t4_tom.c revision 312116
1/*- 2 * Copyright (c) 2012 Chelsio Communications, Inc. 3 * All rights reserved. 4 * Written by: Navdeep Parhar <np@FreeBSD.org> 5 * 6 * Redistribution and use in source and binary forms, with or without 7 * modification, are permitted provided that the following conditions 8 * are met: 9 * 1. Redistributions of source code must retain the above copyright 10 * notice, this list of conditions and the following disclaimer. 11 * 2. Redistributions in binary form must reproduce the above copyright 12 * notice, this list of conditions and the following disclaimer in the 13 * documentation and/or other materials provided with the distribution. 14 * 15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 25 * SUCH DAMAGE. 26 */ 27 28#include <sys/cdefs.h> 29__FBSDID("$FreeBSD: stable/11/sys/dev/cxgbe/tom/t4_tom.c 312116 2017-01-14 04:24:49Z np $"); 30 31#include "opt_inet.h" 32#include "opt_inet6.h" 33 34#include <sys/param.h> 35#include <sys/types.h> 36#include <sys/systm.h> 37#include <sys/kernel.h> 38#include <sys/ktr.h> 39#include <sys/lock.h> 40#include <sys/limits.h> 41#include <sys/module.h> 42#include <sys/protosw.h> 43#include <sys/domain.h> 44#include <sys/refcount.h> 45#include <sys/rmlock.h> 46#include <sys/socket.h> 47#include <sys/socketvar.h> 48#include <sys/taskqueue.h> 49#include <net/if.h> 50#include <net/if_var.h> 51#include <netinet/in.h> 52#include <netinet/in_pcb.h> 53#include <netinet/in_var.h> 54#include <netinet/ip.h> 55#include <netinet/ip6.h> 56#include <netinet6/scope6_var.h> 57#define TCPSTATES 58#include <netinet/tcp_fsm.h> 59#include <netinet/tcp_var.h> 60#include <netinet/toecore.h> 61 62#ifdef TCP_OFFLOAD 63#include "common/common.h" 64#include "common/t4_msg.h" 65#include "common/t4_regs.h" 66#include "common/t4_regs_values.h" 67#include "common/t4_tcb.h" 68#include "tom/t4_tom_l2t.h" 69#include "tom/t4_tom.h" 70 71static struct protosw toe_protosw; 72static struct pr_usrreqs toe_usrreqs; 73 74static struct protosw toe6_protosw; 75static struct pr_usrreqs toe6_usrreqs; 76 77/* Module ops */ 78static int t4_tom_mod_load(void); 79static int t4_tom_mod_unload(void); 80static int t4_tom_modevent(module_t, int, void *); 81 82/* ULD ops and helpers */ 83static int t4_tom_activate(struct adapter *); 84static int t4_tom_deactivate(struct adapter *); 85 86static struct uld_info tom_uld_info = { 87 .uld_id = ULD_TOM, 88 .activate = t4_tom_activate, 89 .deactivate = t4_tom_deactivate, 90}; 91 92static void queue_tid_release(struct adapter *, int); 93static void release_offload_resources(struct toepcb *); 94static int alloc_tid_tabs(struct tid_info *); 95static void free_tid_tabs(struct tid_info *); 96static int add_lip(struct adapter *, struct in6_addr *); 97static int delete_lip(struct adapter *, struct in6_addr *); 98static struct clip_entry *search_lip(struct tom_data *, struct in6_addr *); 99static void init_clip_table(struct adapter *, struct tom_data *); 100static void update_clip(struct adapter *, void *); 101static void t4_clip_task(void *, int); 102static void update_clip_table(struct adapter *, struct tom_data *); 103static void destroy_clip_table(struct adapter *, struct tom_data *); 104static void free_tom_data(struct adapter *, struct tom_data *); 105static void reclaim_wr_resources(void *, int); 106 107static int in6_ifaddr_gen; 108static eventhandler_tag ifaddr_evhandler; 109static struct timeout_task clip_task; 110 111struct toepcb * 112alloc_toepcb(struct vi_info *vi, int txqid, int rxqid, int flags) 113{ 114 struct port_info *pi = vi->pi; 115 struct adapter *sc = pi->adapter; 116 struct toepcb *toep; 117 int tx_credits, txsd_total, len; 118 119 /* 120 * The firmware counts tx work request credits in units of 16 bytes 121 * each. Reserve room for an ABORT_REQ so the driver never has to worry 122 * about tx credits if it wants to abort a connection. 123 */ 124 tx_credits = sc->params.ofldq_wr_cred; 125 tx_credits -= howmany(sizeof(struct cpl_abort_req), 16); 126 127 /* 128 * Shortest possible tx work request is a fw_ofld_tx_data_wr + 1 byte 129 * immediate payload, and firmware counts tx work request credits in 130 * units of 16 byte. Calculate the maximum work requests possible. 