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
  91         if (free_space < rxq->refill_threshold)
  92                 return;
  93
  94         bulks = free_space / RTE_DIM(objs);
  95
  96         id = added & rxq->ptr_mask;
  97         while (bulks-- > 0) {
  98                 if (rte_mempool_get_bulk(rxq->refill_mb_pool, objs,
  99                                          RTE_DIM(objs)) < 0) {
 100                         /*
 101                          * It is hardly a safe way to increment counter
 102                          * from different contexts, but all PMDs do it.
 103                          */
 104                         rxq->evq->sa->eth_dev->data->rx_mbuf_alloc_failed +=
 105                                 RTE_DIM(objs);
 106                         break;
 107                 }
 108
 109                 for (i = 0; i < RTE_DIM(objs);
 110                      ++i, id = (id + 1) & rxq->ptr_mask) {
 111                         m = objs[i];
 112
 113                         rxd = &rxq->sw_desc[id];
 114                         rxd->mbuf = m;
 115
 116                         rte_mbuf_refcnt_set(m, 1);
 117                         m->data_off = RTE_PKTMBUF_HEADROOM;
 118                         m->next = NULL;
 119                         m->nb_segs = 1;
 120                         m->port = port_id;
 121
 122                         addr[i] = rte_pktmbuf_mtophys(m);
 123                 }
 124
 125                 efx_rx_qpost(rxq->common, addr, rxq->buf_size,
 126                              RTE_DIM(objs), rxq->completed, added);
 127                 added += RTE_DIM(objs);
 128         }
 129
 130         /* Push doorbell if something is posted */
 131         if (rxq->added != added) {
 132                 rxq->added = added;
 133                 efx_rx_qpush(rxq->common, added, &rxq->pushed);
 134         }
 135 }
 136
 137 static uint64_t
 138 sfc_rx_desc_flags_to_offload_flags(const unsigned int desc_flags)
 139 {
 140         uint64_t mbuf_flags = 0;
 141
 142         switch (desc_flags & (EFX_PKT_IPV4 | EFX_CKSUM_IPV4)) {
 143         case (EFX_PKT_IPV4 | EFX_CKSUM_IPV4):
 144                 mbuf_flags |= PKT_RX_IP_CKSUM_GOOD;
 145                 break;
 146         case EFX_PKT_IPV4:
 147                 mbuf_flags |= PKT_RX_IP_CKSUM_BAD;
 148                 break;
 149         default:
 150                 RTE_BUILD_BUG_ON(PKT_RX_IP_CKSUM_UNKNOWN != 0);
 151                 SFC_ASSERT((mbuf_flags & PKT_RX_IP_CKSUM_MASK) ==
 152                            PKT_RX_IP_CKSUM_UNKNOWN);
 153                 break;
 154         }
 155
 156         switch ((desc_flags &
 157                  (EFX_PKT_TCP | EFX_PKT_UDP | EFX_CKSUM_TCPUDP))) {
 158         case (EFX_PKT_TCP | EFX_CKSUM_TCPUDP):
 159         case (EFX_PKT_UDP | EFX_CKSUM_TCPUDP):
 160                 mbuf_flags |= PKT_RX_L4_CKSUM_GOOD;
 161                 break;
 162         case EFX_PKT_TCP:
 163         case EFX_PKT_UDP:
 164                 mbuf_flags |= PKT_RX_L4_CKSUM_BAD;
 165                 break;
 166         default:
 167                 RTE_BUILD_BUG_ON(PKT_RX_L4_CKSUM_UNKNOWN != 0);
 168                 SFC_ASSERT((mbuf_flags & PKT_RX_L4_CKSUM_MASK) ==
 169                            PKT_RX_L4_CKSUM_UNKNOWN);
 170                 break;
 171         }
 172
 173         return mbuf_flags;
 174 }
 175
 176 static uint32_t
 177 sfc_rx_desc_flags_to_packet_type(const unsigned int desc_flags)
 178 {
 179         return RTE_PTYPE_L2_ETHER |
 180                 ((desc_flags & EFX_PKT_IPV4) ?
