t4_tom.c revision 346850
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 346850 2019-04-28 18:44:29Z 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 <net/if_types.h> 52#include <net/if_vlan_var.h> 53#include <netinet/in.h> 54#include <netinet/in_pcb.h> 55#include <netinet/in_var.h> 56#include <netinet/ip.h> 57#include <netinet/ip6.h> 58#include <netinet6/scope6_var.h> 59#define TCPSTATES 60#include <netinet/tcp_fsm.h> 61#include <netinet/tcp_timer.h> 62#include <netinet/tcp_var.h> 63#include <netinet/toecore.h> 64 65#ifdef TCP_OFFLOAD 66#include "common/common.h" 67#include "common/t4_msg.h" 68#include "common/t4_regs.h" 69#include "common/t4_regs_values.h" 70#include "common/t4_tcb.h" 71#include "tom/t4_tom_l2t.h" 72#include "tom/t4_tom.h" 73#include "tom/t4_tls.h" 74 75static struct protosw toe_protosw; 76static struct pr_usrreqs toe_usrreqs; 77 78static struct protosw toe6_protosw; 79static struct pr_usrreqs toe6_usrreqs; 80 81/* Module ops */ 82static int t4_tom_mod_load(void); 83static int t4_tom_mod_unload(void); 84static int t4_tom_modevent(module_t, int, void *); 85 86/* ULD ops and helpers */ 87static int t4_tom_activate(struct adapter *); 88static int t4_tom_deactivate(struct adapter *); 89 90static struct uld_info tom_uld_info = { 91 .uld_id = ULD_TOM, 92 .activate = t4_tom_activate, 93 .deactivate = t4_tom_deactivate, 94}; 95 96static void release_offload_resources(struct toepcb *); 97static int alloc_tid_tabs(struct tid_info *); 98static void free_tid_tabs(struct tid_info *); 99static int add_lip(struct adapter *, struct in6_addr *); 100static int delete_lip(struct adapter *, struct in6_addr *); 101static struct clip_entry *search_lip(struct tom_data *, struct in6_addr *); 102static void init_clip_table(struct adapter *, struct tom_data *); 103static void update_clip(struct adapter *, void *); 104static void t4_clip_task(void *, int); 105static void update_clip_table(struct adapter *, struct tom_data *); 106static void destroy_clip_table(struct adapter *, struct tom_data *); 107static void free_tom_data(struct adapter *, struct tom_data *); 108static void reclaim_wr_resources(void *, int); 109 110static int in6_ifaddr_gen; 111static eventhandler_tag ifaddr_evhandler; 112static struct timeout_task clip_task; 113 114struct toepcb * 115alloc_toepcb(struct vi_info *vi, int txqid, int rxqid, int flags) 116{ 117 struct port_info *pi = vi->pi; 118 struct adapter *sc = pi->adapter; 119 struct toepcb *toep; 120 int tx_credits, txsd_total, len; 121 122 /* 123 * The firmware counts tx work request credits in units of 16 bytes 124 * each. Reserve room for an ABORT_REQ so the driver never has to worry 125 * about tx credits if it wants to abort a connection. 126 */ 127 tx_credits = sc->params.ofldq_wr_cred; 128 tx_credits -= howmany(sizeof(struct cpl_abort_req), 16); 129 130 /* 131 * Shortest possible tx work request is a fw_ofld_tx_data_wr + 1 byte 132 * immediate payload, and firmware counts tx work request credits in 133 * units of 16 byte. Calculate the maximum work requests possible. 134 */ 135 txsd_total = tx_credits / 136 howmany(sizeof(struct fw_ofld_tx_data_wr) + 1, 16); 137 138 KASSERT(txqid >= vi->first_ofld_txq && 139 txqid < vi->first_ofld_txq + vi->nofldtxq, 140 ("%s: txqid %d for vi %p (first %d, n %d)", __func__, txqid, vi, 141 vi->first_ofld_txq, vi->nofldtxq)); 142 143 KASSERT(rxqid >= vi->first_ofld_rxq && 144 rxqid < vi->first_ofld_rxq + vi->nofldrxq, 145 ("%s: rxqid %d for vi %p (first %d, n %d)", __func__, rxqid, vi, 146 vi->first_ofld_rxq, vi->nofldrxq)); 147 148 len = offsetof(struct toepcb, txsd) + 149 txsd_total * sizeof(struct ofld_tx_sdesc); 150 151 toep = malloc(len, M_CXGBE, M_ZERO | flags); 152 if (toep == NULL) 153 return (NULL); 154 155 refcount_init(&toep->refcount, 1); 156 toep->td = sc->tom_softc; 157 toep->vi = vi; 158 toep->tc_idx = -1; 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); 171 172 return (toep); 173} 174 175struct toepcb * 176hold_toepcb(struct toepcb *toep) 177{ 178 179 refcount_acquire(&toep->refcount); 180 return (toep); 181} 182 183void 184free_toepcb(struct toepcb *toep) 185{ 186 187 if (refcount_release(&toep->refcount) == 0) 188 return; 189 190 KASSERT(!(toep->flags & TPF_ATTACHED), 191 ("%s: attached to an inpcb", __func__)); 192 KASSERT(!(toep->flags & TPF_CPL_PENDING), 193 ("%s: CPL pending", __func__)); 194 195 if (toep->ulp_mode == ULP_MODE_TCPDDP) 196 ddp_uninit_toep(toep); 197 tls_uninit_toep(toep); 198 free(toep, M_CXGBE); 199} 200 201/* 202 * Set up the socket for TCP offload. 