drivers/net/sfc/sfc_rx.c

   1 /*-
   2  * Copyright (c) 2016 Solarflare Communications Inc.
   3  * All rights reserved.
   4  *
   5  * This software was jointly developed between OKTET Labs (under contract
   6  * for Solarflare) and Solarflare Communications, Inc.
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions are met:
  10  *
  11  * 1. Redistributions of source code must retain the above copyright notice,
  12  *    this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright notice,
  14  *    this list of conditions and the following disclaimer in the documentation
  15  *    and/or other materials provided with the distribution.
  16  *
  17  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
  18  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,
  19  * THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
  20  * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
  21  * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
  22  * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
  23  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
  24  * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
  25  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
  26  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE,
  27  * EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  28  */
  29
  30 #include <rte_mempool.h>
  31
  32 #include "efx.h"
  33
  34 #include "sfc.h"
  35 #include "sfc_debug.h"
  36 #include "sfc_log.h"
  37 #include "sfc_ev.h"
  38 #include "sfc_rx.h"
  39 #include "sfc_tweak.h"
  40
  41 /*
  42  * Maximum number of Rx queue flush attempt in the case of failure or
  43  * flush timeout
  44  */
  45 #define SFC_RX_QFLUSH_ATTEMPTS          (3)
  46
  47 /*
  48  * Time to wait between event queue polling attempts when waiting for Rx
  49  * queue flush done or failed events.
  50  */
  51 #define SFC_RX_QFLUSH_POLL_WAIT_MS      (1)
  52
  53 /*
  54  * Maximum number of event queue polling attempts when waiting for Rx queue
  55  * flush done or failed events. It defines Rx queue flush attempt timeout
  56  * together with SFC_RX_QFLUSH_POLL_WAIT_MS.
  57  */
  58 #define SFC_RX_QFLUSH_POLL_ATTEMPTS     (2000)
  59
  60 void
  61 sfc_rx_qflush_done(struct sfc_rxq *rxq)
  62 {
  63         rxq->state |= SFC_RXQ_FLUSHED;
  64         rxq->state &= ~SFC_RXQ_FLUSHING;
  65 }
  66
  67 void
  68 sfc_rx_qflush_failed(struct sfc_rxq *rxq)
  69 {
  70         rxq->state |= SFC_RXQ_FLUSH_FAILED;
  71         rxq->state &= ~SFC_RXQ_FLUSHING;
  72 }
  73
  74 static void
  75 sfc_rx_qrefill(struct sfc_rxq *rxq)
  76 {
  77         unsigned int free_space;
  78         unsigned int bulks;
  79         void *objs[SFC_RX_REFILL_BULK];
  80         efsys_dma_addr_t addr[RTE_DIM(objs)];
  81         unsigned int added = rxq->added;
  82         unsigned int id;
  83         unsigned int i;
  84         struct sfc_rx_sw_desc *rxd;
  85         struct rte_mbuf *m;
  86         uint8_t port_id = rxq->port_id;
  87
  88         free_space = EFX_RXQ_LIMIT(rxq->ptr_mask + 1) -
  89                 (added - rxq->completed);
  90         bulks = free_space / RTE_DIM(objs);
  91
  92         id = added & rxq->ptr_mask;
  93         while (bulks-- > 0) {
  94                 if (rte_mempool_get_bulk(rxq->refill_mb_pool, objs,
  95                                          RTE_DIM(objs)) < 0) {
  96                         /*
  97                          * It is hardly a safe way to increment counter
  98                          * from different contexts, but all PMDs do it.
