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         struct rte_mbuf *scatter_pkt = NULL;
 197
 198         if (unlikely((rxq->state & SFC_RXQ_RUNNING) == 0))
 199                 return 0;
 200
 201         sfc_ev_qpoll(rxq->evq);
 202
 203         completed = rxq->completed;
 204         while (completed != rxq->pending && done_pkts < nb_pkts) {
 205                 unsigned int id;
 206                 struct sfc_rx_sw_desc *rxd;
 207                 struct rte_mbuf *m;
 208                 unsigned int seg_len;
 209                 unsigned int desc_flags;
 210
 211                 id = completed++ & rxq->ptr_mask;
 212                 rxd = &rxq->sw_desc[id];
 213                 m = rxd->mbuf;
 214                 desc_flags = rxd->flags;
 215
 216                 if (discard_next)
 217                         goto discard;
 218
 219                 if (desc_flags & (EFX_ADDR_MISMATCH | EFX_DISCARD))
 220                         goto discard;
 221
 222                 if (desc_flags & EFX_PKT_PREFIX_LEN) {
 223                         uint16_t tmp_size;
 224                         int rc __rte_unused;
 225
 226                         rc = efx_pseudo_hdr_pkt_length_get(rxq->common,
 227                                 rte_pktmbuf_mtod(m, uint8_t *), &tmp_size);
 228                         SFC_ASSERT(rc == 0);
 229                         seg_len = tmp_size;
 230                 } else {
 231                         seg_len = rxd->size - prefix_size;
 232                 }
 233
 234                 m->data_off += prefix_size;
 235                 rte_pktmbuf_data_len(m) = seg_len;
 236                 rte_pktmbuf_pkt_len(m) = seg_len;
 237
 238                 if (scatter_pkt != NULL) {
 239                         if (rte_pktmbuf_chain(scatter_pkt, m) != 0) {
 240                                 rte_mempool_put(rxq->refill_mb_pool,
 241                                                 scatter_pkt);
 242                                 goto discard;
 243                         }
 244                         /* The packet to deliver */
 245                         m = scatter_pkt;
 246                 }
 247
 248                 if (desc_flags & EFX_PKT_CONT) {
 249                         /* The packet is scattered, more fragments to come */
 250                         scatter_pkt = m;
 251                         /* Futher fragments have no prefix */
 252                         prefix_size = 0;
 253                         continue;
 254                 }
 255
 256                 /* Scattered packet is done */
 257                 scatter_pkt = NULL;
 258                 /* The first fragment of the packet has prefix */
 259                 prefix_size = rxq->prefix_size;
 260
 261                 m->ol_flags = sfc_rx_desc_flags_to_offload_flags(desc_flags);
 262                 m->packet_type = sfc_rx_desc_flags_to_packet_type(desc_flags);
 263
 264                 *rx_pkts++ = m;
 265                 done_pkts++;
 266                 continue;
 267
 268 discard:
 269                 discard_next = ((desc_flags & EFX_PKT_CONT) != 0);
 270                 rte_mempool_put(rxq->refill_mb_pool, m);
 271                 rxd->mbuf = NULL;
 272         }
 273
 274         /* pending is only moved when entire packet is received */
 275         SFC_ASSERT(scatter_pkt == NULL);
 276
 277         rxq->completed = completed;
 278
 279         sfc_rx_qrefill(rxq);
 280
 281         return done_pkts;
 282 }
 283
 284 unsigned int
 285 sfc_rx_qdesc_npending(struct sfc_adapter *sa, unsigned int sw_index)
 286 {
 287         struct sfc_rxq *rxq;
 288
 289         SFC_ASSERT(sw_index < sa->rxq_count);
 290         rxq = sa->rxq_info[sw_index].rxq;
 291
 292         if (rxq == NULL || (rxq->state & SFC_RXQ_RUNNING) == 0)
