1/*- 2 * SPDX-License-Identifier: BSD-2-Clause 3 * 4 * Copyright (c) 2015-2023 Amazon.com, Inc. or its affiliates. 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#include "opt_rss.h" 32#include "ena.h" 33#include "ena_datapath.h" 34#ifdef DEV_NETMAP 35#include "ena_netmap.h" 36#endif /* DEV_NETMAP */ 37#ifdef RSS 38#include <net/rss_config.h> 39#endif /* RSS */ 40 41#include <netinet6/ip6_var.h> 42 43/********************************************************************* 44 * Static functions prototypes 45 *********************************************************************/ 46 47static int ena_tx_cleanup(struct ena_ring *); 48static int ena_rx_cleanup(struct ena_ring *); 49static inline int ena_get_tx_req_id(struct ena_ring *tx_ring, 50 struct ena_com_io_cq *io_cq, uint16_t *req_id); 51static void ena_rx_hash_mbuf(struct ena_ring *, struct ena_com_rx_ctx *, 52 struct mbuf *); 53static struct mbuf *ena_rx_mbuf(struct ena_ring *, struct ena_com_rx_buf_info *, 54 struct ena_com_rx_ctx *, uint16_t *); 55static inline void ena_rx_checksum(struct ena_ring *, struct ena_com_rx_ctx *, 56 struct mbuf *); 57static void ena_tx_csum(struct ena_com_tx_ctx *, struct mbuf *, bool); 58static int ena_check_and_collapse_mbuf(struct ena_ring *tx_ring, 59 struct mbuf **mbuf); 60static int ena_xmit_mbuf(struct ena_ring *, struct mbuf **); 61static void ena_start_xmit(struct ena_ring *); 62 63/********************************************************************* 64 * Global functions 65 *********************************************************************/ 66 67void 68ena_cleanup(void *arg, int pending) 69{ 70 struct ena_que *que = arg; 71 struct ena_adapter *adapter = que->adapter; 72 if_t ifp = adapter->ifp; 73 struct ena_ring *tx_ring; 74 struct ena_ring *rx_ring; 75 struct ena_com_io_cq *io_cq; 76 struct ena_eth_io_intr_reg intr_reg; 77 int qid, ena_qid; 78 int txc, rxc, i; 79 80 if (unlikely((if_getdrvflags(ifp) & IFF_DRV_RUNNING) == 0)) 81 return; 82 83 ena_log_io(adapter->pdev, DBG, "MSI-X TX/RX routine\n"); 84 85 tx_ring = que->tx_ring; 86 rx_ring = que->rx_ring; 87 qid = que->id; 88 ena_qid = ENA_IO_TXQ_IDX(qid); 89 io_cq = &adapter->ena_dev->io_cq_queues[ena_qid]; 90 91 atomic_store_8(&tx_ring->first_interrupt, 1); 92 atomic_store_8(&rx_ring->first_interrupt, 1); 93 94 for (i = 0; i < ENA_CLEAN_BUDGET; ++i) { 95 rxc = ena_rx_cleanup(rx_ring); 96 txc = ena_tx_cleanup(tx_ring); 97 98 if (unlikely((if_getdrvflags(ifp) & IFF_DRV_RUNNING) == 0)) 99 return; 100 101 if ((txc != ENA_TX_BUDGET) && (rxc != ENA_RX_BUDGET)) 102 break; 103 } 104 105 /* Signal that work is done and unmask interrupt */ 106 ena_com_update_intr_reg(&intr_reg, ENA_RX_IRQ_INTERVAL, 107 ENA_TX_IRQ_INTERVAL, true, false); 108 counter_u64_add(tx_ring->tx_stats.unmask_interrupt_num, 1); 109 ena_com_unmask_intr(io_cq, &intr_reg); 110} 111 112void 113ena_deferred_mq_start(void *arg, int pending) 114{ 115 struct ena_ring *tx_ring = (struct ena_ring *)arg; 116 if_t ifp = tx_ring->adapter->ifp; 117 118 while (!drbr_empty(ifp, tx_ring->br) && tx_ring->running && 119 (if_getdrvflags(ifp) & IFF_DRV_RUNNING) != 0) { 120 ENA_RING_MTX_LOCK(tx_ring); 121 ena_start_xmit(tx_ring); 122 ENA_RING_MTX_UNLOCK(tx_ring); 123 } 124} 125 126int 127ena_mq_start(if_t ifp, struct mbuf *m) 128{ 129 struct ena_adapter *adapter = if_getsoftc(ifp); 130 struct ena_ring *tx_ring; 131 int ret, is_drbr_empty; 132 uint32_t i; 133#ifdef RSS 134 uint32_t bucket_id; 135#endif 136 137 if (unlikely((if_getdrvflags(adapter->ifp) & IFF_DRV_RUNNING) == 0)) 138 return (ENODEV); 139 140 /* Which queue to use */ 141 /* 142 * If everything is setup correctly, it should be the 143 * same bucket that the current CPU we're on is. 144 * It should improve performance. 