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 31 #include <sys/param.h> 32 #include <sys/systm.h> 33 #include <sys/bus.h> 34 #include <sys/kernel.h> 35 #include <sys/kthread.h> 36 #include <sys/malloc.h> 37 #include <sys/mbuf.h> 38 #include <sys/smp.h> 39 #include <sys/socket.h> 40 #include <sys/sockio.h> 41 #include <sys/time.h> 42 #include <sys/eventhandler.h> 43 44 #include <machine/bus.h> 45 #include <machine/resource.h> 46 #include <machine/in_cksum.h> 47 48 #include <net/if.h> 49 #include <net/if_var.h> 50 #include <net/if_types.h> 51 #include <net/if_vlan_var.h> 52 #ifdef RSS 53 #include <net/rss_config.h> 54 #endif 55 56 #include <netinet/in_systm.h> 57 #include <netinet/in.h> 58 #include <netinet/if_ether.h> 59 #include <netinet/ip.h> 60 #include <netinet/ip6.h> 61 #include <netinet/tcp.h> 62 #include <netinet/udp.h> 63 64 #include "mana.h" 65 #include "mana_sysctl.h" 66 67 static int mana_up(struct mana_port_context *apc); 68 static int mana_down(struct mana_port_context *apc); 69 70 static void 71 mana_rss_key_fill(void *k, size_t size) 72 { 73 static bool rss_key_generated = false; 74 static uint8_t rss_key[MANA_HASH_KEY_SIZE]; 75 76 KASSERT(size <= MANA_HASH_KEY_SIZE, 77 ("Request more buytes than MANA RSS key can hold")); 78 79 if (!rss_key_generated) { 80 arc4random_buf(rss_key, MANA_HASH_KEY_SIZE); 81 rss_key_generated = true; 82 } 83 memcpy(k, rss_key, size); 84 } 85 86 static int 87 mana_ifmedia_change(if_t ifp __unused) 88 { 89 return EOPNOTSUPP; 90 } 91 92 static void 93 mana_ifmedia_status(if_t ifp, struct ifmediareq *ifmr) 94 { 95 struct mana_port_context *apc = if_getsoftc(ifp); 96 97 if (!apc) { 98 if_printf(ifp, "Port not available\n"); 99 return; 100 } 101 102 MANA_APC_LOCK_LOCK(apc); 103 104 ifmr->ifm_status = IFM_AVALID; 105 ifmr->ifm_active = IFM_ETHER; 106 107 if (!apc->port_is_up) { 108 MANA_APC_LOCK_UNLOCK(apc); 109 mana_dbg(NULL, "Port %u link is down\n", apc->port_idx); 110 return; 111 } 112 113 ifmr->ifm_status |= IFM_ACTIVE; 114 ifmr->ifm_active |= IFM_100G_DR | IFM_FDX; 115 116 MANA_APC_LOCK_UNLOCK(apc); 117 } 118 119 static uint64_t 120 mana_get_counter(if_t ifp, ift_counter cnt) 121 { 122 struct mana_port_context *apc = if_getsoftc(ifp); 123 struct mana_port_stats *stats = &apc->port_stats; 124 125 switch (cnt) { 126 case IFCOUNTER_IPACKETS: 127 return (counter_u64_fetch(stats->rx_packets)); 128 case IFCOUNTER_OPACKETS: 129 return (counter_u64_fetch(stats->tx_packets)); 130 case IFCOUNTER_IBYTES: 131 return (counter_u64_fetch(stats->rx_bytes)); 132 case IFCOUNTER_OBYTES: 133 return (counter_u64_fetch(stats->tx_bytes)); 134 case IFCOUNTER_IQDROPS: 135 return (counter_u64_fetch(stats->rx_drops)); 136 case IFCOUNTER_OQDROPS: 137 return (counter_u64_fetch(stats->tx_drops)); 138 default: 139 return (if_get_counter_default(ifp, cnt)); 140 } 141 } 142 143 static void 144 mana_qflush(if_t ifp) 145 { 146 if_qflush(ifp); 147 } 148 149 int 150 mana_restart(struct mana_port_context *apc) 151 { 152 int rc = 0; 153 154 MANA_APC_LOCK_LOCK(apc); 155 if (apc->port_is_up) 156 mana_down(apc); 157 158 rc = mana_up(apc); 159 MANA_APC_LOCK_UNLOCK(apc); 160 161 return (rc); 162 } 163 164 static int 165 mana_ioctl(if_t ifp, u_long command, caddr_t data) 166 { 167 struct mana_port_context *apc = if_getsoftc(ifp); 168 struct ifrsskey *ifrk; 169 struct ifrsshash *ifrh; 170 struct ifreq *ifr; 171 uint16_t new_mtu; 172 int rc = 0, mask; 173 174 switch (command) { 175 case SIOCSIFMTU: 176 ifr = (struct ifreq *)data; 177 new_mtu = ifr->ifr_mtu; 178 if (if_getmtu(ifp) == new_mtu) 179 break; 180 if ((new_mtu + 18 > MAX_FRAME_SIZE) || 181 (new_mtu + 18 < MIN_FRAME_SIZE)) { 182 if_printf(ifp, "Invalid MTU. new_mtu: %d, " 183 "max allowed: %d, min allowed: %d\n", 184 new_mtu, MAX_FRAME_SIZE - 18, MIN_FRAME_SIZE - 18); 185 return EINVAL; 186 } 187 MANA_APC_LOCK_LOCK(apc); 188 if (apc->port_is_up) 189 mana_down(apc); 190 191 apc->frame_size = new_mtu + 18; 192 if_setmtu(ifp, new_mtu); 193 mana_dbg(NULL, "Set MTU to %d\n", new_mtu); 194 195 rc = mana_up(apc); 196 MANA_APC_LOCK_UNLOCK(apc); 197 break; 198 199 case SIOCSIFFLAGS: 200 if (if_getflags(ifp) & IFF_UP) { 201 if ((if_getdrvflags(ifp) & IFF_DRV_RUNNING) == 0) { 202 MANA_APC_LOCK_LOCK(apc); 203 if (!apc->port_is_up) 204 rc = mana_up(apc); 205 MANA_APC_LOCK_UNLOCK(apc); 206 } 207 } else { 208 if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) { 209 MANA_APC_LOCK_LOCK(apc); 210 if (apc->port_is_up) 211 mana_down(apc); 212 MANA_APC_LOCK_UNLOCK(apc); 213 } 214 } 215 break; 216 217 case SIOCSIFCAP: 218 MANA_APC_LOCK_LOCK(apc); 219 ifr = (struct ifreq *)data; 220 /* 221 * Fix up requested capabilities w/ supported capabilities, 222 * since the supported capabilities could have been changed. 223 */ 224 mask = (ifr->ifr_reqcap & if_getcapabilities(ifp)) ^ 225 if_getcapenable(ifp); 226 227 if (mask & IFCAP_TXCSUM) { 228 if_togglecapenable(ifp, IFCAP_TXCSUM); 229 if_togglehwassist(ifp, (CSUM_TCP | CSUM_UDP | CSUM_IP)); 230 231 if ((IFCAP_TSO4 & if_getcapenable(ifp)) && 232 !(IFCAP_TXCSUM & if_getcapenable(ifp))) { 233 mask &= ~IFCAP_TSO4; 234 if_setcapenablebit(ifp, 0, IFCAP_TSO4); 235 if_sethwassistbits(ifp, 0, CSUM_IP_TSO); 236 mana_warn(NULL, 237 "Also disabled tso4 due to -txcsum.\n"); 238 } 239 } 240 241 if (mask & IFCAP_TXCSUM_IPV6) { 242 if_togglecapenable(ifp, IFCAP_TXCSUM_IPV6); 243 if_togglehwassist(ifp, (CSUM_UDP_IPV6 | CSUM_TCP_IPV6)); 244 245 if ((IFCAP_TSO6 & if_getcapenable(ifp)) && 246 !(IFCAP_TXCSUM_IPV6 & if_getcapenable(ifp))) { 247 mask &= ~IFCAP_TSO6; 248 if_setcapenablebit(ifp, 0, IFCAP_TSO6); 249 if_sethwassistbits(ifp, 0, CSUM_IP6_TSO); 250 mana_warn(ifp, 251 "Also disabled tso6 due to -txcsum6.\n"); 252 } 253 } 254 255 if (mask & IFCAP_RXCSUM) 256 if_togglecapenable(ifp, IFCAP_RXCSUM); 257 /* We can't diff IPv6 packets from IPv4 packets on RX path. */ 258 if (mask & IFCAP_RXCSUM_IPV6) 259 if_togglecapenable(ifp, IFCAP_RXCSUM_IPV6); 260 261 if (mask & IFCAP_LRO) 262 if_togglecapenable(ifp, IFCAP_LRO); 263 264 if (mask & IFCAP_TSO4) { 265 if (!(IFCAP_TSO4 & if_getcapenable(ifp)) && 266 !(IFCAP_TXCSUM & if_getcapenable(ifp))) { 267 MANA_APC_LOCK_UNLOCK(apc); 268 if_printf(ifp, "Enable txcsum first.\n"); 269 rc = EAGAIN; 270 goto out; 271 } 272 if_togglecapenable(ifp, IFCAP_TSO4); 273 if_togglehwassist(ifp, CSUM_IP_TSO); 274 } 275 276 if (mask & IFCAP_TSO6) { 277 if (!(IFCAP_TSO6 & if_getcapenable(ifp)) && 278 !(IFCAP_TXCSUM_IPV6 & if_getcapenable(ifp))) { 279 MANA_APC_LOCK_UNLOCK(apc); 280 if_printf(ifp, "Enable txcsum6 first.\n"); 281 rc = EAGAIN; 282 goto out; 283 } 284 if_togglecapenable(ifp, IFCAP_TSO6); 285 if_togglehwassist(ifp, CSUM_IP6_TSO); 286 } 287 288 MANA_APC_LOCK_UNLOCK(apc); 289 out: 290 break; 291 292 case SIOCSIFMEDIA: 293 case SIOCGIFMEDIA: 294 case SIOCGIFXMEDIA: 295 ifr = (struct ifreq *)data; 296 rc = ifmedia_ioctl(ifp, ifr, &apc->media, command); 297 break; 298 299 case SIOCGIFRSSKEY: 300 ifrk = (struct ifrsskey *)data; 301 ifrk->ifrk_func = RSS_FUNC_TOEPLITZ; 302 ifrk->ifrk_keylen = MANA_HASH_KEY_SIZE; 303 memcpy(ifrk->ifrk_key, apc->hashkey, MANA_HASH_KEY_SIZE); 304 break; 305 306 case SIOCGIFRSSHASH: 307 ifrh = (struct ifrsshash *)data; 308 ifrh->ifrh_func = RSS_FUNC_TOEPLITZ; 309 ifrh->ifrh_types = 310 RSS_TYPE_TCP_IPV4 | 311 RSS_TYPE_UDP_IPV4 | 312 RSS_TYPE_TCP_IPV6 | 313 RSS_TYPE_UDP_IPV6; 314 break; 315 316 default: 317 rc = ether_ioctl(ifp, command, data); 318 break; 319 } 320 321 return (rc); 322 } 323 324 static inline void 325 mana_alloc_counters(counter_u64_t *begin, int size) 326 { 327 counter_u64_t *end = (counter_u64_t *)((char *)begin + size); 328 329 for (; begin < end; ++begin) 330 *begin = counter_u64_alloc(M_WAITOK); 331 } 332 333 static inline void 334 mana_free_counters(counter_u64_t *begin, int size) 335 { 336 counter_u64_t *end = (counter_u64_t *)((char *)begin + size); 337 338 for (; begin < end; ++begin) 339 counter_u64_free(*begin); 340 } 341 342 static bool 343 mana_can_tx(struct gdma_queue *wq) 344 { 345 return mana_gd_wq_avail_space(wq) >= MAX_TX_WQE_SIZE; 346 } 347 348 static inline int 349 mana_tx_map_mbuf(struct mana_port_context *apc, 350 struct mana_send_buf_info *tx_info, 351 struct mbuf **m_head, struct mana_tx_package *tp, 352 struct mana_stats *tx_stats) 353 { 354 struct gdma_dev *gd = apc->ac->gdma_dev; 355 bus_dma_segment_t segs[MAX_MBUF_FRAGS]; 356 struct mbuf *m = *m_head; 357 int err, nsegs, i; 358 359 err = bus_dmamap_load_mbuf_sg(apc->tx_buf_tag, tx_info->dma_map, 360 m, segs, &nsegs, BUS_DMA_NOWAIT); 361 if (err == EFBIG) { 362 struct mbuf *m_new; 363 364 counter_u64_add(tx_stats->collapse, 1); 365 m_new = m_collapse(m, M_NOWAIT, MAX_MBUF_FRAGS); 366 if (unlikely(m_new == NULL)) { 367 counter_u64_add(tx_stats->collapse_err, 1); 368 return ENOBUFS; 369 } else { 370 *m_head = m = m_new; 371 } 372 373 mana_warn(NULL, 374 "Too many segs in orig mbuf, m_collapse called\n"); 375 376 err = bus_dmamap_load_mbuf_sg(apc->tx_buf_tag, 377 tx_info->dma_map, m, segs, &nsegs, BUS_DMA_NOWAIT); 378 } 379 if (!err) { 380 for (i = 0; i < nsegs; i++) { 381 tp->wqe_req.sgl[i].address = segs[i].ds_addr; 382 tp->wqe_req.sgl[i].mem_key = gd->gpa_mkey; 383 tp->wqe_req.sgl[i].size = segs[i].ds_len; 384 } 385 tp->wqe_req.num_sge = nsegs; 386 387 tx_info->mbuf = *m_head; 388 389 bus_dmamap_sync(apc->tx_buf_tag, tx_info->dma_map, 390 BUS_DMASYNC_PREWRITE); 391 } 392 393 return err; 394 } 395 396 static inline void 397 mana_tx_unmap_mbuf(struct mana_port_context *apc, 398 struct mana_send_buf_info *tx_info) 399 { 400 bus_dmamap_sync(apc->tx_buf_tag, tx_info->dma_map, 401 BUS_DMASYNC_POSTWRITE); 402 bus_dmamap_unload(apc->tx_buf_tag, tx_info->dma_map); 403 if (tx_info->mbuf) { 404 m_freem(tx_info->mbuf); 405 tx_info->mbuf = NULL; 406 } 407 } 408 409 static inline int 410 mana_load_rx_mbuf(struct mana_port_context *apc, struct mana_rxq *rxq, 411 struct