drivers/net/sfc/sfc_rx.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  * Copyright (c) 2016-2017 Solarflare Communications Inc.
   5  * All rights reserved.
   6  *
   7  * This software was jointly developed between OKTET Labs (under contract
   8  * for Solarflare) and Solarflare Communications, Inc.
   9  *
  10  * Redistribution and use in source and binary forms, with or without
  11  * modification, are permitted provided that the following conditions are met:
  12  *
  13  * 1. Redistributions of source code must retain the above copyright notice,
  14  *    this list of conditions and the following disclaimer.
  15  * 2. Redistributions in binary form must reproduce the above copyright notice,
  16  *    this list of conditions and the following disclaimer in the documentation
  17  *    and/or other materials provided with the distribution.
  18  *
  19  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
  20  * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO,
  21  * THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
  22  * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
  23  * CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
  24  * EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
  25  * PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
  26  * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
  27  * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR
  28  * OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE,
  29  * EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  30  */
  31
  32 #include <rte_mempool.h>
  33
  34 #include "efx.h"
  35
  36 #include "sfc.h"
  37 #include "sfc_debug.h"
  38 #include "sfc_log.h"
  39 #include "sfc_ev.h"
  40 #include "sfc_rx.h"
  41 #include "sfc_tweak.h"
  42
  43 /*
  44  * Maximum number of Rx queue flush attempt in the case of failure or
  45  * flush timeout
  46  */
  47 #define SFC_RX_QFLUSH_ATTEMPTS          (3)
  48
  49 /*
  50  * Time to wait between event queue polling attempts when waiting for Rx
  51  * queue flush done or failed events.
  52  */
  53 #define SFC_RX_QFLUSH_POLL_WAIT_MS      (1)
  54
  55 /*
  56  * Maximum number of event queue polling attempts when waiting for Rx queue
  57  * flush done or failed events. It defines Rx queue flush attempt timeout
  58  * together with SFC_RX_QFLUSH_POLL_WAIT_MS.
  59  */
  60 #define SFC_RX_QFLUSH_POLL_ATTEMPTS     (2000)
  61
  62 void
  63 sfc_rx_qflush_done(struct sfc_rxq *rxq)
  64 {
  65         rxq->state |= SFC_RXQ_FLUSHED;
  66         rxq->state &= ~SFC_RXQ_FLUSHING;
  67 }
  68
  69 void
  70 sfc_rx_qflush_failed(struct sfc_rxq *rxq)
  71 {
  72         rxq->state |= SFC_RXQ_FLUSH_FAILED;
  73         rxq->state &= ~SFC_RXQ_FLUSHING;
  74 }
  75
  76 static void
  77 sfc_rx_qrefill(struct sfc_rxq *rxq)
  78 {
  79         unsigned int free_space;
  80         unsigned int bulks;
  81         void *objs[SFC_RX_REFILL_BULK];
  82         efsys_dma_addr_t addr[RTE_DIM(objs)];
  83         unsigned int added = rxq->added;
  84         unsigned int id;
  85         unsigned int i;
  86         struct sfc_rx_sw_desc *rxd;
  87         struct rte_mbuf *m;
  88         uint8_t port_id = rxq->port_id;
  89
  90         free_space = EFX_RXQ_LIMIT(rxq->ptr_mask + 1) -
  91                 (added - rxq->completed);
  92
  93         if (free_space < rxq->refill_threshold)
  94                 return;
  95
  96         bulks = free_space / RTE_DIM(objs);
  97
  98         id = added & rxq->ptr_mask;
  99         while (bulks-- > 0) {
 100                 if (rte_mempool_get_bulk(rxq->refill_mb_pool, objs,
 101                                          RTE_DIM(objs)) < 0) {
 102                         /*
 103                          * It is hardly a safe way to increment counter
 104                          * from different contexts, but all PMDs do it.
 105                          */
 106                         rxq->evq->sa->eth_dev->data->rx_mbuf_alloc_failed +=
 107                                 RTE_DIM(objs);
 108                         break;
 109                 }
 110
 111                 for (i = 0; i < RTE_DIM(objs);
 112                      ++i, id = (id + 1) & rxq->ptr_mask) {
 113                         m = objs[i];
 114
 115                         rxd = &rxq->sw_desc[id];
 116                         rxd->mbuf = m;
 117
 118                         rte_mbuf_refcnt_set(m, 1);
 119                         m->data_off = RTE_PKTMBUF_HEADROOM;
 120                         m->next = NULL;
 121                         m->nb_segs = 1;
 122                         m->port = port_id;
 123
 124                         addr[i] = rte_pktmbuf_mtophys(m);
 125                 }
 126
 127                 efx_rx_qpost(rxq->common, addr, rxq->buf_size,