145 */ 146 if (M_HASHTYPE_GET(m) != M_HASHTYPE_NONE) { 147#ifdef RSS 148 if (rss_hash2bucket(m->m_pkthdr.flowid, M_HASHTYPE_GET(m), 149 &bucket_id) == 0) 150 i = bucket_id % adapter->num_io_queues; 151 else 152#endif 153 i = m->m_pkthdr.flowid % adapter->num_io_queues; 154 } else { 155 i = curcpu % adapter->num_io_queues; 156 } 157 tx_ring = &adapter->tx_ring[i]; 158 159 /* Check if drbr is empty before putting packet */ 160 is_drbr_empty = drbr_empty(ifp, tx_ring->br); 161 ret = drbr_enqueue(ifp, tx_ring->br, m); 162 if (unlikely(ret != 0)) { 163 taskqueue_enqueue(tx_ring->enqueue_tq, &tx_ring->enqueue_task); 164 return (ret); 165 } 166 167 if (is_drbr_empty && (ENA_RING_MTX_TRYLOCK(tx_ring) != 0)) { 168 ena_start_xmit(tx_ring); 169 ENA_RING_MTX_UNLOCK(tx_ring); 170 } else { 171 taskqueue_enqueue(tx_ring->enqueue_tq, &tx_ring->enqueue_task); 172 } 173 174 return (0); 175} 176 177void 178ena_qflush(if_t ifp) 179{ 180 struct ena_adapter *adapter = if_getsoftc(ifp); 181 struct ena_ring *tx_ring = adapter->tx_ring; 182 int i; 183 184 for (i = 0; i < adapter->num_io_queues; ++i, ++tx_ring) 185 if (!drbr_empty(ifp, tx_ring->br)) { 186 ENA_RING_MTX_LOCK(tx_ring); 187 drbr_flush(ifp, tx_ring->br); 188 ENA_RING_MTX_UNLOCK(tx_ring); 189 } 190 191 if_qflush(ifp); 192} 193 194/********************************************************************* 195 * Static functions 196 *********************************************************************/ 197 198static inline int 199ena_get_tx_req_id(struct ena_ring *tx_ring, struct ena_com_io_cq *io_cq, 200 uint16_t *req_id) 201{ 202 struct ena_adapter *adapter = tx_ring->adapter; 203 int rc; 204 205 rc = ena_com_tx_comp_req_id_get(io_cq, req_id); 206 if (rc == ENA_COM_TRY_AGAIN) 207 return (EAGAIN); 208 209 if (unlikely(rc != 0)) { 210 ena_log(adapter->pdev, ERR, "Invalid req_id %hu in qid %hu\n", 211 *req_id, tx_ring->qid); 212 counter_u64_add(tx_ring->tx_stats.bad_req_id, 1); 213 goto err; 214 } 215 216 if (tx_ring->tx_buffer_info[*req_id].mbuf != NULL) 217 return (0); 218 219 ena_log(adapter->pdev, ERR, 220 "tx_info doesn't have valid mbuf. req_id %hu qid %hu\n", 221 *req_id, tx_ring->qid); 222err: 223 ena_trigger_reset(adapter, ENA_REGS_RESET_INV_TX_REQ_ID); 224 225 return (EFAULT); 226} 227 228/** 229 * ena_tx_cleanup - clear sent packets and corresponding descriptors 230 * @tx_ring: ring for which we want to clean packets 231 * 232 * Once packets are sent, we ask the device in a loop for no longer used 233 * descriptors. We find the related mbuf chain in a map (index in an array) 234 * and free it, then update ring state. 235 * This is performed in "endless" loop, updating ring pointers every 236 * TX_COMMIT. The first check of free descriptor is performed before the actual 237 * loop, then repeated at the loop end. 238 **/ 239static int 240ena_tx_cleanup(struct ena_ring *tx_ring) 241{ 242 struct ena_adapter *adapter; 243 struct ena_com_io_cq *io_cq; 244 uint16_t next_to_clean; 245 uint16_t req_id; 246 uint16_t ena_qid; 247 unsigned int total_done = 0; 248 int rc; 249 int commit = ENA_TX_COMMIT; 250 int budget = ENA_TX_BUDGET; 251 int work_done; 252 bool above_thresh; 253 254 adapter = tx_ring->que->adapter; 255 ena_qid = ENA_IO_TXQ_IDX(tx_ring->que->id); 256 io_cq = &adapter->ena_dev->io_cq_queues[ena_qid]; 257 next_to_clean = tx_ring->next_to_clean; 258 259#ifdef DEV_NETMAP 260 if (netmap_tx_irq(adapter->ifp, tx_ring->qid) != NM_IRQ_PASS) 261 return (0); 262#endif /* DEV_NETMAP */ 263 264 do { 265 struct ena_tx_buffer *tx_info; 266 struct mbuf *mbuf; 267 268 rc = ena_get_tx_req_id(tx_ring, io_cq, &req_id); 269 if (unlikely(rc != 0)) 270 break; 271 272 tx_info = &tx_ring->tx_buffer_info[req_id]; 273 274 mbuf = tx_info->mbuf; 275 276 tx_info->mbuf = NULL; 277 bintime_clear(&tx_info->timestamp); 278 279 bus_dmamap_sync(adapter->tx_buf_tag, tx_info->dmamap, 280 BUS_DMASYNC_POSTWRITE); 281 bus_dmamap_unload(adapter->tx_buf_tag, tx_info->dmamap); 282 283 ena_log_io(adapter->pdev, DBG, "tx: q %d mbuf %p completed\n", 284 tx_ring->qid, mbuf); 285 286 m_freem(mbuf); 287 288 total_done += tx_info->tx_descs; 289 290 tx_ring->free_tx_ids[next_to_clean] = req_id; 291 next_to_clean = ENA_TX_RING_IDX_NEXT(next_to_clean, 292 tx_ring->ring_size); 293 294 if (unlikely(--commit == 0)) { 295 commit = ENA_TX_COMMIT; 296 /* update ring state every ENA_TX_COMMIT descriptor */ 297 tx_ring->next_to_clean = next_to_clean; 298 ena_com_comp_ack( 299 &adapter->ena_dev->io_sq_queues[ena_qid], 300 total_done); 301 total_done = 0; 302 } 303 } while (likely(--budget)); 304 305 work_done = ENA_TX_BUDGET - budget; 306 307 ena_log_io(adapter->pdev, DBG, "tx: q %d done. total pkts: %d\n", 308 tx_ring->qid, work_done); 309 310 /* If there is still something to commit update ring state */ 311 if (likely(commit != ENA_TX_COMMIT)) { 312 tx_ring->next_to_clean = next_to_clean; 313 ena_com_comp_ack(&adapter->ena_dev->io_sq_queues[ena_qid], 314 total_done); 315 } 316 317 /* 318 * Need to make the rings circular update visible to 319 * ena_xmit_mbuf() before checking for tx_ring->running. 