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_seq.h> 67 #include <netinet/tcp_timer.h> 68 #include <netinet/tcp_var.h> 69 #include <netinet/toecore.h> 70 #include <netinet/cc/cc.h> 71 72 #ifdef TCP_OFFLOAD 73 #include "common/common.h" 74 #include "common/t4_msg.h" 75 #include "common/t4_regs.h" 76 #include "common/t4_regs_values.h" 77 #include "common/t4_tcb.h" 78 #include "t4_clip.h" 79 #include "tom/t4_tom_l2t.h" 80 #include "tom/t4_tom.h" 81 #include "tom/t4_tls.h" 82 83 static struct protosw toe_protosw; 84 static struct protosw toe6_protosw; 85 86 /* Module ops */ 87 static int t4_tom_mod_load(void); 88 static int t4_tom_mod_unload(void); 89 static int t4_tom_modevent(module_t, int, void *); 90 91 /* ULD ops and helpers */ 92 static int t4_tom_activate(struct adapter *); 93 static int t4_tom_deactivate(struct adapter *); 94 95 static struct uld_info tom_uld_info = { 96 .uld_id = ULD_TOM, 97 .activate = t4_tom_activate, 98 .deactivate = t4_tom_deactivate, 99 }; 100 101 static void release_offload_resources(struct toepcb *); 102 static int alloc_tid_tabs(struct tid_info *); 103 static void free_tid_tabs(struct tid_info *); 104 static void free_tom_data(struct adapter *, struct tom_data *); 105 static void reclaim_wr_resources(void *, int); 106 107 struct toepcb * 108 alloc_toepcb(struct vi_info *vi, int flags) 109 { 110 struct port_info *pi = vi->pi; 111 struct adapter *sc = pi->adapter; 112 struct toepcb *toep; 113 int tx_credits, txsd_total, len; 114 115 /* 116 * The firmware counts tx work request credits in units of 16 bytes 117 * each. Reserve room for an ABORT_REQ so the driver never has to worry 118 * about tx credits if it wants to abort a connection. 119 */ 120 tx_credits = sc->params.ofldq_wr_cred; 121 tx_credits -= howmany(sizeof(struct cpl_abort_req), 16); 122 123 /* 124 * Shortest possible tx work request is a fw_ofld_tx_data_wr + 1 byte 125 * immediate payload, and firmware counts tx work request credits in 126 * units of 16 byte. Calculate the maximum work requests possible. 127 */ 128 txsd_total = tx_credits / 129 howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16); 130 131 len = offsetof(struct toepcb, txsd) + 132 txsd_total * sizeof(struct ofld_tx_sdesc); 133 134 toep = malloc(len, M_CXGBE, M_ZERO | flags); 135 if (toep == NULL) 136 return (NULL); 137 138 refcount_init(&toep->refcount, 1); 139 toep->td = sc->tom_softc; 140 toep->vi = vi; 141 toep->tid = -1; 142 toep->tx_total = tx_credits; 143 toep->tx_credits = tx_credits; 144 mbufq_init(&toep->ulp_pduq, INT_MAX); 145 mbufq_init(&toep->ulp_pdu_reclaimq, INT_MAX); 146 toep->txsd_total = txsd_total; 147 toep->txsd_avail = txsd_total; 148 toep->txsd_pidx = 0; 149 toep->txsd_cidx = 0; 150 aiotx_init_toep(toep); 151 152 return (toep); 153 } 154 155 /* 156 * Initialize a toepcb after its params have been filled out. 157 */ 158 int 159 init_toepcb(struct vi_info *vi, struct toepcb *toep) 160 { 161 struct conn_params *cp = &toep->params; 162 struct port_info *pi = vi->pi; 163 struct adapter *sc = pi->adapter; 164 struct tx_cl_rl_params *tc; 165 166 if (cp->tc_idx >= 0 && cp->tc_idx < sc->params.nsched_cls) { 167 tc = &pi->sched_params->cl_rl[cp->tc_idx]; 168 mtx_lock(&sc->tc_lock); 169 if (tc->state != CS_HW_CONFIGURED) { 170 CH_ERR(vi, "tid %d cannot be bound to traffic class %d " 171 "because it is not configured (its state is %d)\n", 172 toep->tid, cp->tc_idx, tc->state); 173 cp->tc_idx = -1; 174 } else { 175 tc->refcount++; 176 } 177 mtx_unlock(&sc->tc_lock); 178 } 179 toep->ofld_txq = &sc->sge.ofld_txq[cp->txq_idx]; 180 toep->ofld_rxq = &sc->sge.ofld_rxq[cp->rxq_idx]; 181 toep->ctrlq = &sc->sge.ctrlq[pi->port_id]; 182 183 tls_init_toep(toep); 184 if (ulp_mode(toep) == ULP_MODE_TCPDDP) 185 ddp_init_toep(toep); 186 187 toep->flags |= TPF_INITIALIZED; 188 189 return (0); 190 } 191 192 struct toepcb * 193 hold_toepcb(struct toepcb *toep) 194 { 195 196 refcount_acquire(&toep->refcount); 197 return (toep); 198 } 199 200 void 201 free_toepcb(struct toepcb *toep) 202 { 203 204 if (refcount_release(&toep->refcount) == 0) 205 return; 206 207 KASSERT(!(toep->flags & TPF_ATTACHED), 208 ("%s: attached to an inpcb", __func__)); 209 KASSERT(!(toep->flags & TPF_CPL_PENDING), 210 ("%s: CPL pending", __func__)); 211 212 if (toep->flags & TPF_INITIALIZED) { 213 if (ulp_mode(toep) == ULP_MODE_TCPDDP) 214 ddp_uninit_toep(toep); 215 tls_uninit_toep(toep); 216 } 217 free(toep, M_CXGBE); 218 } 219 220 /* 221 * Set up the socket for TCP offload. 222 */ 223 void 224 offload_socket(struct socket *so, struct toepcb *toep) 225 { 226 struct tom_data *td = toep->td; 227 struct inpcb *inp = sotoinpcb(so); 228 struct tcpcb *tp = intotcpcb(inp); 229 struct sockbuf *sb; 230 231 INP_WLOCK_ASSERT(inp); 232 233 /* Update socket */ 234 sb = &so->so_snd; 235 SOCKBUF_LOCK(sb); 236 sb->sb_flags |= SB_NOCOALESCE; 237 SOCKBUF_UNLOCK(sb); 238 sb = &so->so_rcv; 239 SOCKBUF_LOCK(sb); 240 sb->sb_flags |= SB_NOCOALESCE; 241 if (inp->inp_vflag & INP_IPV6) 242 so->so_proto = &toe6_protosw; 243 else 244 so->so_proto = &toe_protosw; 245 SOCKBUF_UNLOCK(sb); 246 247 /* Update TCP PCB */ 248 tp->tod = &td->tod; 249 tp->t_toe = toep; 250 tp->t_flags |= TF_TOE; 251 252 /* Install an extra hold on inp */ 253 toep->inp = inp; 254 toep->flags |= TPF_ATTACHED; 255 in_pcbref(inp); 256 257 /* Add the TOE PCB to the active list */ 258 mtx_lock(&td->toep_list_lock); 259 TAILQ_INSERT_HEAD(&td->toep_list, toep, link); 260 mtx_unlock(&td->toep_list_lock); 261 } 262 263 void 264 restore_so_proto(struct socket *so, bool v6) 265 { 266 if (v6) 267 so->so_proto = &tcp6_protosw; 268 else 269 so->so_proto = &tcp_protosw; 270 } 271 272 /* This is _not_ the normal way to "unoffload" a socket. */ 273 void 274 undo_offload_socket(struct socket *so) 275 { 276 struct inpcb *inp = sotoinpcb(so); 277 struct tcpcb *tp = intotcpcb(inp); 278 struct toepcb *toep = tp->t_toe; 279 struct tom_data *td = toep->td; 280 struct sockbuf *sb; 281 282 INP_WLOCK_ASSERT(inp); 283 284 sb = &so->so_snd; 285 SOCKBUF_LOCK(sb); 286 sb->sb_flags &= ~SB_NOCOALESCE; 287 SOCKBUF_UNLOCK(sb); 288 sb = &so->so_rcv; 289 SOCKBUF_LOCK(sb); 290 sb->sb_flags &= ~SB_NOCOALESCE; 291 restore_so_proto(so, inp->inp_vflag & INP_IPV6); 292 SOCKBUF_UNLOCK(sb); 293 294 tp->tod = NULL; 295 tp->t_toe = NULL; 296 tp->t_flags &= ~TF_TOE; 297 298 toep->inp = NULL; 299 toep->flags &= ~TPF_ATTACHED; 300 if (in_pcbrele_wlocked(inp)) 301 panic("%s: inp freed.", __func__); 302 303 mtx_lock(&td->toep_list_lock); 304 TAILQ_REMOVE(&td->toep_list, toep, link); 305 mtx_unlock(&td->toep_list_lock); 306 } 307 308 static void 309 release_offload_resources(struct toepcb *toep) 310 { 311 struct tom_data *td = toep->td; 312 struct adapter *sc = td_adapter(td); 313 int tid = toep->tid; 314 315 KASSERT(!(toep->flags & TPF_CPL_PENDING), 316 ("%s: %p has CPL pending.", __func__, toep)); 317 KASSERT(!