 181                         RTE_PTYPE_L3_IPV4_EXT_UNKNOWN : 0) |
 182                 ((desc_flags & EFX_PKT_IPV6) ?
 183                         RTE_PTYPE_L3_IPV6_EXT_UNKNOWN : 0) |
 184                 ((desc_flags & EFX_PKT_TCP) ? RTE_PTYPE_L4_TCP : 0) |
 185                 ((desc_flags & EFX_PKT_UDP) ? RTE_PTYPE_L4_UDP : 0);
 186 }
 187
 188 uint16_t
 189 sfc_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 190 {
 191         struct sfc_rxq *rxq = rx_queue;
 192         unsigned int completed;
 193         unsigned int prefix_size = rxq->prefix_size;
 194         unsigned int done_pkts = 0;
 195         boolean_t discard_next = B_FALSE;
 196
 197         if (unlikely((rxq->state & SFC_RXQ_RUNNING) == 0))
 198                 return 0;
 199
 200         sfc_ev_qpoll(rxq->evq);
 201
 202         completed = rxq->completed;
 203         while (completed != rxq->pending && done_pkts < nb_pkts) {
 204                 unsigned int id;
 205                 struct sfc_rx_sw_desc *rxd;
 206                 struct rte_mbuf *m;
 207                 unsigned int seg_len;
 208                 unsigned int desc_flags;
 209
 210                 id = completed++ & rxq->ptr_mask;
 211                 rxd = &rxq->sw_desc[id];
 212                 m = rxd->mbuf;
 213                 desc_flags = rxd->flags;
 214
 215                 if (discard_next)
 216                         goto discard;
 217
 218                 if (desc_flags & (EFX_ADDR_MISMATCH | EFX_DISCARD))
 219                         goto discard;
 220
 221                 if (desc_flags & EFX_PKT_CONT)
 222                         goto discard;
 223
 224                 if (desc_flags & EFX_PKT_PREFIX_LEN) {
 225                         uint16_t tmp_size;
 226                         int rc __rte_unused;
 227
 228                         rc = efx_pseudo_hdr_pkt_length_get(rxq->common,
 229                                 rte_pktmbuf_mtod(m, uint8_t *), &tmp_size);
 230                         SFC_ASSERT(rc == 0);
 231                         seg_len = tmp_size;
 232                 } else {
 233                         seg_len = rxd->size - prefix_size;
 234                 }
 235
 236                 m->data_off += prefix_size;
 237                 rte_pktmbuf_data_len(m) = seg_len;
 238                 rte_pktmbuf_pkt_len(m) = seg_len;
 239
 240                 m->ol_flags = sfc_rx_desc_flags_to_offload_flags(desc_flags);
 241                 m->packet_type = sfc_rx_desc_flags_to_packet_type(desc_flags);
 242
 243                 *rx_pkts++ = m;
 244                 done_pkts++;
 245                 continue;
 246
 247 discard:
 248                 discard_next = ((desc_flags & EFX_PKT_CONT) != 0);
 249                 rte_mempool_put(rxq->refill_mb_pool, m);
 250                 rxd->mbuf = NULL;
 251         }
 252
 253         rxq->completed = completed;
 254
 255         sfc_rx_qrefill(rxq);
 256
 257         return done_pkts;
 258 }
 259
 260 static void
 261 sfc_rx_qpurge(struct sfc_rxq *rxq)
 262 {
 263         unsigned int i;
 264         struct sfc_rx_sw_desc *rxd;
 265
 266         for (i = rxq->completed; i != rxq->added; ++i) {
 267                 rxd = &rxq->sw_desc[i & rxq->ptr_mask];
 268                 rte_mempool_put(rxq->refill_mb_pool, rxd->mbuf);
 269                 rxd->mbuf = NULL;
 270         }
 271 }
 272
 273 static void
 274 sfc_rx_qflush(struct sfc_adapter *sa, unsigned int sw_index)
 275 {
 276         struct sfc_rxq *rxq;
 277         unsigned int retry_count;
 278         unsigned int wait_count;
 279
 280         rxq = sa->rxq_info[sw_index].rxq;
 281         SFC_ASSERT(rxq->state & SFC_RXQ_STARTED);
 282
 283         /*
 284          * Retry Rx queue flushing in the case of flush failed or
 285          * timeout. In the worst case it can delay for 6 seconds.