  99                          */
 100                         rxq->evq->sa->eth_dev->data->rx_mbuf_alloc_failed +=
 101                                 RTE_DIM(objs);
 102                         break;
 103                 }
 104
 105                 for (i = 0; i < RTE_DIM(objs);
 106                      ++i, id = (id + 1) & rxq->ptr_mask) {
 107                         m = objs[i];
 108
 109                         rxd = &rxq->sw_desc[id];
 110                         rxd->mbuf = m;
 111
 112                         rte_mbuf_refcnt_set(m, 1);
 113                         m->data_off = RTE_PKTMBUF_HEADROOM;
 114                         m->next = NULL;
 115                         m->nb_segs = 1;
 116                         m->port = port_id;
 117
 118                         addr[i] = rte_pktmbuf_mtophys(m);
 119                 }
 120
 121                 efx_rx_qpost(rxq->common, addr, rxq->buf_size,
 122                              RTE_DIM(objs), rxq->completed, added);
 123                 added += RTE_DIM(objs);
 124         }
 125
 126         /* Push doorbell if something is posted */
 127         if (rxq->added != added) {
 128                 rxq->added = added;
 129                 efx_rx_qpush(rxq->common, added, &rxq->pushed);
 130         }
 131 }
 132
 133 uint16_t
 134 sfc_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 135 {
 136         struct sfc_rxq *rxq = rx_queue;
 137         unsigned int completed;
 138         unsigned int prefix_size = rxq->prefix_size;
 139         unsigned int done_pkts = 0;
 140         boolean_t discard_next = B_FALSE;
 141
 142         if (unlikely((rxq->state & SFC_RXQ_RUNNING) == 0))
 143                 return 0;
 144
 145         sfc_ev_qpoll(rxq->evq);
 146
 147         completed = rxq->completed;
 148         while (completed != rxq->pending && done_pkts < nb_pkts) {
 149                 unsigned int id;
 150                 struct sfc_rx_sw_desc *rxd;
 151                 struct rte_mbuf *m;
 152                 unsigned int seg_len;
 153                 unsigned int desc_flags;
 154
 155                 id = completed++ & rxq->ptr_mask;
 156                 rxd = &rxq->sw_desc[id];
 157                 m = rxd->mbuf;
 158                 desc_flags = rxd->flags;
 159
 160                 if (discard_next)
 161                         goto discard;
 162
 163                 if (desc_flags & (EFX_ADDR_MISMATCH | EFX_DISCARD))
 164                         goto discard;
 165
 166                 if (desc_flags & EFX_PKT_CONT)
 167                         goto discard;
 168
 169                 if (desc_flags & EFX_PKT_PREFIX_LEN) {
 170                         uint16_t tmp_size;
 171                         int rc __rte_unused;
 172
 173                         rc = efx_pseudo_hdr_pkt_length_get(rxq->common,
 174                                 rte_pktmbuf_mtod(m, uint8_t *), &tmp_size);
 175                         SFC_ASSERT(rc == 0);
 176                         seg_len = tmp_size;
 177                 } else {
 178                         seg_len = rxd->size - prefix_size;
 179                 }
 180
 181                 m->data_off += prefix_size;
 182                 rte_pktmbuf_data_len(m) = seg_len;
 183                 rte_pktmbuf_pkt_len(m) = seg_len;
 184
 185                 m->packet_type = RTE_PTYPE_L2_ETHER;
 186
 187                 *rx_pkts++ = m;
 188                 done_pkts++;
 189                 continue;
 190
 191 discard:
 192                 discard_next = ((desc_flags & EFX_PKT_CONT) != 0);
 193                 rte_mempool_put(rxq->refill_mb_pool, m);
 194                 rxd->mbuf = NULL;
 195         }
 196
 197         rxq->completed = completed;
 198
 199         sfc_rx_qrefill(rxq);
 200
 201         return done_pkts;
 202 }
 203
 204 static void
 205 sfc_rx_qpurge(struct sfc_rxq *rxq)
 206 {
 207         unsigned int i;
 208         struct sfc_rx_sw_desc *rxd;
 209
 210         for (i = rxq->completed; i != rxq->added; ++i) {
 211                 rxd = &rxq->sw_desc[i & rxq->ptr_mask];
 212                 rte_mempool_put(rxq->refill_mb_pool, rxd->mbuf);
 213                 rxd->mbuf = NULL;
 214         }
 215 }
 216
 217 static void
 218 sfc_rx_qflush(struct sfc_adapter *sa, unsigned int sw_index)
 219 {
 220         struct sfc_rxq *rxq;
 221         unsigned int retry_count;
 222         unsigned int wait_count;
 223
 224         rxq = sa->rxq_info[sw_index].rxq;
 225         SFC_ASSERT(rxq->state & SFC_RXQ_STARTED);
 226
 227         /*
 228          * Retry Rx queue flushing in the case of flush failed or
 229          * timeout. In the worst case it can delay for 6 seconds.