 293                 return 0;
 294
 295         sfc_ev_qpoll(rxq->evq);
 296
 297         return rxq->pending - rxq->completed;
 298 }
 299
 300 int
 301 sfc_rx_qdesc_done(struct sfc_rxq *rxq, unsigned int offset)
 302 {
 303         if ((rxq->state & SFC_RXQ_RUNNING) == 0)
 304                 return 0;
 305
 306         sfc_ev_qpoll(rxq->evq);
 307
 308         return offset < (rxq->pending - rxq->completed);
 309 }
 310
 311 static void
 312 sfc_rx_qpurge(struct sfc_rxq *rxq)
 313 {
 314         unsigned int i;
 315         struct sfc_rx_sw_desc *rxd;
 316
 317         for (i = rxq->completed; i != rxq->added; ++i) {
 318                 rxd = &rxq->sw_desc[i & rxq->ptr_mask];
 319                 rte_mempool_put(rxq->refill_mb_pool, rxd->mbuf);
 320                 rxd->mbuf = NULL;
 321         }
 322 }
 323
 324 static void
 325 sfc_rx_qflush(struct sfc_adapter *sa, unsigned int sw_index)
 326 {
 327         struct sfc_rxq *rxq;
 328         unsigned int retry_count;
 329         unsigned int wait_count;
 330
 331         rxq = sa->rxq_info[sw_index].rxq;
 332         SFC_ASSERT(rxq->state & SFC_RXQ_STARTED);
 333
 334         /*
 335          * Retry Rx queue flushing in the case of flush failed or
 336          * timeout. In the worst case it can delay for 6 seconds.
 337          */
 338         for (retry_count = 0;
 339              ((rxq->state & SFC_RXQ_FLUSHED) == 0) &&
 340              (retry_count < SFC_RX_QFLUSH_ATTEMPTS);
 341              ++retry_count) {
 342                 if (efx_rx_qflush(rxq->common) != 0) {
 343                         rxq->state |= SFC_RXQ_FLUSH_FAILED;
 344                         break;
 345                 }
 346                 rxq->state &= ~SFC_RXQ_FLUSH_FAILED;
 347                 rxq->state |= SFC_RXQ_FLUSHING;
 348
 349                 /*
 350                  * Wait for Rx queue flush done or failed event at least
 351                  * SFC_RX_QFLUSH_POLL_WAIT_MS milliseconds and not more
 352                  * than 2 seconds (SFC_RX_QFLUSH_POLL_WAIT_MS multiplied
 353                  * by SFC_RX_QFLUSH_POLL_ATTEMPTS).
 354                  */
 355                 wait_count = 0;
 356                 do {
 357                         rte_delay_ms(SFC_RX_QFLUSH_POLL_WAIT_MS);
 358                         sfc_ev_qpoll(rxq->evq);
 359                 } while ((rxq->state & SFC_RXQ_FLUSHING) &&
 360                          (wait_count++ < SFC_RX_QFLUSH_POLL_ATTEMPTS));
 361
 362                 if (rxq->state & SFC_RXQ_FLUSHING)
 363                         sfc_err(sa, "RxQ %u flush timed out", sw_index);
 364
 365                 if (rxq->state & SFC_RXQ_FLUSH_FAILED)
 366                         sfc_err(sa, "RxQ %u flush failed", sw_index);
 367
 368                 if (rxq->state & SFC_RXQ_FLUSHED)
 369                         sfc_info(sa, "RxQ %u flushed", sw_index);
 370         }
 371
 372         sfc_rx_qpurge(rxq);
 373 }
 374
 375 int
 376 sfc_rx_qstart(struct sfc_adapter *sa, unsigned int sw_index)
 377 {
 378         struct sfc_rxq_info *rxq_info;
 379         struct sfc_rxq *rxq;
 380         struct sfc_evq *evq;
 381         int rc;
 382
 383         sfc_log_init(sa, "sw_index=%u", sw_index);
 384
 385         SFC_ASSERT(sw_index < sa->rxq_count);
 386
 387         rxq_info = &sa->rxq_info[sw_index];
 388         rxq = rxq_info->rxq;
 389         SFC_ASSERT(rxq->state == SFC_RXQ_INITIALIZED);
 390
 391         evq = rxq->evq;
 392
 393         rc = sfc_ev_qstart(sa, evq->evq_index);
 394         if (rc != 0)
 395                 goto fail_ev_qstart;
 396
 397         rc = efx_rx_qcreate(sa->nic, rxq->hw_index, 0, rxq_info->type,
 398                             &rxq->mem, rxq_info->entries,
 399                             0 /* not used on EF10 */, evq->common,