 286          */
 287         for (retry_count = 0;
 288              ((rxq->state & SFC_RXQ_FLUSHED) == 0) &&
 289              (retry_count < SFC_RX_QFLUSH_ATTEMPTS);
 290              ++retry_count) {
 291                 if (efx_rx_qflush(rxq->common) != 0) {
 292                         rxq->state |= SFC_RXQ_FLUSH_FAILED;
 293                         break;
 294                 }
 295                 rxq->state &= ~SFC_RXQ_FLUSH_FAILED;
 296                 rxq->state |= SFC_RXQ_FLUSHING;
 297
 298                 /*
 299                  * Wait for Rx queue flush done or failed event at least
 300                  * SFC_RX_QFLUSH_POLL_WAIT_MS milliseconds and not more
 301                  * than 2 seconds (SFC_RX_QFLUSH_POLL_WAIT_MS multiplied
 302                  * by SFC_RX_QFLUSH_POLL_ATTEMPTS).
 303                  */
 304                 wait_count = 0;
 305                 do {
 306                         rte_delay_ms(SFC_RX_QFLUSH_POLL_WAIT_MS);
 307                         sfc_ev_qpoll(rxq->evq);
 308                 } while ((rxq->state & SFC_RXQ_FLUSHING) &&
 309                          (wait_count++ < SFC_RX_QFLUSH_POLL_ATTEMPTS));
 310
 311                 if (rxq->state & SFC_RXQ_FLUSHING)
 312                         sfc_err(sa, "RxQ %u flush timed out", sw_index);
 313
 314                 if (rxq->state & SFC_RXQ_FLUSH_FAILED)
 315                         sfc_err(sa, "RxQ %u flush failed", sw_index);
 316
 317                 if (rxq->state & SFC_RXQ_FLUSHED)
 318                         sfc_info(sa, "RxQ %u flushed", sw_index);
 319         }
 320
 321         sfc_rx_qpurge(rxq);
 322 }
 323
 324 int
 325 sfc_rx_qstart(struct sfc_adapter *sa, unsigned int sw_index)
 326 {
 327         struct sfc_rxq_info *rxq_info;
 328         struct sfc_rxq *rxq;
 329         struct sfc_evq *evq;
 330         int rc;
 331
 332         sfc_log_init(sa, "sw_index=%u", sw_index);
 333
 334         SFC_ASSERT(sw_index < sa->rxq_count);
 335
 336         rxq_info = &sa->rxq_info[sw_index];
 337         rxq = rxq_info->rxq;
 338         SFC_ASSERT(rxq->state == SFC_RXQ_INITIALIZED);
 339
 340         evq = rxq->evq;
 341
 342         rc = sfc_ev_qstart(sa, evq->evq_index);
 343         if (rc != 0)
 344                 goto fail_ev_qstart;
 345
 346         rc = efx_rx_qcreate(sa->nic, rxq->hw_index, 0, rxq_info->type,
 347                             &rxq->mem, rxq_info->entries,
 348                             0 /* not used on EF10 */, evq->common,
 349                             &rxq->common);
 350         if (rc != 0)
 351                 goto fail_rx_qcreate;
 352
 353         efx_rx_qenable(rxq->common);
 354
 355         rxq->pending = rxq->completed = rxq->added = rxq->pushed = 0;
 356
 357         rxq->state |= (SFC_RXQ_STARTED | SFC_RXQ_RUNNING);
 358
 359         sfc_rx_qrefill(rxq);
 360
 361         if (sw_index == 0) {
 362                 rc = efx_mac_filter_default_rxq_set(sa->nic, rxq->common,
 363                                                     B_FALSE);
 364                 if (rc != 0)
 365                         goto fail_mac_filter_default_rxq_set;
 366         }
 367
 368         /* It seems to be used by DPDK for debug purposes only ('rte_ether') */
 369         sa->eth_dev->data->rx_queue_state[sw_index] =
 370                 RTE_ETH_QUEUE_STATE_STARTED;
 371
 372         return 0;
