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