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 static void
 189 sfc_rx_set_rss_hash(struct sfc_rxq *rxq, unsigned int flags, struct rte_mbuf *m)
 190 {
 191 #if EFSYS_OPT_RX_SCALE
 192         uint8_t *mbuf_data;
 193
 194
 195         if ((rxq->flags & SFC_RXQ_RSS_HASH) == 0)
 196                 return;
 197
 198         mbuf_data = rte_pktmbuf_mtod(m, uint8_t *);
 199
 200         if (flags & (EFX_PKT_IPV4 | EFX_PKT_IPV6)) {
 201                 m->hash.rss = efx_pseudo_hdr_hash_get(rxq->common,
 202                                                       EFX_RX_HASHALG_TOEPLITZ,
 203                                                       mbuf_data);
 204
 205                 m->ol_flags |= PKT_RX_RSS_HASH;
 206         }
 207 #endif
 208 }
 209
 210 uint16_t
 211 sfc_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 212 {
 213         struct sfc_rxq *rxq = rx_queue;
 214         unsigned int completed;
 215         unsigned int prefix_size = rxq->prefix_size;
 216         unsigned int done_pkts = 0;
 217         boolean_t discard_next = B_FALSE;
 218         struct rte_mbuf *scatter_pkt = NULL;
 219
 220         if (unlikely((rxq->state & SFC_RXQ_RUNNING) == 0))
 221                 return 0;
 222
 223         sfc_ev_qpoll(rxq->evq);
 224
 225         completed = rxq->completed;
 226         while (completed != rxq->pending && done_pkts < nb_pkts) {
 227                 unsigned int id;
 228                 struct sfc_rx_sw_desc *rxd;
 229                 struct rte_mbuf *m;
 230                 unsigned int seg_len;
 231                 unsigned int desc_flags;
 232
 233                 id = completed++ & rxq->ptr_mask;
 234                 rxd = &rxq->sw_desc[id];
 235                 m = rxd->mbuf;
 236                 desc_flags = rxd->flags;
 237
 238                 if (discard_next)
 239                         goto discard;
 240
 241                 if (desc_flags & (EFX_ADDR_MISMATCH | EFX_DISCARD))
 242                         goto discard;
 243
 244                 if (desc_flags & EFX_PKT_PREFIX_LEN) {
 245                         uint16_t tmp_size;
 246                         int rc __rte_unused;
 247
 248                         rc = efx_pseudo_hdr_pkt_length_get(rxq->common,
 249                                 rte_pktmbuf_mtod(m, uint8_t *), &tmp_size);
 250                         SFC_ASSERT(rc == 0);
 251                         seg_len = tmp_size;
 252                 } else {
 253                         seg_len = rxd->size - prefix_size;
 254                 }
 255
 256                 rte_pktmbuf_data_len(m) = seg_len;
 257                 rte_pktmbuf_pkt_len(m) = seg_len;
 258
 259                 if (scatter_pkt != NULL) {
 260                         if (rte_pktmbuf_chain(scatter_pkt, m) != 0) {
 261                                 rte_mempool_put(rxq->refill_mb_pool,
 262                                                 scatter_pkt);
 263                                 goto discard;
 264                         }
 265                         /* The packet to deliver */
 266                         m = scatter_pkt;
 267                 }
 268
 269                 if (desc_flags & EFX_PKT_CONT) {
 270                         /* The packet is scattered, more fragments to come */
 271                         scatter_pkt = m;
 272                         /* Futher fragments have no prefix */
 273                         prefix_size = 0;
 274                         continue;
 275                 }
 276
 277                 /* Scattered packet is done */
 278                 scatter_pkt = NULL;
