1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause 3 * 4 * Copyright (c) 2021 Microsoft Corp. 5 * All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 11 * 1. Redistributions of source code must retain the above copyright 12 * notice, this list of conditions and the following disclaimer. 13 * 14 * 2. Redistributions in binary form must reproduce the above copyright 15 * notice, this list of conditions and the following disclaimer in the 16 * documentation and/or other materials provided with the distribution. 17 * 18 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS 19 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT 20 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR 21 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT 22 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, 23 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT 24 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, 25 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY 26 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT 27 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE 28 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. 29 */ 30 #include <sys/cdefs.h> 31 __FBSDID("$FreeBSD$"); 32 33 #include <sys/param.h> 34 #include <sys/systm.h> 35 #include <sys/bus.h> 36 #include <sys/kernel.h> 37 #include <sys/kthread.h> 38 #include <sys/malloc.h> 39 #include <sys/mbuf.h> 40 #include <sys/smp.h> 41 #include <sys/socket.h> 42 #include <sys/sockio.h> 43 #include <sys/time.h> 44 #include <sys/eventhandler.h> 45 46 #include <machine/bus.h> 47 #include <machine/resource.h> 48 #include <machine/in_cksum.h> 49 50 #include <net/if.h> 51 #include <net/if_var.h> 52 #include <net/if_types.h> 53 #include <net/if_vlan_var.h> 54 #ifdef RSS 55 #include <net/rss_config.h> 56 #endif 57 58 #include <netinet/in_systm.h> 59 #include <netinet/in.h> 60 #include <netinet/if_ether.h> 61 #include <netinet/ip.h> 62 #include <netinet/ip6.h> 63 #include <netinet/tcp.h> 64 #include <netinet/udp.h> 65 66 #include "mana.h" 67 #include "mana_sysctl.h" 68 69 static int mana_up(struct mana_port_context *apc); 70 static int mana_down(struct mana_port_context *apc); 71 72 static void 73 mana_rss_key_fill(void *k, size_t size) 74 { 75 static bool rss_key_generated = false; 76 static uint8_t rss_key[MANA_HASH_KEY_SIZE]; 77 78 KASSERT(size <= MANA_HASH_KEY_SIZE, 79 ("Request more buytes than MANA RSS key can hold")); 80 81 if (!rss_key_generated) { 82 arc4random_buf(rss_key, MANA_HASH_KEY_SIZE); 83 rss_key_generated = true; 84 } 85 memcpy(k, rss_key, size); 86 } 87 88 static int 89 mana_ifmedia_change(struct ifnet *ifp __unused) 90 { 91 return EOPNOTSUPP; 92 } 93 94 static void 95 mana_ifmedia_status(struct ifnet *ifp, struct ifmediareq *ifmr) 96 { 97 struct mana_port_context *apc = if_getsoftc(ifp); 98 99 if (!apc) { 100 if_printf(ifp, "Port not available\n"); 101 return; 102 } 103 104 MANA_APC_LOCK_LOCK(apc); 105 106 ifmr->ifm_status = IFM_AVALID; 107 ifmr->ifm_active = IFM_ETHER; 108 109 if (!apc->port_is_up) { 110 MANA_APC_LOCK_UNLOCK(apc); 111 mana_dbg(NULL, "Port %u link is down\n", apc->port_idx); 112 return; 113 } 114 115 ifmr->ifm_status |= IFM_ACTIVE; 116 ifmr->ifm_active |= IFM_100G_DR | IFM_FDX; 117 118 MANA_APC_LOCK_UNLOCK(apc); 119 } 120 121 static uint64_t 122 mana_get_counter(struct ifnet *ifp, ift_counter cnt) 123 { 124 struct mana_port_context *apc = if_getsoftc(ifp); 125 struct mana_port_stats *stats = &apc->port_stats; 126 127 switch (cnt) { 128 case IFCOUNTER_IPACKETS: 129 return (counter_u64_fetch(stats->rx_packets)); 130 case IFCOUNTER_OPACKETS: 131 return (counter_u64_fetch(stats->tx_packets)); 132 case IFCOUNTER_IBYTES: 133 return (counter_u64_fetch(stats->rx_bytes)); 134 case IFCOUNTER_OBYTES: 135 return (counter_u64_fetch(stats->tx_bytes)); 136 case IFCOUNTER_IQDROPS: 137 return (counter_u64_fetch(stats->rx_drops)); 138 case IFCOUNTER_OQDROPS: 139 return (counter_u64_fetch(stats->tx_drops)); 140 default: 141 return (if_get_counter_default(ifp, cnt)); 142 } 143 } 144 145 static void 146 mana_qflush(struct ifnet *ifp) 147 { 148 if_qflush(ifp); 149 } 150 151 int 152 mana_restart(struct mana_port_context *apc) 153 { 154 int rc = 0; 155 156 MANA_APC_LOCK_LOCK(apc); 157 if (apc->port_is_up) 158 mana_down(apc); 159 160 rc = mana_up(apc); 161 MANA_APC_LOCK_UNLOCK(apc); 162 163 return (rc); 164 } 165 166 static int 167 mana_ioctl(struct ifnet *ifp, u_long command, caddr_t data) 168 { 169 struct mana_port_context *apc = if_getsoftc(ifp); 170 struct ifrsskey *ifrk; 171 struct ifrsshash *ifrh; 172 struct ifreq *ifr; 173 uint16_t new_mtu; 174 int rc = 0; 175 176 switch (command) { 177 case SIOCSIFMTU: 178 ifr = (struct ifreq *)data; 179 new_mtu = ifr->ifr_mtu; 180 if (ifp->if_mtu == new_mtu) 181 break; 182 if ((new_mtu + 18 > MAX_FRAME_SIZE) || 183 (new_mtu + 18 < MIN_FRAME_SIZE)) { 184 if_printf(ifp, "Invalid MTU. new_mtu: %d, " 185 "max allowed: %d, min allowed: %d\n", 186 new_mtu, MAX_FRAME_SIZE - 18, MIN_FRAME_SIZE - 18); 187 return EINVAL; 188 } 189 MANA_APC_LOCK_LOCK(apc); 190 if (apc->port_is_up) 191 mana_down(apc); 192 193 apc->frame_size = new_mtu + 18; 194 if_setmtu(ifp, new_mtu); 195 mana_dbg(NULL, "Set MTU to %d\n", new_mtu); 196 197 rc = mana_up(apc); 198 MANA_APC_LOCK_UNLOCK(apc); 199 break; 200 201 case SIOCSIFFLAGS: 202 if (ifp->if_flags & IFF_UP) { 203 if ((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0) { 204 MANA_APC_LOCK_LOCK(apc); 205 if (!apc->port_is_up) 206 rc = mana_up(apc); 207 MANA_APC_LOCK_UNLOCK(apc); 208 } 209 } else { 210 if (ifp->if_drv_flags & IFF_DRV_RUNNING) { 211 MANA_APC_LOCK_LOCK(apc); 212 if (apc->port_is_up) 213 mana_down(apc); 214 MANA_APC_LOCK_UNLOCK(apc); 215 } 216 } 217 break; 218 219 case SIOCSIFMEDIA: 220 case SIOCGIFMEDIA: 221 case SIOCGIFXMEDIA: 222 ifr = (struct ifreq *)data; 223 rc = ifmedia_ioctl(ifp, ifr, &apc->media, command); 224 break; 225 226 case SIOCGIFRSSKEY: 227 ifrk = (struct ifrsskey *)data; 228 ifrk->ifrk_func = RSS_FUNC_TOEPLITZ; 229 ifrk->ifrk_keylen = MANA_HASH_KEY_SIZE; 230 memcpy(ifrk->ifrk_key, apc->hashkey, MANA_HASH_KEY_SIZE); 231 break; 232 233 case SIOCGIFRSSHASH: 234 ifrh = (struct ifrsshash *)data; 235 ifrh->ifrh_func = RSS_FUNC_TOEPLITZ; 236 ifrh->ifrh_types = 237 RSS_TYPE_TCP_IPV4 | 238 RSS_TYPE_UDP_IPV4 | 239 RSS_TYPE_TCP_IPV6 | 240 RSS_TYPE_UDP_IPV6; 241 break; 242 243 default: 244 rc = ether_ioctl(ifp, command, data); 245 break; 246 } 247 248 return (rc); 249 } 250 251 static inline void 252 mana_alloc_counters(counter_u64_t *begin, int size) 253 { 254 counter_u64_t *end = (counter_u64_t *)((char *)begin + size); 255 256 for (; begin < end; ++begin) 257 *begin = counter_u64_alloc(M_WAITOK); 258 } 259 260 static inline void 261 mana_free_counters(counter_u64_t *begin, int size) 262 { 263 counter_u64_t *end = (counter_u64_t *)((char *)begin + size); 264 265 for (; begin < end; ++begin) 266 counter_u64_free(*begin); 267 } 268 269 static bool 270 mana_can_tx(struct gdma_queue *wq) 271 { 272 return mana_gd_wq_avail_space(wq) >= MAX_TX_WQE_SIZE; 273 } 274 275 static inline int 276 mana_tx_map_mbuf(struct mana_port_context *apc, 277 struct mana_send_buf_info *tx_info, 278 struct mbuf **m_head, struct mana_tx_package *tp, 279 struct mana_stats *tx_stats) 280 { 281 struct gdma_dev *gd = apc->ac->gdma_dev; 282 bus_dma_segment_t segs[MAX_MBUF_FRAGS]; 283 struct mbuf *m = *m_head; 284 int err, nsegs, i; 285 286 err = bus_dmamap_load_mbuf_sg(apc->tx_buf_tag, tx_info->dma_map, 287 m, segs, &nsegs, BUS_DMA_NOWAIT); 288 if (err == EFBIG) { 289 struct mbuf *m_new; 290 291 counter_u64_add(tx_stats->collapse, 1); 292 m_new = m_collapse(m, M_NOWAIT, MAX_MBUF_FRAGS); 293 if (unlikely(m_new == NULL)) { 294 counter_u64_add(tx_stats->collapse_err, 1); 295 return ENOBUFS; 296 } else { 297 *m_head = m = m_new; 298 } 299 300 mana_warn(NULL, 301 "Too many segs in orig mbuf, m_collapse called\n"); 302 303 err = bus_dmamap_load_mbuf_sg(apc->tx_buf_tag, 304 tx_info->dma_map, m, segs, &nsegs, BUS_DMA_NOWAIT); 305 } 306 if (!err) { 307 for (i = 0; i < nsegs; i++) { 308 tp->wqe_req.sgl[i].address = segs[i].ds_addr; 309 tp->wqe_req.sgl[i].mem_key = gd->gpa_mkey; 310 tp->wqe_req.sgl[i].size = segs[i].ds_len; 311 } 312 tp->wqe_req.num_sge = nsegs; 313 314 tx_info->mbuf = *m_head; 315 316 bus_dmamap_sync(apc->tx_buf_tag, tx_info->dma_map, 317 BUS_DMASYNC_PREWRITE); 318 } 319 320 return err; 321 } 322 323 static inline void 324 mana_tx_unmap_mbuf(struct mana_port_context *apc, 325 struct mana_send_buf_info *tx_info) 326 { 327 bus_dmamap_sync(apc->tx_buf_tag, tx_info->dma_map, 328 BUS_DMASYNC_POSTWRITE); 329 bus_dmamap_unload(apc->tx_buf_tag, tx_info->dma_map); 330 if (tx_info->mbuf) { 331 m_freem(tx_info->mbuf); 332 tx_info->mbuf = NULL; 333 } 334 } 335 336 static inline int 337 mana_load_rx_mbuf(struct mana_port_context *apc, struct mana_rxq *rxq, 338 struct mana_recv_buf_oob *rx_oob, bool alloc_mbuf) 339 { 340 bus_dma_segment_t segs[1]; 341 struct mbuf *mbuf; 342 int nsegs, err; 343 uint32_t mlen; 344 345 if (alloc_mbuf) { 346 mbuf = m_getjcl(M_NOWAIT, MT_DATA, M_PKTHDR, rxq->datasize); 347 if (unlikely(mbuf == NULL)) { 348 mbuf = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); 349 if (unlikely(mbuf == NULL)) { 350 return ENOMEM; 351 } 352 mlen = MCLBYTES; 353 } else { 354 mlen = rxq->datasize; 355 } 356 357 mbuf->m_pkthdr.len = mbuf->m_len = mlen; 358 } else { 359 if (rx_oob->mbuf) { 360 mbuf = rx_oob->mbuf; 361 mlen = rx_oob->mbuf->m_pkthdr.len; 362 } else { 363 return ENOMEM; 364 } 365 } 366 367 err = bus_dmamap_load_mbuf_sg(apc->rx_buf_tag, rx_oob->dma_map, 