 373
 374 fail_mac_filter_default_rxq_set:
 375         sfc_rx_qflush(sa, sw_index);
 376
 377 fail_rx_qcreate:
 378         sfc_ev_qstop(sa, evq->evq_index);
 379
 380 fail_ev_qstart:
 381         return rc;
 382 }
 383
 384 void
 385 sfc_rx_qstop(struct sfc_adapter *sa, unsigned int sw_index)
 386 {
 387         struct sfc_rxq_info *rxq_info;
 388         struct sfc_rxq *rxq;
 389
 390         sfc_log_init(sa, "sw_index=%u", sw_index);
 391
 392         SFC_ASSERT(sw_index < sa->rxq_count);
 393
 394         rxq_info = &sa->rxq_info[sw_index];
 395         rxq = rxq_info->rxq;
 396         SFC_ASSERT(rxq->state & SFC_RXQ_STARTED);
 397
 398         /* It seems to be used by DPDK for debug purposes only ('rte_ether') */
 399         sa->eth_dev->data->rx_queue_state[sw_index] =
 400                 RTE_ETH_QUEUE_STATE_STOPPED;
 401
 402         rxq->state &= ~SFC_RXQ_RUNNING;
 403
 404         if (sw_index == 0)
 405                 efx_mac_filter_default_rxq_clear(sa->nic);
 406
 407         sfc_rx_qflush(sa, sw_index);
 408
 409         rxq->state = SFC_RXQ_INITIALIZED;
 410
 411         efx_rx_qdestroy(rxq->common);
 412
 413         sfc_ev_qstop(sa, rxq->evq->evq_index);
 414 }
 415
 416 static int
 417 sfc_rx_qcheck_conf(struct sfc_adapter *sa, uint16_t nb_rx_desc,
 418                    const struct rte_eth_rxconf *rx_conf)
 419 {
 420         const uint16_t rx_free_thresh_max = EFX_RXQ_LIMIT(nb_rx_desc);
 421         int rc = 0;
 422
 423         if (rx_conf->rx_thresh.pthresh != 0 ||
 424             rx_conf->rx_thresh.hthresh != 0 ||
 425             rx_conf->rx_thresh.wthresh != 0) {
 426                 sfc_err(sa,
 427                         "RxQ prefetch/host/writeback thresholds are not supported");
 428                 rc = EINVAL;
 429         }
 430
 431         if (rx_conf->rx_free_thresh > rx_free_thresh_max) {
 432                 sfc_err(sa,
 433                         "RxQ free threshold too large: %u vs maximum %u",
 434                         rx_conf->rx_free_thresh, rx_free_thresh_max);
 435                 rc = EINVAL;
 436         }
 437
 438         if (rx_conf->rx_drop_en == 0) {
 439                 sfc_err(sa, "RxQ drop disable is not supported");
 440                 rc = EINVAL;
 441         }
 442
 443         if (rx_conf->rx_deferred_start != 0) {
 444                 sfc_err(sa, "RxQ deferred start is not supported");
 445                 rc = EINVAL;
 446         }
 447
 448         return rc;
 449 }
 450
 451 static unsigned int
 452 sfc_rx_mbuf_data_alignment(struct rte_mempool *mb_pool)
 453 {
 454         uint32_t data_off;
 455         uint32_t order;
 456
 457         /* The mbuf object itself is always cache line aligned */
 458         order = rte_bsf32(RTE_CACHE_LINE_SIZE);
 459
 460         /* Data offset from mbuf object start */
 461         data_off = sizeof(struct rte_mbuf) + rte_pktmbuf_priv_size(mb_pool) +
 462                 RTE_PKTMBUF_HEADROOM;
 463
 464         order = MIN(order, rte_bsf32(data_off));
 465
 466         return 1u << (order - 1);
 467 }
 468
 469 static uint16_t
 470 sfc_rx_mb_pool_buf_size(struct sfc_adapter *sa, struct rte_mempool *mb_pool)
 471 {
 472         const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
 473         const uint32_t nic_align_start = MAX(1, encp->enc_rx_buf_align_start);
