t4_tom.c revision 309560
1/*- 2 * Copyright (c) 2012 Chelsio Communications, Inc. 3 * All rights reserved. 4 * Written by: Navdeep Parhar <np@FreeBSD.org> 5 * 6 * Redistribution and use in source and binary forms, with or without 7 * modification, are permitted provided that the following conditions 8 * are met: 9 * 1. Redistributions of source code must retain the above copyright 10 * notice, this list of conditions and the following disclaimer. 11 * 2. Redistributions in binary form must reproduce the above copyright 12 * notice, this list of conditions and the following disclaimer in the 13 * documentation and/or other materials provided with the distribution. 14 * 15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 16 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 17 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 19 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 21 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 22 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 23 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 24 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 25 * SUCH DAMAGE. 26 */ 27 28#include <sys/cdefs.h> 29__FBSDID("$FreeBSD: stable/10/sys/dev/cxgbe/tom/t4_tom.c 309560 2016-12-05 20:43:25Z jhb $"); 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/limits.h> 40#include <sys/module.h> 41#include <sys/protosw.h> 42#include <sys/domain.h> 43#include <sys/socket.h> 44#include <sys/socketvar.h> 45#include <sys/taskqueue.h> 46#include <net/if.h> 47#include <netinet/in.h> 48#include <netinet/in_pcb.h> 49#include <netinet/in_var.h> 50#include <netinet/ip.h> 51#include <netinet/ip6.h> 52#include <netinet/tcp_var.h> 53#include <netinet6/scope6_var.h> 54#define TCPSTATES 55#include <netinet/tcp_fsm.h> 56#include <netinet/toecore.h> 57 58#ifdef TCP_OFFLOAD 59#include "common/common.h" 60#include "common/t4_msg.h" 61#include "common/t4_regs.h" 62#include "common/t4_regs_values.h" 63#include "common/t4_tcb.h" 64#include "tom/t4_tom_l2t.h" 65#include "tom/t4_tom.h" 66 67static struct protosw ddp_protosw; 68static struct pr_usrreqs ddp_usrreqs; 69 70static struct protosw ddp6_protosw; 71static struct pr_usrreqs ddp6_usrreqs; 72 73/* Module ops */ 74static int t4_tom_mod_load(void); 75static int t4_tom_mod_unload(void); 76static int t4_tom_modevent(module_t, int, void *); 77 78/* ULD ops and helpers */ 79static int t4_tom_activate(struct adapter *); 80static int t4_tom_deactivate(struct adapter *); 81 82static struct uld_info tom_uld_info = { 83 .uld_id = ULD_TOM, 84 .activate = t4_tom_activate, 85 .deactivate = t4_tom_deactivate, 86}; 87 88static void queue_tid_release(struct adapter *, int); 89static void release_offload_resources(struct toepcb *); 90static int alloc_tid_tabs(struct tid_info *); 91static void free_tid_tabs(struct tid_info *); 92static int add_lip(struct adapter *, struct in6_addr *); 93static int delete_lip(struct adapter *, struct in6_addr *); 94static struct clip_entry *search_lip(struct tom_data *, struct in6_addr *); 95static void init_clip_table(struct adapter *, struct tom_data *); 96static void update_clip(struct adapter *, void *); 97static void t4_clip_task(void *, int); 98static void update_clip_table(struct adapter *, struct tom_data *); 99static void destroy_clip_table(struct adapter *, struct tom_data *); 100static void free_tom_data(struct adapter *, struct tom_data *); 101static void reclaim_wr_resources(void *, int); 102 103static int in6_ifaddr_gen; 104static eventhandler_tag ifaddr_evhandler; 105static struct timeout_task clip_task; 106 107static void 108mbufq_init(struct mbufq *q, int limit) 109{ 110 111 q->head = q->tail = NULL; 112} 113 114static void 115mbufq_drain(struct mbufq *q) 116{ 117 struct mbuf *m; 118 119 while ((m = q->head) != NULL) { 120 q->head = m->m_nextpkt; 121 m_freem(m); 122 } 123 q->tail = NULL; 124} 125 126#ifdef INVARIANTS 127static inline int 128mbufq_len(const struct mbufq *q) 129{ 130 struct mbuf *m; 131 int len; 132 133 len = 0; 134 for (m = q->head; m != NULL; m = m->m_nextpkt) 135 len++; 136 return (len); 137} 138#endif 139 140struct toepcb * 141alloc_toepcb(struct vi_info *vi, int txqid, int rxqid, int flags) 142{ 143 struct port_info *pi = vi->pi; 144 struct adapter *sc = pi->adapter; 145 struct toepcb *toep; 146 int tx_credits, txsd_total, len; 147 148 /* 149 * The firmware counts tx work request credits in units of 16 bytes 150 * each. Reserve room for an ABORT_REQ so the driver never has to worry 151 * about tx credits if it wants to abort a connection. 