(&sc->policy_lock, RA_LOCKED); 1575 1576 /* 1577 * If there's no Connection Offloading Policy attached to the device 1578 * then we need to return a default static policy. If 1579 * "cop_managed_offloading" is true, then we need to disallow 1580 * offloading until a COP is attached to the device. Otherwise we 1581 * allow offloading ... 1582 */ 1583 op = sc->policy; 1584 if (op == NULL) { 1585 if (sc->tt.cop_managed_offloading) 1586 return (&disallow_offloading_settings); 1587 else 1588 return (&allow_offloading_settings); 1589 } 1590 1591 switch (open_type) { 1592 case OPEN_TYPE_ACTIVE: 1593 case OPEN_TYPE_LISTEN: 1594 pkt = prepare_pkt(open_type, vtag, inp, &pktlen, &buflen); 1595 break; 1596 case OPEN_TYPE_PASSIVE: 1597 MPASS(m != NULL); 1598 pkt = mtod(m, char *); 1599 MPASS(*pkt == CPL_PASS_ACCEPT_REQ); 1600 pkt += sizeof(struct cpl_pass_accept_req); 1601 pktlen = m->m_pkthdr.len - sizeof(struct cpl_pass_accept_req); 1602 buflen = m->m_len - sizeof(struct cpl_pass_accept_req); 1603 break; 1604 default: 1605 MPASS(0); 1606 return (&disallow_offloading_settings); 1607 } 1608 1609 if (pkt == NULL || pktlen == 0 || buflen == 0) 1610 return (&disallow_offloading_settings); 1611 1612 matched = 0; 1613 r = &op->rule[0]; 1614 for (i = 0; i < op->nrules; i++, r++) { 1615 if (r->open_type != open_type && 1616 r->open_type != OPEN_TYPE_DONTCARE) { 1617 continue; 1618 } 1619 matched = bpf_filter(r->bpf_prog.bf_insns, pkt, pktlen, buflen); 1620 if (matched) 1621 break; 1622 } 1623 1624 if (open_type == OPEN_TYPE_ACTIVE || open_type == OPEN_TYPE_LISTEN) 1625 free(pkt, M_CXGBE); 1626 1627 return (matched ? &r->settings : &disallow_offloading_settings); 1628 } 1629 1630 static void 1631 reclaim_wr_resources(void *arg, int count) 1632 { 1633 struct tom_data *td = arg; 1634 STAILQ_HEAD(, wrqe) twr_list = STAILQ_HEAD_INITIALIZER(twr_list); 1635 struct cpl_act_open_req *cpl; 1636 u_int opcode, atid, tid; 1637 struct wrqe *wr; 1638 struct adapter *sc = td_adapter(td); 1639 1640 mtx_lock(&td->unsent_wr_lock); 1641 STAILQ_SWAP(&td->unsent_wr_list, &twr_list, wrqe); 1642 mtx_unlock(&td->unsent_wr_lock); 1643 1644 while ((wr = STAILQ_FIRST(&twr_list)) != NULL) { 1645 STAILQ_REMOVE_HEAD(&twr_list, link); 1646 1647 cpl = wrtod(wr); 1648 opcode = GET_OPCODE(cpl); 1649 1650 switch (opcode) { 1651 case CPL_ACT_OPEN_REQ: 1652 case CPL_ACT_OPEN_REQ6: 1653 atid = G_TID_TID(be32toh(OPCODE_TID(cpl))); 1654 CTR2(KTR_CXGBE, "%s: atid %u ", __func__, atid); 1655 act_open_failure_cleanup(sc, atid, EHOSTUNREACH); 1656 free(wr, M_CXGBE); 1657 break; 1658 case CPL_PASS_ACCEPT_RPL: 1659 tid = GET_TID(cpl); 1660 CTR2(KTR_CXGBE, "%s: tid %u ", __func__, tid); 1661 synack_failure_cleanup(sc, tid); 1662 free(wr, M_CXGBE); 1663 break; 1664 default: 1665 log(LOG_ERR, "%s: leaked work request %p, wr_len %d, " 1666 "opcode %x\n", __func__, wr, wr->wr_len, opcode); 1667 /* WR not freed here; go look at it with a debugger. */ 1668 } 1669 } 1670 } 1671 1672 /* 1673 * Ground control to Major TOM 1674 * Commencing countdown, engines on 1675 */ 1676 static int 1677 t4_tom_activate(struct adapter *sc) 1678 { 1679 struct tom_data *td; 1680 struct toedev *tod; 1681 struct vi_info *vi; 1682 int i, rc, v; 1683 1684 ASSERT_SYNCHRONIZED_OP(sc); 1685 1686 /* per-adapter softc for TOM */ 1687 td = malloc(sizeof(*td), M_CXGBE, M_ZERO | M_NOWAIT); 1688 if (td == NULL) 1689 return (ENOMEM); 1690 1691 /* List of TOE PCBs and associated lock */ 1692 mtx_init(&td->toep_list_lock, "PCB list lock", NULL, MTX_DEF); 1693 TAILQ_INIT(&td->toep_list); 1694 1695 /* Listen context */ 1696 mtx_init(&td->lctx_hash_lock, "lctx hash lock", NULL, MTX_DEF); 1697 td->listen_hash = hashinit_flags(LISTEN_HASH_SIZE, M_CXGBE, 1698 &td->listen_mask, HASH_NOWAIT); 1699 1700 /* List of WRs for which L2 resolution failed */ 1701 mtx_init(&td->unsent_wr_lock, "Unsent WR list lock", NULL, MTX_DEF); 1702 STAILQ_INIT(&td->unsent_wr_list); 1703 TASK_INIT(&td->reclaim_wr_resources, 0, reclaim_wr_resources, td); 1704 1705 /* TID tables */ 1706 rc = alloc_tid_tabs(&sc->tids); 1707 if (rc != 0) 1708 goto done; 1709 1710 rc = t4_init_ppod_region(&td->pr, &sc->vres.ddp, 1711 t4_read_reg(sc, A_ULP_RX_TDDP_PSZ), "TDDP page pods"); 1712 if (rc != 0) 1713 goto done; 1714 t4_set_reg_field(sc, A_ULP_RX_TDDP_TAGMASK, 1715 V_TDDPTAGMASK(M_TDDPTAGMASK), td->pr.pr_tag_mask); 1716 1717 alloc_tcb_history(sc, td); 1718 1719 /* toedev ops */ 1720 tod = &td->tod; 1721 init_toedev(tod); 1722 tod->tod_softc = sc; 1723 tod->tod_connect = t4_connect; 1724 tod->tod_listen_start = t4_listen_start; 1725 tod->tod_listen_stop = t4_listen_stop; 1726 tod->tod_rcvd = t4_rcvd; 1727 tod->tod_output = t4_tod_output; 1728 tod->tod_send_rst = t4_send_rst; 1729 tod->tod_send_fin = t4_send_fin; 1730 tod->tod_pcb_detach = t4_pcb_detach; 1731 tod->tod_l2_update = t4_l2_update; 1732 tod->tod_syncache_added = t4_syncache_added; 1733 tod->tod_syncache_removed = t4_syncache_removed; 1734 tod->tod_syncache_respond = t4_syncache_respond; 1735 tod->tod_offload_socket = t4_offload_socket; 1736 tod->tod_ctloutput = t4_ctloutput; 1737 tod->tod_tcp_info = t4_tcp_info; 1738 #ifdef KERN_TLS 1739 tod->tod_alloc_tls_session = t4_alloc_tls_session; 1740 #endif 