203 */ 204void 205offload_socket(struct socket *so, struct toepcb *toep) 206{ 207 struct tom_data *td = toep->td; 208 struct inpcb *inp = sotoinpcb(so); 209 struct tcpcb *tp = intotcpcb(inp); 210 struct sockbuf *sb; 211 212 INP_WLOCK_ASSERT(inp); 213 214 /* Update socket */ 215 sb = &so->so_snd; 216 SOCKBUF_LOCK(sb); 217 sb->sb_flags |= SB_NOCOALESCE; 218 SOCKBUF_UNLOCK(sb); 219 sb = &so->so_rcv; 220 SOCKBUF_LOCK(sb); 221 sb->sb_flags |= SB_NOCOALESCE; 222 if (inp->inp_vflag & INP_IPV6) 223 so->so_proto = &toe6_protosw; 224 else 225 so->so_proto = &toe_protosw; 226 SOCKBUF_UNLOCK(sb); 227 228 /* Update TCP PCB */ 229 tp->tod = &td->tod; 230 tp->t_toe = toep; 231 tp->t_flags |= TF_TOE; 232 233 /* Install an extra hold on inp */ 234 toep->inp = inp; 235 toep->flags |= TPF_ATTACHED; 236 in_pcbref(inp); 237 238 /* Add the TOE PCB to the active list */ 239 mtx_lock(&td->toep_list_lock); 240 TAILQ_INSERT_HEAD(&td->toep_list, toep, link); 241 mtx_unlock(&td->toep_list_lock); 242} 243 244/* This is _not_ the normal way to "unoffload" a socket. */ 245void 246undo_offload_socket(struct socket *so) 247{ 248 struct inpcb *inp = sotoinpcb(so); 249 struct tcpcb *tp = intotcpcb(inp); 250 struct toepcb *toep = tp->t_toe; 251 struct tom_data *td = toep->td; 252 struct sockbuf *sb; 253 254 INP_WLOCK_ASSERT(inp); 255 256 sb = &so->so_snd; 257 SOCKBUF_LOCK(sb); 258 sb->sb_flags &= ~SB_NOCOALESCE; 259 SOCKBUF_UNLOCK(sb); 260 sb = &so->so_rcv; 261 SOCKBUF_LOCK(sb); 262 sb->sb_flags &= ~SB_NOCOALESCE; 263 SOCKBUF_UNLOCK(sb); 264 265 tp->tod = NULL; 266 tp->t_toe = NULL; 267 tp->t_flags &= ~TF_TOE; 268 269 toep->inp = NULL; 270 toep->flags &= ~TPF_ATTACHED; 271 if (in_pcbrele_wlocked(inp)) 272 panic("%s: inp freed.", __func__); 273 274 mtx_lock(&td->toep_list_lock); 275 TAILQ_REMOVE(&td->toep_list, toep, link); 276 mtx_unlock(&td->toep_list_lock); 277} 278 279static void 280release_offload_resources(struct toepcb *toep) 281{ 282 struct tom_data *td = toep->td; 283 struct adapter *sc = td_adapter(td); 284 int tid = toep->tid; 285 286 KASSERT(!(toep->flags & TPF_CPL_PENDING), 287 ("%s: %p has CPL pending.", __func__, toep)); 288 KASSERT(!(toep->flags & TPF_ATTACHED), 289 ("%s: %p is still attached.", __func__, toep)); 290 291 CTR5(KTR_CXGBE, "%s: toep %p (tid %d, l2te %p, ce %p)", 292 __func__, toep, tid, toep->l2te, toep->ce); 293 294 /* 295 * These queues should have been emptied at approximately the same time 296 * that a normal connection's socket's so_snd would have been purged or 297 * drained. Do _not_ clean up here. 298 */ 299 MPASS(mbufq_len(&toep->ulp_pduq) == 0); 300 MPASS(mbufq_len(&toep->ulp_pdu_reclaimq) == 0); 301#ifdef INVARIANTS 302 if (toep->ulp_mode == ULP_MODE_TCPDDP) 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#ifdef RATELIMIT 318 if (toep->tc_idx != -1) 319 t4_release_cl_rl_kbps(sc, toep->vi->pi->port_id, toep->tc_idx); 320#endif 321 mtx_lock(&td->toep_list_lock); 322 TAILQ_REMOVE(&td->toep_list, toep, link); 323 mtx_unlock(&td->toep_list_lock); 324 325 free_toepcb(toep); 326} 327 328/* 329 * The kernel is done with the TCP PCB and this is our opportunity to unhook the 330 * toepcb hanging off of it. If the TOE driver is also done with the toepcb (no 331 * pending CPL) then it is time to release all resources tied to the toepcb. 332 * 333 * Also gets called when an offloaded active open fails and the TOM wants the 334 * kernel to take the TCP PCB back. 335 */ 336static void 337t4_pcb_detach(struct toedev *tod __unused, struct tcpcb *tp) 338{ 339#if defined(KTR) || defined(INVARIANTS) 340 struct inpcb *inp = tp->t_inpcb; 341#endif 342 struct toepcb *toep = tp->t_toe; 343 344 INP_WLOCK_ASSERT(inp); 345 346 KASSERT(toep != NULL, ("%s: toep is NULL", __func__)); 347 KASSERT(toep->flags & TPF_ATTACHED, 348 ("%s: not attached", __func__)); 349 350#ifdef KTR 351 if (tp->t_state == TCPS_SYN_SENT) { 352 CTR6(KTR_CXGBE, "%s: atid %d, toep %p (0x%x), inp %p (0x%x)", 353 __func__, toep->tid, toep, toep->flags, inp, 354 inp->inp_flags); 355 } else { 356 CTR6(KTR_CXGBE, 357 "t4_pcb_detach: tid %d (%s), toep %p (0x%x), inp %p (0x%x)", 358 toep->tid, tcpstates[tp->t_state], toep, toep->flags, inp, 359 inp->inp_flags); 360 } 361#endif 362 363 tp->t_toe = NULL; 364 tp->t_flags &= ~TF_TOE; 365 toep->flags &= ~TPF_ATTACHED; 366 367 if (!(toep->flags & TPF_CPL_PENDING)) 368 release_offload_resources(toep); 369} 370 371/* 372 * setsockopt handler. 373 */ 374static void 375t4_ctloutput(struct toedev *tod, struct tcpcb *tp, int dir, int name) 376{ 377 struct adapter *sc = tod->tod_softc; 378 struct toepcb *toep = tp->t_toe; 379 380 if (dir == SOPT_GET) 381 return; 382 383 CTR4(KTR_CXGBE, "%s: tp %p, dir %u, name %u", __func__, tp, dir, name); 384 385 switch (name) { 386 case TCP_NODELAY: 387 if (tp->t_state != TCPS_ESTABLISHED) 388 break; 389 t4_set_tcb_field(sc, toep->ctrlq, toep, W_TCB_T_FLAGS, 390 V_TF_NAGLE(1), V_TF_NAGLE(tp->t_flags & TF_NODELAY ? 0 : 1), 391 0, 0); 392 break; 393 default: 394 break; 395 } 396} 397 398static inline int 399get_tcb_bit(u_char *tcb, int bit) 400{ 401 int ix, shift; 402 403 ix = 127 - (bit >> 3); 404 shift = bit & 0x7; 405 406 return ((tcb[ix] >> shift) & 1); 407} 408 409static inline uint64_t 410get_tcb_bits(u_char *tcb, int hi, int lo) 411{ 412 uint64_t rc = 0; 413 414 while (hi >= lo) { 415 rc = (rc << 1) | get_tcb_bit(tcb, hi); 416 --hi; 417 } 418 419 return (rc); 420} 421 422/* 423 * Called by the kernel to allow the TOE driver to "refine" values filled up in 424 * the tcp_info for an offloaded connection. 