(toep->flags & TPF_ATTACHED), 318 ("%s: %p is still attached.", __func__, toep)); 319 320 CTR5(KTR_CXGBE, "%s: toep %p (tid %d, l2te %p, ce %p)", 321 __func__, toep, tid, toep->l2te, toep->ce); 322 323 /* 324 * These queues should have been emptied at approximately the same time 325 * that a normal connection's socket's so_snd would have been purged or 326 * drained. Do _not_ clean up here. 327 */ 328 MPASS(mbufq_len(&toep->ulp_pduq) == 0); 329 MPASS(mbufq_len(&toep->ulp_pdu_reclaimq) == 0); 330 #ifdef INVARIANTS 331 if (ulp_mode(toep) == ULP_MODE_TCPDDP) 332 ddp_assert_empty(toep); 333 #endif 334 MPASS(TAILQ_EMPTY(&toep->aiotx_jobq)); 335 336 if (toep->l2te) 337 t4_l2t_release(toep->l2te); 338 339 if (tid >= 0) { 340 remove_tid(sc, tid, toep->ce ? 2 : 1); 341 release_tid(sc, tid, toep->ctrlq); 342 } 343 344 if (toep->ce) 345 t4_release_clip_entry(sc, toep->ce); 346 347 if (toep->params.tc_idx != -1) 348 t4_release_cl_rl(sc, toep->vi->pi->port_id, toep->params.tc_idx); 349 350 mtx_lock(&td->toep_list_lock); 351 TAILQ_REMOVE(&td->toep_list, toep, link); 352 mtx_unlock(&td->toep_list_lock); 353 354 free_toepcb(toep); 355 } 356 357 /* 358 * The kernel is done with the TCP PCB and this is our opportunity to unhook the 359 * toepcb hanging off of it. If the TOE driver is also done with the toepcb (no 360 * pending CPL) then it is time to release all resources tied to the toepcb. 361 * 362 * Also gets called when an offloaded active open fails and the TOM wants the 363 * kernel to take the TCP PCB back. 364 */ 365 static void 366 t4_pcb_detach(struct toedev *tod __unused, struct tcpcb *tp) 367 { 368 #if defined(KTR) || defined(INVARIANTS) 369 struct inpcb *inp = tptoinpcb(tp); 370 #endif 371 struct toepcb *toep = tp->t_toe; 372 373 INP_WLOCK_ASSERT(inp); 374 375 KASSERT(toep != NULL, ("%s: toep is NULL", __func__)); 376 KASSERT(toep->flags & TPF_ATTACHED, 377 ("%s: not attached", __func__)); 378 379 #ifdef KTR 380 if (tp->t_state == TCPS_SYN_SENT) { 381 CTR6(KTR_CXGBE, "%s: atid %d, toep %p (0x%x), inp %p (0x%x)", 382 __func__, toep->tid, toep, toep->flags, inp, 383 inp->inp_flags); 384 } else { 385 CTR6(KTR_CXGBE, 386 "t4_pcb_detach: tid %d (%s), toep %p (0x%x), inp %p (0x%x)", 387 toep->tid, tcpstates[tp->t_state], toep, toep->flags, inp, 388 inp->inp_flags); 389 } 390 #endif 391 392 tp->tod = NULL; 393 tp->t_toe = NULL; 394 tp->t_flags &= ~TF_TOE; 395 toep->flags &= ~TPF_ATTACHED; 396 397 if (!(toep->flags & TPF_CPL_PENDING)) 398 release_offload_resources(toep); 399 } 400 401 /* 402 * setsockopt handler. 403 */ 404 static void 405 t4_ctloutput(struct toedev *tod, struct tcpcb *tp, int dir, int name) 406 { 407 struct adapter *sc = tod->tod_softc; 408 struct toepcb *toep = tp->t_toe; 409 410 if (dir == SOPT_GET) 411 return; 412 413 CTR4(KTR_CXGBE, "%s: tp %p, dir %u, name %u", __func__, tp, dir, name); 414 415 switch (name) { 416 case TCP_NODELAY: 417 if (tp->t_state != TCPS_ESTABLISHED) 418 break; 419 toep->params.nagle = tp->t_flags & TF_NODELAY ? 0 : 1; 420 t4_set_tcb_field(sc, toep->ctrlq, toep, W_TCB_T_FLAGS, 421 V_TF_NAGLE(1), V_TF_NAGLE(toep->params.nagle), 0, 0); 422 break; 423 default: 424 break; 425 } 426 } 427 428 static inline uint64_t 429 get_tcb_tflags(const uint64_t *tcb) 430 { 431 432 return ((be64toh(tcb[14]) << 32) | (be64toh(tcb[15]) >> 32)); 433 } 434 435 static inline uint32_t 436 get_tcb_field(const uint64_t *tcb, u_int word, uint32_t mask, u_int shift) 437 { 438 #define LAST_WORD ((TCB_SIZE / 4) - 1) 439 uint64_t t1, t2; 440 int flit_idx; 441 442 MPASS(mask != 0); 443 MPASS(word <= LAST_WORD); 444 MPASS(shift < 32); 445 446 flit_idx = (LAST_WORD - word) / 2; 447 if (word & 0x1) 448 shift += 32; 449 t1 = be64toh(tcb[flit_idx]) >> shift; 450 t2 = 0; 451 if (fls(mask) > 64 - shift) { 452 /* 453 * Will spill over into the next logical flit, which is the flit 454 * before this one. The flit_idx before this one must be valid. 455 */ 456 MPASS(flit_idx > 0); 457 t2 = be64toh(tcb[flit_idx - 1]) << (64 - shift); 458 } 459 return ((t2 | t1) & mask); 460 #undef LAST_WORD 461 } 462 #define GET_TCB_FIELD(tcb, F) \ 463 get_tcb_field(tcb, W_TCB_##F, M_TCB_##F, S_TCB_##F) 464 465 /* 466 * Issues a CPL_GET_TCB to read the entire TCB for the tid. 467 */ 468 static int 469 send_get_tcb(struct adapter *sc, u_int tid) 470 { 471 struct cpl_get_tcb *cpl; 472 struct wrq_cookie cookie; 473 474 MPASS(tid >= sc->tids.tid_base); 475 MPASS(tid - sc->tids.tid_base < sc->tids.ntids); 476 477 cpl = start_wrq_wr(&sc->sge.ctrlq[0], howmany(sizeof(*cpl), 16), 478 &cookie); 479 if (__predict_false(cpl == NULL)) 480 return (ENOMEM); 481 bzero(cpl, sizeof(*cpl)); 482 INIT_TP_WR(cpl, tid); 483 OPCODE_TID(cpl) = htobe32(MK_OPCODE_TID(CPL_GET_TCB, tid)); 484 cpl->reply_ctrl = htobe16(V_REPLY_CHAN(0) | 485 V_QUEUENO(sc->sge.ofld_rxq[0].iq.cntxt_id)); 486 cpl->cookie = 0xff; 487 commit_wrq_wr(&sc->sge.ctrlq[0], cpl, &cookie); 488 489 return (0); 490 } 491 492 static struct tcb_histent * 493 alloc_tcb_histent(struct adapter *sc, u_int tid, int flags) 494 { 495 struct tcb_histent *te; 496 497 MPASS(flags == M_NOWAIT || flags == M_WAITOK); 498 499 te = malloc(sizeof(*te), M_CXGBE, M_ZERO | flags); 500 if (te == NULL) 501 return (NULL); 502 mtx_init(&te->te_lock, "TCB entry", NULL, MTX_DEF); 503 callout_init_mtx(&te->te_callout, &te->te_lock, 0); 504 te->te_adapter = sc; 505 te->te_tid = tid; 506 507 return (te); 508 } 509 510 static void 511 free_tcb_histent(struct tcb_histent *te) 512 { 513 514 mtx_destroy(&te->te_lock); 515 free(te, M_CXGBE); 516 } 517 518 /* 519 * Start tracking the tid in the TCB history. 520 */ 521 int 522 add_tid_to_history(struct adapter *sc, u_int tid) 523 { 524 struct tcb_histent *te = NULL; 525 struct tom_data *td = sc->tom_softc; 526 int rc; 527 528 MPASS(tid >= sc->tids.tid_base); 529 MPASS(tid - sc->tids.tid_base < sc->tids.ntids); 530 531 if (td->tcb_history == NULL) 532 return (ENXIO); 533 534 rw_wlock(&td->tcb_history_lock); 535 if (td->tcb_history[tid] != NULL) { 536 rc = EEXIST; 537 goto done; 538 } 539 te = alloc_tcb_histent(sc, tid, M_NOWAIT); 540 if (te == NULL) { 541 rc = ENOMEM; 542 goto done; 543 } 544 mtx_lock(&te->te_lock); 545 rc = send_get_tcb(sc, tid); 546 if (rc == 0) { 547 te->te_flags |= TE_RPL_PENDING; 548 td->tcb_history[tid] = te; 549 } else { 550 free(te, M_CXGBE); 551 } 552 mtx_unlock(&te->te_lock); 553 done: 554 rw_wunlock(&td->tcb_history_lock); 555 return (rc); 556 } 557 558 static void 559 remove_tcb_histent(struct tcb_histent *te) 560 { 561 struct adapter *sc = te->te_adapter; 562 struct tom_data *td = sc->tom_softc; 563 564 rw_assert(&td->tcb_history_lock, RA_WLOCKED); 565 mtx_assert(&te->te_lock, MA_OWNED); 566 MPASS(td->tcb_history[te->te_tid] == te); 567 568 td->tcb_history[te->te_tid] = NULL; 569 free_tcb_histent(te); 570 rw_wunlock(&td->tcb_history_lock); 571 } 572 573 static inline struct tcb_histent * 574 lookup_tcb_histent(struct adapter *sc, u_int tid, bool addrem) 575 { 576 struct tcb_histent *te; 577 struct tom_data *td = sc->tom_softc; 578 579 MPASS(tid >= sc->tids.tid_base); 580 MPASS(tid - sc->tids.tid_base < sc->tids.ntids); 581 582 if (td->tcb_history == NULL) 583 return (NULL); 584 585 if (addrem) 586 rw_wlock(&td->tcb_history_lock); 587 else 588 rw_rlock(&td->tcb_history_lock); 589 te = td->tcb_history[tid]; 590 if (te != NULL) { 591 mtx_lock(&te->te_lock); 592 return (te); /* with both locks held */ 593 } 594 if (addrem) 595 rw_wunlock(&td->tcb_history_lock); 596 else 597 rw_runlock(&td->tcb_history_lock); 598 599 return (te); 600 } 601 602 static inline void 603 release_tcb_histent(struct tcb_histent *te) 604 { 605 struct adapter *sc = te->te_adapter; 606 struct tom_data *td = sc->tom_softc; 607 608 mtx_assert(&te->te_lock, MA_OWNED); 609 mtx_unlock(&te->te_lock); 610 rw_assert(&td->tcb_history_lock, RA_RLOCKED); 611 rw_runlock(&td->tcb_history_lock); 612 } 613 614 static void 615 request_tcb(void *arg) 616 { 617 struct tcb_histent *te = arg; 618 619 mtx_assert(&te->te_lock, MA_OWNED); 620 621 /* Noone else is supposed to update the histent. */ 622 MPASS(!