mana_recv_buf_oob *rx_oob, bool alloc_mbuf) 412 { 413 bus_dma_segment_t segs[1]; 414 struct mbuf *mbuf; 415 int nsegs, err; 416 uint32_t mlen; 417 418 if (alloc_mbuf) { 419 mbuf = m_getjcl(M_NOWAIT, MT_DATA, M_PKTHDR, rxq->datasize); 420 if (unlikely(mbuf == NULL)) { 421 mbuf = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR); 422 if (unlikely(mbuf == NULL)) { 423 return ENOMEM; 424 } 425 mlen = MCLBYTES; 426 } else { 427 mlen = rxq->datasize; 428 } 429 430 mbuf->m_pkthdr.len = mbuf->m_len = mlen; 431 } else { 432 if (rx_oob->mbuf) { 433 mbuf = rx_oob->mbuf; 434 mlen = rx_oob->mbuf->m_pkthdr.len; 435 } else { 436 return ENOMEM; 437 } 438 } 439 440 err = bus_dmamap_load_mbuf_sg(apc->rx_buf_tag, rx_oob->dma_map, 441 mbuf, segs, &nsegs, BUS_DMA_NOWAIT); 442 443 if (unlikely((err != 0) || (nsegs != 1))) { 444 mana_warn(NULL, "Failed to map mbuf, error: %d, " 445 "nsegs: %d\n", err, nsegs); 446 counter_u64_add(rxq->stats.dma_mapping_err, 1); 447 goto error; 448 } 449 450 bus_dmamap_sync(apc->rx_buf_tag, rx_oob->dma_map, 451 BUS_DMASYNC_PREREAD); 452 453 rx_oob->mbuf = mbuf; 454 rx_oob->num_sge = 1; 455 rx_oob->sgl[0].address = segs[0].ds_addr; 456 rx_oob->sgl[0].size = mlen; 457 rx_oob->sgl[0].mem_key = apc->ac->gdma_dev->gpa_mkey; 458 459 return 0; 460 461 error: 462 m_freem(mbuf); 463 return EFAULT; 464 } 465 466 static inline void 467 mana_unload_rx_mbuf(struct mana_port_context *apc, struct mana_rxq *rxq, 468 struct mana_recv_buf_oob *rx_oob, bool free_mbuf) 469 { 470 bus_dmamap_sync(apc->rx_buf_tag, rx_oob->dma_map, 471 BUS_DMASYNC_POSTREAD); 472 bus_dmamap_unload(apc->rx_buf_tag, rx_oob->dma_map); 473 474 if (free_mbuf && rx_oob->mbuf) { 475 m_freem(rx_oob->mbuf); 476 rx_oob->mbuf = NULL; 477 } 478 } 479 480 481 /* Use couple mbuf PH_loc spaces for l3 and l4 protocal type */ 482 #define MANA_L3_PROTO(_mbuf) ((_mbuf)->m_pkthdr.PH_loc.sixteen[0]) 483 #define MANA_L4_PROTO(_mbuf) ((_mbuf)->m_pkthdr.PH_loc.sixteen[1]) 484 485 #define MANA_TXQ_FULL (IFF_DRV_RUNNING | IFF_DRV_OACTIVE) 486 487 static void 488 mana_xmit(struct mana_txq *txq) 489 { 490 enum mana_tx_pkt_format pkt_fmt = MANA_SHORT_PKT_FMT; 491 struct mana_send_buf_info *tx_info; 492 if_t ndev = txq->ndev; 493 struct mbuf *mbuf; 494 struct mana_port_context *apc = if_getsoftc(ndev); 495 struct mana_port_stats *port_stats = &apc->port_stats; 496 struct gdma_dev *gd = apc->ac->gdma_dev; 497 uint64_t packets, bytes; 498 uint16_t next_to_use; 499 struct mana_tx_package pkg = {}; 500 struct mana_stats *tx_stats; 501 struct gdma_queue *gdma_sq; 502 struct mana_cq *cq; 503 int err, len; 504 bool is_tso; 505 506 gdma_sq = txq->gdma_sq; 507 cq = &apc->tx_qp[txq->idx].tx_cq; 508 tx_stats = &txq->stats; 509 510 packets = 0; 511 bytes = 0; 512 next_to_use = txq->next_to_use; 513 514 while ((mbuf = drbr_peek(ndev, txq->txq_br)) != NULL) { 515 if (!apc->port_is_up || 516 (if_getdrvflags(ndev) & MANA_TXQ_FULL) != IFF_DRV_RUNNING) { 517 drbr_putback(ndev, txq->txq_br, mbuf); 518 break; 519 } 520 521 if (!mana_can_tx(gdma_sq)) { 522 /* SQ is full. Set the IFF_DRV_OACTIVE flag */ 523 if_setdrvflagbits(apc->ndev, IFF_DRV_OACTIVE, 0); 524 counter_u64_add(tx_stats->stop, 1); 525 uint64_t stops = counter_u64_fetch(tx_stats->stop); 526 uint64_t wakeups = counter_u64_fetch(tx_stats->wakeup); 527 #define MANA_TXQ_STOP_THRESHOLD 50 528 if (stops > MANA_TXQ_STOP_THRESHOLD && wakeups > 0 && 529 stops > wakeups && txq->alt_txq_idx == txq->idx) { 530 txq->alt_txq_idx = 531 (txq->idx + (stops / wakeups)) 532 % apc->num_queues; 533 counter_u64_add(tx_stats->alt_chg, 1); 534 } 535 536 drbr_putback(ndev, txq->txq_br, mbuf); 537 538 taskqueue_enqueue(cq->cleanup_tq, &cq->cleanup_task); 539 break; 540 } 541 542 tx_info = &txq->tx_buf_info[next_to_use]; 543 544 memset(&pkg, 0, sizeof(struct mana_tx_package)); 545 pkg.wqe_req.sgl = pkg.sgl_array; 546 547 err = mana_tx_map_mbuf(apc, tx_info, &mbuf, &pkg, tx_stats); 548 if (unlikely(err)) { 549 mana_dbg(NULL, 550 "Failed to map tx mbuf, err %d\n", err); 551 552 counter_u64_add(tx_stats->dma_mapping_err, 1); 553 554 /* The mbuf is still there. Free it */ 555 m_freem(mbuf); 556 /* Advance the drbr queue */ 557 drbr_advance(ndev, txq->txq_br); 558 continue; 559 } 560 561 pkg.tx_oob.s_oob.vcq_num = cq->gdma_id; 562 pkg.tx_oob.s_oob.vsq_frame = txq->vsq_frame; 563 564 if (txq->vp_offset > MANA_SHORT_VPORT_OFFSET_MAX) { 565 pkg.tx_oob.l_oob.long_vp_offset = txq->vp_offset; 566 pkt_fmt = MANA_LONG_PKT_FMT; 567 } else { 568 pkg.tx_oob.s_oob.short_vp_offset = txq->vp_offset; 569 } 570 571 pkg.tx_oob.s_oob.pkt_fmt = pkt_fmt; 572 573 if (pkt_fmt == MANA_SHORT_PKT_FMT) 574 pkg.wqe_req.inline_oob_size = sizeof(struct mana_tx_short_oob); 575 else 576 pkg.wqe_req.inline_oob_size = sizeof(struct mana_tx_oob); 577 578 pkg.wqe_req.inline_oob_data = &pkg.tx_oob; 579 pkg.wqe_req.flags = 0; 580 pkg.wqe_req.client_data_unit = 0; 581 582 is_tso = false; 583 if (mbuf->m_pkthdr.csum_flags & CSUM_TSO) { 584 is_tso = true; 585 586 if (MANA_L3_PROTO(mbuf) == ETHERTYPE_IP) 587 pkg.tx_oob.s_oob.is_outer_ipv4 = 1; 588 else 589 pkg.tx_oob.s_oob.is_outer_ipv6 = 1; 590 591 pkg.tx_oob.s_oob.comp_iphdr_csum = 1; 592 pkg.tx_oob.s_oob.comp_tcp_csum = 1; 593 pkg.tx_oob.s_oob.trans_off = mbuf->m_pkthdr.l3hlen; 594 595 pkg.wqe_req.client_data_unit = mbuf->m_pkthdr.tso_segsz; 596 pkg.wqe_req.flags = GDMA_WR_OOB_IN_SGL | GDMA_WR_PAD_BY_SGE0; 597 } else if (mbuf->m_pkthdr.csum_flags & 598 (CSUM_IP_UDP | CSUM_IP_TCP | CSUM_IP6_UDP | CSUM_IP6_TCP)) { 599 if (MANA_L3_PROTO(mbuf) == ETHERTYPE_IP) { 600 pkg.tx_oob.s_oob.is_outer_ipv4 = 1; 601 pkg.tx_oob.s_oob.comp_iphdr_csum = 1; 602 } else { 603 pkg.tx_oob.s_oob.is_outer_ipv6 = 1; 604 } 605 606 if (MANA_L4_PROTO(mbuf) == IPPROTO_TCP) { 607 pkg.tx_oob.s_oob.comp_tcp_csum = 1; 608 pkg.tx_oob.s_oob.trans_off = 609 mbuf->m_pkthdr.l3hlen; 610 } else { 611 pkg.tx_oob.s_oob.comp_udp_csum = 1; 612 } 613 } else if (mbuf->m_pkthdr.csum_flags & CSUM_IP) { 614 pkg.tx_oob.s_oob.is_outer_ipv4 = 1; 615 pkg.tx_oob.s_oob.comp_iphdr_csum = 1; 616 } else { 617 if (MANA_L3_PROTO(mbuf) == ETHERTYPE_IP) 618 pkg.tx_oob.s_oob.is_outer_ipv4 = 1; 619 else if (MANA_L3_PROTO(mbuf) == ETHERTYPE_IPV6) 620 pkg.tx_oob.s_oob.is_outer_ipv6 = 1; 621 } 622 623 len = mbuf->m_pkthdr.len; 624 625 err = mana_gd_post_work_request(gdma_sq, &pkg.wqe_req, 626 (struct gdma_posted_wqe_info *)&tx_info->wqe_inf); 627 if (unlikely(err)) { 628 /* Should not happen */ 629 if_printf(ndev, "Failed to post TX OOB: %d\n", err); 630 631 mana_tx_unmap_mbuf(apc, tx_info); 632 633 drbr_advance(ndev, txq->txq_br); 634 continue; 635 } 636 637 next_to_use = 638 (next_to_use + 1) % MAX_SEND_BUFFERS_PER_QUEUE; 639 640 (void)atomic_inc_return(&txq->pending_sends); 641 642 drbr_advance(ndev, txq->txq_br); 643 644 mana_gd_wq_ring_doorbell(gd->gdma_context, gdma_sq); 645 646 packets++; 647 bytes += len; 648 649 if (is_tso) { 650 txq->tso_pkts++; 651 txq->tso_bytes += len; 652 } 653 } 654 655 counter_enter(); 656 counter_u64_add_protected(tx_stats->packets, packets); 657 counter_u64_add_protected(port_stats->tx_packets, packets); 658 counter_u64_add_protected(tx_stats->bytes, bytes); 659 counter_u64_add_protected(port_stats->tx_bytes, bytes); 660 counter_exit(); 661 662 txq->next_to_use = next_to_use; 663 } 664 665 static void 666 mana_xmit_taskfunc(void *arg, int pending) 667 { 668 struct mana_txq *txq = (struct mana_txq *)arg; 669 if_t ndev = txq->ndev; 670 struct mana_port_context *apc = if_getsoftc(ndev); 671 672 while (!drbr_empty(ndev, txq->txq_br) && apc->port_is_up && 673 (if_getdrvflags(ndev) & MANA_TXQ_FULL) == IFF_DRV_RUNNING) { 674 mtx_lock(&txq->txq_mtx); 675 mana_xmit(txq); 676 mtx_unlock(&txq->txq_mtx); 677 } 678 } 679 680 #define PULLUP_HDR(m, len) \ 681 do { \ 682 if (unlikely((m)->m_len < (len))) { \ 683 (m) = m_pullup((m), (len)); \ 684 if ((m) == NULL) \ 685 return (NULL); \ 686 } \ 687 } while (0) 688 689 /* 690 * If this function failed, the mbuf would be freed. 691 */ 692 static inline struct mbuf * 693 mana_tso_fixup(struct mbuf *mbuf) 694 { 695 struct ether_vlan_header *eh = mtod(mbuf, struct ether_vlan_header *); 696 struct tcphdr *th; 697 uint16_t etype; 698 int ehlen; 699 700 if (eh->evl_encap_proto == ntohs(ETHERTYPE_VLAN)) { 701 etype = ntohs(eh->evl_proto); 702 ehlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN; 703 } else { 704 etype = ntohs(eh->evl_encap_proto); 705 ehlen = ETHER_HDR_LEN; 706 } 707 708 if (etype == ETHERTYPE_IP) { 709 struct ip *ip; 710 int iphlen; 711 712 PULLUP_HDR(mbuf, ehlen + sizeof(*ip)); 713 ip = mtodo(mbuf, ehlen); 714 iphlen = ip->ip_hl << 2; 715 mbuf->m_pkthdr.l3hlen = ehlen + iphlen; 716 717 PULLUP_HDR(mbuf, ehlen + iphlen + sizeof(*th)); 718 th = mtodo(mbuf, ehlen + iphlen); 719 720 ip->ip_len = 0; 721 ip->ip_sum = 0; 722 th->th_sum = in_pseudo(ip->ip_src.s_addr, 723 ip->ip_dst.s_addr, htons(IPPROTO_TCP)); 724 } else if (etype == ETHERTYPE_IPV6) { 725 struct ip6_hdr *ip6; 726 727 PULLUP_HDR(mbuf, ehlen + sizeof(*ip6) + sizeof(*th)); 728 ip6 = mtodo(mbuf, ehlen); 729 if (ip6->ip6_nxt != IPPROTO_TCP) { 730 /* Realy something wrong, just return */ 731 mana_dbg(NULL, "TSO mbuf not TCP, freed.\n"); 732 m_freem(mbuf); 733 return NULL; 734 } 735 mbuf->m_pkthdr.l3hlen = ehlen + sizeof(*ip6); 736 737 th = mtodo(mbuf, ehlen + sizeof(*ip6)); 738 739 ip6->ip6_plen = 0; 740 th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0); 741 } else { 742 /* CSUM_TSO is set but not IP protocol. */ 743 mana_warn(NULL, "TSO mbuf not right, freed.\n"); 744 m_freem(mbuf); 745 return NULL; 746 } 747 748 MANA_L3_PROTO(mbuf) = etype; 749 750 return (mbuf); 751 } 752 753 /* 754 * If this function failed, the mbuf would be freed. 