425 */ 426static void 427t4_tcp_info(struct toedev *tod, struct tcpcb *tp, struct tcp_info *ti) 428{ 429 int i, j, k, rc; 430 struct adapter *sc = tod->tod_softc; 431 struct toepcb *toep = tp->t_toe; 432 uint32_t addr, v; 433 uint32_t buf[TCB_SIZE / sizeof(uint32_t)]; 434 u_char *tcb, tmp; 435 436 INP_WLOCK_ASSERT(tp->t_inpcb); 437 MPASS(ti != NULL); 438 439 addr = t4_read_reg(sc, A_TP_CMM_TCB_BASE) + toep->tid * TCB_SIZE; 440 rc = read_via_memwin(sc, 2, addr, &buf[0], TCB_SIZE); 441 if (rc != 0) 442 return; 443 444 tcb = (u_char *)&buf[0]; 445 for (i = 0, j = TCB_SIZE - 16; i < j; i += 16, j -= 16) { 446 for (k = 0; k < 16; k++) { 447 tmp = tcb[i + k]; 448 tcb[i + k] = tcb[j + k]; 449 tcb[j + k] = tmp; 450 } 451 } 452 453 ti->tcpi_state = get_tcb_bits(tcb, 115, 112); 454 455 v = get_tcb_bits(tcb, 271, 256); 456 ti->tcpi_rtt = tcp_ticks_to_us(sc, v); 457 458 v = get_tcb_bits(tcb, 287, 272); 459 ti->tcpi_rttvar = tcp_ticks_to_us(sc, v); 460 461 ti->tcpi_snd_ssthresh = get_tcb_bits(tcb, 487, 460); 462 ti->tcpi_snd_cwnd = get_tcb_bits(tcb, 459, 432); 463 ti->tcpi_rcv_nxt = get_tcb_bits(tcb, 553, 522); 464 465 ti->tcpi_snd_nxt = get_tcb_bits(tcb, 319, 288) - 466 get_tcb_bits(tcb, 375, 348); 467 468 /* Receive window being advertised by us. */ 469 ti->tcpi_rcv_space = get_tcb_bits(tcb, 581, 554); 470 471 /* Send window ceiling. */ 472 v = get_tcb_bits(tcb, 159, 144) << get_tcb_bits(tcb, 131, 128); 473 ti->tcpi_snd_wnd = min(v, ti->tcpi_snd_cwnd); 474} 475 476/* 477 * The TOE driver will not receive any more CPLs for the tid associated with the 478 * toepcb; release the hold on the inpcb. 479 */ 480void 481final_cpl_received(struct toepcb *toep) 482{ 483 struct inpcb *inp = toep->inp; 484 485 KASSERT(inp != NULL, ("%s: inp is NULL", __func__)); 486 INP_WLOCK_ASSERT(inp); 487 KASSERT(toep->flags & TPF_CPL_PENDING, 488 ("%s: CPL not pending already?", __func__)); 489 490 CTR6(KTR_CXGBE, "%s: tid %d, toep %p (0x%x), inp %p (0x%x)", 491 __func__, toep->tid, toep, toep->flags, inp, inp->inp_flags); 492 493 if (toep->ulp_mode == ULP_MODE_TCPDDP) 494 release_ddp_resources(toep); 495 toep->inp = NULL; 496 toep->flags &= ~TPF_CPL_PENDING; 497 mbufq_drain(&toep->ulp_pdu_reclaimq); 498 499 if (!(toep->flags & TPF_ATTACHED)) 500 release_offload_resources(toep); 501 502 if (!in_pcbrele_wlocked(inp)) 503 INP_WUNLOCK(inp); 504} 505 506void 507insert_tid(struct adapter *sc, int tid, void *ctx, int ntids) 508{ 509 struct tid_info *t = &sc->tids; 510 511 t->tid_tab[tid] = ctx; 512 atomic_add_int(&t->tids_in_use, ntids); 513} 514 515void * 516lookup_tid(struct adapter *sc, int tid) 517{ 518 struct tid_info *t = &sc->tids; 519 520 return (t->tid_tab[tid]); 521} 522 523void 524update_tid(struct adapter *sc, int tid, void *ctx) 525{ 526 struct tid_info *t = &sc->tids; 527 528 t->tid_tab[tid] = ctx; 529} 530 531void 532remove_tid(struct adapter *sc, int tid, int ntids) 533{ 534 struct tid_info *t = &sc->tids; 535 536 t->tid_tab[tid] = NULL; 537 atomic_subtract_int(&t->tids_in_use, ntids); 538} 539 540/* 541 * What mtu_idx to use, given a 4-tuple. Note that both s->mss and tcp_mssopt 542 * have the MSS that we should advertise in our SYN. Advertised MSS doesn't 543 * account for any TCP options so the effective MSS (only payload, no headers or 544 * options) could be different. We fill up tp->t_maxseg with the effective MSS 545 * at the end of the 3-way handshake. 546 */ 547int 548find_best_mtu_idx(struct adapter *sc, struct in_conninfo *inc, 549 struct offload_settings *s) 550{ 551 unsigned short *mtus = &sc->params.mtus[0]; 552 int i, mss, mtu; 553 554 MPASS(inc != NULL); 555 556 mss = s->mss > 0 ? s->mss : tcp_mssopt(inc); 557 if (inc->inc_flags & INC_ISIPV6) 558 mtu = mss + sizeof(struct ip6_hdr) + sizeof(struct tcphdr); 559 else 560 mtu = mss + sizeof(struct ip) + sizeof(struct tcphdr); 561 562 for (i = 0; i < NMTUS - 1 && mtus[i + 1] <= mtu; i++) 563 continue; 564 565 return (i); 566} 567 568/* 569 * Determine the receive window size for a socket. 570 */ 571u_long 572select_rcv_wnd(struct socket *so) 573{ 574 unsigned long wnd; 575 576 SOCKBUF_LOCK_ASSERT(&so->so_rcv); 577 578 wnd = sbspace(&so->so_rcv); 579 if (wnd < MIN_RCV_WND) 580 wnd = MIN_RCV_WND; 581 582 return min(wnd, MAX_RCV_WND); 583} 584 585int 586select_rcv_wscale(void) 587{ 588 int wscale = 0; 589 unsigned long space = sb_max; 590 591 if (space > MAX_RCV_WND) 592 space = MAX_RCV_WND; 593 594 while (wscale < TCP_MAX_WINSHIFT && (TCP_MAXWIN << wscale) < space) 595 wscale++; 596 597 return (wscale); 598} 599 600/* 601 * socket so could be a listening socket too. 602 */ 603uint64_t 604calc_opt0(struct socket *so, struct vi_info *vi, struct l2t_entry *e, 605 int mtu_idx, int rscale, int rx_credits, int ulp_mode, 606 struct offload_settings *s) 607{ 608 int keepalive; 609 uint64_t opt0; 610 611 MPASS(so != NULL); 612 MPASS(vi != NULL); 613 KASSERT(rx_credits <= M_RCV_BUFSIZ, 614 ("%s: rcv_bufsiz too high", __func__)); 615 616 opt0 = F_TCAM_BYPASS | V_WND_SCALE(rscale) | V_MSS_IDX(mtu_idx) | 617 V_ULP_MODE(ulp_mode) | V_RCV_BUFSIZ(rx_credits) | 618 V_L2T_IDX(e->idx) | V_SMAC_SEL(vi->smt_idx) | 619 V_TX_CHAN(vi->pi->tx_chan); 620 621 keepalive = tcp_always_keepalive || so_options_get(so) & SO_KEEPALIVE; 622 opt0 |= V_KEEP_ALIVE(keepalive != 0); 623 624 if (s->nagle < 0) { 625 struct inpcb *inp = sotoinpcb(so); 626 struct tcpcb *tp = intotcpcb(inp); 627 628 opt0 |= V_NAGLE((tp->t_flags & TF_NODELAY) == 0); 629 } else 630 opt0 |= V_NAGLE(s->nagle != 0); 631 632 return htobe64(opt0); 633} 634 635uint64_t 636select_ntuple(struct vi_info *vi, struct l2t_entry *e) 637{ 638 struct adapter *sc = vi->pi->adapter; 639 struct tp_params *tp = &sc->params.tp; 640 uint16_t viid = vi->viid; 641 uint64_t ntuple = 0; 642 643 /* 644 * Initialize each of the fields which we care about which are present 645 * in the Compressed Filter Tuple. 