 230          */
 231         for (retry_count = 0;
 232              ((rxq->state & SFC_RXQ_FLUSHED) == 0) &&
 233              (retry_count < SFC_RX_QFLUSH_ATTEMPTS);
 234              ++retry_count) {
 235                 if (efx_rx_qflush(rxq->common) != 0) {
 236                         rxq->state |= SFC_RXQ_FLUSH_FAILED;
 237                         break;
 238                 }
 239                 rxq->state &= ~SFC_RXQ_FLUSH_FAILED;
 240                 rxq->state |= SFC_RXQ_FLUSHING;
 241
 242                 /*
 243                  * Wait for Rx queue flush done or failed event at least
 244                  * SFC_RX_QFLUSH_POLL_WAIT_MS milliseconds and not more
 245                  * than 2 seconds (SFC_RX_QFLUSH_POLL_WAIT_MS multiplied
 246                  * by SFC_RX_QFLUSH_POLL_ATTEMPTS).
 247                  */
 248                 wait_count = 0;
 249                 do {
 250                         rte_delay_ms(SFC_RX_QFLUSH_POLL_WAIT_MS);
 251                         sfc_ev_qpoll(rxq->evq);
 252                 } while ((rxq->state & SFC_RXQ_FLUSHING) &&
 253                          (wait_count++ < SFC_RX_QFLUSH_POLL_ATTEMPTS));
 254
 255                 if (rxq->state & SFC_RXQ_FLUSHING)
 256                         sfc_err(sa, "RxQ %u flush timed out", sw_index);
 257
 258                 if (rxq->state & SFC_RXQ_FLUSH_FAILED)
 259                         sfc_err(sa, "RxQ %u flush failed", sw_index);
 260
 261                 if (rxq->state & SFC_RXQ_FLUSHED)
 262                         sfc_info(sa, "RxQ %u flushed", sw_index);
 263         }
 264
 265         sfc_rx_qpurge(rxq);
 266 }
 267
 268 int
 269 sfc_rx_qstart(struct sfc_adapter *sa, unsigned int sw_index)
 270 {
 271         struct sfc_rxq_info *rxq_info;
 272         struct sfc_rxq *rxq;
 273         struct sfc_evq *evq;
 274         int rc;
 275
 276         sfc_log_init(sa, "sw_index=%u", sw_index);
 277
 278         SFC_ASSERT(sw_index < sa->rxq_count);
 279
 280         rxq_info = &sa->rxq_info[sw_index];
 281         rxq = rxq_info->rxq;
 282         SFC_ASSERT(rxq->state == SFC_RXQ_INITIALIZED);
 283
 284         evq = rxq->evq;
 285
 286         rc = sfc_ev_qstart(sa, evq->evq_index);
 287         if (rc != 0)
 288                 goto fail_ev_qstart;
 289
 290         rc = efx_rx_qcreate(sa->nic, rxq->hw_index, 0, rxq_info->type,
 291                             &rxq->mem, rxq_info->entries,
 292                             0 /* not used on EF10 */, evq->common,
 293                             &rxq->common);
 294         if (rc != 0)
 295                 goto fail_rx_qcreate;
 296
 297         efx_rx_qenable(rxq->common);
 298
 299         rxq->pending = rxq->completed = rxq->added = rxq->pushed = 0;
 300
 301         rxq->state |= (SFC_RXQ_STARTED | SFC_RXQ_RUNNING);
 302
 303         sfc_rx_qrefill(rxq);
 304
 305         if (sw_index == 0) {
 306                 rc = efx_mac_filter_default_rxq_set(sa->nic, rxq->common,
 307                                                     B_FALSE);
 308                 if (rc != 0)
 309                         goto fail_mac_filter_default_rxq_set;
 310         }
 311
 312         /* It seems to be used by DPDK for debug purposes only ('rte_ether') */
 313         sa->eth_dev->data->rx_queue_state[sw_index] =
 314                 RTE_ETH_QUEUE_STATE_STARTED;