 279                 /* The first fragment of the packet has prefix */
 280                 prefix_size = rxq->prefix_size;
 281
 282                 m->ol_flags = sfc_rx_desc_flags_to_offload_flags(desc_flags);
 283                 m->packet_type = sfc_rx_desc_flags_to_packet_type(desc_flags);
 284
 285                 /*
 286                  * Extract RSS hash from the packet prefix and
 287                  * set the corresponding field (if needed and possible)
 288                  */
 289                 sfc_rx_set_rss_hash(rxq, desc_flags, m);
 290
 291                 m->data_off += prefix_size;
 292
 293                 *rx_pkts++ = m;
 294                 done_pkts++;
 295                 continue;
 296
 297 discard:
 298                 discard_next = ((desc_flags & EFX_PKT_CONT) != 0);
 299                 rte_mempool_put(rxq->refill_mb_pool, m);
 300                 rxd->mbuf = NULL;
 301         }
 302
 303         /* pending is only moved when entire packet is received */
 304         SFC_ASSERT(scatter_pkt == NULL);
 305
 306         rxq->completed = completed;
 307
 308         sfc_rx_qrefill(rxq);
 309
 310         return done_pkts;
 311 }
 312
 313 unsigned int
 314 sfc_rx_qdesc_npending(struct sfc_adapter *sa, unsigned int sw_index)
 315 {
 316         struct sfc_rxq *rxq;
 317
 318         SFC_ASSERT(sw_index < sa->rxq_count);
 319         rxq = sa->rxq_info[sw_index].rxq;
 320
 321         if (rxq == NULL || (rxq->state & SFC_RXQ_RUNNING) == 0)
 322                 return 0;
 323
 324         sfc_ev_qpoll(rxq->evq);
 325
 326         return rxq->pending - rxq->completed;
 327 }
 328
 329 int
 330 sfc_rx_qdesc_done(struct sfc_rxq *rxq, unsigned int offset)
 331 {
 332         if ((rxq->state & SFC_RXQ_RUNNING) == 0)
 333                 return 0;
 334
 335         sfc_ev_qpoll(rxq->evq);
 336
 337         return offset < (rxq->pending - rxq->completed);
 338 }
 339
 340 static void
 341 sfc_rx_qpurge(struct sfc_rxq *rxq)
 342 {
 343         unsigned int i;
 344         struct sfc_rx_sw_desc *rxd;
 345
 346         for (i = rxq->completed; i != rxq->added; ++i) {
 347                 rxd = &rxq->sw_desc[i & rxq->ptr_mask];
 348                 rte_mempool_put(rxq->refill_mb_pool, rxd->mbuf);
 349                 rxd->mbuf = NULL;
 350         }
 351 }
 352
 353 static void
 354 sfc_rx_qflush(struct sfc_adapter *sa, unsigned int sw_index)
 355 {
 356         struct sfc_rxq *rxq;
 357         unsigned int retry_count;
 358         unsigned int wait_count;
 359
 360         rxq = sa->rxq_info[sw_index].rxq;
 361         SFC_ASSERT(rxq->state & SFC_RXQ_STARTED);
 362
 363         /*
 364          * Retry Rx queue flushing in the case of flush failed or
 365          * timeout. In the worst case it can delay for 6 seconds.
 366          */
 367         for (retry_count = 0;
 368              ((rxq->state & SFC_RXQ_FLUSHED) == 0) &&
 369              (retry_count < SFC_RX_QFLUSH_ATTEMPTS);
 370              ++retry_count) {
 371                 if (efx_rx_qflush(rxq->common) != 0) {
 372                         rxq->state |= SFC_RXQ_FLUSH_FAILED;
 373                         break;
 374                 }
 375                 rxq->state &= ~SFC_RXQ_FLUSH_FAILED;
 376                 rxq->state |= SFC_RXQ_FLUSHING;
 377
 378                 /*
 379                  * Wait for Rx queue flush done or failed event at least
 380                  * SFC_RX_QFLUSH_POLL_WAIT_MS milliseconds and not more
 381                  * than 2 seconds (SFC_RX_QFLUSH_POLL_WAIT_MS multiplied
 382                  * by SFC_RX_QFLUSH_POLL_ATTEMPTS).
