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 unsigned int
 261 sfc_rx_qdesc_npending(struct sfc_adapter *sa, unsigned int sw_index)
 262 {
 263         struct sfc_rxq *rxq;
 264
 265         SFC_ASSERT(sw_index < sa->rxq_count);
 266         rxq = sa->rxq_info[sw_index].rxq;
 267
 268         if (rxq == NULL || (rxq->state & SFC_RXQ_RUNNING) == 0)
 269                 return 0;
 270
 271         sfc_ev_qpoll(rxq->evq);
 272
 273         return rxq->pending - rxq->completed;
 274 }
 275
 276 static void
 277 sfc_rx_qpurge(struct sfc_rxq *rxq)
 278 {
 279         unsigned int i;
 280         struct sfc_rx_sw_desc *rxd;
 281
 282         for (i = rxq->completed; i != rxq->added; ++i) {
 283                 rxd = &rxq->sw_desc[i & rxq->ptr_mask];
 284                 rte_mempool_put(rxq->refill_mb_pool, rxd->mbuf);
 285                 rxd->mbuf = NULL;
 286         }
 287 }
 288
 289 static void
 290 sfc_rx_qflush(struct sfc_adapter *sa, unsigned int sw_index)
 291 {
 292         struct sfc_rxq *rxq;
 293         unsigned int retry_count;
 294         unsigned int wait_count;
 295
 296         rxq = sa->rxq_info[sw_index].rxq;
 297         SFC_ASSERT(rxq->state & SFC_RXQ_STARTED);
 298
 299         /*
 300          * Retry Rx queue flushing in the case of flush failed or
 301          * timeout. In the worst case it can delay for 6 seconds.
 302          */
 303         for (retry_count = 0;
 304              ((rxq->state & SFC_RXQ_FLUSHED) == 0) &&
 305              (retry_count < SFC_RX_QFLUSH_ATTEMPTS);
 306              ++retry_count) {
 307                 if (efx_rx_qflush(rxq->common) != 0) {
 308                         rxq->state |= SFC_RXQ_FLUSH_FAILED;
 309                         break;
 310                 }
 311                 rxq->state &= ~SFC_RXQ_FLUSH_FAILED;
 312                 rxq->state |= SFC_RXQ_FLUSHING;
 313
 314                 /*
 315                  * Wait for Rx queue flush done or failed event at least
 316                  * SFC_RX_QFLUSH_POLL_WAIT_MS milliseconds and not more
 317                  * than 2 seconds (SFC_RX_QFLUSH_POLL_WAIT_MS multiplied
 318                  * by SFC_RX_QFLUSH_POLL_ATTEMPTS).
 319                  */
 320                 wait_count = 0;
 321                 do {
 322                         rte_delay_ms(SFC_RX_QFLUSH_POLL_WAIT_MS);
 323                         sfc_ev_qpoll(rxq->evq);
 324                 } while ((rxq->state & SFC_RXQ_FLUSHING) &&
 325                          (wait_count++ < SFC_RX_QFLUSH_POLL_ATTEMPTS));
 326
 327                 if (rxq->state & SFC_RXQ_FLUSHING)
 328                         sfc_err(sa, "RxQ %u flush timed out", sw_index);
 329
 330                 if (rxq->state & SFC_RXQ_FLUSH_FAILED)
 331                         sfc_err(sa, "RxQ %u flush failed", sw_index);
 332
 333                 if (rxq->state & SFC_RXQ_FLUSHED)
 334                         sfc_info(sa, "RxQ %u flushed", sw_index);
 335         }
 336
 337         sfc_rx_qpurge(rxq);
 338 }
 339
 340 int
 341 sfc_rx_qstart(struct sfc_adapter *sa, unsigned int sw_index)
 342 {
 343         struct sfc_rxq_info *rxq_info;
 344         struct sfc_rxq *rxq;
 345         struct sfc_evq *evq;
 346         int rc;
 347
 348         sfc_log_init(sa, "sw_index=%u", sw_index);
 349
 350         SFC_ASSERT(sw_index < sa->rxq_count);
 351
 352         rxq_info = &sa->rxq_info[sw_index];
 353         rxq = rxq_info->rxq;
 354         SFC_ASSERT(rxq->state == SFC_RXQ_INITIALIZED);
 355
 356         evq = rxq->evq;
 357
