drivers/net/sfc/sfc_rx.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  *
   3  * Copyright (c) 2016-2018 Solarflare Communications Inc.
   4  * All rights reserved.
   5  *
   6  * This software was jointly developed between OKTET Labs (under contract
   7  * for Solarflare) and Solarflare Communications, Inc.
   8  */
   9
  10 #include <rte_mempool.h>
  11
  12 #include "efx.h"
  13
  14 #include "sfc.h"
  15 #include "sfc_debug.h"
  16 #include "sfc_log.h"
  17 #include "sfc_ev.h"
  18 #include "sfc_rx.h"
  19 #include "sfc_kvargs.h"
  20 #include "sfc_tweak.h"
  21
  22 /*
  23  * Maximum number of Rx queue flush attempt in the case of failure or
  24  * flush timeout
  25  */
  26 #define SFC_RX_QFLUSH_ATTEMPTS          (3)
  27
  28 /*
  29  * Time to wait between event queue polling attempts when waiting for Rx
  30  * queue flush done or failed events.
  31  */
  32 #define SFC_RX_QFLUSH_POLL_WAIT_MS      (1)
  33
  34 /*
  35  * Maximum number of event queue polling attempts when waiting for Rx queue
  36  * flush done or failed events. It defines Rx queue flush attempt timeout
  37  * together with SFC_RX_QFLUSH_POLL_WAIT_MS.
  38  */
  39 #define SFC_RX_QFLUSH_POLL_ATTEMPTS     (2000)
  40
  41 void
  42 sfc_rx_qflush_done(struct sfc_rxq_info *rxq_info)
  43 {
  44         rxq_info->state |= SFC_RXQ_FLUSHED;
  45         rxq_info->state &= ~SFC_RXQ_FLUSHING;
  46 }
  47
  48 void
  49 sfc_rx_qflush_failed(struct sfc_rxq_info *rxq_info)
  50 {
  51         rxq_info->state |= SFC_RXQ_FLUSH_FAILED;
  52         rxq_info->state &= ~SFC_RXQ_FLUSHING;
  53 }
  54
  55 static void
  56 sfc_efx_rx_qrefill(struct sfc_efx_rxq *rxq)
  57 {
  58         unsigned int free_space;
  59         unsigned int bulks;
  60         void *objs[SFC_RX_REFILL_BULK];
  61         efsys_dma_addr_t addr[RTE_DIM(objs)];
  62         unsigned int added = rxq->added;
  63         unsigned int id;
  64         unsigned int i;
  65         struct sfc_efx_rx_sw_desc *rxd;
  66         struct rte_mbuf *m;
  67         uint16_t port_id = rxq->dp.dpq.port_id;
  68
  69         free_space = rxq->max_fill_level - (added - rxq->completed);
  70
  71         if (free_space < rxq->refill_threshold)
  72                 return;
  73
  74         bulks = free_space / RTE_DIM(objs);
  75         /* refill_threshold guarantees that bulks is positive */
  76         SFC_ASSERT(bulks > 0);
  77
  78         id = added & rxq->ptr_mask;
  79         do {
  80                 if (unlikely(rte_mempool_get_bulk(rxq->refill_mb_pool, objs,
  81                                                   RTE_DIM(objs)) < 0)) {
  82                         /*
  83                          * It is hardly a safe way to increment counter
  84                          * from different contexts, but all PMDs do it.
  85                          */
  86                         rxq->evq->sa->eth_dev->data->rx_mbuf_alloc_failed +=
  87                                 RTE_DIM(objs);
  88                         /* Return if we have posted nothing yet */
  89                         if (added == rxq->added)
  90                                 return;
  91                         /* Push posted */
  92                         break;
  93                 }
  94
  95                 for (i = 0; i < RTE_DIM(objs);
  96                      ++i, id = (id + 1) & rxq->ptr_mask) {
  97                         m = objs[i];
  98
  99                         MBUF_RAW_ALLOC_CHECK(m);
 100
 101                         rxd = &rxq->sw_desc[id];
 102                         rxd->mbuf = m;
 103
 104                         m->data_off = RTE_PKTMBUF_HEADROOM;
 105                         m->port = port_id;
 106
 107                         addr[i] = rte_pktmbuf_iova(m);
 108                 }
 109
 110                 efx_rx_qpost(rxq->common, addr, rxq->buf_size,
 111                              RTE_DIM(objs), rxq->completed, added);
 112                 added += RTE_DIM(objs);
 113         } while (--bulks > 0);
 114
 115         SFC_ASSERT(added != rxq->added);
 116         rxq->added = added;
 117         efx_rx_qpush(rxq->common, added, &rxq->pushed);
 118 }
 119
 120 static uint64_t
 121 sfc_efx_rx_desc_flags_to_offload_flags(const unsigned int desc_flags)
 122 {
 123         uint64_t mbuf_flags = 0;
 124
 125         switch (desc_flags & (EFX_PKT_IPV4 | EFX_CKSUM_IPV4)) {
 126         case (EFX_PKT_IPV4 | EFX_CKSUM_IPV4):
 127                 mbuf_flags |= PKT_RX_IP_CKSUM_GOOD;
 128                 break;
 129         case EFX_PKT_IPV4:
 130                 mbuf_flags |= PKT_RX_IP_CKSUM_BAD;
 131                 break;
 132         default:
 133                 RTE_BUILD_BUG_ON(PKT_RX_IP_CKSUM_UNKNOWN != 0);
 134                 SFC_ASSERT((mbuf_flags & PKT_RX_IP_CKSUM_MASK) ==
 135                            PKT_RX_IP_CKSUM_UNKNOWN);
 136                 break;
 137         }
 138
 139         switch ((desc_flags &
 140                  (EFX_PKT_TCP | EFX_PKT_UDP | EFX_CKSUM_TCPUDP))) {
 141         case (EFX_PKT_TCP | EFX_CKSUM_TCPUDP):
 142         case (EFX_PKT_UDP | EFX_CKSUM_TCPUDP):
 143                 mbuf_flags |= PKT_RX_L4_CKSUM_GOOD;
 144                 break;
 145         case EFX_PKT_TCP:
 146         case EFX_PKT_UDP:
 147                 mbuf_flags |= PKT_RX_L4_CKSUM_BAD;
 148                 break;
 149         default:
 150                 RTE_BUILD_BUG_ON(PKT_RX_L4_CKSUM_UNKNOWN != 0);
 151                 SFC_ASSERT((mbuf_flags & PKT_RX_L4_CKSUM_MASK) ==
 152                            PKT_RX_L4_CKSUM_UNKNOWN);
 153                 break;
 154         }
 155
 156         return mbuf_flags;
 157 }
 158
 159 static uint32_t
 160 sfc_efx_rx_desc_flags_to_packet_type(const unsigned int desc_flags)
 161 {
 162         return RTE_PTYPE_L2_ETHER |
 163                 ((desc_flags & EFX_PKT_IPV4) ?
 164                         RTE_PTYPE_L3_IPV4_EXT_UNKNOWN : 0) |
 165                 ((desc_flags & EFX_PKT_IPV6) ?