 383                  */
 384                 wait_count = 0;
 385                 do {
 386                         rte_delay_ms(SFC_RX_QFLUSH_POLL_WAIT_MS);
 387                         sfc_ev_qpoll(rxq->evq);
 388                 } while ((rxq->state & SFC_RXQ_FLUSHING) &&
 389                          (wait_count++ < SFC_RX_QFLUSH_POLL_ATTEMPTS));
 390
 391                 if (rxq->state & SFC_RXQ_FLUSHING)
 392                         sfc_err(sa, "RxQ %u flush timed out", sw_index);
 393
 394                 if (rxq->state & SFC_RXQ_FLUSH_FAILED)
 395                         sfc_err(sa, "RxQ %u flush failed", sw_index);
 396
 397                 if (rxq->state & SFC_RXQ_FLUSHED)
 398                         sfc_info(sa, "RxQ %u flushed", sw_index);
 399         }
 400
 401         sfc_rx_qpurge(rxq);
 402 }
 403
 404 static int
 405 sfc_rx_default_rxq_set_filter(struct sfc_adapter *sa, struct sfc_rxq *rxq)
 406 {
 407         boolean_t rss = (sa->rss_channels > 1) ? B_TRUE : B_FALSE;
 408         struct sfc_port *port = &sa->port;
 409         int rc;
 410
 411         /*
 412          * If promiscuous or all-multicast mode has been requested, setting
 413          * filter for the default Rx queue might fail, in particular, while
 414          * running over PCI function which is not a member of corresponding
 415          * privilege groups; if this occurs, few iterations will be made to
 416          * repeat this step without promiscuous and all-multicast flags set
 417          */
 418 retry:
 419         rc = efx_mac_filter_default_rxq_set(sa->nic, rxq->common, rss);
 420         if (rc == 0)
 421                 return 0;
 422         else if (rc != EOPNOTSUPP)
 423                 return rc;
 424
 425         if (port->promisc) {
 426                 sfc_warn(sa, "promiscuous mode has been requested, "
 427                              "but the HW rejects it");
 428                 sfc_warn(sa, "promiscuous mode will be disabled");
 429
 430                 port->promisc = B_FALSE;
 431                 rc = sfc_set_rx_mode(sa);
 432                 if (rc != 0)
 433                         return rc;
 434
 435                 goto retry;
 436         }
 437
 438         if (port->allmulti) {
 439                 sfc_warn(sa, "all-multicast mode has been requested, "
 440                              "but the HW rejects it");
 441                 sfc_warn(sa, "all-multicast mode will be disabled");
 442
 443                 port->allmulti = B_FALSE;
 444                 rc = sfc_set_rx_mode(sa);
 445                 if (rc != 0)
 446                         return rc;
 447
 448                 goto retry;
 449         }
 450
 451         return rc;
 452 }
 453
 454 int
 455 sfc_rx_qstart(struct sfc_adapter *sa, unsigned int sw_index)
 456 {
 457         struct sfc_rxq_info *rxq_info;
 458         struct sfc_rxq *rxq;
 459         struct sfc_evq *evq;
 460         int rc;
 461
 462         sfc_log_init(sa, "sw_index=%u", sw_index);
 463
 464         SFC_ASSERT(sw_index < sa->rxq_count);
 465
 466         rxq_info = &sa->rxq_info[sw_index];
 467         rxq = rxq_info->rxq;
 468         SFC_ASSERT(rxq->state == SFC_RXQ_INITIALIZED);
 469
 470         evq = rxq->evq;
 471
 472         rc = sfc_ev_qstart(sa, evq->evq_index);
 473         if (rc != 0)
 474                 goto fail_ev_qstart;
 475
 476         rc = efx_rx_qcreate(sa->nic, rxq->hw_index, 0, rxq_info->type,
 477                             &rxq->mem, rxq_info->entries,
 478                             0 /* not used on EF10 */, evq->common,
 479                             &rxq->common);
 480         if (rc != 0)
 481                 goto fail_rx_qcreate;
 482
 483         efx_rx_qenable(rxq->common);
 484
 485         rxq->pending = rxq->completed = rxq->added = rxq->pushed = 0;
 486
 487         rxq->state |= (SFC_RXQ_STARTED | SFC_RXQ_RUNNING);
 488
 489         sfc_rx_qrefill(rxq);
 490
 491         if (sw_index == 0) {
 492                 rc = sfc_rx_default_rxq_set_filter(sa, rxq);
 493                 if (rc != 0)
 494                         goto fail_mac_filter_default_rxq_set;
 495         }
 496
 497         /* It seems to be used by DPDK for debug purposes only ('rte_ether') */
 498         sa->eth_dev->data->rx_queue_state[sw_index] =
 499                 RTE_ETH_QUEUE_STATE_STARTED;
 500
 501         return 0;
 502
 503 fail_mac_filter_default_rxq_set:
 504         sfc_rx_qflush(sa, sw_index);
 505
 506 fail_rx_qcreate:
 507         sfc_ev_qstop(sa, evq->evq_index);
 508
 509 fail_ev_qstart:
 510         return rc;
 511 }
 512
 513 void
 514 sfc_rx_qstop(struct sfc_adapter *sa, unsigned int sw_index)
 515 {
 516         struct sfc_rxq_info *rxq_info;
 517         struct sfc_rxq *rxq;
 518
 519         sfc_log_init(sa, "sw_index=%u", sw_index);
 520
 521         SFC_ASSERT(sw_index < sa->rxq_count);
 522
 523         rxq_info = &sa->rxq_info[sw_index];
 524         rxq = rxq_info->rxq;
 525
 526         if (rxq->state == SFC_RXQ_INITIALIZED)
 527                 return;