 315
 316         return 0;
 317
 318 fail_mac_filter_default_rxq_set:
 319         sfc_rx_qflush(sa, sw_index);
 320
 321 fail_rx_qcreate:
 322         sfc_ev_qstop(sa, evq->evq_index);
 323
 324 fail_ev_qstart:
 325         return rc;
 326 }
 327
 328 void
 329 sfc_rx_qstop(struct sfc_adapter *sa, unsigned int sw_index)
 330 {
 331         struct sfc_rxq_info *rxq_info;
 332         struct sfc_rxq *rxq;
 333
 334         sfc_log_init(sa, "sw_index=%u", sw_index);
 335
 336         SFC_ASSERT(sw_index < sa->rxq_count);
 337
 338         rxq_info = &sa->rxq_info[sw_index];
 339         rxq = rxq_info->rxq;
 340         SFC_ASSERT(rxq->state & SFC_RXQ_STARTED);
 341
 342         /* It seems to be used by DPDK for debug purposes only ('rte_ether') */
 343         sa->eth_dev->data->rx_queue_state[sw_index] =
 344                 RTE_ETH_QUEUE_STATE_STOPPED;
 345
 346         rxq->state &= ~SFC_RXQ_RUNNING;
 347
 348         if (sw_index == 0)
 349                 efx_mac_filter_default_rxq_clear(sa->nic);
 350
 351         sfc_rx_qflush(sa, sw_index);
 352
 353         rxq->state = SFC_RXQ_INITIALIZED;
 354
 355         efx_rx_qdestroy(rxq->common);
 356
 357         sfc_ev_qstop(sa, rxq->evq->evq_index);
 358 }
 359
 360 static int
 361 sfc_rx_qcheck_conf(struct sfc_adapter *sa,
 362                    const struct rte_eth_rxconf *rx_conf)
 363 {
 364         int rc = 0;
 365
 366         if (rx_conf->rx_thresh.pthresh != 0 ||
 367             rx_conf->rx_thresh.hthresh != 0 ||
 368             rx_conf->rx_thresh.wthresh != 0) {
 369                 sfc_err(sa,
 370                         "RxQ prefetch/host/writeback thresholds are not supported");
 371                 rc = EINVAL;
 372         }
 373
 374         if (rx_conf->rx_free_thresh != 0) {
 375                 sfc_err(sa, "RxQ free threshold is not supported");
 376                 rc = EINVAL;
 377         }
 378
 379         if (rx_conf->rx_drop_en == 0) {
 380                 sfc_err(sa, "RxQ drop disable is not supported");
 381                 rc = EINVAL;
 382         }
 383
 384         if (rx_conf->rx_deferred_start != 0) {
 385                 sfc_err(sa, "RxQ deferred start is not supported");
 386                 rc = EINVAL;
 387         }
 388
 389         return rc;
 390 }
 391
 392 static unsigned int
 393 sfc_rx_mbuf_data_alignment(struct rte_mempool *mb_pool)
 394 {
 395         uint32_t data_off;
 396         uint32_t order;
 397
 398         /* The mbuf object itself is always cache line aligned */
 399         order = rte_bsf32(RTE_CACHE_LINE_SIZE);
 400
 401         /* Data offset from mbuf object start */
 402         data_off = sizeof(struct rte_mbuf) + rte_pktmbuf_priv_size(mb_pool) +
 403                 RTE_PKTMBUF_HEADROOM;
 404
 405         order = MIN(order, rte_bsf32(data_off));
 406
 407         return 1u << (order - 1);
 408 }
 409
 410 static uint16_t
 411 sfc_rx_mb_pool_buf_size(struct sfc_adapter *sa, struct rte_mempool *mb_pool)
 412 {
 413         const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
 414         const uint32_t nic_align_start = MAX(1, encp->enc_rx_buf_align_start);
 415         const uint32_t nic_align_end = MAX(1, encp->enc_rx_buf_align_end);
 416         uint16_t buf_size;
 417         unsigned int buf_aligned;