755 */ 756 static inline struct mbuf * 757 mana_mbuf_csum_check(struct mbuf *mbuf) 758 { 759 struct ether_vlan_header *eh = mtod(mbuf, struct ether_vlan_header *); 760 struct mbuf *mbuf_next; 761 uint16_t etype; 762 int offset; 763 int ehlen; 764 765 if (eh->evl_encap_proto == ntohs(ETHERTYPE_VLAN)) { 766 etype = ntohs(eh->evl_proto); 767 ehlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN; 768 } else { 769 etype = ntohs(eh->evl_encap_proto); 770 ehlen = ETHER_HDR_LEN; 771 } 772 773 mbuf_next = m_getptr(mbuf, ehlen, &offset); 774 775 MANA_L4_PROTO(mbuf) = 0; 776 if (etype == ETHERTYPE_IP) { 777 const struct ip *ip; 778 int iphlen; 779 780 ip = (struct ip *)(mtodo(mbuf_next, offset)); 781 iphlen = ip->ip_hl << 2; 782 mbuf->m_pkthdr.l3hlen = ehlen + iphlen; 783 784 MANA_L4_PROTO(mbuf) = ip->ip_p; 785 } else if (etype == ETHERTYPE_IPV6) { 786 const struct ip6_hdr *ip6; 787 788 ip6 = (struct ip6_hdr *)(mtodo(mbuf_next, offset)); 789 mbuf->m_pkthdr.l3hlen = ehlen + sizeof(*ip6); 790 791 MANA_L4_PROTO(mbuf) = ip6->ip6_nxt; 792 } else { 793 MANA_L4_PROTO(mbuf) = 0; 794 } 795 796 MANA_L3_PROTO(mbuf) = etype; 797 798 return (mbuf); 799 } 800 801 static int 802 mana_start_xmit(if_t ifp, struct mbuf *m) 803 { 804 struct mana_port_context *apc = if_getsoftc(ifp); 805 struct mana_txq *txq; 806 int is_drbr_empty; 807 uint16_t txq_id; 808 int err; 809 810 if (unlikely((!apc->port_is_up) || 811 (if_getdrvflags(ifp) & IFF_DRV_RUNNING) == 0)) 812 return ENODEV; 813 814 if (m->m_pkthdr.csum_flags & CSUM_TSO) { 815 m = mana_tso_fixup(m); 816 if (unlikely(m == NULL)) { 817 counter_enter(); 818 counter_u64_add_protected(apc->port_stats.tx_drops, 1); 819 counter_exit(); 820 return EIO; 821 } 822 } else { 823 m = mana_mbuf_csum_check(m); 824 if (unlikely(m == NULL)) { 825 counter_enter(); 826 counter_u64_add_protected(apc->port_stats.tx_drops, 1); 827 counter_exit(); 828 return EIO; 829 } 830 } 831 832 if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) { 833 uint32_t hash = m->m_pkthdr.flowid; 834 txq_id = apc->indir_table[(hash) & MANA_INDIRECT_TABLE_MASK] % 835 apc->num_queues; 836 } else { 837 txq_id = m->m_pkthdr.flowid % apc->num_queues; 838 } 839 840 if (apc->enable_tx_altq) 841 txq_id = apc->tx_qp[txq_id].txq.alt_txq_idx; 842 843 txq = &apc->tx_qp[txq_id].txq; 844 845 is_drbr_empty = drbr_empty(ifp, txq->txq_br); 846 err = drbr_enqueue(ifp, txq->txq_br, m); 847 if (unlikely(err)) { 848 mana_warn(NULL, "txq %u failed to enqueue: %d\n", 849 txq_id, err); 850 taskqueue_enqueue(txq->enqueue_tq, &txq->enqueue_task); 851 return err; 852 } 853 854 if (is_drbr_empty && mtx_trylock(&txq->txq_mtx)) { 855 mana_xmit(txq); 856 mtx_unlock(&txq->txq_mtx); 857 } else { 858 taskqueue_enqueue(txq->enqueue_tq, &txq->enqueue_task); 859 } 860 861 return 0; 862 } 863 864 static void 865 mana_cleanup_port_context(struct mana_port_context *apc) 866 { 867 bus_dma_tag_destroy(apc->tx_buf_tag); 868 bus_dma_tag_destroy(apc->rx_buf_tag); 869 apc->rx_buf_tag = NULL; 870 871 free(apc->rxqs, M_DEVBUF); 872 apc->rxqs = NULL; 873 874 mana_free_counters((counter_u64_t *)&apc->port_stats, 875 sizeof(struct mana_port_stats)); 876 } 877 878 static int 879 mana_init_port_context(struct mana_port_context *apc) 880 { 881 device_t dev = apc->ac->gdma_dev->gdma_context->dev; 882 uint32_t tso_maxsize; 883 int err; 884 885 tso_maxsize = MANA_TSO_MAX_SZ; 886 887 /* Create DMA tag for tx bufs */ 888 err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */ 889 1, 0, /* alignment, boundary */ 890 BUS_SPACE_MAXADDR, /* lowaddr */ 891 BUS_SPACE_MAXADDR, /* highaddr */ 892 NULL, NULL, /* filter, filterarg */ 893 tso_maxsize, /* maxsize */ 894 MAX_MBUF_FRAGS, /* nsegments */ 895 tso_maxsize, /* maxsegsize */ 896 0, /* flags */ 897 NULL, NULL, /* lockfunc, lockfuncarg*/ 898 &apc->tx_buf_tag); 899 if (unlikely(err)) { 900 device_printf(dev, "Feiled to create TX DMA tag\n"); 901 return err; 902 } 903 904 /* Create DMA tag for rx bufs */ 905 err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */ 906 64, 0, /* alignment, boundary */ 907 BUS_SPACE_MAXADDR, /* lowaddr */ 908 BUS_SPACE_MAXADDR, /* highaddr */ 909 NULL, NULL, /* filter, filterarg */ 910 MJUMPAGESIZE, /* maxsize */ 911 1, /* nsegments */ 912 MJUMPAGESIZE, /* maxsegsize */ 913 0, /* flags */ 914 NULL, NULL, /* lockfunc, lockfuncarg*/ 915 &apc->rx_buf_tag); 916 if (unlikely(err)) { 917 device_printf(dev, "Feiled to create RX DMA tag\n"); 918 return err; 919 } 920 921 apc->rxqs = mallocarray(apc->num_queues, sizeof(struct mana_rxq *), 922 M_DEVBUF, M_WAITOK | M_ZERO); 923 924 return 0; 925 } 926 927 static int 928 mana_send_request(struct mana_context *ac, void *in_buf, 929 uint32_t in_len, void *out_buf, uint32_t out_len) 930 { 931 struct gdma_context *gc = ac->gdma_dev->gdma_context; 932 struct gdma_resp_hdr *resp = out_buf; 933 struct gdma_req_hdr *req = in_buf; 934 device_t dev = gc->dev; 935 static atomic_t activity_id; 936 int err; 937 938 req->dev_id = gc->mana.dev_id; 939 req->activity_id = atomic_inc_return(&activity_id); 940 941 mana_dbg(NULL, "activity_id = %u\n", activity_id); 942 943 err = mana_gd_send_request(gc, in_len, in_buf, out_len, 944 out_buf); 945 if (err || resp->status) { 946 device_printf(dev, "Failed to send mana message: %d, 0x%x\n", 947 err, resp->status); 948 return err ? err : EPROTO; 949 } 950 951 if (req->dev_id.as_uint32 != resp->dev_id.as_uint32 || 952 req->activity_id != resp->activity_id) { 953 device_printf(dev, 954 "Unexpected mana message response: %x,%x,%x,%x\n", 955 req->dev_id.as_uint32, resp->dev_id.as_uint32, 956 req->activity_id, resp->activity_id); 957 return EPROTO; 958 } 959 960 return 0; 961 } 962 963 static int 964 mana_verify_resp_hdr(const struct gdma_resp_hdr *resp_hdr, 965 const enum mana_command_code expected_code, 966 const uint32_t min_size) 967 { 968 if (resp_hdr->response.msg_type != expected_code) 969 return EPROTO; 970 971 if (resp_hdr->response.msg_version < GDMA_MESSAGE_V1) 972 return EPROTO; 973 974 if (resp_hdr->response.msg_size < min_size) 975 return EPROTO; 976 977 return 0; 978 } 979 980 static int 981 mana_query_device_cfg(struct mana_context *ac, uint32_t proto_major_ver, 982 uint32_t proto_minor_ver, uint32_t proto_micro_ver, 983 uint16_t *max_num_vports) 984 { 985 struct gdma_context *gc = ac->gdma_dev->gdma_context; 986 struct mana_query_device_cfg_resp resp = {}; 987 struct mana_query_device_cfg_req req = {}; 988 device_t dev = gc->dev; 989 int err = 0; 990 991 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_DEV_CONFIG, 992 sizeof(req), sizeof(resp)); 993 req.proto_major_ver = proto_major_ver; 994 req.proto_minor_ver = proto_minor_ver; 995 req.proto_micro_ver = proto_micro_ver; 996 997 err = mana_send_request(ac, &req, sizeof(req), &resp, sizeof(resp)); 998 if (err) { 999 device_printf(dev, "Failed to query config: %d", err); 1000 return err; 1001 } 1002 1003 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_DEV_CONFIG, 1004 sizeof(resp)); 1005 if (err || resp.hdr.status) { 1006 device_printf(dev, "Invalid query result: %d, 0x%x\n", err, 1007 resp.hdr.status); 1008 if (!err) 1009 err = EPROTO; 1010 return err; 1011 } 1012 1013 *max_num_vports = resp.max_num_vports; 1014 1015 mana_dbg(NULL, "mana max_num_vports from device = %d\n", 1016 *max_num_vports); 1017 1018 return 0; 1019 } 1020 1021 static int 1022 mana_query_vport_cfg(struct mana_port_context *apc, uint32_t vport_index, 1023 uint32_t *max_sq, uint32_t *max_rq, uint32_t *num_indir_entry) 1024 { 1025 struct mana_query_vport_cfg_resp resp = {}; 1026 struct mana_query_vport_cfg_req req = {}; 1027 int err; 1028 1029 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_VPORT_CONFIG, 1030 sizeof(req), sizeof(resp)); 1031 1032 req.vport_index = vport_index; 1033 1034 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 1035 sizeof(resp)); 1036 if (err) 1037 return err; 1038 1039 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_VPORT_CONFIG, 1040 sizeof(resp)); 1041 if (err) 1042 return err; 1043 1044 if (resp.hdr.status) 1045 return EPROTO; 1046 1047 *max_sq = resp.max_num_sq; 1048 *max_rq = resp.max_num_rq; 1049 *num_indir_entry = resp.num_indirection_ent; 1050 1051 apc->port_handle = resp.vport; 1052 memcpy(apc->mac_addr, resp.mac_addr, ETHER_ADDR_LEN); 1053 1054 return 0; 1055 } 1056 1057 void 1058 mana_uncfg_vport(struct mana_port_context *apc) 1059 { 1060 apc->vport_use_count--; 1061 if (apc->vport_use_count < 0) { 1062 mana_err(NULL, 1063 "WARNING: vport_use_count less than 0: %u\n", 1064 apc->vport_use_count); 1065 } 1066 } 1067 1068 int 1069 mana_cfg_vport(struct mana_port_context *apc, uint32_t protection_dom_id, 1070 uint32_t doorbell_pg_id) 1071 { 1072 struct mana_config_vport_resp resp = {}; 1073 struct mana_config_vport_req req = {}; 1074 int err; 1075 1076 /* This function is used to program the Ethernet port in the hardware 1077 * table. It can be called from the Ethernet driver or the RDMA driver. 1078 * 1079 * For Ethernet usage, the hardware supports only one active user on a 1080 * physical port. The driver checks on the port usage before programming 1081 * the hardware when creating the RAW QP (RDMA driver) or exposing the 1082 * device to kernel NET layer (Ethernet driver). 1083 * 1084 * Because the RDMA driver doesn't know in advance which QP type the 1085 * user will create, it exposes the device with all its ports. The user 1086 * may not be able to create RAW QP on a port if this port is already 1087 * in used by the Ethernet driver from the kernel. 