152 */ 153 tx_credits = sc->params.ofldq_wr_cred; 154 tx_credits -= howmany(sizeof(struct cpl_abort_req), 16); 155 156 /* 157 * Shortest possible tx work request is a fw_ofld_tx_data_wr + 1 byte 158 * immediate payload, and firmware counts tx work request credits in 159 * units of 16 byte. Calculate the maximum work requests possible. 160 */ 161 txsd_total = tx_credits / 162 howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16); 163 164 if (txqid < 0) 165 txqid = (arc4random() % vi->nofldtxq) + vi->first_ofld_txq; 166 KASSERT(txqid >= vi->first_ofld_txq && 167 txqid < vi->first_ofld_txq + vi->nofldtxq, 168 ("%s: txqid %d for vi %p (first %d, n %d)", __func__, txqid, vi, 169 vi->first_ofld_txq, vi->nofldtxq)); 170 171 if (rxqid < 0) 172 rxqid = (arc4random() % vi->nofldrxq) + vi->first_ofld_rxq; 173 KASSERT(rxqid >= vi->first_ofld_rxq && 174 rxqid < vi->first_ofld_rxq + vi->nofldrxq, 175 ("%s: rxqid %d for vi %p (first %d, n %d)", __func__, rxqid, vi, 176 vi->first_ofld_rxq, vi->nofldrxq)); 177 178 len = offsetof(struct toepcb, txsd) + 179 txsd_total * sizeof(struct ofld_tx_sdesc); 180 181 toep = malloc(len, M_CXGBE, M_ZERO | flags); 182 if (toep == NULL) 183 return (NULL); 184 185 toep->td = sc->tom_softc; 186 toep->vi = vi; 187 toep->tx_total = tx_credits; 188 toep->tx_credits = tx_credits; 189 toep->ofld_txq = &sc->sge.ofld_txq[txqid]; 190 toep->ofld_rxq = &sc->sge.ofld_rxq[rxqid]; 191 toep->ctrlq = &sc->sge.ctrlq[pi->port_id]; 192 mbufq_init(&toep->ulp_pduq, INT_MAX); 193 mbufq_init(&toep->ulp_pdu_reclaimq, INT_MAX); 194 toep->txsd_total = txsd_total; 195 toep->txsd_avail = txsd_total; 196 toep->txsd_pidx = 0; 197 toep->txsd_cidx = 0; 198 199 return (toep); 200} 201 202void 203free_toepcb(struct toepcb *toep) 204{ 205 206 KASSERT(!(toep->flags & TPF_ATTACHED), 207 ("%s: attached to an inpcb", __func__)); 208 KASSERT(!(toep->flags & TPF_CPL_PENDING), 209 ("%s: CPL pending", __func__)); 210 211 free(toep, M_CXGBE); 212} 213 214/* 215 * Set up the socket for TCP offload. 216 */ 217void 218offload_socket(struct socket *so, struct toepcb *toep) 219{ 220 struct tom_data *td = toep->td; 221 struct inpcb *inp = sotoinpcb(so); 222 struct tcpcb *tp = intotcpcb(inp); 223 struct sockbuf *sb; 224 225 INP_WLOCK_ASSERT(inp); 226 227 /* Update socket */ 228 sb = &so->so_snd; 229 SOCKBUF_LOCK(sb); 230 sb->sb_flags |= SB_NOCOALESCE; 231 SOCKBUF_UNLOCK(sb); 232 sb = &so->so_rcv; 233 SOCKBUF_LOCK(sb); 234 sb->sb_flags |= SB_NOCOALESCE; 235 if (toep->ulp_mode == ULP_MODE_TCPDDP) { 236 if (inp->inp_vflag & INP_IPV6) 237 so->so_proto = &ddp6_protosw; 238 else 239 so->so_proto = &ddp_protosw; 240 } 241 SOCKBUF_UNLOCK(sb); 242 243 /* Update TCP PCB */ 244 tp->tod = &td->tod; 245 tp->t_toe = toep; 246 tp->t_flags |= TF_TOE; 247 248 /* Install an extra hold on inp */ 249 toep->inp = inp; 250 toep->flags |= TPF_ATTACHED; 251 in_pcbref(inp); 252 253 /* Add the TOE PCB to the active list */ 254 mtx_lock(&td->toep_list_lock); 255 TAILQ_INSERT_HEAD(&td->toep_list, toep, link); 256 mtx_unlock(&td->toep_list_lock); 257} 258 259/* This is _not_ the normal way to "unoffload" a socket. */ 260void 261undo_offload_socket(struct socket *so) 262{ 263 struct inpcb *inp = sotoinpcb(so); 264 struct tcpcb *tp = intotcpcb(inp); 265 struct toepcb *toep = tp->t_toe; 266 struct tom_data *td = toep->td; 267 struct sockbuf *sb; 268 269 INP_WLOCK_ASSERT(inp); 270 271 sb = &so->so_snd; 272 SOCKBUF_LOCK(sb); 273 sb->sb_flags &= ~SB_NOCOALESCE; 274 SOCKBUF_UNLOCK(sb); 275 sb = &so->so_rcv; 276 SOCKBUF_LOCK(sb); 277 sb->sb_flags &= ~SB_NOCOALESCE; 278 SOCKBUF_UNLOCK(sb); 279 280 tp->tod = NULL; 281 tp->t_toe = NULL; 282 tp->t_flags &= ~TF_TOE; 283 284 toep->inp = NULL; 285 toep->flags &= ~TPF_ATTACHED; 286 if (in_pcbrele_wlocked(inp)) 287 panic("%s: inp freed.", __func__); 288 289 mtx_lock(&td->toep_list_lock); 290 TAILQ_REMOVE(&td->toep_list, toep, link); 291 mtx_unlock(&td->toep_list_lock); 292} 293 294static void 295release_offload_resources(struct toepcb *toep) 296{ 297 struct tom_data *td = toep->td; 298 struct adapter *sc = td_adapter(td); 299 int tid = toep->tid; 300 301 KASSERT(!