646 */ 647 if (tp->vlan_shift >= 0 && e->vlan != CPL_L2T_VLAN_NONE) 648 ntuple |= (uint64_t)(F_FT_VLAN_VLD | e->vlan) << tp->vlan_shift; 649 650 if (tp->port_shift >= 0) 651 ntuple |= (uint64_t)e->lport << tp->port_shift; 652 653 if (tp->protocol_shift >= 0) 654 ntuple |= (uint64_t)IPPROTO_TCP << tp->protocol_shift; 655 656 if (tp->vnic_shift >= 0) { 657 uint32_t vf = G_FW_VIID_VIN(viid); 658 uint32_t pf = G_FW_VIID_PFN(viid); 659 uint32_t vld = G_FW_VIID_VIVLD(viid); 660 661 ntuple |= (uint64_t)(V_FT_VNID_ID_VF(vf) | V_FT_VNID_ID_PF(pf) | 662 V_FT_VNID_ID_VLD(vld)) << tp->vnic_shift; 663 } 664 665 if (is_t4(sc)) 666 return (htobe32((uint32_t)ntuple)); 667 else 668 return (htobe64(V_FILTER_TUPLE(ntuple))); 669} 670 671static int 672is_tls_sock(struct socket *so, struct adapter *sc) 673{ 674 struct inpcb *inp = sotoinpcb(so); 675 int i, rc; 676 677 /* XXX: Eventually add a SO_WANT_TLS socket option perhaps? */ 678 rc = 0; 679 ADAPTER_LOCK(sc); 680 for (i = 0; i < sc->tt.num_tls_rx_ports; i++) { 681 if (inp->inp_lport == htons(sc->tt.tls_rx_ports[i]) || 682 inp->inp_fport == htons(sc->tt.tls_rx_ports[i])) { 683 rc = 1; 684 break; 685 } 686 } 687 ADAPTER_UNLOCK(sc); 688 return (rc); 689} 690 691int 692select_ulp_mode(struct socket *so, struct adapter *sc, 693 struct offload_settings *s) 694{ 695 696 if (can_tls_offload(sc) && 697 (s->tls > 0 || (s->tls < 0 && is_tls_sock(so, sc)))) 698 return (ULP_MODE_TLS); 699 else if (s->ddp > 0 || 700 (s->ddp < 0 && sc->tt.ddp && (so->so_options & SO_NO_DDP) == 0)) 701 return (ULP_MODE_TCPDDP); 702 else 703 return (ULP_MODE_NONE); 704} 705 706void 707set_ulp_mode(struct toepcb *toep, int ulp_mode) 708{ 709 710 CTR4(KTR_CXGBE, "%s: toep %p (tid %d) ulp_mode %d", 711 __func__, toep, toep->tid, ulp_mode); 712 toep->ulp_mode = ulp_mode; 713 tls_init_toep(toep); 714 if (toep->ulp_mode == ULP_MODE_TCPDDP) 715 ddp_init_toep(toep); 716} 717 718int 719negative_advice(int status) 720{ 721 722 return (status == CPL_ERR_RTX_NEG_ADVICE || 723 status == CPL_ERR_PERSIST_NEG_ADVICE || 724 status == CPL_ERR_KEEPALV_NEG_ADVICE); 725} 726 727static int 728alloc_tid_tab(struct tid_info *t, int flags) 729{ 730 731 MPASS(t->ntids > 0); 732 MPASS(t->tid_tab == NULL); 733 734 t->tid_tab = malloc(t->ntids * sizeof(*t->tid_tab), M_CXGBE, 735 M_ZERO | flags); 736 if (t->tid_tab == NULL) 737 return (ENOMEM); 738 atomic_store_rel_int(&t->tids_in_use, 0); 739 740 return (0); 741} 742 743static void 744free_tid_tab(struct tid_info *t) 745{ 746 747 KASSERT(t->tids_in_use == 0, 748 ("%s: %d tids still in use.", __func__, t->tids_in_use)); 749 750 free(t->tid_tab, M_CXGBE); 751 t->tid_tab = NULL; 752} 753 754static int 755alloc_stid_tab(struct tid_info *t, int flags) 756{ 757 758 MPASS(t->nstids > 0); 759 MPASS(t->stid_tab == NULL); 760 761 t->stid_tab = malloc(t->nstids * sizeof(*t->stid_tab), M_CXGBE, 762 M_ZERO | flags); 763 if (t->stid_tab == NULL) 764 return (ENOMEM); 765 mtx_init(&t->stid_lock, "stid lock", NULL, MTX_DEF); 766 t->stids_in_use = 0; 767 TAILQ_INIT(&t->stids); 768 t->nstids_free_head = t->nstids; 769 770 return (0); 771} 772 773static void 774free_stid_tab(struct tid_info *t) 775{ 776 777 KASSERT(t->stids_in_use == 0, 778 ("%s: %d tids still in use.", __func__, t->stids_in_use)); 779 780 if (mtx_initialized(&t->stid_lock)) 781 mtx_destroy(&t->stid_lock); 782 free(t->stid_tab, M_CXGBE); 783 t->stid_tab = NULL; 784} 785 786static void 787free_tid_tabs(struct tid_info *t) 788{ 789 790 free_tid_tab(t); 791 free_atid_tab(t); 792 free_stid_tab(t); 793} 794 795static int 796alloc_tid_tabs(struct tid_info *t) 797{ 798 int rc; 799 800 rc = alloc_tid_tab(t, M_NOWAIT); 801 if (rc != 0) 802 goto failed; 803 804 rc = alloc_atid_tab(t, M_NOWAIT); 805 if (rc != 0) 806 goto failed; 807 808 rc = alloc_stid_tab(t, M_NOWAIT); 809 if (rc != 0) 810 goto failed; 811 812 return (0); 813failed: 814 free_tid_tabs(t); 815 return (rc); 816} 817 818static int 819add_lip(struct adapter *sc, struct in6_addr *lip) 820{ 821 struct fw_clip_cmd c; 822 823 ASSERT_SYNCHRONIZED_OP(sc); 824 /* mtx_assert(&td->clip_table_lock, MA_OWNED); */ 825 826 memset(&c, 0, sizeof(c)); 827 c.op_to_write = htonl(V_FW_CMD_OP(FW_CLIP_CMD) | F_FW_CMD_REQUEST | 828 F_FW_CMD_WRITE); 829 c.alloc_to_len16 = htonl(F_FW_CLIP_CMD_ALLOC | FW_LEN16(c)); 830 c.ip_hi = *(uint64_t *)&lip->s6_addr[0]; 831 c.ip_lo = *(uint64_t *)&lip->s6_addr[8]; 832 833 return (-t4_wr_mbox_ns(sc, sc->mbox, &c, sizeof(c), &c)); 834} 835 836static int 837delete_lip(struct adapter *sc, struct in6_addr *lip) 838{ 839 struct fw_clip_cmd c; 840 841 ASSERT_SYNCHRONIZED_OP(sc); 842 /* mtx_assert(&td->clip_table_lock, MA_OWNED); */ 843 844 memset(&c, 0, sizeof(c)); 845 c.op_to_write = htonl(V_FW_CMD_OP(FW_CLIP_CMD) | F_FW_CMD_REQUEST | 846 F_FW_CMD_READ); 847 