 358         rc = sfc_ev_qstart(sa, evq->evq_index);
 359         if (rc != 0)
 360                 goto fail_ev_qstart;
 361
 362         rc = efx_rx_qcreate(sa->nic, rxq->hw_index, 0, rxq_info->type,
 363                             &rxq->mem, rxq_info->entries,
 364                             0 /* not used on EF10 */, evq->common,
 365                             &rxq->common);
 366         if (rc != 0)
 367                 goto fail_rx_qcreate;
 368
 369         efx_rx_qenable(rxq->common);
 370
 371         rxq->pending = rxq->completed = rxq->added = rxq->pushed = 0;
 372
 373         rxq->state |= (SFC_RXQ_STARTED | SFC_RXQ_RUNNING);
 374
 375         sfc_rx_qrefill(rxq);
 376
 377         if (sw_index == 0) {
 378                 rc = efx_mac_filter_default_rxq_set(sa->nic, rxq->common,
 379                                                     B_FALSE);
 380                 if (rc != 0)
 381                         goto fail_mac_filter_default_rxq_set;
 382         }
 383
 384         /* It seems to be used by DPDK for debug purposes only ('rte_ether') */
 385         sa->eth_dev->data->rx_queue_state[sw_index] =
 386                 RTE_ETH_QUEUE_STATE_STARTED;
 387
 388         return 0;
 389
 390 fail_mac_filter_default_rxq_set:
 391         sfc_rx_qflush(sa, sw_index);
 392
 393 fail_rx_qcreate:
 394         sfc_ev_qstop(sa, evq->evq_index);
 395
 396 fail_ev_qstart:
 397         return rc;
 398 }
 399
 400 void
 401 sfc_rx_qstop(struct sfc_adapter *sa, unsigned int sw_index)
 402 {
 403         struct sfc_rxq_info *rxq_info;
 404         struct sfc_rxq *rxq;
 405
 406         sfc_log_init(sa, "sw_index=%u", sw_index);
 407
 408         SFC_ASSERT(sw_index < sa->rxq_count);
 409
 410         rxq_info = &sa->rxq_info[sw_index];
 411         rxq = rxq_info->rxq;
 412         SFC_ASSERT(rxq->state & SFC_RXQ_STARTED);
 413
 414         /* It seems to be used by DPDK for debug purposes only ('rte_ether') */
 415         sa->eth_dev->data->rx_queue_state[sw_index] =
 416                 RTE_ETH_QUEUE_STATE_STOPPED;
 417
 418         rxq->state &= ~SFC_RXQ_RUNNING;
 419
 420         if (sw_index == 0)
 421                 efx_mac_filter_default_rxq_clear(sa->nic);
 422
 423         sfc_rx_qflush(sa, sw_index);
 424
 425         rxq->state = SFC_RXQ_INITIALIZED;
 426
 427         efx_rx_qdestroy(rxq->common);
 428
 429         sfc_ev_qstop(sa, rxq->evq->evq_index);
 430 }
 431
 432 static int
 433 sfc_rx_qcheck_conf(struct sfc_adapter *sa, uint16_t nb_rx_desc,
 434                    const struct rte_eth_rxconf *rx_conf)
 435 {
 436         const uint16_t rx_free_thresh_max = EFX_RXQ_LIMIT(nb_rx_desc);
 437         int rc = 0;
 438
 439         if (rx_conf->rx_thresh.pthresh != 0 ||
 440             rx_conf->rx_thresh.hthresh != 0 ||
 441             rx_conf->rx_thresh.wthresh != 0) {
 442                 sfc_err(sa,
 443                         "RxQ prefetch/host/writeback thresholds are not supported");
 444                 rc = EINVAL;
 445         }
 446
 447         if (rx_conf->rx_free_thresh > rx_free_thresh_max) {
 448                 sfc_err(sa,
 449                         "RxQ free threshold too large: %u vs maximum %u",
 450                         rx_conf->rx_free_thresh, rx_free_thresh_max);
 451                 rc = EINVAL;
 452         }
 453
 454         if (rx_conf->rx_drop_en == 0) {
 455                 sfc_err(sa, "RxQ drop disable is not supported");
 456                 rc = EINVAL;
 457         }
 458
 459         if (rx_conf->rx_deferred_start != 0) {