(te->te_flags & TE_RPL_PENDING)); 623 if (send_get_tcb(te->te_adapter, te->te_tid) == 0) 624 te->te_flags |= TE_RPL_PENDING; 625 else 626 callout_schedule(&te->te_callout, hz / 100); 627 } 628 629 static void 630 update_tcb_histent(struct tcb_histent *te, const uint64_t *tcb) 631 { 632 struct tom_data *td = te->te_adapter->tom_softc; 633 uint64_t tflags = get_tcb_tflags(tcb); 634 uint8_t sample = 0; 635 636 if (GET_TCB_FIELD(tcb, SND_MAX_RAW) != GET_TCB_FIELD(tcb, SND_UNA_RAW)) { 637 if (GET_TCB_FIELD(tcb, T_RXTSHIFT) != 0) 638 sample |= TS_RTO; 639 if (GET_TCB_FIELD(tcb, T_DUPACKS) != 0) 640 sample |= TS_DUPACKS; 641 if (GET_TCB_FIELD(tcb, T_DUPACKS) >= td->dupack_threshold) 642 sample |= TS_FASTREXMT; 643 } 644 645 if (GET_TCB_FIELD(tcb, SND_MAX_RAW) != 0) { 646 uint32_t snd_wnd; 647 648 sample |= TS_SND_BACKLOGGED; /* for whatever reason. */ 649 650 snd_wnd = GET_TCB_FIELD(tcb, RCV_ADV); 651 if (tflags & V_TF_RECV_SCALE(1)) 652 snd_wnd <<= GET_TCB_FIELD(tcb, RCV_SCALE); 653 if (GET_TCB_FIELD(tcb, SND_CWND) < snd_wnd) 654 sample |= TS_CWND_LIMITED; /* maybe due to CWND */ 655 } 656 657 if (tflags & V_TF_CCTRL_ECN(1)) { 658 659 /* 660 * CE marker on incoming IP hdr, echoing ECE back in the TCP 661 * hdr. Indicates congestion somewhere on the way from the peer 662 * to this node. 663 */ 664 if (tflags & V_TF_CCTRL_ECE(1)) 665 sample |= TS_ECN_ECE; 666 667 /* 668 * ECE seen and CWR sent (or about to be sent). Might indicate 669 * congestion on the way to the peer. This node is reducing its 670 * congestion window in response. 671 */ 672 if (tflags & (V_TF_CCTRL_CWR(1) | V_TF_CCTRL_RFR(1))) 673 sample |= TS_ECN_CWR; 674 } 675 676 te->te_sample[te->te_pidx] = sample; 677 if (++te->te_pidx == nitems(te->te_sample)) 678 te->te_pidx = 0; 679 memcpy(te->te_tcb, tcb, TCB_SIZE); 680 te->te_flags |= TE_ACTIVE; 681 } 682 683 static int 684 do_get_tcb_rpl(struct sge_iq *iq, const struct rss_header *rss, struct mbuf *m) 685 { 686 struct adapter *sc = iq->adapter; 687 const struct cpl_get_tcb_rpl *cpl = mtod(m, const void *); 688 const uint64_t *tcb = (const uint64_t *)(const void *)(cpl + 1); 689 struct tcb_histent *te; 690 const u_int tid = GET_TID(cpl); 691 bool remove; 692 693 remove = GET_TCB_FIELD(tcb, T_STATE) == TCPS_CLOSED; 694 te = lookup_tcb_histent(sc, tid, remove); 695 if (te == NULL) { 696 /* Not in the history. Who issued the GET_TCB for this? */ 697 device_printf(sc->dev, "tcb %u: flags 0x%016jx, state %u, " 698 "srtt %u, sscale %u, rscale %u, cookie 0x%x\n", tid, 699 (uintmax_t)get_tcb_tflags(tcb), GET_TCB_FIELD(tcb, T_STATE), 700 GET_TCB_FIELD(tcb, T_SRTT), GET_TCB_FIELD(tcb, SND_SCALE), 701 GET_TCB_FIELD(tcb, RCV_SCALE), cpl->cookie); 702 goto done; 703 } 704 705 MPASS(te->te_flags & TE_RPL_PENDING); 706 te->te_flags &= ~TE_RPL_PENDING; 707 if (remove) { 708 remove_tcb_histent(te); 709 } else { 710 update_tcb_histent(te, tcb); 711 callout_reset(&te->te_callout, hz / 10, request_tcb, te); 712 release_tcb_histent(te); 713 } 714 done: 715 m_freem(m); 716 return (0); 717 } 718 719 static void 720 fill_tcp_info_from_tcb(struct adapter *sc, uint64_t *tcb, struct tcp_info *ti) 721 { 722 uint32_t v; 723 724 ti->tcpi_state = GET_TCB_FIELD(tcb, T_STATE); 725 726 v = GET_TCB_FIELD(tcb, T_SRTT); 727 ti->tcpi_rtt = tcp_ticks_to_us(sc, v); 728 729 v = GET_TCB_FIELD(tcb, T_RTTVAR); 730 ti->tcpi_rttvar = tcp_ticks_to_us(sc, v); 731 732 ti->tcpi_snd_ssthresh = GET_TCB_FIELD(tcb, SND_SSTHRESH); 733 ti->tcpi_snd_cwnd = GET_TCB_FIELD(tcb, SND_CWND); 734 ti->tcpi_rcv_nxt = GET_TCB_FIELD(tcb, RCV_NXT); 735 736 v = GET_TCB_FIELD(tcb, TX_MAX); 737 ti->tcpi_snd_nxt = v - GET_TCB_FIELD(tcb, SND_NXT_RAW); 738 739 /* Receive window being advertised by us. */ 740 ti->tcpi_rcv_wscale = GET_TCB_FIELD(tcb, SND_SCALE); /* Yes, SND. */ 741 ti->tcpi_rcv_space = GET_TCB_FIELD(tcb, RCV_WND); 742 743 /* Send window */ 744 ti->tcpi_snd_wscale = GET_TCB_FIELD(tcb, RCV_SCALE); /* Yes, RCV. */ 745 ti->tcpi_snd_wnd = GET_TCB_FIELD(tcb, RCV_ADV); 746 if (get_tcb_tflags(tcb) & V_TF_RECV_SCALE(1)) 747 ti->tcpi_snd_wnd <<= ti->tcpi_snd_wscale; 748 else 749 ti->tcpi_snd_wscale = 0; 750 751 } 752 753 static void 754 fill_tcp_info_from_history(struct adapter *sc, struct tcb_histent *te, 755 struct tcp_info *ti) 756 { 757 758 fill_tcp_info_from_tcb(sc, te->te_tcb, ti); 759 } 760 761 /* 762 * Reads the TCB for the given tid using a memory window and copies it to 'buf' 763 * in the same format as CPL_GET_TCB_RPL. 764 */ 765 static void 766 read_tcb_using_memwin(struct adapter *sc, u_int tid, uint64_t *buf) 767 { 768 int i, j, k, rc; 769 uint32_t addr; 770 u_char *tcb, tmp; 771 772 MPASS(tid >= sc->tids.tid_base); 773 MPASS(tid - sc->tids.tid_base < sc->tids.ntids); 774 775 addr = t4_read_reg(sc, A_TP_CMM_TCB_BASE) + tid * TCB_SIZE; 776 rc = read_via_memwin(sc, 2, addr, (uint32_t *)buf, TCB_SIZE); 777 if (rc != 0) 778 return; 779 780 tcb = (u_char *)buf; 781 for (i = 0, j = TCB_SIZE - 16; i < j; i += 16, j -= 16) { 782 for (k = 0; k < 16; k++) { 783 tmp = tcb[i + k]; 784 tcb[i + k] = tcb[j + k]; 785 tcb[j + k] = tmp; 786 } 787 } 788 } 789 790 static void 791 fill_tcp_info(struct adapter *sc, u_int tid, struct tcp_info *ti) 792 { 793 uint64_t tcb[TCB_SIZE / sizeof(uint64_t)]; 794 struct tcb_histent *te; 795 796 ti->tcpi_toe_tid = tid; 797 te = lookup_tcb_histent(sc, tid, false); 798 if (te != NULL) { 799 fill_tcp_info_from_history(sc, te, ti); 800 release_tcb_histent(te); 801 } else { 802 if (!(sc->debug_flags & DF_DISABLE_TCB_CACHE)) { 803 /* XXX: tell firmware to flush TCB cache. */ 804 } 805 read_tcb_using_memwin(sc, tid, tcb); 806 fill_tcp_info_from_tcb(sc, tcb, ti); 807 } 808 } 809 810 /* 811 * Called by the kernel to allow the TOE driver to "refine" values filled up in 812 * the tcp_info for an offloaded connection. 