c.alloc_to_len16 = htonl(F_FW_CLIP_CMD_FREE | FW_LEN16(c)); 848 c.ip_hi = *(uint64_t *)&lip->s6_addr[0]; 849 c.ip_lo = *(uint64_t *)&lip->s6_addr[8]; 850 851 return (-t4_wr_mbox_ns(sc, sc->mbox, &c, sizeof(c), &c)); 852} 853 854static struct clip_entry * 855search_lip(struct tom_data *td, struct in6_addr *lip) 856{ 857 struct clip_entry *ce; 858 859 mtx_assert(&td->clip_table_lock, MA_OWNED); 860 861 TAILQ_FOREACH(ce, &td->clip_table, link) { 862 if (IN6_ARE_ADDR_EQUAL(&ce->lip, lip)) 863 return (ce); 864 } 865 866 return (NULL); 867} 868 869struct clip_entry * 870hold_lip(struct tom_data *td, struct in6_addr *lip, struct clip_entry *ce) 871{ 872 873 mtx_lock(&td->clip_table_lock); 874 if (ce == NULL) 875 ce = search_lip(td, lip); 876 if (ce != NULL) 877 ce->refcount++; 878 mtx_unlock(&td->clip_table_lock); 879 880 return (ce); 881} 882 883void 884release_lip(struct tom_data *td, struct clip_entry *ce) 885{ 886 887 mtx_lock(&td->clip_table_lock); 888 KASSERT(search_lip(td, &ce->lip) == ce, 889 ("%s: CLIP entry %p p not in CLIP table.", __func__, ce)); 890 KASSERT(ce->refcount > 0, 891 ("%s: CLIP entry %p has refcount 0", __func__, ce)); 892 --ce->refcount; 893 mtx_unlock(&td->clip_table_lock); 894} 895 896static void 897init_clip_table(struct adapter *sc, struct tom_data *td) 898{ 899 900 ASSERT_SYNCHRONIZED_OP(sc); 901 902 mtx_init(&td->clip_table_lock, "CLIP table lock", NULL, MTX_DEF); 903 TAILQ_INIT(&td->clip_table); 904 td->clip_gen = -1; 905 906 update_clip_table(sc, td); 907} 908 909static void 910update_clip(struct adapter *sc, void *arg __unused) 911{ 912 913 if (begin_synchronized_op(sc, NULL, HOLD_LOCK, "t4tomuc")) 914 return; 915 916 if (uld_active(sc, ULD_TOM)) 917 update_clip_table(sc, sc->tom_softc); 918 919 end_synchronized_op(sc, LOCK_HELD); 920} 921 922static void 923t4_clip_task(void *arg, int count) 924{ 925 926 t4_iterate(update_clip, NULL); 927} 928 929static void 930update_clip_table(struct adapter *sc, struct tom_data *td) 931{ 932 struct rm_priotracker in6_ifa_tracker; 933 struct in6_ifaddr *ia; 934 struct in6_addr *lip, tlip; 935 struct clip_head stale; 936 struct clip_entry *ce, *ce_temp; 937 struct vi_info *vi; 938 int rc, gen, i, j; 939 uintptr_t last_vnet; 940 941 ASSERT_SYNCHRONIZED_OP(sc); 942 943 IN6_IFADDR_RLOCK(&in6_ifa_tracker); 944 mtx_lock(&td->clip_table_lock); 945 946 gen = atomic_load_acq_int(&in6_ifaddr_gen); 947 if (gen == td->clip_gen) 948 goto done; 949 950 TAILQ_INIT(&stale); 951 TAILQ_CONCAT(&stale, &td->clip_table, link); 952 953 /* 954 * last_vnet optimizes the common cases where all if_vnet = NULL (no 955 * VIMAGE) or all if_vnet = vnet0. 956 */ 957 last_vnet = (uintptr_t)(-1); 958 for_each_port(sc, i) 959 for_each_vi(sc->port[i], j, vi) { 960 if (last_vnet == (uintptr_t)vi->ifp->if_vnet) 961 continue; 962 963 /* XXX: races with if_vmove */ 964 CURVNET_SET(vi->ifp->if_vnet); 965 TAILQ_FOREACH(ia, &V_in6_ifaddrhead, ia_link) { 966 lip = &ia->ia_addr.sin6_addr; 967 968 KASSERT(!IN6_IS_ADDR_MULTICAST(lip), 969 ("%s: mcast address in in6_ifaddr list", __func__)); 970 971 if (IN6_IS_ADDR_LOOPBACK(lip)) 972 continue; 973 if (IN6_IS_SCOPE_EMBED(lip)) { 974 /* Remove the embedded scope */ 975 tlip = *lip; 976 lip = &tlip; 977 in6_clearscope(lip); 978 } 979 /* 980 * XXX: how to weed out the link local address for the 981 * loopback interface? It's fe80::1 usually (always?). 982 */ 983 984 /* 985 * If it's in the main list then we already know it's 986 * not stale. 987 */ 988 TAILQ_FOREACH(ce, &td->clip_table, link) { 989 if (IN6_ARE_ADDR_EQUAL(&ce->lip, lip)) 990 goto next; 991 } 992 993 /* 994 * If it's in the stale list we should move it to the 995 * main list. 996 */ 997 TAILQ_FOREACH(ce, &stale, link) { 998 if (IN6_ARE_ADDR_EQUAL(&ce->lip, lip)) { 999 TAILQ_REMOVE(&stale, ce, link); 1000 TAILQ_INSERT_TAIL(&td->clip_table, ce, 1001 link); 1002 goto next; 1003 } 1004 } 1005 1006 /* A new IP6 address; add it to the CLIP table */ 1007 ce = malloc(sizeof(*ce), M_CXGBE, M_NOWAIT); 1008 memcpy(&ce->lip, lip, sizeof(ce->lip)); 1009 ce->refcount = 0; 1010 rc = add_lip(sc, lip); 1011 if (rc == 0) 1012 TAILQ_INSERT_TAIL(&td->clip_table, ce, link); 1013 else { 1014 char ip[INET6_ADDRSTRLEN]; 1015 1016 inet_ntop(AF_INET6, &ce->lip, &ip[0], 1017 sizeof(ip)); 1018 log(LOG_ERR, "%s: could not add %s (%d)\n", 1019 __func__, ip, rc); 1020 free(ce, M_CXGBE); 1021 } 1022next: 1023 continue; 1024 } 1025 CURVNET_RESTORE(); 1026 last_vnet = (uintptr_t)vi->ifp->if_vnet; 1027 } 1028 1029 /* 1030 * Remove stale addresses (those no longer in V_in6_ifaddrhead) that are 1031 * no longer referenced by the driver. 