 166                         RTE_PTYPE_L3_IPV6_EXT_UNKNOWN : 0) |
 167                 ((desc_flags & EFX_PKT_TCP) ? RTE_PTYPE_L4_TCP : 0) |
 168                 ((desc_flags & EFX_PKT_UDP) ? RTE_PTYPE_L4_UDP : 0);
 169 }
 170
 171 static const uint32_t *
 172 sfc_efx_supported_ptypes_get(__rte_unused uint32_t tunnel_encaps)
 173 {
 174         static const uint32_t ptypes[] = {
 175                 RTE_PTYPE_L2_ETHER,
 176                 RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
 177                 RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
 178                 RTE_PTYPE_L4_TCP,
 179                 RTE_PTYPE_L4_UDP,
 180                 RTE_PTYPE_UNKNOWN
 181         };
 182
 183         return ptypes;
 184 }
 185
 186 static void
 187 sfc_efx_rx_set_rss_hash(struct sfc_efx_rxq *rxq, unsigned int flags,
 188                         struct rte_mbuf *m)
 189 {
 190         uint8_t *mbuf_data;
 191
 192
 193         if ((rxq->flags & SFC_EFX_RXQ_FLAG_RSS_HASH) == 0)
 194                 return;
 195
 196         mbuf_data = rte_pktmbuf_mtod(m, uint8_t *);
 197
 198         if (flags & (EFX_PKT_IPV4 | EFX_PKT_IPV6)) {
 199                 m->hash.rss = efx_pseudo_hdr_hash_get(rxq->common,
 200                                                       EFX_RX_HASHALG_TOEPLITZ,
 201                                                       mbuf_data);
 202
 203                 m->ol_flags |= PKT_RX_RSS_HASH;
 204         }
 205 }
 206
 207 static uint16_t
 208 sfc_efx_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 209 {
 210         struct sfc_dp_rxq *dp_rxq = rx_queue;
 211         struct sfc_efx_rxq *rxq = sfc_efx_rxq_by_dp_rxq(dp_rxq);
 212         unsigned int completed;
 213         unsigned int prefix_size = rxq->prefix_size;
 214         unsigned int done_pkts = 0;
 215         boolean_t discard_next = B_FALSE;
 216         struct rte_mbuf *scatter_pkt = NULL;
 217
 218         if (unlikely((rxq->flags & SFC_EFX_RXQ_FLAG_RUNNING) == 0))
 219                 return 0;
 220
 221         sfc_ev_qpoll(rxq->evq);
 222
 223         completed = rxq->completed;
 224         while (completed != rxq->pending && done_pkts < nb_pkts) {
 225                 unsigned int id;
 226                 struct sfc_efx_rx_sw_desc *rxd;
 227                 struct rte_mbuf *m;
 228                 unsigned int seg_len;
 229                 unsigned int desc_flags;
 230
 231                 id = completed++ & rxq->ptr_mask;
 232                 rxd = &rxq->sw_desc[id];
 233                 m = rxd->mbuf;
 234                 desc_flags = rxd->flags;
 235
 236                 if (discard_next)
 237                         goto discard;
 238
 239                 if (desc_flags & (EFX_ADDR_MISMATCH | EFX_DISCARD))
 240                         goto discard;
 241
 242                 if (desc_flags & EFX_PKT_PREFIX_LEN) {
 243                         uint16_t tmp_size;
 244                         int rc __rte_unused;
 245
 246                         rc = efx_pseudo_hdr_pkt_length_get(rxq->common,
 247                                 rte_pktmbuf_mtod(m, uint8_t *), &tmp_size);
 248                         SFC_ASSERT(rc == 0);
 249                         seg_len = tmp_size;
 250                 } else {
 251                         seg_len = rxd->size - prefix_size;
 252                 }
 253
 254                 rte_pktmbuf_data_len(m) = seg_len;
 255                 rte_pktmbuf_pkt_len(m) = seg_len;
 256
 257                 if (scatter_pkt != NULL) {
 258                         if (rte_pktmbuf_chain(scatter_pkt, m) != 0) {
 259                                 rte_pktmbuf_free(scatter_pkt);
 260                                 goto discard;
 261                         }
 262                         /* The packet to deliver */
 263                         m = scatter_pkt;
 264                 }
 265
 266                 if (desc_flags & EFX_PKT_CONT) {
 267                         /* The packet is scattered, more fragments to come */
 268                         scatter_pkt = m;
 269                         /* Further fragments have no prefix */
 270                         prefix_size = 0;
 271                         continue;
 272                 }
 273
 274                 /* Scattered packet is done */
 275                 scatter_pkt = NULL;
 276                 /* The first fragment of the packet has prefix */
 277                 prefix_size = rxq->prefix_size;
 278
 279                 m->ol_flags =
 280                         sfc_efx_rx_desc_flags_to_offload_flags(desc_flags);
 281                 m->packet_type =
 282                         sfc_efx_rx_desc_flags_to_packet_type(desc_flags);
 283
 284                 /*
 285                  * Extract RSS hash from the packet prefix and
 286                  * set the corresponding field (if needed and possible)
 287                  */
 288                 sfc_efx_rx_set_rss_hash(rxq, desc_flags, m);
 289
 290                 m->data_off += prefix_size;
 291
 292                 *rx_pkts++ = m;
 293                 done_pkts++;
 294                 continue;
 295
 296 discard:
 297                 discard_next = ((desc_flags & EFX_PKT_CONT) != 0);
 298                 rte_mbuf_raw_free(m);
 299                 rxd->mbuf = NULL;
 300         }
 301
 302         /* pending is only moved when entire packet is received */
 303         SFC_ASSERT(scatter_pkt == NULL);
 304
 305         rxq->completed = completed;
 306
 307         sfc_efx_rx_qrefill(rxq);
 308
 309         return done_pkts;
 310 }
 311
 312 static sfc_dp_rx_qdesc_npending_t sfc_efx_rx_qdesc_npending;
 313 static unsigned int
 314 sfc_efx_rx_qdesc_npending(struct sfc_dp_rxq *dp_rxq)
 315 {
 316         struct sfc_efx_rxq *rxq = sfc_efx_rxq_by_dp_rxq(dp_rxq);
 317
 318         if ((rxq->flags & SFC_EFX_RXQ_FLAG_RUNNING) == 0)
 319                 return 0;
 320
 321         sfc_ev_qpoll(rxq->evq);
 322
 323         return rxq->pending - rxq->completed;
 324 }
 325
 326 static sfc_dp_rx_qdesc_status_t sfc_efx_rx_qdesc_status;
 327 static int
 328 sfc_efx_rx_qdesc_status(struct sfc_dp_rxq *dp_rxq, uint16_t offset)
 329 {
 330         struct sfc_efx_rxq *rxq = sfc_efx_rxq_by_dp_rxq(dp_rxq);
 331
 332         if (unlikely(offset > rxq->ptr_mask))
 333                 return -EINVAL;
 334
 335         /*
 336          * Poll EvQ to derive up-to-date 'rxq->pending' figure;
 337          * it is required for the queue to be running, but the
 338          * check is omitted because API design assumes that it
 339          * is the duty of the caller to satisfy all conditions
 340          */
 341         SFC_ASSERT((rxq->flags & SFC_EFX_RXQ_FLAG_RUNNING) ==
 342                    SFC_EFX_RXQ_FLAG_RUNNING);
 343         sfc_ev_qpoll(rxq->evq);
 344
 345         /*
 346          * There is a handful of reserved entries in the ring,
 347          * but an explicit check whether the offset points to
 348          * a reserved entry is neglected since the two checks
 349          * below rely on the figures which take the HW limits
 350          * into account and thus if an entry is reserved, the
 351          * checks will fail and UNAVAIL code will be returned
 352          */
 353
 354         if (offset < (rxq->pending - rxq->completed))
 355                 return RTE_ETH_RX_DESC_DONE;
 356
 357         if (offset < (rxq->added - rxq->completed))
 358                 return RTE_ETH_RX_DESC_AVAIL;
 359
 360         return RTE_ETH_RX_DESC_UNAVAIL;
 361 }
 362
 363 /** Get Rx datapath ops by the datapath RxQ handle */
 364 const struct sfc_dp_rx *
 365 sfc_dp_rx_by_dp_rxq(const struct sfc_dp_rxq *dp_rxq)
 366 {
 367         const struct sfc_dp_queue *dpq = &dp_rxq->dpq;
 368         struct rte_eth_dev *eth_dev;
 369         struct sfc_adapter_priv *sap;
 370
 371         SFC_ASSERT(rte_eth_dev_is_valid_port(dpq->port_id));
 372         eth_dev = &rte_eth_devices[dpq->port_id];
 373
 374         sap = sfc_adapter_priv_by_eth_dev(eth_dev);
 375
 376         return sap->dp_rx;
 377 }
 378
 379 struct sfc_rxq_info *
 380 sfc_rxq_info_by_dp_rxq(const struct sfc_dp_rxq *dp_rxq)
 381 {
 382         const struct sfc_dp_queue *dpq = &dp_rxq->dpq;
 383         struct rte_eth_dev *eth_dev;
 384         struct sfc_adapter_shared *sas;
 385