 128                              RTE_DIM(objs), rxq->completed, added);
 129                 added += RTE_DIM(objs);
 130         }
 131
 132         /* Push doorbell if something is posted */
 133         if (rxq->added != added) {
 134                 rxq->added = added;
 135                 efx_rx_qpush(rxq->common, added, &rxq->pushed);
 136         }
 137 }
 138
 139 static uint64_t
 140 sfc_rx_desc_flags_to_offload_flags(const unsigned int desc_flags)
 141 {
 142         uint64_t mbuf_flags = 0;
 143
 144         switch (desc_flags & (EFX_PKT_IPV4 | EFX_CKSUM_IPV4)) {
 145         case (EFX_PKT_IPV4 | EFX_CKSUM_IPV4):
 146                 mbuf_flags |= PKT_RX_IP_CKSUM_GOOD;
 147                 break;
 148         case EFX_PKT_IPV4:
 149                 mbuf_flags |= PKT_RX_IP_CKSUM_BAD;
 150                 break;
 151         default:
 152                 RTE_BUILD_BUG_ON(PKT_RX_IP_CKSUM_UNKNOWN != 0);
 153                 SFC_ASSERT((mbuf_flags & PKT_RX_IP_CKSUM_MASK) ==
 154                            PKT_RX_IP_CKSUM_UNKNOWN);
 155                 break;
 156         }
 157
 158         switch ((desc_flags &
 159                  (EFX_PKT_TCP | EFX_PKT_UDP | EFX_CKSUM_TCPUDP))) {
 160         case (EFX_PKT_TCP | EFX_CKSUM_TCPUDP):
 161         case (EFX_PKT_UDP | EFX_CKSUM_TCPUDP):
 162                 mbuf_flags |= PKT_RX_L4_CKSUM_GOOD;
 163                 break;
 164         case EFX_PKT_TCP:
 165         case EFX_PKT_UDP:
 166                 mbuf_flags |= PKT_RX_L4_CKSUM_BAD;
 167                 break;
 168         default:
 169                 RTE_BUILD_BUG_ON(PKT_RX_L4_CKSUM_UNKNOWN != 0);
 170                 SFC_ASSERT((mbuf_flags & PKT_RX_L4_CKSUM_MASK) ==
 171                            PKT_RX_L4_CKSUM_UNKNOWN);
 172                 break;
 173         }
 174
 175         return mbuf_flags;
 176 }
 177
 178 static uint32_t
 179 sfc_rx_desc_flags_to_packet_type(const unsigned int desc_flags)
 180 {
 181         return RTE_PTYPE_L2_ETHER |
 182                 ((desc_flags & EFX_PKT_IPV4) ?
 183                         RTE_PTYPE_L3_IPV4_EXT_UNKNOWN : 0) |
 184                 ((desc_flags & EFX_PKT_IPV6) ?
 185                         RTE_PTYPE_L3_IPV6_EXT_UNKNOWN : 0) |
 186                 ((desc_flags & EFX_PKT_TCP) ? RTE_PTYPE_L4_TCP : 0) |
 187                 ((desc_flags & EFX_PKT_UDP) ? RTE_PTYPE_L4_UDP : 0);
 188 }
 189
 190 static void
 191 sfc_rx_set_rss_hash(struct sfc_rxq *rxq, unsigned int flags, struct rte_mbuf *m)
 192 {
 193 #if EFSYS_OPT_RX_SCALE
 194         uint8_t *mbuf_data;
 195
 196
 197         if ((rxq->flags & SFC_RXQ_FLAG_RSS_HASH) == 0)
 198                 return;
 199
 200         mbuf_data = rte_pktmbuf_mtod(m, uint8_t *);
 201
 202         if (flags & (EFX_PKT_IPV4 | EFX_PKT_IPV6)) {
 203                 m->hash.rss = efx_pseudo_hdr_hash_get(rxq->common,
 204                                                       EFX_RX_HASHALG_TOEPLITZ,
 205                                                       mbuf_data);
 206
 207                 m->ol_flags |= PKT_RX_RSS_HASH;
 208         }
 209 #endif
 210 }
 211
 212 uint16_t
 213 sfc_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 214 {
 215         struct sfc_rxq *rxq = rx_queue;
 216         unsigned int completed;
 217         unsigned int prefix_size = rxq->prefix_size;
 218         unsigned int done_pkts = 0;
 219         boolean_t discard_next = B_FALSE;
 220         struct rte_mbuf *scatter_pkt = NULL;
 221
 222         if (unlikely((rxq->flags & SFC_RXQ_FLAG_RUNNING) == 0))
 223                 return 0;
 224
 225         sfc_ev_qpoll(rxq->evq);
 226
 227         completed = rxq->completed;
 228         while (completed != rxq->pending && done_pkts < nb_pkts) {
 229                 unsigned int id;
 230                 struct sfc_rx_sw_desc *rxd;
 231                 struct rte_mbuf *m;
 232                 unsigned int seg_len;
 233                 unsigned int desc_flags;
 234
 235                 id = completed++ & rxq->ptr_mask;
 236                 rxd = &rxq->sw_desc[id];
 237                 m = rxd->mbuf;
 238                 desc_flags = rxd->flags;
 239
 240                 if (discard_next)
 241                         goto discard;
 242
 243                 if (desc_flags & (EFX_ADDR_MISMATCH | EFX_DISCARD))
 244                         goto discard;
 245
 246                 if (desc_flags & EFX_PKT_PREFIX_LEN) {
 247                         uint16_t tmp_size;
 248                         int rc __rte_unused;
 249
 250                         rc = efx_pseudo_hdr_pkt_length_get(rxq->common,
 251                                 rte_pktmbuf_mtod(m, uint8_t *), &tmp_size);
 252                         SFC_ASSERT(rc == 0);
 253                         seg_len = tmp_size;