1032 */ 1033 TAILQ_FOREACH_SAFE(ce, &stale, link, ce_temp) { 1034 if (ce->refcount == 0) { 1035 rc = delete_lip(sc, &ce->lip); 1036 if (rc == 0) { 1037 TAILQ_REMOVE(&stale, ce, link); 1038 free(ce, M_CXGBE); 1039 } else { 1040 char ip[INET6_ADDRSTRLEN]; 1041 1042 inet_ntop(AF_INET6, &ce->lip, &ip[0], 1043 sizeof(ip)); 1044 log(LOG_ERR, "%s: could not delete %s (%d)\n", 1045 __func__, ip, rc); 1046 } 1047 } 1048 } 1049 /* The ones that are still referenced need to stay in the CLIP table */ 1050 TAILQ_CONCAT(&td->clip_table, &stale, link); 1051 1052 td->clip_gen = gen; 1053done: 1054 mtx_unlock(&td->clip_table_lock); 1055 IN6_IFADDR_RUNLOCK(&in6_ifa_tracker); 1056} 1057 1058static void 1059destroy_clip_table(struct adapter *sc, struct tom_data *td) 1060{ 1061 struct clip_entry *ce, *ce_temp; 1062 1063 if (mtx_initialized(&td->clip_table_lock)) { 1064 mtx_lock(&td->clip_table_lock); 1065 TAILQ_FOREACH_SAFE(ce, &td->clip_table, link, ce_temp) { 1066 KASSERT(ce->refcount == 0, 1067 ("%s: CLIP entry %p still in use (%d)", __func__, 1068 ce, ce->refcount)); 1069 TAILQ_REMOVE(&td->clip_table, ce, link); 1070 delete_lip(sc, &ce->lip); 1071 free(ce, M_CXGBE); 1072 } 1073 mtx_unlock(&td->clip_table_lock); 1074 mtx_destroy(&td->clip_table_lock); 1075 } 1076} 1077 1078static void 1079free_tom_data(struct adapter *sc, struct tom_data *td) 1080{ 1081 1082 ASSERT_SYNCHRONIZED_OP(sc); 1083 1084 KASSERT(TAILQ_EMPTY(&td->toep_list), 1085 ("%s: TOE PCB list is not empty.", __func__)); 1086 KASSERT(td->lctx_count == 0, 1087 ("%s: lctx hash table is not empty.", __func__)); 1088 1089 t4_free_ppod_region(&td->pr); 1090 destroy_clip_table(sc, td); 1091 1092 if (td->listen_mask != 0) 1093 hashdestroy(td->listen_hash, M_CXGBE, td->listen_mask); 1094 1095 if (mtx_initialized(&td->unsent_wr_lock)) 1096 mtx_destroy(&td->unsent_wr_lock); 1097 if (mtx_initialized(&td->lctx_hash_lock)) 1098 mtx_destroy(&td->lctx_hash_lock); 1099 if (mtx_initialized(&td->toep_list_lock)) 1100 mtx_destroy(&td->toep_list_lock); 1101 1102 free_tid_tabs(&sc->tids); 1103 free(td, M_CXGBE); 1104} 1105 1106static char * 1107prepare_pkt(int open_type, uint16_t vtag, struct inpcb *inp, int *pktlen, 1108 int *buflen) 1109{ 1110 char *pkt; 1111 struct tcphdr *th; 1112 int ipv6, len; 1113 const int maxlen = 1114 max(sizeof(struct ether_header), sizeof(struct ether_vlan_header)) + 1115 max(sizeof(struct ip), sizeof(struct ip6_hdr)) + 1116 sizeof(struct tcphdr); 1117 1118 MPASS(open_type == OPEN_TYPE_ACTIVE || open_type == OPEN_TYPE_LISTEN); 1119 1120 pkt = malloc(maxlen, M_CXGBE, M_ZERO | M_NOWAIT); 1121 if (pkt == NULL) 1122 return (NULL); 1123 1124 ipv6 = inp->inp_vflag & INP_IPV6; 1125 len = 0; 1126 1127 if (vtag == 0xffff) { 1128 struct ether_header *eh = (void *)pkt; 1129 1130 if (ipv6) 1131 eh->ether_type = htons(ETHERTYPE_IPV6); 1132 else 1133 eh->ether_type = htons(ETHERTYPE_IP); 1134 1135 len += sizeof(*eh); 1136 } else { 1137 struct ether_vlan_header *evh = (void *)pkt; 1138 1139 evh->evl_encap_proto = htons(ETHERTYPE_VLAN); 1140 evh->evl_tag = htons(vtag); 1141 if (ipv6) 1142 evh->evl_proto = htons(ETHERTYPE_IPV6); 1143 else 1144 evh->evl_proto = htons(ETHERTYPE_IP); 1145 1146 len += sizeof(*evh); 1147 } 1148 1149 if (ipv6) { 1150 struct ip6_hdr *ip6 = (void *)&pkt[len]; 1151 1152 ip6->ip6_vfc = IPV6_VERSION; 1153 ip6->ip6_plen = htons(sizeof(struct tcphdr)); 1154 ip6->ip6_nxt = IPPROTO_TCP; 1155 if (open_type == OPEN_TYPE_ACTIVE) { 1156 ip6->ip6_src = inp->in6p_laddr; 1157 ip6->ip6_dst = inp->in6p_faddr; 1158 } else if (open_type == OPEN_TYPE_LISTEN) { 1159 ip6->ip6_src = inp->in6p_laddr; 1160 ip6->ip6_dst = ip6->ip6_src; 1161 } 1162 1163 len += sizeof(*ip6); 1164 } else { 1165 struct ip *ip = (void *)&pkt[len]; 1166 1167 ip->ip_v = IPVERSION; 1168 ip->ip_hl = sizeof(*ip) >> 2; 1169 ip->ip_tos = inp->inp_ip_tos; 1170 ip->ip_len = htons(sizeof(struct ip) + sizeof(struct tcphdr)); 1171 ip->ip_ttl = inp->inp_ip_ttl; 1172 ip->ip_p = IPPROTO_TCP; 1173 if (open_type == OPEN_TYPE_ACTIVE) { 1174 ip->ip_src = inp->inp_laddr; 1175 ip->ip_dst = inp->inp_faddr; 1176 } else if (open_type == OPEN_TYPE_LISTEN) { 1177 ip->ip_src = inp->inp_laddr; 1178 ip->ip_dst = ip->ip_src; 1179 } 1180 1181 len += sizeof(*ip); 1182 } 1183 1184 th = (void *)&pkt[len]; 1185 if (open_type == OPEN_TYPE_ACTIVE) { 1186 th->th_sport = inp->inp_lport; /* network byte order already */ 1187 th->th_dport = inp->inp_fport; /* ditto */ 1188 } else if (open_type == OPEN_TYPE_LISTEN) { 1189 th->th_sport = inp->inp_lport; /* network byte order already */ 1190 th->th_dport = th->th_sport; 1191 } 1192 len += sizeof(th); 1193 1194 *pktlen = *buflen = len; 1195 return (pkt); 1196} 1197 1198const struct offload_settings * 1199lookup_offload_policy(struct adapter *sc, int open_type, struct mbuf *m, 1200 uint16_t vtag, struct inpcb *inp) 1201{ 1202 const struct t4_offload_policy *op; 1203 char *pkt; 1204 struct offload_rule *r; 1205 int i, matched, pktlen, buflen; 1206 static const struct offload_settings allow_offloading_settings = { 1207 .offload = 1, 1208 .rx_coalesce = -1, 1209 .cong_algo = -1, 1210 .sched_class = -1, 1211 .tstamp = -1, 1212 .sack = -1, 1213 .nagle = -1, 1214 .ecn = -1, 1215 .ddp = -1, 1216 .tls = -1, 1217 .txq = -1, 1218 .rxq = -1, 1219 .mss = -1, 1220 }; 1221 static const struct offload_settings disallow_offloading_settings = { 1222 .offload = 0, 1223 /* rest is irrelevant when offload is off. */ 1224 }; 1225 1226 rw_assert(&sc->policy_lock, RA_LOCKED); 1227 1228 /* 1229 * If there's no Connection Offloading Policy attached to the device 1230 * then we need to return a default static policy. If 1231 * "cop_managed_offloading" is true, then we need to disallow 1232 * offloading until a COP is attached to the device. Otherwise we 1233 * allow offloading ... 