 474         const uint32_t nic_align_end = MAX(1, encp->enc_rx_buf_align_end);
 475         uint16_t buf_size;
 476         unsigned int buf_aligned;
 477         unsigned int start_alignment;
 478         unsigned int end_padding_alignment;
 479
 480         /* Below it is assumed that both alignments are power of 2 */
 481         SFC_ASSERT(rte_is_power_of_2(nic_align_start));
 482         SFC_ASSERT(rte_is_power_of_2(nic_align_end));
 483
 484         /*
 485          * mbuf is always cache line aligned, double-check
 486          * that it meets rx buffer start alignment requirements.
 487          */
 488
 489         /* Start from mbuf pool data room size */
 490         buf_size = rte_pktmbuf_data_room_size(mb_pool);
 491
 492         /* Remove headroom */
 493         if (buf_size <= RTE_PKTMBUF_HEADROOM) {
 494                 sfc_err(sa,
 495                         "RxQ mbuf pool %s object data room size %u is smaller than headroom %u",
 496                         mb_pool->name, buf_size, RTE_PKTMBUF_HEADROOM);
 497                 return 0;
 498         }
 499         buf_size -= RTE_PKTMBUF_HEADROOM;
 500
 501         /* Calculate guaranteed data start alignment */
 502         buf_aligned = sfc_rx_mbuf_data_alignment(mb_pool);
 503
 504         /* Reserve space for start alignment */
 505         if (buf_aligned < nic_align_start) {
 506                 start_alignment = nic_align_start - buf_aligned;
 507                 if (buf_size <= start_alignment) {
 508                         sfc_err(sa,
 509                                 "RxQ mbuf pool %s object data room size %u is insufficient for headroom %u and buffer start alignment %u required by NIC",
 510                                 mb_pool->name,
 511                                 rte_pktmbuf_data_room_size(mb_pool),
 512                                 RTE_PKTMBUF_HEADROOM, start_alignment);
 513                         return 0;
 514                 }
 515                 buf_aligned = nic_align_start;
 516                 buf_size -= start_alignment;
 517         } else {
 518                 start_alignment = 0;
 519         }
 520
 521         /* Make sure that end padding does not write beyond the buffer */
 522         if (buf_aligned < nic_align_end) {
 523                 /*
 524                  * Estimate space which can be lost. If guarnteed buffer
 525                  * size is odd, lost space is (nic_align_end - 1). More
 526                  * accurate formula is below.
 527                  */
 528                 end_padding_alignment = nic_align_end -
 529                         MIN(buf_aligned, 1u << (rte_bsf32(buf_size) - 1));
 530                 if (buf_size <= end_padding_alignment) {
 531                         sfc_err(sa,
 532                                 "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",
 533                                 mb_pool->name,
 534                                 rte_pktmbuf_data_room_size(mb_pool),
 535                                 RTE_PKTMBUF_HEADROOM, start_alignment,
 536                                 end_padding_alignment);
 537                         return 0;
 538                 }
 539                 buf_size -= end_padding_alignment;
 540         } else {
 541                 /*
 542                  * Start is aligned the same or better than end,
 543                  * just align length.