 386         SFC_ASSERT(rte_eth_dev_is_valid_port(dpq->port_id));
 387         eth_dev = &rte_eth_devices[dpq->port_id];
 388
 389         sas = sfc_adapter_shared_by_eth_dev(eth_dev);
 390
 391         SFC_ASSERT(dpq->queue_id < sas->rxq_count);
 392         return &sas->rxq_info[dpq->queue_id];
 393 }
 394
 395 struct sfc_rxq *
 396 sfc_rxq_by_dp_rxq(const struct sfc_dp_rxq *dp_rxq)
 397 {
 398         const struct sfc_dp_queue *dpq = &dp_rxq->dpq;
 399         struct rte_eth_dev *eth_dev;
 400         struct sfc_adapter *sa;
 401
 402         SFC_ASSERT(rte_eth_dev_is_valid_port(dpq->port_id));
 403         eth_dev = &rte_eth_devices[dpq->port_id];
 404
 405         sa = sfc_adapter_by_eth_dev(eth_dev);
 406
 407         SFC_ASSERT(dpq->queue_id < sfc_sa2shared(sa)->rxq_count);
 408         return &sa->rxq_ctrl[dpq->queue_id];
 409 }
 410
 411 static sfc_dp_rx_qsize_up_rings_t sfc_efx_rx_qsize_up_rings;
 412 static int
 413 sfc_efx_rx_qsize_up_rings(uint16_t nb_rx_desc,
 414                           __rte_unused struct sfc_dp_rx_hw_limits *limits,
 415                           __rte_unused struct rte_mempool *mb_pool,
 416                           unsigned int *rxq_entries,
 417                           unsigned int *evq_entries,
 418                           unsigned int *rxq_max_fill_level)
 419 {
 420         *rxq_entries = nb_rx_desc;
 421         *evq_entries = nb_rx_desc;
 422         *rxq_max_fill_level = EFX_RXQ_LIMIT(*rxq_entries);
 423         return 0;
 424 }
 425
 426 static sfc_dp_rx_qcreate_t sfc_efx_rx_qcreate;
 427 static int
 428 sfc_efx_rx_qcreate(uint16_t port_id, uint16_t queue_id,
 429                    const struct rte_pci_addr *pci_addr, int socket_id,
 430                    const struct sfc_dp_rx_qcreate_info *info,
 431                    struct sfc_dp_rxq **dp_rxqp)
 432 {
 433         struct sfc_efx_rxq *rxq;
 434         int rc;
 435
 436         rc = ENOMEM;
 437         rxq = rte_zmalloc_socket("sfc-efx-rxq", sizeof(*rxq),
 438                                  RTE_CACHE_LINE_SIZE, socket_id);
 439         if (rxq == NULL)
 440                 goto fail_rxq_alloc;
 441
 442         sfc_dp_queue_init(&rxq->dp.dpq, port_id, queue_id, pci_addr);
 443
 444         rc = ENOMEM;
 445         rxq->sw_desc = rte_calloc_socket("sfc-efx-rxq-sw_desc",
 446                                          info->rxq_entries,
 447                                          sizeof(*rxq->sw_desc),
 448                                          RTE_CACHE_LINE_SIZE, socket_id);
 449         if (rxq->sw_desc == NULL)
 450                 goto fail_desc_alloc;
 451
 452         /* efx datapath is bound to efx control path */
 453         rxq->evq = sfc_rxq_by_dp_rxq(&rxq->dp)->evq;
 454         if (info->flags & SFC_RXQ_FLAG_RSS_HASH)
 455                 rxq->flags |= SFC_EFX_RXQ_FLAG_RSS_HASH;
 456         rxq->ptr_mask = info->rxq_entries - 1;
 457         rxq->batch_max = info->batch_max;
 458         rxq->prefix_size = info->prefix_size;
 459         rxq->max_fill_level = info->max_fill_level;
 460         rxq->refill_threshold = info->refill_threshold;
 461         rxq->buf_size = info->buf_size;
 462         rxq->refill_mb_pool = info->refill_mb_pool;
 463
 464         *dp_rxqp = &rxq->dp;
 465         return 0;
 466
 467 fail_desc_alloc:
 468         rte_free(rxq);
 469
 470 fail_rxq_alloc:
 471         return rc;
 472 }
 473
 474 static sfc_dp_rx_qdestroy_t sfc_efx_rx_qdestroy;
 475 static void
 476 sfc_efx_rx_qdestroy(struct sfc_dp_rxq *dp_rxq)
 477 {
 478         struct sfc_efx_rxq *rxq = sfc_efx_rxq_by_dp_rxq(dp_rxq);
 479
 480         rte_free(rxq->sw_desc);
 481         rte_free(rxq);
 482 }
 483
 484 static sfc_dp_rx_qstart_t sfc_efx_rx_qstart;
 485 static int
 486 sfc_efx_rx_qstart(struct sfc_dp_rxq *dp_rxq,
 487                   __rte_unused unsigned int evq_read_ptr)
 488 {
 489         /* libefx-based datapath is specific to libefx-based PMD */
 490         struct sfc_efx_rxq *rxq = sfc_efx_rxq_by_dp_rxq(dp_rxq);
 491         struct sfc_rxq *crxq = sfc_rxq_by_dp_rxq(dp_rxq);
 492
 493         rxq->common = crxq->common;
 494
 495         rxq->pending = rxq->completed = rxq->added = rxq->pushed = 0;
 496
 497         sfc_efx_rx_qrefill(rxq);
 498
 499         rxq->flags |= (SFC_EFX_RXQ_FLAG_STARTED | SFC_EFX_RXQ_FLAG_RUNNING);
 500
 501         return 0;
 502 }
 503
 504 static sfc_dp_rx_qstop_t sfc_efx_rx_qstop;
 505 static void
 506 sfc_efx_rx_qstop(struct sfc_dp_rxq *dp_rxq,
 507                  __rte_unused unsigned int *evq_read_ptr)
 508 {
 509         struct sfc_efx_rxq *rxq = sfc_efx_rxq_by_dp_rxq(dp_rxq);
 510
 511         rxq->flags &= ~SFC_EFX_RXQ_FLAG_RUNNING;
 512
 513         /* libefx-based datapath is bound to libefx-based PMD and uses
 514          * event queue structure directly. So, there is no necessity to
 515          * return EvQ read pointer.
 516          */
 517 }
 518
 519 static sfc_dp_rx_qpurge_t sfc_efx_rx_qpurge;
 520 static void
 521 sfc_efx_rx_qpurge(struct sfc_dp_rxq *dp_rxq)
 522 {
 523         struct sfc_efx_rxq *rxq = sfc_efx_rxq_by_dp_rxq(dp_rxq);
 524         unsigned int i;
 525         struct sfc_efx_rx_sw_desc *rxd;
 526
 527         for (i = rxq->completed; i != rxq->added; ++i) {
 528                 rxd = &rxq->sw_desc[i & rxq->ptr_mask];
 529                 rte_mbuf_raw_free(rxd->mbuf);
 530                 rxd->mbuf = NULL;
 531                 /* Packed stream relies on 0 in inactive SW desc.
 532                  * Rx queue stop is not performance critical, so
 533                  * there is no harm to do it always.
 534                  */
 535                 rxd->flags = 0;
 536                 rxd->size = 0;
 537         }
 538
 539         rxq->flags &= ~SFC_EFX_RXQ_FLAG_STARTED;
 540 }
 541
 542 struct sfc_dp_rx sfc_efx_rx = {
 543         .dp = {
 544                 .name           = SFC_KVARG_DATAPATH_EFX,
 545                 .type           = SFC_DP_RX,
 546                 .hw_fw_caps     = 0,
 547         },
 548         .features               = SFC_DP_RX_FEAT_SCATTER |
 549                                   SFC_DP_RX_FEAT_CHECKSUM,
 550         .qsize_up_rings         = sfc_efx_rx_qsize_up_rings,
 551         .qcreate                = sfc_efx_rx_qcreate,
 552         .qdestroy               = sfc_efx_rx_qdestroy,
 553         .qstart                 = sfc_efx_rx_qstart,
 554         .qstop                  = sfc_efx_rx_qstop,
 555         .qpurge                 = sfc_efx_rx_qpurge,
 556         .supported_ptypes_get   = sfc_efx_supported_ptypes_get,
 557         .qdesc_npending         = sfc_efx_rx_qdesc_npending,
 558         .qdesc_status           = sfc_efx_rx_qdesc_status,
 559         .pkt_burst              = sfc_efx_recv_pkts,
 560 };
 561
 562 static void
 563 sfc_rx_qflush(struct sfc_adapter *sa, unsigned int sw_index)
 564 {
 565         struct sfc_rxq_info *rxq_info;
 566         struct sfc_rxq *rxq;
 567         unsigned int retry_count;
 568         unsigned int wait_count;
 569         int rc;
 570
 571         rxq_info = &sfc_sa2shared(sa)->rxq_info[sw_index];
 572         SFC_ASSERT(rxq_info->state & SFC_RXQ_STARTED);
 573
 574         rxq = &sa->rxq_ctrl[sw_index];
 575
 576         /*
 577          * Retry Rx queue flushing in the case of flush failed or
 578          * timeout. In the worst case it can delay for 6 seconds.
 579          */
 580         for (retry_count = 0;
 581              ((rxq_info->state & SFC_RXQ_FLUSHED) == 0) &&
 582              (retry_count < SFC_RX_QFLUSH_ATTEMPTS);
 583              ++retry_count) {
 584                 rc = efx_rx_qflush(rxq->common);
 585                 if (rc != 0) {
 586                         rxq_info->state |= (rc == EALREADY) ?
 587                                 SFC_RXQ_FLUSHED : SFC_RXQ_FLUSH_FAILED;
 588                         break;
 589                 }
 590                 rxq_info->state &= ~SFC_RXQ_FLUSH_FAILED;
 591                 rxq_info->state |= SFC_RXQ_FLUSHING;
 592
 593                 /*
 594                  * Wait for Rx queue flush done or failed event at least
 595                  * SFC_RX_QFLUSH_POLL_WAIT_MS milliseconds and not more
 596                  * than 2 seconds (SFC_RX_QFLUSH_POLL_WAIT_MS multiplied
 597                  * by SFC_RX_QFLUSH_POLL_ATTEMPTS).