1088 * 1089 * This physical port limitation only applies to the RAW QP. For RC QP, 1090 * the hardware doesn't have this limitation. The user can create RC 1091 * QPs on a physical port up to the hardware limits independent of the 1092 * Ethernet usage on the same port. 1093 */ 1094 if (apc->vport_use_count > 0) { 1095 return EBUSY; 1096 } 1097 apc->vport_use_count++; 1098 1099 mana_gd_init_req_hdr(&req.hdr, MANA_CONFIG_VPORT_TX, 1100 sizeof(req), sizeof(resp)); 1101 req.vport = apc->port_handle; 1102 req.pdid = protection_dom_id; 1103 req.doorbell_pageid = doorbell_pg_id; 1104 1105 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 1106 sizeof(resp)); 1107 if (err) { 1108 if_printf(apc->ndev, "Failed to configure vPort: %d\n", err); 1109 goto out; 1110 } 1111 1112 err = mana_verify_resp_hdr(&resp.hdr, MANA_CONFIG_VPORT_TX, 1113 sizeof(resp)); 1114 if (err || resp.hdr.status) { 1115 if_printf(apc->ndev, "Failed to configure vPort: %d, 0x%x\n", 1116 err, resp.hdr.status); 1117 if (!err) 1118 err = EPROTO; 1119 1120 goto out; 1121 } 1122 1123 apc->tx_shortform_allowed = resp.short_form_allowed; 1124 apc->tx_vp_offset = resp.tx_vport_offset; 1125 1126 if_printf(apc->ndev, "Configured vPort %ju PD %u DB %u\n", 1127 apc->port_handle, protection_dom_id, doorbell_pg_id); 1128 1129 out: 1130 if (err) 1131 mana_uncfg_vport(apc); 1132 1133 return err; 1134 } 1135 1136 static int 1137 mana_cfg_vport_steering(struct mana_port_context *apc, 1138 enum TRI_STATE rx, 1139 bool update_default_rxobj, bool update_key, 1140 bool update_tab) 1141 { 1142 uint16_t num_entries = MANA_INDIRECT_TABLE_SIZE; 1143 struct mana_cfg_rx_steer_req *req = NULL; 1144 struct mana_cfg_rx_steer_resp resp = {}; 1145 if_t ndev = apc->ndev; 1146 mana_handle_t *req_indir_tab; 1147 uint32_t req_buf_size; 1148 int err; 1149 1150 req_buf_size = sizeof(*req) + sizeof(mana_handle_t) * num_entries; 1151 req = malloc(req_buf_size, M_DEVBUF, M_WAITOK | M_ZERO); 1152 1153 mana_gd_init_req_hdr(&req->hdr, MANA_CONFIG_VPORT_RX, req_buf_size, 1154 sizeof(resp)); 1155 1156 req->vport = apc->port_handle; 1157 req->num_indir_entries = num_entries; 1158 req->indir_tab_offset = sizeof(*req); 1159 req->rx_enable = rx; 1160 req->rss_enable = apc->rss_state; 1161 req->update_default_rxobj = update_default_rxobj; 1162 req->update_hashkey = update_key; 1163 req->update_indir_tab = update_tab; 1164 req->default_rxobj = apc->default_rxobj; 1165 1166 if (update_key) 1167 memcpy(&req->hashkey, apc->hashkey, MANA_HASH_KEY_SIZE); 1168 1169 if (update_tab) { 1170 req_indir_tab = (mana_handle_t *)(req + 1); 1171 memcpy(req_indir_tab, apc->rxobj_table, 1172 req->num_indir_entries * sizeof(mana_handle_t)); 1173 } 1174 1175 err = mana_send_request(apc->ac, req, req_buf_size, &resp, 1176 sizeof(resp)); 1177 if (err) { 1178 if_printf(ndev, "Failed to configure vPort RX: %d\n", err); 1179 goto out; 1180 } 1181 1182 err = mana_verify_resp_hdr(&resp.hdr, MANA_CONFIG_VPORT_RX, 1183 sizeof(resp)); 1184 if (err) { 1185 if_printf(ndev, "vPort RX configuration failed: %d\n", err); 1186 goto out; 1187 } 1188 1189 if (resp.hdr.status) { 1190 if_printf(ndev, "vPort RX configuration failed: 0x%x\n", 1191 resp.hdr.status); 1192 err = EPROTO; 1193 } 1194 1195 if_printf(ndev, "Configured steering vPort %ju entries %u\n", 1196 apc->port_handle, num_entries); 1197 1198 out: 1199 free(req, M_DEVBUF); 1200 return err; 1201 } 1202 1203 int 1204 mana_create_wq_obj(struct mana_port_context *apc, 1205 mana_handle_t vport, 1206 uint32_t wq_type, struct mana_obj_spec *wq_spec, 1207 struct mana_obj_spec *cq_spec, 1208 mana_handle_t *wq_obj) 1209 { 1210 struct mana_create_wqobj_resp resp = {}; 1211 struct mana_create_wqobj_req req = {}; 1212 if_t ndev = apc->ndev; 1213 int err; 1214 1215 mana_gd_init_req_hdr(&req.hdr, MANA_CREATE_WQ_OBJ, 1216 sizeof(req), sizeof(resp)); 1217 req.vport = vport; 1218 req.wq_type = wq_type; 1219 req.wq_gdma_region = wq_spec->gdma_region; 1220 req.cq_gdma_region = cq_spec->gdma_region; 1221 req.wq_size = wq_spec->queue_size; 1222 req.cq_size = cq_spec->queue_size; 1223 req.cq_moderation_ctx_id = cq_spec->modr_ctx_id; 1224 req.cq_parent_qid = cq_spec->attached_eq; 1225 1226 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 1227 sizeof(resp)); 1228 if (err) { 1229 if_printf(ndev, "Failed to create WQ object: %d\n", err); 1230 goto out; 1231 } 1232 1233 err = mana_verify_resp_hdr(&resp.hdr, MANA_CREATE_WQ_OBJ, 1234 sizeof(resp)); 1235 if (err || resp.hdr.status) { 1236 if_printf(ndev, "Failed to create WQ object: %d, 0x%x\n", err, 1237 resp.hdr.status); 1238 if (!err) 1239 err = EPROTO; 1240 goto out; 1241 } 1242 1243 if (resp.wq_obj == INVALID_MANA_HANDLE) { 1244 if_printf(ndev, "Got an invalid WQ object handle\n"); 1245 err = EPROTO; 1246 goto out; 1247 } 1248 1249 *wq_obj = resp.wq_obj; 1250 wq_spec->queue_index = resp.wq_id; 1251 cq_spec->queue_index = resp.cq_id; 1252 1253 return 0; 1254 out: 1255 return err; 1256 } 1257 1258 void 1259 mana_destroy_wq_obj(struct mana_port_context *apc, uint32_t wq_type, 1260 mana_handle_t wq_obj) 1261 { 1262 struct mana_destroy_wqobj_resp resp = {}; 1263 struct mana_destroy_wqobj_req req = {}; 1264 if_t ndev = apc->ndev; 1265 int err; 1266 1267 mana_gd_init_req_hdr(&req.hdr, MANA_DESTROY_WQ_OBJ, 1268 sizeof(req), sizeof(resp)); 1269 req.wq_type = wq_type; 1270 req.wq_obj_handle = wq_obj; 1271 1272 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 1273 sizeof(resp)); 1274 if (err) { 1275 if_printf(ndev, "Failed to destroy WQ object: %d\n", err); 1276 return; 1277 } 1278 1279 err = mana_verify_resp_hdr(&resp.hdr, MANA_DESTROY_WQ_OBJ, 1280 sizeof(resp)); 1281 if (err || resp.hdr.status) 1282 if_printf(ndev, "Failed to destroy WQ object: %d, 0x%x\n", 1283 err, resp.hdr.status); 1284 } 1285 1286 static void 1287 mana_destroy_eq(struct mana_context *ac) 1288 { 1289 struct gdma_context *gc = ac->gdma_dev->gdma_context; 1290 struct gdma_queue *eq; 1291 int i; 1292 1293 if (!ac->eqs) 1294 return; 1295 1296 for (i = 0; i < gc->max_num_queues; i++) { 1297 eq = ac->eqs[i].eq; 1298 if (!eq) 1299 continue; 1300 1301 mana_gd_destroy_queue(gc, eq); 1302 } 1303 1304 free(ac->eqs, M_DEVBUF); 1305 ac->eqs = NULL; 1306 } 1307 1308 static int 1309 mana_create_eq(struct mana_context *ac) 1310 { 1311 struct gdma_dev *gd = ac->gdma_dev; 1312 struct gdma_context *gc = gd->gdma_context; 1313 struct gdma_queue_spec spec = {}; 1314 int err; 1315 int i; 1316 1317 ac->eqs = mallocarray(gc->max_num_queues, sizeof(struct mana_eq), 1318 M_DEVBUF, M_WAITOK | M_ZERO); 1319 1320 spec.type = GDMA_EQ; 1321 spec.monitor_avl_buf = false; 1322 spec.queue_size = EQ_SIZE; 1323 spec.eq.callback = NULL; 1324 spec.eq.context = ac->eqs; 1325 spec.eq.log2_throttle_limit = LOG2_EQ_THROTTLE; 1326 1327 for (i = 0; i < gc->max_num_queues; i++) { 1328 err = mana_gd_create_mana_eq(gd, &spec, &ac->eqs[i].eq); 1329 if (err) 1330 goto out; 1331 } 1332 1333 return 0; 1334 out: 1335 mana_destroy_eq(ac); 1336 return err; 1337 } 1338 1339 static int 1340 mana_fence_rq(struct mana_port_context *apc, struct mana_rxq *rxq) 1341 { 1342 struct mana_fence_rq_resp resp = {}; 1343 struct mana_fence_rq_req req = {}; 1344 int err; 1345 1346 init_completion(&rxq->fence_event); 1347 1348 mana_gd_init_req_hdr(&req.hdr, MANA_FENCE_RQ, 1349 sizeof(req), sizeof(resp)); 1350 req.wq_obj_handle = rxq->rxobj; 1351 1352 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 1353 sizeof(resp)); 1354 if (err) { 1355 if_printf(apc->ndev, "Failed to fence RQ %u: %d\n", 1356 rxq->rxq_idx, err); 1357 return err; 1358 } 1359 1360 err = mana_verify_resp_hdr(&resp.hdr, MANA_FENCE_RQ, sizeof(resp)); 1361 if (err || resp.hdr.status) { 1362 if_printf(apc->ndev, "Failed to fence RQ %u: %d, 0x%x\n", 1363 rxq->rxq_idx, err, resp.hdr.status); 1364 if (!err) 1365 err = EPROTO; 1366 1367 return err; 1368 } 1369 1370 if (wait_for_completion_timeout(&rxq->fence_event, 10 * hz)) { 1371 if_printf(apc->ndev, "Failed to fence RQ %u: timed out\n", 1372 rxq->rxq_idx); 1373 return ETIMEDOUT; 1374 } 1375 1376 return 0; 1377 } 1378 1379 static void 1380 mana_fence_rqs(struct mana_port_context *apc) 1381 { 1382 unsigned int rxq_idx; 1383 struct mana_rxq *rxq; 1384 int err; 1385 1386 for (rxq_idx = 0; rxq_idx < apc->num_queues; rxq_idx++) { 1387 rxq = apc->rxqs[rxq_idx]; 1388 err = mana_fence_rq(apc, rxq); 1389 1390 /* In case of any error, use sleep instead. */ 1391 if (err) 1392 gdma_msleep(100); 1393 } 1394 } 1395 1396 static int 1397 mana_move_wq_tail(struct gdma_queue *wq, uint32_t num_units) 1398 { 1399 uint32_t used_space_old; 1400 uint32_t used_space_new; 1401 1402 used_space_old = wq->head - wq->tail; 1403 used_space_new = wq->head - (wq->tail + num_units); 1404 1405 if (used_space_new > used_space_old) { 1406 mana_err(NULL, 1407 "WARNING: new used space %u greater than old one %u\n", 1408 used_space_new, used_space_old); 1409 return ERANGE; 1410 } 1411 1412 wq->tail += num_units; 1413 return 0; 1414 } 1415 1416 static void 1417 mana_poll_tx_cq(struct mana_cq *cq) 1418 { 1419 struct gdma_comp *completions = cq->gdma_comp_buf; 1420 struct gdma_posted_wqe_info *wqe_info; 1421 struct mana_send_buf_info *tx_info; 1422 unsigned int pkt_transmitted = 0; 1423 unsigned int wqe_unit_cnt = 0; 1424 struct mana_txq *txq = cq->txq; 1425 struct mana_port_context *apc; 1426 uint16_t next_to_complete; 1427 if_t ndev; 1428 int comp_read; 1429 int txq_idx = txq->idx;; 1430 int i; 1431 int sa_drop = 0; 1432 1433 struct gdma_queue *gdma_wq; 1434 unsigned int avail_space; 1435 bool txq_full = false; 1436 1437 ndev = txq->ndev; 1438 apc = if_getsoftc(ndev); 1439 1440 comp_read = mana_gd_poll_cq(cq->gdma_cq, completions, 1441 CQE_POLLING_BUFFER); 1442 1443 if (comp_read < 1) 1444 return; 1445 1446 next_to_complete = txq->next_to_complete; 1447 1448 for (i = 0; i < comp_read; i++) { 1449 struct mana_tx_comp_oob *cqe_oob; 1450 1451 if (!completions[i].is_sq) { 1452 mana_err(NULL, "WARNING: Not for SQ\n"); 1453 return; 1454 } 1455 1456 cqe_oob = (struct mana_tx_comp_oob *)completions[i].cqe_data; 1457 if (cqe_oob->cqe_hdr.client_type != 1458 MANA_CQE_COMPLETION) { 1459 mana_err(NULL, 1460 "WARNING: Invalid CQE client type %u\n", 1461 cqe_oob->cqe_hdr.client_type); 1462 return; 1463 } 1464 1465 switch (cqe_oob->cqe_hdr.cqe_type) { 1466 case CQE_TX_OKAY: 1467 break; 1468 1469 case CQE_TX_SA_DROP: 1470 case CQE_TX_MTU_DROP: 1471 case CQE_TX_INVALID_OOB: 1472 case CQE_TX_INVALID_ETH_TYPE: 1473 case CQE_TX_HDR_PROCESSING_ERROR: 1474 case CQE_TX_VF_DISABLED: 1475 case CQE_TX_VPORT_IDX_OUT_OF_RANGE: 1476 case CQE_TX_VPORT_DISABLED: 1477 case CQE_TX_VLAN_TAGGING_VIOLATION: 1478 sa_drop ++; 1479 mana_dbg(NULL, 1480 "TX: txq %d CQE error %d, ntc = %d, " 1481 "pending sends = %d: err ignored.