 400                             &rxq->common);
 401         if (rc != 0)
 402                 goto fail_rx_qcreate;
 403
 404         efx_rx_qenable(rxq->common);
 405
 406         rxq->pending = rxq->completed = rxq->added = rxq->pushed = 0;
 407
 408         rxq->state |= (SFC_RXQ_STARTED | SFC_RXQ_RUNNING);
 409
 410         sfc_rx_qrefill(rxq);
 411
 412         if (sw_index == 0) {
 413                 rc = efx_mac_filter_default_rxq_set(sa->nic, rxq->common,
 414                                                     B_FALSE);
 415                 if (rc != 0)
 416                         goto fail_mac_filter_default_rxq_set;
 417         }
 418
 419         /* It seems to be used by DPDK for debug purposes only ('rte_ether') */
 420         sa->eth_dev->data->rx_queue_state[sw_index] =
 421                 RTE_ETH_QUEUE_STATE_STARTED;
 422
 423         return 0;
 424
 425 fail_mac_filter_default_rxq_set:
 426         sfc_rx_qflush(sa, sw_index);
 427
 428 fail_rx_qcreate:
 429         sfc_ev_qstop(sa, evq->evq_index);
 430
 431 fail_ev_qstart:
 432         return rc;
 433 }
 434
 435 void
 436 sfc_rx_qstop(struct sfc_adapter *sa, unsigned int sw_index)
 437 {
 438         struct sfc_rxq_info *rxq_info;
 439         struct sfc_rxq *rxq;
 440
 441         sfc_log_init(sa, "sw_index=%u", sw_index);
 442
 443         SFC_ASSERT(sw_index < sa->rxq_count);
 444
 445         rxq_info = &sa->rxq_info[sw_index];
 446         rxq = rxq_info->rxq;
 447         SFC_ASSERT(rxq->state & SFC_RXQ_STARTED);
 448
 449         /* It seems to be used by DPDK for debug purposes only ('rte_ether') */
 450         sa->eth_dev->data->rx_queue_state[sw_index] =
 451                 RTE_ETH_QUEUE_STATE_STOPPED;
 452
 453         rxq->state &= ~SFC_RXQ_RUNNING;
 454
 455         if (sw_index == 0)
 456                 efx_mac_filter_default_rxq_clear(sa->nic);
 457
 458         sfc_rx_qflush(sa, sw_index);
 459
 460         rxq->state = SFC_RXQ_INITIALIZED;
 461
 462         efx_rx_qdestroy(rxq->common);
 463
 464         sfc_ev_qstop(sa, rxq->evq->evq_index);
 465 }
 466
 467 static int
 468 sfc_rx_qcheck_conf(struct sfc_adapter *sa, uint16_t nb_rx_desc,
 469                    const struct rte_eth_rxconf *rx_conf)
 470 {
 471         const uint16_t rx_free_thresh_max = EFX_RXQ_LIMIT(nb_rx_desc);
 472         int rc = 0;
 473
 474         if (rx_conf->rx_thresh.pthresh != 0 ||
 475             rx_conf->rx_thresh.hthresh != 0 ||
 476             rx_conf->rx_thresh.wthresh != 0) {
 477                 sfc_err(sa,
 478                         "RxQ prefetch/host/writeback thresholds are not supported");
 479                 rc = EINVAL;
 480         }
 481
 482         if (rx_conf->rx_free_thresh > rx_free_thresh_max) {
 483                 sfc_err(sa,
 484                         "RxQ free threshold too large: %u vs maximum %u",
 485                         rx_conf->rx_free_thresh, rx_free_thresh_max);
 486                 rc = EINVAL;
 487         }
 488
 489         if (rx_conf->rx_drop_en == 0) {
 490                 sfc_err(sa, "RxQ drop disable is not supported");
 491                 rc = EINVAL;
 492         }
 493
 494         if (rx_conf->rx_deferred_start != 0) {
 495                 sfc_err(sa, "RxQ deferred start is not supported");
 496                 rc = EINVAL;
 497         }
 498
 499         return rc;
 500 }
 501
 502 static unsigned int
 503 sfc_rx_mbuf_data_alignment(struct rte_mempool *mb_pool)
 504 {
 505         uint32_t data_off;
 506         uint32_t order;
 507
 508         /* The mbuf object itself is always cache line aligned */
 509         order = rte_bsf32(RTE_CACHE_LINE_SIZE);
 510
 511         /* Data offset from mbuf object start */
 512         data_off = sizeof(struct rte_mbuf) + rte_pktmbuf_priv_size(mb_pool) +
 513                 RTE_PKTMBUF_HEADROOM;