(toep->flags & TPF_CPL_PENDING), 302 ("%s: %p has CPL pending.", __func__, toep)); 303 KASSERT(!(toep->flags & TPF_ATTACHED), 304 ("%s: %p is still attached.", __func__, toep)); 305 306 CTR5(KTR_CXGBE, "%s: toep %p (tid %d, l2te %p, ce %p)", 307 __func__, toep, tid, toep->l2te, toep->ce); 308 309 /* 310 * These queues should have been emptied at approximately the same time 311 * that a normal connection's socket's so_snd would have been purged or 312 * drained. Do _not_ clean up here. 313 */ 314 MPASS(mbufq_len(&toep->ulp_pduq) == 0); 315 MPASS(mbufq_len(&toep->ulp_pdu_reclaimq) == 0); 316 317 if (toep->ulp_mode == ULP_MODE_TCPDDP) 318 release_ddp_resources(toep); 319 320 if (toep->l2te) 321 t4_l2t_release(toep->l2te); 322 323 if (tid >= 0) { 324 remove_tid(sc, tid); 325 release_tid(sc, tid, toep->ctrlq); 326 } 327 328 if (toep->ce) 329 release_lip(td, toep->ce); 330 331 mtx_lock(&td->toep_list_lock); 332 TAILQ_REMOVE(&td->toep_list, toep, link); 333 mtx_unlock(&td->toep_list_lock); 334 335 free_toepcb(toep); 336} 337 338/* 339 * The kernel is done with the TCP PCB and this is our opportunity to unhook the 340 * toepcb hanging off of it. If the TOE driver is also done with the toepcb (no 341 * pending CPL) then it is time to release all resources tied to the toepcb. 342 * 343 * Also gets called when an offloaded active open fails and the TOM wants the 344 * kernel to take the TCP PCB back. 345 */ 346static void 347t4_pcb_detach(struct toedev *tod __unused, struct tcpcb *tp) 348{ 349#if defined(KTR) || defined(INVARIANTS) 350 struct inpcb *inp = tp->t_inpcb; 351#endif 352 struct toepcb *toep = tp->t_toe; 353 354 INP_WLOCK_ASSERT(inp); 355 356 KASSERT(toep != NULL, ("%s: toep is NULL", __func__)); 357 KASSERT(toep->flags & TPF_ATTACHED, 358 ("%s: not attached", __func__)); 359 360#ifdef KTR 361 if (tp->t_state == TCPS_SYN_SENT) { 362 CTR6(KTR_CXGBE, "%s: atid %d, toep %p (0x%x), inp %p (0x%x)", 363 __func__, toep->tid, toep, toep->flags, inp, 364 inp->inp_flags); 365 } else { 366 CTR6(KTR_CXGBE, 367 "t4_pcb_detach: tid %d (%s), toep %p (0x%x), inp %p (0x%x)", 368 toep->tid, tcpstates[tp->t_state], toep, toep->flags, inp, 369 inp->inp_flags); 370 } 371#endif 372 373 tp->t_toe = NULL; 374 tp->t_flags &= ~TF_TOE; 375 toep->flags &= ~TPF_ATTACHED; 376 377 if (!(toep->flags & TPF_CPL_PENDING)) 378 release_offload_resources(toep); 379} 380 381/* 382 * setsockopt handler. 383 */ 384static void 385t4_ctloutput(struct toedev *tod, struct tcpcb *tp, int dir, int name) 386{ 387 struct adapter *sc = tod->tod_softc; 388 struct toepcb *toep = tp->t_toe; 389 390 if (dir == SOPT_GET) 391 return; 392 393 CTR4(KTR_CXGBE, "%s: tp %p, dir %u, name %u", __func__, tp, dir, name); 394 395 switch (name) { 396 case TCP_NODELAY: 397 t4_set_tcb_field(sc, toep->ctrlq, toep->tid, W_TCB_T_FLAGS, 398 V_TF_NAGLE(1), V_TF_NAGLE(tp->t_flags & TF_NODELAY ? 0 : 1), 399 0, 0, toep->ofld_rxq->iq.abs_id); 400 break; 401 default: 402 break; 403 } 404} 405 406/* 407 * The TOE driver will not receive any more CPLs for the tid associated with the 408 * toepcb; release the hold on the inpcb. 409 */ 410void 411final_cpl_received(struct toepcb *toep) 412{ 413 struct inpcb *inp = toep->inp; 414 415 KASSERT(inp != NULL, ("%s: inp is NULL", __func__)); 416 INP_WLOCK_ASSERT(inp); 417 KASSERT(toep->flags & TPF_CPL_PENDING, 418 ("%s: CPL not pending already?", __func__)); 419 420 CTR6(KTR_CXGBE, "%s: tid %d, toep %p (0x%x), inp %p (0x%x)", 421 __func__, toep->tid, toep, toep->flags, inp, inp->inp_flags); 422 423 toep->inp = NULL; 424 toep->flags &= ~TPF_CPL_PENDING; 425 mbufq_drain(&toep->ulp_pdu_reclaimq); 426 427 if (!