368 mbuf, segs, &nsegs, BUS_DMA_NOWAIT); 369 370 if (unlikely((err != 0) || (nsegs != 1))) { 371 mana_warn(NULL, "Failed to map mbuf, error: %d, " 372 "nsegs: %d\n", err, nsegs); 373 counter_u64_add(rxq->stats.dma_mapping_err, 1); 374 goto error; 375 } 376 377 bus_dmamap_sync(apc->rx_buf_tag, rx_oob->dma_map, 378 BUS_DMASYNC_PREREAD); 379 380 rx_oob->mbuf = mbuf; 381 rx_oob->num_sge = 1; 382 rx_oob->sgl[0].address = segs[0].ds_addr; 383 rx_oob->sgl[0].size = mlen; 384 rx_oob->sgl[0].mem_key = apc->ac->gdma_dev->gpa_mkey; 385 386 return 0; 387 388 error: 389 m_freem(mbuf); 390 return EFAULT; 391 } 392 393 static inline void 394 mana_unload_rx_mbuf(struct mana_port_context *apc, struct mana_rxq *rxq, 395 struct mana_recv_buf_oob *rx_oob, bool free_mbuf) 396 { 397 bus_dmamap_sync(apc->rx_buf_tag, rx_oob->dma_map, 398 BUS_DMASYNC_POSTREAD); 399 bus_dmamap_unload(apc->rx_buf_tag, rx_oob->dma_map); 400 401 if (free_mbuf && rx_oob->mbuf) { 402 m_freem(rx_oob->mbuf); 403 rx_oob->mbuf = NULL; 404 } 405 } 406 407 408 /* Use couple mbuf PH_loc spaces for l3 and l4 protocal type */ 409 #define MANA_L3_PROTO(_mbuf) ((_mbuf)->m_pkthdr.PH_loc.sixteen[0]) 410 #define MANA_L4_PROTO(_mbuf) ((_mbuf)->m_pkthdr.PH_loc.sixteen[1]) 411 412 #define MANA_TXQ_FULL (IFF_DRV_RUNNING | IFF_DRV_OACTIVE) 413 414 static void 415 mana_xmit(struct mana_txq *txq) 416 { 417 enum mana_tx_pkt_format pkt_fmt = MANA_SHORT_PKT_FMT; 418 struct mana_send_buf_info *tx_info; 419 struct ifnet *ndev = txq->ndev; 420 struct mbuf *mbuf; 421 struct mana_port_context *apc = if_getsoftc(ndev); 422 struct mana_port_stats *port_stats = &apc->port_stats; 423 struct gdma_dev *gd = apc->ac->gdma_dev; 424 uint64_t packets, bytes; 425 uint16_t next_to_use; 426 struct mana_tx_package pkg = {}; 427 struct mana_stats *tx_stats; 428 struct gdma_queue *gdma_sq; 429 struct mana_cq *cq; 430 int err, len; 431 432 gdma_sq = txq->gdma_sq; 433 cq = &apc->tx_qp[txq->idx].tx_cq; 434 tx_stats = &txq->stats; 435 436 packets = 0; 437 bytes = 0; 438 next_to_use = txq->next_to_use; 439 440 while ((mbuf = drbr_peek(ndev, txq->txq_br)) != NULL) { 441 if (!apc->port_is_up || 442 (if_getdrvflags(ndev) & MANA_TXQ_FULL) != IFF_DRV_RUNNING) { 443 drbr_putback(ndev, txq->txq_br, mbuf); 444 break; 445 } 446 447 if (!mana_can_tx(gdma_sq)) { 448 /* SQ is full. Set the IFF_DRV_OACTIVE flag */ 449 if_setdrvflagbits(apc->ndev, IFF_DRV_OACTIVE, 0); 450 counter_u64_add(tx_stats->stop, 1); 451 uint64_t stops = counter_u64_fetch(tx_stats->stop); 452 uint64_t wakeups = counter_u64_fetch(tx_stats->wakeup); 453 #define MANA_TXQ_STOP_THRESHOLD 50 454 if (stops > MANA_TXQ_STOP_THRESHOLD && wakeups > 0 && 455 stops > wakeups && txq->alt_txq_idx == txq->idx) { 456 txq->alt_txq_idx = 457 (txq->idx + (stops / wakeups)) 458 % apc->num_queues; 459 counter_u64_add(tx_stats->alt_chg, 1); 460 } 461 462 drbr_putback(ndev, txq->txq_br, mbuf); 463 464 taskqueue_enqueue(cq->cleanup_tq, &cq->cleanup_task); 465 break; 466 } 467 468 tx_info = &txq->tx_buf_info[next_to_use]; 469 470 memset(&pkg, 0, sizeof(struct mana_tx_package)); 471 pkg.wqe_req.sgl = pkg.sgl_array; 472 473 err = mana_tx_map_mbuf(apc, tx_info, &mbuf, &pkg, tx_stats); 474 if (unlikely(err)) { 475 mana_dbg(NULL, 476 "Failed to map tx mbuf, err %d\n", err); 477 478 counter_u64_add(tx_stats->dma_mapping_err, 1); 479 480 /* The mbuf is still there. Free it */ 481 m_freem(mbuf); 482 /* Advance the drbr queue */ 483 drbr_advance(ndev, txq->txq_br); 484 continue; 485 } 486 487 pkg.tx_oob.s_oob.vcq_num = cq->gdma_id; 488 pkg.tx_oob.s_oob.vsq_frame = txq->vsq_frame; 489 490 if (txq->vp_offset > MANA_SHORT_VPORT_OFFSET_MAX) { 491 pkg.tx_oob.l_oob.long_vp_offset = txq->vp_offset; 492 pkt_fmt = MANA_LONG_PKT_FMT; 493 } else { 494 pkg.tx_oob.s_oob.short_vp_offset = txq->vp_offset; 495 } 496 497 pkg.tx_oob.s_oob.pkt_fmt = pkt_fmt; 498 499 if (pkt_fmt == MANA_SHORT_PKT_FMT) 500 pkg.wqe_req.inline_oob_size = sizeof(struct mana_tx_short_oob); 501 else 502 pkg.wqe_req.inline_oob_size = sizeof(struct mana_tx_oob); 503 504 pkg.wqe_req.inline_oob_data = &pkg.tx_oob; 505 pkg.wqe_req.flags = 0; 506 pkg.wqe_req.client_data_unit = 0; 507 508 if (mbuf->m_pkthdr.csum_flags & CSUM_TSO) { 509 if (MANA_L3_PROTO(mbuf) == ETHERTYPE_IP) 510 pkg.tx_oob.s_oob.is_outer_ipv4 = 1; 511 else 512 pkg.tx_oob.s_oob.is_outer_ipv6 = 1; 513 514 pkg.tx_oob.s_oob.comp_iphdr_csum = 1; 515 pkg.tx_oob.s_oob.comp_tcp_csum = 1; 516 pkg.tx_oob.s_oob.trans_off = mbuf->m_pkthdr.l3hlen; 517 518 pkg.wqe_req.client_data_unit = mbuf->m_pkthdr.tso_segsz; 519 pkg.wqe_req.flags = GDMA_WR_OOB_IN_SGL | GDMA_WR_PAD_BY_SGE0; 520 } else if (mbuf->m_pkthdr.csum_flags & 521 (CSUM_IP_UDP | CSUM_IP_TCP | CSUM_IP6_UDP | CSUM_IP6_TCP)) { 522 if (MANA_L3_PROTO(mbuf) == ETHERTYPE_IP) { 523 pkg.tx_oob.s_oob.is_outer_ipv4 = 1; 524 pkg.tx_oob.s_oob.comp_iphdr_csum = 1; 525 } else { 526 pkg.tx_oob.s_oob.is_outer_ipv6 = 1; 527 } 528 529 if (MANA_L4_PROTO(mbuf) == IPPROTO_TCP) { 530 pkg.tx_oob.s_oob.comp_tcp_csum = 1; 531 pkg.tx_oob.s_oob.trans_off = 532 mbuf->m_pkthdr.l3hlen; 533 } else { 534 pkg.tx_oob.s_oob.comp_udp_csum = 1; 535 } 536 } else if (mbuf->m_pkthdr.csum_flags & CSUM_IP) { 537 pkg.tx_oob.s_oob.is_outer_ipv4 = 1; 538 pkg.tx_oob.s_oob.comp_iphdr_csum = 1; 539 } else { 540 if (MANA_L3_PROTO(mbuf) == ETHERTYPE_IP) 541 pkg.tx_oob.s_oob.is_outer_ipv4 = 1; 542 else if (MANA_L3_PROTO(mbuf) == ETHERTYPE_IPV6) 543 pkg.tx_oob.s_oob.is_outer_ipv6 = 1; 544 } 545 546 len = mbuf->m_pkthdr.len; 547 548 err = mana_gd_post_work_request(gdma_sq, &pkg.wqe_req, 549 (struct gdma_posted_wqe_info *)&tx_info->wqe_inf); 550 if (unlikely(err)) { 551 /* Should not happen */ 552 if_printf(ndev, "Failed to post TX OOB: %d\n", err); 553 554 mana_tx_unmap_mbuf(apc, tx_info); 555 556 drbr_advance(ndev, txq->txq_br); 557 continue; 558 } 559 560 next_to_use = 561 (next_to_use + 1) % MAX_SEND_BUFFERS_PER_QUEUE; 562 563 (void)atomic_inc_return(&txq->pending_sends); 564 565 drbr_advance(ndev, txq->txq_br); 566 567 mana_gd_wq_ring_doorbell(gd->gdma_context, gdma_sq); 568 569 packets++; 570 bytes += len; 571 } 572 573 counter_enter(); 574 counter_u64_add_protected(tx_stats->packets, packets); 575 counter_u64_add_protected(port_stats->tx_packets, packets); 576 counter_u64_add_protected(tx_stats->bytes, bytes); 577 counter_u64_add_protected(port_stats->tx_bytes, bytes); 578 counter_exit(); 579 580 txq->next_to_use = next_to_use; 581 } 582 583 static void 584 mana_xmit_taskfunc(void *arg, int pending) 585 { 586 struct mana_txq *txq = (struct mana_txq *)arg; 587 struct ifnet *ndev = txq->ndev; 588 struct mana_port_context *apc = if_getsoftc(ndev); 589 590 while (!drbr_empty(ndev, txq->txq_br) && apc->port_is_up && 591 (if_getdrvflags(ndev) & MANA_TXQ_FULL) == IFF_DRV_RUNNING) { 592 mtx_lock(&txq->txq_mtx); 593 mana_xmit(txq); 594 mtx_unlock(&txq->txq_mtx); 595 } 596 } 597 598 #define PULLUP_HDR(m, len) \ 599 do { \ 600 if (unlikely((m)->m_len < (len))) { \ 601 (m) = m_pullup((m), (len)); \ 602 if ((m) == NULL) \ 603 return (NULL); \ 604 } \ 605 } while (0) 606 607 /* 608 * If this function failed, the mbuf would be freed. 609 */ 610 static inline struct mbuf * 611 mana_tso_fixup(struct mbuf *mbuf) 612 { 613 struct ether_vlan_header *eh = mtod(mbuf, struct ether_vlan_header *); 614 struct tcphdr *th; 615 uint16_t etype; 616 int ehlen; 617 618 if (eh->evl_encap_proto == ntohs(ETHERTYPE_VLAN)) { 619 etype = ntohs(eh->evl_proto); 620 ehlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN; 621 } else { 622 etype = ntohs(eh->evl_encap_proto); 623 ehlen = ETHER_HDR_LEN; 624 } 625 626 if (etype == ETHERTYPE_IP) { 627 struct ip *ip; 628 int iphlen; 629 630 PULLUP_HDR(mbuf, ehlen + sizeof(*ip)); 631 ip = mtodo(mbuf, ehlen); 632 iphlen = ip->ip_hl << 2; 633 mbuf->m_pkthdr.l3hlen = ehlen + iphlen; 634 635 PULLUP_HDR(mbuf, ehlen + iphlen + sizeof(*th)); 636 th = mtodo(mbuf, ehlen + iphlen); 637 638 ip->ip_len = 0; 639 ip->ip_sum = 0; 640 th->th_sum = in_pseudo(ip->ip_src.s_addr, 641 ip->ip_dst.s_addr, htons(IPPROTO_TCP)); 642 } else if (etype == ETHERTYPE_IPV6) { 643 struct ip6_hdr *ip6; 644 645 PULLUP_HDR(mbuf, ehlen + sizeof(*ip6) + sizeof(*th)); 646 ip6 = mtodo(mbuf, ehlen); 647 if (ip6->ip6_nxt != IPPROTO_TCP) { 648 /* Realy something wrong, just return */ 649 mana_dbg(NULL, "TSO mbuf not TCP, freed.\n"); 650 m_freem(mbuf); 651 return NULL; 652 } 653 mbuf->m_pkthdr.l3hlen = ehlen + sizeof(*ip6); 654 655 th = mtodo(mbuf, ehlen + sizeof(*ip6)); 656 657 ip6->ip6_plen = 0; 658 th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0); 659 } else { 660 /* CSUM_TSO is set but not IP protocol. */ 661 mana_warn(NULL, "TSO mbuf not right, freed.\n"); 662 m_freem(mbuf); 663 return NULL; 664 } 665 666 MANA_L3_PROTO(mbuf) = etype; 667 668 return (mbuf); 669 } 670 671 /* 672 * If this function failed, the mbuf would be freed. 