 514
 515         order = MIN(order, rte_bsf32(data_off));
 516
 517         return 1u << (order - 1);
 518 }
 519
 520 static uint16_t
 521 sfc_rx_mb_pool_buf_size(struct sfc_adapter *sa, struct rte_mempool *mb_pool)
 522 {
 523         const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
 524         const uint32_t nic_align_start = MAX(1, encp->enc_rx_buf_align_start);
 525         const uint32_t nic_align_end = MAX(1, encp->enc_rx_buf_align_end);
 526         uint16_t buf_size;
 527         unsigned int buf_aligned;
 528         unsigned int start_alignment;
 529         unsigned int end_padding_alignment;
 530
 531         /* Below it is assumed that both alignments are power of 2 */
 532         SFC_ASSERT(rte_is_power_of_2(nic_align_start));
 533         SFC_ASSERT(rte_is_power_of_2(nic_align_end));
 534
 535         /*
 536          * mbuf is always cache line aligned, double-check
 537          * that it meets rx buffer start alignment requirements.
 538          */
 539
 540         /* Start from mbuf pool data room size */
 541         buf_size = rte_pktmbuf_data_room_size(mb_pool);
 542
 543         /* Remove headroom */
 544         if (buf_size <= RTE_PKTMBUF_HEADROOM) {
 545                 sfc_err(sa,
 546                         "RxQ mbuf pool %s object data room size %u is smaller than headroom %u",
 547                         mb_pool->name, buf_size, RTE_PKTMBUF_HEADROOM);
 548                 return 0;
 549         }
 550         buf_size -= RTE_PKTMBUF_HEADROOM;
 551
 552         /* Calculate guaranteed data start alignment */
 553         buf_aligned = sfc_rx_mbuf_data_alignment(mb_pool);
 554
 555         /* Reserve space for start alignment */
 556         if (buf_aligned < nic_align_start) {
 557                 start_alignment = nic_align_start - buf_aligned;
 558                 if (buf_size <= start_alignment) {
 559                         sfc_err(sa,
 560                                 "RxQ mbuf pool %s object data room size %u is insufficient for headroom %u and buffer start alignment %u required by NIC",
 561                                 mb_pool->name,
 562                                 rte_pktmbuf_data_room_size(mb_pool),
 563                                 RTE_PKTMBUF_HEADROOM, start_alignment);
 564                         return 0;
 565                 }
 566                 buf_aligned = nic_align_start;
 567                 buf_size -= start_alignment;
 568         } else {
 569                 start_alignment = 0;
 570         }
 571
 572         /* Make sure that end padding does not write beyond the buffer */
 573         if (buf_aligned < nic_align_end) {
 574                 /*
 575                  * Estimate space which can be lost. If guarnteed buffer
 576                  * size is odd, lost space is (nic_align_end - 1). More
 577                  * accurate formula is below.
 578                  */
 579                 end_padding_alignment = nic_align_end -
 580                         MIN(buf_aligned, 1u << (rte_bsf32(buf_size) - 1));
 581                 if (buf_size <= end_padding_alignment) {
 582                         sfc_err(sa,
 583                                 "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",
 584                                 mb_pool->name,
 585                                 rte_pktmbuf_data_room_size(mb_pool),
 586                                 RTE_PKTMBUF_HEADROOM, start_alignment,
 587                                 end_padding_alignment);
 588                         return 0;
 589                 }
 590                 buf_size -= end_padding_alignment;
 591         } else {
 592                 /*
 593                  * Start is aligned the same or better than end,
 594                  * just align length.