813 */ 814 static void 815 t4_tcp_info(struct toedev *tod, struct tcpcb *tp, struct tcp_info *ti) 816 { 817 struct adapter *sc = tod->tod_softc; 818 struct toepcb *toep = tp->t_toe; 819 820 INP_WLOCK_ASSERT(tptoinpcb(tp)); 821 MPASS(ti != NULL); 822 823 fill_tcp_info(sc, toep->tid, ti); 824 } 825 826 #ifdef KERN_TLS 827 static int 828 t4_alloc_tls_session(struct toedev *tod, struct tcpcb *tp, 829 struct ktls_session *tls, int direction) 830 { 831 struct toepcb *toep = tp->t_toe; 832 833 INP_WLOCK_ASSERT(tptoinpcb(tp)); 834 MPASS(tls != NULL); 835 836 return (tls_alloc_ktls(toep, tls, direction)); 837 } 838 #endif 839 840 /* SET_TCB_FIELD sent as a ULP command looks like this */ 841 #define LEN__SET_TCB_FIELD_ULP (sizeof(struct ulp_txpkt) + \ 842 sizeof(struct ulptx_idata) + sizeof(struct cpl_set_tcb_field_core)) 843 844 static void * 845 mk_set_tcb_field_ulp(struct ulp_txpkt *ulpmc, uint64_t word, uint64_t mask, 846 uint64_t val, uint32_t tid) 847 { 848 struct ulptx_idata *ulpsc; 849 struct cpl_set_tcb_field_core *req; 850 851 ulpmc->cmd_dest = htonl(V_ULPTX_CMD(ULP_TX_PKT) | V_ULP_TXPKT_DEST(0)); 852 ulpmc->len = htobe32(howmany(LEN__SET_TCB_FIELD_ULP, 16)); 853 854 ulpsc = (struct ulptx_idata *)(ulpmc + 1); 855 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_IMM)); 856 ulpsc->len = htobe32(sizeof(*req)); 857 858 req = (struct cpl_set_tcb_field_core *)(ulpsc + 1); 859 OPCODE_TID(req) = htobe32(MK_OPCODE_TID(CPL_SET_TCB_FIELD, tid)); 860 req->reply_ctrl = htobe16(V_NO_REPLY(1)); 861 req->word_cookie = htobe16(V_WORD(word) | V_COOKIE(0)); 862 req->mask = htobe64(mask); 863 req->val = htobe64(val); 864 865 ulpsc = (struct ulptx_idata *)(req + 1); 866 if (LEN__SET_TCB_FIELD_ULP % 16) { 867 ulpsc->cmd_more = htobe32(V_ULPTX_CMD(ULP_TX_SC_NOOP)); 868 ulpsc->len = htobe32(0); 869 return (ulpsc + 1); 870 } 871 return (ulpsc); 872 } 873 874 static void 875 send_mss_flowc_wr(struct adapter *sc, struct toepcb *toep) 876 { 877 struct wrq_cookie cookie; 878 struct fw_flowc_wr *flowc; 879 struct ofld_tx_sdesc *txsd; 880 const int flowclen = sizeof(*flowc) + sizeof(struct fw_flowc_mnemval); 881 const int flowclen16 = howmany(flowclen, 16); 882 883 if (toep->tx_credits < flowclen16 || toep->txsd_avail == 0) { 884 CH_ERR(sc, "%s: tid %u out of tx credits (%d, %d).\n", __func__, 885 toep->tid, toep->tx_credits, toep->txsd_avail); 886 return; 887 } 888 889 flowc = start_wrq_wr(&toep->ofld_txq->wrq, flowclen16, &cookie); 890 if (__predict_false(flowc == NULL)) { 891 CH_ERR(sc, "ENOMEM in %s for tid %u.\n", __func__, toep->tid); 892 return; 893 } 894 flowc->op_to_nparams = htobe32(V_FW_WR_OP(FW_FLOWC_WR) | 895 V_FW_FLOWC_WR_NPARAMS(1)); 896 flowc->flowid_len16 = htonl(V_FW_WR_LEN16(flowclen16) | 897 V_FW_WR_FLOWID(toep->tid)); 898 flowc->mnemval[0].mnemonic = FW_FLOWC_MNEM_MSS; 899 flowc->mnemval[0].val = htobe32(toep->params.emss); 900 901 txsd = &toep->txsd[toep->txsd_pidx]; 902 txsd->tx_credits = flowclen16; 903 txsd->plen = 0; 904 toep->tx_credits -= txsd->tx_credits; 905 if (__predict_false(++toep->txsd_pidx == toep->txsd_total)) 906 toep->txsd_pidx = 0; 907 toep->txsd_avail--; 908 commit_wrq_wr(&toep->ofld_txq->wrq, flowc, &cookie); 909 } 910 911 static void 912 t4_pmtu_update(struct toedev *tod, struct tcpcb *tp, tcp_seq seq, int mtu) 913 { 914 struct work_request_hdr *wrh; 915 struct ulp_txpkt *ulpmc; 916 int idx, len; 917 struct wrq_cookie cookie; 918 struct inpcb *inp = tptoinpcb(tp); 919 struct toepcb *toep = tp->t_toe; 920 struct adapter *sc = td_adapter(toep->td); 921 unsigned short *mtus = &sc->params.mtus[0]; 922 923 INP_WLOCK_ASSERT(inp); 924 MPASS(mtu > 0); /* kernel is supposed to provide something usable. */ 925 926 /* tp->snd_una and snd_max are in host byte order too. */ 927 seq = be32toh(seq); 928 929 CTR6(KTR_CXGBE, "%s: tid %d, seq 0x%08x, mtu %u, mtu_idx %u (%d)", 930 __func__, toep->tid, seq, mtu, toep->params.mtu_idx, 931 mtus[toep->params.mtu_idx]); 932 933 if (ulp_mode(toep) == ULP_MODE_NONE && /* XXX: Read TCB otherwise? */ 934 (SEQ_LT(seq, tp->snd_una) || SEQ_GEQ(seq, tp->snd_max))) { 935 CTR5(KTR_CXGBE, 936 "%s: tid %d, seq 0x%08x not in range [0x%08x, 0x%08x).", 937 __func__, toep->tid, seq, tp->snd_una, tp->snd_max); 938 return; 939 } 940 941 /* Find the best mtu_idx for the suggested MTU. */ 942 for (idx = 0; idx < NMTUS - 1 && mtus[idx + 1] <= mtu; idx++) 943 continue; 944 if (idx >= toep->params.mtu_idx) 945 return; /* Never increase the PMTU (just like the kernel). */ 946 947 /* 948 * We'll send a compound work request with 2 SET_TCB_FIELDs -- the first 949 * one updates the mtu_idx and the second one triggers a retransmit. 950 */ 951 len = sizeof(*wrh) + 2 * roundup2(LEN__SET_TCB_FIELD_ULP, 16); 952 wrh = start_wrq_wr(toep->ctrlq, howmany(len, 16), &cookie); 953 if (wrh == NULL) { 954 CH_ERR(sc, "failed to change mtu_idx of tid %d (%u -> %u).\n", 955 toep->tid, toep->params.mtu_idx, idx); 956 return; 957 } 958 INIT_ULPTX_WRH(wrh, len, 1, 0); /* atomic */ 959 ulpmc = (struct ulp_txpkt *)(wrh + 1); 960 ulpmc = mk_set_tcb_field_ulp(ulpmc, W_TCB_T_MAXSEG, 961 V_TCB_T_MAXSEG(M_TCB_T_MAXSEG), V_TCB_T_MAXSEG(idx), toep->tid); 962 ulpmc = mk_set_tcb_field_ulp(ulpmc, W_TCB_TIMESTAMP, 963 V_TCB_TIMESTAMP(0x7FFFFULL << 11), 0, toep->tid); 964 commit_wrq_wr(toep->ctrlq, wrh, &cookie); 965 966 /* Update the software toepcb and tcpcb. */ 967 toep->params.mtu_idx = idx; 968 tp->t_maxseg = mtus[toep->params.mtu_idx]; 969 if (inp->inp_inc.inc_flags & INC_ISIPV6) 970 tp->t_maxseg -= sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 971 else 972 tp->t_maxseg -= sizeof(struct ip) + sizeof(struct tcphdr); 973 toep->params.emss = tp->t_maxseg; 974 if (tp->t_flags & TF_RCVD_TSTMP) 975 toep->params.emss -= TCPOLEN_TSTAMP_APPA; 976 977 /* Update the firmware flowc. */ 978 send_mss_flowc_wr(sc, toep); 979 980 /* Update the MTU in the kernel's hostcache. */ 981 if (sc->tt.update_hc_on_pmtu_change != 0) { 982 struct in_conninfo inc = {0}; 983 984 inc.inc_fibnum = inp->inp_inc.inc_fibnum; 985 if (inp->inp_inc.inc_flags & INC_ISIPV6) { 986 inc.inc_flags |= INC_ISIPV6; 987 inc.inc6_faddr = inp->inp_inc.inc6_faddr; 988 } else { 989 inc.inc_faddr = inp->inp_inc.inc_faddr; 990 } 991 tcp_hc_updatemtu(&inc, mtu); 992 } 993 994 CTR6(KTR_CXGBE, "%s: tid %d, mtu_idx %u (%u), t_maxseg %u, emss %u", 995 __func__, toep->tid, toep->params.mtu_idx, 996 mtus[toep->params.mtu_idx], tp->t_maxseg, toep->params.emss); 997 } 998 999 /* 1000 * The TOE driver will not receive any more CPLs for the tid associated with the 1001 * toepcb; release the hold on the inpcb. 1002 */ 1003 void 1004 final_cpl_received(struct toepcb *toep) 1005 { 1006 struct inpcb *inp = toep->inp; 1007 bool need_wakeup; 1008 1009 KASSERT(inp != NULL, ("%s: inp is NULL", __func__)); 1010 INP_WLOCK_ASSERT(inp); 1011 KASSERT(toep->flags & TPF_CPL_PENDING, 1012 ("%s: CPL not pending already?", __func__)); 1013 1014 CTR6(KTR_CXGBE, "%s: tid %d, toep %p (0x%x), inp %p (0x%x)", 1015 __func__, toep->tid, toep, toep->flags, inp, inp->inp_flags); 1016 1017 if (ulp_mode(toep) == ULP_MODE_TCPDDP) 1018 release_ddp_resources(toep); 1019 toep->inp = NULL; 1020 need_wakeup = (toep->flags & TPF_WAITING_FOR_FINAL) != 0; 1021 toep->flags &= ~(TPF_CPL_PENDING | TPF_WAITING_FOR_FINAL); 1022 mbufq_drain(&toep->ulp_pduq); 1023 mbufq_drain(&toep->ulp_pdu_reclaimq); 1024 1025 if (!