 528         SFC_ASSERT(rxq->state & SFC_RXQ_STARTED);
 529
 530         /* It seems to be used by DPDK for debug purposes only ('rte_ether') */
 531         sa->eth_dev->data->rx_queue_state[sw_index] =
 532                 RTE_ETH_QUEUE_STATE_STOPPED;
 533
 534         rxq->state &= ~SFC_RXQ_RUNNING;
 535
 536         if (sw_index == 0)
 537                 efx_mac_filter_default_rxq_clear(sa->nic);
 538
 539         sfc_rx_qflush(sa, sw_index);
 540
 541         rxq->state = SFC_RXQ_INITIALIZED;
 542
 543         efx_rx_qdestroy(rxq->common);
 544
 545         sfc_ev_qstop(sa, rxq->evq->evq_index);
 546 }
 547
 548 static int
 549 sfc_rx_qcheck_conf(struct sfc_adapter *sa, uint16_t nb_rx_desc,
 550                    const struct rte_eth_rxconf *rx_conf)
 551 {
 552         const uint16_t rx_free_thresh_max = EFX_RXQ_LIMIT(nb_rx_desc);
 553         int rc = 0;
 554
 555         if (rx_conf->rx_thresh.pthresh != 0 ||
 556             rx_conf->rx_thresh.hthresh != 0 ||
 557             rx_conf->rx_thresh.wthresh != 0) {
 558                 sfc_err(sa,
 559                         "RxQ prefetch/host/writeback thresholds are not supported");
 560                 rc = EINVAL;
 561         }
 562
 563         if (rx_conf->rx_free_thresh > rx_free_thresh_max) {
 564                 sfc_err(sa,
 565                         "RxQ free threshold too large: %u vs maximum %u",
 566                         rx_conf->rx_free_thresh, rx_free_thresh_max);
 567                 rc = EINVAL;
 568         }
 569
 570         if (rx_conf->rx_drop_en == 0) {
 571                 sfc_err(sa, "RxQ drop disable is not supported");
 572                 rc = EINVAL;
 573         }
 574
 575         return rc;
 576 }
 577
 578 static unsigned int
 579 sfc_rx_mbuf_data_alignment(struct rte_mempool *mb_pool)
 580 {
 581         uint32_t data_off;
 582         uint32_t order;
 583
 584         /* The mbuf object itself is always cache line aligned */
 585         order = rte_bsf32(RTE_CACHE_LINE_SIZE);
 586
 587         /* Data offset from mbuf object start */
 588         data_off = sizeof(struct rte_mbuf) + rte_pktmbuf_priv_size(mb_pool) +
 589                 RTE_PKTMBUF_HEADROOM;
 590
 591         order = MIN(order, rte_bsf32(data_off));
 592
 593         return 1u << (order - 1);
 594 }
 595
 596 static uint16_t
 597 sfc_rx_mb_pool_buf_size(struct sfc_adapter *sa, struct rte_mempool *mb_pool)
 598 {
 599         const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
 600         const uint32_t nic_align_start = MAX(1, encp->enc_rx_buf_align_start);
 601         const uint32_t nic_align_end = MAX(1, encp->enc_rx_buf_align_end);
 602         uint16_t buf_size;
 603         unsigned int buf_aligned;
 604         unsigned int start_alignment;
 605         unsigned int end_padding_alignment;
 606
 607         /* Below it is assumed that both alignments are power of 2 */
 608         SFC_ASSERT(rte_is_power_of_2(nic_align_start));
 609         SFC_ASSERT(rte_is_power_of_2(nic_align_end));
 610
 611         /*
 612          * mbuf is always cache line aligned, double-check
 613          * that it meets rx buffer start alignment requirements.
 614          */
 615
 616         /* Start from mbuf pool data room size */
 617         buf_size = rte_pktmbuf_data_room_size(mb_pool);
 618
 619         /* Remove headroom */
 620         if (buf_size <= RTE_PKTMBUF_HEADROOM) {
 621                 sfc_err(sa,
 622                         "RxQ mbuf pool %s object data room size %u is smaller than headroom %u",
 623                         mb_pool->name, buf_size, RTE_PKTMBUF_HEADROOM);
 624                 return 0;
 625         }
 626         buf_size -= RTE_PKTMBUF_HEADROOM;
 627
 628         /* Calculate guaranteed data start alignment */
 629         buf_aligned = sfc_rx_mbuf_data_alignment(mb_pool);
 630
 631         /* Reserve space for start alignment */
 632         if (buf_aligned < nic_align_start) {
 633                 start_alignment = nic_align_start - buf_aligned;
 634                 if (buf_size <= start_alignment) {
 635                         sfc_err(sa,
 636                                 "RxQ mbuf pool %s object data room size %u is insufficient for headroom %u and buffer start alignment %u required by NIC",
 637                                 mb_pool->name,
 638                                 rte_pktmbuf_data_room_size(mb_pool),
 639                                 RTE_PKTMBUF_HEADROOM, start_alignment);
 640                         return 0;
 641                 }
 642                 buf_aligned = nic_align_start;
 643                 buf_size -= start_alignment;
 644         } else {
 645                 start_alignment = 0;
 646         }
 647
 648         /* Make sure that end padding does not write beyond the buffer */
 649         if (buf_aligned < nic_align_end) {
 650                 /*
 651                  * Estimate space which can be lost. If guarnteed buffer
 652                  * size is odd, lost space is (nic_align_end - 1). More
 653                  * accurate formula is below.