 598                  */
 599                 wait_count = 0;
 600                 do {
 601                         rte_delay_ms(SFC_RX_QFLUSH_POLL_WAIT_MS);
 602                         sfc_ev_qpoll(rxq->evq);
 603                 } while ((rxq_info->state & SFC_RXQ_FLUSHING) &&
 604                          (wait_count++ < SFC_RX_QFLUSH_POLL_ATTEMPTS));
 605
 606                 if (rxq_info->state & SFC_RXQ_FLUSHING)
 607                         sfc_err(sa, "RxQ %u flush timed out", sw_index);
 608
 609                 if (rxq_info->state & SFC_RXQ_FLUSH_FAILED)
 610                         sfc_err(sa, "RxQ %u flush failed", sw_index);
 611
 612                 if (rxq_info->state & SFC_RXQ_FLUSHED)
 613                         sfc_notice(sa, "RxQ %u flushed", sw_index);
 614         }
 615
 616         sa->priv.dp_rx->qpurge(rxq_info->dp);
 617 }
 618
 619 static int
 620 sfc_rx_default_rxq_set_filter(struct sfc_adapter *sa, struct sfc_rxq *rxq)
 621 {
 622         struct sfc_rss *rss = &sfc_sa2shared(sa)->rss;
 623         boolean_t need_rss = (rss->channels > 0) ? B_TRUE : B_FALSE;
 624         struct sfc_port *port = &sa->port;
 625         int rc;
 626
 627         /*
 628          * If promiscuous or all-multicast mode has been requested, setting
 629          * filter for the default Rx queue might fail, in particular, while
 630          * running over PCI function which is not a member of corresponding
 631          * privilege groups; if this occurs, few iterations will be made to
 632          * repeat this step without promiscuous and all-multicast flags set
 633          */
 634 retry:
 635         rc = efx_mac_filter_default_rxq_set(sa->nic, rxq->common, need_rss);
 636         if (rc == 0)
 637                 return 0;
 638         else if (rc != EOPNOTSUPP)
 639                 return rc;
 640
 641         if (port->promisc) {
 642                 sfc_warn(sa, "promiscuous mode has been requested, "
 643                              "but the HW rejects it");
 644                 sfc_warn(sa, "promiscuous mode will be disabled");
 645
 646                 port->promisc = B_FALSE;
 647                 rc = sfc_set_rx_mode(sa);
 648                 if (rc != 0)
 649                         return rc;
 650
 651                 goto retry;
 652         }
 653
 654         if (port->allmulti) {
 655                 sfc_warn(sa, "all-multicast mode has been requested, "
 656                              "but the HW rejects it");
 657                 sfc_warn(sa, "all-multicast mode will be disabled");
 658
 659                 port->allmulti = B_FALSE;
 660                 rc = sfc_set_rx_mode(sa);
 661                 if (rc != 0)
 662                         return rc;
 663
 664                 goto retry;
 665         }
 666
 667         return rc;
 668 }
 669
 670 int
 671 sfc_rx_qstart(struct sfc_adapter *sa, unsigned int sw_index)
 672 {
 673         struct sfc_rxq_info *rxq_info;
 674         struct sfc_rxq *rxq;
 675         struct sfc_evq *evq;
 676         int rc;
 677
 678         sfc_log_init(sa, "sw_index=%u", sw_index);
 679
 680         SFC_ASSERT(sw_index < sfc_sa2shared(sa)->rxq_count);
 681
 682         rxq_info = &sfc_sa2shared(sa)->rxq_info[sw_index];
 683         SFC_ASSERT(rxq_info->state == SFC_RXQ_INITIALIZED);
 684
 685         rxq = &sa->rxq_ctrl[sw_index];
 686         evq = rxq->evq;
 687
 688         rc = sfc_ev_qstart(evq, sfc_evq_index_by_rxq_sw_index(sa, sw_index));
 689         if (rc != 0)
 690                 goto fail_ev_qstart;
 691
 692         switch (rxq_info->type) {
 693         case EFX_RXQ_TYPE_DEFAULT:
 694                 rc = efx_rx_qcreate(sa->nic, rxq->hw_index, 0, rxq_info->type,
 695                         rxq->buf_size,
 696                         &rxq->mem, rxq_info->entries, 0 /* not used on EF10 */,
 697                         rxq_info->type_flags, evq->common, &rxq->common);
 698                 break;
 699         case EFX_RXQ_TYPE_ES_SUPER_BUFFER: {
 700                 struct rte_mempool *mp = rxq_info->refill_mb_pool;
 701                 struct rte_mempool_info mp_info;
 702
 703                 rc = rte_mempool_ops_get_info(mp, &mp_info);
 704                 if (rc != 0) {
 705                         /* Positive errno is used in the driver */
 706                         rc = -rc;
 707                         goto fail_mp_get_info;
 708                 }
 709                 if (mp_info.contig_block_size <= 0) {
 710                         rc = EINVAL;
 711                         goto fail_bad_contig_block_size;
 712                 }
 713                 rc = efx_rx_qcreate_es_super_buffer(sa->nic, rxq->hw_index, 0,
 714                         mp_info.contig_block_size, rxq->buf_size,
 715                         mp->header_size + mp->elt_size + mp->trailer_size,
 716                         sa->rxd_wait_timeout_ns,
 717                         &rxq->mem, rxq_info->entries, rxq_info->type_flags,
 718                         evq->common, &rxq->common);
 719                 break;
 720         }
 721         default:
 722                 rc = ENOTSUP;
 723         }
 724         if (rc != 0)
 725                 goto fail_rx_qcreate;
 726
 727         efx_rx_qenable(rxq->common);
 728
 729         rc = sa->priv.dp_rx->qstart(rxq_info->dp, evq->read_ptr);
 730         if (rc != 0)
 731                 goto fail_dp_qstart;
 732
 733         rxq_info->state |= SFC_RXQ_STARTED;
 734
 735         if (sw_index == 0 && !sfc_sa2shared(sa)->isolated) {
 736                 rc = sfc_rx_default_rxq_set_filter(sa, rxq);
 737                 if (rc != 0)
 738                         goto fail_mac_filter_default_rxq_set;
 739         }
 740
 741         /* It seems to be used by DPDK for debug purposes only ('rte_ether') */
 742         sa->eth_dev->data->rx_queue_state[sw_index] =
 743                 RTE_ETH_QUEUE_STATE_STARTED;
 744
 745         return 0;
 746
 747 fail_mac_filter_default_rxq_set:
 748         sa->priv.dp_rx->qstop(rxq_info->dp, &rxq->evq->read_ptr);
 749
 750 fail_dp_qstart:
 751         sfc_rx_qflush(sa, sw_index);
 752
 753 fail_rx_qcreate:
 754 fail_bad_contig_block_size:
 755 fail_mp_get_info:
 756         sfc_ev_qstop(evq);
 757
 758 fail_ev_qstart:
 759         return rc;
 760 }
 761
 762 void
 763 sfc_rx_qstop(struct sfc_adapter *sa, unsigned int sw_index)
 764 {
 765         struct sfc_rxq_info *rxq_info;
 766         struct sfc_rxq *rxq;
 767
 768         sfc_log_init(sa, "sw_index=%u", sw_index);
 769
 770         SFC_ASSERT(sw_index < sfc_sa2shared(sa)->rxq_count);
 771
 772         rxq_info = &sfc_sa2shared(sa)->rxq_info[sw_index];
 773
 774         if (rxq_info->state == SFC_RXQ_INITIALIZED)
 775                 return;
 776         SFC_ASSERT(rxq_info->state & SFC_RXQ_STARTED);
 777
 778         /* It seems to be used by DPDK for debug purposes only ('rte_ether') */
 779         sa->eth_dev->data->rx_queue_state[sw_index] =
 780                 RTE_ETH_QUEUE_STATE_STOPPED;
 781
 782         rxq = &sa->rxq_ctrl[sw_index];
 783         sa->priv.dp_rx->qstop(rxq_info->dp, &rxq->evq->read_ptr);
 784
 785         if (sw_index == 0)
 786                 efx_mac_filter_default_rxq_clear(sa->nic);
 787
 788         sfc_rx_qflush(sa, sw_index);
 789
 790         rxq_info->state = SFC_RXQ_INITIALIZED;
 791
 792         efx_rx_qdestroy(rxq->common);
 793
 794         sfc_ev_qstop(rxq->evq);
 795 }
 796
 797 uint64_t
 798 sfc_rx_get_dev_offload_caps(struct sfc_adapter *sa)
 799 {
 800         const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
 801         uint64_t caps = 0;
 802
 803         caps |= DEV_RX_OFFLOAD_JUMBO_FRAME;
 804
 805         if (sa->priv.dp_rx->features & SFC_DP_RX_FEAT_CHECKSUM) {
 806                 caps |= DEV_RX_OFFLOAD_IPV4_CKSUM;
 807                 caps |= DEV_RX_OFFLOAD_UDP_CKSUM;
 808                 caps |= DEV_RX_OFFLOAD_TCP_CKSUM;
 809         }
 810
 811         if (encp->enc_tunnel_encapsulations_supported &&
 812             (sa->priv.dp_rx->features & SFC_DP_RX_FEAT_TUNNELS))
 813                 caps |= DEV_RX_OFFLOAD_OUTER_IPV4_CKSUM;
 814
 815         return caps;
 816 }
 817
 818 uint64_t
 819 sfc_rx_get_queue_offload_caps(struct sfc_adapter *sa)
 820 {
 821         uint64_t caps = 0;
 822
 823         if (sa->priv.dp_rx->features & SFC_DP_RX_FEAT_SCATTER)
 824                 caps |= DEV_RX_OFFLOAD_SCATTER;
 825
 826         return caps;
 827 }
 828
 829 static int
 830 sfc_rx_qcheck_conf(struct sfc_adapter *sa, unsigned int rxq_max_fill_level,
 831                    const struct rte_eth_rxconf *rx_conf,
 832                    __rte_unused uint64_t offloads)