 254                 } else {
 255                         seg_len = rxd->size - prefix_size;
 256                 }
 257
 258                 rte_pktmbuf_data_len(m) = seg_len;
 259                 rte_pktmbuf_pkt_len(m) = seg_len;
 260
 261                 if (scatter_pkt != NULL) {
 262                         if (rte_pktmbuf_chain(scatter_pkt, m) != 0) {
 263                                 rte_mempool_put(rxq->refill_mb_pool,
 264                                                 scatter_pkt);
 265                                 goto discard;
 266                         }
 267                         /* The packet to deliver */
 268                         m = scatter_pkt;
 269                 }
 270
 271                 if (desc_flags & EFX_PKT_CONT) {
 272                         /* The packet is scattered, more fragments to come */
 273                         scatter_pkt = m;
 274                         /* Futher fragments have no prefix */
 275                         prefix_size = 0;
 276                         continue;
 277                 }
 278
 279                 /* Scattered packet is done */
 280                 scatter_pkt = NULL;
 281                 /* The first fragment of the packet has prefix */
 282                 prefix_size = rxq->prefix_size;
 283
 284                 m->ol_flags = sfc_rx_desc_flags_to_offload_flags(desc_flags);
 285                 m->packet_type = sfc_rx_desc_flags_to_packet_type(desc_flags);
 286
 287                 /*
 288                  * Extract RSS hash from the packet prefix and
 289                  * set the corresponding field (if needed and possible)
 290                  */
 291                 sfc_rx_set_rss_hash(rxq, desc_flags, m);
 292
 293                 m->data_off += prefix_size;
 294
 295                 *rx_pkts++ = m;
 296                 done_pkts++;
 297                 continue;
 298
 299 discard:
 300                 discard_next = ((desc_flags & EFX_PKT_CONT) != 0);
 301                 rte_mempool_put(rxq->refill_mb_pool, m);
 302                 rxd->mbuf = NULL;
 303         }
 304
 305         /* pending is only moved when entire packet is received */
 306         SFC_ASSERT(scatter_pkt == NULL);
 307
 308         rxq->completed = completed;
 309
 310         sfc_rx_qrefill(rxq);
 311
 312         return done_pkts;
 313 }
 314
 315 unsigned int
 316 sfc_rx_qdesc_npending(struct sfc_adapter *sa, unsigned int sw_index)
 317 {
 318         struct sfc_rxq *rxq;
 319
 320         SFC_ASSERT(sw_index < sa->rxq_count);
 321         rxq = sa->rxq_info[sw_index].rxq;
 322
 323         if (rxq == NULL || (rxq->flags & SFC_RXQ_FLAG_RUNNING) == 0)
 324                 return 0;
 325
 326         sfc_ev_qpoll(rxq->evq);
 327
 328         return rxq->pending - rxq->completed;
 329 }
 330
 331 int
 332 sfc_rx_qdesc_done(struct sfc_rxq *rxq, unsigned int offset)
 333 {
 334         if ((rxq->flags & SFC_RXQ_FLAG_RUNNING) == 0)
 335                 return 0;
 336
 337         sfc_ev_qpoll(rxq->evq);
 338
 339         return offset < (rxq->pending - rxq->completed);
 340 }
 341
 342 static void
 343 sfc_rx_qpurge(struct sfc_rxq *rxq)
 344 {
 345         unsigned int i;
 346         struct sfc_rx_sw_desc *rxd;
 347
 348         for (i = rxq->completed; i != rxq->added; ++i) {
 349                 rxd = &rxq->sw_desc[i & rxq->ptr_mask];
 350                 rte_mempool_put(rxq->refill_mb_pool, rxd->mbuf);
 351                 rxd->mbuf = NULL;
 352         }
 353 }
 354
 355 static void
 356 sfc_rx_qflush(struct sfc_adapter *sa, unsigned int sw_index)
 357 {
 358         struct sfc_rxq *rxq;
 359         unsigned int retry_count;
 360         unsigned int wait_count;
 361
 362         rxq = sa->rxq_info[sw_index].rxq;
 363         SFC_ASSERT(rxq->state & SFC_RXQ_STARTED);
 364
 365         /*
 366          * Retry Rx queue flushing in the case of flush failed or
 367          * timeout. In the worst case it can delay for 6 seconds.
 368          */
 369         for (retry_count = 0;
 370              ((rxq->state & SFC_RXQ_FLUSHED) == 0) &&
 371              (retry_count < SFC_RX_QFLUSH_ATTEMPTS);
 372              ++retry_count) {
 373                 if (efx_rx_qflush(rxq->common) != 0) {
 374                         rxq->state |= SFC_RXQ_FLUSH_FAILED;
 375                         break;
 376                 }
 377                 rxq->state &= ~SFC_RXQ_FLUSH_FAILED;
 378                 rxq->state |= SFC_RXQ_FLUSHING;
 379
 380                 /*
 381                  * Wait for Rx queue flush done or failed event at least
 382                  * SFC_RX_QFLUSH_POLL_WAIT_MS milliseconds and not more
 383                  * than 2 seconds (SFC_RX_QFLUSH_POLL_WAIT_MS multiplied
 384                  * by SFC_RX_QFLUSH_POLL_ATTEMPTS).