 654                  */
 655                 end_padding_alignment = nic_align_end -
 656                         MIN(buf_aligned, 1u << (rte_bsf32(buf_size) - 1));
 657                 if (buf_size <= end_padding_alignment) {
 658                         sfc_err(sa,
 659                                 "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",
 660                                 mb_pool->name,
 661                                 rte_pktmbuf_data_room_size(mb_pool),
 662                                 RTE_PKTMBUF_HEADROOM, start_alignment,
 663                                 end_padding_alignment);
 664                         return 0;
 665                 }
 666                 buf_size -= end_padding_alignment;
 667         } else {
 668                 /*
 669                  * Start is aligned the same or better than end,
 670                  * just align length.
 671                  */
 672                 buf_size = P2ALIGN(buf_size, nic_align_end);
 673         }
 674
 675         return buf_size;
 676 }
 677
 678 int
 679 sfc_rx_qinit(struct sfc_adapter *sa, unsigned int sw_index,
 680              uint16_t nb_rx_desc, unsigned int socket_id,
 681              const struct rte_eth_rxconf *rx_conf,
 682              struct rte_mempool *mb_pool)
 683 {
 684         const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
 685         int rc;
 686         uint16_t buf_size;
 687         struct sfc_rxq_info *rxq_info;
 688         unsigned int evq_index;
 689         struct sfc_evq *evq;
 690         struct sfc_rxq *rxq;
 691
 692         rc = sfc_rx_qcheck_conf(sa, nb_rx_desc, rx_conf);
 693         if (rc != 0)
 694                 goto fail_bad_conf;
 695
 696         buf_size = sfc_rx_mb_pool_buf_size(sa, mb_pool);
 697         if (buf_size == 0) {
 698                 sfc_err(sa, "RxQ %u mbuf pool object size is too small",
 699                         sw_index);
 700                 rc = EINVAL;
 701                 goto fail_bad_conf;
 702         }
 703
 704         if ((buf_size < sa->port.pdu + encp->enc_rx_prefix_size) &&
 705             !sa->eth_dev->data->dev_conf.rxmode.enable_scatter) {
 706                 sfc_err(sa, "Rx scatter is disabled and RxQ %u mbuf pool "
 707                         "object size is too small", sw_index);
 708                 sfc_err(sa, "RxQ %u calculated Rx buffer size is %u vs "
 709                         "PDU size %u plus Rx prefix %u bytes",
 710                         sw_index, buf_size, (unsigned int)sa->port.pdu,
 711                         encp->enc_rx_prefix_size);
 712                 rc = EINVAL;
 713                 goto fail_bad_conf;
 714         }
 715
 716         SFC_ASSERT(sw_index < sa->rxq_count);
 717         rxq_info = &sa->rxq_info[sw_index];
 718
 719         SFC_ASSERT(nb_rx_desc <= rxq_info->max_entries);
 720         rxq_info->entries = nb_rx_desc;
 721         rxq_info->type =
 722                 sa->eth_dev->data->dev_conf.rxmode.enable_scatter ?