 833 {
 834         int rc = 0;
 835
 836         if (rx_conf->rx_thresh.pthresh != 0 ||
 837             rx_conf->rx_thresh.hthresh != 0 ||
 838             rx_conf->rx_thresh.wthresh != 0) {
 839                 sfc_warn(sa,
 840                         "RxQ prefetch/host/writeback thresholds are not supported");
 841         }
 842
 843         if (rx_conf->rx_free_thresh > rxq_max_fill_level) {
 844                 sfc_err(sa,
 845                         "RxQ free threshold too large: %u vs maximum %u",
 846                         rx_conf->rx_free_thresh, rxq_max_fill_level);
 847                 rc = EINVAL;
 848         }
 849
 850         if (rx_conf->rx_drop_en == 0) {
 851                 sfc_err(sa, "RxQ drop disable is not supported");
 852                 rc = EINVAL;
 853         }
 854
 855         return rc;
 856 }
 857
 858 static unsigned int
 859 sfc_rx_mbuf_data_alignment(struct rte_mempool *mb_pool)
 860 {
 861         uint32_t data_off;
 862         uint32_t order;
 863
 864         /* The mbuf object itself is always cache line aligned */
 865         order = rte_bsf32(RTE_CACHE_LINE_SIZE);
 866
 867         /* Data offset from mbuf object start */
 868         data_off = sizeof(struct rte_mbuf) + rte_pktmbuf_priv_size(mb_pool) +
 869                 RTE_PKTMBUF_HEADROOM;
 870
 871         order = MIN(order, rte_bsf32(data_off));
 872
 873         return 1u << order;
 874 }
 875
 876 static uint16_t
 877 sfc_rx_mb_pool_buf_size(struct sfc_adapter *sa, struct rte_mempool *mb_pool)
 878 {
 879         const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
 880         const uint32_t nic_align_start = MAX(1, encp->enc_rx_buf_align_start);
 881         const uint32_t nic_align_end = MAX(1, encp->enc_rx_buf_align_end);
 882         uint16_t buf_size;
 883         unsigned int buf_aligned;
 884         unsigned int start_alignment;
 885         unsigned int end_padding_alignment;
 886
 887         /* Below it is assumed that both alignments are power of 2 */
 888         SFC_ASSERT(rte_is_power_of_2(nic_align_start));
 889         SFC_ASSERT(rte_is_power_of_2(nic_align_end));
 890
 891         /*
 892          * mbuf is always cache line aligned, double-check
 893          * that it meets rx buffer start alignment requirements.
 894          */
 895
 896         /* Start from mbuf pool data room size */
 897         buf_size = rte_pktmbuf_data_room_size(mb_pool);
 898
 899         /* Remove headroom */
 900         if (buf_size <= RTE_PKTMBUF_HEADROOM) {
 901                 sfc_err(sa,
 902                         "RxQ mbuf pool %s object data room size %u is smaller than headroom %u",
 903                         mb_pool->name, buf_size, RTE_PKTMBUF_HEADROOM);
 904                 return 0;
 905         }
 906         buf_size -= RTE_PKTMBUF_HEADROOM;
 907
 908         /* Calculate guaranteed data start alignment */
 909         buf_aligned = sfc_rx_mbuf_data_alignment(mb_pool);
 910
 911         /* Reserve space for start alignment */
 912         if (buf_aligned < nic_align_start) {
 913                 start_alignment = nic_align_start - buf_aligned;
 914                 if (buf_size <= start_alignment) {
 915                         sfc_err(sa,
 916                                 "RxQ mbuf pool %s object data room size %u is insufficient for headroom %u and buffer start alignment %u required by NIC",
 917                                 mb_pool->name,
 918                                 rte_pktmbuf_data_room_size(mb_pool),
 919                                 RTE_PKTMBUF_HEADROOM, start_alignment);
 920                         return 0;
 921                 }
 922                 buf_aligned = nic_align_start;
 923                 buf_size -= start_alignment;
 924         } else {
 925                 start_alignment = 0;
 926         }
 927
 928         /* Make sure that end padding does not write beyond the buffer */
 929         if (buf_aligned < nic_align_end) {
 930                 /*
 931                  * Estimate space which can be lost. If guarnteed buffer
 932                  * size is odd, lost space is (nic_align_end - 1). More
 933                  * accurate formula is below.
 934                  */
 935                 end_padding_alignment = nic_align_end -
 936                         MIN(buf_aligned, 1u << (rte_bsf32(buf_size) - 1));
 937                 if (buf_size <= end_padding_alignment) {
 938                         sfc_err(sa,
 939                                 "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",
 940                                 mb_pool->name,
 941                                 rte_pktmbuf_data_room_size(mb_pool),
 942                                 RTE_PKTMBUF_HEADROOM, start_alignment,
 943                                 end_padding_alignment);
 944                         return 0;
 945                 }
 946                 buf_size -= end_padding_alignment;
 947         } else {
 948                 /*
 949                  * Start is aligned the same or better than end,
 950                  * just align length.
 951                  */
 952                 buf_size = P2ALIGN(buf_size, nic_align_end);
 953         }
 954
 955         return buf_size;
 956 }
 957
 958 int
 959 sfc_rx_qinit(struct sfc_adapter *sa, unsigned int sw_index,
 960              uint16_t nb_rx_desc, unsigned int socket_id,
 961              const struct rte_eth_rxconf *rx_conf,
 962              struct rte_mempool *mb_pool)
 963 {
 964         const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
 965         struct sfc_rss *rss = &sfc_sa2shared(sa)->rss;
 966         int rc;
 967         unsigned int rxq_entries;
 968         unsigned int evq_entries;
 969         unsigned int rxq_max_fill_level;
 970         uint64_t offloads;
 971         uint16_t buf_size;
 972         struct sfc_rxq_info *rxq_info;
 973         struct sfc_evq *evq;
 974         struct sfc_rxq *rxq;
 975         struct sfc_dp_rx_qcreate_info info;
 976         struct sfc_dp_rx_hw_limits hw_limits;
 977
 978         memset(&hw_limits, 0, sizeof(hw_limits));
 979         hw_limits.rxq_max_entries = sa->rxq_max_entries;
 980         hw_limits.rxq_min_entries = sa->rxq_min_entries;
 981         hw_limits.evq_max_entries = sa->evq_max_entries;
 982         hw_limits.evq_min_entries = sa->evq_min_entries;
 983
 984         rc = sa->priv.dp_rx->qsize_up_rings(nb_rx_desc, &hw_limits, mb_pool,
 985                                             &rxq_entries, &evq_entries,
 986                                             &rxq_max_fill_level);
 987         if (rc != 0)
 988                 goto fail_size_up_rings;
 989         SFC_ASSERT(rxq_entries >= sa->rxq_min_entries);
 990         SFC_ASSERT(rxq_entries <= sa->rxq_max_entries);
 991         SFC_ASSERT(rxq_max_fill_level <= nb_rx_desc);
 992
 993         offloads = rx_conf->offloads |
 994                 sa->eth_dev->data->dev_conf.rxmode.offloads;
 995         rc = sfc_rx_qcheck_conf(sa, rxq_max_fill_level, rx_conf, offloads);
 996         if (rc != 0)
 997                 goto fail_bad_conf;
 998
 999         buf_size = sfc_rx_mb_pool_buf_size(sa, mb_pool);
1000         if (buf_size == 0) {
1001                 sfc_err(sa, "RxQ %u mbuf pool object size is too small",
1002                         sw_index);
1003                 rc = EINVAL;
1004                 goto fail_bad_conf;
1005         }
1006
1007         if ((buf_size < sa->port.pdu + encp->enc_rx_prefix_size) &&
1008             (~offloads & DEV_RX_OFFLOAD_SCATTER)) {
1009                 sfc_err(sa, "Rx scatter is disabled and RxQ %u mbuf pool "
1010                         "object size is too small", sw_index);
1011                 sfc_err(sa, "RxQ %u calculated Rx buffer size is %u vs "
1012                         "PDU size %u plus Rx prefix %u bytes",
1013                         sw_index, buf_size, (unsigned int)sa->port.pdu,
1014                         encp->enc_rx_prefix_size);
1015                 rc = EINVAL;
1016                 goto fail_bad_conf;
1017         }
1018
1019         SFC_ASSERT(sw_index < sfc_sa2shared(sa)->rxq_count);
1020         rxq_info = &sfc_sa2shared(sa)->rxq_info[sw_index];
1021
1022         SFC_ASSERT(rxq_entries <= rxq_info->max_entries);
1023         rxq_info->entries = rxq_entries;
1024
1025         if (sa->priv.dp_rx->dp.hw_fw_caps & SFC_DP_HW_FW_CAP_RX_ES_SUPER_BUFFER)
1026                 rxq_info->type = EFX_RXQ_TYPE_ES_SUPER_BUFFER;
1027         else
1028                 rxq_info->type = EFX_RXQ_TYPE_DEFAULT;
1029
1030         rxq_info->type_flags =
1031                 (offloads & DEV_RX_OFFLOAD_SCATTER) ?