 385                  */
 386                 wait_count = 0;
 387                 do {
 388                         rte_delay_ms(SFC_RX_QFLUSH_POLL_WAIT_MS);
 389                         sfc_ev_qpoll(rxq->evq);
 390                 } while ((rxq->state & SFC_RXQ_FLUSHING) &&
 391                          (wait_count++ < SFC_RX_QFLUSH_POLL_ATTEMPTS));
 392
 393                 if (rxq->state & SFC_RXQ_FLUSHING)
 394                         sfc_err(sa, "RxQ %u flush timed out", sw_index);
 395
 396                 if (rxq->state & SFC_RXQ_FLUSH_FAILED)
 397                         sfc_err(sa, "RxQ %u flush failed", sw_index);
 398
 399                 if (rxq->state & SFC_RXQ_FLUSHED)
 400                         sfc_info(sa, "RxQ %u flushed", sw_index);
 401         }
 402
 403         sfc_rx_qpurge(rxq);
 404 }
 405
 406 static int
 407 sfc_rx_default_rxq_set_filter(struct sfc_adapter *sa, struct sfc_rxq *rxq)
 408 {
 409         boolean_t rss = (sa->rss_channels > 1) ? B_TRUE : B_FALSE;
 410         struct sfc_port *port = &sa->port;
 411         int rc;
 412
 413         /*
 414          * If promiscuous or all-multicast mode has been requested, setting
 415          * filter for the default Rx queue might fail, in particular, while
 416          * running over PCI function which is not a member of corresponding
 417          * privilege groups; if this occurs, few iterations will be made to
 418          * repeat this step without promiscuous and all-multicast flags set
 419          */
 420 retry:
 421         rc = efx_mac_filter_default_rxq_set(sa->nic, rxq->common, rss);
 422         if (rc == 0)
 423                 return 0;
 424         else if (rc != EOPNOTSUPP)
 425                 return rc;
 426
 427         if (port->promisc) {
 428                 sfc_warn(sa, "promiscuous mode has been requested, "
 429                              "but the HW rejects it");
 430                 sfc_warn(sa, "promiscuous mode will be disabled");
 431
 432                 port->promisc = B_FALSE;
 433                 rc = sfc_set_rx_mode(sa);
 434                 if (rc != 0)
 435                         return rc;
 436
 437                 goto retry;
 438         }
 439
 440         if (port->allmulti) {
 441                 sfc_warn(sa, "all-multicast mode has been requested, "
 442                              "but the HW rejects it");
 443                 sfc_warn(sa, "all-multicast mode will be disabled");
 444
 445                 port->allmulti = B_FALSE;
 446                 rc = sfc_set_rx_mode(sa);
 447                 if (rc != 0)
 448                         return rc;
 449
 450                 goto retry;
 451         }
 452
 453         return rc;
 454 }
 455
 456 int
 457 sfc_rx_qstart(struct sfc_adapter *sa, unsigned int sw_index)
 458 {
 459         struct sfc_rxq_info *rxq_info;
 460         struct sfc_rxq *rxq;
 461         struct sfc_evq *evq;
 462         int rc;
 463
 464         sfc_log_init(sa, "sw_index=%u", sw_index);
 465
 466         SFC_ASSERT(sw_index < sa->rxq_count);
 467
 468         rxq_info = &sa->rxq_info[sw_index];
 469         rxq = rxq_info->rxq;
 470         SFC_ASSERT(rxq->state == SFC_RXQ_INITIALIZED);
 471
 472         evq = rxq->evq;
 473
 474         rc = sfc_ev_qstart(sa, evq->evq_index);
 475         if (rc != 0)
 476                 goto fail_ev_qstart;
 477
 478         rc = efx_rx_qcreate(sa->nic, rxq->hw_index, 0, rxq_info->type,
 479                             &rxq->mem, rxq_info->entries,
 480                             0 /* not used on EF10 */, evq->common,
 481                             &rxq->common);
 482         if (rc != 0)
 483                 goto fail_rx_qcreate;
 484
 485         efx_rx_qenable(rxq->common);
 486
 487         rxq->pending = rxq->completed = rxq->added = rxq->pushed = 0;
 488
 489         rxq->state |= SFC_RXQ_STARTED;
 490         rxq->flags |= SFC_RXQ_FLAG_STARTED | SFC_RXQ_FLAG_RUNNING;
 491
 492         sfc_rx_qrefill(rxq);
 493
 494         if (sw_index == 0) {
 495                 rc = sfc_rx_default_rxq_set_filter(sa, rxq);
 496                 if (rc != 0)
 497                         goto fail_mac_filter_default_rxq_set;
 498         }
 499
 500         /* It seems to be used by DPDK for debug purposes only ('rte_ether') */
 501         sa->eth_dev->data->rx_queue_state[sw_index] =
 502                 RTE_ETH_QUEUE_STATE_STARTED;
 503
 504         return 0;
 505
 506 fail_mac_filter_default_rxq_set:
 507         sfc_rx_qflush(sa, sw_index);
 508
 509 fail_rx_qcreate:
 510         sfc_ev_qstop(sa, evq->evq_index);
 511
 512 fail_ev_qstart:
 513         return rc;
 514 }
 515
 516 void
 517 sfc_rx_qstop(struct sfc_adapter *sa, unsigned int sw_index)
 518 {
 519         struct sfc_rxq_info *rxq_info;
 520         struct sfc_rxq *rxq;
 521
 522         sfc_log_init(sa, "sw_index=%u", sw_index);
 523
 524         SFC_ASSERT(sw_index < sa->rxq_count);
 525
 526         rxq_info = &sa->rxq_info[sw_index];
 527         rxq = rxq_info->rxq;
 528
 529         if (rxq->state == SFC_RXQ_INITIALIZED)
 530                 return;
 531         SFC_ASSERT(rxq->state & SFC_RXQ_STARTED);
 532
 533         /* It seems to be used by DPDK for debug purposes only ('rte_ether') */
 534         sa->eth_dev->data->rx_queue_state[sw_index] =
 535                 RTE_ETH_QUEUE_STATE_STOPPED;
 536
 537         rxq->flags &= ~SFC_RXQ_FLAG_RUNNING;
 538
 539         if (sw_index == 0)
 540                 efx_mac_filter_default_rxq_clear(sa->nic);
 541
 542         sfc_rx_qflush(sa, sw_index);
 543