1741 1742 for_each_port(sc, i) { 1743 for_each_vi(sc->port[i], v, vi) { 1744 TOEDEV(vi->ifp) = &td->tod; 1745 } 1746 } 1747 1748 sc->tom_softc = td; 1749 register_toedev(sc->tom_softc); 1750 1751 done: 1752 if (rc != 0) 1753 free_tom_data(sc, td); 1754 return (rc); 1755 } 1756 1757 static int 1758 t4_tom_deactivate(struct adapter *sc) 1759 { 1760 int rc = 0; 1761 struct tom_data *td = sc->tom_softc; 1762 1763 ASSERT_SYNCHRONIZED_OP(sc); 1764 1765 if (td == NULL) 1766 return (0); /* XXX. KASSERT? */ 1767 1768 if (sc->offload_map != 0) 1769 return (EBUSY); /* at least one port has IFCAP_TOE enabled */ 1770 1771 if (uld_active(sc, ULD_IWARP) || uld_active(sc, ULD_ISCSI)) 1772 return (EBUSY); /* both iWARP and iSCSI rely on the TOE. */ 1773 1774 mtx_lock(&td->toep_list_lock); 1775 if (!TAILQ_EMPTY(&td->toep_list)) 1776 rc = EBUSY; 1777 mtx_unlock(&td->toep_list_lock); 1778 1779 mtx_lock(&td->lctx_hash_lock); 1780 if (td->lctx_count > 0) 1781 rc = EBUSY; 1782 mtx_unlock(&td->lctx_hash_lock); 1783 1784 taskqueue_drain(taskqueue_thread, &td->reclaim_wr_resources); 1785 mtx_lock(&td->unsent_wr_lock); 1786 if (!STAILQ_EMPTY(&td->unsent_wr_list)) 1787 rc = EBUSY; 1788 mtx_unlock(&td->unsent_wr_lock); 1789 1790 if (rc == 0) { 1791 unregister_toedev(sc->tom_softc); 1792 free_tom_data(sc, td); 1793 sc->tom_softc = NULL; 1794 } 1795 1796 return (rc); 1797 } 1798 1799 static int 1800 t4_aio_queue_tom(struct socket *so, struct kaiocb *job) 1801 { 1802 struct tcpcb *tp = so_sototcpcb(so); 1803 struct toepcb *toep = tp->t_toe; 1804 int error; 1805 1806 if (ulp_mode(toep) == ULP_MODE_TCPDDP) { 1807 error = t4_aio_queue_ddp(so, job); 1808 if (error != EOPNOTSUPP) 1809 return (error); 1810 } 1811 1812 return (t4_aio_queue_aiotx(so, job)); 1813 } 1814 1815 static int 1816 t4_ctloutput_tom(struct socket *so, struct sockopt *sopt) 1817 { 1818 1819 if (sopt->sopt_level != IPPROTO_TCP) 1820 return (tcp_ctloutput(so, sopt)); 1821 1822 switch (sopt->sopt_name) { 1823 case TCP_TLSOM_SET_TLS_CONTEXT: 1824 case TCP_TLSOM_GET_TLS_TOM: 1825 case TCP_TLSOM_CLR_TLS_TOM: 1826 case TCP_TLSOM_CLR_QUIES: 1827 return (t4_ctloutput_tls(so, sopt)); 1828 default: 1829 return (tcp_ctloutput(so, sopt)); 1830 } 1831 } 1832 1833 static int 1834 t4_tom_mod_load(void) 1835 { 1836 struct protosw *tcp_protosw, *tcp6_protosw; 1837 1838 /* CPL handlers */ 1839 t4_register_cpl_handler(CPL_GET_TCB_RPL, do_get_tcb_rpl); 1840 t4_register_shared_cpl_handler(CPL_L2T_WRITE_RPL, do_l2t_write_rpl2, 1841 CPL_COOKIE_TOM); 1842 t4_init_connect_cpl_handlers(); 1843 t4_init_listen_cpl_handlers(); 1844 t4_init_cpl_io_handlers(); 