131 */ 132 txsd_total = tx_credits / 133 howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16); 134 135 if (txqid < 0) 136 txqid = (arc4random() % vi->nofldtxq) + vi->first_ofld_txq; 137 KASSERT(txqid >= vi->first_ofld_txq && 138 txqid < vi->first_ofld_txq + vi->nofldtxq, 139 ("%s: txqid %d for vi %p (first %d, n %d)", __func__, txqid, vi, 140 vi->first_ofld_txq, vi->nofldtxq)); 141 142 if (rxqid < 0) 143 rxqid = (arc4random() % vi->nofldrxq) + vi->first_ofld_rxq; 144 KASSERT(rxqid >= vi->first_ofld_rxq && 145 rxqid < vi->first_ofld_rxq + vi->nofldrxq, 146 ("%s: rxqid %d for vi %p (first %d, n %d)", __func__, rxqid, vi, 147 vi->first_ofld_rxq, vi->nofldrxq)); 148 149 len = offsetof(struct toepcb, txsd) + 150 txsd_total * sizeof(struct ofld_tx_sdesc); 151 152 toep = malloc(len, M_CXGBE, M_ZERO | flags); 153 if (toep == NULL) 154 return (NULL); 155 156 refcount_init(&toep->refcount, 1); 157 toep->td = sc->tom_softc; 158 toep->vi = vi; 159 toep->tx_total = tx_credits; 160 toep->tx_credits = tx_credits; 161 toep->ofld_txq = &sc->sge.ofld_txq[txqid]; 162 toep->ofld_rxq = &sc->sge.ofld_rxq[rxqid]; 163 toep->ctrlq = &sc->sge.ctrlq[pi->port_id]; 164 mbufq_init(&toep->ulp_pduq, INT_MAX); 165 mbufq_init(&toep->ulp_pdu_reclaimq, INT_MAX); 166 toep->txsd_total = txsd_total; 167 toep->txsd_avail = txsd_total; 168 toep->txsd_pidx = 0; 169 toep->txsd_cidx = 0; 170 aiotx_init_toep(toep); 171 ddp_init_toep(toep); 172 173 return (toep); 174} 175 176struct toepcb * 177hold_toepcb(struct toepcb *toep) 178{ 179 180 refcount_acquire(&toep->refcount); 181 return (toep); 182} 183 184void 185free_toepcb(struct toepcb *toep) 186{ 187 188 if (refcount_release(&toep->refcount) == 0) 189 return; 190 191 KASSERT(!(toep->flags & TPF_ATTACHED), 192 ("%s: attached to an inpcb", __func__)); 193 KASSERT(!(toep->flags & TPF_CPL_PENDING), 194 ("%s: CPL pending", __func__)); 195 196 ddp_uninit_toep(toep); 197 free(toep, M_CXGBE); 198} 199 200/* 201 * Set up the socket for TCP offload. 202 */ 203void 204offload_socket(struct socket *so, struct toepcb *toep) 205{ 206 struct tom_data *td = toep->td; 207 struct inpcb *inp = sotoinpcb(so); 208 struct tcpcb *tp = intotcpcb(inp); 209 struct sockbuf *sb; 210 211 INP_WLOCK_ASSERT(inp); 212 213 /* Update socket */ 214 sb = &so->so_snd; 215 SOCKBUF_LOCK(sb); 216 sb->sb_flags |= SB_NOCOALESCE; 217 SOCKBUF_UNLOCK(sb); 218 sb = &so->so_rcv; 219 SOCKBUF_LOCK(sb); 220 sb->sb_flags |= SB_NOCOALESCE; 221 if (inp->inp_vflag & INP_IPV6) 222 so->so_proto = &toe6_protosw; 223 else 224 so->so_proto = &toe_protosw; 225 SOCKBUF_UNLOCK(sb); 226 227 /* Update TCP PCB */ 228 tp->tod = &td->tod; 229 tp->t_toe = toep; 230 tp->t_flags |= TF_TOE; 231 232 /* Install an extra hold on inp */ 233 toep->inp = inp; 234 toep->flags |= TPF_ATTACHED; 235 in_pcbref(inp); 236 237 /* Add the TOE PCB to the active list */ 238 mtx_lock(&td->toep_list_lock); 239 TAILQ_INSERT_HEAD(&td->toep_list, toep, link); 240 mtx_unlock(&td->toep_list_lock); 241} 242 243/* This is _not_ the normal way to "unoffload" a socket. */ 244void 245undo_offload_socket(struct socket *so) 246{ 247 struct inpcb *inp = sotoinpcb(so); 248 struct tcpcb *tp = intotcpcb(inp); 249 struct toepcb *toep = tp->t_toe; 250 struct tom_data *td = toep->td; 251 struct sockbuf *sb; 252 253 INP_WLOCK_ASSERT(inp); 254 255 sb = &so->so_snd; 256 SOCKBUF_LOCK(sb); 257 sb->sb_flags &= ~SB_NOCOALESCE; 258 SOCKBUF_UNLOCK(sb); 259 sb = &so->so_rcv; 260 SOCKBUF_LOCK(sb); 261 sb->sb_flags &= ~SB_NOCOALESCE; 262 SOCKBUF_UNLOCK(sb); 263 264 tp->tod = NULL; 265 tp->t_toe = NULL; 266 tp->t_flags &= ~TF_TOE; 267 268 toep->inp = NULL; 269 toep->flags &= ~TPF_ATTACHED; 270 if (in_pcbrele_wlocked(inp)) 271 panic("%s: inp freed.", __func__); 272 273 mtx_lock(&td->toep_list_lock); 274 TAILQ_REMOVE(&td->toep_list, toep, link); 275 mtx_unlock(&td->toep_list_lock); 276 277 free_toepcb(toep); 278} 279 280static void 281release_offload_resources(struct toepcb *toep) 282{ 283 struct tom_data *td = toep->td; 284 struct adapter *sc = td_adapter(td); 285 int tid = toep->tid; 286 287 KASSERT(!(toep->flags & TPF_CPL_PENDING), 288 ("%s: %p has CPL pending.", __func__, toep)); 289 KASSERT(!