(toep->flags & TPF_ATTACHED)) 1026 release_offload_resources(toep); 1027 1028 if (!in_pcbrele_wlocked(inp)) 1029 INP_WUNLOCK(inp); 1030 1031 if (need_wakeup) { 1032 struct mtx *lock = mtx_pool_find(mtxpool_sleep, toep); 1033 1034 mtx_lock(lock); 1035 wakeup(toep); 1036 mtx_unlock(lock); 1037 } 1038 } 1039 1040 void 1041 insert_tid(struct adapter *sc, int tid, void *ctx, int ntids) 1042 { 1043 struct tid_info *t = &sc->tids; 1044 1045 MPASS(tid >= t->tid_base); 1046 MPASS(tid - t->tid_base < t->ntids); 1047 1048 t->tid_tab[tid - t->tid_base] = ctx; 1049 atomic_add_int(&t->tids_in_use, ntids); 1050 } 1051 1052 void * 1053 lookup_tid(struct adapter *sc, int tid) 1054 { 1055 struct tid_info *t = &sc->tids; 1056 1057 return (t->tid_tab[tid - t->tid_base]); 1058 } 1059 1060 void 1061 update_tid(struct adapter *sc, int tid, void *ctx) 1062 { 1063 struct tid_info *t = &sc->tids; 1064 1065 t->tid_tab[tid - t->tid_base] = ctx; 1066 } 1067 1068 void 1069 remove_tid(struct adapter *sc, int tid, int ntids) 1070 { 1071 struct tid_info *t = &sc->tids; 1072 1073 t->tid_tab[tid - t->tid_base] = NULL; 1074 atomic_subtract_int(&t->tids_in_use, ntids); 1075 } 1076 1077 /* 1078 * What mtu_idx to use, given a 4-tuple. Note that both s->mss and tcp_mssopt 1079 * have the MSS that we should advertise in our SYN. Advertised MSS doesn't 1080 * account for any TCP options so the effective MSS (only payload, no headers or 1081 * options) could be different. 1082 */ 1083 static int 1084 find_best_mtu_idx(struct adapter *sc, struct in_conninfo *inc, 1085 struct offload_settings *s) 1086 { 1087 unsigned short *mtus = &sc->params.mtus[0]; 1088 int i, mss, mtu; 1089 1090 MPASS(inc != NULL); 1091 1092 mss = s->mss > 0 ? s->mss : tcp_mssopt(inc); 1093 if (inc->inc_flags & INC_ISIPV6) 1094 mtu = mss + sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 1095 else 1096 mtu = mss + sizeof(struct ip) + sizeof(struct tcphdr); 1097 1098 for (i = 0; i < NMTUS - 1 && mtus[i + 1] <= mtu; i++) 1099 continue; 1100 1101 return (i); 1102 } 1103 1104 /* 1105 * Determine the receive window size for a socket. 1106 */ 1107 u_long 1108 select_rcv_wnd(struct socket *so) 1109 { 1110 unsigned long wnd; 1111 1112 SOCKBUF_LOCK_ASSERT(&so->so_rcv); 1113 1114 wnd = sbspace(&so->so_rcv); 1115 if (wnd < MIN_RCV_WND) 1116 wnd = MIN_RCV_WND; 1117 1118 return min(wnd, MAX_RCV_WND); 1119 } 1120 1121 int 1122 select_rcv_wscale(void) 1123 { 1124 int wscale = 0; 1125 unsigned long space = sb_max; 1126 1127 if (space > MAX_RCV_WND) 1128 space = MAX_RCV_WND; 1129 1130 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < space) 1131 wscale++; 1132 1133 return (wscale); 1134 } 1135 1136 __be64 1137 calc_options0(struct vi_info *vi, struct conn_params *cp) 1138 { 1139 uint64_t opt0 = 0; 1140 1141 opt0 |= F_TCAM_BYPASS; 1142 1143 MPASS(cp->wscale >= 0 && cp->wscale <= M_WND_SCALE); 1144 opt0 |= V_WND_SCALE(cp->wscale); 1145 1146 MPASS(cp->mtu_idx >= 0 && cp->mtu_idx < NMTUS); 1147 opt0 |= V_MSS_IDX(cp->mtu_idx); 1148 1149 MPASS(cp->ulp_mode >= 0 && cp->ulp_mode <= M_ULP_MODE); 1150 opt0 |= V_ULP_MODE(cp->ulp_mode); 1151 1152 MPASS(cp->opt0_bufsize >= 0 && cp->opt0_bufsize <= M_RCV_BUFSIZ); 1153 opt0 |= V_RCV_BUFSIZ(cp->opt0_bufsize); 1154 1155 MPASS(cp->l2t_idx >= 0 && cp->l2t_idx < vi->adapter->vres.l2t.size); 1156 opt0 |= V_L2T_IDX(cp->l2t_idx); 1157 1158 opt0 |= V_SMAC_SEL(vi->smt_idx); 1159 opt0 |= V_TX_CHAN(vi->pi->tx_chan); 1160 1161 MPASS(cp->keepalive == 0 || cp->keepalive == 1); 1162 opt0 |= V_KEEP_ALIVE(cp->keepalive); 1163 1164 MPASS(cp->nagle == 0 || cp->nagle == 1); 1165 opt0 |= V_NAGLE(cp->nagle); 1166 1167 return (htobe64(opt0)); 1168 } 1169 1170 __be32 1171 calc_options2(struct vi_info *vi, struct conn_params *cp) 1172 { 1173 uint32_t opt2 = 0; 1174 struct port_info *pi = vi->pi; 1175 struct adapter *sc = pi->adapter; 1176 1177 /* 1178 * rx flow control, rx coalesce, congestion control, and tx pace are all 1179 * explicitly set by the driver. On T5+ the ISS is also set by the 1180 * driver to the value picked by the kernel. 1181 */ 1182 if (is_t4(sc)) { 1183 opt2 |= F_RX_FC_VALID | F_RX_COALESCE_VALID; 1184 opt2 |= F_CONG_CNTRL_VALID | F_PACE_VALID; 1185 } else { 1186 opt2 |= F_T5_OPT_2_VALID; /* all 4 valid */ 1187 opt2 |= F_T5_ISS; /* ISS provided in CPL */ 1188 } 1189 1190 MPASS(cp->sack == 0 || cp->sack == 1); 1191 opt2 |= V_SACK_EN(cp->sack); 1192 1193 MPASS(cp->tstamp == 0 || cp->tstamp == 1); 1194 opt2 |= V_TSTAMPS_EN(cp->tstamp); 1195 1196 if (cp->wscale > 0) 1197 opt2 |= F_WND_SCALE_EN; 1198 1199 MPASS(cp->ecn == 0 || cp->ecn == 1); 1200 opt2 |= V_CCTRL_ECN(cp->ecn); 1201 1202 /* XXX: F_RX_CHANNEL for multiple rx c-chan support goes here. */ 1203 1204 opt2 |= V_TX_QUEUE(sc->params.tp.tx_modq[pi->tx_chan]); 1205 opt2 |= V_PACE(0); 1206 opt2 |= F_RSS_QUEUE_VALID; 1207 opt2 |= V_RSS_QUEUE(sc->sge.ofld_rxq[cp->rxq_idx].iq.abs_id); 1208 1209 MPASS(cp->cong_algo >= 0 && cp->cong_algo <= M_CONG_CNTRL); 1210 opt2 |= V_CONG_CNTRL(cp->cong_algo); 1211 1212 MPASS(cp->rx_coalesce == 0 || cp->rx_coalesce == 1); 1213 if (cp->rx_coalesce == 1) 1214 opt2 |= V_RX_COALESCE(M_RX_COALESCE); 1215 1216 opt2 |= V_RX_FC_DDP(0) | V_RX_FC_DISABLE(0); 1217 #ifdef USE_DDP_RX_FLOW_CONTROL 1218 if (cp->ulp_mode == ULP_MODE_TCPDDP) 1219 opt2 |= F_RX_FC_DDP; 1220 #endif 1221 1222 return (htobe32(opt2)); 1223 } 1224 1225 uint64_t 1226 select_ntuple(struct vi_info *vi, struct l2t_entry *e) 1227 { 1228 struct adapter *sc = vi->adapter; 1229 struct tp_params *tp = &sc->params.tp; 1230 uint64_t ntuple = 0; 1231 1232 /* 1233 * Initialize each of the fields which we care about which are present 1234 * in the Compressed Filter Tuple. 1235 */ 1236 if (tp->vlan_shift >= 0 && EVL_VLANOFTAG(e->vlan) != CPL_L2T_VLAN_NONE) 1237 ntuple |= (uint64_t)(F_FT_VLAN_VLD | e->vlan) << tp->vlan_shift; 1238 1239 if (tp->port_shift >= 0) 1240 ntuple |= (uint64_t)e->lport << tp->port_shift; 1241 1242 if (tp->protocol_shift >= 0) 1243 ntuple |= (uint64_t)IPPROTO_TCP << tp->protocol_shift; 1244 1245 if (tp->vnic_shift >= 0 && tp->vnic_mode == FW_VNIC_MODE_PF_VF) { 1246 ntuple |= (uint64_t)(V_FT_VNID_ID_VF(vi->vin) | 1247 V_FT_VNID_ID_PF(sc->pf) | V_FT_VNID_ID_VLD(vi->vfvld)) << 1248 tp->vnic_shift; 1249 } 1250 1251 if (is_t4(sc)) 1252 return (htobe32((uint32_t)ntuple)); 1253 else 1254 return (htobe64(V_FILTER_TUPLE(ntuple))); 1255 } 1256 1257 /* 1258 * Initialize various connection parameters. 1259 */ 1260 void 1261 init_conn_params(struct vi_info *vi , struct offload_settings *s, 1262 struct in_conninfo *inc, struct socket *so, 1263 const struct tcp_options *tcpopt, int16_t l2t_idx, struct conn_params *cp) 1264 { 1265 struct port_info *pi = vi->pi; 1266 struct adapter *sc = pi->adapter; 1267 struct tom_tunables *tt = &sc->tt; 1268 struct inpcb *inp = sotoinpcb(so); 1269 struct tcpcb *tp = intotcpcb(inp); 1270 u_long wnd; 1271 u_int q_idx; 1272 1273 MPASS(s->offload != 0); 1274 1275 /* Congestion control algorithm */ 1276 if (s->cong_algo >= 0) 1277 cp->cong_algo = s->cong_algo & M_CONG_CNTRL; 1278 else if (sc->tt.cong_algorithm >= 0) 1279 cp->cong_algo = tt->cong_algorithm & M_CONG_CNTRL; 1280 else { 1281 struct cc_algo *cc = CC_ALGO(tp); 1282 1283 if (strcasecmp(cc->name, "reno") == 0) 1284 cp->cong_algo = CONG_ALG_RENO; 1285 else if (strcasecmp(cc->name, "tahoe") == 0) 1286 cp->cong_algo = CONG_ALG_TAHOE; 1287 if (strcasecmp(cc->name, "newreno") == 0) 1288 cp->cong_algo = CONG_ALG_NEWRENO; 1289 if (strcasecmp(cc->name, "highspeed") == 0) 1290 cp->cong_algo = CONG_ALG_HIGHSPEED; 1291 else { 1292 /* 1293 * Use newreno in case the algorithm selected by the 1294 * host stack is not supported by the hardware. 