 544                  */
 545                 buf_size = P2ALIGN(buf_size, nic_align_end);
 546         }
 547
 548         return buf_size;
 549 }
 550
 551 int
 552 sfc_rx_qinit(struct sfc_adapter *sa, unsigned int sw_index,
 553              uint16_t nb_rx_desc, unsigned int socket_id,
 554              const struct rte_eth_rxconf *rx_conf,
 555              struct rte_mempool *mb_pool)
 556 {
 557         const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
 558         int rc;
 559         uint16_t buf_size;
 560         struct sfc_rxq_info *rxq_info;
 561         unsigned int evq_index;
 562         struct sfc_evq *evq;
 563         struct sfc_rxq *rxq;
 564
 565         rc = sfc_rx_qcheck_conf(sa, nb_rx_desc, rx_conf);
 566         if (rc != 0)
 567                 goto fail_bad_conf;
 568
 569         buf_size = sfc_rx_mb_pool_buf_size(sa, mb_pool);
 570         if (buf_size == 0) {
 571                 sfc_err(sa, "RxQ %u mbuf pool object size is too small",
 572                         sw_index);
 573                 rc = EINVAL;
 574                 goto fail_bad_conf;
 575         }
 576
 577         if ((buf_size < sa->port.pdu + encp->enc_rx_prefix_size) &&
 578             !sa->eth_dev->data->dev_conf.rxmode.enable_scatter) {
 579                 sfc_err(sa, "Rx scatter is disabled and RxQ %u mbuf pool "
 580                         "object size is too small", sw_index);
 581                 sfc_err(sa, "RxQ %u calculated Rx buffer size is %u vs "
 582                         "PDU size %u plus Rx prefix %u bytes",
 583                         sw_index, buf_size, (unsigned int)sa->port.pdu,
 584                         encp->enc_rx_prefix_size);
 585                 rc = EINVAL;
 586                 goto fail_bad_conf;
 587         }
 588
 589         SFC_ASSERT(sw_index < sa->rxq_count);
 590         rxq_info = &sa->rxq_info[sw_index];
 591
 592         SFC_ASSERT(nb_rx_desc <= rxq_info->max_entries);
 593         rxq_info->entries = nb_rx_desc;
 594         rxq_info->type = EFX_RXQ_TYPE_DEFAULT;
 595
 596         evq_index = sfc_evq_index_by_rxq_sw_index(sa, sw_index);
 597
 598         rc = sfc_ev_qinit(sa, evq_index, rxq_info->entries, socket_id);
 599         if (rc != 0)
 600                 goto fail_ev_qinit;
 601
 602         evq = sa->evq_info[evq_index].evq;
 603
 604         rc = ENOMEM;
 605         rxq = rte_zmalloc_socket("sfc-rxq", sizeof(*rxq), RTE_CACHE_LINE_SIZE,
 606                                  socket_id);
 607         if (rxq == NULL)
 608                 goto fail_rxq_alloc;
 609
 610         rc = sfc_dma_alloc(sa, "rxq", sw_index, EFX_RXQ_SIZE(rxq_info->entries),
 611                            socket_id, &rxq->mem);
 612         if (rc != 0)
 613                 goto fail_dma_alloc;
 614
 615         rc = ENOMEM;
 616         rxq->sw_desc = rte_calloc_socket("sfc-rxq-sw_desc", rxq_info->entries,
 617                                          sizeof(*rxq->sw_desc),
 618                                          RTE_CACHE_LINE_SIZE, socket_id);
 619         if (rxq->sw_desc == NULL)
 620                 goto fail_desc_alloc;
 621
 622         evq->rxq = rxq;
 623         rxq->evq = evq;
 624         rxq->ptr_mask = rxq_info->entries - 1;
 625         rxq->refill_threshold = rx_conf->rx_free_thresh;
 626         rxq->refill_mb_pool = mb_pool;
 627         rxq->buf_size = buf_size;
 628         rxq->hw_index = sw_index;