673 */ 674 static inline struct mbuf * 675 mana_mbuf_csum_check(struct mbuf *mbuf) 676 { 677 struct ether_vlan_header *eh = mtod(mbuf, struct ether_vlan_header *); 678 struct mbuf *mbuf_next; 679 uint16_t etype; 680 int offset; 681 int ehlen; 682 683 if (eh->evl_encap_proto == ntohs(ETHERTYPE_VLAN)) { 684 etype = ntohs(eh->evl_proto); 685 ehlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN; 686 } else { 687 etype = ntohs(eh->evl_encap_proto); 688 ehlen = ETHER_HDR_LEN; 689 } 690 691 mbuf_next = m_getptr(mbuf, ehlen, &offset); 692 693 MANA_L4_PROTO(mbuf) = 0; 694 if (etype == ETHERTYPE_IP) { 695 const struct ip *ip; 696 int iphlen; 697 698 ip = (struct ip *)(mtodo(mbuf_next, offset)); 699 iphlen = ip->ip_hl << 2; 700 mbuf->m_pkthdr.l3hlen = ehlen + iphlen; 701 702 MANA_L4_PROTO(mbuf) = ip->ip_p; 703 } else if (etype == ETHERTYPE_IPV6) { 704 const struct ip6_hdr *ip6; 705 706 ip6 = (struct ip6_hdr *)(mtodo(mbuf_next, offset)); 707 mbuf->m_pkthdr.l3hlen = ehlen + sizeof(*ip6); 708 709 MANA_L4_PROTO(mbuf) = ip6->ip6_nxt; 710 } else { 711 MANA_L4_PROTO(mbuf) = 0; 712 } 713 714 MANA_L3_PROTO(mbuf) = etype; 715 716 return (mbuf); 717 } 718 719 static int 720 mana_start_xmit(struct ifnet *ifp, struct mbuf *m) 721 { 722 struct mana_port_context *apc = if_getsoftc(ifp); 723 struct mana_txq *txq; 724 int is_drbr_empty; 725 uint16_t txq_id; 726 int err; 727 728 if (unlikely((!apc->port_is_up) || 729 (if_getdrvflags(ifp) & IFF_DRV_RUNNING) == 0)) 730 return ENODEV; 731 732 if (m->m_pkthdr.csum_flags & CSUM_TSO) { 733 m = mana_tso_fixup(m); 734 if (unlikely(m == NULL)) { 735 counter_enter(); 736 counter_u64_add_protected(apc->port_stats.tx_drops, 1); 737 counter_exit(); 738 return EIO; 739 } 740 } else { 741 m = mana_mbuf_csum_check(m); 742 if (unlikely(m == NULL)) { 743 counter_enter(); 744 counter_u64_add_protected(apc->port_stats.tx_drops, 1); 745 counter_exit(); 746 return EIO; 747 } 748 } 749 750 if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) { 751 uint32_t hash = m->m_pkthdr.flowid; 752 txq_id = apc->indir_table[(hash) & MANA_INDIRECT_TABLE_MASK] % 753 apc->num_queues; 754 } else { 755 txq_id = m->m_pkthdr.flowid % apc->num_queues; 756 } 757 758 if (apc->enable_tx_altq) 759 txq_id = apc->tx_qp[txq_id].txq.alt_txq_idx; 760 761 txq = &apc->tx_qp[txq_id].txq; 762 763 is_drbr_empty = drbr_empty(ifp, txq->txq_br); 764 err = drbr_enqueue(ifp, txq->txq_br, m); 765 if (unlikely(err)) { 766 mana_warn(NULL, "txq %u failed to enqueue: %d\n", 767 txq_id, err); 768 taskqueue_enqueue(txq->enqueue_tq, &txq->enqueue_task); 769 return err; 770 } 771 772 if (is_drbr_empty && mtx_trylock(&txq->txq_mtx)) { 773 mana_xmit(txq); 774 mtx_unlock(&txq->txq_mtx); 775 } else { 776 taskqueue_enqueue(txq->enqueue_tq, &txq->enqueue_task); 777 } 778 779 return 0; 780 } 781 782 static void 783 mana_cleanup_port_context(struct mana_port_context *apc) 784 { 785 bus_dma_tag_destroy(apc->tx_buf_tag); 786 bus_dma_tag_destroy(apc->rx_buf_tag); 787 apc->rx_buf_tag = NULL; 788 789 free(apc->rxqs, M_DEVBUF); 790 apc->rxqs = NULL; 791 792 mana_free_counters((counter_u64_t *)&apc->port_stats, 793 sizeof(struct mana_port_stats)); 794 } 795 796 static int 797 mana_init_port_context(struct mana_port_context *apc) 798 { 799 device_t dev = apc->ac->gdma_dev->gdma_context->dev; 800 uint32_t tso_maxsize; 801 int err; 802 803 tso_maxsize = MAX_MBUF_FRAGS * MANA_TSO_MAXSEG_SZ - 804 (ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN); 805 806 /* Create DMA tag for tx bufs */ 807 err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */ 808 1, 0, /* alignment, boundary */ 809 BUS_SPACE_MAXADDR, /* lowaddr */ 810 BUS_SPACE_MAXADDR, /* highaddr */ 811 NULL, NULL, /* filter, filterarg */ 812 tso_maxsize, /* maxsize */ 813 MAX_MBUF_FRAGS, /* nsegments */ 814 tso_maxsize, /* maxsegsize */ 815 0, /* flags */ 816 NULL, NULL, /* lockfunc, lockfuncarg*/ 817 &apc->tx_buf_tag); 818 if (unlikely(err)) { 819 device_printf(dev, "Feiled to create TX DMA tag\n"); 820 return err; 821 } 822 823 /* Create DMA tag for rx bufs */ 824 err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */ 825 64, 0, /* alignment, boundary */ 826 BUS_SPACE_MAXADDR, /* lowaddr */ 827 BUS_SPACE_MAXADDR, /* highaddr */ 828 NULL, NULL, /* filter, filterarg */ 829 MJUMPAGESIZE, /* maxsize */ 830 1, /* nsegments */ 831 MJUMPAGESIZE, /* maxsegsize */ 832 0, /* flags */ 833 NULL, NULL, /* lockfunc, lockfuncarg*/ 834 &apc->rx_buf_tag); 835 if (unlikely(err)) { 836 device_printf(dev, "Feiled to create RX DMA tag\n"); 837 return err; 838 } 839 840 apc->rxqs = mallocarray(apc->num_queues, sizeof(struct mana_rxq *), 841 M_DEVBUF, M_WAITOK | M_ZERO); 842 843 if (!apc->rxqs) { 844 bus_dma_tag_destroy(apc->tx_buf_tag); 845 bus_dma_tag_destroy(apc->rx_buf_tag); 846 apc->rx_buf_tag = NULL; 847 return ENOMEM; 848 } 849 850 return 0; 851 } 852 853 static int 854 mana_send_request(struct mana_context *ac, void *in_buf, 855 uint32_t in_len, void *out_buf, uint32_t out_len) 856 { 857 struct gdma_context *gc = ac->gdma_dev->gdma_context; 858 struct gdma_resp_hdr *resp = out_buf; 859 struct gdma_req_hdr *req = in_buf; 860 device_t dev = gc->dev; 861 static atomic_t activity_id; 862 int err; 863 864 req->dev_id = gc->mana.dev_id; 865 req->activity_id = atomic_inc_return(&activity_id); 866 867 mana_dbg(NULL, "activity_id = %u\n", activity_id); 868 869 err = mana_gd_send_request(gc, in_len, in_buf, out_len, 870 out_buf); 871 if (err || resp->status) { 872 device_printf(dev, "Failed to send mana message: %d, 0x%x\n", 873 err, resp->status); 874 return err ? err : EPROTO; 875 } 876 877 if (req->dev_id.as_uint32 != resp->dev_id.as_uint32 || 878 req->activity_id != resp->activity_id) { 879 device_printf(dev, 880 "Unexpected mana message response: %x,%x,%x,%x\n", 881 req->dev_id.as_uint32, resp->dev_id.as_uint32, 882 req->activity_id, resp->activity_id); 883 return EPROTO; 884 } 885 886 return 0; 887 } 888 889 static int 890 mana_verify_resp_hdr(const struct gdma_resp_hdr *resp_hdr, 891 const enum mana_command_code expected_code, 892 const uint32_t min_size) 893 { 894 if (resp_hdr->response.msg_type != expected_code) 895 return EPROTO; 896 897 if (resp_hdr->response.msg_version < GDMA_MESSAGE_V1) 898 return EPROTO; 899 900 if (resp_hdr->response.msg_size < min_size) 901 return EPROTO; 902 903 return 0; 904 } 905 906 static int 907 mana_query_device_cfg(struct mana_context *ac, uint32_t proto_major_ver, 908 uint32_t proto_minor_ver, uint32_t proto_micro_ver, 909 uint16_t *max_num_vports) 910 { 911 struct gdma_context *gc = ac->gdma_dev->gdma_context; 912 struct mana_query_device_cfg_resp resp = {}; 913 struct mana_query_device_cfg_req req = {}; 914 device_t dev = gc->dev; 915 int err = 0; 916 917 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_DEV_CONFIG, 918 sizeof(req), sizeof(resp)); 919 req.proto_major_ver = proto_major_ver; 920 req.proto_minor_ver = proto_minor_ver; 921 req.proto_micro_ver = proto_micro_ver; 922 923 err = mana_send_request(ac, &req, sizeof(req), &resp, sizeof(resp)); 924 if (err) { 925 device_printf(dev, "Failed to query config: %d", err); 926 return err; 927 } 928 929 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_DEV_CONFIG, 930 sizeof(resp)); 931 if (err || resp.hdr.status) { 932 device_printf(dev, "Invalid query result: %d, 0x%x\n", err, 933 resp.hdr.status); 934 if (!err) 935 err = EPROTO; 936 return err; 937 } 938 939 *max_num_vports = resp.max_num_vports; 940 941 mana_dbg(NULL, "mana max_num_vports from device = %d\n", 942 *max_num_vports); 943 944 return 0; 945 } 946 947 static int 948 mana_query_vport_cfg(struct mana_port_context *apc, uint32_t vport_index, 949 uint32_t *max_sq, uint32_t *max_rq, uint32_t *num_indir_entry) 950 { 951 struct mana_query_vport_cfg_resp resp = {}; 952 struct mana_query_vport_cfg_req req = {}; 953 int err; 954 955 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_VPORT_CONFIG, 956 sizeof(req), sizeof(resp)); 957 958 req.vport_index = vport_index; 959 960 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 961 sizeof(resp)); 962 if (err) 963 return err; 964 965 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_VPORT_CONFIG, 966 sizeof(resp)); 967 if (err) 968 return err; 969 970 if (resp.hdr.status) 971 return EPROTO; 972 973 *max_sq = resp.max_num_sq; 974 *max_rq = resp.max_num_rq; 975 *num_indir_entry = resp.num_indirection_ent; 976 977 apc->port_handle = resp.vport; 978 memcpy(apc->mac_addr, resp.mac_addr, ETHER_ADDR_LEN); 979 980 return 0; 981 } 982 983 static int 984 mana_cfg_vport(struct mana_port_context *apc, uint32_t protection_dom_id, 985 uint32_t doorbell_pg_id) 986 { 987 struct mana_config_vport_resp resp = {}; 988 struct mana_config_vport_req req = {}; 989 int err; 990 991 mana_gd_init_req_hdr(&req.hdr, MANA_CONFIG_VPORT_TX, 992 sizeof(req), sizeof(resp)); 993 req.vport = apc->port_handle; 994 req.pdid = protection_dom_id; 995 req.doorbell_pageid = doorbell_pg_id; 996 997 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 998 sizeof(resp)); 999 if (err) { 1000 if_printf(apc->ndev, "Failed to configure vPort: %d\n", err); 1001 goto out; 1002 } 1003 1004 err = mana_verify_resp_hdr(&resp.hdr, MANA_CONFIG_VPORT_TX, 1005 sizeof(resp)); 1006 if (err || resp.hdr.status) { 1007 if_printf(apc->ndev, "Failed to configure vPort: %d, 0x%x\n", 1008 err, resp.hdr.status); 1009 if (!err) 1010 err = EPROTO; 1011 1012 goto out; 1013 } 1014 1015 apc->tx_shortform_allowed = resp.short_form_allowed; 1016 apc->tx_vp_offset = resp.tx_vport_offset; 1017 out: 1018 return err; 1019 } 1020 1021 static int 1022 mana_cfg_vport_steering(struct mana_port_context *apc, 1023 enum TRI_STATE rx, 1024 bool update_default_rxobj, bool update_key, 1025 bool update_tab) 1026 { 1027 uint16_t num_entries = MANA_INDIRECT_TABLE_SIZE; 1028 struct mana_cfg_rx_steer_req *req = NULL; 1029 struct mana_cfg_rx_steer_resp resp = {}; 1030 struct ifnet *ndev = apc->ndev; 1031 mana_handle_t *req_indir_tab; 1032 uint32_t req_buf_size; 1033 int err; 1034 1035 req_buf_size = sizeof(*req) + sizeof(mana_handle_t) * num_entries; 1036 req = malloc(req_buf_size, M_DEVBUF, M_WAITOK | M_ZERO); 1037 if (!req) 1038 return ENOMEM; 1039 1040 mana_gd_init_req_hdr(&req->hdr, MANA_CONFIG_VPORT_RX, req_buf_size, 1041 sizeof(resp)); 1042 1043 req->vport = apc->port_handle; 1044 req->num_indir_entries = num_entries; 1045 