 460                 sfc_err(sa, "RxQ deferred start is not supported");
 461                 rc = EINVAL;
 462         }
 463
 464         return rc;
 465 }
 466
 467 static unsigned int
 468 sfc_rx_mbuf_data_alignment(struct rte_mempool *mb_pool)
 469 {
 470         uint32_t data_off;
 471         uint32_t order;
 472
 473         /* The mbuf object itself is always cache line aligned */
 474         order = rte_bsf32(RTE_CACHE_LINE_SIZE);
 475
 476         /* Data offset from mbuf object start */
 477         data_off = sizeof(struct rte_mbuf) + rte_pktmbuf_priv_size(mb_pool) +
 478                 RTE_PKTMBUF_HEADROOM;
 479
 480         order = MIN(order, rte_bsf32(data_off));
 481
 482         return 1u << (order - 1);
 483 }
 484
 485 static uint16_t
 486 sfc_rx_mb_pool_buf_size(struct sfc_adapter *sa, struct rte_mempool *mb_pool)
 487 {
 488         const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
 489         const uint32_t nic_align_start = MAX(1, encp->enc_rx_buf_align_start);
 490         const uint32_t nic_align_end = MAX(1, encp->enc_rx_buf_align_end);
 491         uint16_t buf_size;
 492         unsigned int buf_aligned;
 493         unsigned int start_alignment;
 494         unsigned int end_padding_alignment;
 495
 496         /* Below it is assumed that both alignments are power of 2 */
 497         SFC_ASSERT(rte_is_power_of_2(nic_align_start));
 498         SFC_ASSERT(rte_is_power_of_2(nic_align_end));
 499
 500         /*
 501          * mbuf is always cache line aligned, double-check
 502          * that it meets rx buffer start alignment requirements.
 503          */
 504
 505         /* Start from mbuf pool data room size */
 506         buf_size = rte_pktmbuf_data_room_size(mb_pool);
 507
 508         /* Remove headroom */
 509         if (buf_size <= RTE_PKTMBUF_HEADROOM) {
 510                 sfc_err(sa,
 511                         "RxQ mbuf pool %s object data room size %u is smaller than headroom %u",
 512                         mb_pool->name, buf_size, RTE_PKTMBUF_HEADROOM);
 513                 return 0;
 514         }
 515         buf_size -= RTE_PKTMBUF_HEADROOM;
 516
 517         /* Calculate guaranteed data start alignment */
 518         buf_aligned = sfc_rx_mbuf_data_alignment(mb_pool);
 519
 520         /* Reserve space for start alignment */
 521         if (buf_aligned < nic_align_start) {
 522                 start_alignment = nic_align_start - buf_aligned;
 523                 if (buf_size <= start_alignment) {
 524                         sfc_err(sa,
 525                                 "RxQ mbuf pool %s object data room size %u is insufficient for headroom %u and buffer start alignment %u required by NIC",
 526                                 mb_pool->name,
 527                                 rte_pktmbuf_data_room_size(mb_pool),
 528                                 RTE_PKTMBUF_HEADROOM, start_alignment);
 529                         return 0;
 530                 }
 531                 buf_aligned = nic_align_start;
 532                 buf_size -= start_alignment;
 533         } else {
 534                 start_alignment = 0;
 535         }
 536
 537         /* Make sure that end padding does not write beyond the buffer */
 538         if (buf_aligned < nic_align_end) {
 539                 /*
 540                  * Estimate space which can be lost. If guarnteed buffer
 541                  * size is odd, lost space is (nic_align_end - 1). More
 542                  * accurate formula is below.