 723                 EFX_RXQ_TYPE_SCATTER : EFX_RXQ_TYPE_DEFAULT;
 724
 725         evq_index = sfc_evq_index_by_rxq_sw_index(sa, sw_index);
 726
 727         rc = sfc_ev_qinit(sa, evq_index, rxq_info->entries, socket_id);
 728         if (rc != 0)
 729                 goto fail_ev_qinit;
 730
 731         evq = sa->evq_info[evq_index].evq;
 732
 733         rc = ENOMEM;
 734         rxq = rte_zmalloc_socket("sfc-rxq", sizeof(*rxq), RTE_CACHE_LINE_SIZE,
 735                                  socket_id);
 736         if (rxq == NULL)
 737                 goto fail_rxq_alloc;
 738
 739         rc = sfc_dma_alloc(sa, "rxq", sw_index, EFX_RXQ_SIZE(rxq_info->entries),
 740                            socket_id, &rxq->mem);
 741         if (rc != 0)
 742                 goto fail_dma_alloc;
 743
 744         rc = ENOMEM;
 745         rxq->sw_desc = rte_calloc_socket("sfc-rxq-sw_desc", rxq_info->entries,
 746                                          sizeof(*rxq->sw_desc),
 747                                          RTE_CACHE_LINE_SIZE, socket_id);
 748         if (rxq->sw_desc == NULL)
 749                 goto fail_desc_alloc;
 750
 751         evq->rxq = rxq;
 752         rxq->evq = evq;
 753         rxq->ptr_mask = rxq_info->entries - 1;
 754         rxq->refill_threshold = rx_conf->rx_free_thresh;
 755         rxq->refill_mb_pool = mb_pool;
 756         rxq->buf_size = buf_size;
 757         rxq->hw_index = sw_index;
 758         rxq->port_id = sa->eth_dev->data->port_id;
 759
 760         /* Cache limits required on datapath in RxQ structure */
 761         rxq->batch_max = encp->enc_rx_batch_max;
 762         rxq->prefix_size = encp->enc_rx_prefix_size;
 763
 764 #if EFSYS_OPT_RX_SCALE
 765         if (sa->hash_support == EFX_RX_HASH_AVAILABLE)
 766                 rxq->flags |= SFC_RXQ_RSS_HASH;
 767 #endif
 768
 769         rxq->state = SFC_RXQ_INITIALIZED;
 770
 771         rxq_info->rxq = rxq;
 772         rxq_info->deferred_start = (rx_conf->rx_deferred_start != 0);
 773
 774         return 0;
 775
 776 fail_desc_alloc:
 777         sfc_dma_free(sa, &rxq->mem);
 778
 779 fail_dma_alloc:
 780         rte_free(rxq);
 781
 782 fail_rxq_alloc:
 783         sfc_ev_qfini(sa, evq_index);
 784
 785 fail_ev_qinit:
 786         rxq_info->entries = 0;
 787
 788 fail_bad_conf:
 789         sfc_log_init(sa, "failed %d", rc);
 790         return rc;
 791 }
 792
 793 void
 794 sfc_rx_qfini(struct sfc_adapter *sa, unsigned int sw_index)
 795 {
 796         struct sfc_rxq_info *rxq_info;
 797         struct sfc_rxq *rxq;
 798
 799         SFC_ASSERT(sw_index < sa->rxq_count);
 800
 801         rxq_info = &sa->rxq_info[sw_index];
 802
 803         rxq = rxq_info->rxq;
 804         SFC_ASSERT(rxq->state == SFC_RXQ_INITIALIZED);
 805
 806         rxq_info->rxq = NULL;
 807         rxq_info->entries = 0;
 808
 809         rte_free(rxq->sw_desc);
 810         sfc_dma_free(sa, &rxq->mem);
 811         rte_free(rxq);
 812 }
 813
 814 #if EFSYS_OPT_RX_SCALE
 815 efx_rx_hash_type_t
 816 sfc_rte_to_efx_hash_type(uint64_t rss_hf)
 817 {
 818         efx_rx_hash_type_t efx_hash_types = 0;
 819
 820         if ((rss_hf & (ETH_RSS_IPV4 | ETH_RSS_FRAG_IPV4 |
 821                        ETH_RSS_NONFRAG_IPV4_OTHER)) != 0)
 822                 efx_hash_types |= EFX_RX_HASH_IPV4;
 823
 824         if ((rss_hf & ETH_RSS_NONFRAG_IPV4_TCP) != 0)
 825                 efx_hash_types |= EFX_RX_HASH_TCPIPV4;
 826
 827         if ((rss_hf & (ETH_RSS_IPV6 | ETH_RSS_FRAG_IPV6 |
 828                         ETH_RSS_NONFRAG_IPV6_OTHER | ETH_RSS_IPV6_EX)) != 0)
 829                 efx_hash_types |= EFX_RX_HASH_IPV6;
 830
 831         if ((rss_hf & (ETH_RSS_NONFRAG_IPV6_TCP | ETH_RSS_IPV6_TCP_EX)) != 0)
 832                 efx_hash_types |= EFX_RX_HASH_TCPIPV6;
 833
 834         return efx_hash_types;
 835 }
 836
 837 uint64_t
 838 sfc_efx_to_rte_hash_type(efx_rx_hash_type_t efx_hash_types)
 839 {
 840         uint64_t rss_hf = 0;
 841