1234 */ 1235 op = sc->policy; 1236 if (op == NULL) { 1237 if (sc->tt.cop_managed_offloading) 1238 return (&disallow_offloading_settings); 1239 else 1240 return (&allow_offloading_settings); 1241 } 1242 1243 switch (open_type) { 1244 case OPEN_TYPE_ACTIVE: 1245 case OPEN_TYPE_LISTEN: 1246 pkt = prepare_pkt(open_type, vtag, inp, &pktlen, &buflen); 1247 break; 1248 case OPEN_TYPE_PASSIVE: 1249 MPASS(m != NULL); 1250 pkt = mtod(m, char *); 1251 MPASS(*pkt == CPL_PASS_ACCEPT_REQ); 1252 pkt += sizeof(struct cpl_pass_accept_req); 1253 pktlen = m->m_pkthdr.len - sizeof(struct cpl_pass_accept_req); 1254 buflen = m->m_len - sizeof(struct cpl_pass_accept_req); 1255 break; 1256 default: 1257 MPASS(0); 1258 return (&disallow_offloading_settings); 1259 } 1260 1261 if (pkt == NULL || pktlen == 0 || buflen == 0) 1262 return (&disallow_offloading_settings); 1263 1264 r = &op->rule[0]; 1265 for (i = 0; i < op->nrules; i++, r++) { 1266 if (r->open_type != open_type && 1267 r->open_type != OPEN_TYPE_DONTCARE) { 1268 continue; 1269 } 1270 matched = bpf_filter(r->bpf_prog.bf_insns, pkt, pktlen, buflen); 1271 if (matched) 1272 break; 1273 } 1274 1275 if (open_type == OPEN_TYPE_ACTIVE || open_type == OPEN_TYPE_LISTEN) 1276 free(pkt, M_CXGBE); 1277 1278 return (matched ? &r->settings : &disallow_offloading_settings); 1279} 1280 1281static void 1282reclaim_wr_resources(void *arg, int count) 1283{ 1284 struct tom_data *td = arg; 1285 STAILQ_HEAD(, wrqe) twr_list = STAILQ_HEAD_INITIALIZER(twr_list); 1286 struct cpl_act_open_req *cpl; 1287 u_int opcode, atid; 1288 struct wrqe *wr; 1289 struct adapter *sc; 1290 1291 mtx_lock(&td->unsent_wr_lock); 1292 STAILQ_SWAP(&td->unsent_wr_list, &twr_list, wrqe); 1293 mtx_unlock(&td->unsent_wr_lock); 1294 1295 while ((wr = STAILQ_FIRST(&twr_list)) != NULL) { 1296 STAILQ_REMOVE_HEAD(&twr_list, link); 1297 1298 cpl = wrtod(wr); 1299 opcode = GET_OPCODE(cpl); 1300 1301 switch (opcode) { 1302 case CPL_ACT_OPEN_REQ: 1303 case CPL_ACT_OPEN_REQ6: 1304 atid = G_TID_TID(be32toh(OPCODE_TID(cpl))); 1305 sc = td_adapter(td); 1306 1307 CTR2(KTR_CXGBE, "%s: atid %u ", __func__, atid); 1308 act_open_failure_cleanup(sc, atid, EHOSTUNREACH); 1309 free(wr, M_CXGBE); 1310 break; 1311 default: 1312 log(LOG_ERR, "%s: leaked work request %p, wr_len %d, " 1313 "opcode %x\n", __func__, wr, wr->wr_len, opcode); 1314 /* WR not freed here; go look at it with a debugger. */ 1315 } 1316 } 1317} 1318 1319/* 1320 * Ground control to Major TOM 1321 * Commencing countdown, engines on 1322 */ 1323static int 1324t4_tom_activate(struct adapter *sc) 1325{ 1326 struct tom_data *td; 1327 struct toedev *tod; 1328 struct vi_info *vi; 1329 struct sge_ofld_rxq *ofld_rxq; 1330 int i, j, rc, v; 1331 1332 ASSERT_SYNCHRONIZED_OP(sc); 1333 1334 /* per-adapter softc for TOM */ 1335 td = malloc(sizeof(*td), M_CXGBE, M_ZERO | M_NOWAIT); 1336 if (td == NULL) 1337 return (ENOMEM); 1338 1339 /* List of TOE PCBs and associated lock */ 1340 mtx_init(&td->toep_list_lock, "PCB list lock", NULL, MTX_DEF); 1341 TAILQ_INIT(&td->toep_list); 1342 1343 /* Listen context */ 1344 mtx_init(&td->lctx_hash_lock, "lctx hash lock", NULL, MTX_DEF); 1345 td->listen_hash = hashinit_flags(LISTEN_HASH_SIZE, M_CXGBE, 1346 &td->listen_mask, HASH_NOWAIT); 1347 1348 /* List of WRs for which L2 resolution failed */ 1349 mtx_init(&td->unsent_wr_lock, "Unsent WR list lock", NULL, MTX_DEF); 1350 STAILQ_INIT(&td->unsent_wr_list); 1351 TASK_INIT(&td->reclaim_wr_resources, 0, reclaim_wr_resources, td); 1352 1353 /* TID tables */ 1354 rc = alloc_tid_tabs(&sc->tids); 1355 if (rc != 0) 1356 goto done; 1357 1358 rc = t4_init_ppod_region(&td->pr, &sc->vres.ddp, 1359 t4_read_reg(sc, A_ULP_RX_TDDP_PSZ), "TDDP page pods"); 1360 if (rc != 0) 1361 goto done; 1362 t4_set_reg_field(sc, A_ULP_RX_TDDP_TAGMASK, 1363 V_TDDPTAGMASK(M_TDDPTAGMASK), td->pr.pr_tag_mask); 1364 1365 /* CLIP table for IPv6 offload */ 1366 init_clip_table(sc, td); 1367 1368 /* toedev ops */ 1369 tod = &td->tod; 1370 init_toedev(tod); 1371 tod->tod_softc = sc; 1372 tod->tod_connect = t4_connect; 1373 tod->tod_listen_start = t4_listen_start; 1374 tod->tod_listen_stop = t4_listen_stop; 1375 tod->tod_rcvd = t4_rcvd; 1376 tod->tod_output = t4_tod_output; 1377 tod->tod_send_rst = t4_send_rst; 1378 tod->tod_send_fin = t4_send_fin; 1379 tod->tod_pcb_detach = t4_pcb_detach; 1380 tod->tod_l2_update = t4_l2_update; 1381 tod->tod_syncache_added = t4_syncache_added; 1382 tod->tod_syncache_removed = t4_syncache_removed; 1383 tod->tod_syncache_respond = t4_syncache_respond; 1384 tod->tod_offload_socket = t4_offload_socket; 1385 tod->tod_ctloutput = t4_ctloutput; 1386#if 0 1387 tod->tod_tcp_info = t4_tcp_info; 1388#else 1389 (void)&t4_tcp_info; 1390#endif 1391 1392 for_each_port(sc, i) { 1393 for_each_vi(sc->port[i], v, vi) { 1394 TOEDEV(vi->ifp) = &td->tod; 1395 for_each_ofld_rxq(vi, j, ofld_rxq) { 1396 ofld_rxq->iq.set_tcb_rpl = do_set_tcb_rpl; 1397 ofld_rxq->iq.l2t_write_rpl = do_l2t_write_rpl2; 1398 } 1399 } 1400 } 1401 1402 sc->tom_softc = td; 1403 register_toedev(sc->tom_softc); 1404 1405done: 1406 if (rc != 0) 1407 free_tom_data(sc, td); 1408 return (rc); 1409} 1410 