320 */ 321 mb(); 322 323 above_thresh = ena_com_sq_have_enough_space(tx_ring->ena_com_io_sq, 324 ENA_TX_RESUME_THRESH); 325 if (unlikely(!tx_ring->running && above_thresh)) { 326 ENA_RING_MTX_LOCK(tx_ring); 327 above_thresh = ena_com_sq_have_enough_space( 328 tx_ring->ena_com_io_sq, ENA_TX_RESUME_THRESH); 329 if (!tx_ring->running && above_thresh) { 330 tx_ring->running = true; 331 counter_u64_add(tx_ring->tx_stats.queue_wakeup, 1); 332 taskqueue_enqueue(tx_ring->enqueue_tq, 333 &tx_ring->enqueue_task); 334 } 335 ENA_RING_MTX_UNLOCK(tx_ring); 336 } 337 338 tx_ring->tx_last_cleanup_ticks = ticks; 339 340 return (work_done); 341} 342 343static void 344ena_rx_hash_mbuf(struct ena_ring *rx_ring, struct ena_com_rx_ctx *ena_rx_ctx, 345 struct mbuf *mbuf) 346{ 347 struct ena_adapter *adapter = rx_ring->adapter; 348 349 if (likely(ENA_FLAG_ISSET(ENA_FLAG_RSS_ACTIVE, adapter))) { 350 mbuf->m_pkthdr.flowid = ena_rx_ctx->hash; 351 352#ifdef RSS 353 /* 354 * Hardware and software RSS are in agreement only when both are 355 * configured to Toeplitz algorithm. This driver configures 356 * that algorithm only when software RSS is enabled and uses it. 357 */ 358 if (adapter->ena_dev->rss.hash_func != ENA_ADMIN_TOEPLITZ && 359 ena_rx_ctx->l3_proto != ENA_ETH_IO_L3_PROTO_UNKNOWN) { 360 M_HASHTYPE_SET(mbuf, M_HASHTYPE_OPAQUE_HASH); 361 return; 362 } 363#endif 364 365 if (ena_rx_ctx->frag && 366 (ena_rx_ctx->l3_proto != ENA_ETH_IO_L3_PROTO_UNKNOWN)) { 367 M_HASHTYPE_SET(mbuf, M_HASHTYPE_OPAQUE_HASH); 368 return; 369 } 370 371 switch (ena_rx_ctx->l3_proto) { 372 case ENA_ETH_IO_L3_PROTO_IPV4: 373 switch (ena_rx_ctx->l4_proto) { 374 case ENA_ETH_IO_L4_PROTO_TCP: 375 M_HASHTYPE_SET(mbuf, M_HASHTYPE_RSS_TCP_IPV4); 376 break; 377 case ENA_ETH_IO_L4_PROTO_UDP: 378 M_HASHTYPE_SET(mbuf, M_HASHTYPE_RSS_UDP_IPV4); 379 break; 380 default: 381 M_HASHTYPE_SET(mbuf, M_HASHTYPE_RSS_IPV4); 382 } 383 break; 384 case ENA_ETH_IO_L3_PROTO_IPV6: 385 switch (ena_rx_ctx->l4_proto) { 386 case ENA_ETH_IO_L4_PROTO_TCP: 387 M_HASHTYPE_SET(mbuf, M_HASHTYPE_RSS_TCP_IPV6); 388 break; 389 case ENA_ETH_IO_L4_PROTO_UDP: 390 M_HASHTYPE_SET(mbuf, M_HASHTYPE_RSS_UDP_IPV6); 391 break; 392 default: 393 M_HASHTYPE_SET(mbuf, M_HASHTYPE_RSS_IPV6); 394 } 395 break; 396 case ENA_ETH_IO_L3_PROTO_UNKNOWN: 397 M_HASHTYPE_SET(mbuf, M_HASHTYPE_NONE); 398 break; 399 default: 400 M_HASHTYPE_SET(mbuf, M_HASHTYPE_OPAQUE_HASH); 401 } 402 } else { 403 mbuf->m_pkthdr.flowid = rx_ring->qid; 404 M_HASHTYPE_SET(mbuf, M_HASHTYPE_NONE); 405 } 406} 407 408/** 409 * ena_rx_mbuf - assemble mbuf from descriptors 410 * @rx_ring: ring for which we want to clean packets 411 * @ena_bufs: buffer info 412 * @ena_rx_ctx: metadata for this packet(s) 413 * @next_to_clean: ring pointer, will be updated only upon success 414 * 415 **/ 416static struct mbuf * 417ena_rx_mbuf(struct ena_ring *rx_ring, struct ena_com_rx_buf_info *ena_bufs, 418 struct ena_com_rx_ctx *ena_rx_ctx, uint16_t *next_to_clean) 419{ 420 struct mbuf *mbuf; 421 struct ena_rx_buffer *rx_info; 422 struct ena_adapter *adapter; 423 device_t pdev; 424 unsigned int descs = ena_rx_ctx->descs; 425 uint16_t ntc, len, req_id, buf = 0; 426 427 ntc = *next_to_clean; 428 adapter = rx_ring->adapter; 429 pdev = adapter->pdev; 430 431 len = ena_bufs[buf].len; 432 req_id = ena_bufs[buf].req_id; 433 rx_info = &rx_ring->rx_buffer_info[req_id]; 434 if (unlikely(rx_info->mbuf == NULL)) { 435 ena_log(pdev, ERR, "NULL mbuf in rx_info"); 436 return (NULL); 437 } 438 439 ena_log_io(pdev, DBG, "rx_info %p, mbuf %p, paddr %jx\n", rx_info, 440 rx_info->mbuf, (uintmax_t)rx_info->ena_buf.paddr); 441 442 bus_dmamap_sync(adapter->rx_buf_tag, rx_info->map, 443 BUS_DMASYNC_POSTREAD); 444 mbuf = rx_info->mbuf; 445 mbuf->m_flags |= M_PKTHDR; 446 mbuf->m_pkthdr.len = len; 447 mbuf->m_len = len; 448 /* Only for the first segment the data starts at specific