 595                  */
 596                 buf_size = P2ALIGN(buf_size, nic_align_end);
 597         }
 598
 599         return buf_size;
 600 }
 601
 602 int
 603 sfc_rx_qinit(struct sfc_adapter *sa, unsigned int sw_index,
 604              uint16_t nb_rx_desc, unsigned int socket_id,
 605              const struct rte_eth_rxconf *rx_conf,
 606              struct rte_mempool *mb_pool)
 607 {
 608         const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
 609         int rc;
 610         uint16_t buf_size;
 611         struct sfc_rxq_info *rxq_info;
 612         unsigned int evq_index;
 613         struct sfc_evq *evq;
 614         struct sfc_rxq *rxq;
 615
 616         rc = sfc_rx_qcheck_conf(sa, nb_rx_desc, rx_conf);
 617         if (rc != 0)
 618                 goto fail_bad_conf;
 619
 620         buf_size = sfc_rx_mb_pool_buf_size(sa, mb_pool);
 621         if (buf_size == 0) {
 622                 sfc_err(sa, "RxQ %u mbuf pool object size is too small",
 623                         sw_index);
 624                 rc = EINVAL;
 625                 goto fail_bad_conf;
 626         }
 627
 628         if ((buf_size < sa->port.pdu + encp->enc_rx_prefix_size) &&
 629             !sa->eth_dev->data->dev_conf.rxmode.enable_scatter) {
 630                 sfc_err(sa, "Rx scatter is disabled and RxQ %u mbuf pool "
 631                         "object size is too small", sw_index);
 632                 sfc_err(sa, "RxQ %u calculated Rx buffer size is %u vs "
 633                         "PDU size %u plus Rx prefix %u bytes",
 634                         sw_index, buf_size, (unsigned int)sa->port.pdu,
 635                         encp->enc_rx_prefix_size);
 636                 rc = EINVAL;
 637                 goto fail_bad_conf;
 638         }
 639
 640         SFC_ASSERT(sw_index < sa->rxq_count);
 641         rxq_info = &sa->rxq_info[sw_index];
 642
 643         SFC_ASSERT(nb_rx_desc <= rxq_info->max_entries);
 644         rxq_info->entries = nb_rx_desc;
 645         rxq_info->type =
 646                 sa->eth_dev->data->dev_conf.rxmode.enable_scatter ?
 647                 EFX_RXQ_TYPE_SCATTER : EFX_RXQ_TYPE_DEFAULT;
 648
 649         evq_index = sfc_evq_index_by_rxq_sw_index(sa, sw_index);
 650
 651         rc = sfc_ev_qinit(sa, evq_index, rxq_info->entries, socket_id);
 652         if (rc != 0)
 653                 goto fail_ev_qinit;
 654
 655         evq = sa->evq_info[evq_index].evq;
 656
 657         rc = ENOMEM;
 658         rxq = rte_zmalloc_socket("sfc-rxq", sizeof(*rxq), RTE_CACHE_LINE_SIZE,
 659                                  socket_id);
 660         if (rxq == NULL)
 661                 goto fail_rxq_alloc;
 662
 663         rc = sfc_dma_alloc(sa, "rxq", sw_index, EFX_RXQ_SIZE(rxq_info->entries),
 664                            socket_id, &rxq->mem);
 665         if (rc != 0)
 666                 goto fail_dma_alloc;
 667
 668         rc = ENOMEM;
 669         rxq->sw_desc = rte_calloc_socket("sfc-rxq-sw_desc", rxq_info->entries,
 670                                          sizeof(*rxq->sw_desc),
 671                                          RTE_CACHE_LINE_SIZE, socket_id);
 672         if (rxq->sw_desc == NULL)
 673                 goto fail_desc_alloc;
 674
 675         evq->rxq = rxq;
 676         rxq->evq = evq;
 677         rxq->ptr_mask = rxq_info->entries - 1;
 678         rxq->refill_threshold = rx_conf->rx_free_thresh;