\n", 1482 txq_idx, cqe_oob->cqe_hdr.cqe_type, 1483 next_to_complete, txq->pending_sends); 1484 counter_u64_add(txq->stats.cqe_err, 1); 1485 break; 1486 1487 default: 1488 /* If the CQE type is unknown, log a debug msg, 1489 * and still free the mbuf, etc. 1490 */ 1491 mana_dbg(NULL, 1492 "ERROR: TX: Unknown CQE type %d\n", 1493 cqe_oob->cqe_hdr.cqe_type); 1494 counter_u64_add(txq->stats.cqe_unknown_type, 1); 1495 break; 1496 } 1497 if (txq->gdma_txq_id != completions[i].wq_num) { 1498 mana_dbg(NULL, 1499 "txq gdma id not match completion wq num: " 1500 "%d != %d\n", 1501 txq->gdma_txq_id, completions[i].wq_num); 1502 break; 1503 } 1504 1505 tx_info = &txq->tx_buf_info[next_to_complete]; 1506 if (!tx_info->mbuf) { 1507 mana_err(NULL, 1508 "WARNING: txq %d Empty mbuf on tx_info: %u, " 1509 "ntu = %u, pending_sends = %d, " 1510 "transmitted = %d, sa_drop = %d, i = %d, comp_read = %d\n", 1511 txq_idx, next_to_complete, txq->next_to_use, 1512 txq->pending_sends, pkt_transmitted, sa_drop, 1513 i, comp_read); 1514 break; 1515 } 1516 1517 wqe_info = &tx_info->wqe_inf; 1518 wqe_unit_cnt += wqe_info->wqe_size_in_bu; 1519 1520 mana_tx_unmap_mbuf(apc, tx_info); 1521 mb(); 1522 1523 next_to_complete = 1524 (next_to_complete + 1) % MAX_SEND_BUFFERS_PER_QUEUE; 1525 1526 pkt_transmitted++; 1527 } 1528 1529 txq->next_to_complete = next_to_complete; 1530 1531 if (wqe_unit_cnt == 0) { 1532 mana_err(NULL, 1533 "WARNING: TX ring not proceeding!\n"); 1534 return; 1535 } 1536 1537 mana_move_wq_tail(txq->gdma_sq, wqe_unit_cnt); 1538 1539 /* Ensure tail updated before checking q stop */ 1540 wmb(); 1541 1542 gdma_wq = txq->gdma_sq; 1543 avail_space = mana_gd_wq_avail_space(gdma_wq); 1544 1545 1546 if ((if_getdrvflags(ndev) & MANA_TXQ_FULL) == MANA_TXQ_FULL) { 1547 txq_full = true; 1548 } 1549 1550 /* Ensure checking txq_full before apc->port_is_up. */ 1551 rmb(); 1552 1553 if (txq_full && apc->port_is_up && avail_space >= MAX_TX_WQE_SIZE) { 1554 /* Grab the txq lock and re-test */ 1555 mtx_lock(&txq->txq_mtx); 1556 avail_space = mana_gd_wq_avail_space(gdma_wq); 1557 1558 if ((if_getdrvflags(ndev) & MANA_TXQ_FULL) == MANA_TXQ_FULL && 1559 apc->port_is_up && avail_space >= MAX_TX_WQE_SIZE) { 1560 /* Clear the Q full flag */ 1561 if_setdrvflagbits(apc->ndev, IFF_DRV_RUNNING, 1562 IFF_DRV_OACTIVE); 1563 counter_u64_add(txq->stats.wakeup, 1); 1564 if (txq->alt_txq_idx != txq->idx) { 1565 uint64_t stops = counter_u64_fetch(txq->stats.stop); 1566 uint64_t wakeups = counter_u64_fetch(txq->stats.wakeup); 1567 /* Reset alt_txq_idx back if it is not overloaded */ 1568 if (stops < wakeups) { 1569 txq->alt_txq_idx = txq->idx; 1570 counter_u64_add(txq->stats.alt_reset, 1); 1571 } 1572 } 1573 rmb(); 1574 /* Schedule a tx enqueue task */ 1575 taskqueue_enqueue(txq->enqueue_tq, &txq->enqueue_task); 1576 } 1577 mtx_unlock(&txq->txq_mtx); 1578 } 1579 1580 if (atomic_sub_return(pkt_transmitted, &txq->pending_sends) < 0) 1581 mana_err(NULL, 1582 "WARNING: TX %d pending_sends error: %d\n", 1583 txq->idx, txq->pending_sends); 1584 1585 cq->work_done = pkt_transmitted; 1586 } 1587 1588 static void 1589 mana_post_pkt_rxq(struct mana_rxq *rxq) 1590 { 1591 struct mana_recv_buf_oob *recv_buf_oob; 1592 uint32_t curr_index; 1593 int err; 1594 1595 curr_index = rxq->buf_index++; 1596 if (rxq->buf_index == rxq->num_rx_buf) 1597 rxq->buf_index = 0; 1598 1599 recv_buf_oob = &rxq->rx_oobs[curr_index]; 1600 1601 err = mana_gd_post_work_request(rxq->gdma_rq, &recv_buf_oob->wqe_req, 1602 &recv_buf_oob->wqe_inf); 1603 if (err) { 1604 mana_err(NULL, "WARNING: rxq %u post pkt err %d\n", 1605 rxq->rxq_idx, err); 1606 return; 1607 } 1608 1609 if (recv_buf_oob->wqe_inf.wqe_size_in_bu != 1) { 1610 mana_err(NULL, "WARNING: rxq %u wqe_size_in_bu %u\n", 1611 rxq->rxq_idx, recv_buf_oob->wqe_inf.wqe_size_in_bu); 1612 } 1613 } 1614 1615 static void 1616 mana_rx_mbuf(struct mbuf *mbuf, struct mana_rxcomp_oob *cqe, 1617 struct mana_rxq *rxq) 1618 { 1619 struct mana_stats *rx_stats = &rxq->stats; 1620 if_t ndev = rxq->ndev; 1621 uint32_t pkt_len = cqe->ppi[0].pkt_len; 1622 uint16_t rxq_idx = rxq->rxq_idx; 1623 struct mana_port_context *apc; 1624 bool do_lro = false; 1625 bool do_if_input; 1626 1627 apc = if_getsoftc(ndev); 1628 rxq->rx_cq.work_done++; 1629 1630 if (!mbuf) { 1631 return; 1632 } 1633 1634 mbuf->m_flags |= M_PKTHDR; 1635 mbuf->m_pkthdr.len = pkt_len; 1636 mbuf->m_len = pkt_len; 1637 mbuf->m_pkthdr.rcvif = ndev; 1638 1639 if ((if_getcapenable(ndev) & IFCAP_RXCSUM || 1640 if_getcapenable(ndev) & IFCAP_RXCSUM_IPV6) && 1641 (cqe->rx_iphdr_csum_succeed)) { 1642 mbuf->m_pkthdr.csum_flags = CSUM_IP_CHECKED; 1643 mbuf->m_pkthdr.csum_flags |= CSUM_IP_VALID; 1644 if (cqe->rx_tcp_csum_succeed || cqe->rx_udp_csum_succeed) { 1645 mbuf->m_pkthdr.csum_flags |= 1646 (CSUM_DATA_VALID | CSUM_PSEUDO_HDR); 1647 mbuf->m_pkthdr.csum_data = 0xffff; 1648 1649 if (cqe->rx_tcp_csum_succeed) 1650 do_lro = true; 1651 } 1652 } 1653 1654 if (cqe->rx_hashtype != 0) { 1655 mbuf->m_pkthdr.flowid = cqe->ppi[0].pkt_hash; 1656 1657 uint16_t hashtype = cqe->rx_hashtype; 1658 if (hashtype & NDIS_HASH_IPV4_MASK) { 1659 hashtype &= NDIS_HASH_IPV4_MASK; 1660 switch (hashtype) { 1661 case NDIS_HASH_TCP_IPV4: 1662 M_HASHTYPE_SET(mbuf, M_HASHTYPE_RSS_TCP_IPV4); 1663 break; 1664 case NDIS_HASH_UDP_IPV4: 1665 M_HASHTYPE_SET(mbuf, M_HASHTYPE_RSS_UDP_IPV4); 1666 break; 1667 default: 1668 M_HASHTYPE_SET(mbuf, M_HASHTYPE_RSS_IPV4); 1669 } 1670 } else if (hashtype & NDIS_HASH_IPV6_MASK) { 1671 hashtype &= NDIS_HASH_IPV6_MASK; 1672 switch (hashtype) { 1673 case NDIS_HASH_TCP_IPV6: 1674 M_HASHTYPE_SET(mbuf, M_HASHTYPE_RSS_TCP_IPV6); 1675 break; 1676 case NDIS_HASH_TCP_IPV6_EX: 1677 M_HASHTYPE_SET(mbuf, 1678 M_HASHTYPE_RSS_TCP_IPV6_EX); 1679 break; 1680 case NDIS_HASH_UDP_IPV6: 1681 M_HASHTYPE_SET(mbuf, M_HASHTYPE_RSS_UDP_IPV6); 1682 break; 1683 case NDIS_HASH_UDP_IPV6_EX: 1684 M_HASHTYPE_SET(mbuf, 1685 M_HASHTYPE_RSS_UDP_IPV6_EX); 1686 break; 1687 default: 1688 M_HASHTYPE_SET(mbuf, M_HASHTYPE_RSS_IPV6); 1689 } 1690 } else { 1691 M_HASHTYPE_SET(mbuf, M_HASHTYPE_OPAQUE_HASH); 1692 } 1693 } else { 1694 mbuf->m_pkthdr.flowid = rxq_idx; 1695 M_HASHTYPE_SET(mbuf, M_HASHTYPE_NONE); 1696 } 1697 1698 do_if_input = true; 1699 if ((if_getcapenable(ndev) & IFCAP_LRO) && do_lro) { 1700 rxq->lro_tried++; 1701 if (rxq->lro.lro_cnt != 0 && 1702 tcp_lro_rx(&rxq->lro, mbuf, 0) == 0) 1703 do_if_input = false; 1704 else 1705 rxq->lro_failed++; 1706 } 1707 if (do_if_input) { 1708 if_input(ndev, mbuf); 1709 } 1710 1711 counter_enter(); 1712 counter_u64_add_protected(rx_stats->packets, 1); 1713 counter_u64_add_protected(apc->port_stats.rx_packets, 1); 1714 counter_u64_add_protected(rx_stats->bytes, pkt_len); 1715 counter_u64_add_protected(apc->port_stats.rx_bytes, pkt_len); 1716 counter_exit(); 1717 } 1718 1719 static void 1720 mana_process_rx_cqe(struct mana_rxq *rxq, struct mana_cq *cq, 1721 struct gdma_comp *cqe) 1722 { 1723 struct mana_rxcomp_oob *oob = (struct mana_rxcomp_oob *)cqe->cqe_data; 1724 struct mana_recv_buf_oob *rxbuf_oob; 1725 if_t ndev = rxq->ndev; 1726 struct mana_port_context *apc; 1727 struct mbuf *old_mbuf; 1728 uint32_t curr, pktlen; 1729 int err; 1730 1731 switch (oob->cqe_hdr.cqe_type) { 1732 case CQE_RX_OKAY: 1733 break; 1734 1735 case CQE_RX_TRUNCATED: 1736 apc = if_getsoftc(ndev); 1737 counter_u64_add(apc->port_stats.rx_drops, 1); 1738 rxbuf_oob = &rxq->rx_oobs[rxq->buf_index]; 1739 if_printf(ndev, "Dropped a truncated packet\n"); 1740 goto drop; 1741 1742 case CQE_RX_COALESCED_4: 1743 if_printf(ndev, "RX coalescing is unsupported\n"); 1744 return; 1745 1746 case CQE_RX_OBJECT_FENCE: 1747 complete(&rxq->fence_event); 1748 return; 1749 1750 default: 1751 if_printf(ndev, "Unknown RX CQE type = %d\n", 1752 oob->cqe_hdr.cqe_type); 1753 return; 1754 } 1755 1756 if (oob->cqe_hdr.cqe_type != CQE_RX_OKAY) 1757 return; 1758 1759 pktlen = oob->ppi[0].pkt_len; 1760 1761 if (pktlen == 0) { 1762 /* data packets should never have packetlength of zero */ 1763 if_printf(ndev, "RX pkt len=0, rq=%u, cq=%u, rxobj=0x%jx\n", 1764 rxq->gdma_id, cq->gdma_id, rxq->rxobj); 1765 return; 1766 } 1767 1768 curr = rxq->buf_index; 1769 rxbuf_oob = &rxq->rx_oobs[curr]; 1770 if (rxbuf_oob->wqe_inf.wqe_size_in_bu != 1) { 1771 mana_err(NULL, "WARNING: Rx Incorrect complete " 1772 "WQE size %u\n", 1773 rxbuf_oob->wqe_inf.wqe_size_in_bu); 1774 } 1775 1776 apc = if_getsoftc(ndev); 1777 1778 old_mbuf = rxbuf_oob->mbuf; 1779 1780 /* Unload DMA map for the old mbuf */ 1781 mana_unload_rx_mbuf(apc, rxq, rxbuf_oob, false); 1782 1783 /* Load a new mbuf to replace the old one */ 1784 err = mana_load_rx_mbuf(apc, rxq, rxbuf_oob, true); 1785 if (err) { 1786 mana_dbg(NULL, 1787 "failed to load rx mbuf, err = %d, packet dropped.\n", 1788 err); 1789 counter_u64_add(rxq->stats.mbuf_alloc_fail, 1); 1790 /* 1791 * Failed to load new mbuf, rxbuf_oob->mbuf is still 1792 * pointing to the old one. Drop the packet. 