(toep->flags & TPF_ATTACHED)) 428 release_offload_resources(toep); 429 430 if (!in_pcbrele_wlocked(inp)) 431 INP_WUNLOCK(inp); 432} 433 434void 435insert_tid(struct adapter *sc, int tid, void *ctx) 436{ 437 struct tid_info *t = &sc->tids; 438 439 t->tid_tab[tid] = ctx; 440 atomic_add_int(&t->tids_in_use, 1); 441} 442 443void * 444lookup_tid(struct adapter *sc, int tid) 445{ 446 struct tid_info *t = &sc->tids; 447 448 return (t->tid_tab[tid]); 449} 450 451void 452update_tid(struct adapter *sc, int tid, void *ctx) 453{ 454 struct tid_info *t = &sc->tids; 455 456 t->tid_tab[tid] = ctx; 457} 458 459void 460remove_tid(struct adapter *sc, int tid) 461{ 462 struct tid_info *t = &sc->tids; 463 464 t->tid_tab[tid] = NULL; 465 atomic_subtract_int(&t->tids_in_use, 1); 466} 467 468void 469release_tid(struct adapter *sc, int tid, struct sge_wrq *ctrlq) 470{ 471 struct wrqe *wr; 472 struct cpl_tid_release *req; 473 474 wr = alloc_wrqe(sizeof(*req), ctrlq); 475 if (wr == NULL) { 476 queue_tid_release(sc, tid); /* defer */ 477 return; 478 } 479 req = wrtod(wr); 480 481 INIT_TP_WR_MIT_CPL(req, CPL_TID_RELEASE, tid); 482 483 t4_wrq_tx(sc, wr); 484} 485 486static void 487queue_tid_release(struct adapter *sc, int tid) 488{ 489 490 CXGBE_UNIMPLEMENTED("deferred tid release"); 491} 492 493/* 494 * What mtu_idx to use, given a 4-tuple and/or an MSS cap 495 */ 496int 497find_best_mtu_idx(struct adapter *sc, struct in_conninfo *inc, int pmss) 498{ 499 unsigned short *mtus = &sc->params.mtus[0]; 500 int i, mss, n; 501 502 KASSERT(inc != NULL || pmss > 0, 503 ("%s: at least one of inc/pmss must be specified", __func__)); 504 505 mss = inc ? tcp_mssopt(inc) : pmss; 506 if (pmss > 0 && mss > pmss) 507 mss = pmss; 508 509 if (inc->inc_flags & INC_ISIPV6) 510 n = sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 511 else 512 n = sizeof(struct ip) + sizeof(struct tcphdr); 513 514 for (i = 0; i < NMTUS - 1 && mtus[i + 1] <= mss + n; i++) 515 continue; 516 517 return (i); 518} 519 520/* 521 * Determine the receive window size for a socket. 522 */ 523u_long 524select_rcv_wnd(struct socket *so) 525{ 526 unsigned long wnd; 527 528 SOCKBUF_LOCK_ASSERT(&so->so_rcv); 529 530 wnd = sbspace(&so->so_rcv); 531 if (wnd < MIN_RCV_WND) 532 wnd = MIN_RCV_WND; 533 534 return min(wnd, MAX_RCV_WND); 535} 536 537int 538select_rcv_wscale(void) 539{ 540 int wscale = 0; 541 unsigned long space = sb_max; 542 543 if (space > MAX_RCV_WND) 544 space = MAX_RCV_WND; 545 546 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < space) 547 wscale++; 548 549 return (wscale); 550} 551 552extern int always_keepalive; 553 554/* 555 * socket so could be a listening socket too. 556 */ 557uint64_t 558calc_opt0(struct socket *so, struct vi_info *vi, struct l2t_entry *e, 559 int mtu_idx, int rscale, int rx_credits, int ulp_mode) 560{ 561 uint64_t opt0; 562 563 KASSERT(rx_credits <= M_RCV_BUFSIZ, 564 ("%s: rcv_bufsiz too high", __func__)); 565 566 opt0 = F_TCAM_BYPASS | V_WND_SCALE(rscale) | V_MSS_IDX(mtu_idx) | 567 V_ULP_MODE(ulp_mode) | V_RCV_BUFSIZ(rx_credits); 568 569 if (so != NULL) { 570 struct inpcb *inp = sotoinpcb(so); 571 struct tcpcb *tp = intotcpcb(inp); 572 int keepalive = always_keepalive || 573 so_options_get(so) & SO_KEEPALIVE; 574 575 opt0 |= V_NAGLE((tp->t_flags & TF_NODELAY) == 0); 576 opt0 |= V_KEEP_ALIVE(keepalive != 0); 577 } 578 579 if (e != NULL) 580 opt0 |= V_L2T_IDX(e->idx); 581 582 if (vi != NULL) { 583 opt0 |= V_SMAC_SEL(vi->smt_idx); 584 opt0 |= V_TX_CHAN(vi->pi->tx_chan); 585 } 586 587 return htobe64(opt0); 588} 589 590uint64_t 591select_ntuple(struct vi_info *vi, struct l2t_entry *e) 592{ 593 struct adapter *sc = vi->pi->adapter; 594 struct tp_params *tp = &sc->params.tp; 595 uint16_t viid = vi->viid; 596 uint64_t ntuple = 0; 597 598 /* 599 * Initialize each of the fields which we care about which are present 600 * in the Compressed Filter Tuple. 