offset */ 449 mbuf->m_data = mtodo(mbuf, ena_rx_ctx->pkt_offset); 450 ena_log_io(pdev, DBG, "Mbuf data offset=%u\n", ena_rx_ctx->pkt_offset); 451 mbuf->m_pkthdr.rcvif = rx_ring->que->adapter->ifp; 452 453 /* Fill mbuf with hash key and it's interpretation for optimization */ 454 ena_rx_hash_mbuf(rx_ring, ena_rx_ctx, mbuf); 455 456 ena_log_io(pdev, DBG, "rx mbuf 0x%p, flags=0x%x, len: %d\n", mbuf, 457 mbuf->m_flags, mbuf->m_pkthdr.len); 458 459 /* DMA address is not needed anymore, unmap it */ 460 bus_dmamap_unload(rx_ring->adapter->rx_buf_tag, rx_info->map); 461 462 rx_info->mbuf = NULL; 463 rx_ring->free_rx_ids[ntc] = req_id; 464 ntc = ENA_RX_RING_IDX_NEXT(ntc, rx_ring->ring_size); 465 466 /* 467 * While we have more than 1 descriptors for one rcvd packet, append 468 * other mbufs to the main one 469 */ 470 while (--descs) { 471 ++buf; 472 len = ena_bufs[buf].len; 473 req_id = ena_bufs[buf].req_id; 474 rx_info = &rx_ring->rx_buffer_info[req_id]; 475 476 if (unlikely(rx_info->mbuf == NULL)) { 477 ena_log(pdev, ERR, "NULL mbuf in rx_info"); 478 /* 479 * If one of the required mbufs was not allocated yet, 480 * we can break there. 481 * All earlier used descriptors will be reallocated 482 * later and not used mbufs can be reused. 483 * The next_to_clean pointer will not be updated in case 484 * of an error, so caller should advance it manually 485 * in error handling routine to keep it up to date 486 * with hw ring. 487 */ 488 m_freem(mbuf); 489 return (NULL); 490 } 491 492 bus_dmamap_sync(adapter->rx_buf_tag, rx_info->map, 493 BUS_DMASYNC_POSTREAD); 494 if (unlikely(m_append(mbuf, len, rx_info->mbuf->m_data) == 0)) { 495 counter_u64_add(rx_ring->rx_stats.mbuf_alloc_fail, 1); 496 ena_log_io(pdev, WARN, "Failed to append Rx mbuf %p\n", 497 mbuf); 498 } 499 500 ena_log_io(pdev, DBG, "rx mbuf updated. len %d\n", 501 mbuf->m_pkthdr.len); 502 503 /* Free already appended mbuf, it won't be useful anymore */ 504 bus_dmamap_unload(rx_ring->adapter->rx_buf_tag, rx_info->map); 505 m_freem(rx_info->mbuf); 506 rx_info->mbuf = NULL; 507 508 rx_ring->free_rx_ids[ntc] = req_id; 509 ntc = ENA_RX_RING_IDX_NEXT(ntc, rx_ring->ring_size); 510 } 511 512 *next_to_clean = ntc; 513 514 return (mbuf); 515} 516 517/** 518 * ena_rx_checksum - indicate in mbuf if hw indicated a good cksum 519 **/ 520static inline void 521ena_rx_checksum(struct ena_ring *rx_ring, struct ena_com_rx_ctx *ena_rx_ctx, 522 struct mbuf *mbuf) 523{ 524 device_t pdev = rx_ring->adapter->pdev; 525 526 /* if IP and error */ 527 if (unlikely((ena_rx_ctx->l3_proto == ENA_ETH_IO_L3_PROTO_IPV4) && 528 ena_rx_ctx->l3_csum_err)) { 529 /* ipv4 checksum error */ 530 mbuf->m_pkthdr.csum_flags = 0; 531 counter_u64_add(rx_ring->rx_stats.csum_bad, 1); 532 ena_log_io(pdev, DBG, "RX IPv4 header checksum error\n"); 533 return; 534 } 535 536 /* if TCP/UDP */ 537 if ((ena_rx_ctx->l4_proto == ENA_ETH_IO_L4_PROTO_TCP) || 538 (ena_rx_ctx->l4_proto == ENA_ETH_IO_L4_PROTO_UDP)) { 539 if (ena_rx_ctx->l4_csum_err) { 540 /* TCP/UDP checksum error */ 541 mbuf->m_pkthdr.csum_flags = 0; 542 counter_u64_add(rx_ring->rx_stats.csum_bad, 1); 543 ena_log_io(pdev, DBG, "RX L4 checksum error\n"); 544 } else { 545 mbuf->m_pkthdr.csum_flags = CSUM_IP_CHECKED; 546 mbuf->m_pkthdr.csum_flags |= CSUM_IP_VALID; 547 counter_u64_add(rx_ring->rx_stats.csum_good, 1); 548 } 549 } 550} 551 552/** 553 * ena_rx_cleanup - handle rx irq 554 * @arg: ring for which irq is being handled 555 **/ 556static int 557ena_rx_cleanup(struct ena_ring *rx_ring) 558{ 559 struct ena_adapter *adapter; 560 device_t pdev; 561 struct mbuf *mbuf; 562 struct ena_com_rx_ctx ena_rx_ctx; 563 struct ena_com_io_cq *io_cq; 564 struct ena_com_io_sq *io_sq; 565 enum ena_regs_reset_reason_types reset_reason; 566 if_t ifp; 567 uint16_t ena_qid; 568 uint16_t next_to_clean; 569 uint32_t refill_required; 570 uint32_t refill_threshold; 571 uint32_t do_if_input = 0; 572 unsigned int qid; 573 int rc, i; 574 int budget = ENA_RX_BUDGET; 575#ifdef DEV_NETMAP 576 int done; 577#endif /* DEV_NETMAP */ 578 579 adapter = rx_ring->que->adapter; 580 pdev = adapter->pdev; 581 ifp = adapter->ifp; 582 qid = rx_ring->que->id; 583 ena_qid = ENA_IO_RXQ_IDX(qid); 584 io_cq = &adapter->ena_dev->io_cq_queues[ena_qid]; 