1845 1846 t4_ddp_mod_load(); 1847 t4_tls_mod_load(); 1848 1849 tcp_protosw = pffindproto(PF_INET, IPPROTO_TCP, SOCK_STREAM); 1850 if (tcp_protosw == NULL) 1851 return (ENOPROTOOPT); 1852 bcopy(tcp_protosw, &toe_protosw, sizeof(toe_protosw)); 1853 bcopy(tcp_protosw->pr_usrreqs, &toe_usrreqs, sizeof(toe_usrreqs)); 1854 toe_usrreqs.pru_aio_queue = t4_aio_queue_tom; 1855 toe_protosw.pr_ctloutput = t4_ctloutput_tom; 1856 toe_protosw.pr_usrreqs = &toe_usrreqs; 1857 1858 tcp6_protosw = pffindproto(PF_INET6, IPPROTO_TCP, SOCK_STREAM); 1859 if (tcp6_protosw == NULL) 1860 return (ENOPROTOOPT); 1861 bcopy(tcp6_protosw, &toe6_protosw, sizeof(toe6_protosw)); 1862 bcopy(tcp6_protosw->pr_usrreqs, &toe6_usrreqs, sizeof(toe6_usrreqs)); 1863 toe6_usrreqs.pru_aio_queue = t4_aio_queue_tom; 1864 toe6_protosw.pr_ctloutput = t4_ctloutput_tom; 1865 toe6_protosw.pr_usrreqs = &toe6_usrreqs; 1866 1867 return (t4_register_uld(&tom_uld_info)); 1868 } 1869 1870 static void 1871 tom_uninit(struct adapter *sc, void *arg __unused) 1872 { 1873 if (begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4tomun")) 1874 return; 1875 1876 /* Try to free resources (works only if no port has IFCAP_TOE) */ 1877 if (uld_active(sc, ULD_TOM)) 1878 t4_deactivate_uld(sc, ULD_TOM); 1879 1880 end_synchronized_op(sc, 0); 1881 } 1882 1883 static int 1884 t4_tom_mod_unload(void) 1885 { 1886 t4_iterate(tom_uninit, NULL); 1887 1888 if (t4_unregister_uld(&tom_uld_info) == EBUSY) 1889 return (EBUSY); 1890 1891 t4_tls_mod_unload(); 1892 t4_ddp_mod_unload(); 1893 1894 t4_uninit_connect_cpl_handlers(); 1895 t4_uninit_listen_cpl_handlers(); 1896 t4_uninit_cpl_io_handlers(); 1897 t4_register_shared_cpl_handler(CPL_L2T_WRITE_RPL, NULL, CPL_COOKIE_TOM); 1898 t4_register_cpl_handler(CPL_GET_TCB_RPL, NULL); 1899 1900 return (0); 1901 } 1902 #endif /* TCP_OFFLOAD */ 1903 1904 static int 1905 t4_tom_modevent(module_t mod, int cmd, void *arg) 1906 { 1907 int rc = 0; 1908 1909 #ifdef TCP_OFFLOAD 1910 switch (cmd) { 1911 case MOD_LOAD: 1912 rc = t4_tom_mod_load(); 1913 break; 1914 1915 case MOD_UNLOAD: 1916 rc = t4_tom_mod_unload(); 1917 break; 1918 1919 default: 1920 rc = EINVAL; 1921 } 1922 #else 1923 printf("t4_tom: compiled without TCP_OFFLOAD support.\n"); 1924 rc = EOPNOTSUPP; 1925 #endif 1926 return (rc); 1927 } 1928 1929 static moduledata_t t4_tom_moddata= { 1930 "t4_tom", 1931 t4_tom_modevent, 1932 0 1933 }; 1934 1935 MODULE_VERSION(t4_tom, 1); 1936 MODULE_DEPEND(t4_tom, toecore, 1, 1, 1); 1937 MODULE_DEPEND(t4_tom, t4nex, 1, 1, 1); 1938 DECLARE_MODULE(t4_tom, t4_tom_moddata, SI_SUB_EXEC, SI_ORDER_ANY); 1939