601 */ 602 if (tp->vlan_shift >= 0 && e->vlan != CPL_L2T_VLAN_NONE) 603 ntuple |= (uint64_t)(F_FT_VLAN_VLD | e->vlan) << tp->vlan_shift; 604 605 if (tp->port_shift >= 0) 606 ntuple |= (uint64_t)e->lport << tp->port_shift; 607 608 if (tp->protocol_shift >= 0) 609 ntuple |= (uint64_t)IPPROTO_TCP << tp->protocol_shift; 610 611 if (tp->vnic_shift >= 0) { 612 uint32_t vf = G_FW_VIID_VIN(viid); 613 uint32_t pf = G_FW_VIID_PFN(viid); 614 uint32_t vld = G_FW_VIID_VIVLD(viid); 615 616 ntuple |= (uint64_t)(V_FT_VNID_ID_VF(vf) | V_FT_VNID_ID_PF(pf) | 617 V_FT_VNID_ID_VLD(vld)) << tp->vnic_shift; 618 } 619 620 if (is_t4(sc)) 621 return (htobe32((uint32_t)ntuple)); 622 else 623 return (htobe64(V_FILTER_TUPLE(ntuple))); 624} 625 626void 627set_tcpddp_ulp_mode(struct toepcb *toep) 628{ 629 630 toep->ulp_mode = ULP_MODE_TCPDDP; 631 toep->ddp_flags = DDP_OK; 632 toep->ddp_score = DDP_LOW_SCORE; 633} 634 635int 636negative_advice(int status) 637{ 638 639 return (status == CPL_ERR_RTX_NEG_ADVICE || 640 status == CPL_ERR_PERSIST_NEG_ADVICE || 641 status == CPL_ERR_KEEPALV_NEG_ADVICE); 642} 643 644static int 645alloc_tid_tabs(struct tid_info *t) 646{ 647 size_t size; 648 unsigned int i; 649 650 size = t->ntids * sizeof(*t->tid_tab) + 651 t->natids * sizeof(*t->atid_tab) + 652 t->nstids * sizeof(*t->stid_tab); 653 654 t->tid_tab = malloc(size, M_CXGBE, M_ZERO | M_NOWAIT); 655 if (t->tid_tab == NULL) 656 return (ENOMEM); 657 658 mtx_init(&t->atid_lock, "atid lock", NULL, MTX_DEF); 659 t->atid_tab = (union aopen_entry *)&t->tid_tab[t->ntids]; 660 t->afree = t->atid_tab; 661 t->atids_in_use = 0; 662 for (i = 1; i < t->natids; i++) 663 t->atid_tab[i - 1].next = &t->atid_tab[i]; 664 t->atid_tab[t->natids - 1].next = NULL; 665 666 mtx_init(&t->stid_lock, "stid lock", NULL, MTX_DEF); 667 t->stid_tab = (struct listen_ctx **)&t->atid_tab[t->natids]; 668 t->stids_in_use = 0; 669 TAILQ_INIT(&t->stids); 670 t->nstids_free_head = t->nstids; 671 672 atomic_store_rel_int(&t->tids_in_use, 0); 673 674 return (0); 675} 676 677static void 678free_tid_tabs(struct tid_info *t) 679{ 680 KASSERT(t->tids_in_use == 0, 681 ("%s: %d tids still in use.", __func__, t->tids_in_use)); 682 KASSERT(t->atids_in_use == 0, 683 ("%s: %d atids still in use.", __func__, t->atids_in_use)); 684 KASSERT(t->stids_in_use == 0, 685 ("%s: %d tids still in use.", __func__, t->stids_in_use)); 686 687 free(t->tid_tab, M_CXGBE); 688 t->tid_tab = NULL; 689 690 if (mtx_initialized(&t->atid_lock)) 691 mtx_destroy(&t->atid_lock); 692 if (mtx_initialized(&t->stid_lock)) 693 mtx_destroy(&t->stid_lock); 694} 695 696static int 697add_lip(struct adapter *sc, struct in6_addr *lip) 698{ 699 struct fw_clip_cmd c; 700 701 ASSERT_SYNCHRONIZED_OP(sc); 702 /* mtx_assert(&td->clip_table_lock, MA_OWNED); */ 703 704 memset(&c, 0, sizeof(c)); 705 c.op_to_write = htonl(V_FW_CMD_OP(FW_CLIP_CMD) | F_FW_CMD_REQUEST | 706 F_FW_CMD_WRITE); 707 c.alloc_to_len16 = htonl(F_FW_CLIP_CMD_ALLOC | FW_LEN16(c)); 708 c.ip_hi = *(uint64_t *)&lip->s6_addr[0]; 709 c.ip_lo = *(uint64_t *)&lip->s6_addr[8]; 710 711 return (-t4_wr_mbox_ns(sc, sc->mbox, &c, sizeof(c), &c)); 712} 713 714static int 715delete_lip(struct adapter *sc, struct in6_addr *lip) 716{ 717 struct fw_clip_cmd c; 718 719 ASSERT_SYNCHRONIZED_OP(sc); 720 /* mtx_assert(&td->clip_table_lock, MA_OWNED); */ 721 722 memset(&c, 0, sizeof(c)); 723 c.op_to_write = htonl(V_FW_CMD_OP(FW_CLIP_CMD) | F_FW_CMD_REQUEST | 724 F_FW_CMD_READ); 725 c.alloc_to_len16 = htonl(F_FW_CLIP_CMD_FREE | FW_LEN16(c)); 726 c.ip_hi = *(uint64_t *)&lip->s6_addr[0]; 727 c.ip_lo = *(uint64_t *)&lip->s6_addr[8]; 728 729 return (-t4_wr_mbox_ns(sc, sc->mbox, &c, sizeof(c), &c)); 730} 731 732static struct clip_entry * 733search_lip(struct tom_data *td, struct in6_addr *lip) 734{ 735 struct clip_entry *ce; 736 737 mtx_assert(&td->clip_table_lock, MA_OWNED); 738 739 TAILQ_FOREACH(ce, &td->clip_table, link) { 740 if (IN6_ARE_ADDR_EQUAL(&ce->lip, lip)) 741 return (ce); 742 } 743 744 return (NULL); 745} 746 747struct clip_entry * 748hold_lip(struct tom_data *td, struct in6_addr *lip) 749{ 750 struct clip_entry *ce; 751 752 mtx_lock(&td->clip_table_lock); 753 ce = search_lip(td, lip); 754 if (ce != NULL) 755 ce->refcount++; 756 mtx_unlock(&td->clip_table_lock); 757 758 return (ce); 759} 760 761void 762release_lip(struct tom_data *td, struct clip_entry *ce) 763{ 764 765 mtx_lock(&td->clip_table_lock); 766 KASSERT(search_lip(td, &ce->lip) == ce, 767 ("%s: CLIP entry %p p not in CLIP table.", __func__, ce)); 768 KASSERT(ce->refcount > 0, 769 ("%s: CLIP entry %p has refcount 0", __func__, ce)); 770 --ce->refcount; 771 mtx_unlock(&td->clip_table_lock); 772} 773 774static void 775init_clip_table(struct adapter *sc, struct tom_data *td) 776{ 777 778 