1793 */ 1794 old_mbuf = NULL; 1795 /* Reload the existing mbuf */ 1796 mana_load_rx_mbuf(apc, rxq, rxbuf_oob, false); 1797 } 1798 1799 mana_rx_mbuf(old_mbuf, oob, rxq); 1800 1801 drop: 1802 mana_move_wq_tail(rxq->gdma_rq, rxbuf_oob->wqe_inf.wqe_size_in_bu); 1803 1804 mana_post_pkt_rxq(rxq); 1805 } 1806 1807 static void 1808 mana_poll_rx_cq(struct mana_cq *cq) 1809 { 1810 struct gdma_comp *comp = cq->gdma_comp_buf; 1811 int comp_read, i; 1812 1813 comp_read = mana_gd_poll_cq(cq->gdma_cq, comp, CQE_POLLING_BUFFER); 1814 KASSERT(comp_read <= CQE_POLLING_BUFFER, 1815 ("comp_read %d great than buf size %d", 1816 comp_read, CQE_POLLING_BUFFER)); 1817 1818 for (i = 0; i < comp_read; i++) { 1819 if (comp[i].is_sq == true) { 1820 mana_err(NULL, 1821 "WARNING: CQE not for receive queue\n"); 1822 return; 1823 } 1824 1825 /* verify recv cqe references the right rxq */ 1826 if (comp[i].wq_num != cq->rxq->gdma_id) { 1827 mana_err(NULL, 1828 "WARNING: Received CQE %d not for " 1829 "this receive queue %d\n", 1830 comp[i].wq_num, cq->rxq->gdma_id); 1831 return; 1832 } 1833 1834 mana_process_rx_cqe(cq->rxq, cq, &comp[i]); 1835 } 1836 1837 if (comp_read > 0) { 1838 struct gdma_context *gc = 1839 cq->rxq->gdma_rq->gdma_dev->gdma_context; 1840 1841 mana_gd_wq_ring_doorbell(gc, cq->rxq->gdma_rq); 1842 } 1843 1844 tcp_lro_flush_all(&cq->rxq->lro); 1845 } 1846 1847 static void 1848 mana_cq_handler(void *context, struct gdma_queue *gdma_queue) 1849 { 1850 struct mana_cq *cq = context; 1851 uint8_t arm_bit; 1852 1853 KASSERT(cq->gdma_cq == gdma_queue, 1854 ("cq do not match %p, %p", cq->gdma_cq, gdma_queue)); 1855 1856 if (cq->type == MANA_CQ_TYPE_RX) { 1857 mana_poll_rx_cq(cq); 1858 } else { 1859 mana_poll_tx_cq(cq); 1860 } 1861 1862 if (cq->work_done < cq->budget && cq->do_not_ring_db == false) 1863 arm_bit = SET_ARM_BIT; 1864 else 1865 arm_bit = 0; 1866 1867 mana_gd_ring_cq(gdma_queue, arm_bit); 1868 } 1869 1870 #define MANA_POLL_BUDGET 8 1871 #define MANA_RX_BUDGET 256 1872 #define MANA_TX_BUDGET MAX_SEND_BUFFERS_PER_QUEUE 1873 1874 static void 1875 mana_poll(void *arg, int pending) 1876 { 1877 struct mana_cq *cq = arg; 1878 int i; 1879 1880 cq->work_done = 0; 1881 if (cq->type == MANA_CQ_TYPE_RX) { 1882 cq->budget = MANA_RX_BUDGET; 1883 } else { 1884 cq->budget = MANA_TX_BUDGET; 1885 } 1886 1887 for (i = 0; i < MANA_POLL_BUDGET; i++) { 1888 /* 1889 * If this is the last loop, set the budget big enough 1890 * so it will arm the CQ any way. 1891 */ 1892 if (i == (MANA_POLL_BUDGET - 1)) 1893 cq->budget = CQE_POLLING_BUFFER + 1; 1894 1895 mana_cq_handler(cq, cq->gdma_cq); 1896 1897 if (cq->work_done < cq->budget) 1898 break; 1899 1900 cq->work_done = 0; 1901 } 1902 } 1903 1904 static void 1905 mana_schedule_task(void *arg, struct gdma_queue *gdma_queue) 1906 { 1907 struct mana_cq *cq = arg; 1908 1909 taskqueue_enqueue(cq->cleanup_tq, &cq->cleanup_task); 1910 } 1911 1912 static void 1913 mana_deinit_cq(struct mana_port_context *apc, struct mana_cq *cq) 1914 { 1915 struct gdma_dev *gd = apc->ac->gdma_dev; 1916 1917 if (!cq->gdma_cq) 1918 return; 1919 1920 /* Drain cleanup taskqueue */ 1921 if (cq->cleanup_tq) { 1922 while (taskqueue_cancel(cq->cleanup_tq, 1923 &cq->cleanup_task, NULL)) { 1924 taskqueue_drain(cq->cleanup_tq, 1925 &cq->cleanup_task); 1926 } 1927 1928 taskqueue_free(cq->cleanup_tq); 1929 } 1930 1931 mana_gd_destroy_queue(gd->gdma_context, cq->gdma_cq); 1932 } 1933 1934 static void 1935 mana_deinit_txq(struct mana_port_context *apc, struct mana_txq *txq) 1936 { 1937 struct gdma_dev *gd = apc->ac->gdma_dev; 1938 struct mana_send_buf_info *txbuf_info; 1939 uint32_t pending_sends; 1940 int i; 1941 1942 if (!txq->gdma_sq) 1943 return; 1944 1945 if ((pending_sends = atomic_read(&txq->pending_sends)) > 0) { 1946 mana_err(NULL, 1947 "WARNING: txq pending sends not zero: %u\n", 1948 pending_sends); 1949 } 1950 1951 if (txq->next_to_use != txq->next_to_complete) { 1952 mana_err(NULL, 1953 "WARNING: txq buf not completed, " 1954 "next use %u, next complete %u\n", 1955 txq->next_to_use, txq->next_to_complete); 1956 } 1957 1958 /* Flush buf ring. Grab txq mtx lock */ 1959 if (txq->txq_br) { 1960 mtx_lock(&txq->txq_mtx); 1961 drbr_flush(apc->ndev, txq->txq_br); 1962 mtx_unlock(&txq->txq_mtx); 1963 buf_ring_free(txq->txq_br, M_DEVBUF); 1964 } 1965 1966 /* Drain taskqueue */ 1967 if (txq->enqueue_tq) { 1968 while (taskqueue_cancel(txq->enqueue_tq, 1969 &txq->enqueue_task, NULL)) { 1970 taskqueue_drain(txq->enqueue_tq, 1971 &txq->enqueue_task); 1972 } 1973 1974 taskqueue_free(txq->enqueue_tq); 1975 } 1976 1977 if (txq->tx_buf_info) { 1978 /* Free all mbufs which are still in-flight */ 1979 for (i = 0; i < MAX_SEND_BUFFERS_PER_QUEUE; i++) { 1980 txbuf_info = &txq->tx_buf_info[i]; 1981 if (txbuf_info->mbuf) { 1982 mana_tx_unmap_mbuf(apc, txbuf_info); 1983 } 1984 } 1985 1986 free(txq->tx_buf_info, M_DEVBUF); 1987 } 1988 1989 mana_free_counters((counter_u64_t *)&txq->stats, 1990 sizeof(txq->stats)); 1991 1992 mana_gd_destroy_queue(gd->gdma_context, txq->gdma_sq); 1993 1994 mtx_destroy(&txq->txq_mtx); 1995 } 1996 1997 static void 1998 mana_destroy_txq(struct mana_port_context *apc) 1999 { 2000 int i; 2001 2002 if (!apc->tx_qp) 2003 return; 2004 2005 for (i = 0; i < apc->num_queues; i++) { 2006 mana_destroy_wq_obj(apc, GDMA_SQ, apc->tx_qp[i].tx_object); 2007 2008 mana_deinit_cq(apc, &apc->tx_qp[i].tx_cq); 2009 2010 mana_deinit_txq(apc, &apc->tx_qp[i].txq); 2011 } 2012 2013 free(apc->tx_qp, M_DEVBUF); 2014 apc->tx_qp = NULL; 2015 } 2016 2017 static int 2018 mana_create_txq(struct mana_port_context *apc, if_t net) 2019 { 2020 struct mana_context *ac = apc->ac; 2021 struct gdma_dev *gd = ac->gdma_dev; 2022 struct mana_obj_spec wq_spec; 2023 struct mana_obj_spec cq_spec; 2024 struct gdma_queue_spec spec; 2025 struct gdma_context *gc; 2026 struct mana_txq *txq; 2027 struct mana_cq *cq; 2028 uint32_t txq_size; 2029 uint32_t cq_size; 2030 int err; 2031 int i; 2032 2033 apc->tx_qp = mallocarray(apc->num_queues, sizeof(struct mana_tx_qp), 2034 M_DEVBUF, M_WAITOK | M_ZERO); 2035 2036 /* The minimum size of the WQE is 32 bytes, hence 2037 * MAX_SEND_BUFFERS_PER_QUEUE represents the maximum number of WQEs 2038 * the SQ can store. This value is then used to size other queues 2039 * to prevent overflow. 2040 */ 2041 txq_size = MAX_SEND_BUFFERS_PER_QUEUE * 32; 2042 KASSERT(IS_ALIGNED(txq_size, PAGE_SIZE), 2043 ("txq size not page aligned")); 2044 2045 cq_size = MAX_SEND_BUFFERS_PER_QUEUE * COMP_ENTRY_SIZE; 2046 cq_size = ALIGN(cq_size, PAGE_SIZE); 2047 2048 gc = gd->gdma_context; 2049 2050 for (i = 0; i < apc->num_queues; i++) { 2051 apc->tx_qp[i].tx_object = INVALID_MANA_HANDLE; 2052 2053 /* Create SQ */ 2054 txq = &apc->tx_qp[i].txq; 2055 2056 txq->ndev = net; 2057 txq->vp_offset = apc->tx_vp_offset; 2058 txq->idx = i; 2059 txq->alt_txq_idx = i; 2060 2061 memset(&spec, 0, sizeof(spec)); 2062 spec.type = GDMA_SQ; 2063 spec.monitor_avl_buf = true; 2064 spec.queue_size = txq_size; 2065 err = mana_gd_create_mana_wq_cq(gd, &spec, &txq->gdma_sq); 2066 if (err) 2067 goto out; 2068 2069 /* Create SQ's CQ */ 2070 cq = &apc->tx_qp[i].tx_cq; 2071 cq->type = MANA_CQ_TYPE_TX; 2072 2073 cq->txq = txq; 2074 2075 memset(&spec, 0, sizeof(spec)); 2076 spec.type = GDMA_CQ; 2077 spec.monitor_avl_buf = false; 2078 spec.queue_size = cq_size; 2079 spec.cq.callback = mana_schedule_task; 2080 spec.cq.parent_eq = ac->eqs[i].eq; 2081 spec.cq.context = cq; 2082 err = mana_gd_create_mana_wq_cq(gd, &spec, &cq->gdma_cq); 2083 if (err) 2084 goto out; 2085 2086 memset(&wq_spec, 0, sizeof(wq_spec)); 2087 memset(&cq_spec, 0, sizeof(cq_spec)); 2088 2089 wq_spec.gdma_region = txq->gdma_sq->mem_info.dma_region_handle; 2090 wq_spec.queue_size = txq->gdma_sq->queue_size; 2091 2092 cq_spec.gdma_region = cq->gdma_cq->mem_info.dma_region_handle; 2093 cq_spec.queue_size = cq->gdma_cq->queue_size; 2094 cq_spec.modr_ctx_id = 0; 2095 cq_spec.attached_eq = cq->gdma_cq->cq.parent->id; 2096 2097 err = mana_create_wq_obj(apc, apc->port_handle, GDMA_SQ, 2098 &wq_spec, &cq_spec, &apc->tx_qp[i].tx_object); 2099 2100 if (err) 2101 goto out; 2102 2103 txq->gdma_sq->id = wq_spec.queue_index; 2104 cq->gdma_cq->id = cq_spec.queue_index; 2105 2106 txq->gdma_sq->mem_info.dma_region_handle = 2107 GDMA_INVALID_DMA_REGION; 2108 cq->gdma_cq->mem_info.dma_region_handle = 2109 GDMA_INVALID_DMA_REGION; 2110 2111 txq->gdma_txq_id = txq->gdma_sq->id; 2112 2113 cq->gdma_id = cq->gdma_cq->id; 2114 2115 mana_dbg(NULL, 2116 "txq %d, txq gdma id %d, txq cq gdma id %d\n", 2117 i, txq->gdma_txq_id, cq->gdma_id);; 2118 2119 if (cq->gdma_id >= gc->max_num_cqs) { 2120 if_printf(net, "CQ id %u too large.\n", cq->gdma_id); 2121 err = EINVAL; 2122 goto out; 2123 } 2124 2125 gc->cq_table[cq->gdma_id] = cq->gdma_cq; 2126 2127 /* Initialize tx specific data */ 2128 txq->tx_buf_info = malloc(MAX_SEND_BUFFERS_PER_QUEUE * 2129 sizeof(struct mana_send_buf_info), 2130 M_DEVBUF, M_WAITOK | M_ZERO); 2131 2132 snprintf(txq->txq_mtx_name, nitems(txq->txq_mtx_name), 2133 "mana:tx(%d)", i); 2134 mtx_init(&txq->txq_mtx, txq->txq_mtx_name, NULL, MTX_DEF); 2135 2136 txq->txq_br = buf_ring_alloc(4 * MAX_SEND_BUFFERS_PER_QUEUE, 2137 M_DEVBUF, M_WAITOK, &txq->txq_mtx); 2138 2139 /* Allocate taskqueue for deferred send */ 2140 TASK_INIT(&txq->enqueue_task, 0, mana_xmit_taskfunc, txq); 2141 txq->enqueue_tq = taskqueue_create_fast("mana_tx_enque", 2142 M_NOWAIT, taskqueue_thread_enqueue, &txq->enqueue_tq); 2143 if (unlikely(txq->enqueue_tq == NULL)) { 2144 if_printf(net, 2145 "Unable to create tx %d enqueue task queue\n", i); 2146 err = ENOMEM; 2147 goto out; 2148 } 2149 taskqueue_start_threads(&txq->enqueue_tq, 1, PI_NET, 2150 "mana txq p%u-tx%d", apc->port_idx, i); 2151 2152 mana_alloc_counters((counter_u64_t *)&txq->stats, 2153 sizeof(txq->stats)); 2154 2155 /* Allocate and start the cleanup task on CQ */ 2156 cq->do_not_ring_db = false; 2157 2158 NET_TASK_INIT(&cq->cleanup_task, 0, mana_poll, cq); 2159 cq->cleanup_tq = 2160 taskqueue_create_fast("mana tx cq cleanup", 2161 M_WAITOK, taskqueue_thread_enqueue, 2162 &cq->cleanup_tq); 2163 2164 if (apc->last_tx_cq_bind_cpu < 0) 2165 apc->last_tx_cq_bind_cpu = CPU_FIRST(); 2166 cq->cpu = apc->last_tx_cq_bind_cpu; 2167 apc->last_tx_cq_bind_cpu = CPU_NEXT(apc->last_tx_cq_bind_cpu); 2168 2169 if (apc->bind_cleanup_thread_cpu) { 2170 cpuset_t cpu_mask; 2171 CPU_SETOF(cq->cpu, &cpu_mask); 2172 taskqueue_start_threads_cpuset(&cq->cleanup_tq, 2173 1, PI_NET, &cpu_mask, 2174 "mana cq p%u-tx%u-cpu%d", 2175 