(toep->flags & TPF_ATTACHED), 290 ("%s: %p is still attached.", __func__, toep)); 291 292 CTR5(KTR_CXGBE, "%s: toep %p (tid %d, l2te %p, ce %p)", 293 __func__, toep, tid, toep->l2te, toep->ce); 294 295 /* 296 * These queues should have been emptied at approximately the same time 297 * that a normal connection's socket's so_snd would have been purged or 298 * drained. Do _not_ clean up here. 299 */ 300 MPASS(mbufq_len(&toep->ulp_pduq) == 0); 301 MPASS(mbufq_len(&toep->ulp_pdu_reclaimq) == 0); 302#ifdef INVARIANTS 303 ddp_assert_empty(toep); 304#endif 305 306 if (toep->l2te) 307 t4_l2t_release(toep->l2te); 308 309 if (tid >= 0) { 310 remove_tid(sc, tid, toep->ce ? 2 : 1); 311 release_tid(sc, tid, toep->ctrlq); 312 } 313 314 if (toep->ce) 315 release_lip(td, toep->ce); 316 317 mtx_lock(&td->toep_list_lock); 318 TAILQ_REMOVE(&td->toep_list, toep, link); 319 mtx_unlock(&td->toep_list_lock); 320 321 free_toepcb(toep); 322} 323 324/* 325 * The kernel is done with the TCP PCB and this is our opportunity to unhook the 326 * toepcb hanging off of it. If the TOE driver is also done with the toepcb (no 327 * pending CPL) then it is time to release all resources tied to the toepcb. 328 * 329 * Also gets called when an offloaded active open fails and the TOM wants the 330 * kernel to take the TCP PCB back. 331 */ 332static void 333t4_pcb_detach(struct toedev *tod __unused, struct tcpcb *tp) 334{ 335#if defined(KTR) || defined(INVARIANTS) 336 struct inpcb *inp = tp->t_inpcb; 337#endif 338 struct toepcb *toep = tp->t_toe; 339 340 INP_WLOCK_ASSERT(inp); 341 342 KASSERT(toep != NULL, ("%s: toep is NULL", __func__)); 343 KASSERT(toep->flags & TPF_ATTACHED, 344 ("%s: not attached", __func__)); 345 346#ifdef KTR 347 if (tp->t_state == TCPS_SYN_SENT) { 348 CTR6(KTR_CXGBE, "%s: atid %d, toep %p (0x%x), inp %p (0x%x)", 349 __func__, toep->tid, toep, toep->flags, inp, 350 inp->inp_flags); 351 } else { 352 CTR6(KTR_CXGBE, 353 "t4_pcb_detach: tid %d (%s), toep %p (0x%x), inp %p (0x%x)", 354 toep->tid, tcpstates[tp->t_state], toep, toep->flags, inp, 355 inp->inp_flags); 356 } 357#endif 358 359 tp->t_toe = NULL; 360 tp->t_flags &= ~TF_TOE; 361 toep->flags &= ~TPF_ATTACHED; 362 363 if (!(toep->flags & TPF_CPL_PENDING)) 364 release_offload_resources(toep); 365} 366 367/* 368 * setsockopt handler. 369 */ 370static void 371t4_ctloutput(struct toedev *tod, struct tcpcb *tp, int dir, int name) 372{ 373 struct adapter *sc = tod->tod_softc; 374 struct toepcb *toep = tp->t_toe; 375 376 if (dir == SOPT_GET) 377 return; 378 379 CTR4(KTR_CXGBE, "%s: tp %p, dir %u, name %u", __func__, tp, dir, name); 380 381 switch (name) { 382 case TCP_NODELAY: 383 t4_set_tcb_field(sc, toep->ctrlq, toep->tid, W_TCB_T_FLAGS, 384 V_TF_NAGLE(1), V_TF_NAGLE(tp->t_flags & TF_NODELAY ? 0 : 1), 385 0, 0, toep->ofld_rxq->iq.abs_id); 386 break; 387 default: 388 break; 389 } 390} 391 392/* 393 * The TOE driver will not receive any more CPLs for the tid associated with the 394 * toepcb; release the hold on the inpcb. 395 */ 396void 397final_cpl_received(struct toepcb *toep) 398{ 399 struct inpcb *inp = toep->inp; 400 401 KASSERT(inp != NULL, ("%s: inp is NULL", __func__)); 402 INP_WLOCK_ASSERT(inp); 403 KASSERT(toep->flags & TPF_CPL_PENDING, 404 ("%s: CPL not pending already?", __func__)); 405 406 CTR6(KTR_CXGBE, "%s: tid %d, toep %p (0x%x), inp %p (0x%x)", 407 __func__, toep->tid, toep, toep->flags, inp, inp->inp_flags); 408 409 if (toep->ulp_mode == ULP_MODE_TCPDDP) 410 release_ddp_resources(toep); 411 toep->inp = NULL; 412 toep->flags &= ~TPF_CPL_PENDING; 413 mbufq_drain(&toep->ulp_pdu_reclaimq); 414 415 if (!(toep->flags & TPF_ATTACHED)) 416 release_offload_resources(toep); 417 418 if (!in_pcbrele_wlocked(inp)) 419 INP_WUNLOCK(inp); 420} 421 422void 423insert_tid(struct adapter *sc, int tid, void *ctx, int ntids) 424{ 425 struct tid_info *t = &sc->tids; 426 427 t->tid_tab[tid] = ctx; 428 atomic_add_int(&t->tids_in_use, ntids); 429} 430 431void * 432lookup_tid(struct adapter *sc, int tid) 433{ 434 struct tid_info *t = &sc->tids; 435 436 return (t->tid_tab[tid]); 437} 438 439void 440update_tid(struct adapter *sc, int tid, void *ctx) 441{ 442 struct tid_info *t = &sc->tids; 443 444 t->tid_tab[tid] = ctx; 445} 446 447void 448remove_tid(struct adapter *sc, int tid, int ntids) 449{ 450 struct tid_info *t = &sc->tids; 451 452 t->tid_tab[tid] = NULL; 453 atomic_subtract_int(&t->tids_in_use, ntids); 454} 455 456void 