1295 */ 1296 cp->cong_algo = CONG_ALG_NEWRENO; 1297 } 1298 } 1299 1300 /* Tx traffic scheduling class. */ 1301 if (s->sched_class >= 0 && s->sched_class < sc->params.nsched_cls) 1302 cp->tc_idx = s->sched_class; 1303 else 1304 cp->tc_idx = -1; 1305 1306 /* Nagle's algorithm. */ 1307 if (s->nagle >= 0) 1308 cp->nagle = s->nagle > 0 ? 1 : 0; 1309 else 1310 cp->nagle = tp->t_flags & TF_NODELAY ? 0 : 1; 1311 1312 /* TCP Keepalive. */ 1313 if (V_tcp_always_keepalive || so_options_get(so) & SO_KEEPALIVE) 1314 cp->keepalive = 1; 1315 else 1316 cp->keepalive = 0; 1317 1318 /* Optimization that's specific to T5 @ 40G. */ 1319 if (tt->tx_align >= 0) 1320 cp->tx_align = tt->tx_align > 0 ? 1 : 0; 1321 else if (chip_id(sc) == CHELSIO_T5 && 1322 (port_top_speed(pi) > 10 || sc->params.nports > 2)) 1323 cp->tx_align = 1; 1324 else 1325 cp->tx_align = 0; 1326 1327 /* ULP mode. */ 1328 if (s->ddp > 0 || 1329 (s->ddp < 0 && sc->tt.ddp && (so_options_get(so) & SO_NO_DDP) == 0)) 1330 cp->ulp_mode = ULP_MODE_TCPDDP; 1331 else 1332 cp->ulp_mode = ULP_MODE_NONE; 1333 1334 /* Rx coalescing. */ 1335 if (s->rx_coalesce >= 0) 1336 cp->rx_coalesce = s->rx_coalesce > 0 ? 1 : 0; 1337 else if (tt->rx_coalesce >= 0) 1338 cp->rx_coalesce = tt->rx_coalesce > 0 ? 1 : 0; 1339 else 1340 cp->rx_coalesce = 1; /* default */ 1341 1342 /* 1343 * Index in the PMTU table. This controls the MSS that we announce in 1344 * our SYN initially, but after ESTABLISHED it controls the MSS that we 1345 * use to send data. 1346 */ 1347 cp->mtu_idx = find_best_mtu_idx(sc, inc, s); 1348 1349 /* Tx queue for this connection. */ 1350 if (s->txq == QUEUE_RANDOM) 1351 q_idx = arc4random(); 1352 else if (s->txq == QUEUE_ROUNDROBIN) 1353 q_idx = atomic_fetchadd_int(&vi->txq_rr, 1); 1354 else 1355 q_idx = s->txq; 1356 cp->txq_idx = vi->first_ofld_txq + q_idx % vi->nofldtxq; 1357 1358 /* Rx queue for this connection. */ 1359 if (s->rxq == QUEUE_RANDOM) 1360 q_idx = arc4random(); 1361 else if (s->rxq == QUEUE_ROUNDROBIN) 1362 q_idx = atomic_fetchadd_int(&vi->rxq_rr, 1); 1363 else 1364 q_idx = s->rxq; 1365 cp->rxq_idx = vi->first_ofld_rxq + q_idx % vi->nofldrxq; 1366 1367 if (SOLISTENING(so)) { 1368 /* Passive open */ 1369 MPASS(tcpopt != NULL); 1370 1371 /* TCP timestamp option */ 1372 if (tcpopt->tstamp && 1373 (s->tstamp > 0 || (s->tstamp < 0 && V_tcp_do_rfc1323))) 1374 cp->tstamp = 1; 1375 else 1376 cp->tstamp = 0; 1377 1378 /* SACK */ 1379 if (tcpopt->sack && 1380 (s->sack > 0 || (s->sack < 0 && V_tcp_do_sack))) 1381 cp->sack = 1; 1382 else 1383 cp->sack = 0; 1384 1385 /* Receive window scaling. */ 1386 if (tcpopt->wsf > 0 && tcpopt->wsf < 15 && V_tcp_do_rfc1323) 1387 cp->wscale = select_rcv_wscale(); 1388 else 1389 cp->wscale = 0; 1390 1391 /* ECN */ 1392 if (tcpopt->ecn && /* XXX: review. */ 1393 (s->ecn > 0 || (s->ecn < 0 && V_tcp_do_ecn))) 1394 cp->ecn = 1; 1395 else 1396 cp->ecn = 0; 1397 1398 wnd = max(so->sol_sbrcv_hiwat, MIN_RCV_WND); 1399 cp->opt0_bufsize = min(wnd >> 10, M_RCV_BUFSIZ); 1400 1401 if (tt->sndbuf > 0) 1402 cp->sndbuf = tt->sndbuf; 1403 else if (so->sol_sbsnd_flags & SB_AUTOSIZE && 1404 V_tcp_do_autosndbuf) 1405 cp->sndbuf = 256 * 1024; 1406 else 1407 cp->sndbuf = so->sol_sbsnd_hiwat; 1408 } else { 1409 /* Active open */ 1410 1411 /* TCP timestamp option */ 1412 if (s->tstamp > 0 || 1413 (s->tstamp < 0 && (tp->t_flags & TF_REQ_TSTMP))) 1414 cp->tstamp = 1; 1415 else 1416 cp->tstamp = 0; 1417 1418 /* SACK */ 1419 if (s->sack > 0 || 1420 (s->sack < 0 && (tp->t_flags & TF_SACK_PERMIT))) 1421 cp->sack = 1; 1422 else 1423 cp->sack = 0; 1424 1425 /* Receive window scaling */ 1426 if (tp->t_flags & TF_REQ_SCALE) 1427 cp->wscale = select_rcv_wscale(); 1428 else 1429 cp->wscale = 0; 1430 1431 /* ECN */ 1432 if (s->ecn > 0 || (s->ecn < 0 && V_tcp_do_ecn == 1)) 1433 cp->ecn = 1; 1434 else 1435 cp->ecn = 0; 1436 1437 SOCKBUF_LOCK(&so->so_rcv); 1438 wnd = max(select_rcv_wnd(so), MIN_RCV_WND); 1439 SOCKBUF_UNLOCK(&so->so_rcv); 1440 cp->opt0_bufsize = min(wnd >> 10, M_RCV_BUFSIZ); 1441 1442 if (tt->sndbuf > 0) 1443 cp->sndbuf = tt->sndbuf; 1444 else { 1445 SOCKBUF_LOCK(&so->so_snd); 1446 if (so->so_snd.sb_flags & SB_AUTOSIZE && 1447 V_tcp_do_autosndbuf) 1448 cp->sndbuf = 256 * 1024; 1449 else 1450 cp->sndbuf = so->so_snd.sb_hiwat; 1451 SOCKBUF_UNLOCK(&so->so_snd); 1452 } 1453 } 1454 1455 cp->l2t_idx = l2t_idx; 1456 1457 /* This will be initialized on ESTABLISHED. */ 1458 cp->emss = 0; 1459 } 1460 1461 int 1462 negative_advice(int status) 1463 { 1464 1465 return (status == CPL_ERR_RTX_NEG_ADVICE || 1466 status == CPL_ERR_PERSIST_NEG_ADVICE || 1467 status == CPL_ERR_KEEPALV_NEG_ADVICE); 1468 } 1469 1470 static int 1471 alloc_tid_tab(struct tid_info *t, int flags) 1472 { 1473 1474 MPASS(t->ntids > 0); 1475 MPASS(t->tid_tab == NULL); 1476 1477 t->tid_tab = malloc(t->ntids * sizeof(*t->tid_tab), M_CXGBE, 1478 M_ZERO | flags); 1479 if (t->tid_tab == NULL) 1480 return (ENOMEM); 1481 atomic_store_rel_int(&t->tids_in_use, 0); 1482 1483 return (0); 1484 } 1485 1486 static void 1487 free_tid_tab(struct tid_info *t) 1488 { 1489 1490 KASSERT(t->tids_in_use == 0, 1491 ("%s: %d tids still in use.", __func__, t->tids_in_use)); 1492 1493 free(t->tid_tab, M_CXGBE); 1494 t->tid_tab = NULL; 1495 } 1496 1497 static int 1498 alloc_stid_tab(struct tid_info *t, int flags) 1499 { 1500 1501 MPASS(t->nstids > 0); 1502 MPASS(t->stid_tab == NULL); 1503 1504 t->stid_tab = malloc(t->nstids * sizeof(*t->stid_tab), M_CXGBE, 1505 M_ZERO | flags); 1506 if (t->stid_tab == NULL) 1507 return (ENOMEM); 1508 mtx_init(&t->stid_lock, "stid lock", NULL, MTX_DEF); 1509 t->stids_in_use = 0; 1510 TAILQ_INIT(&t->stids); 1511 t->nstids_free_head = t->nstids; 1512 1513 return (0); 1514 } 1515 1516 static void 1517 free_stid_tab(struct tid_info *t) 1518 { 1519 1520 KASSERT(t->stids_in_use == 0, 1521 ("%s: %d tids still in use.", __func__, t->stids_in_use)); 1522 1523 if (mtx_initialized(&t->stid_lock)) 1524 mtx_destroy(&t->stid_lock); 1525 free(t->stid_tab, M_CXGBE); 1526 t->stid_tab = NULL; 1527 } 1528 1529 static void 1530 free_tid_tabs(struct tid_info *t) 1531 { 1532 1533 free_tid_tab(t); 1534 free_stid_tab(t); 1535 } 1536 1537 static int 1538 alloc_tid_tabs(struct tid_info *t) 1539 { 1540 int rc; 1541 1542 rc = alloc_tid_tab(t, M_NOWAIT); 1543 if (rc != 0) 1544 goto failed; 1545 1546 rc = alloc_stid_tab(t, M_NOWAIT); 1547 if (rc != 0) 1548 goto failed; 1549 1550 return (0); 1551 failed: 1552 free_tid_tabs(t); 1553 return (rc); 1554 } 1555 1556 static inline void 1557 alloc_tcb_history(struct adapter *sc, struct tom_data *td) 1558 { 1559 1560 if (sc->tids.ntids == 0 || sc->tids.ntids > 1024) 1561 return; 1562 rw_init(&td->tcb_history_lock, "TCB history"); 1563 td->tcb_history = malloc(sc->tids.ntids * sizeof(*td->tcb_history), 1564 M_CXGBE, M_ZERO | M_NOWAIT); 1565 td->dupack_threshold = G_DUPACKTHRESH(t4_read_reg(sc, A_TP_PARA_REG0)); 1566 } 1567 1568 static inline void 1569 free_tcb_history(struct adapter *sc, struct tom_data *td) 1570 { 1571 #ifdef INVARIANTS 1572 int i; 1573 1574 if (td->tcb_history != NULL) { 1575 for (i = 0; i < sc->tids.ntids; i++) { 1576 MPASS(td->tcb_history[i] == NULL); 1577 } 1578 } 1579 #endif 1580 free(td->tcb_history, M_CXGBE); 1581 if (rw_initialized(&td->tcb_history_lock)) 1582 rw_destroy(&td->tcb_history_lock); 1583 } 1584 1585 static void 1586 free_tom_data(struct adapter *sc, struct tom_data *td) 1587 { 1588 1589 ASSERT_SYNCHRONIZED_OP(sc); 1590 1591 KASSERT(TAILQ_EMPTY(&td->toep_list), 1592 ("%s: TOE PCB list is not empty.", __func__)); 1593 KASSERT(td->lctx_count == 0, 1594 ("%s: lctx hash table is not empty.", __func__)); 1595 1596 t4_free_ppod_region(&td->pr); 1597 1598 if (td->listen_mask != 0) 1599 hashdestroy(td->listen_hash, M_CXGBE, td->listen_mask); 1600 1601 if (mtx_initialized(&td->unsent_wr_lock)) 1602 mtx_destroy(&td->unsent_wr_lock); 1603 if (mtx_initialized(&td->lctx_hash_lock)) 1604 mtx_destroy(&td->lctx_hash_lock); 1605 if (mtx_initialized(&td->toep_list_lock)) 1606 mtx_destroy(&td->toep_list_lock); 1607 1608 free_tcb_history(sc, td); 1609 free_tid_tabs(&sc->tids); 1610 free(td, M_CXGBE); 1611 } 1612 1613 static char * 1614 prepare_pkt(int open_type, uint16_t vtag, struct inpcb *inp, int *pktlen, 1615 int *buflen) 1616 { 1617 char *pkt; 1618 struct tcphdr *th; 1619 int ipv6, len; 1620 const int maxlen = 1621 max(sizeof(struct ether_header), sizeof(struct ether_vlan_header)) + 1622 max(sizeof(struct ip), sizeof(struct ip6_hdr)) + 1623 sizeof(struct tcphdr); 1624 1625 MPASS(open_type == OPEN_TYPE_ACTIVE || open_type == OPEN_TYPE_LISTEN); 1626 1627 pkt = malloc(maxlen, M_CXGBE, M_ZERO | M_NOWAIT); 1628 if (pkt == NULL) 1629 return (NULL); 1630 1631 ipv6 = inp->inp_vflag & INP_IPV6; 1632 len = 0; 1633 1634 if (EVL_VLANOFTAG(vtag) == 0xfff) { 1635 struct ether_header *eh = (void *)pkt; 1636 1637 if (ipv6) 1638 eh->ether_type = htons(ETHERTYPE_IPV6); 1639 else 1640 eh->ether_type = htons(ETHERTYPE_IP); 1641 1642 len += sizeof(*eh); 1643 } else { 1644 struct ether_vlan_header *evh = (void *)pkt; 1645 1646 evh->evl_encap_proto = htons(ETHERTYPE_VLAN); 1647 evh->evl_tag = htons(vtag); 1648 if (ipv6) 1649 evh->evl_proto = htons(ETHERTYPE_IPV6); 1650 else 1651 evh->evl_proto = htons(ETHERTYPE_IP); 1652 1653 len += sizeof(*evh); 1654 } 1655 1656 if (ipv6) { 1657 struct ip6_hdr *ip6 = (void *)&pkt[len]; 1658 1659 ip6->ip6_vfc = IPV6_VERSION; 1660 ip6->ip6_plen = htons(sizeof(struct tcphdr)); 1661 ip6->ip6_nxt = IPPROTO_TCP; 1662 if (open_type == OPEN_TYPE_ACTIVE) { 1663 ip6->ip6_src = inp->in6p_laddr; 1664 ip6->ip6_dst = inp->in6p_faddr; 1665 } else if (open_type == OPEN_TYPE_LISTEN) { 1666 ip6->ip6_src = inp->in6p_laddr; 1667 ip6->ip6_dst = ip6->ip6_src; 1668 } 1669 1670 len += sizeof(*ip6); 1671 } else { 1672 struct ip *ip = (void *)&pkt[len]; 1673 1674 ip->ip_v = IPVERSION; 1675 ip->ip_hl = sizeof(*ip) >> 2; 1676 ip->ip_tos = inp->inp_ip_tos; 1677 ip->ip_len = htons(sizeof(struct ip) + sizeof(struct tcphdr)); 1678 ip->ip_ttl = inp->inp_ip_ttl; 1679 ip->ip_p = IPPROTO_TCP; 1680 if (open_type == OPEN_TYPE_ACTIVE) { 1681 ip->ip_src = inp->inp_laddr; 1682 ip->ip_dst = inp->inp_faddr; 1683 } else if (open_type == OPEN_TYPE_LISTEN) { 1684 ip->ip_src = inp->inp_laddr; 1685 ip->ip_dst = ip->ip_src; 1686 } 1687 1688 len += sizeof(*ip); 1689 } 1690 1691 th = (void *)&pkt[len]; 1692 if (open_type == OPEN_TYPE_ACTIVE) { 1693 th->th_sport = inp->inp_lport; /* network byte order already */ 1694 th->th_dport = inp->inp_fport; /* ditto */ 1695 } else if (open_type == OPEN_TYPE_LISTEN) { 1696 th->th_sport = inp->inp_lport; /* network byte order already */ 1697 th->th_dport = th->th_sport; 1698 } 1699 len += sizeof(th); 1700 1701 *pktlen = *buflen = len; 1702 return (pkt); 1703 } 1704 1705 const struct offload_settings * 1706 lookup_offload_policy(struct adapter *sc, int open_type, struct mbuf *m, 1707 uint16_t vtag, struct inpcb *inp) 1708 { 1709 const struct t4_offload_policy *op; 1710 char *pkt; 1711 struct offload_rule *r; 1712 int i, matched, pktlen, buflen; 1713 static const struct offload_settings allow_offloading_settings = { 1714 .offload = 1, 1715 .rx_coalesce = -1, 1716 .cong_algo = -1, 1717 .sched_class = -1, 1718 .tstamp = -1, 1719 .sack = -1, 1720 .nagle = -1, 1721 .ecn = -1, 1722 .ddp = -1, 1723 .tls = -1, 1724 .txq = QUEUE_RANDOM, 1725 .rxq = QUEUE_RANDOM, 1726 .mss = -1, 1727 }; 1728 static const struct offload_settings disallow_offloading_settings = { 1729 .offload = 0, 1730 /* rest is irrelevant when offload is off. */ 1731 }; 1732 1733 rw_assert(&sc->policy_lock, RA_LOCKED); 1734 1735 /* 1736 * If there's no Connection Offloading Policy attached to the device 1737 * then we need to return a default static policy. If 1738 * "cop_managed_offloading" is true, then we need to disallow 1739 * offloading until a COP is attached to the device. Otherwise we 1740 * allow offloading ... 1741 */ 1742 op = sc->policy; 1743 if (op == NULL) { 1744 if (sc->tt.cop_managed_offloading) 1745 return (&disallow_offloading_settings); 1746 else 1747 return (&allow_offloading_settings); 1748 } 1749 1750 switch (open_type) { 1751 case OPEN_TYPE_ACTIVE: 1752 case OPEN_TYPE_LISTEN: 1753 pkt = prepare_pkt(open_type, vtag, inp, &pktlen, &buflen); 1754 break; 1755 case OPEN_TYPE_PASSIVE: 1756 MPASS(m != NULL); 1757 pkt = mtod(m, char *); 1758 MPASS(*pkt == CPL_PASS_ACCEPT_REQ); 1759 pkt += sizeof(struct cpl_pass_accept_req); 1760 pktlen = m->m_pkthdr.len - sizeof(struct cpl_pass_accept_req); 1761 buflen = m->m_len - sizeof(struct cpl_pass_accept_req); 1762 break; 1763 default: 1764 MPASS(0); 1765 return (&disallow_offloading_settings); 1766 } 1767 1768 if (pkt == NULL || pktlen == 0 || buflen == 0) 1769 return (&disallow_offloading_settings); 1770 1771 matched = 0; 1772 r = &op->rule[0]; 1773 for (i = 0; i < op->nrules; i++, r++) { 1774 if (r->open_type != open_type && 1775 r->open_type != OPEN_TYPE_DONTCARE) { 1776 continue; 1777 } 1778 matched = bpf_filter(r->bpf_prog.bf_insns, pkt, pktlen, buflen); 1779 if (matched) 1780 break; 1781 } 1782 1783 if (open_type == OPEN_TYPE_ACTIVE || open_type == OPEN_TYPE_LISTEN) 1784 free(pkt, M_CXGBE); 1785 1786 return (matched ? &r->settings : &disallow_offloading_settings); 1787 } 1788 1789 static void 1790 reclaim_wr_resources(void *arg, int count) 1791 { 1792 struct tom_data *td = arg; 1793 STAILQ_HEAD(, wrqe) twr_list = STAILQ_HEAD_INITIALIZER(twr_list); 1794 struct cpl_act_open_req *cpl; 1795 u_int opcode, atid, tid; 1796 struct wrqe *wr; 1797 struct adapter *sc = td_adapter(td); 1798 1799 mtx_lock(&td->unsent_wr_lock); 1800 STAILQ_SWAP(&td->unsent_wr_list, &twr_list, wrqe); 1801 mtx_unlock(&td->unsent_wr_lock); 1802 1803 while ((wr = STAILQ_FIRST(&twr_list)) != NULL) { 1804 STAILQ_REMOVE_HEAD(&twr_list, link); 1805 1806 cpl = wrtod(wr); 1807 opcode = GET_OPCODE(cpl); 1808 1809 switch (opcode) { 1810 case CPL_ACT_OPEN_REQ: 1811 case CPL_ACT_OPEN_REQ6: 1812 atid = G_TID_TID(be32toh(OPCODE_TID(cpl))); 1813 