 418         unsigned int start_alignment;
 419         unsigned int end_padding_alignment;
 420
 421         /* Below it is assumed that both alignments are power of 2 */
 422         SFC_ASSERT(rte_is_power_of_2(nic_align_start));
 423         SFC_ASSERT(rte_is_power_of_2(nic_align_end));
 424
 425         /*
 426          * mbuf is always cache line aligned, double-check
 427          * that it meets rx buffer start alignment requirements.
 428          */
 429
 430         /* Start from mbuf pool data room size */
 431         buf_size = rte_pktmbuf_data_room_size(mb_pool);
 432
 433         /* Remove headroom */
 434         if (buf_size <= RTE_PKTMBUF_HEADROOM) {
 435                 sfc_err(sa,
 436                         "RxQ mbuf pool %s object data room size %u is smaller than headroom %u",
 437                         mb_pool->name, buf_size, RTE_PKTMBUF_HEADROOM);
 438                 return 0;
 439         }
 440         buf_size -= RTE_PKTMBUF_HEADROOM;
 441
 442         /* Calculate guaranteed data start alignment */
 443         buf_aligned = sfc_rx_mbuf_data_alignment(mb_pool);
 444
 445         /* Reserve space for start alignment */
 446         if (buf_aligned < nic_align_start) {
 447                 start_alignment = nic_align_start - buf_aligned;
 448                 if (buf_size <= start_alignment) {
 449                         sfc_err(sa,
 450                                 "RxQ mbuf pool %s object data room size %u is insufficient for headroom %u and buffer start alignment %u required by NIC",
 451                                 mb_pool->name,
 452                                 rte_pktmbuf_data_room_size(mb_pool),
 453                                 RTE_PKTMBUF_HEADROOM, start_alignment);
 454                         return 0;
 455                 }
 456                 buf_aligned = nic_align_start;
 457                 buf_size -= start_alignment;
 458         } else {
 459                 start_alignment = 0;
 460         }
 461
 462         /* Make sure that end padding does not write beyond the buffer */
 463         if (buf_aligned < nic_align_end) {
 464                 /*
 465                  * Estimate space which can be lost. If guarnteed buffer
 466                  * size is odd, lost space is (nic_align_end - 1). More
 467                  * accurate formula is below.
 468                  */
 469                 end_padding_alignment = nic_align_end -
 470                         MIN(buf_aligned, 1u << (rte_bsf32(buf_size) - 1));
 471                 if (buf_size <= end_padding_alignment) {
 472                         sfc_err(sa,
 473                                 "RxQ mbuf pool %s object data room size %u is insufficient for headroom %u, buffer start alignment %u and end padding alignment %u required by NIC",
 474                                 mb_pool->name,
 475                                 rte_pktmbuf_data_room_size(mb_pool),
 476                                 RTE_PKTMBUF_HEADROOM, start_alignment,
 477                                 end_padding_alignment);
 478                         return 0;
 479                 }
 480                 buf_size -= end_padding_alignment;
 481         } else {
 482                 /*
 483                  * Start is aligned the same or better than end,
 484                  * just align length.