585 io_sq = &adapter->ena_dev->io_sq_queues[ena_qid]; 586 next_to_clean = rx_ring->next_to_clean; 587 588#ifdef DEV_NETMAP 589 if (netmap_rx_irq(adapter->ifp, rx_ring->qid, &done) != NM_IRQ_PASS) 590 return (0); 591#endif /* DEV_NETMAP */ 592 593 ena_log_io(pdev, DBG, "rx: qid %d\n", qid); 594 595 do { 596 ena_rx_ctx.ena_bufs = rx_ring->ena_bufs; 597 ena_rx_ctx.max_bufs = adapter->max_rx_sgl_size; 598 ena_rx_ctx.descs = 0; 599 ena_rx_ctx.pkt_offset = 0; 600 601 bus_dmamap_sync(io_cq->cdesc_addr.mem_handle.tag, 602 io_cq->cdesc_addr.mem_handle.map, BUS_DMASYNC_POSTREAD); 603 rc = ena_com_rx_pkt(io_cq, io_sq, &ena_rx_ctx); 604 if (unlikely(rc != 0)) { 605 if (rc == ENA_COM_NO_SPACE) { 606 counter_u64_add(rx_ring->rx_stats.bad_desc_num, 607 1); 608 reset_reason = ENA_REGS_RESET_TOO_MANY_RX_DESCS; 609 } else { 610 counter_u64_add(rx_ring->rx_stats.bad_req_id, 611 1); 612 reset_reason = ENA_REGS_RESET_INV_RX_REQ_ID; 613 } 614 ena_trigger_reset(adapter, reset_reason); 615 return (0); 616 } 617 618 if (unlikely(ena_rx_ctx.descs == 0)) 619 break; 620 621 ena_log_io(pdev, DBG, 622 "rx: q %d got packet from ena. descs #: %d l3 proto %d l4 proto %d hash: %x\n", 623 rx_ring->qid, ena_rx_ctx.descs, ena_rx_ctx.l3_proto, 624 ena_rx_ctx.l4_proto, ena_rx_ctx.hash); 625 626 /* Receive mbuf from the ring */ 627 mbuf = ena_rx_mbuf(rx_ring, rx_ring->ena_bufs, &ena_rx_ctx, 628 &next_to_clean); 629 bus_dmamap_sync(io_cq->cdesc_addr.mem_handle.tag, 630 io_cq->cdesc_addr.mem_handle.map, BUS_DMASYNC_PREREAD); 631 /* Exit if we failed to retrieve a buffer */ 632 if (unlikely(mbuf == NULL)) { 633 for (i = 0; i < ena_rx_ctx.descs; ++i) { 634 rx_ring->free_rx_ids[next_to_clean] = 635 rx_ring->ena_bufs[i].req_id; 636 next_to_clean = ENA_RX_RING_IDX_NEXT( 637 next_to_clean, rx_ring->ring_size); 638 } 639 break; 640 } 641 642 if (((if_getcapenable(ifp) & IFCAP_RXCSUM) != 0) || 643 ((if_getcapenable(ifp) & IFCAP_RXCSUM_IPV6) != 0)) { 644 ena_rx_checksum(rx_ring, &ena_rx_ctx, mbuf); 645 } 646 647 counter_enter(); 648 counter_u64_add_protected(rx_ring->rx_stats.bytes, 649 mbuf->m_pkthdr.len); 650 counter_u64_add_protected(adapter->hw_stats.rx_bytes, 651 mbuf->m_pkthdr.len); 652 counter_exit(); 653 /* 654 * LRO is only for IP/TCP packets and TCP checksum of the packet 655 * should be computed by hardware. 656 */ 657 do_if_input = 1; 658 if (((if_getcapenable(ifp) & IFCAP_LRO) != 0) && 659 ((mbuf->m_pkthdr.csum_flags & CSUM_IP_VALID) != 0) && 660 (ena_rx_ctx.l4_proto == ENA_ETH_IO_L4_PROTO_TCP)) { 661 /* 662 * Send to the stack if: 663 * - LRO not enabled, or 664 * - no LRO resources, or 665 * - lro enqueue fails 666 */ 667 if ((rx_ring->lro.lro_cnt != 0) && 668 (tcp_lro_rx(&rx_ring->lro, mbuf, 0) == 0)) 669 do_if_input = 0; 670 } 671 if (do_if_input != 0) { 672 ena_log_io(pdev, DBG, 673 "calling if_input() with mbuf %p\n", mbuf); 674 if_input(ifp, mbuf); 675 } 676 677 counter_enter(); 678 counter_u64_add_protected(rx_ring->rx_stats.cnt, 1); 679 counter_u64_add_protected(adapter->hw_stats.rx_packets, 1); 680 counter_exit(); 681 } while (--budget); 682 683 rx_ring->next_to_clean = next_to_clean; 684 685 refill_required = ena_com_free_q_entries(io_sq); 686 refill_threshold = min_t(int, 687 rx_ring->ring_size / ENA_RX_REFILL_THRESH_DIVIDER, 688 ENA_RX_REFILL_THRESH_PACKET); 689 690 if (refill_required > refill_threshold) { 691 ena_refill_rx_bufs(rx_ring, refill_required); 692 } 693 694 tcp_lro_flush_all(&rx_ring->lro); 695 696 return (ENA_RX_BUDGET - budget); 697} 698 699static void 700ena_tx_csum(struct ena_com_tx_ctx *ena_tx_ctx, struct mbuf *mbuf, 701 bool disable_meta_caching) 702{ 703 struct ena_com_tx_meta *ena_meta; 704 struct ether_vlan_header *eh; 705 struct mbuf *mbuf_next; 706 u32 mss; 707 bool offload; 708 uint16_t etype; 709 int ehdrlen; 710 struct ip *ip; 711 int ipproto; 712 int iphlen; 713 struct tcphdr *th; 714 int offset; 715 716 offload = false; 717 ena_meta = &ena_tx_ctx->ena_meta; 718 mss = mbuf->m_pkthdr.tso_segsz; 719 720 if (mss != 0) 721 offload = true; 722 723 if ((mbuf->m_pkthdr.csum_flags & CSUM_TSO) != 0) 724 offload = true; 725 726 if ((mbuf->m_pkthdr.csum_flags & CSUM_OFFLOAD) != 0) 727 offload = true; 728 729 if ((mbuf->m_pkthdr.csum_flags & CSUM6_OFFLOAD) != 0) 730 offload = true; 731 732 if (!offload) { 733 if (disable_meta_caching) { 734 memset(ena_meta, 0, sizeof(*ena_meta)); 735 ena_tx_ctx->meta_valid = 1; 736 } else { 737 ena_tx_ctx->meta_valid = 0; 738 } 739 return; 740 } 741 742 /* Determine where frame payload starts. */ 743 eh = mtod(mbuf, struct ether_vlan_header *); 744 if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) { 745 etype = ntohs(eh->evl_proto); 746 ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN; 747 } else { 748 etype = ntohs(eh->evl_encap_proto); 749 ehdrlen = ETHER_HDR_LEN; 750 } 751 752 mbuf_next = m_getptr(mbuf, ehdrlen, &offset); 753 754 switch (etype) { 755 case ETHERTYPE_IP: 756 ip = (struct ip *)(mtodo(mbuf_next, offset)); 757 iphlen = ip->ip_hl << 2; 758 ipproto = ip->ip_p; 759 ena_tx_ctx->l3_proto = ENA_ETH_IO_L3_PROTO_IPV4; 760 if ((ip->ip_off & htons(IP_DF)) != 0) 761 ena_tx_ctx->df = 1; 762 break; 763 case ETHERTYPE_IPV6: 764 ena_tx_ctx->l3_proto = ENA_ETH_IO_L3_PROTO_IPV6; 765 iphlen = ip6_lasthdr(mbuf, ehdrlen, IPPROTO_IPV6, &ipproto); 766 iphlen -= ehdrlen; 767 ena_tx_ctx->df = 1; 768 break; 769 default: 770 iphlen = 0; 771 ipproto = 0; 772 break; 773 } 774 775 mbuf_next = m_getptr(mbuf, iphlen + ehdrlen, &offset); 776 th = (struct tcphdr *)(mtodo(mbuf_next, offset)); 777 778 if ((mbuf->m_pkthdr.csum_flags & CSUM_IP) != 0) { 779 ena_tx_ctx->l3_csum_enable = 1; 780 } 781 if ((mbuf->m_pkthdr.csum_flags & CSUM_TSO) != 0) { 782 ena_tx_ctx->tso_enable = 1; 783 ena_meta->l4_hdr_len = (th->th_off); 784 } 785 786 if (ipproto == IPPROTO_TCP) { 787 ena_tx_ctx->l4_proto = ENA_ETH_IO_L4_PROTO_TCP; 788 if ((mbuf->m_pkthdr.csum_flags & 789 (CSUM_IP_TCP | CSUM_IP6_TCP)) != 0) 790 ena_tx_ctx->l4_csum_enable = 1; 791 else 792 ena_tx_ctx->l4_csum_enable = 0; 793 } else if (ipproto == IPPROTO_UDP) { 794 ena_tx_ctx->l4_proto = ENA_ETH_IO_L4_PROTO_UDP; 795 if ((mbuf->m_pkthdr.csum_flags & 796 (CSUM_IP_UDP | CSUM_IP6_UDP)) != 0) 797 ena_tx_ctx->l4_csum_enable = 1; 798 else 799 ena_tx_ctx->l4_csum_enable = 0; 800 } else { 801 ena_tx_ctx->l4_proto = ENA_ETH_IO_L4_PROTO_UNKNOWN; 802 ena_tx_ctx->l4_csum_enable = 0; 803 } 804 805 ena_meta->mss = mss; 806 ena_meta->l3_hdr_len = iphlen; 807 ena_meta->l3_hdr_offset = ehdrlen; 808 ena_tx_ctx->meta_valid = 1; 809} 810 811static int 812ena_check_and_collapse_mbuf(struct ena_ring *tx_ring, struct mbuf **mbuf) 813{ 814 struct ena_adapter *adapter; 815 struct mbuf *collapsed_mbuf; 816 int num_frags; 817 818 adapter = tx_ring->adapter; 819 num_frags = ena_mbuf_count(*mbuf); 820 821 /* One segment must be reserved for configuration descriptor. */ 822 if (num_frags < adapter->max_tx_sgl_size) 823 return (0); 824 825 if ((num_frags == adapter->max_tx_sgl_size) && 826 ((*mbuf)->m_pkthdr.len < tx_ring->tx_max_header_size)) 827 return (0); 828 829 counter_u64_add(tx_ring->tx_stats.collapse, 1); 830 831 collapsed_mbuf = m_collapse(*mbuf, M_NOWAIT, 832 adapter->max_tx_sgl_size - 1); 833 if (unlikely(collapsed_mbuf == NULL)) { 834 counter_u64_add(tx_ring->tx_stats.collapse_err, 1); 835 return (ENOMEM); 836 } 837 838 /* If mbuf was collapsed succesfully, original mbuf is released. */ 839 *mbuf = collapsed_mbuf; 840 841 return (0); 842} 843 844static int 845ena_tx_map_mbuf(struct ena_ring *tx_ring, struct ena_tx_buffer *tx_info, 846 struct mbuf *mbuf, void **push_hdr, u16 *header_len) 847{ 848 struct ena_adapter *adapter = tx_ring->adapter; 849 struct ena_com_buf *ena_buf; 850 bus_dma_segment_t segs[ENA_BUS_DMA_SEGS]; 851 size_t iseg = 0; 852 uint32_t mbuf_head_len; 853 uint16_t offset; 854 int rc, nsegs; 855 856 mbuf_head_len = mbuf->m_len; 857 tx_info->mbuf = mbuf; 858 ena_buf = tx_info->bufs; 859 860 /* 861 * For easier maintaining of the DMA map, map the whole mbuf even if 862 * the LLQ is used. The descriptors will be filled using the segments. 863 */ 864 rc = bus_dmamap_load_mbuf_sg(adapter->tx_buf_tag, 865 tx_info->dmamap, mbuf, segs, &nsegs, BUS_DMA_NOWAIT); 866 if (unlikely((rc != 0) || (nsegs == 0))) { 867 ena_log_io(adapter->pdev, WARN, 868 "dmamap load failed! err: %d nsegs: %d\n", rc, nsegs); 869 goto dma_error; 870 } 871 872 if (tx_ring->tx_mem_queue_type == ENA_ADMIN_PLACEMENT_POLICY_DEV) { 873 /* 874 * When the device is LLQ mode, the driver will copy 875 * the header into the device memory space. 