1411static int 1412t4_tom_deactivate(struct adapter *sc) 1413{ 1414 int rc = 0; 1415 struct tom_data *td = sc->tom_softc; 1416 1417 ASSERT_SYNCHRONIZED_OP(sc); 1418 1419 if (td == NULL) 1420 return (0); /* XXX. KASSERT? */ 1421 1422 if (sc->offload_map != 0) 1423 return (EBUSY); /* at least one port has IFCAP_TOE enabled */ 1424 1425 if (uld_active(sc, ULD_IWARP) || uld_active(sc, ULD_ISCSI)) 1426 return (EBUSY); /* both iWARP and iSCSI rely on the TOE. */ 1427 1428 mtx_lock(&td->toep_list_lock); 1429 if (!TAILQ_EMPTY(&td->toep_list)) 1430 rc = EBUSY; 1431 mtx_unlock(&td->toep_list_lock); 1432 1433 mtx_lock(&td->lctx_hash_lock); 1434 if (td->lctx_count > 0) 1435 rc = EBUSY; 1436 mtx_unlock(&td->lctx_hash_lock); 1437 1438 taskqueue_drain(taskqueue_thread, &td->reclaim_wr_resources); 1439 mtx_lock(&td->unsent_wr_lock); 1440 if (!STAILQ_EMPTY(&td->unsent_wr_list)) 1441 rc = EBUSY; 1442 mtx_unlock(&td->unsent_wr_lock); 1443 1444 if (rc == 0) { 1445 unregister_toedev(sc->tom_softc); 1446 free_tom_data(sc, td); 1447 sc->tom_softc = NULL; 1448 } 1449 1450 return (rc); 1451} 1452 1453static void 1454t4_tom_ifaddr_event(void *arg __unused, struct ifnet *ifp) 1455{ 1456 1457 atomic_add_rel_int(&in6_ifaddr_gen, 1); 1458 taskqueue_enqueue_timeout(taskqueue_thread, &clip_task, -hz / 4); 1459} 1460 1461static int 1462t4_aio_queue_tom(struct socket *so, struct kaiocb *job) 1463{ 1464 struct tcpcb *tp = so_sototcpcb(so); 1465 struct toepcb *toep = tp->t_toe; 1466 int error; 1467 1468 if (toep->ulp_mode == ULP_MODE_TCPDDP) { 1469 error = t4_aio_queue_ddp(so, job); 1470 if (error != EOPNOTSUPP) 1471 return (error); 1472 } 1473 1474 return (t4_aio_queue_aiotx(so, job)); 1475} 1476 1477static int 1478t4_ctloutput_tom(struct socket *so, struct sockopt *sopt) 1479{ 1480 1481 if (sopt->sopt_level != IPPROTO_TCP) 1482 return (tcp_ctloutput(so, sopt)); 1483 1484 switch (sopt->sopt_name) { 1485 case TCP_TLSOM_SET_TLS_CONTEXT: 1486 case TCP_TLSOM_GET_TLS_TOM: 1487 case TCP_TLSOM_CLR_TLS_TOM: 1488 case TCP_TLSOM_CLR_QUIES: 1489 return (t4_ctloutput_tls(so, sopt)); 1490 default: 1491 return (tcp_ctloutput(so, sopt)); 1492 } 1493} 1494 1495static int 1496t4_tom_mod_load(void) 1497{ 1498 struct protosw *tcp_protosw, *tcp6_protosw; 1499 1500 /* CPL handlers */ 1501 t4_init_connect_cpl_handlers(); 1502 t4_init_listen_cpl_handlers(); 1503 t4_init_cpl_io_handlers(); 1504 1505 t4_ddp_mod_load(); 1506 t4_tls_mod_load(); 1507 1508 tcp_protosw = pffindproto(PF_INET, IPPROTO_TCP, SOCK_STREAM); 1509 if (tcp_protosw == NULL) 1510 return (ENOPROTOOPT); 1511 bcopy(tcp_protosw, &toe_protosw, sizeof(toe_protosw)); 1512 bcopy(tcp_protosw->pr_usrreqs, &toe_usrreqs, sizeof(toe_usrreqs)); 1513 toe_usrreqs.pru_aio_queue = t4_aio_queue_tom; 1514 toe_protosw.pr_ctloutput = t4_ctloutput_tom; 1515 toe_protosw.pr_usrreqs = &toe_usrreqs; 1516 1517 tcp6_protosw = pffindproto(PF_INET6, IPPROTO_TCP, SOCK_STREAM); 1518 if (tcp6_protosw == NULL) 1519 return (ENOPROTOOPT); 1520 bcopy(tcp6_protosw, &toe6_protosw, sizeof(toe6_protosw)); 1521 bcopy(tcp6_protosw->pr_usrreqs, &toe6_usrreqs, sizeof(toe6_usrreqs)); 1522 toe6_usrreqs.pru_aio_queue = t4_aio_queue_tom; 1523 toe6_protosw.pr_ctloutput = t4_ctloutput_tom; 1524 toe6_protosw.pr_usrreqs = &toe6_usrreqs; 1525 1526 TIMEOUT_TASK_INIT(taskqueue_thread, &clip_task, 0, t4_clip_task, NULL); 1527 ifaddr_evhandler = EVENTHANDLER_REGISTER(ifaddr_event, 1528 t4_tom_ifaddr_event, NULL, EVENTHANDLER_PRI_ANY); 1529 1530 return (t4_register_uld(&tom_uld_info)); 1531} 1532 1533static void 1534tom_uninit(struct adapter *sc, void *arg __unused) 1535{ 1536 if (begin_synchronized_op(sc, NULL, SLEEP_OK | INTR_OK, "t4tomun")) 1537 return; 1538 1539 /* Try to free resources (works only if no port has IFCAP_TOE) */ 1540 if (uld_active(sc, ULD_TOM)) 1541 t4_deactivate_uld(sc, ULD_TOM); 1542 1543 end_synchronized_op(sc, 0); 1544} 1545 1546static int 1547t4_tom_mod_unload(void) 1548{ 1549 t4_iterate(tom_uninit, NULL); 1550 1551 if (t4_unregister_uld(&tom_uld_info) == EBUSY) 1552 return (EBUSY); 1553 1554 if (ifaddr_evhandler) { 1555 EVENTHANDLER_DEREGISTER(ifaddr_event, ifaddr_evhandler); 1556 taskqueue_cancel_timeout(taskqueue_thread, &clip_task, NULL); 1557 } 1558 1559 t4_tls_mod_unload(); 1560 t4_ddp_mod_unload(); 1561 1562 t4_uninit_connect_cpl_handlers(); 1563 t4_uninit_listen_cpl_handlers(); 1564 t4_uninit_cpl_io_handlers(); 1565 1566 return (0); 1567} 1568#endif /* TCP_OFFLOAD */ 1569 1570static int 1571t4_tom_modevent(module_t mod, int cmd, void *arg) 1572{ 1573 int rc = 0; 1574 1575#ifdef TCP_OFFLOAD 1576 switch (cmd) { 1577 case MOD_LOAD: 1578 rc = t4_tom_mod_load(); 1579 break; 1580 1581 case MOD_UNLOAD: 1582 rc = t4_tom_mod_unload(); 1583 break; 1584 1585 default: 1586 rc = EINVAL; 1587 } 1588#else 1589 printf("t4_tom: compiled without TCP_OFFLOAD support.\n"); 1590 rc = EOPNOTSUPP; 1591#endif 1592 return (rc); 1593} 1594 1595static moduledata_t t4_tom_moddata= { 1596 "t4_tom", 1597 t4_tom_modevent, 1598 0 1599}; 1600 1601MODULE_VERSION(t4_tom, 1); 1602MODULE_DEPEND(t4_tom, toecore, 1, 1, 1); 1603MODULE_DEPEND(t4_tom, t4nex, 1, 1, 1); 1604DECLARE_MODULE(t4_tom, t4_tom_moddata, SI_SUB_EXEC, SI_ORDER_ANY); 1605