 543                  */
 544                 end_padding_alignment = nic_align_end -
 545                         MIN(buf_aligned, 1u << (rte_bsf32(buf_size) - 1));
 546                 if (buf_size <= end_padding_alignment) {
 547                         sfc_err(sa,
 548                                 "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",
 549                                 mb_pool->name,
 550                                 rte_pktmbuf_data_room_size(mb_pool),
 551                                 RTE_PKTMBUF_HEADROOM, start_alignment,
 552                                 end_padding_alignment);
 553                         return 0;
 554                 }
 555                 buf_size -= end_padding_alignment;
 556         } else {
 557                 /*
 558                  * Start is aligned the same or better than end,
 559                  * just align length.
 560                  */
 561                 buf_size = P2ALIGN(buf_size, nic_align_end);
 562         }
 563
 564         return buf_size;
 565 }
 566
 567 int
 568 sfc_rx_qinit(struct sfc_adapter *sa, unsigned int sw_index,
 569              uint16_t nb_rx_desc, unsigned int socket_id,
 570              const struct rte_eth_rxconf *rx_conf,
 571              struct rte_mempool *mb_pool)
 572 {
 573         const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
 574         int rc;
 575         uint16_t buf_size;
 576         struct sfc_rxq_info *rxq_info;
 577         unsigned int evq_index;
 578         struct sfc_evq *evq;
 579         struct sfc_rxq *rxq;
 580
 581         rc = sfc_rx_qcheck_conf(sa, nb_rx_desc, rx_conf);
 582         if (rc != 0)
 583                 goto fail_bad_conf;
 584
 585         buf_size = sfc_rx_mb_pool_buf_size(sa, mb_pool);
 586         if (buf_size == 0) {
 587                 sfc_err(sa, "RxQ %u mbuf pool object size is too small",
 588                         sw_index);
 589                 rc = EINVAL;
 590                 goto fail_bad_conf;
 591         }
 592
 593         if ((buf_size < sa->port.pdu + encp->enc_rx_prefix_size) &&
 594             !sa->eth_dev->data->dev_conf.rxmode.enable_scatter) {
 595                 sfc_err(sa, "Rx scatter is disabled and RxQ %u mbuf pool "
 596                         "object size is too small", sw_index);
 597                 sfc_err(sa, "RxQ %u calculated Rx buffer size is %u vs "
 598                         "PDU size %u plus Rx prefix %u bytes",
 599                         sw_index, buf_size, (unsigned int)sa->port.pdu,
 600                         encp->enc_rx_prefix_size);
 601                 rc = EINVAL;
 602                 goto fail_bad_conf;
 603         }
 604
 605         SFC_ASSERT(sw_index < sa->rxq_count);
 606         rxq_info = &sa->rxq_info[sw_index];
 607
 608         SFC_ASSERT(nb_rx_desc <= rxq_info->max_entries);
 609         rxq_info->entries = nb_rx_desc;
 610         rxq_info->type = EFX_RXQ_TYPE_DEFAULT;
 611
 612         evq_index = sfc_evq_index_by_rxq_sw_index(sa, sw_index);
 613
 614         rc = sfc_ev_qinit(sa, evq_index, rxq_info->entries, socket_id);
 615         if (rc != 0)
 616                 goto fail_ev_qinit;
 617
 618         evq = sa->evq_info[evq_index].evq;
 619
 620         rc = ENOMEM;
 621         rxq = rte_zmalloc_socket("sfc-rxq", sizeof(*rxq), RTE_CACHE_LINE_SIZE,
 622                                  socket_id);
 623         if (rxq == NULL)
 624                 goto fail_rxq_alloc;
 625
 626         rc = sfc_dma_alloc(sa, "rxq", sw_index, EFX_RXQ_SIZE(rxq_info->entries),
 627                            socket_id, &rxq->mem);
 628         if (rc != 0)
 629                 goto fail_dma_alloc;
 630
 631         rc = ENOMEM;