 544         rxq->flags &= ~SFC_RXQ_FLAG_STARTED;
 545         rxq->state = SFC_RXQ_INITIALIZED;
 546
 547         efx_rx_qdestroy(rxq->common);
 548
 549         sfc_ev_qstop(sa, rxq->evq->evq_index);
 550 }
 551
 552 static int
 553 sfc_rx_qcheck_conf(struct sfc_adapter *sa, uint16_t nb_rx_desc,
 554                    const struct rte_eth_rxconf *rx_conf)
 555 {
 556         const uint16_t rx_free_thresh_max = EFX_RXQ_LIMIT(nb_rx_desc);
 557         int rc = 0;
 558
 559         if (rx_conf->rx_thresh.pthresh != 0 ||
 560             rx_conf->rx_thresh.hthresh != 0 ||
 561             rx_conf->rx_thresh.wthresh != 0) {
 562                 sfc_err(sa,
 563                         "RxQ prefetch/host/writeback thresholds are not supported");
 564                 rc = EINVAL;
 565         }
 566
 567         if (rx_conf->rx_free_thresh > rx_free_thresh_max) {
 568                 sfc_err(sa,
 569                         "RxQ free threshold too large: %u vs maximum %u",
 570                         rx_conf->rx_free_thresh, rx_free_thresh_max);
 571                 rc = EINVAL;
 572         }
 573
 574         if (rx_conf->rx_drop_en == 0) {
 575                 sfc_err(sa, "RxQ drop disable is not supported");
 576                 rc = EINVAL;
 577         }
 578
 579         return rc;
 580 }
 581
 582 static unsigned int
 583 sfc_rx_mbuf_data_alignment(struct rte_mempool *mb_pool)
 584 {
 585         uint32_t data_off;
 586         uint32_t order;
 587
 588         /* The mbuf object itself is always cache line aligned */
 589         order = rte_bsf32(RTE_CACHE_LINE_SIZE);
 590
 591         /* Data offset from mbuf object start */
 592         data_off = sizeof(struct rte_mbuf) + rte_pktmbuf_priv_size(mb_pool) +
 593                 RTE_PKTMBUF_HEADROOM;
 594
 595         order = MIN(order, rte_bsf32(data_off));
 596
 597         return 1u << (order - 1);
 598 }
 599
 600 static uint16_t
 601 sfc_rx_mb_pool_buf_size(struct sfc_adapter *sa, struct rte_mempool *mb_pool)
 602 {
 603         const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
 604         const uint32_t nic_align_start = MAX(1, encp->enc_rx_buf_align_start);
 605         const uint32_t nic_align_end = MAX(1, encp->enc_rx_buf_align_end);
 606         uint16_t buf_size;
 607         unsigned int buf_aligned;
 608         unsigned int start_alignment;
 609         unsigned int end_padding_alignment;
 610
 611         /* Below it is assumed that both alignments are power of 2 */
 612         SFC_ASSERT(rte_is_power_of_2(nic_align_start));
 613         SFC_ASSERT(rte_is_power_of_2(nic_align_end));
 614
 615         /*
 616          * mbuf is always cache line aligned, double-check
 617          * that it meets rx buffer start alignment requirements.
 618          */
 619
 620         /* Start from mbuf pool data room size */
 621         buf_size = rte_pktmbuf_data_room_size(mb_pool);
 622
 623         /* Remove headroom */
 624         if (buf_size <= RTE_PKTMBUF_HEADROOM) {
 625                 sfc_err(sa,
 626                         "RxQ mbuf pool %s object data room size %u is smaller than headroom %u",
 627                         mb_pool->name, buf_size, RTE_PKTMBUF_HEADROOM);
 628                 return 0;
 629         }
 630         buf_size -= RTE_PKTMBUF_HEADROOM;
 631
 632         /* Calculate guaranteed data start alignment */
 633         buf_aligned = sfc_rx_mbuf_data_alignment(mb_pool);
 634
 635         /* Reserve space for start alignment */
 636         if (buf_aligned < nic_align_start) {
 637                 start_alignment = nic_align_start - buf_aligned;
 638                 if (buf_size <= start_alignment) {
 639                         sfc_err(sa,
 640                                 "RxQ mbuf pool %s object data room size %u is insufficient for headroom %u and buffer start alignment %u required by NIC",
 641                                 mb_pool->name,
 642                                 rte_pktmbuf_data_room_size(mb_pool),
 643                                 RTE_PKTMBUF_HEADROOM, start_alignment);
 644                         return 0;
 645                 }
 646                 buf_aligned = nic_align_start;
 647                 buf_size -= start_alignment;
 648         } else {
 649                 start_alignment = 0;
 650         }
 651
 652         /* Make sure that end padding does not write beyond the buffer */
 653         if (buf_aligned < nic_align_end) {
 654                 /*
 655                  * Estimate space which can be lost. If guarnteed buffer
 656                  * size is odd, lost space is (nic_align_end - 1). More
 657                  * accurate formula is below.
 658                  */
 659                 end_padding_alignment = nic_align_end -
 660                         MIN(buf_aligned, 1u << (rte_bsf32(buf_size) - 1));
 661                 if (buf_size <= end_padding_alignment) {
 662                         sfc_err(sa,
 663                                 "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",
 664                                 mb_pool->name,
 665                                 rte_pktmbuf_data_room_size(mb_pool),
 666                                 RTE_PKTMBUF_HEADROOM, start_alignment,
 667                                 end_padding_alignment);
 668                         return 0;
 669                 }
 670                 buf_size -= end_padding_alignment;
 671         } else {
 672                 /*
 673                  * Start is aligned the same or better than end,
 674                  * just align length.