CTR2(KTR_CXGBE, "%s: atid %u ", __func__, atid); 1814 act_open_failure_cleanup(sc, atid, EHOSTUNREACH); 1815 free(wr, M_CXGBE); 1816 break; 1817 case CPL_PASS_ACCEPT_RPL: 1818 tid = GET_TID(cpl); 1819 CTR2(KTR_CXGBE, "%s: tid %u ", __func__, tid); 1820 synack_failure_cleanup(sc, tid); 1821 free(wr, M_CXGBE); 1822 break; 1823 default: 1824 log(LOG_ERR, "%s: leaked work request %p, wr_len %d, " 1825 "opcode %x\n", __func__, wr, wr->wr_len, opcode); 1826 /* WR not freed here; go look at it with a debugger. */ 1827 } 1828 } 1829 } 1830 1831 /* 1832 * Ground control to Major TOM 1833 * Commencing countdown, engines on 1834 */ 1835 static int 1836 t4_tom_activate(struct adapter *sc) 1837 { 1838 struct tom_data *td; 1839 struct toedev *tod; 1840 struct vi_info *vi; 1841 int i, rc, v; 1842 1843 ASSERT_SYNCHRONIZED_OP(sc); 1844 1845 /* per-adapter softc for TOM */ 1846 td = malloc(sizeof(*td), M_CXGBE, M_ZERO | M_NOWAIT); 1847 if (td == NULL) 1848 return (ENOMEM); 1849 1850 /* List of TOE PCBs and associated lock */ 1851 mtx_init(&td->toep_list_lock, "PCB list lock", NULL, MTX_DEF); 1852 TAILQ_INIT(&td->toep_list); 1853 1854 /* Listen context */ 1855 mtx_init(&td->lctx_hash_lock, "lctx hash lock", NULL, MTX_DEF); 1856 td->listen_hash = hashinit_flags(LISTEN_HASH_SIZE, M_CXGBE, 1857 &td->listen_mask, HASH_NOWAIT); 1858 1859 /* List of WRs for which L2 resolution failed */ 1860 mtx_init(&td->unsent_wr_lock, "Unsent WR list lock", NULL, MTX_DEF); 1861 STAILQ_INIT(&td->unsent_wr_list); 1862 TASK_INIT(&td->reclaim_wr_resources, 0, reclaim_wr_resources, td); 1863 1864 /* TID tables */ 1865 rc = alloc_tid_tabs(&sc->tids); 1866 if (rc != 0) 1867 goto done; 1868 1869 rc = t4_init_ppod_region(&td->pr, &sc->vres.ddp, 1870 t4_read_reg(sc, A_ULP_RX_TDDP_PSZ), "TDDP page pods"); 1871 if (rc != 0) 1872 goto done; 1873 t4_set_reg_field(sc, A_ULP_RX_TDDP_TAGMASK, 1874 V_TDDPTAGMASK(M_TDDPTAGMASK), td->pr.pr_tag_mask); 1875 1876 alloc_tcb_history(sc, td); 1877 1878 /* toedev ops */ 1879 tod = &td->tod; 1880 init_toedev(tod); 1881 tod->tod_softc = sc; 1882 tod->tod_connect = t4_connect; 1883 tod->tod_listen_start = t4_listen_start; 1884 tod->tod_listen_stop = t4_listen_stop; 1885 tod->tod_rcvd = t4_rcvd; 1886 tod->tod_output = t4_tod_output; 1887 tod->tod_send_rst = t4_send_rst; 1888 tod->tod_send_fin = t4_send_fin; 1889 tod->tod_pcb_detach = t4_pcb_detach; 1890 tod->tod_l2_update = t4_l2_update; 1891 tod->tod_syncache_added = t4_syncache_added; 1892 tod->tod_syncache_removed = t4_syncache_removed; 1893 tod->tod_syncache_respond = t4_syncache_respond; 1894 tod->tod_offload_socket = t4_offload_socket; 1895 tod->tod_ctloutput = t4_ctloutput; 1896 tod->tod_tcp_info = t4_tcp_info; 1897 #ifdef KERN_TLS 1898 tod->tod_alloc_tls_session = t4_alloc_tls_session; 1899 #endif 1900 tod->tod_pmtu_update = t4_pmtu_update; 1901 1902 for_each_port(sc, i) { 1903 for_each_vi(sc->port[i], v, vi) { 1904 SETTOEDEV(vi->ifp, &td->tod); 1905 } 1906 } 1907 1908 sc->tom_softc = td; 1909 register_toedev(sc->tom_softc); 1910 1911 done: 1912 if (rc != 0) 1913 free_tom_data(sc, td); 1914 return (rc); 1915 } 1916 1917 static int 1918 t4_tom_deactivate(struct adapter *sc) 1919 { 1920 int rc = 0; 1921 struct tom_data *td = sc->tom_softc; 1922 1923 ASSERT_SYNCHRONIZED_OP(sc); 1924 1925 if (td == NULL) 1926 return (0); /* XXX. KASSERT? */ 1927 1928 if (sc->offload_map != 0) 1929 return (EBUSY); /* at least one port has IFCAP_TOE enabled */ 1930 1931 if (uld_active(sc, ULD_IWARP) || uld_active(sc, ULD_ISCSI)) 1932 return (EBUSY); /* both iWARP and iSCSI rely on the TOE. */ 1933 1934 mtx_lock(&td->toep_list_lock); 1935 if (!TAILQ_EMPTY(&td->toep_list)) 1936 rc = EBUSY; 1937 mtx_unlock(&td->toep_list_lock); 1938 1939 mtx_lock(&td->lctx_hash_lock); 1940 if (td->lctx_count > 0) 1941 rc = EBUSY; 1942 mtx_unlock(&td->lctx_hash_lock); 1943 1944 taskqueue_drain(taskqueue_thread, &td->reclaim_wr_resources); 1945 mtx_lock(&td->unsent_wr_lock); 1946 if (!STAILQ_EMPTY(&td->unsent_wr_list)) 1947 rc = EBUSY; 1948 mtx_unlock(&td->unsent_wr_lock); 1949 1950 if (rc == 0) { 1951 unregister_toedev(sc->tom_softc); 1952 free_tom_data(sc, td); 1953 sc->tom_softc = NULL; 1954 } 1955 1956 return (rc); 1957 } 1958 1959 static int 1960 t4_aio_queue_tom(struct socket *so, struct kaiocb *job) 1961 { 1962 struct tcpcb *tp = sototcpcb(so); 1963 struct toepcb *toep = tp->t_toe; 1964 int error; 1965 1966 /* 1967 * No lock is needed as TOE sockets never change between 1968 * active and passive. 1969 */ 1970 if (SOLISTENING(so)) 1971 return (EINVAL); 1972 1973 if (ulp_mode(toep) == ULP_MODE_TCPDDP) { 1974 error = t4_aio_queue_ddp(so, job); 1975 if (error != EOPNOTSUPP) 1976 return (error); 1977 } 1978 1979 return (t4_aio_queue_aiotx(so, job)); 1980 } 1981 1982 static int 1983 t4_tom_mod_load(void) 1984 { 1985 /* CPL handlers */ 1986 t4_register_cpl_handler(CPL_GET_TCB_RPL, do_get_tcb_rpl); 1987 t4_register_shared_cpl_handler(CPL_L2T_WRITE_RPL, do_l2t_write_rpl2, 1988 CPL_COOKIE_TOM); 1989 t4_init_connect_cpl_handlers(); 1990 t4_init_listen_cpl_handlers(); 1991 t4_init_cpl_io_handlers(); 1992 1993 t4_ddp_mod_load(); 1994 t4_tls_mod_load(); 1995 1996 bcopy(&tcp_protosw, &toe_protosw, sizeof(toe_protosw)); 1997 toe_protosw.pr_aio_queue = t4_aio_queue_tom; 1998 1999 bcopy(&tcp6_protosw, &toe6_protosw, sizeof(toe6_protosw)); 2000 toe6_protosw.pr_aio_queue = t4_aio_queue_tom; 2001 2002 return (t4_register_uld(&tom_uld_info)); 2003 } 2004 2005 static void 2006 tom_uninit(struct adapter *sc, void *arg __unused) 2007 { 2008 if (begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4tomun")) 2009 return; 2010 2011 /* Try to free resources (works only if no port has IFCAP_TOE) */ 2012 if (uld_active(sc, ULD_TOM)) 2013 t4_deactivate_uld(sc, ULD_TOM); 2014 2015 end_synchronized_op(sc, 0); 2016 } 2017 2018 static int 2019 t4_tom_mod_unload(void) 2020 { 2021 t4_iterate(tom_uninit, NULL); 2022 2023 if (t4_unregister_uld(&tom_uld_info) == EBUSY) 2024 return (EBUSY); 2025 2026 t4_tls_mod_unload(); 2027 t4_ddp_mod_unload(); 2028 2029 t4_uninit_connect_cpl_handlers(); 2030 t4_uninit_listen_cpl_handlers(); 2031 t4_uninit_cpl_io_handlers(); 2032 t4_register_shared_cpl_handler(CPL_L2T_WRITE_RPL, NULL, CPL_COOKIE_TOM); 2033 t4_register_cpl_handler(CPL_GET_TCB_RPL, NULL); 2034 2035 return (0); 2036 } 2037 #endif /* TCP_OFFLOAD */ 2038 2039 static int 2040 t4_tom_modevent(module_t mod, int cmd, void *arg) 2041 { 2042 int rc = 0; 2043 2044 #ifdef TCP_OFFLOAD 2045 switch (cmd) { 2046 case MOD_LOAD: 2047 rc = t4_tom_mod_load(); 2048 break; 2049 2050 case MOD_UNLOAD: 2051 rc = t4_tom_mod_unload(); 2052 break; 2053 2054 default: 2055 rc = EINVAL; 2056 } 2057 #else 2058 printf("t4_tom: compiled without TCP_OFFLOAD support.\n"); 2059 rc = EOPNOTSUPP; 2060 #endif 2061 return (rc); 2062 } 2063 2064 static moduledata_t t4_tom_moddata= { 2065 "t4_tom", 2066 t4_tom_modevent, 2067 0 2068 }; 2069 2070 MODULE_VERSION(t4_tom, 1); 2071 MODULE_DEPEND(t4_tom, toecore, 1, 1, 1); 2072 MODULE_DEPEND(t4_tom, t4nex, 1, 1, 1); 2073 DECLARE_MODULE(t4_tom, t4_tom_moddata, SI_SUB_EXEC, SI_ORDER_ANY); 2074