1032                 EFX_RXQ_FLAG_SCATTER : EFX_RXQ_FLAG_NONE;
1033
1034         if ((encp->enc_tunnel_encapsulations_supported != 0) &&
1035             (sa->priv.dp_rx->features & SFC_DP_RX_FEAT_TUNNELS))
1036                 rxq_info->type_flags |= EFX_RXQ_FLAG_INNER_CLASSES;
1037
1038         rc = sfc_ev_qinit(sa, SFC_EVQ_TYPE_RX, sw_index,
1039                           evq_entries, socket_id, &evq);
1040         if (rc != 0)
1041                 goto fail_ev_qinit;
1042
1043         rxq = &sa->rxq_ctrl[sw_index];
1044         rxq->evq = evq;
1045         rxq->hw_index = sw_index;
1046         rxq_info->refill_threshold =
1047                 RTE_MAX(rx_conf->rx_free_thresh, SFC_RX_REFILL_BULK);
1048         rxq_info->refill_mb_pool = mb_pool;
1049         rxq->buf_size = buf_size;
1050
1051         rc = sfc_dma_alloc(sa, "rxq", sw_index,
1052                            efx_rxq_size(sa->nic, rxq_info->entries),
1053                            socket_id, &rxq->mem);
1054         if (rc != 0)
1055                 goto fail_dma_alloc;
1056
1057         memset(&info, 0, sizeof(info));
1058         info.refill_mb_pool = rxq_info->refill_mb_pool;
1059         info.max_fill_level = rxq_max_fill_level;
1060         info.refill_threshold = rxq_info->refill_threshold;
1061         info.buf_size = buf_size;
1062         info.batch_max = encp->enc_rx_batch_max;
1063         info.prefix_size = encp->enc_rx_prefix_size;
1064
1065         if (rss->hash_support == EFX_RX_HASH_AVAILABLE && rss->channels > 0)
1066                 info.flags |= SFC_RXQ_FLAG_RSS_HASH;
1067
1068         info.rxq_entries = rxq_info->entries;
1069         info.rxq_hw_ring = rxq->mem.esm_base;
1070         info.evq_entries = evq_entries;
1071         info.evq_hw_ring = evq->mem.esm_base;
1072         info.hw_index = rxq->hw_index;
1073         info.mem_bar = sa->mem_bar.esb_base;
1074         info.vi_window_shift = encp->enc_vi_window_shift;
1075
1076         rc = sa->priv.dp_rx->qcreate(sa->eth_dev->data->port_id, sw_index,
1077                                      &RTE_ETH_DEV_TO_PCI(sa->eth_dev)->addr,
1078                                      socket_id, &info, &rxq_info->dp);
1079         if (rc != 0)
1080                 goto fail_dp_rx_qcreate;
1081
1082         evq->dp_rxq = rxq_info->dp;
1083
1084         rxq_info->state = SFC_RXQ_INITIALIZED;
1085
1086         rxq_info->deferred_start = (rx_conf->rx_deferred_start != 0);
1087
1088         return 0;
1089
1090 fail_dp_rx_qcreate:
1091         sfc_dma_free(sa, &rxq->mem);
1092
1093 fail_dma_alloc:
1094         sfc_ev_qfini(evq);
1095
1096 fail_ev_qinit:
1097         rxq_info->entries = 0;
1098
1099 fail_bad_conf:
1100 fail_size_up_rings:
1101         sfc_log_init(sa, "failed %d", rc);
1102         return rc;
1103 }
1104
1105 void
1106 sfc_rx_qfini(struct sfc_adapter *sa, unsigned int sw_index)
1107 {
1108         struct sfc_rxq_info *rxq_info;
1109         struct sfc_rxq *rxq;
1110
1111         SFC_ASSERT(sw_index < sfc_sa2shared(sa)->rxq_count);
1112         sa->eth_dev->data->rx_queues[sw_index] = NULL;
1113
1114         rxq_info = &sfc_sa2shared(sa)->rxq_info[sw_index];
1115
1116         SFC_ASSERT(rxq_info->state == SFC_RXQ_INITIALIZED);
1117
1118         sa->priv.dp_rx->qdestroy(rxq_info->dp);
1119         rxq_info->dp = NULL;
1120
1121         rxq_info->state &= ~SFC_RXQ_INITIALIZED;
1122         rxq_info->entries = 0;
1123
1124         rxq = &sa->rxq_ctrl[sw_index];
1125
1126         sfc_dma_free(sa, &rxq->mem);
1127
1128         sfc_ev_qfini(rxq->evq);
1129         rxq->evq = NULL;
1130 }
1131
1132 /*
1133  * Mapping between RTE RSS hash functions and their EFX counterparts.
1134  */
1135 static const struct sfc_rss_hf_rte_to_efx sfc_rss_hf_map[] = {
1136         { ETH_RSS_NONFRAG_IPV4_TCP,
1137           EFX_RX_HASH(IPV4_TCP, 4TUPLE) },
1138         { ETH_RSS_NONFRAG_IPV4_UDP,
1139           EFX_RX_HASH(IPV4_UDP, 4TUPLE) },
1140         { ETH_RSS_NONFRAG_IPV6_TCP | ETH_RSS_IPV6_TCP_EX,
1141           EFX_RX_HASH(IPV6_TCP, 4TUPLE) },
1142         { ETH_RSS_NONFRAG_IPV6_UDP | ETH_RSS_IPV6_UDP_EX,
1143           EFX_RX_HASH(IPV6_UDP, 4TUPLE) },
1144         { ETH_RSS_IPV4 | ETH_RSS_FRAG_IPV4 | ETH_RSS_NONFRAG_IPV4_OTHER,
1145           EFX_RX_HASH(IPV4_TCP, 2TUPLE) | EFX_RX_HASH(IPV4_UDP, 2TUPLE) |
1146           EFX_RX_HASH(IPV4, 2TUPLE) },
1147         { ETH_RSS_IPV6 | ETH_RSS_FRAG_IPV6 | ETH_RSS_NONFRAG_IPV6_OTHER |
1148           ETH_RSS_IPV6_EX,
1149           EFX_RX_HASH(IPV6_TCP, 2TUPLE) | EFX_RX_HASH(IPV6_UDP, 2TUPLE) |
1150           EFX_RX_HASH(IPV6, 2TUPLE) }
1151 };
1152
1153 static efx_rx_hash_type_t
1154 sfc_rx_hash_types_mask_supp(efx_rx_hash_type_t hash_type,
1155                             unsigned int *hash_type_flags_supported,
1156                             unsigned int nb_hash_type_flags_supported)
1157 {
1158         efx_rx_hash_type_t hash_type_masked = 0;
1159         unsigned int i, j;
1160
1161         for (i = 0; i < nb_hash_type_flags_supported; ++i) {
1162                 unsigned int class_tuple_lbn[] = {
1163                         EFX_RX_CLASS_IPV4_TCP_LBN,
1164                         EFX_RX_CLASS_IPV4_UDP_LBN,
1165                         EFX_RX_CLASS_IPV4_LBN,
1166                         EFX_RX_CLASS_IPV6_TCP_LBN,
1167                         EFX_RX_CLASS_IPV6_UDP_LBN,
1168                         EFX_RX_CLASS_IPV6_LBN
1169                 };
1170
1171                 for (j = 0; j < RTE_DIM(class_tuple_lbn); ++j) {
1172                         unsigned int tuple_mask = EFX_RX_CLASS_HASH_4TUPLE;
1173                         unsigned int flag;
1174
1175                         tuple_mask <<= class_tuple_lbn[j];
1176                         flag = hash_type & tuple_mask;
1177
1178                         if (flag == hash_type_flags_supported[i])
1179                                 hash_type_masked |= flag;
1180                 }
1181         }
1182
1183         return hash_type_masked;
1184 }
1185
1186 int
1187 sfc_rx_hash_init(struct sfc_adapter *sa)
1188 {
1189         struct sfc_rss *rss = &sfc_sa2shared(sa)->rss;
1190         const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
1191         uint32_t alg_mask = encp->enc_rx_scale_hash_alg_mask;
1192         efx_rx_hash_alg_t alg;
1193         unsigned int flags_supp[EFX_RX_HASH_NFLAGS];
1194         unsigned int nb_flags_supp;
1195         struct sfc_rss_hf_rte_to_efx *hf_map;
1196         struct sfc_rss_hf_rte_to_efx *entry;
1197         efx_rx_hash_type_t efx_hash_types;
1198         unsigned int i;
1199         int rc;
1200
1201         if (alg_mask & (1U << EFX_RX_HASHALG_TOEPLITZ))
1202                 alg = EFX_RX_HASHALG_TOEPLITZ;
1203         else if (alg_mask & (1U << EFX_RX_HASHALG_PACKED_STREAM))
1204                 alg = EFX_RX_HASHALG_PACKED_STREAM;
1205         else
1206                 return EINVAL;
1207
1208         rc = efx_rx_scale_hash_flags_get(sa->nic, alg, flags_supp,
1209                                          RTE_DIM(flags_supp), &nb_flags_supp);
1210         if (rc != 0)
1211                 return rc;
1212
1213         hf_map = rte_calloc_socket("sfc-rss-hf-map",
1214                                    RTE_DIM(sfc_rss_hf_map),
1215                                    sizeof(*hf_map), 0, sa->socket_id);
1216         if (hf_map == NULL)
1217                 return ENOMEM;
1218
1219         entry = hf_map;
1220         efx_hash_types = 0;
1221         for (i = 0; i < RTE_DIM(sfc_rss_hf_map); ++i) {
1222                 efx_rx_hash_type_t ht;
1223
1224                 ht = sfc_rx_hash_types_mask_supp(sfc_rss_hf_map[i].efx,
1225                                                  flags_supp, nb_flags_supp);
1226                 if (ht != 0) {
1227                         entry->rte = sfc_rss_hf_map[i].rte;