457release_tid(struct adapter *sc, int tid, struct sge_wrq *ctrlq) 458{ 459 struct wrqe *wr; 460 struct cpl_tid_release *req; 461 462 wr = alloc_wrqe(sizeof(*req), ctrlq); 463 if (wr == NULL) { 464 queue_tid_release(sc, tid); /* defer */ 465 return; 466 } 467 req = wrtod(wr); 468 469 INIT_TP_WR_MIT_CPL(req, CPL_TID_RELEASE, tid); 470 471 t4_wrq_tx(sc, wr); 472} 473 474static void 475queue_tid_release(struct adapter *sc, int tid) 476{ 477 478 CXGBE_UNIMPLEMENTED("deferred tid release"); 479} 480 481/* 482 * What mtu_idx to use, given a 4-tuple and/or an MSS cap 483 */ 484int 485find_best_mtu_idx(struct adapter *sc, struct in_conninfo *inc, int pmss) 486{ 487 unsigned short *mtus = &sc->params.mtus[0]; 488 int i, mss, n; 489 490 KASSERT(inc != NULL || pmss > 0, 491 ("%s: at least one of inc/pmss must be specified", __func__)); 492 493 mss = inc ? tcp_mssopt(inc) : pmss; 494 if (pmss > 0 && mss > pmss) 495 mss = pmss; 496 497 if (inc->inc_flags & INC_ISIPV6) 498 n = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 499 else 500 n = sizeof(struct ip) + sizeof(struct tcphdr); 501 502 for (i = 0; i < NMTUS - 1 && mtus[i + 1] <= mss + n; i++) 503 continue; 504 505 return (i); 506} 507 508/* 509 * Determine the receive window size for a socket. 510 */ 511u_long 512select_rcv_wnd(struct socket *so) 513{ 514 unsigned long wnd; 515 516 SOCKBUF_LOCK_ASSERT(&so->so_rcv); 517 518 wnd = sbspace(&so->so_rcv); 519 if (wnd < MIN_RCV_WND) 520 wnd = MIN_RCV_WND; 521 522 return min(wnd, MAX_RCV_WND); 523} 524 525int 526select_rcv_wscale(void) 527{ 528 int wscale = 0; 529 unsigned long space = sb_max; 530 531 if (space > MAX_RCV_WND) 532 space = MAX_RCV_WND; 533 534 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < space) 535 wscale++; 536 537 return (wscale); 538} 539 540extern int always_keepalive; 541 542/* 543 * socket so could be a listening socket too. 544 */ 545uint64_t 546calc_opt0(struct socket *so, struct vi_info *vi, struct l2t_entry *e, 547 int mtu_idx, int rscale, int rx_credits, int ulp_mode) 548{ 549 uint64_t opt0; 550 551 KASSERT(rx_credits <= M_RCV_BUFSIZ, 552 ("%s: rcv_bufsiz too high", __func__)); 553 554 opt0 = F_TCAM_BYPASS | V_WND_SCALE(rscale) | V_MSS_IDX(mtu_idx) | 555 V_ULP_MODE(ulp_mode) | V_RCV_BUFSIZ(rx_credits); 556 557 if (so != NULL) { 558 struct inpcb *inp = sotoinpcb(so); 559 struct tcpcb *tp = intotcpcb(inp); 560 int keepalive = always_keepalive || 561 so_options_get(so) & SO_KEEPALIVE; 562 563 opt0 |= V_NAGLE((tp->t_flags & TF_NODELAY) == 0); 564 opt0 |= V_KEEP_ALIVE(keepalive != 0); 565 } 566 567 if (e != NULL) 568 opt0 |= V_L2T_IDX(e->idx); 569 570 if (vi != NULL) { 571 opt0 |= V_SMAC_SEL(vi->smt_idx); 572 opt0 |= V_TX_CHAN(vi->pi->tx_chan); 573 } 574 575 return htobe64(opt0); 576} 577 578uint64_t 579select_ntuple(struct vi_info *vi, struct l2t_entry *e) 580{ 581 struct adapter *sc = vi->pi->adapter; 582 struct tp_params *tp = &sc->params.tp; 583 uint16_t viid = vi->viid; 584 uint64_t ntuple = 0; 585 586 /* 587 * Initialize each of the fields which we care about which are present 588 * in the Compressed Filter Tuple. 589 */ 590 if (tp->vlan_shift >= 0 && e->vlan != CPL_L2T_VLAN_NONE) 591 ntuple |= (uint64_t)(F_FT_VLAN_VLD | e->vlan) << tp->vlan_shift; 592 593 if (tp->port_shift >= 0) 594 ntuple |= (uint64_t)e->lport << tp->port_shift; 595 596 if (tp->protocol_shift >= 0) 597 ntuple |= (uint64_t)IPPROTO_TCP << tp->protocol_shift; 598 599 if (tp->vnic_shift >= 0) { 600 uint32_t vf = G_FW_VIID_VIN(viid); 601 uint32_t pf = G_FW_VIID_PFN(viid); 602 uint32_t vld = G_FW_VIID_VIVLD(viid); 603 604 ntuple |= (uint64_t)(V_FT_VNID_ID_VF(vf) | V_FT_VNID_ID_PF(pf) | 605 V_FT_VNID_ID_VLD(vld)) << tp->vnic_shift; 606 } 607 608 if (is_t4(sc)) 609 return (htobe32((uint32_t)ntuple)); 610 else 611 return (htobe64(V_FILTER_TUPLE(ntuple))); 612} 613 614void 615set_tcpddp_ulp_mode(struct toepcb *toep) 616{ 617 618 toep->ulp_mode = ULP_MODE_TCPDDP; 619 toep->ddp_flags = DDP_OK; 620} 621 622int 623negative_advice(int status) 624{ 625 626 return (status == CPL_ERR_RTX_NEG_ADVICE || 627 status == CPL_ERR_PERSIST_NEG_ADVICE || 628 status == CPL_ERR_KEEPALV_NEG_ADVICE); 629} 630 631static int 632alloc_tid_tabs(struct tid_info *t) 633{ 634 size_t size; 635 unsigned int i; 636 637 size = t->ntids * sizeof(*t->tid_tab) + 638 t->natids * sizeof(*t->atid_tab) + 639 t->nstids * sizeof(*t->stid_tab); 640 641 t->tid_tab = malloc(size, M_CXGBE, M_ZERO | M_NOWAIT); 642 if (t->tid_tab == NULL) 643 return (ENOMEM); 644 645 mtx_init(&t->atid_lock, "atid lock", NULL, MTX_DEF); 646 t->atid_tab = (union aopen_entry *)&t->tid_tab[t->ntids]; 647 t->afree = t->atid_tab; 648 t->atids_in_use = 0; 649 for (i = 1; i < t->natids; i++) 650 t->atid_tab[i - 1].next = &t->atid_tab[i]; 651 t->atid_tab[t->natids - 1].next = NULL; 652 653 mtx_init(&t->stid_lock, "stid lock", NULL, MTX_DEF); 654 t->stid_tab = (struct listen_ctx **)&t->atid_tab[t->natids]; 655 t->stids_in_use = 0; 656 TAILQ_INIT(&t->stids); 657 t->nstids_free_head = t->nstids; 658 659 atomic_store_rel_int(&t->tids_in_use, 0); 660 661 return (0); 662} 663 664static void 665free_tid_tabs(struct tid_info *t) 666{ 667 KASSERT(t->tids_in_use == 0, 668 ("%s: %d tids still in use.", __func__, t->tids_in_use)); 669 KASSERT(t->atids_in_use == 0, 670 ("%s: %d atids still in use.", __func__, t->atids_in_use)); 671 KASSERT(t->stids_in_use == 0, 672 ("%s: %d tids still in use.", __func__, t->stids_in_use)); 673 674 free(t->tid_tab, M_CXGBE); 675 t->tid_tab = NULL; 676 677 if (mtx_initialized(&t->atid_lock)) 678 mtx_destroy(&t->atid_lock); 679 if (mtx_initialized(&t->stid_lock)) 680 mtx_destroy(&t->stid_lock); 681} 682 683static int 684add_lip(struct adapter *sc, struct in6_addr *lip) 685{ 686 struct fw_clip_cmd c; 687 688 ASSERT_SYNCHRONIZED_OP(sc); 689 /* mtx_assert(&td->clip_table_lock, MA_OWNED); */ 690 691 memset(&c, 0, sizeof(c)); 692 c.op_to_write = htonl(V_FW_CMD_OP(FW_CLIP_CMD) | F_FW_CMD_REQUEST | 693 F_FW_CMD_WRITE); 694 c.alloc_to_len16 = htonl(F_FW_CLIP_CMD_ALLOC | FW_LEN16(c)); 695 c.ip_hi = *(uint64_t *)&lip->s6_addr[0]; 696 c.ip_lo = *(uint64_t *)&lip->s6_addr[8]; 697 698 return (-t4_wr_mbox_ns(sc, sc->mbox, &c, sizeof(c), &c)); 699} 700 701static int 702delete_lip(struct adapter *sc, struct in6_addr *lip) 703{ 704 struct fw_clip_cmd c; 705 706 ASSERT_SYNCHRONIZED_OP(sc); 707 /* mtx_assert(&td->clip_table_lock, MA_OWNED); */ 708 709 memset(&c, 0, sizeof(c)); 710 c.op_to_write = htonl(V_FW_CMD_OP(FW_CLIP_CMD) | F_FW_CMD_REQUEST | 711 F_FW_CMD_READ); 712 c.alloc_to_len16 = htonl(F_FW_CLIP_CMD_FREE | FW_LEN16(c)); 713 c.ip_hi = *(uint64_t *)&lip->s6_addr[0]; 714 c.ip_lo = *(uint64_t *)&lip->s6_addr[8]; 715 716 return (-t4_wr_mbox_ns(sc, sc->mbox, &c, sizeof(c), &c)); 717} 718 719static struct clip_entry * 720search_lip(struct tom_data *td, struct in6_addr *lip) 721{ 722 struct clip_entry *ce; 723 724 mtx_assert(&td->clip_table_lock, MA_OWNED); 725 726 TAILQ_FOREACH(ce, &td->clip_table, link) { 727 if (IN6_ARE_ADDR_EQUAL(&ce->lip, lip)) 728 return (ce); 729 } 730 731 return (NULL); 732} 733 734struct clip_entry * 735hold_lip(struct tom_data *td, struct in6_addr *lip) 736{ 737 struct clip_entry *ce; 738 739 mtx_lock(&td->clip_table_lock); 740 ce = search_lip(td, lip); 741 if (ce != NULL) 742 ce->refcount++; 743 mtx_unlock(&td->clip_table_lock); 744 745 return (ce); 746} 747 748void 749release_lip(struct tom_data *td, struct clip_entry *ce) 750{ 751 752 mtx_lock(&td->clip_table_lock); 753 KASSERT(search_lip(td, &ce->lip) == ce, 754 ("%s: CLIP entry %p p not in CLIP table.", __func__, ce)); 755 KASSERT(ce->refcount > 0, 756 ("%s: CLIP entry %p has refcount 0", __func__, ce)); 757 --ce->refcount; 758 mtx_unlock(&td->clip_table_lock); 759} 760 761static void 762init_clip_table(struct adapter *sc, struct tom_data *td) 763{ 764 765 ASSERT_SYNCHRONIZED_OP(sc); 766 767 mtx_init(&td->clip_table_lock, "CLIP table lock", NULL, MTX_DEF); 768 TAILQ_INIT(&td->clip_table); 769 td->clip_gen = -1; 770 771 update_clip_table(sc, td); 772} 773 774static void 775update_clip(struct adapter *sc, void *arg __unused) 776{ 777 778 if (begin_synchronized_op(sc, NULL, HOLD_LOCK, "t4tomuc")) 779 return; 780 781 if (uld_active(sc, ULD_TOM)) 782 update_clip_table(sc, sc->tom_softc); 783 784 end_synchronized_op(sc, LOCK_HELD); 785} 786 787static void 788t4_clip_task(void *arg, int count) 789{ 790 791 t4_iterate(update_clip, NULL); 792} 793 794static void 795update_clip_table(struct adapter *sc, struct tom_data *td) 796{ 797 struct rm_priotracker in6_ifa_tracker; 798 struct in6_ifaddr *ia; 799 struct in6_addr *lip, tlip; 800 struct clip_head stale; 801 struct clip_entry *ce, *ce_temp; 802 struct vi_info *vi; 803 int rc, gen, i, j; 804 uintptr_t last_vnet; 805 806 ASSERT_SYNCHRONIZED_OP(sc); 807 808 IN6_IFADDR_RLOCK(&in6_ifa_tracker); 809 mtx_lock(&td->clip_table_lock); 810 811 gen = atomic_load_acq_int(&in6_ifaddr_gen); 812 if (gen == td->clip_gen) 813 goto done; 814 815 TAILQ_INIT(&stale); 816 TAILQ_CONCAT(&stale, &td->clip_table, link); 817 818 /* 819 * last_vnet optimizes the common cases where all if_vnet = NULL (no 820 * VIMAGE) or all if_vnet = vnet0. 