req->indir_tab_offset = sizeof(*req); 1046 req->rx_enable = rx; 1047 req->rss_enable = apc->rss_state; 1048 req->update_default_rxobj = update_default_rxobj; 1049 req->update_hashkey = update_key; 1050 req->update_indir_tab = update_tab; 1051 req->default_rxobj = apc->default_rxobj; 1052 1053 if (update_key) 1054 memcpy(&req->hashkey, apc->hashkey, MANA_HASH_KEY_SIZE); 1055 1056 if (update_tab) { 1057 req_indir_tab = (mana_handle_t *)(req + 1); 1058 memcpy(req_indir_tab, apc->rxobj_table, 1059 req->num_indir_entries * sizeof(mana_handle_t)); 1060 } 1061 1062 err = mana_send_request(apc->ac, req, req_buf_size, &resp, 1063 sizeof(resp)); 1064 if (err) { 1065 if_printf(ndev, "Failed to configure vPort RX: %d\n", err); 1066 goto out; 1067 } 1068 1069 err = mana_verify_resp_hdr(&resp.hdr, MANA_CONFIG_VPORT_RX, 1070 sizeof(resp)); 1071 if (err) { 1072 if_printf(ndev, "vPort RX configuration failed: %d\n", err); 1073 goto out; 1074 } 1075 1076 if (resp.hdr.status) { 1077 if_printf(ndev, "vPort RX configuration failed: 0x%x\n", 1078 resp.hdr.status); 1079 err = EPROTO; 1080 } 1081 out: 1082 free(req, M_DEVBUF); 1083 return err; 1084 } 1085 1086 static int 1087 mana_create_wq_obj(struct mana_port_context *apc, 1088 mana_handle_t vport, 1089 uint32_t wq_type, struct mana_obj_spec *wq_spec, 1090 struct mana_obj_spec *cq_spec, 1091 mana_handle_t *wq_obj) 1092 { 1093 struct mana_create_wqobj_resp resp = {}; 1094 struct mana_create_wqobj_req req = {}; 1095 struct ifnet *ndev = apc->ndev; 1096 int err; 1097 1098 mana_gd_init_req_hdr(&req.hdr, MANA_CREATE_WQ_OBJ, 1099 sizeof(req), sizeof(resp)); 1100 req.vport = vport; 1101 req.wq_type = wq_type; 1102 req.wq_gdma_region = wq_spec->gdma_region; 1103 req.cq_gdma_region = cq_spec->gdma_region; 1104 req.wq_size = wq_spec->queue_size; 1105 req.cq_size = cq_spec->queue_size; 1106 req.cq_moderation_ctx_id = cq_spec->modr_ctx_id; 1107 req.cq_parent_qid = cq_spec->attached_eq; 1108 1109 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 1110 sizeof(resp)); 1111 if (err) { 1112 if_printf(ndev, "Failed to create WQ object: %d\n", err); 1113 goto out; 1114 } 1115 1116 err = mana_verify_resp_hdr(&resp.hdr, MANA_CREATE_WQ_OBJ, 1117 sizeof(resp)); 1118 if (err || resp.hdr.status) { 1119 if_printf(ndev, "Failed to create WQ object: %d, 0x%x\n", err, 1120 resp.hdr.status); 1121 if (!err) 1122 err = EPROTO; 1123 goto out; 1124 } 1125 1126 if (resp.wq_obj == INVALID_MANA_HANDLE) { 1127 if_printf(ndev, "Got an invalid WQ object handle\n"); 1128 err = EPROTO; 1129 goto out; 1130 } 1131 1132 *wq_obj = resp.wq_obj; 1133 wq_spec->queue_index = resp.wq_id; 1134 cq_spec->queue_index = resp.cq_id; 1135 1136 return 0; 1137 out: 1138 return err; 1139 } 1140 1141 static void 1142 mana_destroy_wq_obj(struct mana_port_context *apc, uint32_t wq_type, 1143 mana_handle_t wq_obj) 1144 { 1145 struct mana_destroy_wqobj_resp resp = {}; 1146 struct mana_destroy_wqobj_req req = {}; 1147 struct ifnet *ndev = apc->ndev; 1148 int err; 1149 1150 mana_gd_init_req_hdr(&req.hdr, MANA_DESTROY_WQ_OBJ, 1151 sizeof(req), sizeof(resp)); 1152 req.wq_type = wq_type; 1153 req.wq_obj_handle = wq_obj; 1154 1155 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 1156 sizeof(resp)); 1157 if (err) { 1158 if_printf(ndev, "Failed to destroy WQ object: %d\n", err); 1159 return; 1160 } 1161 1162 err = mana_verify_resp_hdr(&resp.hdr, MANA_DESTROY_WQ_OBJ, 1163 sizeof(resp)); 1164 if (err || resp.hdr.status) 1165 if_printf(ndev, "Failed to destroy WQ object: %d, 0x%x\n", 1166 err, resp.hdr.status); 1167 } 1168 1169 static void 1170 mana_destroy_eq(struct mana_context *ac) 1171 { 1172 struct gdma_context *gc = ac->gdma_dev->gdma_context; 1173 struct gdma_queue *eq; 1174 int i; 1175 1176 if (!ac->eqs) 1177 return; 1178 1179 for (i = 0; i < gc->max_num_queues; i++) { 1180 eq = ac->eqs[i].eq; 1181 if (!eq) 1182 continue; 1183 1184 mana_gd_destroy_queue(gc, eq); 1185 } 1186 1187 free(ac->eqs, M_DEVBUF); 1188 ac->eqs = NULL; 1189 } 1190 1191 static int 1192 mana_create_eq(struct mana_context *ac) 1193 { 1194 struct gdma_dev *gd = ac->gdma_dev; 1195 struct gdma_context *gc = gd->gdma_context; 1196 struct gdma_queue_spec spec = {}; 1197 int err; 1198 int i; 1199 1200 ac->eqs = mallocarray(gc->max_num_queues, sizeof(struct mana_eq), 1201 M_DEVBUF, M_WAITOK | M_ZERO); 1202 if (!ac->eqs) 1203 return ENOMEM; 1204 1205 spec.type = GDMA_EQ; 1206 spec.monitor_avl_buf = false; 1207 spec.queue_size = EQ_SIZE; 1208 spec.eq.callback = NULL; 1209 spec.eq.context = ac->eqs; 1210 spec.eq.log2_throttle_limit = LOG2_EQ_THROTTLE; 1211 1212 for (i = 0; i < gc->max_num_queues; i++) { 1213 err = mana_gd_create_mana_eq(gd, &spec, &ac->eqs[i].eq); 1214 if (err) 1215 goto out; 1216 } 1217 1218 return 0; 1219 out: 1220 mana_destroy_eq(ac); 1221 return err; 1222 } 1223 1224 static int 1225 mana_fence_rq(struct mana_port_context *apc, struct mana_rxq *rxq) 1226 { 1227 struct mana_fence_rq_resp resp = {}; 1228 struct mana_fence_rq_req req = {}; 1229 int err; 1230 1231 init_completion(&rxq->fence_event); 1232 1233 mana_gd_init_req_hdr(&req.hdr, MANA_FENCE_RQ, 1234 sizeof(req), sizeof(resp)); 1235 req.wq_obj_handle = rxq->rxobj; 1236 1237 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 1238 sizeof(resp)); 1239 if (err) { 1240 if_printf(apc->ndev, "Failed to fence RQ %u: %d\n", 1241 rxq->rxq_idx, err); 1242 return err; 1243 } 1244 1245 err = mana_verify_resp_hdr(&resp.hdr, MANA_FENCE_RQ, sizeof(resp)); 1246 if (err || resp.hdr.status) { 1247 if_printf(apc->ndev, "Failed to fence RQ %u: %d, 0x%x\n", 1248 rxq->rxq_idx, err, resp.hdr.status); 1249 if (!err) 1250 err = EPROTO; 1251 1252 return err; 1253 } 1254 1255 if (wait_for_completion_timeout(&rxq->fence_event, 10 * hz)) { 1256 if_printf(apc->ndev, "Failed to fence RQ %u: timed out\n", 1257 rxq->rxq_idx); 1258 return ETIMEDOUT; 1259 } 1260 1261 return 0; 1262 } 1263 1264 static void 1265 mana_fence_rqs(struct mana_port_context *apc) 1266 { 1267 unsigned int rxq_idx; 1268 struct mana_rxq *rxq; 1269 int err; 1270 1271 for (rxq_idx = 0; rxq_idx < apc->num_queues; rxq_idx++) { 1272 rxq = apc->rxqs[rxq_idx]; 1273 err = mana_fence_rq(apc, rxq); 1274 1275 /* In case of any error, use sleep instead. */ 1276 if (err) 1277 gdma_msleep(100); 1278 } 1279 } 1280 1281 static int 1282 mana_move_wq_tail(struct gdma_queue *wq, uint32_t num_units) 1283 { 1284 uint32_t used_space_old; 1285 uint32_t used_space_new; 1286 1287 used_space_old = wq->head - wq->tail; 1288 used_space_new = wq->head - (wq->tail + num_units); 1289 1290 if (used_space_new > used_space_old) { 1291 mana_err(NULL, 1292 "WARNING: new used space %u greater than old one %u\n", 1293 used_space_new, used_space_old); 1294 return ERANGE; 1295 } 1296 1297 wq->tail += num_units; 1298 return 0; 1299 } 1300 1301 static void 1302 mana_poll_tx_cq(struct mana_cq *cq) 1303 { 1304 struct gdma_comp *completions = cq->gdma_comp_buf; 1305 struct gdma_posted_wqe_info *wqe_info; 1306 struct mana_send_buf_info *tx_info; 1307 unsigned int pkt_transmitted = 0; 1308 unsigned int wqe_unit_cnt = 0; 1309 struct mana_txq *txq = cq->txq; 1310 struct mana_port_context *apc; 1311 uint16_t next_to_complete; 1312 struct ifnet *ndev; 1313 int comp_read; 1314 int txq_idx = txq->idx;; 1315 int i; 1316 int sa_drop = 0; 1317 1318 struct gdma_queue *gdma_wq; 1319 unsigned int avail_space; 1320 bool txq_full = false; 1321 1322 ndev = txq->ndev; 1323 apc = if_getsoftc(ndev); 1324 1325 comp_read = mana_gd_poll_cq(cq->gdma_cq, completions, 1326 CQE_POLLING_BUFFER); 1327 1328 if (comp_read < 1) 1329 return; 1330 1331 next_to_complete = txq->next_to_complete; 1332 1333 for (i = 0; i < comp_read; i++) { 1334 struct mana_tx_comp_oob *cqe_oob; 1335 1336 if (!completions[i].is_sq) { 1337 mana_err(NULL, "WARNING: Not for SQ\n"); 1338 return; 1339 } 1340 1341 cqe_oob = (struct mana_tx_comp_oob *)completions[i].cqe_data; 1342 if (cqe_oob->cqe_hdr.client_type != 1343 MANA_CQE_COMPLETION) { 1344 mana_err(NULL, 1345 "WARNING: Invalid CQE client type %u\n", 1346 cqe_oob->cqe_hdr.client_type); 1347 return; 1348 } 1349 1350 switch (cqe_oob->cqe_hdr.cqe_type) { 1351 case CQE_TX_OKAY: 1352 break; 1353 1354 case CQE_TX_SA_DROP: 1355 case CQE_TX_MTU_DROP: 1356 case CQE_TX_INVALID_OOB: 1357 case CQE_TX_INVALID_ETH_TYPE: 1358 case CQE_TX_HDR_PROCESSING_ERROR: 1359 case CQE_TX_VF_DISABLED: 1360 case CQE_TX_VPORT_IDX_OUT_OF_RANGE: 1361 case CQE_TX_VPORT_DISABLED: 1362 case CQE_TX_VLAN_TAGGING_VIOLATION: 1363 sa_drop ++; 1364 mana_err(NULL, 1365 "TX: txq %d CQE error %d, ntc = %d, " 1366 "pending sends = %d: err ignored.\n", 1367 txq_idx, cqe_oob->cqe_hdr.cqe_type, 1368 next_to_complete, txq->pending_sends); 1369 break; 1370 1371 default: 1372 /* If the CQE type is unexpected, log an error, 1373 * and go through the error path. 1374 */ 1375 mana_err(NULL, 1376 "ERROR: TX: Unexpected CQE type %d: HW BUG?\n", 1377 cqe_oob->cqe_hdr.cqe_type); 1378 return; 1379 } 1380 if (txq->gdma_txq_id != completions[i].wq_num) { 1381 mana_dbg(NULL, 1382 "txq gdma id not match completion wq num: " 1383 "%d != %d\n", 1384 txq->gdma_txq_id, completions[i].wq_num); 1385 break; 1386 } 1387 1388 tx_info = &txq->tx_buf_info[next_to_complete]; 1389 if (!tx_info->mbuf) { 1390 mana_err(NULL, 1391 "WARNING: txq %d Empty mbuf on tx_info: %u, " 1392 "ntu = %u, pending_sends = %d, " 1393 "transmitted = %d, sa_drop = %d, i = %d, comp_read = %d\n", 1394 txq_idx, next_to_complete, txq->next_to_use, 1395 txq->pending_sends, pkt_transmitted, sa_drop, 1396 i, comp_read); 1397 break; 1398 } 1399 1400 wqe_info = &tx_info->wqe_inf; 1401 wqe_unit_cnt += wqe_info->wqe_size_in_bu; 1402 1403 mana_tx_unmap_mbuf(apc, tx_info); 1404 mb(); 1405 1406 next_to_complete = 1407 (next_to_complete + 1) % MAX_SEND_BUFFERS_PER_QUEUE; 1408 1409 pkt_transmitted++; 1410 } 1411 1412 txq->next_to_complete = next_to_complete; 1413 1414 if (wqe_unit_cnt == 0) { 1415 mana_err(NULL, 1416 "WARNING: TX ring not proceeding!