 485                  */
 486                 buf_size = P2ALIGN(buf_size, nic_align_end);
 487         }
 488
 489         return buf_size;
 490 }
 491
 492 int
 493 sfc_rx_qinit(struct sfc_adapter *sa, unsigned int sw_index,
 494              uint16_t nb_rx_desc, unsigned int socket_id,
 495              const struct rte_eth_rxconf *rx_conf,
 496              struct rte_mempool *mb_pool)
 497 {
 498         const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
 499         int rc;
 500         uint16_t buf_size;
 501         struct sfc_rxq_info *rxq_info;
 502         unsigned int evq_index;
 503         struct sfc_evq *evq;
 504         struct sfc_rxq *rxq;
 505
 506         rc = sfc_rx_qcheck_conf(sa, rx_conf);
 507         if (rc != 0)
 508                 goto fail_bad_conf;
 509
 510         buf_size = sfc_rx_mb_pool_buf_size(sa, mb_pool);
 511         if (buf_size == 0) {
 512                 sfc_err(sa, "RxQ %u mbuf pool object size is too small",
 513                         sw_index);
 514                 rc = EINVAL;
 515                 goto fail_bad_conf;
 516         }
 517
 518         if ((buf_size < sa->port.pdu + encp->enc_rx_prefix_size) &&
 519             !sa->eth_dev->data->dev_conf.rxmode.enable_scatter) {
 520                 sfc_err(sa, "Rx scatter is disabled and RxQ %u mbuf pool "
 521                         "object size is too small", sw_index);
 522                 sfc_err(sa, "RxQ %u calculated Rx buffer size is %u vs "
 523                         "PDU size %u plus Rx prefix %u bytes",
 524                         sw_index, buf_size, (unsigned int)sa->port.pdu,
 525                         encp->enc_rx_prefix_size);
 526                 rc = EINVAL;
 527                 goto fail_bad_conf;
 528         }
 529
 530         SFC_ASSERT(sw_index < sa->rxq_count);
 531         rxq_info = &sa->rxq_info[sw_index];
 532
 533         SFC_ASSERT(nb_rx_desc <= rxq_info->max_entries);
 534         rxq_info->entries = nb_rx_desc;
 535         rxq_info->type = EFX_RXQ_TYPE_DEFAULT;
 536
 537         evq_index = sfc_evq_index_by_rxq_sw_index(sa, sw_index);
 538
 539         rc = sfc_ev_qinit(sa, evq_index, rxq_info->entries, socket_id);
 540         if (rc != 0)
 541                 goto fail_ev_qinit;
 542
 543         evq = sa->evq_info[evq_index].evq;
 544
 545         rc = ENOMEM;
 546         rxq = rte_zmalloc_socket("sfc-rxq", sizeof(*rxq), RTE_CACHE_LINE_SIZE,
 547                                  socket_id);
 548         if (rxq == NULL)
 549                 goto fail_rxq_alloc;
 550
 551         rc = sfc_dma_alloc(sa, "rxq", sw_index, EFX_RXQ_SIZE(rxq_info->entries),
 552                            socket_id, &rxq->mem);
 553         if (rc != 0)
 554                 goto fail_dma_alloc;
 555
 556         rc = ENOMEM;
 557         rxq->sw_desc = rte_calloc_socket("sfc-rxq-sw_desc", rxq_info->entries,
 558                                          sizeof(*rxq->sw_desc),
 559                                          RTE_CACHE_LINE_SIZE, socket_id);
 560         if (rxq->sw_desc == NULL)
 561                 goto fail_desc_alloc;
 562
 563         evq->rxq = rxq;
 564         rxq->evq = evq;
 565         rxq->ptr_mask = rxq_info->entries - 1;
 566         rxq->refill_mb_pool = mb_pool;
 567         rxq->buf_size = buf_size;
 568         rxq->hw_index = sw_index;
 569         rxq->port_id = sa->eth_dev->data->port_id;