1228                         entry->efx = ht;
1229                         efx_hash_types |= ht;
1230                         ++entry;
1231                 }
1232         }
1233
1234         rss->hash_alg = alg;
1235         rss->hf_map_nb_entries = (unsigned int)(entry - hf_map);
1236         rss->hf_map = hf_map;
1237         rss->hash_types = efx_hash_types;
1238
1239         return 0;
1240 }
1241
1242 void
1243 sfc_rx_hash_fini(struct sfc_adapter *sa)
1244 {
1245         struct sfc_rss *rss = &sfc_sa2shared(sa)->rss;
1246
1247         rte_free(rss->hf_map);
1248 }
1249
1250 int
1251 sfc_rx_hf_rte_to_efx(struct sfc_adapter *sa, uint64_t rte,
1252                      efx_rx_hash_type_t *efx)
1253 {
1254         struct sfc_rss *rss = &sfc_sa2shared(sa)->rss;
1255         efx_rx_hash_type_t hash_types = 0;
1256         unsigned int i;
1257
1258         for (i = 0; i < rss->hf_map_nb_entries; ++i) {
1259                 uint64_t rte_mask = rss->hf_map[i].rte;
1260
1261                 if ((rte & rte_mask) != 0) {
1262                         rte &= ~rte_mask;
1263                         hash_types |= rss->hf_map[i].efx;
1264                 }
1265         }
1266
1267         if (rte != 0) {
1268                 sfc_err(sa, "unsupported hash functions requested");
1269                 return EINVAL;
1270         }
1271
1272         *efx = hash_types;
1273
1274         return 0;
1275 }
1276
1277 uint64_t
1278 sfc_rx_hf_efx_to_rte(struct sfc_rss *rss, efx_rx_hash_type_t efx)
1279 {
1280         uint64_t rte = 0;
1281         unsigned int i;
1282
1283         for (i = 0; i < rss->hf_map_nb_entries; ++i) {
1284                 efx_rx_hash_type_t hash_type = rss->hf_map[i].efx;
1285
1286                 if ((efx & hash_type) == hash_type)
1287                         rte |= rss->hf_map[i].rte;
1288         }
1289
1290         return rte;
1291 }
1292
1293 static int
1294 sfc_rx_process_adv_conf_rss(struct sfc_adapter *sa,
1295                             struct rte_eth_rss_conf *conf)
1296 {
1297         struct sfc_rss *rss = &sfc_sa2shared(sa)->rss;
1298         efx_rx_hash_type_t efx_hash_types = rss->hash_types;
1299         uint64_t rss_hf = sfc_rx_hf_efx_to_rte(rss, efx_hash_types);
1300         int rc;
1301
1302         if (rss->context_type != EFX_RX_SCALE_EXCLUSIVE) {
1303                 if ((conf->rss_hf != 0 && conf->rss_hf != rss_hf) ||
1304                     conf->rss_key != NULL)
1305                         return EINVAL;
1306         }
1307
1308         if (conf->rss_hf != 0) {
1309                 rc = sfc_rx_hf_rte_to_efx(sa, conf->rss_hf, &efx_hash_types);
1310                 if (rc != 0)
1311                         return rc;
1312         }
1313
1314         if (conf->rss_key != NULL) {
1315                 if (conf->rss_key_len != sizeof(rss->key)) {
1316                         sfc_err(sa, "RSS key size is wrong (should be %lu)",
1317                                 sizeof(rss->key));
1318                         return EINVAL;
1319                 }
1320                 rte_memcpy(rss->key, conf->rss_key, sizeof(rss->key));
1321         }
1322
1323         rss->hash_types = efx_hash_types;
1324
1325         return 0;
1326 }
1327
1328 static int
1329 sfc_rx_rss_config(struct sfc_adapter *sa)
1330 {
1331         struct sfc_rss *rss = &sfc_sa2shared(sa)->rss;
1332         int rc = 0;
1333
1334         if (rss->channels > 0) {
1335                 rc = efx_rx_scale_mode_set(sa->nic, EFX_RSS_CONTEXT_DEFAULT,
1336                                            rss->hash_alg, rss->hash_types,
1337                                            B_TRUE);
1338                 if (rc != 0)
1339                         goto finish;
1340
1341                 rc = efx_rx_scale_key_set(sa->nic, EFX_RSS_CONTEXT_DEFAULT,
1342                                           rss->key, sizeof(rss->key));
1343                 if (rc != 0)
1344                         goto finish;
1345
1346                 rc = efx_rx_scale_tbl_set(sa->nic, EFX_RSS_CONTEXT_DEFAULT,
1347                                           rss->tbl, RTE_DIM(rss->tbl));
1348         }
1349
1350 finish:
1351         return rc;
1352 }
1353
1354 int
1355 sfc_rx_start(struct sfc_adapter *sa)
1356 {
1357         struct sfc_adapter_shared * const sas = sfc_sa2shared(sa);
1358         unsigned int sw_index;
1359         int rc;
1360
1361         sfc_log_init(sa, "rxq_count=%u", sas->rxq_count);
1362
1363         rc = efx_rx_init(sa->nic);
1364         if (rc != 0)
1365                 goto fail_rx_init;
1366
1367         rc = sfc_rx_rss_config(sa);
1368         if (rc != 0)
1369                 goto fail_rss_config;
1370
1371         for (sw_index = 0; sw_index < sas->rxq_count; ++sw_index) {
1372                 if (sas->rxq_info[sw_index].state == SFC_RXQ_INITIALIZED &&
1373                     (!sas->rxq_info[sw_index].deferred_start ||
1374                      sas->rxq_info[sw_index].deferred_started)) {
1375                         rc = sfc_rx_qstart(sa, sw_index);
1376                         if (rc != 0)
1377                                 goto fail_rx_qstart;
1378                 }
1379         }
1380
1381         return 0;
1382
1383 fail_rx_qstart:
1384         while (sw_index-- > 0)
1385                 sfc_rx_qstop(sa, sw_index);
1386
1387 fail_rss_config:
1388         efx_rx_fini(sa->nic);
1389
1390 fail_rx_init:
1391         sfc_log_init(sa, "failed %d", rc);
1392         return rc;
1393 }
1394
1395 void
1396 sfc_rx_stop(struct sfc_adapter *sa)
1397 {
1398         struct sfc_adapter_shared * const sas = sfc_sa2shared(sa);
1399         unsigned int sw_index;
1400
1401         sfc_log_init(sa, "rxq_count=%u", sas->rxq_count);
1402
1403         sw_index = sas->rxq_count;
1404         while (sw_index-- > 0) {
1405                 if (sas->rxq_info[sw_index].state & SFC_RXQ_STARTED)
1406                         sfc_rx_qstop(sa, sw_index);
1407         }
1408
1409         efx_rx_fini(sa->nic);
1410 }
1411
1412 static int
1413 sfc_rx_qinit_info(struct sfc_adapter *sa, unsigned int sw_index)
1414 {
1415         struct sfc_adapter_shared * const sas = sfc_sa2shared(sa);
1416         struct sfc_rxq_info *rxq_info = &sas->rxq_info[sw_index];
1417         const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
1418         unsigned int max_entries;
1419
1420         max_entries = encp->enc_rxq_max_ndescs;
1421         SFC_ASSERT(rte_is_power_of_2(max_entries));
1422
1423         rxq_info->max_entries = max_entries;
1424
1425         return 0;
1426 }
1427
1428 static int
1429 sfc_rx_check_mode(struct sfc_adapter *sa, struct rte_eth_rxmode *rxmode)
1430 {
1431         struct sfc_adapter_shared * const sas = sfc_sa2shared(sa);
1432         uint64_t offloads_supported = sfc_rx_get_dev_offload_caps(sa) |
1433                                       sfc_rx_get_queue_offload_caps(sa);
1434         struct sfc_rss *rss = &sas->rss;
1435         int rc = 0;
1436
1437         switch (rxmode->mq_mode) {
1438         case ETH_MQ_RX_NONE:
1439                 /* No special checks are required */
1440                 break;
1441         case ETH_MQ_RX_RSS:
1442                 if (rss->context_type == EFX_RX_SCALE_UNAVAILABLE) {
1443                         sfc_err(sa, "RSS is not available");
1444                         rc = EINVAL;
1445                 }
1446                 break;
1447         default:
1448                 sfc_err(sa, "Rx multi-queue mode %u not supported",
1449                         rxmode->mq_mode);
1450                 rc = EINVAL;
1451         }
1452
1453         /*
1454          * Requested offloads are validated against supported by ethdev,
1455          * so unsupported offloads cannot be added as the result of
1456          * below check.