821 */ 822 last_vnet = (uintptr_t)(-1); 823 for_each_port(sc, i) 824 for_each_vi(sc->port[i], j, vi) { 825 if (last_vnet == (uintptr_t)vi->ifp->if_vnet) 826 continue; 827 828 /* XXX: races with if_vmove */ 829 CURVNET_SET(vi->ifp->if_vnet); 830 TAILQ_FOREACH(ia, &V_in6_ifaddrhead, ia_link) { 831 lip = &ia->ia_addr.sin6_addr; 832 833 KASSERT(!IN6_IS_ADDR_MULTICAST(lip), 834 ("%s: mcast address in in6_ifaddr list", __func__)); 835 836 if (IN6_IS_ADDR_LOOPBACK(lip)) 837 continue; 838 if (IN6_IS_SCOPE_EMBED(lip)) { 839 /* Remove the embedded scope */ 840 tlip = *lip; 841 lip = &tlip; 842 in6_clearscope(lip); 843 } 844 /* 845 * XXX: how to weed out the link local address for the 846 * loopback interface? It's fe80::1 usually (always?). 847 */ 848 849 /* 850 * If it's in the main list then we already know it's 851 * not stale. 852 */ 853 TAILQ_FOREACH(ce, &td->clip_table, link) { 854 if (IN6_ARE_ADDR_EQUAL(&ce->lip, lip)) 855 goto next; 856 } 857 858 /* 859 * If it's in the stale list we should move it to the 860 * main list. 861 */ 862 TAILQ_FOREACH(ce, &stale, link) { 863 if (IN6_ARE_ADDR_EQUAL(&ce->lip, lip)) { 864 TAILQ_REMOVE(&stale, ce, link); 865 TAILQ_INSERT_TAIL(&td->clip_table, ce, 866 link); 867 goto next; 868 } 869 } 870 871 /* A new IP6 address; add it to the CLIP table */ 872 ce = malloc(sizeof(*ce), M_CXGBE, M_NOWAIT); 873 memcpy(&ce->lip, lip, sizeof(ce->lip)); 874 ce->refcount = 0; 875 rc = add_lip(sc, lip); 876 if (rc == 0) 877 TAILQ_INSERT_TAIL(&td->clip_table, ce, link); 878 else { 879 char ip[INET6_ADDRSTRLEN]; 880 881 inet_ntop(AF_INET6, &ce->lip, &ip[0], 882 sizeof(ip)); 883 log(LOG_ERR, "%s: could not add %s (%d)\n", 884 __func__, ip, rc); 885 free(ce, M_CXGBE); 886 } 887next: 888 continue; 889 } 890 CURVNET_RESTORE(); 891 last_vnet = (uintptr_t)vi->ifp->if_vnet; 892 } 893 894 /* 895 * Remove stale addresses (those no longer in V_in6_ifaddrhead) that are 896 * no longer referenced by the driver. 897 */ 898 TAILQ_FOREACH_SAFE(ce, &stale, link, ce_temp) { 899 if (ce->refcount == 0) { 900 rc = delete_lip(sc, &ce->lip); 901 if (rc == 0) { 902 TAILQ_REMOVE(&stale, ce, link); 903 free(ce, M_CXGBE); 904 } else { 905 char ip[INET6_ADDRSTRLEN]; 906 907 inet_ntop(AF_INET6, &ce->lip, &ip[0], 908 sizeof(ip)); 909 log(LOG_ERR, "%s: could not delete %s (%d)\n", 910 __func__, ip, rc); 911 } 912 } 913 } 914 /* The ones that are still referenced need to stay in the CLIP table */ 915 TAILQ_CONCAT(&td->clip_table, &stale, link); 916 917 td->clip_gen = gen; 918done: 919 mtx_unlock(&td->clip_table_lock); 920 IN6_IFADDR_RUNLOCK(&in6_ifa_tracker); 921} 922 923static void 924destroy_clip_table(struct adapter *sc, struct tom_data *td) 925{ 926 struct clip_entry *ce, *ce_temp; 927 928 if (mtx_initialized(&td->clip_table_lock)) { 929 mtx_lock(&td->clip_table_lock); 930 TAILQ_FOREACH_SAFE(ce, &td->clip_table, link, ce_temp) { 931 KASSERT(ce->refcount == 0, 932 ("%s: CLIP entry %p still in use (%d)", __func__, 933 ce, ce->refcount)); 934 TAILQ_REMOVE(&td->clip_table, ce, link); 935 delete_lip(sc, &ce->lip); 936 free(ce, M_CXGBE); 937 } 938 mtx_unlock(&td->clip_table_lock); 939 mtx_destroy(&td->clip_table_lock); 940 } 941} 942 943static void 944free_tom_data(struct adapter *sc, struct tom_data *td) 945{ 946 947 ASSERT_SYNCHRONIZED_OP(sc); 948 949 KASSERT(TAILQ_EMPTY(&td->toep_list), 950 ("%s: TOE PCB list is not empty.", __func__)); 951 KASSERT(td->lctx_count == 0, 952 ("%s: lctx hash table is not empty.", __func__)); 953 954 t4_free_ppod_region(&td->pr); 955 destroy_clip_table(sc, td); 956 957 if (td->listen_mask != 0) 958 hashdestroy(td->listen_hash, M_CXGBE, td->listen_mask); 959 960 if (mtx_initialized(&td->unsent_wr_lock)) 