apc->port_idx, txq->idx, cq->cpu); 2176 } else { 2177 taskqueue_start_threads(&cq->cleanup_tq, 1, 2178 PI_NET, "mana cq p%u-tx%u", 2179 apc->port_idx, txq->idx); 2180 } 2181 2182 mana_gd_ring_cq(cq->gdma_cq, SET_ARM_BIT); 2183 } 2184 2185 return 0; 2186 out: 2187 mana_destroy_txq(apc); 2188 return err; 2189 } 2190 2191 static void 2192 mana_destroy_rxq(struct mana_port_context *apc, struct mana_rxq *rxq, 2193 bool validate_state) 2194 { 2195 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context; 2196 struct mana_recv_buf_oob *rx_oob; 2197 int i; 2198 2199 if (!rxq) 2200 return; 2201 2202 if (validate_state) { 2203 /* 2204 * XXX Cancel and drain cleanup task queue here. 2205 */ 2206 ; 2207 } 2208 2209 mana_destroy_wq_obj(apc, GDMA_RQ, rxq->rxobj); 2210 2211 mana_deinit_cq(apc, &rxq->rx_cq); 2212 2213 mana_free_counters((counter_u64_t *)&rxq->stats, 2214 sizeof(rxq->stats)); 2215 2216 /* Free LRO resources */ 2217 tcp_lro_free(&rxq->lro); 2218 2219 for (i = 0; i < rxq->num_rx_buf; i++) { 2220 rx_oob = &rxq->rx_oobs[i]; 2221 2222 if (rx_oob->mbuf) 2223 mana_unload_rx_mbuf(apc, rxq, rx_oob, true); 2224 2225 bus_dmamap_destroy(apc->rx_buf_tag, rx_oob->dma_map); 2226 } 2227 2228 if (rxq->gdma_rq) 2229 mana_gd_destroy_queue(gc, rxq->gdma_rq); 2230 2231 free(rxq, M_DEVBUF); 2232 } 2233 2234 #define MANA_WQE_HEADER_SIZE 16 2235 #define MANA_WQE_SGE_SIZE 16 2236 2237 static int 2238 mana_alloc_rx_wqe(struct mana_port_context *apc, 2239 struct mana_rxq *rxq, uint32_t *rxq_size, uint32_t *cq_size) 2240 { 2241 struct mana_recv_buf_oob *rx_oob; 2242 uint32_t buf_idx; 2243 int err; 2244 2245 if (rxq->datasize == 0 || rxq->datasize > PAGE_SIZE) { 2246 mana_err(NULL, 2247 "WARNING: Invalid rxq datasize %u\n", rxq->datasize); 2248 } 2249 2250 *rxq_size = 0; 2251 *cq_size = 0; 2252 2253 for (buf_idx = 0; buf_idx < rxq->num_rx_buf; buf_idx++) { 2254 rx_oob = &rxq->rx_oobs[buf_idx]; 2255 memset(rx_oob, 0, sizeof(*rx_oob)); 2256 2257 err = bus_dmamap_create(apc->rx_buf_tag, 0, 2258 &rx_oob->dma_map); 2259 if (err) { 2260 mana_err(NULL, 2261 "Failed to create rx DMA map for buf %d\n", 2262 buf_idx); 2263 return err; 2264 } 2265 2266 err = mana_load_rx_mbuf(apc, rxq, rx_oob, true); 2267 if (err) { 2268 mana_err(NULL, 2269 "Failed to create rx DMA map for buf %d\n", 2270 buf_idx); 2271 bus_dmamap_destroy(apc->rx_buf_tag, rx_oob->dma_map); 2272 return err; 2273 } 2274 2275 rx_oob->wqe_req.sgl = rx_oob->sgl; 2276 rx_oob->wqe_req.num_sge = rx_oob->num_sge; 2277 rx_oob->wqe_req.inline_oob_size = 0; 2278 rx_oob->wqe_req.inline_oob_data = NULL; 2279 rx_oob->wqe_req.flags = 0; 2280 rx_oob->wqe_req.client_data_unit = 0; 2281 2282 *rxq_size += ALIGN(MANA_WQE_HEADER_SIZE + 2283 MANA_WQE_SGE_SIZE * rx_oob->num_sge, 32); 2284 *cq_size += COMP_ENTRY_SIZE; 2285 } 2286 2287 return 0; 2288 } 2289 2290 static int 2291 mana_push_wqe(struct mana_rxq *rxq) 2292 { 2293 struct mana_recv_buf_oob *rx_oob; 2294 uint32_t buf_idx; 2295 int err; 2296 2297 for (buf_idx = 0; buf_idx < rxq->num_rx_buf; buf_idx++) { 2298 rx_oob = &rxq->rx_oobs[buf_idx]; 2299 2300 err = mana_gd_post_and_ring(rxq->gdma_rq, &rx_oob->wqe_req, 2301 &rx_oob->wqe_inf); 2302 if (err) 2303 return ENOSPC; 2304 } 2305 2306 return 0; 2307 } 2308 2309 static struct mana_rxq * 2310 mana_create_rxq(struct mana_port_context *apc, uint32_t rxq_idx, 2311 struct mana_eq *eq, if_t ndev) 2312 { 2313 struct gdma_dev *gd = apc->ac->gdma_dev; 2314 struct mana_obj_spec wq_spec; 2315 struct mana_obj_spec cq_spec; 2316 struct gdma_queue_spec spec; 2317 struct mana_cq *cq = NULL; 2318 uint32_t cq_size, rq_size; 2319 struct gdma_context *gc; 2320 struct mana_rxq *rxq; 2321 int err; 2322 2323 gc = gd->gdma_context; 2324 2325 rxq = malloc(sizeof(*rxq) + 2326 RX_BUFFERS_PER_QUEUE * sizeof(struct mana_recv_buf_oob), 2327 M_DEVBUF, M_WAITOK | M_ZERO); 2328 rxq->ndev = ndev; 2329 rxq->num_rx_buf = RX_BUFFERS_PER_QUEUE; 2330 rxq->rxq_idx = rxq_idx; 2331 /* 2332 * Minimum size is MCLBYTES(2048) bytes for a mbuf cluster. 2333 * Now we just allow maximum size of 4096. 2334 */ 2335 rxq->datasize = ALIGN(apc->frame_size, MCLBYTES); 2336 if (rxq->datasize > MAX_FRAME_SIZE) 2337 rxq->datasize = MAX_FRAME_SIZE; 2338 2339 mana_dbg(NULL, "Setting rxq %d datasize %d\n", 2340 rxq_idx, rxq->datasize); 2341 2342 rxq->rxobj = INVALID_MANA_HANDLE; 2343 2344 err = mana_alloc_rx_wqe(apc, rxq, &rq_size, &cq_size); 2345 if (err) 2346 goto out; 2347 2348 /* Create LRO for the RQ */ 2349 if (if_getcapenable(ndev) & IFCAP_LRO) { 2350 err = tcp_lro_init(&rxq->lro); 2351 if (err) { 2352 if_printf(ndev, "Failed to create LRO for rxq %d\n", 2353 rxq_idx); 2354 } else { 2355 rxq->lro.ifp = ndev; 2356 } 2357 } 2358 2359 mana_alloc_counters((counter_u64_t *)&rxq->stats, 2360 sizeof(rxq->stats)); 2361 2362 rq_size = ALIGN(rq_size, PAGE_SIZE); 2363 cq_size = ALIGN(cq_size, PAGE_SIZE); 2364 2365 /* Create RQ */ 2366 memset(&spec, 0, sizeof(spec)); 2367 spec.type = GDMA_RQ; 2368 spec.monitor_avl_buf = true; 2369 spec.queue_size = rq_size; 2370 err = mana_gd_create_mana_wq_cq(gd, &spec, &rxq->gdma_rq); 2371 if (err) 2372 goto out; 2373 2374 /* Create RQ's CQ */ 2375 cq = &rxq->rx_cq; 2376 cq->type = MANA_CQ_TYPE_RX; 2377 cq->rxq = rxq; 2378 2379 memset(&spec, 0, sizeof(spec)); 2380 spec.type = GDMA_CQ; 2381 spec.monitor_avl_buf = false; 2382 spec.queue_size = cq_size; 2383 spec.cq.callback = mana_schedule_task; 2384 spec.cq.parent_eq = eq->eq; 2385 spec.cq.context = cq; 2386 err = mana_gd_create_mana_wq_cq(gd, &spec, &cq->gdma_cq); 2387 if (err) 2388 goto out; 2389 2390 memset(&wq_spec, 0, sizeof(wq_spec)); 2391 memset(&cq_spec, 0, sizeof(cq_spec)); 2392 wq_spec.gdma_region = rxq->gdma_rq->mem_info.dma_region_handle; 2393 wq_spec.queue_size = rxq->gdma_rq->queue_size; 2394 2395 cq_spec.gdma_region = cq->gdma_cq->mem_info.dma_region_handle; 2396 cq_spec.queue_size = cq->gdma_cq->queue_size; 2397 cq_spec.modr_ctx_id = 0; 2398 cq_spec.attached_eq = cq->gdma_cq->cq.parent->id; 2399 2400 err = mana_create_wq_obj(apc, apc->port_handle, GDMA_RQ, 2401 &wq_spec, &cq_spec, &rxq->rxobj); 2402 if (err) 2403 goto out; 2404 2405 rxq->gdma_rq->id = wq_spec.queue_index; 2406 cq->gdma_cq->id = cq_spec.queue_index; 2407 2408 rxq->gdma_rq->mem_info.dma_region_handle = GDMA_INVALID_DMA_REGION; 2409 cq->gdma_cq->mem_info.dma_region_handle = GDMA_INVALID_DMA_REGION; 2410 2411 rxq->gdma_id = rxq->gdma_rq->id; 2412 cq->gdma_id = cq->gdma_cq->id; 2413 2414 err = mana_push_wqe(rxq); 2415 if (err) 2416 goto out; 2417 2418 if (cq->gdma_id >= gc->max_num_cqs) { 2419 err = EINVAL; 2420 goto out; 2421 } 2422 2423 gc->cq_table[cq->gdma_id] = cq->gdma_cq; 2424 2425 /* Allocate and start the cleanup task on CQ */ 2426 cq->do_not_ring_db = false; 2427 2428 NET_TASK_INIT(&cq->cleanup_task, 0, mana_poll, cq); 2429 cq->cleanup_tq = 2430 taskqueue_create_fast("mana rx cq cleanup", 2431 M_WAITOK, taskqueue_thread_enqueue, 2432 &cq->cleanup_tq); 2433 2434 if (apc->last_rx_cq_bind_cpu < 0) 2435 apc->last_rx_cq_bind_cpu = CPU_FIRST(); 2436 cq->cpu = apc->last_rx_cq_bind_cpu; 2437 apc->last_rx_cq_bind_cpu = CPU_NEXT(apc->last_rx_cq_bind_cpu); 2438 2439 if (apc->bind_cleanup_thread_cpu) { 2440 cpuset_t cpu_mask; 2441 CPU_SETOF(cq->cpu, &cpu_mask); 2442 taskqueue_start_threads_cpuset(&cq->cleanup_tq, 2443 1, PI_NET, &cpu_mask, 2444 "mana cq p%u-rx%u-cpu%d", 2445 apc->port_idx, rxq->rxq_idx, cq->cpu); 2446 } else { 2447 taskqueue_start_threads(&cq->cleanup_tq, 1, 2448 PI_NET, "mana cq p%u-rx%u", 2449 apc->port_idx, rxq->rxq_idx); 2450 } 2451 2452 mana_gd_ring_cq(cq->gdma_cq, SET_ARM_BIT); 2453 out: 2454 if (!err) 2455 return rxq; 2456 2457 if_printf(ndev, "Failed to create RXQ: err = %d\n", err); 2458 2459 mana_destroy_rxq(apc, rxq, false); 2460 2461 if (cq) 2462 mana_deinit_cq(apc, cq); 2463 2464 return NULL; 2465 } 2466 2467 static int 2468 mana_add_rx_queues(struct mana_port_context *apc, if_t ndev) 2469 { 2470 struct mana_context *ac = apc->ac; 2471 struct mana_rxq *rxq; 2472 int err = 0; 2473 int i; 2474 2475 for (i = 0; i < apc->num_queues; i++) { 2476 rxq = mana_create_rxq(apc, i, &ac->eqs[i], ndev); 2477 if (!rxq) { 2478 err = ENOMEM; 2479 goto out; 2480 } 2481 2482 apc->rxqs[i] = rxq; 2483 } 2484 2485 apc->default_rxobj = apc->rxqs[0]->rxobj; 2486 out: 2487 return err; 2488 } 2489 2490 static void 2491 mana_destroy_vport(struct mana_port_context *apc) 2492 { 2493 struct mana_rxq *rxq; 2494 uint32_t rxq_idx; 2495 2496 for (rxq_idx = 0; rxq_idx < apc->num_queues; rxq_idx++) { 2497 rxq = apc->rxqs[rxq_idx]; 2498 if (!rxq) 2499 continue; 2500 2501 mana_destroy_rxq(apc, rxq, true); 2502 apc->rxqs[rxq_idx] = NULL; 2503 } 2504 2505 mana_destroy_txq(apc); 2506 2507 mana_uncfg_vport(apc); 2508 } 2509 2510 static int 2511 mana_create_vport(struct mana_port_context *apc, if_t net) 2512 { 2513 struct gdma_dev *gd = apc->ac->gdma_dev; 2514 int err; 2515 2516 apc->default_rxobj = INVALID_MANA_HANDLE; 2517 2518 err = mana_cfg_vport(apc, gd->pdid, gd->doorbell); 2519 if (err) 2520 return err; 2521 2522 return mana_create_txq(apc, net); 2523 } 2524 2525 2526 static void mana_rss_table_init(struct mana_port_context *apc) 2527 { 2528 int i; 2529 2530 for (i = 0; i < MANA_INDIRECT_TABLE_SIZE; i++) 2531 apc->indir_table[i] = i % apc->num_queues; 2532 } 2533 2534 int mana_config_rss(struct mana_port_context *apc, enum TRI_STATE rx, 2535 bool update_hash, bool update_tab) 2536 { 2537 uint32_t queue_idx; 2538 int err; 2539 int i; 2540 2541 if (update_tab) { 2542 for (i = 0; i < MANA_INDIRECT_TABLE_SIZE; i++) { 2543 queue_idx = apc->indir_table[i]; 2544 apc->rxobj_table[i] = apc->rxqs[queue_idx]->rxobj; 2545 } 2546 } 2547 2548 err = mana_cfg_vport_steering(apc, rx, true, update_hash, update_tab); 2549 if (err) 2550 return err; 2551 2552 mana_fence_rqs(apc); 2553 2554 return 0; 2555 } 2556 2557 static int 2558 mana_init_port(if_t ndev) 2559 { 2560 struct mana_port_context *apc = if_getsoftc(ndev); 2561 uint32_t max_txq, max_rxq, max_queues; 2562 int port_idx = apc->port_idx; 2563 uint32_t num_indirect_entries; 2564 int err; 2565 2566 err = mana_init_port_context(apc); 2567 if (err) 2568 return err; 2569 2570 err = mana_query_vport_cfg(apc, port_idx, &max_txq, &max_rxq, 2571 &num_indirect_entries); 2572 if (err) { 2573 if_printf(ndev, "Failed to query info for vPort %d\n", 2574 port_idx); 2575 goto reset_apc; 2576 } 2577 2578 max_queues = min_t(uint32_t, max_txq, max_rxq); 2579 if (apc->max_queues > max_queues) 2580 apc->max_queues = max_queues; 2581 2582 if (apc->num_queues > apc->max_queues) 2583 apc->num_queues = apc->max_queues; 2584 2585 return 0; 2586 2587 reset_apc: 2588 bus_dma_tag_destroy(apc->rx_buf_tag); 2589 apc->rx_buf_tag = NULL; 2590 free(apc->rxqs, M_DEVBUF); 2591 apc->rxqs = NULL; 2592 return err; 2593 } 2594 2595 int 2596 mana_alloc_queues(if_t ndev) 2597 { 2598 struct mana_port_context *apc = if_getsoftc(ndev); 2599 int err; 2600 2601 err = mana_create_vport(apc, ndev); 2602 if (err) 2603 return err; 2604 2605 err = mana_add_rx_queues(apc, ndev); 2606 if (err) 2607 goto destroy_vport; 2608 2609 apc->rss_state = apc->num_queues > 1 ? TRI_STATE_TRUE : TRI_STATE_FALSE; 2610 2611 mana_rss_table_init(apc); 2612 2613 err = mana_config_rss(apc, TRI_STATE_TRUE, true, true); 2614 if (err) 2615 goto destroy_vport; 2616 2617 return 0; 2618 2619 destroy_vport: 2620 mana_destroy_vport(apc); 2621 return err; 2622 } 2623 2624 static int 2625 mana_up(struct mana_port_context *apc) 2626 { 2627 int err; 2628 2629 mana_dbg(NULL, "mana_up called\n"); 2630 2631 err = mana_alloc_queues(apc->ndev); 2632 if (err) { 2633 mana_err(NULL, "Faile alloc mana queues: %d\n", err); 2634 return err; 2635 } 2636 2637 /* Add queue specific sysctl */ 2638 mana_sysctl_add_queues(apc); 2639 2640 apc->port_is_up = true; 2641 2642 /* Ensure port state updated before txq state */ 2643 wmb(); 2644 2645 if_link_state_change(apc->ndev, LINK_STATE_UP); 2646 if_setdrvflagbits(apc->ndev, IFF_DRV_RUNNING, IFF_DRV_OACTIVE); 2647 2648 return 0; 2649 } 2650 2651 2652 static void 2653 mana_init(void *arg) 2654 { 2655 struct mana_port_context *apc = (struct mana_port_context *)arg; 2656 2657 MANA_APC_LOCK_LOCK(apc); 2658 if (!apc->port_is_up) { 2659 mana_up(apc); 2660 } 2661 MANA_APC_LOCK_UNLOCK(apc); 2662 } 2663 2664 static int 2665 mana_dealloc_queues(if_t ndev) 2666 { 2667 struct mana_port_context *apc = if_getsoftc(ndev); 2668 struct mana_txq *txq; 2669 int i, err; 2670 2671 if (apc->port_is_up) 2672 return EINVAL; 2673 2674 /* No packet can be transmitted now since apc->port_is_up is false. 2675 * There is still a tiny chance that mana_poll_tx_cq() can re-enable 2676 * a txq because it may not timely see apc->port_is_up being cleared 2677 * to false, but it doesn't matter since mana_start_xmit() drops any 2678 * new packets due to apc->port_is_up being false. 2679 * 2680 * Drain all the in-flight TX packets 2681 */ 2682 for (i = 0; i < apc->num_queues; i++) { 2683 txq = &apc->tx_qp[i].txq; 2684 2685 struct mana_cq *tx_cq = &apc->tx_qp[i].tx_cq; 2686 struct mana_cq *rx_cq = &(apc->rxqs[i]->rx_cq); 2687 2688 tx_cq->do_not_ring_db = true; 2689 rx_cq->do_not_ring_db = true; 2690 2691 /* Schedule a cleanup task */ 2692 taskqueue_enqueue(tx_cq->cleanup_tq, &tx_cq->cleanup_task); 2693 2694 while (atomic_read(&txq->pending_sends) > 0) 2695 usleep_range(1000, 2000); 2696 } 2697 2698 /* We're 100% sure the queues can no longer be woken up, because 2699 * we're sure now mana_poll_tx_cq() can't be running. 2700 */ 2701 2702 apc->rss_state = TRI_STATE_FALSE; 2703 err = mana_config_rss(apc, TRI_STATE_FALSE, false, false); 2704 if (err) { 2705 if_printf(ndev, "Failed to disable vPort: %d\n", err); 2706 return err; 2707 } 2708 2709 mana_destroy_vport(apc); 2710 2711 return 0; 2712 } 2713 2714 static int 2715 mana_down(struct mana_port_context *apc) 2716 { 2717 int err = 0; 2718 2719 apc->port_st_save = apc->port_is_up; 2720 apc->port_is_up = false; 2721 2722 /* Ensure port state updated before txq state */ 2723 wmb(); 2724 2725 if (apc->port_st_save) { 2726 if_setdrvflagbits(apc->ndev, IFF_DRV_OACTIVE, 2727 IFF_DRV_RUNNING); 2728 if_link_state_change(apc->ndev, LINK_STATE_DOWN); 2729 2730 mana_sysctl_free_queues(apc); 2731 2732 err = mana_dealloc_queues(apc->ndev); 2733 if (err) { 2734 if_printf(apc->ndev, 2735 "Failed to bring down mana interface: %d\n", err); 2736 } 2737 } 2738 2739 return err; 2740 } 2741 2742 int 2743 mana_detach(if_t ndev) 2744 { 2745 struct mana_port_context *apc = if_getsoftc(ndev); 2746 int err; 2747 2748 ether_ifdetach(ndev); 2749 2750 if (!apc) 2751 return 0; 2752 2753 MANA_APC_LOCK_LOCK(apc); 2754 err = mana_down(apc); 2755 MANA_APC_LOCK_UNLOCK(apc); 2756 2757 mana_cleanup_port_context(apc); 2758 2759 MANA_APC_LOCK_DESTROY(apc); 2760 2761 free(apc, M_DEVBUF); 2762 2763 return err; 2764 } 2765 2766 static int 2767 mana_probe_port(struct mana_context *ac, int port_idx, 2768 if_t *ndev_storage) 2769 { 2770 struct gdma_context *gc = ac->gdma_dev->gdma_context; 2771 struct mana_port_context *apc; 2772 uint32_t hwassist; 2773 if_t ndev; 2774 int err; 2775 2776 ndev = if_alloc_dev(IFT_ETHER, gc->dev); 2777 *ndev_storage = ndev; 2778 2779 apc = malloc(sizeof(*apc), M_DEVBUF, M_WAITOK | M_ZERO); 2780 apc->ac = ac; 2781 apc->ndev = ndev; 2782 apc->max_queues = gc->max_num_queues; 2783 apc->num_queues = min_t(unsigned int, 2784 gc->max_num_queues, MANA_MAX_NUM_QUEUES); 2785 apc->port_handle = INVALID_MANA_HANDLE; 2786 apc->port_idx = port_idx; 2787 apc->frame_size = DEFAULT_FRAME_SIZE; 2788 apc->last_tx_cq_bind_cpu = -1; 2789 apc->last_rx_cq_bind_cpu = -1; 2790 apc->vport_use_count = 0; 2791 2792 MANA_APC_LOCK_INIT(apc); 2793 2794 if_initname(ndev, device_get_name(gc->dev), port_idx); 2795 if_setdev(ndev,gc->dev); 2796 if_setsoftc(ndev, apc); 2797 2798 if_setflags(ndev, IFF_BROADCAST | IFF_SIMPLEX | IFF_MULTICAST); 2799 if_setinitfn(ndev, mana_init); 2800 if_settransmitfn(ndev, mana_start_xmit); 2801 if_setqflushfn(ndev, mana_qflush); 2802 if_setioctlfn(ndev, mana_ioctl); 2803 if_setgetcounterfn(ndev, mana_get_counter); 2804 2805 if_setmtu(ndev, ETHERMTU); 2806 if_setbaudrate(ndev, IF_Gbps(100)); 2807 2808 mana_rss_key_fill(apc->hashkey, MANA_HASH_KEY_SIZE); 2809 2810 err = mana_init_port(ndev); 2811 if (err) 2812 goto reset_apc; 2813 2814 if_setcapabilitiesbit(ndev, 2815 IFCAP_TXCSUM | IFCAP_TXCSUM_IPV6 | 2816 IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6 | 2817 IFCAP_TSO4 | IFCAP_TSO6 | 2818 IFCAP_LRO | IFCAP_LINKSTATE, 0); 2819 2820 /* Enable all available capabilities by default. */ 2821 if_setcapenable(ndev, if_getcapabilities(ndev)); 2822 2823 /* TSO parameters */ 2824 if_sethwtsomax(ndev, MANA_TSO_MAX_SZ - 2825 (ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN)); 2826 if_sethwtsomaxsegcount(ndev, MAX_MBUF_FRAGS); 2827 if_sethwtsomaxsegsize(ndev, PAGE_SIZE); 2828 2829 hwassist = 0; 2830 if (if_getcapenable(ndev) & (IFCAP_TSO4 | IFCAP_TSO6)) 2831 hwassist |= CSUM_TSO; 2832 if (if_getcapenable(ndev) & IFCAP_TXCSUM) 2833 hwassist |= (CSUM_TCP | CSUM_UDP | CSUM_IP); 2834 if (if_getcapenable(ndev) & IFCAP_TXCSUM_IPV6) 2835 hwassist |= (CSUM_UDP_IPV6 | CSUM_TCP_IPV6); 2836 mana_dbg(NULL, "set hwassist 0x%x\n", hwassist); 2837 if_sethwassist(ndev, hwassist); 2838 2839 ifmedia_init(&apc->media, IFM_IMASK, 2840 mana_ifmedia_change, mana_ifmedia_status); 2841 ifmedia_add(&apc->media, IFM_ETHER | IFM_AUTO, 0, NULL); 2842 ifmedia_set(&apc->media, IFM_ETHER | IFM_AUTO); 2843 2844 ether_ifattach(ndev, apc->mac_addr); 2845 2846 /* Initialize statistics */ 2847 mana_alloc_counters((counter_u64_t *)&apc->port_stats, 2848 sizeof(struct mana_port_stats)); 2849 mana_sysctl_add_port(apc); 2850 2851 /* Tell the stack that the interface is not active */ 2852 if_setdrvflagbits(ndev, IFF_DRV_OACTIVE, IFF_DRV_RUNNING); 2853 2854 return 0; 2855 2856 reset_apc: 2857 free(apc, M_DEVBUF); 2858 *ndev_storage = NULL; 2859 if_printf(ndev, "Failed to probe vPort %d: %d\n", port_idx, err); 2860 if_free(ndev); 2861 return err; 2862 } 2863 2864 int mana_probe(struct gdma_dev *gd) 2865 { 2866 struct gdma_context *gc = gd->gdma_context; 2867 device_t dev = gc->dev; 2868 struct mana_context *ac; 2869 int err; 2870 int i; 2871 2872 device_printf(dev, "%s protocol version: %d.%d.%d\n", DEVICE_NAME, 2873 MANA_MAJOR_VERSION, MANA_MINOR_VERSION, MANA_MICRO_VERSION); 2874 2875 err = mana_gd_register_device(gd); 2876 if (err) 2877 return err; 2878 2879 ac = malloc(sizeof(*ac), M_DEVBUF, M_WAITOK | M_ZERO); 2880 ac->gdma_dev = gd; 2881 ac->num_ports = 1; 2882 gd->driver_data = ac; 2883 2884 err = mana_create_eq(ac); 2885 if (err) 2886 goto out; 2887 2888 err = mana_query_device_cfg(ac, MANA_MAJOR_VERSION, MANA_MINOR_VERSION, 2889 MANA_MICRO_VERSION, &ac->num_ports); 2890 if (err) 2891 goto out; 2892 2893 if (ac->num_ports > MAX_PORTS_IN_MANA_DEV) 2894 ac->num_ports = MAX_PORTS_IN_MANA_DEV; 2895 2896 for (i = 0; i < ac->num_ports; i++) { 2897 err = mana_probe_port(ac, i, &ac->ports[i]); 2898 if (err) { 2899 device_printf(dev, 2900 "Failed to probe mana port %d\n", i); 2901 break; 2902 } 2903 } 2904 2905 out: 2906 if (err) 2907 mana_remove(gd); 2908 2909 return err; 2910 } 2911 2912 void 2913 mana_remove(struct gdma_dev *gd) 2914 { 2915 struct gdma_context *gc = gd->gdma_context; 2916 struct mana_context *ac = gd->driver_data; 2917 device_t dev = gc->dev; 2918 if_t ndev; 2919 int i; 2920 2921 for (i = 0; i < ac->num_ports; i++) { 2922 ndev = ac->ports[i]; 2923 if (!ndev) { 2924 if (i == 0) 2925 device_printf(dev, "No net device to remove\n"); 2926 goto out; 2927 } 2928 2929 mana_detach(ndev); 2930 2931 if_free(ndev); 2932 } 2933 2934 mana_destroy_eq(ac); 2935 2936 out: 2937 mana_gd_deregister_device(gd); 2938 gd->driver_data = NULL; 2939 gd->gdma_context = NULL; 2940 free(ac, M_DEVBUF); 2941 } 2942