 629         rxq->port_id = sa->eth_dev->data->port_id;
 630
 631         /* Cache limits required on datapath in RxQ structure */
 632         rxq->batch_max = encp->enc_rx_batch_max;
 633         rxq->prefix_size = encp->enc_rx_prefix_size;
 634
 635         rxq->state = SFC_RXQ_INITIALIZED;
 636
 637         rxq_info->rxq = rxq;
 638
 639         return 0;
 640
 641 fail_desc_alloc:
 642         sfc_dma_free(sa, &rxq->mem);
 643
 644 fail_dma_alloc:
 645         rte_free(rxq);
 646
 647 fail_rxq_alloc:
 648         sfc_ev_qfini(sa, evq_index);
 649
 650 fail_ev_qinit:
 651         rxq_info->entries = 0;
 652
 653 fail_bad_conf:
 654         sfc_log_init(sa, "failed %d", rc);
 655         return rc;
 656 }
 657
 658 void
 659 sfc_rx_qfini(struct sfc_adapter *sa, unsigned int sw_index)
 660 {
 661         struct sfc_rxq_info *rxq_info;
 662         struct sfc_rxq *rxq;
 663
 664         SFC_ASSERT(sw_index < sa->rxq_count);
 665
 666         rxq_info = &sa->rxq_info[sw_index];
 667
 668         rxq = rxq_info->rxq;
 669         SFC_ASSERT(rxq->state == SFC_RXQ_INITIALIZED);
 670
 671         rxq_info->rxq = NULL;
 672         rxq_info->entries = 0;
 673
 674         rte_free(rxq->sw_desc);
 675         sfc_dma_free(sa, &rxq->mem);
 676         rte_free(rxq);
 677 }
 678
 679 int
 680 sfc_rx_start(struct sfc_adapter *sa)
 681 {
 682         unsigned int sw_index;
 683         int rc;
 684
 685         sfc_log_init(sa, "rxq_count=%u", sa->rxq_count);
 686
 687         rc = efx_rx_init(sa->nic);
 688         if (rc != 0)
 689                 goto fail_rx_init;
 690
 691         for (sw_index = 0; sw_index < sa->rxq_count; ++sw_index) {
 692                 rc = sfc_rx_qstart(sa, sw_index);
 693                 if (rc != 0)
 694                         goto fail_rx_qstart;
 695         }
 696
 697         return 0;
 698
 699 fail_rx_qstart:
 700         while (sw_index-- > 0)
 701                 sfc_rx_qstop(sa, sw_index);
 702
 703         efx_rx_fini(sa->nic);
 704
 705 fail_rx_init:
 706         sfc_log_init(sa, "failed %d", rc);
 707         return rc;
 708 }
 709
 710 void
 711 sfc_rx_stop(struct sfc_adapter *sa)
 712 {
 713         unsigned int sw_index;
 714
 715         sfc_log_init(sa, "rxq_count=%u", sa->rxq_count);
 716
 717         sw_index = sa->rxq_count;
 718         while (sw_index-- > 0) {
 719                 if (sa->rxq_info[sw_index].rxq != NULL)
 720                         sfc_rx_qstop(sa, sw_index);
 721         }
 722
 723         efx_rx_fini(sa->nic);
 724 }
 725
 726 static int
 727 sfc_rx_qinit_info(struct sfc_adapter *sa, unsigned int sw_index)
 728 {
 729         struct sfc_rxq_info *rxq_info = &sa->rxq_info[sw_index];
 730         unsigned int max_entries;
 731
 732         max_entries = EFX_RXQ_MAXNDESCS;
 733         SFC_ASSERT(rte_is_power_of_2(max_entries));
 734
 735         rxq_info->max_entries = max_entries;
 736
 737         return 0;
 738 }
 739
 740 static int
 741 sfc_rx_check_mode(struct sfc_adapter *sa, struct rte_eth_rxmode *rxmode)
 742 {
 743         int rc = 0;
 744
 745         switch (rxmode->mq_mode) {
 746         case ETH_MQ_RX_NONE:
 747                 /* No special checks are required */
 748                 break;
 749         default:
 750                 sfc_err(sa, "Rx multi-queue mode %u not supported",
 751                         rxmode->mq_mode);
 752                 rc = EINVAL;