ASSERT_SYNCHRONIZED_OP(sc); 779 780 mtx_init(&td->clip_table_lock, "CLIP table lock", NULL, MTX_DEF); 781 TAILQ_INIT(&td->clip_table); 782 td->clip_gen = -1; 783 784 update_clip_table(sc, td); 785} 786 787static void 788update_clip(struct adapter *sc, void *arg __unused) 789{ 790 791 if (begin_synchronized_op(sc, NULL, HOLD_LOCK, "t4tomuc")) 792 return; 793 794 if (uld_active(sc, ULD_TOM)) 795 update_clip_table(sc, sc->tom_softc); 796 797 end_synchronized_op(sc, LOCK_HELD); 798} 799 800static void 801t4_clip_task(void *arg, int count) 802{ 803 804 t4_iterate(update_clip, NULL); 805} 806 807static void 808update_clip_table(struct adapter *sc, struct tom_data *td) 809{ 810 struct in6_ifaddr *ia; 811 struct in6_addr *lip, tlip; 812 struct clip_head stale; 813 struct clip_entry *ce, *ce_temp; 814 int rc, gen = atomic_load_acq_int(&in6_ifaddr_gen); 815 816 ASSERT_SYNCHRONIZED_OP(sc); 817 818 IN6_IFADDR_RLOCK(); 819 mtx_lock(&td->clip_table_lock); 820 821 if (gen == td->clip_gen) 822 goto done; 823 824 TAILQ_INIT(&stale); 825 TAILQ_CONCAT(&stale, &td->clip_table, link); 826 827 TAILQ_FOREACH(ia, &V_in6_ifaddrhead, ia_link) { 828 lip = &ia->ia_addr.sin6_addr; 829 830 KASSERT(!IN6_IS_ADDR_MULTICAST(lip), 831 ("%s: mcast address in in6_ifaddr list", __func__)); 832 833 if (IN6_IS_ADDR_LOOPBACK(lip)) 834 continue; 835 if (IN6_IS_SCOPE_EMBED(lip)) { 836 /* Remove the embedded scope */ 837 tlip = *lip; 838 lip = &tlip; 839 in6_clearscope(lip); 840 } 841 /* 842 * XXX: how to weed out the link local address for the loopback 843 * interface? It's fe80::1 usually (always?). 844 */ 845 846 /* 847 * If it's in the main list then we already know it's not stale. 848 */ 849 TAILQ_FOREACH(ce, &td->clip_table, link) { 850 if (IN6_ARE_ADDR_EQUAL(&ce->lip, lip)) 851 goto next; 852 } 853 854 /* 855 * If it's in the stale list we should move it to the main list. 856 */ 857 TAILQ_FOREACH(ce, &stale, link) { 858 if (IN6_ARE_ADDR_EQUAL(&ce->lip, lip)) { 859 TAILQ_REMOVE(&stale, ce, link); 860 TAILQ_INSERT_TAIL(&td->clip_table, ce, link); 861 goto next; 862 } 863 } 864 865 /* A new IP6 address; add it to the CLIP table */ 866 ce = malloc(sizeof(*ce), M_CXGBE, M_NOWAIT); 867 memcpy(&ce->lip, lip, sizeof(ce->lip)); 868 ce->refcount = 0; 869 rc = add_lip(sc, lip); 870 if (rc == 0) 871 TAILQ_INSERT_TAIL(&td->clip_table, ce, link); 872 else { 873 char ip[INET6_ADDRSTRLEN]; 874 875 inet_ntop(AF_INET6, &ce->lip, &ip[0], sizeof(ip)); 876 log(LOG_ERR, "%s: could not add %s (%d)\n", 877 __func__, ip, rc); 878 free(ce, M_CXGBE); 879 } 880next: 881 continue; 882 } 883 884 /* 885 * Remove stale addresses (those no longer in V_in6_ifaddrhead) that are 886 * no longer referenced by the driver. 887 */ 888 TAILQ_FOREACH_SAFE(ce, &stale, link, ce_temp) { 889 if (ce->refcount == 0) { 890 rc = delete_lip(sc, &ce->lip); 891 if (rc == 0) { 892 TAILQ_REMOVE(&stale, ce, link); 893 free(ce, M_CXGBE); 894 } else { 895 char ip[INET6_ADDRSTRLEN]; 896 897 inet_ntop(AF_INET6, &ce->lip, &ip[0], 898 sizeof(ip)); 899 log(LOG_ERR, "%s: could not delete %s (%d)\n", 900 __func__, ip, rc); 901 } 902 } 903 } 904 /* The ones that are still referenced need to stay in the CLIP table */ 905 TAILQ_CONCAT(&td->clip_table, &stale, link); 906 907 td->clip_gen = gen; 908done: 909 mtx_unlock(&td->clip_table_lock); 910 IN6_IFADDR_RUNLOCK(); 911} 912 913static void 914destroy_clip_table(struct adapter *sc, struct tom_data *td) 915{ 916 struct clip_entry *ce, *ce_temp; 917 918 if (mtx_initialized(&td->clip_table_lock)) { 919 mtx_lock(&td->clip_table_lock); 920 TAILQ_FOREACH_SAFE(ce, &td->clip_table, link, ce_temp) { 921 KASSERT(ce->refcount == 0, 922 ("%s: CLIP entry %p still in use (%d)", __func__, 923 ce, ce->refcount)); 924 TAILQ_REMOVE(&td->clip_table, ce, link); 925 delete_lip(sc, &ce->lip); 926 free(ce, M_CXGBE); 927 } 928 mtx_unlock(&td->clip_table_lock); 929 mtx_destroy(&td->clip_table_lock); 930 } 931} 932 933static void 934free_tom_data(struct adapter *sc, struct tom_data *td) 935{ 936 937 ASSERT_SYNCHRONIZED_OP(sc); 938 939 KASSERT(TAILQ_EMPTY(&td->toep_list), 940 ("%s: TOE PCB list is not empty.", __func__)); 941 KASSERT(td->lctx_count == 0, 942 ("%s: lctx hash table is not