 842         if ((efx_hash_types & EFX_RX_HASH_IPV4) != 0)
 843                 rss_hf |= (ETH_RSS_IPV4 | ETH_RSS_FRAG_IPV4 |
 844                            ETH_RSS_NONFRAG_IPV4_OTHER);
 845
 846         if ((efx_hash_types & EFX_RX_HASH_TCPIPV4) != 0)
 847                 rss_hf |= ETH_RSS_NONFRAG_IPV4_TCP;
 848
 849         if ((efx_hash_types & EFX_RX_HASH_IPV6) != 0)
 850                 rss_hf |= (ETH_RSS_IPV6 | ETH_RSS_FRAG_IPV6 |
 851                            ETH_RSS_NONFRAG_IPV6_OTHER | ETH_RSS_IPV6_EX);
 852
 853         if ((efx_hash_types & EFX_RX_HASH_TCPIPV6) != 0)
 854                 rss_hf |= (ETH_RSS_NONFRAG_IPV6_TCP | ETH_RSS_IPV6_TCP_EX);
 855
 856         return rss_hf;
 857 }
 858 #endif
 859
 860 static int
 861 sfc_rx_rss_config(struct sfc_adapter *sa)
 862 {
 863         int rc = 0;
 864
 865 #if EFSYS_OPT_RX_SCALE
 866         if (sa->rss_channels > 1) {
 867                 rc = efx_rx_scale_mode_set(sa->nic, EFX_RX_HASHALG_TOEPLITZ,
 868                                            sa->rss_hash_types, B_TRUE);
 869                 if (rc != 0)
 870                         goto finish;
 871
 872                 rc = efx_rx_scale_key_set(sa->nic, sa->rss_key,
 873                                           sizeof(sa->rss_key));
 874                 if (rc != 0)
 875                         goto finish;
 876
 877                 rc = efx_rx_scale_tbl_set(sa->nic, sa->rss_tbl,
 878                                           sizeof(sa->rss_tbl));
 879         }
 880
 881 finish:
 882 #endif
 883         return rc;
 884 }
 885
 886 int
 887 sfc_rx_start(struct sfc_adapter *sa)
 888 {
 889         unsigned int sw_index;
 890         int rc;
 891
 892         sfc_log_init(sa, "rxq_count=%u", sa->rxq_count);
 893
 894         rc = efx_rx_init(sa->nic);
 895         if (rc != 0)
 896                 goto fail_rx_init;
 897
 898         rc = sfc_rx_rss_config(sa);
 899         if (rc != 0)
 900                 goto fail_rss_config;
 901
 902         for (sw_index = 0; sw_index < sa->rxq_count; ++sw_index) {
 903                 if ((!sa->rxq_info[sw_index].deferred_start ||
 904                      sa->rxq_info[sw_index].deferred_started)) {
 905                         rc = sfc_rx_qstart(sa, sw_index);
 906                         if (rc != 0)
 907                                 goto fail_rx_qstart;
 908                 }
 909         }
 910
 911         return 0;
 912
 913 fail_rx_qstart:
 914         while (sw_index-- > 0)
 915                 sfc_rx_qstop(sa, sw_index);
 916
 917 fail_rss_config:
 918         efx_rx_fini(sa->nic);
 919
 920 fail_rx_init:
 921         sfc_log_init(sa, "failed %d", rc);
 922         return rc;
 923 }
 924
 925 void
 926 sfc_rx_stop(struct sfc_adapter *sa)
 927 {
 928         unsigned int sw_index;
 929
 930         sfc_log_init(sa, "rxq_count=%u", sa->rxq_count);
 931
 932         sw_index = sa->rxq_count;
 933         while (sw_index-- > 0) {
 934                 if (sa->rxq_info[sw_index].rxq != NULL)
 935                         sfc_rx_qstop(sa, sw_index);
 936         }
 937
 938         efx_rx_fini(sa->nic);
 939 }
 940
 941 static int
 942 sfc_rx_qinit_info(struct sfc_adapter *sa, unsigned int sw_index)
 943 {
 944         struct sfc_rxq_info *rxq_info = &sa->rxq_info[sw_index];
 945         unsigned int max_entries;
 946
 947         max_entries = EFX_RXQ_MAXNDESCS;
 948         SFC_ASSERT(rte_is_power_of_2(max_entries));
 949
 950         rxq_info->max_entries = max_entries;
 951
 952         return 0;
 953 }
 954
 955 static int
 956 sfc_rx_check_mode(struct sfc_adapter *sa, struct rte_eth_rxmode *rxmode)
 957 {
 958         int rc = 0;
 959
 960         switch (rxmode->mq_mode) {
 961         case ETH_MQ_RX_NONE:
 962                 /* No special checks are required */
 963                 break;
 964 #if EFSYS_OPT_RX_SCALE
 965         case ETH_MQ_RX_RSS:
 966                 if (sa->rss_support == EFX_RX_SCALE_UNAVAILABLE) {