961 mtx_destroy(&td->unsent_wr_lock); 962 if (mtx_initialized(&td->lctx_hash_lock)) 963 mtx_destroy(&td->lctx_hash_lock); 964 if (mtx_initialized(&td->toep_list_lock)) 965 mtx_destroy(&td->toep_list_lock); 966 967 free_tid_tabs(&sc->tids); 968 free(td, M_CXGBE); 969} 970 971static void 972reclaim_wr_resources(void *arg, int count) 973{ 974 struct tom_data *td = arg; 975 STAILQ_HEAD(, wrqe) twr_list = STAILQ_HEAD_INITIALIZER(twr_list); 976 struct cpl_act_open_req *cpl; 977 u_int opcode, atid; 978 struct wrqe *wr; 979 struct adapter *sc; 980 981 mtx_lock(&td->unsent_wr_lock); 982 STAILQ_SWAP(&td->unsent_wr_list, &twr_list, wrqe); 983 mtx_unlock(&td->unsent_wr_lock); 984 985 while ((wr = STAILQ_FIRST(&twr_list)) != NULL) { 986 STAILQ_REMOVE_HEAD(&twr_list, link); 987 988 cpl = wrtod(wr); 989 opcode = GET_OPCODE(cpl); 990 991 switch (opcode) { 992 case CPL_ACT_OPEN_REQ: 993 case CPL_ACT_OPEN_REQ6: 994 atid = G_TID_TID(be32toh(OPCODE_TID(cpl))); 995 sc = td_adapter(td); 996 997 CTR2(KTR_CXGBE, "%s: atid %u ", __func__, atid); 998 act_open_failure_cleanup(sc, atid, EHOSTUNREACH); 999 free(wr, M_CXGBE); 1000 break; 1001 default: 1002 log(LOG_ERR, "%s: leaked work request %p, wr_len %d, " 1003 "opcode %x\n", __func__, wr, wr->wr_len, opcode); 1004 /* WR not freed here; go look at it with a debugger. */ 1005 } 1006 } 1007} 1008 1009/* 1010 * Ground control to Major TOM 1011 * Commencing countdown, engines on 1012 */ 1013static int 1014t4_tom_activate(struct adapter *sc) 1015{ 1016 struct tom_data *td; 1017 struct toedev *tod; 1018 struct vi_info *vi; 1019 struct sge_ofld_rxq *ofld_rxq; 1020 int i, j, rc, v; 1021 1022 ASSERT_SYNCHRONIZED_OP(sc); 1023 1024 /* per-adapter softc for TOM */ 1025 td = malloc(sizeof(*td), M_CXGBE, M_ZERO | M_NOWAIT); 1026 if (td == NULL) 1027 return (ENOMEM); 1028 1029 /* List of TOE PCBs and associated lock */ 1030 mtx_init(&td->toep_list_lock, "PCB list lock", NULL, MTX_DEF); 1031 TAILQ_INIT(&td->toep_list); 1032 1033 /* Listen context */ 1034 mtx_init(&td->lctx_hash_lock, "lctx hash lock", NULL, MTX_DEF); 1035 td->listen_hash = hashinit_flags(LISTEN_HASH_SIZE, M_CXGBE, 1036 &td->listen_mask, HASH_NOWAIT); 1037 1038 /* List of WRs for which L2 resolution failed */ 1039 mtx_init(&td->unsent_wr_lock, "Unsent WR list lock", NULL, MTX_DEF); 1040 STAILQ_INIT(&td->unsent_wr_list); 1041 TASK_INIT(&td->reclaim_wr_resources, 0, reclaim_wr_resources, td); 1042 1043 /* TID tables */ 1044 rc = alloc_tid_tabs(&sc->tids); 1045 if (rc != 0) 1046 goto done; 1047 1048 rc = t4_init_ppod_region(&td->pr, &sc->vres.ddp, 1049 t4_read_reg(sc, A_ULP_RX_TDDP_PSZ), "TDDP page pods"); 1050 if (rc != 0) 1051 goto done; 1052 t4_set_reg_field(sc, A_ULP_RX_TDDP_TAGMASK, 1053 V_TDDPTAGMASK(M_TDDPTAGMASK), td->pr.pr_tag_mask); 1054 1055 /* CLIP table for IPv6 offload */ 1056 init_clip_table(sc, td); 1057 1058 /* toedev ops */ 1059 tod = &td->tod; 1060 init_toedev(tod); 1061 tod->tod_softc = sc; 1062 tod->tod_connect = t4_connect; 1063 tod->tod_listen_start = t4_listen_start; 1064 tod->tod_listen_stop = t4_listen_stop; 1065 tod->tod_rcvd = t4_rcvd; 1066 tod->tod_output = t4_tod_output; 1067 tod->tod_send_rst = t4_send_rst; 1068 tod->tod_send_fin = t4_send_fin; 1069 tod->tod_pcb_detach = t4_pcb_detach; 1070 tod->tod_l2_update = t4_l2_update; 1071 tod->tod_syncache_added = t4_syncache_added; 1072 tod->tod_syncache_removed = t4_syncache_removed; 1073 tod->tod_syncache_respond = t4_syncache_respond; 1074 tod->tod_offload_socket = t4_offload_socket; 1075 tod->tod_ctloutput = t4_ctloutput; 1076 1077 for_each_port(sc, i) { 1078 for_each_vi(sc->port[i], v, vi) { 1079 TOEDEV(vi->ifp) = &td->tod; 1080 for_each_ofld_rxq(vi, j, ofld_rxq) { 1081 ofld_rxq->iq.set_tcb_rpl = do_set_tcb_rpl; 1082 ofld_rxq->iq.l2t_write_rpl = do_l2t_write_rpl2; 1083 } 1084 } 1085 } 1086 1087 sc->tom_softc = td; 1088 register_toedev(sc->tom_softc); 1089 1090done: 1091 if (rc != 0) 1092 free_tom_data(sc, td); 1093 return (rc); 1094} 1095 1096static int 1097t4_tom_deactivate(struct adapter *sc) 1098{ 1099 int rc = 0; 1100 struct tom_data *td = sc->tom_softc; 1101 1102 ASSERT_SYNCHRONIZED_OP(sc); 1103 1104 if (td == NULL) 1105 return (0); /* XXX. KASSERT? */ 1106 1107 if (sc->offload_map != 0) 1108 return (EBUSY); /* at least one port has