\n"); 1417 return; 1418 } 1419 1420 mana_move_wq_tail(txq->gdma_sq, wqe_unit_cnt); 1421 1422 /* Ensure tail updated before checking q stop */ 1423 wmb(); 1424 1425 gdma_wq = txq->gdma_sq; 1426 avail_space = mana_gd_wq_avail_space(gdma_wq); 1427 1428 1429 if ((if_getdrvflags(ndev) & MANA_TXQ_FULL) == MANA_TXQ_FULL) { 1430 txq_full = true; 1431 } 1432 1433 /* Ensure checking txq_full before apc->port_is_up. */ 1434 rmb(); 1435 1436 if (txq_full && apc->port_is_up && avail_space >= MAX_TX_WQE_SIZE) { 1437 /* Grab the txq lock and re-test */ 1438 mtx_lock(&txq->txq_mtx); 1439 avail_space = mana_gd_wq_avail_space(gdma_wq); 1440 1441 if ((if_getdrvflags(ndev) & MANA_TXQ_FULL) == MANA_TXQ_FULL && 1442 apc->port_is_up && avail_space >= MAX_TX_WQE_SIZE) { 1443 /* Clear the Q full flag */ 1444 if_setdrvflagbits(apc->ndev, IFF_DRV_RUNNING, 1445 IFF_DRV_OACTIVE); 1446 counter_u64_add(txq->stats.wakeup, 1); 1447 if (txq->alt_txq_idx != txq->idx) { 1448 uint64_t stops = counter_u64_fetch(txq->stats.stop); 1449 uint64_t wakeups = counter_u64_fetch(txq->stats.wakeup); 1450 /* Reset alt_txq_idx back if it is not overloaded */ 1451 if (stops < wakeups) { 1452 txq->alt_txq_idx = txq->idx; 1453 counter_u64_add(txq->stats.alt_reset, 1); 1454 } 1455 } 1456 rmb(); 1457 /* Schedule a tx enqueue task */ 1458 taskqueue_enqueue(txq->enqueue_tq, &txq->enqueue_task); 1459 } 1460 mtx_unlock(&txq->txq_mtx); 1461 } 1462 1463 if (atomic_sub_return(pkt_transmitted, &txq->pending_sends) < 0) 1464 mana_err(NULL, 1465 "WARNING: TX %d pending_sends error: %d\n", 1466 txq->idx, txq->pending_sends); 1467 1468 cq->work_done = pkt_transmitted; 1469 } 1470 1471 static void 1472 mana_post_pkt_rxq(struct mana_rxq *rxq) 1473 { 1474 struct mana_recv_buf_oob *recv_buf_oob; 1475 uint32_t curr_index; 1476 int err; 1477 1478 curr_index = rxq->buf_index++; 1479 if (rxq->buf_index == rxq->num_rx_buf) 1480 rxq->buf_index = 0; 1481 1482 recv_buf_oob = &rxq->rx_oobs[curr_index]; 1483 1484 err = mana_gd_post_and_ring(rxq->gdma_rq, &recv_buf_oob->wqe_req, 1485 &recv_buf_oob->wqe_inf); 1486 if (err) { 1487 mana_err(NULL, "WARNING: rxq %u post pkt err %d\n", 1488 rxq->rxq_idx, err); 1489 return; 1490 } 1491 1492 if (recv_buf_oob->wqe_inf.wqe_size_in_bu != 1) { 1493 mana_err(NULL, "WARNING: rxq %u wqe_size_in_bu %u\n", 1494 rxq->rxq_idx, recv_buf_oob->wqe_inf.wqe_size_in_bu); 1495 } 1496 } 1497 1498 static void 1499 mana_rx_mbuf(struct mbuf *mbuf, struct mana_rxcomp_oob *cqe, 1500 struct mana_rxq *rxq) 1501 { 1502 struct mana_stats *rx_stats = &rxq->stats; 1503 struct ifnet *ndev = rxq->ndev; 1504 uint32_t pkt_len = cqe->ppi[0].pkt_len; 1505 uint16_t rxq_idx = rxq->rxq_idx; 1506 struct mana_port_context *apc; 1507 bool do_lro = false; 1508 bool do_if_input; 1509 1510 apc = if_getsoftc(ndev); 1511 rxq->rx_cq.work_done++; 1512 1513 if (!mbuf) { 1514 return; 1515 } 1516 1517 mbuf->m_flags |= M_PKTHDR; 1518 mbuf->m_pkthdr.len = pkt_len; 1519 mbuf->m_len = pkt_len; 1520 mbuf->m_pkthdr.rcvif = ndev; 1521 1522 if ((ndev->if_capenable & IFCAP_RXCSUM || 1523 ndev->if_capenable & IFCAP_RXCSUM_IPV6) && 1524 (cqe->rx_iphdr_csum_succeed)) { 1525 mbuf->m_pkthdr.csum_flags = CSUM_IP_CHECKED; 1526 mbuf->m_pkthdr.csum_flags |= CSUM_IP_VALID; 1527 if (cqe->rx_tcp_csum_succeed || cqe->rx_udp_csum_succeed) { 1528 mbuf->m_pkthdr.csum_flags |= 1529 (CSUM_DATA_VALID | CSUM_PSEUDO_HDR); 1530 mbuf->m_pkthdr.csum_data = 0xffff; 1531 1532 if (cqe->rx_tcp_csum_succeed) 1533 do_lro = true; 1534 } 1535 } 1536 1537 if (cqe->rx_hashtype != 0) { 1538 mbuf->m_pkthdr.flowid = cqe->ppi[0].pkt_hash; 1539 1540 uint16_t hashtype = cqe->rx_hashtype; 1541 if (hashtype & NDIS_HASH_IPV4_MASK) { 1542 hashtype &= NDIS_HASH_IPV4_MASK; 1543 switch (hashtype) { 1544 case NDIS_HASH_TCP_IPV4: 1545 M_HASHTYPE_SET(mbuf, M_HASHTYPE_RSS_TCP_IPV4); 1546 break; 1547 case NDIS_HASH_UDP_IPV4: 1548 M_HASHTYPE_SET(mbuf, M_HASHTYPE_RSS_UDP_IPV4); 1549 break; 1550 default: 1551 M_HASHTYPE_SET(mbuf, M_HASHTYPE_RSS_IPV4); 1552 } 1553 } else if (hashtype & NDIS_HASH_IPV6_MASK) { 1554 hashtype &= NDIS_HASH_IPV6_MASK; 1555 switch (hashtype) { 1556 case NDIS_HASH_TCP_IPV6: 1557 M_HASHTYPE_SET(mbuf, M_HASHTYPE_RSS_TCP_IPV6); 1558 break; 1559 case NDIS_HASH_TCP_IPV6_EX: 1560 M_HASHTYPE_SET(mbuf, 1561 M_HASHTYPE_RSS_TCP_IPV6_EX); 1562 break; 1563 case NDIS_HASH_UDP_IPV6: 1564 M_HASHTYPE_SET(mbuf, M_HASHTYPE_RSS_UDP_IPV6); 1565 break; 1566 case NDIS_HASH_UDP_IPV6_EX: 1567 M_HASHTYPE_SET(mbuf, 1568 M_HASHTYPE_RSS_UDP_IPV6_EX); 1569 break; 1570 default: 1571 M_HASHTYPE_SET(mbuf, M_HASHTYPE_RSS_IPV6); 1572 } 1573 } else { 1574 M_HASHTYPE_SET(mbuf, M_HASHTYPE_OPAQUE_HASH); 1575 } 1576 } else { 1577 mbuf->m_pkthdr.flowid = rxq_idx; 1578 M_HASHTYPE_SET(mbuf, M_HASHTYPE_NONE); 1579 } 1580 1581 do_if_input = true; 1582 if ((ndev->if_capenable & IFCAP_LRO) && do_lro) { 1583 if (rxq->lro.lro_cnt != 0 && 1584 tcp_lro_rx(&rxq->lro, mbuf, 0) == 0) 1585 do_if_input = false; 1586 } 1587 if (do_if_input) { 1588 ndev->if_input(ndev, mbuf); 1589 } 1590 1591 counter_enter(); 1592 counter_u64_add_protected(rx_stats->packets, 1); 1593 counter_u64_add_protected(apc->port_stats.rx_packets, 1); 1594 counter_u64_add_protected(rx_stats->bytes, pkt_len); 1595 counter_u64_add_protected(apc->port_stats.rx_bytes, pkt_len); 1596 counter_exit(); 1597 } 1598 1599 static void 1600 mana_process_rx_cqe(struct mana_rxq *rxq, struct mana_cq *cq, 1601 struct gdma_comp *cqe) 1602 { 1603 struct mana_rxcomp_oob *oob = (struct mana_rxcomp_oob *)cqe->cqe_data; 1604 struct mana_recv_buf_oob *rxbuf_oob; 1605 struct ifnet *ndev = rxq->ndev; 1606 struct mana_port_context *apc; 1607 struct mbuf *old_mbuf; 1608 uint32_t curr, pktlen; 1609 int err; 1610 1611 switch (oob->cqe_hdr.cqe_type) { 1612 case CQE_RX_OKAY: 1613 break; 1614 1615 case CQE_RX_TRUNCATED: 1616 apc = if_getsoftc(ndev); 1617 counter_u64_add(apc->port_stats.rx_drops, 1); 1618 rxbuf_oob = &rxq->rx_oobs[rxq->buf_index]; 1619 if_printf(ndev, "Dropped a truncated packet\n"); 1620 goto drop; 1621 1622 case CQE_RX_COALESCED_4: 1623 if_printf(ndev, "RX coalescing is unsupported\n"); 1624 return; 1625 1626 case CQE_RX_OBJECT_FENCE: 1627 complete(&rxq->fence_event); 1628 return; 1629 1630 default: 1631 if_printf(ndev, "Unknown RX CQE type = %d\n", 1632 oob->cqe_hdr.cqe_type); 1633 return; 1634 } 1635 1636 if (oob->cqe_hdr.cqe_type != CQE_RX_OKAY) 1637 return; 1638 1639 pktlen = oob->ppi[0].pkt_len; 1640 1641 if (pktlen == 0) { 1642 /* data packets should never have packetlength of zero */ 1643 #if defined(__amd64__) 1644 if_printf(ndev, "RX pkt len=0, rq=%u, cq=%u, rxobj=0x%lx\n", 1645 rxq->gdma_id, cq->gdma_id, rxq->rxobj); 1646 #else 1647 if_printf(ndev, "RX pkt len=0, rq=%u, cq=%u, rxobj=0x%llx\n", 1648 rxq->gdma_id, cq->gdma_id, rxq->rxobj); 1649 #endif 1650 return; 1651 } 1652 1653 curr = rxq->buf_index; 1654 rxbuf_oob = &rxq->rx_oobs[curr]; 1655 if (rxbuf_oob->wqe_inf.wqe_size_in_bu != 1) { 1656 mana_err(NULL, "WARNING: Rx Incorrect complete " 1657 "WQE size %u\n", 1658 rxbuf_oob->wqe_inf.wqe_size_in_bu); 1659 } 1660 1661 apc = if_getsoftc(ndev); 1662 1663 old_mbuf = rxbuf_oob->mbuf; 1664 1665 /* Unload DMA map for the old mbuf */ 1666 mana_unload_rx_mbuf(apc, rxq, rxbuf_oob, false); 1667 1668 /* Load a new mbuf to replace the old one */ 1669 err = mana_load_rx_mbuf(apc, rxq, rxbuf_oob, true); 1670 if (err) { 1671 mana_dbg(NULL, 1672 "failed to load rx mbuf, err = %d, packet dropped.\n", 1673 err); 1674 counter_u64_add(rxq->stats.mbuf_alloc_fail, 1); 1675 /* 1676 * Failed to load new mbuf, rxbuf_oob->mbuf is still 1677 * pointing to the old one. Drop the packet. 1678 */ 1679 old_mbuf = NULL; 1680 /* Reload the existing mbuf */ 1681 mana_load_rx_mbuf(apc, rxq, rxbuf_oob, false); 1682 } 1683 1684 mana_rx_mbuf(old_mbuf, oob, rxq); 1685 1686 drop: 1687 mana_move_wq_tail(rxq->gdma_rq, rxbuf_oob->wqe_inf.wqe_size_in_bu); 1688 1689 mana_post_pkt_rxq(rxq); 1690 } 1691 1692 static void 1693 mana_poll_rx_cq(struct mana_cq *cq) 1694 { 1695 struct gdma_comp *comp = cq->gdma_comp_buf; 1696 int comp_read, i; 1697 1698 comp_read = mana_gd_poll_cq(cq->gdma_cq, comp, CQE_POLLING_BUFFER); 1699 KASSERT(comp_read <= CQE_POLLING_BUFFER, 1700 ("comp_read %d great than buf size %d", 1701 comp_read, CQE_POLLING_BUFFER)); 1702 1703 for (i = 0; i < comp_read; i++) { 1704 if (comp[i].is_sq == true) { 1705 mana_err(NULL, 1706 "WARNING: CQE not for receive queue\n"); 1707 return; 1708 } 1709 1710 /* verify recv cqe references the right rxq */ 1711 if (comp[i].wq_num != cq->rxq->gdma_id) { 1712 mana_err(NULL, 1713 "WARNING: Received CQE %d not for " 1714 "this receive queue %d\n", 1715 comp[i].wq_num, cq->rxq->gdma_id); 1716 return; 1717 } 1718 1719 mana_process_rx_cqe(cq->rxq, cq, &comp[i]); 1720 } 1721 1722 tcp_lro_flush_all(&cq->rxq->lro); 1723 } 1724 1725 static void 1726 mana_cq_handler(void *context, struct gdma_queue *gdma_queue) 1727 { 1728 struct mana_cq *cq = context; 1729 uint8_t arm_bit; 1730 1731 KASSERT(cq->gdma_cq == gdma_queue, 1732 ("cq do not match %p, %p", cq->gdma_cq, gdma_queue)); 1733 1734 if (cq->type == MANA_CQ_TYPE_RX) { 1735 mana_poll_rx_cq(cq); 1736 } else { 1737 mana_poll_tx_cq(cq); 1738 } 1739 1740 if (cq->work_done < cq->budget && cq->do_not_ring_db == false) 1741 arm_bit = SET_ARM_BIT; 1742 