 675                  */
 676                 buf_size = P2ALIGN(buf_size, nic_align_end);
 677         }
 678
 679         return buf_size;
 680 }
 681
 682 int
 683 sfc_rx_qinit(struct sfc_adapter *sa, unsigned int sw_index,
 684              uint16_t nb_rx_desc, unsigned int socket_id,
 685              const struct rte_eth_rxconf *rx_conf,
 686              struct rte_mempool *mb_pool)
 687 {
 688         const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
 689         int rc;
 690         uint16_t buf_size;
 691         struct sfc_rxq_info *rxq_info;
 692         unsigned int evq_index;
 693         struct sfc_evq *evq;
 694         struct sfc_rxq *rxq;
 695
 696         rc = sfc_rx_qcheck_conf(sa, nb_rx_desc, rx_conf);
 697         if (rc != 0)
 698                 goto fail_bad_conf;
 699
 700         buf_size = sfc_rx_mb_pool_buf_size(sa, mb_pool);
 701         if (buf_size == 0) {
 702                 sfc_err(sa, "RxQ %u mbuf pool object size is too small",
 703                         sw_index);
 704                 rc = EINVAL;
 705                 goto fail_bad_conf;
 706         }
 707
 708         if ((buf_size < sa->port.pdu + encp->enc_rx_prefix_size) &&
 709             !sa->eth_dev->data->dev_conf.rxmode.enable_scatter) {
 710                 sfc_err(sa, "Rx scatter is disabled and RxQ %u mbuf pool "
 711                         "object size is too small", sw_index);
 712                 sfc_err(sa, "RxQ %u calculated Rx buffer size is %u vs "
 713                         "PDU size %u plus Rx prefix %u bytes",
 714                         sw_index, buf_size, (unsigned int)sa->port.pdu,
 715                         encp->enc_rx_prefix_size);
 716                 rc = EINVAL;
 717                 goto fail_bad_conf;
 718         }
 719
 720         SFC_ASSERT(sw_index < sa->rxq_count);
 721         rxq_info = &sa->rxq_info[sw_index];
 722
 723         SFC_ASSERT(nb_rx_desc <= rxq_info->max_entries);
 724         rxq_info->entries = nb_rx_desc;
 725         rxq_info->type =
 726                 sa->eth_dev->data->dev_conf.rxmode.enable_scatter ?
 727                 EFX_RXQ_TYPE_SCATTER : EFX_RXQ_TYPE_DEFAULT;
 728
 729         evq_index = sfc_evq_index_by_rxq_sw_index(sa, sw_index);
 730
 731         rc = sfc_ev_qinit(sa, evq_index, rxq_info->entries, socket_id);
 732         if (rc != 0)
 733                 goto fail_ev_qinit;
 734
 735         evq = sa->evq_info[evq_index].evq;
 736
 737         rc = ENOMEM;
 738         rxq = rte_zmalloc_socket("sfc-rxq", sizeof(*rxq), RTE_CACHE_LINE_SIZE,
 739                                  socket_id);
 740         if (rxq == NULL)
 741                 goto fail_rxq_alloc;
 742
 743         rc = sfc_dma_alloc(sa, "rxq", sw_index, EFX_RXQ_SIZE(rxq_info->entries),
 744                            socket_id, &rxq->mem);
 745         if (rc != 0)
 746                 goto fail_dma_alloc;
 747
 748         rc = ENOMEM;
 749         rxq->sw_desc = rte_calloc_socket("sfc-rxq-sw_desc", rxq_info->entries,
 750                                          sizeof(*rxq->sw_desc),
 751                                          RTE_CACHE_LINE_SIZE, socket_id);
 752         if (rxq->sw_desc == NULL)
 753                 goto fail_desc_alloc;
 754
 755         evq->rxq = rxq;
 756         rxq->evq = evq;
 757         rxq->ptr_mask = rxq_info->entries - 1;
 758         rxq->refill_threshold = rx_conf->rx_free_thresh;
 759         rxq->refill_mb_pool = mb_pool;
 760         rxq->buf_size = buf_size;
 761         rxq->hw_index = sw_index;
 762         rxq->port_id = sa->eth_dev->data->port_id;
 763
 764         /* Cache limits required on datapath in RxQ structure */
 765         rxq->batch_max = encp->enc_rx_batch_max;
 766         rxq->prefix_size = encp->enc_rx_prefix_size;
 767
 768 #if EFSYS_OPT_RX_SCALE
 769         if (sa->hash_support == EFX_RX_HASH_AVAILABLE)
 770                 rxq->flags |= SFC_RXQ_FLAG_RSS_HASH;
 771 #endif
 772
 773         rxq->state = SFC_RXQ_INITIALIZED;
 774
 775         rxq_info->rxq = rxq;
 776         rxq_info->deferred_start = (rx_conf->rx_deferred_start != 0);
 777
 778         return 0;
 779
 780 fail_desc_alloc:
 781         sfc_dma_free(sa, &rxq->mem);
 782
 783 fail_dma_alloc:
 784         rte_free(rxq);
 785
 786 fail_rxq_alloc:
 787         sfc_ev_qfini(sa, evq_index);
 788
 789 fail_ev_qinit:
 790         rxq_info->entries = 0;
 791
 792 fail_bad_conf:
 793         sfc_log_init(sa, "failed %d", rc);
 794         return rc;
 795 }
 796
 797 void
 798 sfc_rx_qfini(struct sfc_adapter *sa, unsigned int sw_index)
 799 {
 800         struct sfc_rxq_info *rxq_info;
 801         struct sfc_rxq *rxq;
 802