IFCAP_TOE enabled */ 1109 1110 if (uld_active(sc, ULD_IWARP) || uld_active(sc, ULD_ISCSI)) 1111 return (EBUSY); /* both iWARP and iSCSI rely on the TOE. */ 1112 1113 mtx_lock(&td->toep_list_lock); 1114 if (!TAILQ_EMPTY(&td->toep_list)) 1115 rc = EBUSY; 1116 mtx_unlock(&td->toep_list_lock); 1117 1118 mtx_lock(&td->lctx_hash_lock); 1119 if (td->lctx_count > 0) 1120 rc = EBUSY; 1121 mtx_unlock(&td->lctx_hash_lock); 1122 1123 taskqueue_drain(taskqueue_thread, &td->reclaim_wr_resources); 1124 mtx_lock(&td->unsent_wr_lock); 1125 if (!STAILQ_EMPTY(&td->unsent_wr_list)) 1126 rc = EBUSY; 1127 mtx_unlock(&td->unsent_wr_lock); 1128 1129 if (rc == 0) { 1130 unregister_toedev(sc->tom_softc); 1131 free_tom_data(sc, td); 1132 sc->tom_softc = NULL; 1133 } 1134 1135 return (rc); 1136} 1137 1138static void 1139t4_tom_ifaddr_event(void *arg __unused, struct ifnet *ifp) 1140{ 1141 1142 atomic_add_rel_int(&in6_ifaddr_gen, 1); 1143 taskqueue_enqueue_timeout(taskqueue_thread, &clip_task, -hz / 4); 1144} 1145 1146static int 1147t4_aio_queue_tom(struct socket *so, struct kaiocb *job) 1148{ 1149 struct tcpcb *tp = so_sototcpcb(so); 1150 struct toepcb *toep = tp->t_toe; 1151 int error; 1152 1153 if (toep->ulp_mode == ULP_MODE_TCPDDP) { 1154 error = t4_aio_queue_ddp(so, job); 1155 if (error != EOPNOTSUPP) 1156 return (error); 1157 } 1158 1159 return (t4_aio_queue_aiotx(so, job)); 1160} 1161 1162static int 1163t4_tom_mod_load(void) 1164{ 1165 int rc; 1166 struct protosw *tcp_protosw, *tcp6_protosw; 1167 1168 /* CPL handlers */ 1169 t4_init_connect_cpl_handlers(); 1170 t4_init_listen_cpl_handlers(); 1171 t4_init_cpl_io_handlers(); 1172 1173 rc = t4_ddp_mod_load(); 1174 if (rc != 0) 1175 return (rc); 1176 1177 tcp_protosw = pffindproto(PF_INET, IPPROTO_TCP, SOCK_STREAM); 1178 if (tcp_protosw == NULL) 1179 return (ENOPROTOOPT); 1180 bcopy(tcp_protosw, &toe_protosw, sizeof(toe_protosw)); 1181 bcopy(tcp_protosw->pr_usrreqs, &toe_usrreqs, sizeof(toe_usrreqs)); 1182 toe_usrreqs.pru_aio_queue = t4_aio_queue_tom; 1183 toe_protosw.pr_usrreqs = &toe_usrreqs; 1184 1185 tcp6_protosw = pffindproto(PF_INET6, IPPROTO_TCP, SOCK_STREAM); 1186 if (tcp6_protosw == NULL) 1187 return (ENOPROTOOPT); 1188 bcopy(tcp6_protosw, &toe6_protosw, sizeof(toe6_protosw)); 1189 bcopy(tcp6_protosw->pr_usrreqs, &toe6_usrreqs, sizeof(toe6_usrreqs)); 1190 toe6_usrreqs.pru_aio_queue = t4_aio_queue_tom; 1191 toe6_protosw.pr_usrreqs = &toe6_usrreqs; 1192 1193 TIMEOUT_TASK_INIT(taskqueue_thread, &clip_task, 0, t4_clip_task, NULL); 1194 ifaddr_evhandler = EVENTHANDLER_REGISTER(ifaddr_event, 1195 t4_tom_ifaddr_event, NULL, EVENTHANDLER_PRI_ANY); 1196 1197 rc = t4_register_uld(&tom_uld_info); 1198 if (rc != 0) 1199 t4_tom_mod_unload(); 1200 1201 return (rc); 1202} 1203 1204static void 1205tom_uninit(struct adapter *sc, void *arg __unused) 1206{ 1207 if (begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4tomun")) 1208 return; 1209 1210 /* Try to free resources (works only if no port has IFCAP_TOE) */ 1211 if (uld_active(sc, ULD_TOM)) 1212 t4_deactivate_uld(sc, ULD_TOM); 1213 1214 end_synchronized_op(sc, 0); 1215} 1216 1217static int 1218t4_tom_mod_unload(void) 1219{ 1220 t4_iterate(tom_uninit, NULL); 1221 1222 if (t4_unregister_uld(&tom_uld_info) == EBUSY) 1223 return (EBUSY); 1224 1225 if (ifaddr_evhandler) { 1226 EVENTHANDLER_DEREGISTER(ifaddr_event, ifaddr_evhandler); 1227 taskqueue_cancel_timeout(taskqueue_thread, &clip_task, NULL); 1228 } 1229 1230 t4_ddp_mod_unload(); 1231 1232 return (0); 1233} 1234#endif /* TCP_OFFLOAD */ 1235 1236static int 1237t4_tom_modevent(module_t mod, int cmd, void *arg) 1238{ 1239 int rc = 0; 1240 1241#ifdef TCP_OFFLOAD 1242 switch (cmd) { 1243 case MOD_LOAD: 1244 rc = t4_tom_mod_load(); 1245 break; 1246 1247 case MOD_UNLOAD: 1248 rc = t4_tom_mod_unload(); 1249 break; 1250 1251 default: 1252 rc = EINVAL; 1253 } 1254#else 1255 printf("t4_tom: compiled without TCP_OFFLOAD support.\n"); 1256 rc = EOPNOTSUPP; 1257#endif 1258 return (rc); 1259} 1260 1261static moduledata_t t4_tom_moddata= { 1262 "t4_tom", 1263 t4_tom_modevent, 1264 0 1265}; 1266 1267MODULE_VERSION(t4_tom, 1); 1268MODULE_DEPEND(t4_tom, toecore, 1, 1, 1); 1269MODULE_DEPEND(t4_tom, t4nex, 1, 1, 1); 1270DECLARE_MODULE(t4_tom, t4_tom_moddata, SI_SUB_EXEC, SI_ORDER_ANY); 1271