empty.", __func__)); 943 944 t4_free_ppod_region(&td->pr); 945 destroy_clip_table(sc, td); 946 947 if (td->listen_mask != 0) 948 hashdestroy(td->listen_hash, M_CXGBE, td->listen_mask); 949 950 if (mtx_initialized(&td->unsent_wr_lock)) 951 mtx_destroy(&td->unsent_wr_lock); 952 if (mtx_initialized(&td->lctx_hash_lock)) 953 mtx_destroy(&td->lctx_hash_lock); 954 if (mtx_initialized(&td->toep_list_lock)) 955 mtx_destroy(&td->toep_list_lock); 956 957 free_tid_tabs(&sc->tids); 958 free(td, M_CXGBE); 959} 960 961static void 962reclaim_wr_resources(void *arg, int count) 963{ 964 struct tom_data *td = arg; 965 STAILQ_HEAD(, wrqe) twr_list = STAILQ_HEAD_INITIALIZER(twr_list); 966 struct cpl_act_open_req *cpl; 967 u_int opcode, atid; 968 struct wrqe *wr; 969 struct adapter *sc; 970 971 mtx_lock(&td->unsent_wr_lock); 972 STAILQ_SWAP(&td->unsent_wr_list, &twr_list, wrqe); 973 mtx_unlock(&td->unsent_wr_lock); 974 975 while ((wr = STAILQ_FIRST(&twr_list)) != NULL) { 976 STAILQ_REMOVE_HEAD(&twr_list, link); 977 978 cpl = wrtod(wr); 979 opcode = GET_OPCODE(cpl); 980 981 switch (opcode) { 982 case CPL_ACT_OPEN_REQ: 983 case CPL_ACT_OPEN_REQ6: 984 atid = G_TID_TID(be32toh(OPCODE_TID(cpl))); 985 sc = td_adapter(td); 986 987 CTR2(KTR_CXGBE, "%s: atid %u ", __func__, atid); 988 act_open_failure_cleanup(sc, atid, EHOSTUNREACH); 989 free(wr, M_CXGBE); 990 break; 991 default: 992 log(LOG_ERR, "%s: leaked work request %p, wr_len %d, " 993 "opcode %x\n", __func__, wr, wr->wr_len, opcode); 994 /* WR not freed here; go look at it with a debugger. */ 995 } 996 } 997} 998 999/* 1000 * Ground control to Major TOM 1001 * Commencing countdown, engines on 1002 */ 1003static int 1004t4_tom_activate(struct adapter *sc) 1005{ 1006 struct tom_data *td; 1007 struct toedev *tod; 1008 struct vi_info *vi; 1009 struct sge_ofld_rxq *ofld_rxq; 1010 int i, j, rc, v; 1011 1012 ASSERT_SYNCHRONIZED_OP(sc); 1013 1014 /* per-adapter softc for TOM */ 1015 td = malloc(sizeof(*td), M_CXGBE, M_ZERO | M_NOWAIT); 1016 if (td == NULL) 1017 return (ENOMEM); 1018 1019 /* List of TOE PCBs and associated lock */ 1020 mtx_init(&td->toep_list_lock, "PCB list lock", NULL, MTX_DEF); 1021 TAILQ_INIT(&td->toep_list); 1022 1023 /* Listen context */ 1024 mtx_init(&td->lctx_hash_lock, "lctx hash lock", NULL, MTX_DEF); 1025 td->listen_hash = hashinit_flags(LISTEN_HASH_SIZE, M_CXGBE, 1026 &td->listen_mask, HASH_NOWAIT); 1027 1028 /* List of WRs for which L2 resolution failed */ 1029 mtx_init(&td->unsent_wr_lock, "Unsent WR list lock", NULL, MTX_DEF); 1030 STAILQ_INIT(&td->unsent_wr_list); 1031 TASK_INIT(&td->reclaim_wr_resources, 0, reclaim_wr_resources, td); 1032 1033 /* TID tables */ 1034 rc = alloc_tid_tabs(&sc->tids); 1035 if (rc != 0) 1036 goto done; 1037 1038 rc = t4_init_ppod_region(&td->pr, &sc->vres.ddp, 1039 t4_read_reg(sc, A_ULP_RX_TDDP_PSZ), "TDDP page pods"); 1040 if (rc != 0) 1041 goto done; 1042 t4_set_reg_field(sc, A_ULP_RX_TDDP_TAGMASK, 1043 V_TDDPTAGMASK(M_TDDPTAGMASK), td->pr.pr_tag_mask); 1044 1045 /* CLIP table for IPv6 offload */ 1046 init_clip_table(sc, td); 1047 1048 /* toedev ops */ 1049 tod = &td->tod; 1050 init_toedev(tod); 1051 tod->tod_softc = sc; 1052 tod->tod_connect = t4_connect; 1053 tod->tod_listen_start = t4_listen_start; 1054 tod->tod_listen_stop = t4_listen_stop; 1055 tod->tod_rcvd = t4_rcvd; 1056 tod->tod_output = t4_tod_output; 1057 tod->tod_send_rst = t4_send_rst; 1058 tod->tod_send_fin = t4_send_fin; 1059 tod->tod_pcb_detach = t4_pcb_detach; 1060 tod->tod_l2_update = t4_l2_update; 1061 tod->tod_syncache_added = t4_syncache_added; 1062 tod->tod_syncache_removed = t4_syncache_removed; 1063 tod->tod_syncache_respond = t4_syncache_respond; 1064 tod->tod_offload_socket = t4_offload_socket; 1065 tod->tod_ctloutput = t4_ctloutput; 1066 1067 for_each_port(sc, i) { 1068 for_each_vi(sc->port[i], v, vi) { 1069 TOEDEV(vi->ifp) = &td->tod; 1070 for_each_ofld_rxq(vi, j, ofld_rxq) { 1071 ofld_rxq->iq.set_tcb_rpl = do_set_tcb_rpl; 1072 ofld_rxq->iq.l2t_write_rpl = do_l2t_write_rpl2; 1073 } 1074 } 1075 } 1076 1077 sc->tom_softc = td; 1078 register_toedev(sc->tom_softc); 1079 1080done: 1081 if (rc != 0) 1082 free_tom_data(sc, td); 1083 return (rc); 1084} 1085 1086static