 803         SFC_ASSERT(sw_index < sa->rxq_count);
 804
 805         rxq_info = &sa->rxq_info[sw_index];
 806
 807         rxq = rxq_info->rxq;
 808         SFC_ASSERT(rxq->state == SFC_RXQ_INITIALIZED);
 809
 810         rxq_info->rxq = NULL;
 811         rxq_info->entries = 0;
 812
 813         rte_free(rxq->sw_desc);
 814         sfc_dma_free(sa, &rxq->mem);
 815         rte_free(rxq);
 816 }
 817
 818 #if EFSYS_OPT_RX_SCALE
 819 efx_rx_hash_type_t
 820 sfc_rte_to_efx_hash_type(uint64_t rss_hf)
 821 {
 822         efx_rx_hash_type_t efx_hash_types = 0;
 823
 824         if ((rss_hf & (ETH_RSS_IPV4 | ETH_RSS_FRAG_IPV4 |
 825                        ETH_RSS_NONFRAG_IPV4_OTHER)) != 0)
 826                 efx_hash_types |= EFX_RX_HASH_IPV4;
 827
 828         if ((rss_hf & ETH_RSS_NONFRAG_IPV4_TCP) != 0)
 829                 efx_hash_types |= EFX_RX_HASH_TCPIPV4;
 830
 831         if ((rss_hf & (ETH_RSS_IPV6 | ETH_RSS_FRAG_IPV6 |
 832                         ETH_RSS_NONFRAG_IPV6_OTHER | ETH_RSS_IPV6_EX)) != 0)
 833                 efx_hash_types |= EFX_RX_HASH_IPV6;
 834
 835         if ((rss_hf & (ETH_RSS_NONFRAG_IPV6_TCP | ETH_RSS_IPV6_TCP_EX)) != 0)
 836                 efx_hash_types |= EFX_RX_HASH_TCPIPV6;
 837
 838         return efx_hash_types;
 839 }
 840
 841 uint64_t
 842 sfc_efx_to_rte_hash_type(efx_rx_hash_type_t efx_hash_types)
 843 {
 844         uint64_t rss_hf = 0;
 845
 846         if ((efx_hash_types & EFX_RX_HASH_IPV4) != 0)
 847                 rss_hf |= (ETH_RSS_IPV4 | ETH_RSS_FRAG_IPV4 |
 848                            ETH_RSS_NONFRAG_IPV4_OTHER);
 849
 850         if ((efx_hash_types & EFX_RX_HASH_TCPIPV4) != 0)
 851                 rss_hf |= ETH_RSS_NONFRAG_IPV4_TCP;
 852
 853         if ((efx_hash_types & EFX_RX_HASH_IPV6) != 0)
 854                 rss_hf |= (ETH_RSS_IPV6 | ETH_RSS_FRAG_IPV6 |
 855                            ETH_RSS_NONFRAG_IPV6_OTHER | ETH_RSS_IPV6_EX);
 856
 857         if ((efx_hash_types & EFX_RX_HASH_TCPIPV6) != 0)
 858                 rss_hf |= (ETH_RSS_NONFRAG_IPV6_TCP | ETH_RSS_IPV6_TCP_EX);
 859
 860         return rss_hf;
 861 }
 862 #endif
 863
 864 static int
 865 sfc_rx_rss_config(struct sfc_adapter *sa)
 866 {
 867         int rc = 0;
 868
 869 #if EFSYS_OPT_RX_SCALE
 870         if (sa->rss_channels > 1) {
 871                 rc = efx_rx_scale_mode_set(sa->nic, EFX_RX_HASHALG_TOEPLITZ,
 872                                            sa->rss_hash_types, B_TRUE);
 873                 if (rc != 0)
 874                         goto finish;
 875
 876                 rc = efx_rx_scale_key_set(sa->nic, sa->rss_key,
 877                                           sizeof(sa->rss_key));
 878                 if (rc != 0)
 879                         goto finish;
 880
 881                 rc = efx_rx_scale_tbl_set(sa->nic, sa->rss_tbl,
 882                                           sizeof(sa->rss_tbl));
 883         }
 884
 885 finish:
 886 #endif
 887         return rc;
 888 }
 889
 890 int
 891 sfc_rx_start(struct sfc_adapter *sa)
 892 {
 893         unsigned int sw_index;
 894         int rc;
 895
 896         sfc_log_init(sa, "rxq_count=%u", sa->rxq_count);
 897
 898         rc = efx_rx_init(sa->nic);
 899         if (rc != 0)
 900                 goto fail_rx_init;
 901
 902         rc = sfc_rx_rss_config(sa);
 903         if (rc != 0)
 904                 goto fail_rss_config;
 905
 906         for (sw_index = 0; sw_index < sa->rxq_count; ++sw_index) {
 907                 if ((!sa->rxq_info[sw_index].deferred_start ||
 908                      sa->rxq_info[sw_index].deferred_started)) {
 909                         rc = sfc_rx_qstart(sa, sw_index);
 910                         if (rc != 0)
 911                                 goto fail_rx_qstart;
 912                 }
 913         }
 914
 915         return 0;
 916
 917 fail_rx_qstart:
 918         while (sw_index-- > 0)
 919                 sfc_rx_qstop(sa, sw_index);
 920
 921 fail_rss_config:
 922         efx_rx_fini(sa->nic);
 923
 924 fail_rx_init:
 925         sfc_log_init(sa, "failed %d", rc);
 926         return rc;
 927 }
 928
 929 void
 930 sfc_rx_stop(struct sfc_adapter *sa)
 931 {
 932         unsigned int sw_index;
 933
 934         sfc_log_init(sa, "rxq_count=%u", sa->rxq_count);
 935
 936         sw_index = sa->rxq_count;
 937         while (sw_index-- > 0) {
 938                 if (sa->rxq_info[sw_index].rxq != NULL)
 939                         sfc_rx_qstop(sa, sw_index);
 940         }
 941
 942         efx_rx_fini(sa->nic);
 943 }
 944
 945 static int
 946 sfc_rx_qinit_info(struct sfc_adapter *sa, unsigned int sw_index)
 947 {
 948         struct sfc_rxq_info *rxq_info = &sa->rxq_info[sw_index];