876 * the ena_com layer assumes the header is in a linear 877 * memory space. 878 * This assumption might be wrong since part of the header 879 * can be in the fragmented buffers. 880 * First check if header fits in the mbuf. If not, copy it to 881 * separate buffer that will be holding linearized data. 882 */ 883 *header_len = min_t(uint32_t, mbuf->m_pkthdr.len, 884 tx_ring->tx_max_header_size); 885 886 /* If header is in linear space, just point into mbuf's data. */ 887 if (likely(*header_len <= mbuf_head_len)) { 888 *push_hdr = mbuf->m_data; 889 /* 890 * Otherwise, copy whole portion of header from multiple 891 * mbufs to intermediate buffer. 892 */ 893 } else { 894 m_copydata(mbuf, 0, *header_len, 895 tx_ring->push_buf_intermediate_buf); 896 *push_hdr = tx_ring->push_buf_intermediate_buf; 897 898 counter_u64_add(tx_ring->tx_stats.llq_buffer_copy, 1); 899 } 900 901 ena_log_io(adapter->pdev, DBG, 902 "mbuf: %p header_buf->vaddr: %p push_len: %d\n", 903 mbuf, *push_hdr, *header_len); 904 905 /* If packet is fitted in LLQ header, no need for DMA segments. */ 906 if (mbuf->m_pkthdr.len <= tx_ring->tx_max_header_size) { 907 return (0); 908 } else { 909 offset = tx_ring->tx_max_header_size; 910 /* 911 * As Header part is mapped to LLQ header, we can skip 912 * it and just map the residuum of the mbuf to DMA 913 * Segments. 914 */ 915 while (offset > 0) { 916 if (offset >= segs[iseg].ds_len) { 917 offset -= segs[iseg].ds_len; 918 } else { 919 ena_buf->paddr = segs[iseg].ds_addr + 920 offset; 921 ena_buf->len = segs[iseg].ds_len - 922 offset; 923 ena_buf++; 924 tx_info->num_of_bufs++; 925 offset = 0; 926 } 927 iseg++; 928 } 929 } 930 } else { 931 *push_hdr = NULL; 932 /* 933 * header_len is just a hint for the device. Because FreeBSD is 934 * not giving us information about packet header length and it 935 * is not guaranteed that all packet headers will be in the 1st 936 * mbuf, setting header_len to 0 is making the device ignore 937 * this value and resolve header on it's own. 938 */ 939 *header_len = 0; 940 } 941 942 /* Map rest of the mbuf */ 943 while (iseg < nsegs) { 944 ena_buf->paddr = segs[iseg].ds_addr; 945 ena_buf->len = segs[iseg].ds_len; 946 ena_buf++; 947 iseg++; 948 tx_info->num_of_bufs++; 949 } 950 951 return (0); 952 953dma_error: 954 counter_u64_add(tx_ring->tx_stats.dma_mapping_err, 1); 955 tx_info->mbuf = NULL; 956 return (rc); 957} 958 959static int 960ena_xmit_mbuf(struct ena_ring *tx_ring, struct mbuf **mbuf) 961{ 962 struct ena_adapter *adapter; 963 device_t pdev; 964 struct ena_tx_buffer *tx_info; 965 struct ena_com_tx_ctx ena_tx_ctx; 966 struct ena_com_dev *ena_dev; 967 struct ena_com_io_sq *io_sq; 968 void *push_hdr; 969 uint16_t next_to_use; 970 uint16_t req_id; 971 uint16_t ena_qid; 972 uint16_t header_len; 973 int rc; 974 int nb_hw_desc; 975 976 ena_qid = ENA_IO_TXQ_IDX(tx_ring->que->id); 977 adapter = tx_ring->que->adapter; 978 pdev = adapter->pdev; 979 ena_dev = adapter->ena_dev; 980 io_sq = &ena_dev->io_sq_queues[ena_qid]; 981 982 rc = ena_check_and_collapse_mbuf(tx_ring, mbuf); 983 if (unlikely(rc != 0)) { 984 ena_log_io(pdev, WARN, "Failed to collapse mbuf! err: %d\n", 985 rc); 986 return (rc); 987 } 988 989 ena_log_io(pdev, DBG, "Tx: %d bytes\n", (*mbuf)->m_pkthdr.len); 990 991 next_to_use = tx_ring->next_to_use; 992 req_id = tx_ring->free_tx_ids[next_to_use]; 993 tx_info = &tx_ring->tx_buffer_info[req_id]; 994 tx_info->num_of_bufs = 0; 995 996 ENA_WARN(tx_info->mbuf != NULL, adapter->ena_dev, 997 "mbuf isn't NULL for req_id %d\n", req_id); 998 999 rc = ena_tx_map_mbuf(tx_ring, tx_info, *mbuf, &push_hdr, &header_len); 1000 if (unlikely(rc != 0)) { 1001 ena_log_io(pdev, WARN, "Failed to map TX mbuf\n"); 1002 return (rc); 1003 } 1004 memset(&ena_tx_ctx, 0x0, sizeof(struct ena_com_tx_ctx)); 1005 ena_tx_ctx.ena_bufs = tx_info->bufs; 1006 ena_tx_ctx.push_header = push_hdr; 1007 ena_tx_ctx.num_bufs = tx_info->num_of_bufs; 1008 ena_tx_ctx.req_id = req_id; 1009 ena_tx_ctx.header_len = header_len; 1010 1011 /* Set flags and meta data */ 1012 ena_tx_csum(&ena_tx_ctx, *mbuf, adapter->disable_meta_caching); 1013 1014 if (tx_ring->acum_pkts == ENA_DB_THRESHOLD || 1015 ena_com_is_doorbell_needed(tx_ring->ena_com_io_sq, &ena_tx_ctx)) { 1016 ena_log_io(pdev, DBG, 1017 "llq tx max burst size of queue %d achieved, writing doorbell to send burst\n", 1018 tx_ring->que->id); 1019 ena_ring_tx_doorbell(tx_ring); 1020 } 1021 1022 /* Prepare the packet's descriptors and send them to device */ 1023 rc = ena_com_prepare_tx(io_sq, &ena_tx_ctx, &nb_hw_desc); 1024 if (unlikely(rc != 0)) { 1025 if (likely(rc == ENA_COM_NO_MEM)) { 1026 ena_log_io(pdev, DBG, "tx ring[%d] is out of space\n", 1027 tx_ring->que->id); 1028 } else { 1029 ena_log(pdev, ERR, "failed to prepare tx bufs\n"); 1030 ena_trigger_reset(adapter, 1031 ENA_REGS_RESET_DRIVER_INVALID_STATE); 1032 } 1033 counter_u64_add(tx_ring->tx_stats.prepare_ctx_err, 1); 1034 goto dma_error; 1035 } 1036 1037 counter_enter(); 1038 counter_u64_add_protected(tx_ring->tx_stats.cnt, 1); 1039 counter_u64_add_protected(tx_ring->tx_stats.bytes, 1040 (*mbuf)->m_pkthdr.len); 1041 1042 counter_u64_add_protected(adapter->hw_stats.tx_packets, 1); 1043 counter_u64_add_protected(adapter->hw_stats.tx_bytes, 1044 (*mbuf)->m_pkthdr.len); 1045 counter_exit(); 1046 1047 tx_info->tx_descs = nb_hw_desc; 1048 getbinuptime(&tx_info->timestamp); 1049 tx_info->print_once = true; 1050 1051 tx_ring->next_to_use = ENA_TX_RING_IDX_NEXT(next_to_use, 1052 tx_ring->ring_size); 1053 1054 /* stop the queue when no more space available, the packet can have up 1055 * to sgl_size + 2. one for the meta descriptor and one for header 1056 * (if the header is larger than tx_max_header_size). 1057 */ 1058 if (unlikely(!ena_com_sq_have_enough_space(tx_ring->ena_com_io_sq, 1059 adapter->max_tx_sgl_size + 2))) { 1060 ena_log_io(pdev, DBG, "Stop queue %d\n", tx_ring->que->id); 1061 1062 tx_ring->running = false; 1063 counter_u64_add(tx_ring->tx_stats.queue_stop, 1); 1064 1065 /* There is a rare condition where this function decides to 1066 * stop the queue but meanwhile tx_cleanup() updates 1067 * next_to_completion and terminates. 1068 * The queue will remain stopped forever. 1069 * To solve this issue this function performs mb(), checks 1070 * the wakeup condition and wakes up the queue if needed. 1071 */ 1072 mb(); 1073 1074 if (ena_com_sq_have_enough_space(tx_ring->ena_com_io_sq, 1075 ENA_TX_RESUME_THRESH)) { 1076 tx_ring->running = true; 1077 counter_u64_add(tx_ring->tx_stats.queue_wakeup, 1); 1078 } 1079 } 1080 1081 bus_dmamap_sync(adapter->tx_buf_tag, tx_info->dmamap, 1082 BUS_DMASYNC_PREWRITE); 1083 1084 return (0); 1085 1086dma_error: 1087 tx_info->mbuf = NULL; 1088 bus_dmamap_unload(adapter->tx_buf_tag, tx_info->dmamap); 1089 1090 return (rc); 1091} 1092 1093static void 1094ena_start_xmit(struct ena_ring *tx_ring) 1095{ 1096 struct mbuf *mbuf; 1097 struct ena_adapter *adapter = tx_ring->adapter; 1098 int ret = 0; 1099 1100 ENA_RING_MTX_ASSERT(tx_ring); 1101 1102 if (unlikely((if_getdrvflags(adapter->ifp) & IFF_DRV_RUNNING) == 0)) 1103 return; 1104 1105 if (unlikely(!ENA_FLAG_ISSET(ENA_FLAG_LINK_UP, adapter))) 1106 return; 1107 1108 while ((mbuf = drbr_peek(adapter->ifp, tx_ring->br)) != NULL) { 1109 ena_log_io(adapter->pdev, DBG, 1110 "\ndequeued mbuf %p with flags %#x and header csum flags %#jx\n", 1111 mbuf, mbuf->m_flags, (uint64_t)mbuf->m_pkthdr.csum_flags); 1112 1113 if (unlikely(!tx_ring->running)) { 1114 drbr_putback(adapter->ifp, tx_ring->br, mbuf); 1115 break; 1116 } 1117 1118 if (unlikely((ret = ena_xmit_mbuf(tx_ring, &mbuf)) != 0)) { 1119 if (ret == ENA_COM_NO_MEM) { 1120 drbr_putback(adapter->ifp, tx_ring->br, mbuf); 1121 } else if (ret == ENA_COM_NO_SPACE) { 1122 drbr_putback(adapter->ifp, tx_ring->br, mbuf); 1123 } else { 1124 m_freem(mbuf); 1125 drbr_advance(adapter->ifp, tx_ring->br); 1126 } 1127 1128 break; 1129 } 1130 1131 drbr_advance(adapter->ifp, tx_ring->br); 1132 1133 if (unlikely((if_getdrvflags(adapter->ifp) & IFF_DRV_RUNNING) == 0)) 1134 return; 1135 1136 tx_ring->acum_pkts++; 1137 1138 BPF_MTAP(adapter->ifp, mbuf); 1139 } 1140 1141 if (likely(tx_ring->acum_pkts != 0)) { 1142 /* Trigger the dma engine */ 1143 ena_ring_tx_doorbell(tx_ring); 1144 } 1145 1146 if (unlikely(!tx_ring->running)) 1147 taskqueue_enqueue(tx_ring->que->cleanup_tq, 1148 &tx_ring->que->cleanup_task); 1149} 1150