else 1743 arm_bit = 0; 1744 1745 mana_gd_ring_cq(gdma_queue, arm_bit); 1746 } 1747 1748 #define MANA_POLL_BUDGET 8 1749 #define MANA_RX_BUDGET 256 1750 #define MANA_TX_BUDGET MAX_SEND_BUFFERS_PER_QUEUE 1751 1752 static void 1753 mana_poll(void *arg, int pending) 1754 { 1755 struct mana_cq *cq = arg; 1756 int i; 1757 1758 cq->work_done = 0; 1759 if (cq->type == MANA_CQ_TYPE_RX) { 1760 cq->budget = MANA_RX_BUDGET; 1761 } else { 1762 cq->budget = MANA_TX_BUDGET; 1763 } 1764 1765 for (i = 0; i < MANA_POLL_BUDGET; i++) { 1766 /* 1767 * If this is the last loop, set the budget big enough 1768 * so it will arm the CQ any way. 1769 */ 1770 if (i == (MANA_POLL_BUDGET - 1)) 1771 cq->budget = CQE_POLLING_BUFFER + 1; 1772 1773 mana_cq_handler(cq, cq->gdma_cq); 1774 1775 if (cq->work_done < cq->budget) 1776 break; 1777 1778 cq->work_done = 0; 1779 } 1780 } 1781 1782 static void 1783 mana_schedule_task(void *arg, struct gdma_queue *gdma_queue) 1784 { 1785 struct mana_cq *cq = arg; 1786 1787 taskqueue_enqueue(cq->cleanup_tq, &cq->cleanup_task); 1788 } 1789 1790 static void 1791 mana_deinit_cq(struct mana_port_context *apc, struct mana_cq *cq) 1792 { 1793 struct gdma_dev *gd = apc->ac->gdma_dev; 1794 1795 if (!cq->gdma_cq) 1796 return; 1797 1798 /* Drain cleanup taskqueue */ 1799 if (cq->cleanup_tq) { 1800 while (taskqueue_cancel(cq->cleanup_tq, 1801 &cq->cleanup_task, NULL)) { 1802 taskqueue_drain(cq->cleanup_tq, 1803 &cq->cleanup_task); 1804 } 1805 1806 taskqueue_free(cq->cleanup_tq); 1807 } 1808 1809 mana_gd_destroy_queue(gd->gdma_context, cq->gdma_cq); 1810 } 1811 1812 static void 1813 mana_deinit_txq(struct mana_port_context *apc, struct mana_txq *txq) 1814 { 1815 struct gdma_dev *gd = apc->ac->gdma_dev; 1816 struct mana_send_buf_info *txbuf_info; 1817 uint32_t pending_sends; 1818 int i; 1819 1820 if (!txq->gdma_sq) 1821 return; 1822 1823 if ((pending_sends = atomic_read(&txq->pending_sends)) > 0) { 1824 mana_err(NULL, 1825 "WARNING: txq pending sends not zero: %u\n", 1826 pending_sends); 1827 } 1828 1829 if (txq->next_to_use != txq->next_to_complete) { 1830 mana_err(NULL, 1831 "WARNING: txq buf not completed, " 1832 "next use %u, next complete %u\n", 1833 txq->next_to_use, txq->next_to_complete); 1834 } 1835 1836 /* Flush buf ring. Grab txq mtx lock */ 1837 if (txq->txq_br) { 1838 mtx_lock(&txq->txq_mtx); 1839 drbr_flush(apc->ndev, txq->txq_br); 1840 mtx_unlock(&txq->txq_mtx); 1841 buf_ring_free(txq->txq_br, M_DEVBUF); 1842 } 1843 1844 /* Drain taskqueue */ 1845 if (txq->enqueue_tq) { 1846 while (taskqueue_cancel(txq->enqueue_tq, 1847 &txq->enqueue_task, NULL)) { 1848 taskqueue_drain(txq->enqueue_tq, 1849 &txq->enqueue_task); 1850 } 1851 1852 taskqueue_free(txq->enqueue_tq); 1853 } 1854 1855 if (txq->tx_buf_info) { 1856 /* Free all mbufs which are still in-flight */ 1857 for (i = 0; i < MAX_SEND_BUFFERS_PER_QUEUE; i++) { 1858 txbuf_info = &txq->tx_buf_info[i]; 1859 if (txbuf_info->mbuf) { 1860 mana_tx_unmap_mbuf(apc, txbuf_info); 1861 } 1862 } 1863 1864 free(txq->tx_buf_info, M_DEVBUF); 1865 } 1866 1867 mana_free_counters((counter_u64_t *)&txq->stats, 1868 sizeof(txq->stats)); 1869 1870 mana_gd_destroy_queue(gd->gdma_context, txq->gdma_sq); 1871 1872 mtx_destroy(&txq->txq_mtx); 1873 } 1874 1875 static void 1876 mana_destroy_txq(struct mana_port_context *apc) 1877 { 1878 int i; 1879 1880 if (!apc->tx_qp) 1881 return; 1882 1883 for (i = 0; i < apc->num_queues; i++) { 1884 mana_destroy_wq_obj(apc, GDMA_SQ, apc->tx_qp[i].tx_object); 1885 1886 mana_deinit_cq(apc, &apc->tx_qp[i].tx_cq); 1887 1888 mana_deinit_txq(apc, &apc->tx_qp[i].txq); 1889 } 1890 1891 free(apc->tx_qp, M_DEVBUF); 1892 apc->tx_qp = NULL; 1893 } 1894 1895 static int 1896 mana_create_txq(struct mana_port_context *apc, struct ifnet *net) 1897 { 1898 struct mana_context *ac = apc->ac; 1899 struct gdma_dev *gd = ac->gdma_dev; 1900 struct mana_obj_spec wq_spec; 1901 struct mana_obj_spec cq_spec; 1902 struct gdma_queue_spec spec; 1903 struct gdma_context *gc; 1904 struct mana_txq *txq; 1905 struct mana_cq *cq; 1906 uint32_t txq_size; 1907 uint32_t cq_size; 1908 int err; 1909 int i; 1910 1911 apc->tx_qp = mallocarray(apc->num_queues, sizeof(struct mana_tx_qp), 1912 M_DEVBUF, M_WAITOK | M_ZERO); 1913 if (!apc->tx_qp) 1914 return ENOMEM; 1915 1916 /* The minimum size of the WQE is 32 bytes, hence 1917 * MAX_SEND_BUFFERS_PER_QUEUE represents the maximum number of WQEs 1918 * the SQ can store. This value is then used to size other queues 1919 * to prevent overflow. 1920 */ 1921 txq_size = MAX_SEND_BUFFERS_PER_QUEUE * 32; 1922 KASSERT(IS_ALIGNED(txq_size, PAGE_SIZE), 1923 ("txq size not page aligned")); 1924 1925 cq_size = MAX_SEND_BUFFERS_PER_QUEUE * COMP_ENTRY_SIZE; 1926 cq_size = ALIGN(cq_size, PAGE_SIZE); 1927 1928 gc = gd->gdma_context; 1929 1930 for (i = 0; i < apc->num_queues; i++) { 1931 apc->tx_qp[i].tx_object = INVALID_MANA_HANDLE; 1932 1933 /* Create SQ */ 1934 txq = &apc->tx_qp[i].txq; 1935 1936 txq->ndev = net; 1937 txq->vp_offset = apc->tx_vp_offset; 1938 txq->idx = i; 1939 txq->alt_txq_idx = i; 1940 1941 memset(&spec, 0, sizeof(spec)); 1942 spec.type = GDMA_SQ; 1943 spec.monitor_avl_buf = true; 1944 spec.queue_size = txq_size; 1945 err = mana_gd_create_mana_wq_cq(gd, &spec, &txq->gdma_sq); 1946 if (err) 1947 goto out; 1948 1949 /* Create SQ's CQ */ 1950 cq = &apc->tx_qp[i].tx_cq; 1951 cq->type = MANA_CQ_TYPE_TX; 1952 1953 cq->txq = txq; 1954 1955 memset(&spec, 0, sizeof(spec)); 1956 spec.type = GDMA_CQ; 1957 spec.monitor_avl_buf = false; 1958 spec.queue_size = cq_size; 1959 spec.cq.callback = mana_schedule_task; 1960 spec.cq.parent_eq = ac->eqs[i].eq; 1961 spec.cq.context = cq; 1962 err = mana_gd_create_mana_wq_cq(gd, &spec, &cq->gdma_cq); 1963 if (err) 1964 goto out; 1965 1966 memset(&wq_spec, 0, sizeof(wq_spec)); 1967 memset(&cq_spec, 0, sizeof(cq_spec)); 1968 1969 wq_spec.gdma_region = txq->gdma_sq->mem_info.gdma_region; 1970 wq_spec.queue_size = txq->gdma_sq->queue_size; 1971 1972 cq_spec.gdma_region = cq->gdma_cq->mem_info.gdma_region; 1973 cq_spec.queue_size = cq->gdma_cq->queue_size; 1974 cq_spec.modr_ctx_id = 0; 1975 cq_spec.attached_eq = cq->gdma_cq->cq.parent->id; 1976 1977 err = mana_create_wq_obj(apc, apc->port_handle, GDMA_SQ, 1978 &wq_spec, &cq_spec, &apc->tx_qp[i].tx_object); 1979 1980 if (err) 1981 goto out; 1982 1983 txq->gdma_sq->id = wq_spec.queue_index; 1984 cq->gdma_cq->id = cq_spec.queue_index; 1985 1986 txq->gdma_sq->mem_info.gdma_region = GDMA_INVALID_DMA_REGION; 1987 cq->gdma_cq->mem_info.gdma_region = GDMA_INVALID_DMA_REGION; 1988 1989 txq->gdma_txq_id = txq->gdma_sq->id; 1990 1991 cq->gdma_id = cq->gdma_cq->id; 1992 1993 mana_dbg(NULL, 1994 "txq %d, txq gdma id %d, txq cq gdma id %d\n", 1995 i, txq->gdma_txq_id, cq->gdma_id);; 1996 1997 if (cq->gdma_id >= gc->max_num_cqs) { 1998 if_printf(net, "CQ id %u too large.\n", cq->gdma_id); 1999 err = EINVAL; 2000 goto out; 2001 } 2002 2003 gc->cq_table[cq->gdma_id] = cq->gdma_cq; 2004 2005 /* Initialize tx specific data */ 2006 txq->tx_buf_info = malloc(MAX_SEND_BUFFERS_PER_QUEUE * 2007 sizeof(struct mana_send_buf_info), 2008 M_DEVBUF, M_WAITOK | M_ZERO); 2009 if (unlikely(txq->tx_buf_info == NULL)) { 2010 if_printf(net, 2011 "Failed to allocate tx buf info for SQ %u\n", 2012 txq->gdma_sq->id); 2013 err = ENOMEM; 2014 goto out; 2015 } 2016 2017 2018 snprintf(txq->txq_mtx_name, nitems(txq->txq_mtx_name), 2019 "mana:tx(%d)", i); 2020 mtx_init(&txq->txq_mtx, txq->txq_mtx_name, NULL, MTX_DEF); 2021 2022 txq->txq_br = buf_ring_alloc(4 * MAX_SEND_BUFFERS_PER_QUEUE, 2023 M_DEVBUF, M_WAITOK, &txq->txq_mtx); 2024 if (unlikely(txq->txq_br == NULL)) { 2025 if_printf(net, 2026 "Failed to allocate buf ring for SQ %u\n", 2027 txq->gdma_sq->id); 2028 err = ENOMEM; 2029 goto out; 2030 } 2031 2032 /* Allocate taskqueue for deferred send */ 2033 TASK_INIT(&txq->enqueue_task, 0, mana_xmit_taskfunc, txq); 2034 txq->enqueue_tq = taskqueue_create_fast("mana_tx_enque", 2035 M_NOWAIT, taskqueue_thread_enqueue, &txq->enqueue_tq); 2036 if (unlikely(txq->enqueue_tq == NULL)) { 2037 if_printf(net, 2038 "Unable to create tx %d enqueue task queue\n", i); 2039 err = ENOMEM; 2040 goto out; 2041 } 2042 taskqueue_start_threads(&txq->enqueue_tq, 1, PI_NET, 2043 "mana txq p%u-tx%d", apc->port_idx, i); 2044 2045 mana_alloc_counters((counter_u64_t *)&txq->stats, 2046 sizeof(txq->stats)); 2047 2048 /* Allocate and start the cleanup task on CQ */ 2049 cq->do_not_ring_db = false; 2050 2051 NET_TASK_INIT(&cq->cleanup_task, 0, mana_poll, cq); 2052 cq->cleanup_tq = 2053 taskqueue_create_fast("mana tx cq cleanup", 2054 M_WAITOK, taskqueue_thread_enqueue, 2055 &cq->cleanup_tq); 2056 2057 if (apc->last_tx_cq_bind_cpu < 0) 2058 apc->last_tx_cq_bind_cpu = CPU_FIRST(); 2059 cq->cpu = apc->last_tx_cq_bind_cpu; 2060 apc->last_tx_cq_bind_cpu = CPU_NEXT(apc->last_tx_cq_bind_cpu); 2061 2062 if (apc->bind_cleanup_thread_cpu) { 2063 cpuset_t cpu_mask; 2064 CPU_SETOF(cq->cpu, &cpu_mask); 2065 taskqueue_start_threads_cpuset(&cq->cleanup_tq, 2066 1, PI_NET, &cpu_mask, 2067 "mana cq p%u-tx%u-cpu%d", 2068 apc->port_idx, txq->idx, cq->cpu); 2069 } else { 2070 taskqueue_start_threads(&cq->cleanup_tq, 1, 2071 PI_NET, "mana cq p%u-tx%u", 2072 apc->port_idx, txq->idx); 2073 } 2074 2075 mana_gd_ring_cq(cq->gdma_cq, SET_ARM_BIT); 2076 } 2077 2078 return 0; 2079 out: 2080 mana_destroy_txq(apc); 2081 return err; 2082 } 2083 2084 static void 2085 mana_destroy_rxq(struct mana_port_context *apc, struct mana_rxq *rxq, 2086 bool validate_state) 2087 { 2088 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context; 2089 struct mana_recv_buf_oob *rx_oob; 