 949         unsigned int max_entries;
 950
 951         max_entries = EFX_RXQ_MAXNDESCS;
 952         SFC_ASSERT(rte_is_power_of_2(max_entries));
 953
 954         rxq_info->max_entries = max_entries;
 955
 956         return 0;
 957 }
 958
 959 static int
 960 sfc_rx_check_mode(struct sfc_adapter *sa, struct rte_eth_rxmode *rxmode)
 961 {
 962         int rc = 0;
 963
 964         switch (rxmode->mq_mode) {
 965         case ETH_MQ_RX_NONE:
 966                 /* No special checks are required */
 967                 break;
 968 #if EFSYS_OPT_RX_SCALE
 969         case ETH_MQ_RX_RSS:
 970                 if (sa->rss_support == EFX_RX_SCALE_UNAVAILABLE) {
 971                         sfc_err(sa, "RSS is not available");
 972                         rc = EINVAL;
 973                 }
 974                 break;
 975 #endif
 976         default:
 977                 sfc_err(sa, "Rx multi-queue mode %u not supported",
 978                         rxmode->mq_mode);
 979                 rc = EINVAL;
 980         }
 981
 982         if (rxmode->header_split) {
 983                 sfc_err(sa, "Header split on Rx not supported");
 984                 rc = EINVAL;
 985         }
 986
 987         if (rxmode->hw_vlan_filter) {
 988                 sfc_err(sa, "HW VLAN filtering not supported");
 989                 rc = EINVAL;
 990         }
 991
 992         if (rxmode->hw_vlan_strip) {
 993                 sfc_err(sa, "HW VLAN stripping not supported");
 994                 rc = EINVAL;
 995         }
 996
 997         if (rxmode->hw_vlan_extend) {
 998                 sfc_err(sa,
 999                         "Q-in-Q HW VLAN stripping not supported");
1000                 rc = EINVAL;
1001         }
1002
1003         if (!rxmode->hw_strip_crc) {
1004                 sfc_warn(sa,
1005                          "FCS stripping control not supported - always stripped");
1006                 rxmode->hw_strip_crc = 1;
1007         }
1008
1009         if (rxmode->enable_lro) {
1010                 sfc_err(sa, "LRO not supported");
1011                 rc = EINVAL;
1012         }
1013
1014         return rc;
1015 }
1016
1017 /**
1018  * Initialize Rx subsystem.
1019  *
1020  * Called at device configuration stage when number of receive queues is
1021  * specified together with other device level receive configuration.
1022  *
1023  * It should be used to allocate NUMA-unaware resources.
1024  */
1025 int
1026 sfc_rx_init(struct sfc_adapter *sa)
1027 {
1028         struct rte_eth_conf *dev_conf = &sa->eth_dev->data->dev_conf;
1029         unsigned int sw_index;
1030         int rc;
1031
1032         rc = sfc_rx_check_mode(sa, &dev_conf->rxmode);
1033         if (rc != 0)
1034                 goto fail_check_mode;
1035
1036         sa->rxq_count = sa->eth_dev->data->nb_rx_queues;
1037
1038         rc = ENOMEM;
1039         sa->rxq_info = rte_calloc_socket("sfc-rxqs", sa->rxq_count,
1040                                          sizeof(struct sfc_rxq_info), 0,
1041                                          sa->socket_id);
1042         if (sa->rxq_info == NULL)
1043                 goto fail_rxqs_alloc;
1044
1045         for (sw_index = 0; sw_index < sa->rxq_count; ++sw_index) {
1046                 rc = sfc_rx_qinit_info(sa, sw_index);
1047                 if (rc != 0)
1048                         goto fail_rx_qinit_info;
1049         }
1050
1051 #if EFSYS_OPT_RX_SCALE
1052         sa->rss_channels = (dev_conf->rxmode.mq_mode == ETH_MQ_RX_RSS) ?
1053                            MIN(sa->rxq_count, EFX_MAXRSS) : 1;
1054
1055         if (sa->rss_channels > 1) {
1056                 for (sw_index = 0; sw_index < EFX_RSS_TBL_SIZE; ++sw_index)
1057                         sa->rss_tbl[sw_index] = sw_index % sa->rss_channels;
1058         }
1059 #endif
1060
1061         return 0;
1062
1063 fail_rx_qinit_info:
1064         rte_free(sa->rxq_info);
1065         sa->rxq_info = NULL;
1066
1067 fail_rxqs_alloc:
1068         sa->rxq_count = 0;
1069 fail_check_mode:
1070         sfc_log_init(sa, "failed %d", rc);
1071         return rc;
1072 }
1073
1074 /**
1075  * Shutdown Rx subsystem.
1076  *
1077  * Called at device close stage, for example, before device
1078  * reconfiguration or shutdown.
1079  */
1080 void
1081 sfc_rx_fini(struct sfc_adapter *sa)
1082 {
1083         unsigned int sw_index;
1084
1085         sw_index = sa->rxq_count;
1086         while (sw_index-- > 0) {
1087                 if (sa->rxq_info[sw_index].rxq != NULL)
1088                         sfc_rx_qfini(sa, sw_index);
1089         }
1090
1091         rte_free(sa->rxq_info);
1092         sa->rxq_info = NULL;
1093         sa->rxq_count = 0;
1094 }