 632         rxq->sw_desc = rte_calloc_socket("sfc-rxq-sw_desc", rxq_info->entries,
 633                                          sizeof(*rxq->sw_desc),
 634                                          RTE_CACHE_LINE_SIZE, socket_id);
 635         if (rxq->sw_desc == NULL)
 636                 goto fail_desc_alloc;
 637
 638         evq->rxq = rxq;
 639         rxq->evq = evq;
 640         rxq->ptr_mask = rxq_info->entries - 1;
 641         rxq->refill_threshold = rx_conf->rx_free_thresh;
 642         rxq->refill_mb_pool = mb_pool;
 643         rxq->buf_size = buf_size;
 644         rxq->hw_index = sw_index;
 645         rxq->port_id = sa->eth_dev->data->port_id;
 646
 647         /* Cache limits required on datapath in RxQ structure */
 648         rxq->batch_max = encp->enc_rx_batch_max;
 649         rxq->prefix_size = encp->enc_rx_prefix_size;
 650
 651         rxq->state = SFC_RXQ_INITIALIZED;
 652
 653         rxq_info->rxq = rxq;
 654
 655         return 0;
 656
 657 fail_desc_alloc:
 658         sfc_dma_free(sa, &rxq->mem);
 659
 660 fail_dma_alloc:
 661         rte_free(rxq);
 662
 663 fail_rxq_alloc:
 664         sfc_ev_qfini(sa, evq_index);
 665
 666 fail_ev_qinit:
 667         rxq_info->entries = 0;
 668
 669 fail_bad_conf:
 670         sfc_log_init(sa, "failed %d", rc);
 671         return rc;
 672 }
 673
 674 void
 675 sfc_rx_qfini(struct sfc_adapter *sa, unsigned int sw_index)
 676 {
 677         struct sfc_rxq_info *rxq_info;
 678         struct sfc_rxq *rxq;
 679
 680         SFC_ASSERT(sw_index < sa->rxq_count);
 681
 682         rxq_info = &sa->rxq_info[sw_index];
 683
 684         rxq = rxq_info->rxq;
 685         SFC_ASSERT(rxq->state == SFC_RXQ_INITIALIZED);
 686
 687         rxq_info->rxq = NULL;
 688         rxq_info->entries = 0;
 689
 690         rte_free(rxq->sw_desc);
 691         sfc_dma_free(sa, &rxq->mem);
 692         rte_free(rxq);
 693 }
 694
 695 int
 696 sfc_rx_start(struct sfc_adapter *sa)
 697 {
 698         unsigned int sw_index;
 699         int rc;
 700
 701         sfc_log_init(sa, "rxq_count=%u", sa->rxq_count);
 702
 703         rc = efx_rx_init(sa->nic);
 704         if (rc != 0)
 705                 goto fail_rx_init;
 706
 707         for (sw_index = 0; sw_index < sa->rxq_count; ++sw_index) {
 708                 rc = sfc_rx_qstart(sa, sw_index);
 709                 if (rc != 0)
 710                         goto fail_rx_qstart;
 711         }
 712
 713         return 0;
 714
 715 fail_rx_qstart:
 716         while (sw_index-- > 0)
 717                 sfc_rx_qstop(sa, sw_index);
 718
 719         efx_rx_fini(sa->nic);
 720
 721 fail_rx_init:
 722         sfc_log_init(sa, "failed %d", rc);
 723         return rc;
 724 }
 725
 726 void
 727 sfc_rx_stop(struct sfc_adapter *sa)
 728 {
 729         unsigned int sw_index;
 730
 731         sfc_log_init(sa, "rxq_count=%u", sa->rxq_count);
 732
 733         sw_index = sa->rxq_count;
 734         while (sw_index-- > 0) {
 735                 if (sa->rxq_info[sw_index].rxq != NULL)
 736                         sfc_rx_qstop(sa, sw_index);
 737         }
 738
 739         efx_rx_fini(sa->nic);
 740 }
 741
 742 static int
 743 sfc_rx_qinit_info(struct sfc_adapter *sa, unsigned int sw_index)
 744 {
 745         struct sfc_rxq_info *rxq_info = &sa->rxq_info[sw_index];
 746         unsigned int max_entries;
 747
 748         max_entries = EFX_RXQ_MAXNDESCS;
 749         SFC_ASSERT(rte_is_power_of_2(max_entries));
 750
 751         rxq_info->max_entries = max_entries;
 752
 753         return 0;
 754 }
 755
 756 static int
 757 sfc_rx_check_mode(struct sfc_adapter *sa, struct rte_eth_rxmode *rxmode)
 758 {
 759         int rc = 0;
 760
 761         switch (rxmode->mq_mode) {