1457          */
1458         if ((rxmode->offloads & DEV_RX_OFFLOAD_CHECKSUM) !=
1459             (offloads_supported & DEV_RX_OFFLOAD_CHECKSUM)) {
1460                 sfc_warn(sa, "Rx checksum offloads cannot be disabled - always on (IPv4/TCP/UDP)");
1461                 rxmode->offloads |= DEV_RX_OFFLOAD_CHECKSUM;
1462         }
1463
1464         if ((offloads_supported & DEV_RX_OFFLOAD_OUTER_IPV4_CKSUM) &&
1465             (~rxmode->offloads & DEV_RX_OFFLOAD_OUTER_IPV4_CKSUM)) {
1466                 sfc_warn(sa, "Rx outer IPv4 checksum offload cannot be disabled - always on");
1467                 rxmode->offloads |= DEV_RX_OFFLOAD_OUTER_IPV4_CKSUM;
1468         }
1469
1470         return rc;
1471 }
1472
1473 /**
1474  * Destroy excess queues that are no longer needed after reconfiguration
1475  * or complete close.
1476  */
1477 static void
1478 sfc_rx_fini_queues(struct sfc_adapter *sa, unsigned int nb_rx_queues)
1479 {
1480         struct sfc_adapter_shared * const sas = sfc_sa2shared(sa);
1481         int sw_index;
1482
1483         SFC_ASSERT(nb_rx_queues <= sas->rxq_count);
1484
1485         sw_index = sas->rxq_count;
1486         while (--sw_index >= (int)nb_rx_queues) {
1487                 if (sas->rxq_info[sw_index].state & SFC_RXQ_INITIALIZED)
1488                         sfc_rx_qfini(sa, sw_index);
1489         }
1490
1491         sas->rxq_count = nb_rx_queues;
1492 }
1493
1494 /**
1495  * Initialize Rx subsystem.
1496  *
1497  * Called at device (re)configuration stage when number of receive queues is
1498  * specified together with other device level receive configuration.
1499  *
1500  * It should be used to allocate NUMA-unaware resources.
1501  */
1502 int
1503 sfc_rx_configure(struct sfc_adapter *sa)
1504 {
1505         struct sfc_adapter_shared * const sas = sfc_sa2shared(sa);
1506         struct sfc_rss *rss = &sas->rss;
1507         struct rte_eth_conf *dev_conf = &sa->eth_dev->data->dev_conf;
1508         const unsigned int nb_rx_queues = sa->eth_dev->data->nb_rx_queues;
1509         int rc;
1510
1511         sfc_log_init(sa, "nb_rx_queues=%u (old %u)",
1512                      nb_rx_queues, sas->rxq_count);
1513
1514         rc = sfc_rx_check_mode(sa, &dev_conf->rxmode);
1515         if (rc != 0)
1516                 goto fail_check_mode;
1517
1518         if (nb_rx_queues == sas->rxq_count)
1519                 goto configure_rss;
1520
1521         if (sas->rxq_info == NULL) {
1522                 rc = ENOMEM;
1523                 sas->rxq_info = rte_calloc_socket("sfc-rxqs", nb_rx_queues,
1524                                                   sizeof(sas->rxq_info[0]), 0,
1525                                                   sa->socket_id);
1526                 if (sas->rxq_info == NULL)
1527                         goto fail_rxqs_alloc;
1528
1529                 /*
1530                  * Allocate primary process only RxQ control from heap
1531                  * since it should not be shared.
1532                  */
1533                 rc = ENOMEM;
1534                 sa->rxq_ctrl = calloc(nb_rx_queues, sizeof(sa->rxq_ctrl[0]));
1535                 if (sa->rxq_ctrl == NULL)
1536                         goto fail_rxqs_ctrl_alloc;
1537         } else {
1538                 struct sfc_rxq_info *new_rxq_info;
1539                 struct sfc_rxq *new_rxq_ctrl;
1540
1541                 if (nb_rx_queues < sas->rxq_count)
1542                         sfc_rx_fini_queues(sa, nb_rx_queues);
1543
1544                 rc = ENOMEM;
1545                 new_rxq_info =
1546                         rte_realloc(sas->rxq_info,
1547                                     nb_rx_queues * sizeof(sas->rxq_info[0]), 0);
1548                 if (new_rxq_info == NULL && nb_rx_queues > 0)
1549                         goto fail_rxqs_realloc;
1550
1551                 rc = ENOMEM;
1552                 new_rxq_ctrl = realloc(sa->rxq_ctrl,
1553                                        nb_rx_queues * sizeof(sa->rxq_ctrl[0]));
1554                 if (new_rxq_ctrl == NULL && nb_rx_queues > 0)
1555                         goto fail_rxqs_ctrl_realloc;
1556
1557                 sas->rxq_info = new_rxq_info;
1558                 sa->rxq_ctrl = new_rxq_ctrl;
1559                 if (nb_rx_queues > sas->rxq_count) {
1560                         memset(&sas->rxq_info[sas->rxq_count], 0,
1561                                (nb_rx_queues - sas->rxq_count) *
1562                                sizeof(sas->rxq_info[0]));
1563                         memset(&sa->rxq_ctrl[sas->rxq_count], 0,
1564                                (nb_rx_queues - sas->rxq_count) *
1565                                sizeof(sa->rxq_ctrl[0]));
1566                 }
1567         }
1568
1569         while (sas->rxq_count < nb_rx_queues) {
1570                 rc = sfc_rx_qinit_info(sa, sas->rxq_count);
1571                 if (rc != 0)
1572                         goto fail_rx_qinit_info;
1573
1574                 sas->rxq_count++;
1575         }
1576
1577 configure_rss:
1578         rss->channels = (dev_conf->rxmode.mq_mode == ETH_MQ_RX_RSS) ?
1579                          MIN(sas->rxq_count, EFX_MAXRSS) : 0;
1580
1581         if (rss->channels > 0) {
1582                 struct rte_eth_rss_conf *adv_conf_rss;
1583                 unsigned int sw_index;
1584
1585                 for (sw_index = 0; sw_index < EFX_RSS_TBL_SIZE; ++sw_index)
1586                         rss->tbl[sw_index] = sw_index % rss->channels;
1587
1588                 adv_conf_rss = &dev_conf->rx_adv_conf.rss_conf;
1589                 rc = sfc_rx_process_adv_conf_rss(sa, adv_conf_rss);
1590                 if (rc != 0)
1591                         goto fail_rx_process_adv_conf_rss;
1592         }
1593
1594         return 0;
1595
1596 fail_rx_process_adv_conf_rss:
1597 fail_rx_qinit_info:
1598 fail_rxqs_ctrl_realloc:
1599 fail_rxqs_realloc:
1600 fail_rxqs_ctrl_alloc:
1601 fail_rxqs_alloc:
1602         sfc_rx_close(sa);
1603
1604 fail_check_mode:
1605         sfc_log_init(sa, "failed %d", rc);
1606         return rc;
1607 }
1608
1609 /**
1610  * Shutdown Rx subsystem.
1611  *
1612  * Called at device close stage, for example, before device shutdown.
1613  */
1614 void
1615 sfc_rx_close(struct sfc_adapter *sa)
1616 {
1617         struct sfc_rss *rss = &sfc_sa2shared(sa)->rss;
1618
1619         sfc_rx_fini_queues(sa, 0);
1620
1621         rss->channels = 0;
1622
1623         free(sa->rxq_ctrl);
1624         sa->rxq_ctrl = NULL;
1625
1626         rte_free(sfc_sa2shared(sa)->rxq_info);
1627         sfc_sa2shared(sa)->rxq_info = NULL;
1628 }