 570
 571         /* Cache limits required on datapath in RxQ structure */
 572         rxq->batch_max = encp->enc_rx_batch_max;
 573         rxq->prefix_size = encp->enc_rx_prefix_size;
 574
 575         rxq->state = SFC_RXQ_INITIALIZED;
 576
 577         rxq_info->rxq = rxq;
 578
 579         return 0;
 580
 581 fail_desc_alloc:
 582         sfc_dma_free(sa, &rxq->mem);
 583
 584 fail_dma_alloc:
 585         rte_free(rxq);
 586
 587 fail_rxq_alloc:
 588         sfc_ev_qfini(sa, evq_index);
 589
 590 fail_ev_qinit:
 591         rxq_info->entries = 0;
 592
 593 fail_bad_conf:
 594         sfc_log_init(sa, "failed %d", rc);
 595         return rc;
 596 }
 597
 598 void
 599 sfc_rx_qfini(struct sfc_adapter *sa, unsigned int sw_index)
 600 {
 601         struct sfc_rxq_info *rxq_info;
 602         struct sfc_rxq *rxq;
 603
 604         SFC_ASSERT(sw_index < sa->rxq_count);
 605
 606         rxq_info = &sa->rxq_info[sw_index];
 607
 608         rxq = rxq_info->rxq;
 609         SFC_ASSERT(rxq->state == SFC_RXQ_INITIALIZED);
 610
 611         rxq_info->rxq = NULL;
 612         rxq_info->entries = 0;
 613
 614         rte_free(rxq->sw_desc);
 615         sfc_dma_free(sa, &rxq->mem);
 616         rte_free(rxq);
 617 }
 618
 619 int
 620 sfc_rx_start(struct sfc_adapter *sa)
 621 {
 622         unsigned int sw_index;
 623         int rc;
 624
 625         sfc_log_init(sa, "rxq_count=%u", sa->rxq_count);
 626
 627         rc = efx_rx_init(sa->nic);
 628         if (rc != 0)
 629                 goto fail_rx_init;
 630
 631         for (sw_index = 0; sw_index < sa->rxq_count; ++sw_index) {
 632                 rc = sfc_rx_qstart(sa, sw_index);
 633                 if (rc != 0)
 634                         goto fail_rx_qstart;
 635         }
 636
 637         return 0;
 638
 639 fail_rx_qstart:
 640         while (sw_index-- > 0)
 641                 sfc_rx_qstop(sa, sw_index);
 642
 643         efx_rx_fini(sa->nic);
 644
 645 fail_rx_init:
 646         sfc_log_init(sa, "failed %d", rc);
 647         return rc;
 648 }
 649
 650 void
 651 sfc_rx_stop(struct sfc_adapter *sa)
 652 {
 653         unsigned int sw_index;
 654
 655         sfc_log_init(sa, "rxq_count=%u", sa->rxq_count);
 656
 657         sw_index = sa->rxq_count;
 658         while (sw_index-- > 0) {
 659                 if (sa->rxq_info[sw_index].rxq != NULL)
 660                         sfc_rx_qstop(sa, sw_index);
 661         }
 662
 663         efx_rx_fini(sa->nic);
 664 }
 665
 666 static int
 667 sfc_rx_qinit_info(struct sfc_adapter *sa, unsigned int sw_index)
 668 {
 669         struct sfc_rxq_info *rxq_info = &sa->rxq_info[sw_index];
 670         unsigned int max_entries;
 671
 672         max_entries = EFX_RXQ_MAXNDESCS;
 673         SFC_ASSERT(rte_is_power_of_2(max_entries));
 674
 675         rxq_info->max_entries = max_entries;
 676
 677         return 0;
 678 }
 679
 680 static int
 681 sfc_rx_check_mode(struct sfc_adapter *sa, struct rte_eth_rxmode *rxmode)
 682 {
 683         int rc = 0;
 684
 685         switch (rxmode->mq_mode) {
 686         case ETH_MQ_RX_NONE:
 687                 /* No special checks are required */
 688                 break;
 689         default:
 690                 sfc_err(sa, "Rx multi-queue mode %u not supported",
 691                         rxmode->mq_mode);
 692                 rc = EINVAL;
 693         }
 694