 967                         sfc_err(sa, "RSS is not available");
 968                         rc = EINVAL;
 969                 }
 970                 break;
 971 #endif
 972         default:
 973                 sfc_err(sa, "Rx multi-queue mode %u not supported",
 974                         rxmode->mq_mode);
 975                 rc = EINVAL;
 976         }
 977
 978         if (rxmode->header_split) {
 979                 sfc_err(sa, "Header split on Rx not supported");
 980                 rc = EINVAL;
 981         }
 982
 983         if (rxmode->hw_vlan_filter) {
 984                 sfc_err(sa, "HW VLAN filtering not supported");
 985                 rc = EINVAL;
 986         }
 987
 988         if (rxmode->hw_vlan_strip) {
 989                 sfc_err(sa, "HW VLAN stripping not supported");
 990                 rc = EINVAL;
 991         }
 992
 993         if (rxmode->hw_vlan_extend) {
 994                 sfc_err(sa,
 995                         "Q-in-Q HW VLAN stripping not supported");
 996                 rc = EINVAL;
 997         }
 998
 999         if (!rxmode->hw_strip_crc) {
1000                 sfc_warn(sa,
1001                          "FCS stripping control not supported - always stripped");
1002                 rxmode->hw_strip_crc = 1;
1003         }
1004
1005         if (rxmode->enable_lro) {
1006                 sfc_err(sa, "LRO not supported");
1007                 rc = EINVAL;
1008         }
1009
1010         return rc;
1011 }
1012
1013 /**
1014  * Initialize Rx subsystem.
1015  *
1016  * Called at device configuration stage when number of receive queues is
1017  * specified together with other device level receive configuration.
1018  *
1019  * It should be used to allocate NUMA-unaware resources.
1020  */
1021 int
1022 sfc_rx_init(struct sfc_adapter *sa)
1023 {
1024         struct rte_eth_conf *dev_conf = &sa->eth_dev->data->dev_conf;
1025         unsigned int sw_index;
1026         int rc;
1027
1028         rc = sfc_rx_check_mode(sa, &dev_conf->rxmode);
1029         if (rc != 0)
1030                 goto fail_check_mode;
1031
1032         sa->rxq_count = sa->eth_dev->data->nb_rx_queues;
1033
1034         rc = ENOMEM;
1035         sa->rxq_info = rte_calloc_socket("sfc-rxqs", sa->rxq_count,
1036                                          sizeof(struct sfc_rxq_info), 0,
1037                                          sa->socket_id);
1038         if (sa->rxq_info == NULL)
1039                 goto fail_rxqs_alloc;
1040
1041         for (sw_index = 0; sw_index < sa->rxq_count; ++sw_index) {
1042                 rc = sfc_rx_qinit_info(sa, sw_index);
1043                 if (rc != 0)
1044                         goto fail_rx_qinit_info;
1045         }
1046
1047 #if EFSYS_OPT_RX_SCALE
1048         sa->rss_channels = (dev_conf->rxmode.mq_mode == ETH_MQ_RX_RSS) ?
1049                            MIN(sa->rxq_count, EFX_MAXRSS) : 1;
1050
1051         if (sa->rss_channels > 1) {
1052                 for (sw_index = 0; sw_index < EFX_RSS_TBL_SIZE; ++sw_index)
1053                         sa->rss_tbl[sw_index] = sw_index % sa->rss_channels;
1054         }
1055 #endif
1056
1057         return 0;
1058
1059 fail_rx_qinit_info:
1060         rte_free(sa->rxq_info);
1061         sa->rxq_info = NULL;
1062
1063 fail_rxqs_alloc:
1064         sa->rxq_count = 0;
1065 fail_check_mode:
1066         sfc_log_init(sa, "failed %d", rc);
1067         return rc;
1068 }
1069
1070 /**
1071  * Shutdown Rx subsystem.
1072  *
1073  * Called at device close stage, for example, before device
1074  * reconfiguration or shutdown.
1075  */
1076 void
1077 sfc_rx_fini(struct sfc_adapter *sa)
1078 {
1079         unsigned int sw_index;
1080
1081         sw_index = sa->rxq_count;
1082         while (sw_index-- > 0) {
1083                 if (sa->rxq_info[sw_index].rxq != NULL)
1084                         sfc_rx_qfini(sa, sw_index);
1085         }
1086
1087         rte_free(sa->rxq_info);
1088         sa->rxq_info = NULL;
1089         sa->rxq_count = 0;
1090 }