 753         }
 754
 755         if (rxmode->header_split) {
 756                 sfc_err(sa, "Header split on Rx not supported");
 757                 rc = EINVAL;
 758         }
 759
 760         if (rxmode->hw_vlan_filter) {
 761                 sfc_err(sa, "HW VLAN filtering not supported");
 762                 rc = EINVAL;
 763         }
 764
 765         if (rxmode->hw_vlan_strip) {
 766                 sfc_err(sa, "HW VLAN stripping not supported");
 767                 rc = EINVAL;
 768         }
 769
 770         if (rxmode->hw_vlan_extend) {
 771                 sfc_err(sa,
 772                         "Q-in-Q HW VLAN stripping not supported");
 773                 rc = EINVAL;
 774         }
 775
 776         if (!rxmode->hw_strip_crc) {
 777                 sfc_warn(sa,
 778                          "FCS stripping control not supported - always stripped");
 779                 rxmode->hw_strip_crc = 1;
 780         }
 781
 782         if (rxmode->enable_scatter) {
 783                 sfc_err(sa, "Scatter on Rx not supported");
 784                 rc = EINVAL;
 785         }
 786
 787         if (rxmode->enable_lro) {
 788                 sfc_err(sa, "LRO not supported");
 789                 rc = EINVAL;
 790         }
 791
 792         return rc;
 793 }
 794
 795 /**
 796  * Initialize Rx subsystem.
 797  *
 798  * Called at device configuration stage when number of receive queues is
 799  * specified together with other device level receive configuration.
 800  *
 801  * It should be used to allocate NUMA-unaware resources.
 802  */
 803 int
 804 sfc_rx_init(struct sfc_adapter *sa)
 805 {
 806         struct rte_eth_conf *dev_conf = &sa->eth_dev->data->dev_conf;
 807         unsigned int sw_index;
 808         int rc;
 809
 810         rc = sfc_rx_check_mode(sa, &dev_conf->rxmode);
 811         if (rc != 0)
 812                 goto fail_check_mode;
 813
 814         sa->rxq_count = sa->eth_dev->data->nb_rx_queues;
 815
 816         rc = ENOMEM;
 817         sa->rxq_info = rte_calloc_socket("sfc-rxqs", sa->rxq_count,
 818                                          sizeof(struct sfc_rxq_info), 0,
 819                                          sa->socket_id);
 820         if (sa->rxq_info == NULL)
 821                 goto fail_rxqs_alloc;
 822
 823         for (sw_index = 0; sw_index < sa->rxq_count; ++sw_index) {
 824                 rc = sfc_rx_qinit_info(sa, sw_index);
 825                 if (rc != 0)
 826                         goto fail_rx_qinit_info;
 827         }
 828
 829         return 0;
 830
 831 fail_rx_qinit_info:
 832         rte_free(sa->rxq_info);
 833         sa->rxq_info = NULL;
 834
 835 fail_rxqs_alloc:
 836         sa->rxq_count = 0;
 837 fail_check_mode:
 838         sfc_log_init(sa, "failed %d", rc);
 839         return rc;
 840 }
 841
 842 /**
 843  * Shutdown Rx subsystem.
 844  *
 845  * Called at device close stage, for example, before device
 846  * reconfiguration or shutdown.
 847  */
 848 void
 849 sfc_rx_fini(struct sfc_adapter *sa)
 850 {
 851         unsigned int sw_index;
 852
 853         sw_index = sa->rxq_count;
 854         while (sw_index-- > 0) {
 855                 if (sa->rxq_info[sw_index].rxq != NULL)
 856                         sfc_rx_qfini(sa, sw_index);
 857         }
 858
 859         rte_free(sa->rxq_info);
 860         sa->rxq_info = NULL;
 861         sa->rxq_count = 0;
 862 }