int 1087t4_tom_deactivate(struct adapter *sc) 1088{ 1089 int rc = 0; 1090 struct tom_data *td = sc->tom_softc; 1091 1092 ASSERT_SYNCHRONIZED_OP(sc); 1093 1094 if (td == NULL) 1095 return (0); /* XXX. KASSERT? */ 1096 1097 if (sc->offload_map != 0) 1098 return (EBUSY); /* at least one port has IFCAP_TOE enabled */ 1099 1100 if (uld_active(sc, ULD_IWARP) || uld_active(sc, ULD_ISCSI)) 1101 return (EBUSY); /* both iWARP and iSCSI rely on the TOE. */ 1102 1103 mtx_lock(&td->toep_list_lock); 1104 if (!TAILQ_EMPTY(&td->toep_list)) 1105 rc = EBUSY; 1106 mtx_unlock(&td->toep_list_lock); 1107 1108 mtx_lock(&td->lctx_hash_lock); 1109 if (td->lctx_count > 0) 1110 rc = EBUSY; 1111 mtx_unlock(&td->lctx_hash_lock); 1112 1113 taskqueue_drain(taskqueue_thread, &td->reclaim_wr_resources); 1114 mtx_lock(&td->unsent_wr_lock); 1115 if (!STAILQ_EMPTY(&td->unsent_wr_list)) 1116 rc = EBUSY; 1117 mtx_unlock(&td->unsent_wr_lock); 1118 1119 if (rc == 0) { 1120 unregister_toedev(sc->tom_softc); 1121 free_tom_data(sc, td); 1122 sc->tom_softc = NULL; 1123 } 1124 1125 return (rc); 1126} 1127 1128static void 1129t4_tom_ifaddr_event(void *arg __unused, struct ifnet *ifp) 1130{ 1131 1132 atomic_add_rel_int(&in6_ifaddr_gen, 1); 1133 taskqueue_enqueue_timeout(taskqueue_thread, &clip_task, -hz / 4); 1134} 1135 1136static int 1137t4_tom_mod_load(void) 1138{ 1139 int rc; 1140 struct protosw *tcp_protosw, *tcp6_protosw; 1141 1142 /* CPL handlers */ 1143 t4_init_connect_cpl_handlers(); 1144 t4_init_listen_cpl_handlers(); 1145 t4_init_cpl_io_handlers(); 1146 1147 rc = t4_ddp_mod_load(); 1148 if (rc != 0) 1149 return (rc); 1150 1151 tcp_protosw = pffindproto(PF_INET, IPPROTO_TCP, SOCK_STREAM); 1152 if (tcp_protosw == NULL) 1153 return (ENOPROTOOPT); 1154 bcopy(tcp_protosw, &ddp_protosw, sizeof(ddp_protosw)); 1155 bcopy(tcp_protosw->pr_usrreqs, &ddp_usrreqs, sizeof(ddp_usrreqs)); 1156 ddp_usrreqs.pru_soreceive = t4_soreceive_ddp; 1157 ddp_protosw.pr_usrreqs = &ddp_usrreqs; 1158 1159 tcp6_protosw = pffindproto(PF_INET6, IPPROTO_TCP, SOCK_STREAM); 1160 if (tcp6_protosw == NULL) 1161 return (ENOPROTOOPT); 1162 bcopy(tcp6_protosw, &ddp6_protosw, sizeof(ddp6_protosw)); 1163 bcopy(tcp6_protosw->pr_usrreqs, &ddp6_usrreqs, sizeof(ddp6_usrreqs)); 1164 ddp6_usrreqs.pru_soreceive = t4_soreceive_ddp; 1165 ddp6_protosw.pr_usrreqs = &ddp6_usrreqs; 1166 1167 TIMEOUT_TASK_INIT(taskqueue_thread, &clip_task, 0, t4_clip_task, NULL); 1168 ifaddr_evhandler = EVENTHANDLER_REGISTER(ifaddr_event, 1169 t4_tom_ifaddr_event, NULL, EVENTHANDLER_PRI_ANY); 1170 1171 rc = t4_register_uld(&tom_uld_info); 1172 if (rc != 0) 1173 t4_tom_mod_unload(); 1174 1175 return (rc); 1176} 1177 1178static void 1179tom_uninit(struct adapter *sc, void *arg __unused) 1180{ 1181 if (begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4tomun")) 1182 return; 1183 1184 /* Try to free resources (works only if no port has IFCAP_TOE) */ 1185 if (uld_active(sc, ULD_TOM)) 1186 t4_deactivate_uld(sc, ULD_TOM); 1187 1188 end_synchronized_op(sc, 0); 1189} 1190 1191static int 1192t4_tom_mod_unload(void) 1193{ 1194 t4_iterate(tom_uninit, NULL); 1195 1196 if (t4_unregister_uld(&tom_uld_info) == EBUSY) 1197 return (EBUSY); 1198 1199 if (ifaddr_evhandler) { 1200 EVENTHANDLER_DEREGISTER(ifaddr_event, ifaddr_evhandler); 1201 taskqueue_cancel_timeout(taskqueue_thread, &clip_task, NULL); 1202 } 1203 1204 t4_ddp_mod_unload(); 1205 1206 return (0); 1207} 1208#endif /* TCP_OFFLOAD */ 1209 1210static int 1211t4_tom_modevent(module_t mod, int cmd, void *arg) 1212{ 1213 int rc = 0; 1214 1215#ifdef TCP_OFFLOAD 1216 switch (cmd) { 1217 case MOD_LOAD: 1218 rc = t4_tom_mod_load(); 1219 break; 1220 1221 case MOD_UNLOAD: 1222 rc = t4_tom_mod_unload(); 1223 break; 1224 1225 default: 1226 rc = EINVAL; 1227 } 1228#else 1229 printf("t4_tom: compiled without TCP_OFFLOAD support.\n"); 1230 rc = EOPNOTSUPP; 1231#endif 1232 return (rc); 1233} 1234 1235static moduledata_t t4_tom_moddata= { 1236 "t4_tom", 1237 t4_tom_modevent, 1238 0 1239}; 1240 1241MODULE_VERSION(t4_tom, 1); 1242MODULE_DEPEND(t4_tom, toecore, 1, 1, 1); 1243MODULE_DEPEND(t4_tom, t4nex, 1, 1, 1); 1244DECLARE_MODULE(t4_tom, t4_tom_moddata, SI_SUB_EXEC, SI_ORDER_ANY); 1245