 762         case ETH_MQ_RX_NONE:
 763                 /* No special checks are required */
 764                 break;
 765         default:
 766                 sfc_err(sa, "Rx multi-queue mode %u not supported",
 767                         rxmode->mq_mode);
 768                 rc = EINVAL;
 769         }
 770
 771         if (rxmode->header_split) {
 772                 sfc_err(sa, "Header split on Rx not supported");
 773                 rc = EINVAL;
 774         }
 775
 776         if (rxmode->hw_vlan_filter) {
 777                 sfc_err(sa, "HW VLAN filtering not supported");
 778                 rc = EINVAL;
 779         }
 780
 781         if (rxmode->hw_vlan_strip) {
 782                 sfc_err(sa, "HW VLAN stripping not supported");
 783                 rc = EINVAL;
 784         }
 785
 786         if (rxmode->hw_vlan_extend) {
 787                 sfc_err(sa,
 788                         "Q-in-Q HW VLAN stripping not supported");
 789                 rc = EINVAL;
 790         }
 791
 792         if (!rxmode->hw_strip_crc) {
 793                 sfc_warn(sa,
 794                          "FCS stripping control not supported - always stripped");
 795                 rxmode->hw_strip_crc = 1;
 796         }
 797
 798         if (rxmode->enable_scatter) {
 799                 sfc_err(sa, "Scatter on Rx not supported");
 800                 rc = EINVAL;
 801         }
 802
 803         if (rxmode->enable_lro) {
 804                 sfc_err(sa, "LRO not supported");
 805                 rc = EINVAL;
 806         }
 807
 808         return rc;
 809 }
 810
 811 /**
 812  * Initialize Rx subsystem.
 813  *
 814  * Called at device configuration stage when number of receive queues is
 815  * specified together with other device level receive configuration.
 816  *
 817  * It should be used to allocate NUMA-unaware resources.
 818  */
 819 int
 820 sfc_rx_init(struct sfc_adapter *sa)
 821 {
 822         struct rte_eth_conf *dev_conf = &sa->eth_dev->data->dev_conf;
 823         unsigned int sw_index;
 824         int rc;
 825
 826         rc = sfc_rx_check_mode(sa, &dev_conf->rxmode);
 827         if (rc != 0)
 828                 goto fail_check_mode;
 829
 830         sa->rxq_count = sa->eth_dev->data->nb_rx_queues;
 831
 832         rc = ENOMEM;
 833         sa->rxq_info = rte_calloc_socket("sfc-rxqs", sa->rxq_count,
 834                                          sizeof(struct sfc_rxq_info), 0,
 835                                          sa->socket_id);
 836         if (sa->rxq_info == NULL)
 837                 goto fail_rxqs_alloc;
 838
 839         for (sw_index = 0; sw_index < sa->rxq_count; ++sw_index) {
 840                 rc = sfc_rx_qinit_info(sa, sw_index);
 841                 if (rc != 0)
 842                         goto fail_rx_qinit_info;
 843         }
 844
 845         return 0;
 846
 847 fail_rx_qinit_info:
 848         rte_free(sa->rxq_info);
 849         sa->rxq_info = NULL;
 850
 851 fail_rxqs_alloc:
 852         sa->rxq_count = 0;
 853 fail_check_mode:
 854         sfc_log_init(sa, "failed %d", rc);
 855         return rc;
 856 }
 857
 858 /**
 859  * Shutdown Rx subsystem.
 860  *
 861  * Called at device close stage, for example, before device
 862  * reconfiguration or shutdown.
 863  */
 864 void
 865 sfc_rx_fini(struct sfc_adapter *sa)
 866 {
 867         unsigned int sw_index;
 868
 869         sw_index = sa->rxq_count;
 870         while (sw_index-- > 0) {
 871                 if (sa->rxq_info[sw_index].rxq != NULL)
 872                         sfc_rx_qfini(sa, sw_index);
 873         }
 874
 875         rte_free(sa->rxq_info);
 876         sa->rxq_info = NULL;
 877         sa->rxq_count = 0;
 878 }