 695         if (rxmode->header_split) {
 696                 sfc_err(sa, "Header split on Rx not supported");
 697                 rc = EINVAL;
 698         }
 699
 700         if (rxmode->hw_vlan_filter) {
 701                 sfc_err(sa, "HW VLAN filtering not supported");
 702                 rc = EINVAL;
 703         }
 704
 705         if (rxmode->hw_vlan_strip) {
 706                 sfc_err(sa, "HW VLAN stripping not supported");
 707                 rc = EINVAL;
 708         }
 709
 710         if (rxmode->hw_vlan_extend) {
 711                 sfc_err(sa,
 712                         "Q-in-Q HW VLAN stripping not supported");
 713                 rc = EINVAL;
 714         }
 715
 716         if (!rxmode->hw_strip_crc) {
 717                 sfc_warn(sa,
 718                          "FCS stripping control not supported - always stripped");
 719                 rxmode->hw_strip_crc = 1;
 720         }
 721
 722         if (rxmode->enable_scatter) {
 723                 sfc_err(sa, "Scatter on Rx not supported");
 724                 rc = EINVAL;
 725         }
 726
 727         if (rxmode->enable_lro) {
 728                 sfc_err(sa, "LRO not supported");
 729                 rc = EINVAL;
 730         }
 731
 732         return rc;
 733 }
 734
 735 /**
 736  * Initialize Rx subsystem.
 737  *
 738  * Called at device configuration stage when number of receive queues is
 739  * specified together with other device level receive configuration.
 740  *
 741  * It should be used to allocate NUMA-unaware resources.
 742  */
 743 int
 744 sfc_rx_init(struct sfc_adapter *sa)
 745 {
 746         struct rte_eth_conf *dev_conf = &sa->eth_dev->data->dev_conf;
 747         unsigned int sw_index;
 748         int rc;
 749
 750         rc = sfc_rx_check_mode(sa, &dev_conf->rxmode);
 751         if (rc != 0)
 752                 goto fail_check_mode;
 753
 754         sa->rxq_count = sa->eth_dev->data->nb_rx_queues;
 755
 756         rc = ENOMEM;
 757         sa->rxq_info = rte_calloc_socket("sfc-rxqs", sa->rxq_count,
 758                                          sizeof(struct sfc_rxq_info), 0,
 759                                          sa->socket_id);
 760         if (sa->rxq_info == NULL)
 761                 goto fail_rxqs_alloc;
 762
 763         for (sw_index = 0; sw_index < sa->rxq_count; ++sw_index) {
 764                 rc = sfc_rx_qinit_info(sa, sw_index);
 765                 if (rc != 0)
 766                         goto fail_rx_qinit_info;
 767         }
 768
 769         return 0;
 770
 771 fail_rx_qinit_info:
 772         rte_free(sa->rxq_info);
 773         sa->rxq_info = NULL;
 774
 775 fail_rxqs_alloc:
 776         sa->rxq_count = 0;
 777 fail_check_mode:
 778         sfc_log_init(sa, "failed %d", rc);
 779         return rc;
 780 }
 781
 782 /**
 783  * Shutdown Rx subsystem.
 784  *
 785  * Called at device close stage, for example, before device
 786  * reconfiguration or shutdown.
 787  */
 788 void
 789 sfc_rx_fini(struct sfc_adapter *sa)
 790 {
 791         unsigned int sw_index;
 792
 793         sw_index = sa->rxq_count;
 794         while (sw_index-- > 0) {
 795                 if (sa->rxq_info[sw_index].rxq != NULL)
 796                         sfc_rx_qfini(sa, sw_index);
 797         }
 798
 799         rte_free(sa->rxq_info);
 800         sa->rxq_info = NULL;
 801         sa->rxq_count = 0;
 802 }