2090 int i; 2091 2092 if (!rxq) 2093 return; 2094 2095 if (validate_state) { 2096 /* 2097 * XXX Cancel and drain cleanup task queue here. 2098 */ 2099 ; 2100 } 2101 2102 mana_destroy_wq_obj(apc, GDMA_RQ, rxq->rxobj); 2103 2104 mana_deinit_cq(apc, &rxq->rx_cq); 2105 2106 mana_free_counters((counter_u64_t *)&rxq->stats, 2107 sizeof(rxq->stats)); 2108 2109 /* Free LRO resources */ 2110 tcp_lro_free(&rxq->lro); 2111 2112 for (i = 0; i < rxq->num_rx_buf; i++) { 2113 rx_oob = &rxq->rx_oobs[i]; 2114 2115 if (rx_oob->mbuf) 2116 mana_unload_rx_mbuf(apc, rxq, rx_oob, true); 2117 2118 bus_dmamap_destroy(apc->rx_buf_tag, rx_oob->dma_map); 2119 } 2120 2121 if (rxq->gdma_rq) 2122 mana_gd_destroy_queue(gc, rxq->gdma_rq); 2123 2124 free(rxq, M_DEVBUF); 2125 } 2126 2127 #define MANA_WQE_HEADER_SIZE 16 2128 #define MANA_WQE_SGE_SIZE 16 2129 2130 static int 2131 mana_alloc_rx_wqe(struct mana_port_context *apc, 2132 struct mana_rxq *rxq, uint32_t *rxq_size, uint32_t *cq_size) 2133 { 2134 struct mana_recv_buf_oob *rx_oob; 2135 uint32_t buf_idx; 2136 int err; 2137 2138 if (rxq->datasize == 0 || rxq->datasize > PAGE_SIZE) { 2139 mana_err(NULL, 2140 "WARNING: Invalid rxq datasize %u\n", rxq->datasize); 2141 } 2142 2143 *rxq_size = 0; 2144 *cq_size = 0; 2145 2146 for (buf_idx = 0; buf_idx < rxq->num_rx_buf; buf_idx++) { 2147 rx_oob = &rxq->rx_oobs[buf_idx]; 2148 memset(rx_oob, 0, sizeof(*rx_oob)); 2149 2150 err = bus_dmamap_create(apc->rx_buf_tag, 0, 2151 &rx_oob->dma_map); 2152 if (err) { 2153 mana_err(NULL, 2154 "Failed to create rx DMA map for buf %d\n", 2155 buf_idx); 2156 return err; 2157 } 2158 2159 err = mana_load_rx_mbuf(apc, rxq, rx_oob, true); 2160 if (err) { 2161 mana_err(NULL, 2162 "Failed to create rx DMA map for buf %d\n", 2163 buf_idx); 2164 bus_dmamap_destroy(apc->rx_buf_tag, rx_oob->dma_map); 2165 return err; 2166 } 2167 2168 rx_oob->wqe_req.sgl = rx_oob->sgl; 2169 rx_oob->wqe_req.num_sge = rx_oob->num_sge; 2170 rx_oob->wqe_req.inline_oob_size = 0; 2171 rx_oob->wqe_req.inline_oob_data = NULL; 2172 rx_oob->wqe_req.flags = 0; 2173 rx_oob->wqe_req.client_data_unit = 0; 2174 2175 *rxq_size += ALIGN(MANA_WQE_HEADER_SIZE + 2176 MANA_WQE_SGE_SIZE * rx_oob->num_sge, 32); 2177 *cq_size += COMP_ENTRY_SIZE; 2178 } 2179 2180 return 0; 2181 } 2182 2183 static int 2184 mana_push_wqe(struct mana_rxq *rxq) 2185 { 2186 struct mana_recv_buf_oob *rx_oob; 2187 uint32_t buf_idx; 2188 int err; 2189 2190 for (buf_idx = 0; buf_idx < rxq->num_rx_buf; buf_idx++) { 2191 rx_oob = &rxq->rx_oobs[buf_idx]; 2192 2193 err = mana_gd_post_and_ring(rxq->gdma_rq, &rx_oob->wqe_req, 2194 &rx_oob->wqe_inf); 2195 if (err) 2196 return ENOSPC; 2197 } 2198 2199 return 0; 2200 } 2201 2202 static struct mana_rxq * 2203 mana_create_rxq(struct mana_port_context *apc, uint32_t rxq_idx, 2204 struct mana_eq *eq, struct ifnet *ndev) 2205 { 2206 struct gdma_dev *gd = apc->ac->gdma_dev; 2207 struct mana_obj_spec wq_spec; 2208 struct mana_obj_spec cq_spec; 2209 struct gdma_queue_spec spec; 2210 struct mana_cq *cq = NULL; 2211 uint32_t cq_size, rq_size; 2212 struct gdma_context *gc; 2213 struct mana_rxq *rxq; 2214 int err; 2215 2216 gc = gd->gdma_context; 2217 2218 rxq = malloc(sizeof(*rxq) + 2219 RX_BUFFERS_PER_QUEUE * sizeof(struct mana_recv_buf_oob), 2220 M_DEVBUF, M_WAITOK | M_ZERO); 2221 if (!rxq) 2222 return NULL; 2223 2224 rxq->ndev = ndev; 2225 rxq->num_rx_buf = RX_BUFFERS_PER_QUEUE; 2226 rxq->rxq_idx = rxq_idx; 2227 /* 2228 * Minimum size is MCLBYTES(2048) bytes for a mbuf cluster. 2229 * Now we just allow maximum size of 4096. 2230 */ 2231 rxq->datasize = ALIGN(apc->frame_size, MCLBYTES); 2232 if (rxq->datasize > MAX_FRAME_SIZE) 2233 rxq->datasize = MAX_FRAME_SIZE; 2234 2235 mana_dbg(NULL, "Setting rxq %d datasize %d\n", 2236 rxq_idx, rxq->datasize); 2237 2238 rxq->rxobj = INVALID_MANA_HANDLE; 2239 2240 err = mana_alloc_rx_wqe(apc, rxq, &rq_size, &cq_size); 2241 if (err) 2242 goto out; 2243 2244 /* Create LRO for the RQ */ 2245 if (ndev->if_capenable & IFCAP_LRO) { 2246 err = tcp_lro_init(&rxq->lro); 2247 if (err) { 2248 if_printf(ndev, "Failed to create LRO for rxq %d\n", 2249 rxq_idx); 2250 } else { 2251 rxq->lro.ifp = ndev; 2252 } 2253 } 2254 2255 mana_alloc_counters((counter_u64_t *)&rxq->stats, 2256 sizeof(rxq->stats)); 2257 2258 rq_size = ALIGN(rq_size, PAGE_SIZE); 2259 cq_size = ALIGN(cq_size, PAGE_SIZE); 2260 2261 /* Create RQ */ 2262 memset(&spec, 0, sizeof(spec)); 2263 spec.type = GDMA_RQ; 2264 spec.monitor_avl_buf = true; 2265 spec.queue_size = rq_size; 2266 err = mana_gd_create_mana_wq_cq(gd, &spec, &rxq->gdma_rq); 2267 if (err) 2268 goto out; 2269 2270 /* Create RQ's CQ */ 2271 cq = &rxq->rx_cq; 2272 cq->type = MANA_CQ_TYPE_RX; 2273 cq->rxq = rxq; 2274 2275 memset(&spec, 0, sizeof(spec)); 2276 spec.type = GDMA_CQ; 2277 spec.monitor_avl_buf = false; 2278 spec.queue_size = cq_size; 2279 spec.cq.callback = mana_schedule_task; 2280 spec.cq.parent_eq = eq->eq; 2281 spec.cq.context = cq; 2282 err = mana_gd_create_mana_wq_cq(gd, &spec, &cq->gdma_cq); 2283 if (err) 2284 goto out; 2285 2286 memset(&wq_spec, 0, sizeof(wq_spec)); 2287 memset(&cq_spec, 0, sizeof(cq_spec)); 2288 wq_spec.gdma_region = rxq->gdma_rq->mem_info.gdma_region; 2289 wq_spec.queue_size = rxq->gdma_rq->queue_size; 2290 2291 cq_spec.gdma_region = cq->gdma_cq->mem_info.gdma_region; 2292 cq_spec.queue_size = cq->gdma_cq->queue_size; 2293 cq_spec.modr_ctx_id = 0; 2294 cq_spec.attached_eq = cq->gdma_cq->cq.parent->id; 2295 2296 err = mana_create_wq_obj(apc, apc->port_handle, GDMA_RQ, 2297 &wq_spec, &cq_spec, &rxq->rxobj); 2298 if (err) 2299 goto out; 2300 2301 rxq->gdma_rq->id = wq_spec.queue_index; 2302 cq->gdma_cq->id = cq_spec.queue_index; 2303 2304 rxq->gdma_rq->mem_info.gdma_region = GDMA_INVALID_DMA_REGION; 2305 cq->gdma_cq->mem_info.gdma_region = GDMA_INVALID_DMA_REGION; 2306 2307 rxq->gdma_id = rxq->gdma_rq->id; 2308 cq->gdma_id = cq->gdma_cq->id; 2309 2310 err = mana_push_wqe(rxq); 2311 if (err) 2312 goto out; 2313 2314 if (cq->gdma_id >= gc->max_num_cqs) { 2315 err = EINVAL; 2316 goto out; 2317 } 2318 2319 gc->cq_table[cq->gdma_id] = cq->gdma_cq; 2320 2321 /* Allocate and start the cleanup task on CQ */ 2322 cq->do_not_ring_db = false; 2323 2324 NET_TASK_INIT(&cq->cleanup_task, 0, mana_poll, cq); 2325 cq->cleanup_tq = 2326 taskqueue_create_fast("mana rx cq cleanup", 2327 M_WAITOK, taskqueue_thread_enqueue, 2328 &cq->cleanup_tq); 2329 2330 if (apc->last_rx_cq_bind_cpu < 0) 2331 apc->last_rx_cq_bind_cpu = CPU_FIRST(); 2332 cq->cpu = apc->last_rx_cq_bind_cpu; 2333 apc->last_rx_cq_bind_cpu = CPU_NEXT(apc->last_rx_cq_bind_cpu); 2334 2335 if (apc->bind_cleanup_thread_cpu) { 2336 cpuset_t cpu_mask; 2337 CPU_SETOF(cq->cpu, &cpu_mask); 2338 taskqueue_start_threads_cpuset(&cq->cleanup_tq, 2339 1, PI_NET, &cpu_mask, 2340 "mana cq p%u-rx%u-cpu%d", 2341 apc->port_idx, rxq->rxq_idx, cq->cpu); 2342 } else { 2343 taskqueue_start_threads(&cq->cleanup_tq, 1, 2344 PI_NET, "mana cq p%u-rx%u", 2345 apc->port_idx, rxq->rxq_idx); 2346 } 2347 2348 mana_gd_ring_cq(cq->gdma_cq, SET_ARM_BIT); 2349 out: 2350 if (!err) 2351 return rxq; 2352 2353 if_printf(ndev, "Failed to create RXQ: err = %d\n", err); 2354 2355 mana_destroy_rxq(apc, rxq, false); 2356 2357 if (cq) 2358 mana_deinit_cq(apc, cq); 2359 2360 return NULL; 2361 } 2362 2363 static int 2364 mana_add_rx_queues(struct mana_port_context *apc, struct ifnet *ndev) 2365 { 2366 struct mana_context *ac = apc->ac; 2367 struct mana_rxq *rxq; 2368 int err = 0; 2369 int i; 2370 2371 for (i = 0; i < apc->num_queues; i++) { 2372 rxq = mana_create_rxq(apc, i, &ac->eqs[i], ndev); 2373 if (!rxq) { 2374 err = ENOMEM; 2375 goto out; 2376 } 2377 2378 apc->rxqs[i] = rxq; 2379 } 2380 2381 apc->default_rxobj = apc->rxqs[0]->rxobj; 2382 out: 2383 return err; 2384 } 2385 2386 static void 2387 mana_destroy_vport(struct mana_port_context *apc) 2388 { 2389 struct mana_rxq *rxq; 2390 uint32_t rxq_idx; 2391 2392 for (rxq_idx = 0; rxq_idx < apc->num_queues; rxq_idx++) { 2393 rxq = apc->rxqs[rxq_idx]; 2394 if (!rxq) 2395 continue; 2396 2397 mana_destroy_rxq(apc, rxq, true); 2398 apc->rxqs[rxq_idx] = NULL; 2399 } 2400 2401 mana_destroy_txq(apc); 2402 } 2403 2404 static int 2405 mana_create_vport(struct mana_port_context *apc, struct ifnet *net) 2406 { 2407 struct gdma_dev *gd = apc->ac->gdma_dev; 2408 int err; 2409 2410 apc->default_rxobj = INVALID_MANA_HANDLE; 2411 2412 err = mana_cfg_vport(apc, gd->pdid, gd->doorbell); 2413 if (err) 2414 return err; 2415 2416 return mana_create_txq(apc, net); 2417 } 2418 2419 2420 static void mana_rss_table_init(struct mana_port_context *apc) 2421 { 2422 int i; 2423 2424 for (i = 0; i < MANA_INDIRECT_TABLE_SIZE; i++) 2425 apc->indir_table[i] = i % apc->num_queues; 2426 } 2427 2428 int mana_config_rss(struct mana_port_context *apc, enum TRI_STATE rx, 2429 bool update_hash, bool update_tab) 2430 { 2431 uint32_t queue_idx; 2432 int err; 2433 int i; 2434 2435 if (update_tab) { 2436 for (i = 0; i < MANA_INDIRECT_TABLE_SIZE; i++) { 2437 queue_idx = apc->indir_table[i]; 2438 apc->rxobj_table[i] = apc->rxqs[queue_idx]->rxobj; 2439 } 2440 } 2441 2442 err = mana_cfg_vport_steering(apc, rx, true, update_hash, update_tab); 2443 if (err) 2444 return err; 2445 2446 mana_fence_rqs(apc); 2447 2448 return 0; 2449 } 2450 2451 static int 2452 mana_init_port(struct ifnet *ndev) 2453 { 2454 struct mana_port_context *apc = if_getsoftc(ndev); 2455 uint32_t max_txq, max_rxq, max_queues; 2456 int port_idx = apc->port_idx; 2457 uint32_t num_indirect_entries; 2458 int err; 2459 2460 err = mana_init_port_context(apc); 2461 if (err) 2462 return err; 2463 2464 err = mana_query_vport_cfg(apc, port_idx, &max_txq, &max_rxq, 2465 &num_indirect_entries); 2466 if (err) { 2467 if_printf(ndev, "Failed to query info for vPort %d\n", 2468 port_idx); 2469 goto reset_apc; 2470 } 2471 2472 max_queues = min_t(uint32_t, max_txq, max_rxq); 2473 if (apc->max_queues > max_queues) 2474 apc->max_queues = max_queues; 2475 2476 if (apc->num_queues > apc->max_queues) 2477 apc->num_queues = apc->max_queues; 2478 2479 return 0; 2480 2481 reset_apc: 2482 bus_dma_tag_destroy(apc->rx_buf_tag); 2483 apc->rx_buf_tag = NULL; 2484 free(apc->rxqs, M_DEVBUF); 2485 apc->rxqs = NULL; 2486 return err; 2487 } 2488 2489 int 2490 mana_alloc_queues(struct ifnet *ndev) 2491 { 2492 struct mana_port_context *apc = if_getsoftc(ndev); 2493 int err; 2494 2495 err = mana_create_vport(apc, ndev); 2496 if (err) 2497 return err; 2498 2499 err = mana_add_rx_queues(apc, ndev); 2500 if (err) 2501 goto destroy_vport; 2502 2503 apc->rss_state = apc->num_queues > 1 ? TRI_STATE_TRUE : TRI_STATE_FALSE; 2504 2505 mana_rss_table_init(apc); 2506 2507 err = mana_config_rss(apc, TRI_STATE_TRUE, true, true); 2508 if (err) 2509 goto destroy_vport; 2510 2511 return 0; 2512 2513 destroy_vport: 2514 mana_destroy_vport(apc); 2515 return err; 2516 } 2517 2518 static int 2519 mana_up(struct mana_port_context *apc) 2520 { 2521 int err; 2522 2523 mana_dbg(NULL, "mana_up called\n"); 2524 2525 err = mana_alloc_queues(apc->ndev); 2526 if (err) { 2527 mana_err(NULL, "Faile alloc mana queues: %d\n", err); 2528 return err; 2529 } 2530 2531 /* Add queue specific sysctl */ 2532 mana_sysctl_add_queues(apc); 2533 2534 apc->port_is_up = true; 2535 2536 /* Ensure port state updated before txq state */ 2537 wmb(); 2538 2539 if_link_state_change(apc->ndev, LINK_STATE_UP); 2540 if_setdrvflagbits(apc->ndev, IFF_DRV_RUNNING, IFF_DRV_OACTIVE); 2541 2542 return 0; 2543 } 2544 2545 2546 static void 2547 mana_init(void *arg) 2548 { 2549 struct mana_port_context *apc = (struct mana_port_context *)arg; 2550 2551 MANA_APC_LOCK_LOCK(apc); 2552 if (!apc->port_is_up) { 2553 mana_up(apc); 2554 } 2555 MANA_APC_LOCK_UNLOCK(apc); 2556 } 2557 2558 static int 2559 mana_dealloc_queues(struct ifnet *ndev) 2560 { 2561 struct mana_port_context *apc = if_getsoftc(ndev); 2562 struct mana_txq *txq; 2563 int i, err; 2564 2565 if (apc->port_is_up) 2566 return EINVAL; 2567 2568 /* No packet can be transmitted now since apc->port_is_up is false. 2569 * There is still a tiny chance that mana_poll_tx_cq() can re-enable 2570 * a txq because it may not timely see apc->port_is_up being cleared 2571 * to false, but it doesn't matter since mana_start_xmit() drops any 2572 * new packets due to apc->port_is_up being false. 2573 * 2574 * Drain all the in-flight TX packets 2575 */ 2576 for (i = 0; i < apc->num_queues; i++) { 2577 txq = &apc->tx_qp[i].txq; 2578 2579 struct mana_cq *tx_cq = &apc->tx_qp[i].tx_cq; 2580 struct mana_cq *rx_cq = &(apc->rxqs[i]->rx_cq); 2581 2582 tx_cq->do_not_ring_db = true; 2583 rx_cq->do_not_ring_db = true; 2584 2585 /* Schedule a cleanup task */ 2586 taskqueue_enqueue(tx_cq->cleanup_tq, &tx_cq->cleanup_task); 2587 2588 while (atomic_read(&txq->pending_sends) > 0) 2589 usleep_range(1000, 2000); 2590 } 2591 2592 /* We're 100% sure the queues can no longer be woken up, because 2593 * we're sure now mana_poll_tx_cq() can't be running. 2594 */ 2595 2596 apc->rss_state = TRI_STATE_FALSE; 2597 err = mana_config_rss(apc, TRI_STATE_FALSE, false, false); 2598 if (err) { 2599 if_printf(ndev, "Failed to disable vPort: %d\n", err); 2600 return err; 2601 } 2602 2603 mana_destroy_vport(apc); 2604 2605 return 0; 2606 } 2607 2608 static int 2609 mana_down(struct mana_port_context *apc) 2610 { 2611 int err = 0; 2612 2613 apc->port_st_save = apc->port_is_up; 2614 apc->port_is_up = false; 2615 2616 /* Ensure port state updated before txq state */ 2617 wmb(); 2618 2619 if (apc->port_st_save) { 2620 if_setdrvflagbits(apc->ndev, IFF_DRV_OACTIVE, 2621 IFF_DRV_RUNNING); 2622 if_link_state_change(apc->ndev, LINK_STATE_DOWN); 2623 2624 mana_sysctl_free_queues(apc); 2625 2626 err = mana_dealloc_queues(apc->ndev); 2627 if (err) { 2628 if_printf(apc->ndev, 2629 "Failed to bring down mana interface: %d\n", err); 2630 } 2631 } 2632 2633 return err; 2634 } 2635 2636 int 2637 mana_detach(struct ifnet *ndev) 2638 { 2639 struct mana_port_context *apc = if_getsoftc(ndev); 2640 int err; 2641 2642 ether_ifdetach(ndev); 2643 2644 if (!apc) 2645 return 0; 2646 2647 MANA_APC_LOCK_LOCK(apc); 2648 err = mana_down(apc); 2649 MANA_APC_LOCK_UNLOCK(apc); 2650 2651 mana_cleanup_port_context(apc); 2652 2653 MANA_APC_LOCK_DESTROY(apc); 2654 2655 free(apc, M_DEVBUF); 2656 2657 return err; 2658 } 2659 2660 static int 2661 mana_probe_port(struct mana_context *ac, int port_idx, 2662 struct ifnet **ndev_storage) 2663 { 2664 struct gdma_context *gc = ac->gdma_dev->gdma_context; 2665 struct mana_port_context *apc; 2666 struct ifnet *ndev; 2667 int err; 2668 2669 ndev = if_alloc_dev(IFT_ETHER, gc->dev); 2670 if (!ndev) { 2671 mana_err(NULL, "Failed to allocate ifnet struct\n"); 2672 return ENOMEM; 2673 } 2674 2675 *ndev_storage = ndev; 2676 2677 apc = malloc(sizeof(*apc), M_DEVBUF, M_WAITOK | M_ZERO); 2678 if (!apc) { 2679 mana_err(NULL, "Failed to allocate port context\n"); 2680 err = ENOMEM; 2681 goto free_net; 2682 } 2683 2684 apc->ac = ac; 2685 apc->ndev = ndev; 2686 apc->max_queues = gc->max_num_queues; 2687 apc->num_queues = min_t(unsigned int, 2688 gc->max_num_queues, MANA_MAX_NUM_QUEUES); 2689 apc->port_handle = INVALID_MANA_HANDLE; 2690 apc->port_idx = port_idx; 2691 apc->frame_size = DEFAULT_FRAME_SIZE; 2692 apc->last_tx_cq_bind_cpu = -1; 2693 apc->last_rx_cq_bind_cpu = -1; 2694 2695 MANA_APC_LOCK_INIT(apc); 2696 2697 if_initname(ndev, device_get_name(gc->dev), port_idx); 2698 if_setdev(ndev,gc->dev); 2699 if_setsoftc(ndev, apc); 2700 2701 if_setflags(ndev, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST); 2702 if_setinitfn(ndev, mana_init); 2703 if_settransmitfn(ndev, mana_start_xmit); 2704 if_setqflushfn(ndev, mana_qflush); 2705 if_setioctlfn(ndev, mana_ioctl); 2706 if_setgetcounterfn(ndev, mana_get_counter); 2707 2708 if_setmtu(ndev, ETHERMTU); 2709 if_setbaudrate(ndev, IF_Gbps(100)); 2710 2711 mana_rss_key_fill(apc->hashkey, MANA_HASH_KEY_SIZE); 2712 2713 err = mana_init_port(ndev); 2714 if (err) 2715 goto reset_apc; 2716 2717 ndev->if_capabilities |= IFCAP_TXCSUM | IFCAP_TXCSUM_IPV6; 2718 ndev->if_capabilities |= IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6; 2719 ndev->if_capabilities |= IFCAP_TSO4 | IFCAP_TSO6; 2720 2721 ndev->if_capabilities |= IFCAP_LRO | IFCAP_LINKSTATE; 2722 2723 /* Enable all available capabilities by default. */ 2724 ndev->if_capenable = ndev->if_capabilities; 2725 2726 /* TSO parameters */ 2727 ndev->if_hw_tsomax = MAX_MBUF_FRAGS * MANA_TSO_MAXSEG_SZ - 2728 (ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN); 2729 ndev->if_hw_tsomaxsegcount = MAX_MBUF_FRAGS; 2730 ndev->if_hw_tsomaxsegsize = PAGE_SIZE; 2731 2732 ifmedia_init(&apc->media, IFM_IMASK, 2733 mana_ifmedia_change, mana_ifmedia_status); 2734 ifmedia_add(&apc->media, IFM_ETHER | IFM_AUTO, 0, NULL); 2735 ifmedia_set(&apc->media, IFM_ETHER | IFM_AUTO); 2736 2737 ether_ifattach(ndev, apc->mac_addr); 2738 2739 /* Initialize statistics */ 2740 mana_alloc_counters((counter_u64_t *)&apc->port_stats, 2741 sizeof(struct mana_port_stats)); 2742 mana_sysctl_add_port(apc); 2743 2744 /* Tell the stack that the interface is not active */ 2745 if_setdrvflagbits(ndev, IFF_DRV_OACTIVE, IFF_DRV_RUNNING); 2746 2747 return 0; 2748 2749 reset_apc: 2750 free(apc, M_DEVBUF); 2751 free_net: 2752 *ndev_storage = NULL; 2753 if_printf(ndev, "Failed to probe vPort %d: %d\n", port_idx, err); 2754 if_free(ndev); 2755 return err; 2756 } 2757 2758 int mana_probe(struct gdma_dev *gd) 2759 { 2760 struct gdma_context *gc = gd->gdma_context; 2761 device_t dev = gc->dev; 2762 struct mana_context *ac; 2763 int err; 2764 int i; 2765 2766 device_printf(dev, "%s protocol version: %d.%d.%d\n", DEVICE_NAME, 2767 MANA_MAJOR_VERSION, MANA_MINOR_VERSION, MANA_MICRO_VERSION); 2768 2769 err = mana_gd_register_device(gd); 2770 if (err) 2771 return err; 2772 2773 ac = malloc(sizeof(*ac), M_DEVBUF, M_WAITOK | M_ZERO); 2774 if (!ac) 2775 return ENOMEM; 2776 2777 ac->gdma_dev = gd; 2778 ac->num_ports = 1; 2779 gd->driver_data = ac; 2780 2781 err = mana_create_eq(ac); 2782 if (err) 2783 goto out; 2784 2785 err = mana_query_device_cfg(ac, MANA_MAJOR_VERSION, MANA_MINOR_VERSION, 2786 MANA_MICRO_VERSION, &ac->num_ports); 2787 if (err) 2788 goto out; 2789 2790 if (ac->num_ports > MAX_PORTS_IN_MANA_DEV) 2791 ac->num_ports = MAX_PORTS_IN_MANA_DEV; 2792 2793 for (i = 0; i < ac->num_ports; i++) { 2794 err = mana_probe_port(ac, i, &ac->ports[i]); 2795 if (err) { 2796 device_printf(dev, 2797 "Failed to probe mana port %d\n", i); 2798 break; 2799 } 2800 } 2801 2802 out: 2803 if (err) 2804 mana_remove(gd); 2805 2806 return err; 2807 } 2808 2809 void 2810 mana_remove(struct gdma_dev *gd) 2811 { 2812 struct gdma_context *gc = gd->gdma_context; 2813 struct mana_context *ac = gd->driver_data; 2814 device_t dev = gc->dev; 2815 struct ifnet *ndev; 2816 int i; 2817 2818 for (i = 0; i < ac->num_ports; i++) { 2819 ndev = ac->ports[i]; 2820 if (!ndev) { 2821 if (i == 0) 2822 device_printf(dev, "No net device to remove\n"); 2823 goto out; 2824 } 2825 2826 mana_detach(ndev); 2827 2828 if_free(ndev); 2829 } 2830 2831 mana_destroy_eq(ac); 2832 2833 out: 2834 mana_gd_deregister_device(gd); 2835 gd->driver_data = NULL; 2836 gd->gdma_context = NULL; 2837 free(ac, M_DEVBUF); 2838 } 2839