 679         rxq->refill_mb_pool = mb_pool;
 680         rxq->buf_size = buf_size;
 681         rxq->hw_index = sw_index;
 682         rxq->port_id = sa->eth_dev->data->port_id;
 683
 684         /* Cache limits required on datapath in RxQ structure */
 685         rxq->batch_max = encp->enc_rx_batch_max;
 686         rxq->prefix_size = encp->enc_rx_prefix_size;
 687
 688         rxq->state = SFC_RXQ_INITIALIZED;
 689
 690         rxq_info->rxq = rxq;
 691
 692         return 0;
 693
 694 fail_desc_alloc:
 695         sfc_dma_free(sa, &rxq->mem);
 696
 697 fail_dma_alloc:
 698         rte_free(rxq);
 699
 700 fail_rxq_alloc:
 701         sfc_ev_qfini(sa, evq_index);
 702
 703 fail_ev_qinit:
 704         rxq_info->entries = 0;
 705
 706 fail_bad_conf:
 707         sfc_log_init(sa, "failed %d", rc);
 708         return rc;
 709 }
 710
 711 void
 712 sfc_rx_qfini(struct sfc_adapter *sa, unsigned int sw_index)
 713 {
 714         struct sfc_rxq_info *rxq_info;
 715         struct sfc_rxq *rxq;
 716
 717         SFC_ASSERT(sw_index < sa->rxq_count);
 718
 719         rxq_info = &sa->rxq_info[sw_index];
 720
 721         rxq = rxq_info->rxq;
 722         SFC_ASSERT(rxq->state == SFC_RXQ_INITIALIZED);
 723
 724         rxq_info->rxq = NULL;
 725         rxq_info->entries = 0;
 726
 727         rte_free(rxq->sw_desc);
 728         sfc_dma_free(sa, &rxq->mem);
 729         rte_free(rxq);
 730 }
 731
 732 int
 733 sfc_rx_start(struct sfc_adapter *sa)
 734 {
 735         unsigned int sw_index;
 736         int rc;
 737
 738         sfc_log_init(sa, "rxq_count=%u", sa->rxq_count);
 739
 740         rc = efx_rx_init(sa->nic);
 741         if (rc != 0)
 742                 goto fail_rx_init;
 743
 744         for (sw_index = 0; sw_index < sa->rxq_count; ++sw_index) {
 745                 rc = sfc_rx_qstart(sa, sw_index);
 746                 if (rc != 0)
 747                         goto fail_rx_qstart;
 748         }
 749
 750         return 0;
 751
 752 fail_rx_qstart:
 753         while (sw_index-- > 0)
 754                 sfc_rx_qstop(sa, sw_index);
 755
 756         efx_rx_fini(sa->nic);
 757
 758 fail_rx_init:
 759         sfc_log_init(sa, "failed %d", rc);
 760         return rc;
 761 }
 762
 763 void
 764 sfc_rx_stop(struct sfc_adapter *sa)
 765 {
 766         unsigned int sw_index;
 767
 768         sfc_log_init(sa, "rxq_count=%u", sa->rxq_count);
 769
 770         sw_index = sa->rxq_count;
 771         while (sw_index-- > 0) {
 772                 if (sa->rxq_info[sw_index].rxq != NULL)
 773                         sfc_rx_qstop(sa, sw_index);
 774         }
 775
 776         efx_rx_fini(sa->nic);
 777 }
 778
 779 static int
 780 sfc_rx_qinit_info(struct sfc_adapter *sa, unsigned int sw_index)
 781 {
 782         struct sfc_rxq_info *rxq_info = &sa->rxq_info[sw_index];
 783         unsigned int max_entries;
 784
 785         max_entries = EFX_RXQ_MAXNDESCS;
 786         SFC_ASSERT(rte_is_power_of_2(max_entries));
 787
 788         rxq_info->max_entries = max_entries;
 789
 790         return 0;
 791 }
 792
 793 static int
 794 sfc_rx_check_mode(struct sfc_adapter *sa, struct rte_eth_rxmode *rxmode)
 795 {
 796         int rc = 0;
 797
 798         switch (rxmode->mq_mode) {
 799         case ETH_MQ_RX_NONE:
 800                 /* No special checks are required */
 801                 break;
 802         default:
 803                 sfc_err(sa, "Rx multi-queue mode %u not supported",
 804                         rxmode->mq_mode);
 805                 rc = EINVAL;
 806         }
 807
 808         if (rxmode->header_split) {
 809                 sfc_err(sa, "Header split on Rx not supported");
 810                 rc = EINVAL;
 811         }
 812
 813         if (rxmode->hw_vlan_filter) {
 814                 sfc_err(sa, "HW VLAN filtering not supported");
 815                 rc = EINVAL;
 816         }
 817
 818         if (rxmode->hw_vlan_strip) {
 819                 sfc_err(sa, "HW VLAN stripping not supported");
 820                 rc = EINVAL;
 821         }
 822
 823         if (rxmode->hw_vlan_extend) {
 824                 sfc_err(sa,
 825                         "Q-in-Q HW VLAN stripping not supported");
 826                 rc = EINVAL;
 827         }
 828
 829         if (!rxmode->hw_strip_crc) {
 830                 sfc_warn(sa,
 831                          "FCS stripping control not supported - always stripped");
 832                 rxmode->hw_strip_crc = 1;
 833         }
 834
 835         if (rxmode->enable_lro) {
 836                 sfc_err(sa, "LRO not supported");
 837                 rc = EINVAL;
 838         }
 839
 840         return rc;
 841 }
 842
 843 /**
 844  * Initialize Rx subsystem.
 845  *
 846  * Called at device configuration stage when number of receive queues is
 847  * specified together with other device level receive configuration.
 848  *
 849  * It should be used to allocate NUMA-unaware resources.
 850  */
 851 int
 852 sfc_rx_init(struct sfc_adapter *sa)
 853 {
 854         struct rte_eth_conf *dev_conf = &sa->eth_dev->data->dev_conf;
 855         unsigned int sw_index;
 856         int rc;
 857
 858         rc = sfc_rx_check_mode(sa, &dev_conf->rxmode);
 859         if (rc != 0)
 860                 goto fail_check_mode;
 861
 862         sa->rxq_count = sa->eth_dev->data->nb_rx_queues;
 863
 864         rc = ENOMEM;
 865         sa->rxq_info = rte_calloc_socket("sfc-rxqs", sa->rxq_count,
 866                                          sizeof(struct sfc_rxq_info), 0,
 867                                          sa->socket_id);
 868         if (sa->rxq_info == NULL)
 869                 goto fail_rxqs_alloc;
 870
 871         for (sw_index = 0; sw_index < sa->rxq_count; ++sw_index) {
 872                 rc = sfc_rx_qinit_info(sa, sw_index);
 873                 if (rc != 0)
 874                         goto fail_rx_qinit_info;
 875         }
 876
 877         return 0;
 878
 879 fail_rx_qinit_info:
 880         rte_free(sa->rxq_info);
 881         sa->rxq_info = NULL;
 882
 883 fail_rxqs_alloc:
 884         sa->rxq_count = 0;
 885 fail_check_mode:
 886         sfc_log_init(sa, "failed %d", rc);
 887         return rc;
 888 }
 889
 890 /**
 891  * Shutdown Rx subsystem.
 892  *
 893  * Called at device close stage, for example, before device
 894  * reconfiguration or shutdown.
 895  */
 896 void
 897 sfc_rx_fini(struct sfc_adapter *sa)
 898 {
 899         unsigned int sw_index;
 900
 901         sw_index = sa->rxq_count;
 902         while (sw_index-- > 0) {
 903                 if (sa->rxq_info[sw_index].rxq != NULL)
 904                         sfc_rx_qfini(sa, sw_index);
 905         }
 906
 907         rte_free(sa->rxq_info);
 908         sa->rxq_info = NULL;
 909         sa->rxq_count = 0;
 910 }