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 *rxq)
  43 {
  44         rxq->state |= SFC_RXQ_FLUSHED;
  45         rxq->state &= ~SFC_RXQ_FLUSHING;
  46 }
  47
  48 void
  49 sfc_rx_qflush_failed(struct sfc_rxq *rxq)
  50 {
  51         rxq->state |= SFC_RXQ_FLUSH_FAILED;
  52         rxq->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 struct sfc_rxq *
 364 sfc_rxq_by_dp_rxq(const struct sfc_dp_rxq *dp_rxq)
 365 {
 366         const struct sfc_dp_queue *dpq = &dp_rxq->dpq;
 367         struct rte_eth_dev *eth_dev;
 368         struct sfc_adapter *sa;
 369         struct sfc_rxq *rxq;
 370
 371         SFC_ASSERT(rte_eth_dev_is_valid_port(dpq->port_id));
 372         eth_dev = &rte_eth_devices[dpq->port_id];
 373
 374         sa = eth_dev->data->dev_private;
 375
 376         SFC_ASSERT(dpq->queue_id < sa->rxq_count);
 377         rxq = sa->rxq_info[dpq->queue_id].rxq;
 378
 379         SFC_ASSERT(rxq != NULL);
 380         return rxq;
 381 }
 382
 383 static sfc_dp_rx_qsize_up_rings_t sfc_efx_rx_qsize_up_rings;
 384 static int
 385 sfc_efx_rx_qsize_up_rings(uint16_t nb_rx_desc,
 386                           __rte_unused struct rte_mempool *mb_pool,
 387                           unsigned int *rxq_entries,
 388                           unsigned int *evq_entries,
 389                           unsigned int *rxq_max_fill_level)
 390 {
 391         *rxq_entries = nb_rx_desc;
 392         *evq_entries = nb_rx_desc;
 393         *rxq_max_fill_level = EFX_RXQ_LIMIT(*rxq_entries);
 394         return 0;
 395 }
 396
 397 static sfc_dp_rx_qcreate_t sfc_efx_rx_qcreate;
 398 static int
 399 sfc_efx_rx_qcreate(uint16_t port_id, uint16_t queue_id,
 400                    const struct rte_pci_addr *pci_addr, int socket_id,
 401                    const struct sfc_dp_rx_qcreate_info *info,
 402                    struct sfc_dp_rxq **dp_rxqp)
 403 {
 404         struct sfc_efx_rxq *rxq;
 405         int rc;
 406
 407         rc = ENOMEM;
 408         rxq = rte_zmalloc_socket("sfc-efx-rxq", sizeof(*rxq),
 409                                  RTE_CACHE_LINE_SIZE, socket_id);
 410         if (rxq == NULL)
 411                 goto fail_rxq_alloc;
 412
 413         sfc_dp_queue_init(&rxq->dp.dpq, port_id, queue_id, pci_addr);
 414
 415         rc = ENOMEM;
 416         rxq->sw_desc = rte_calloc_socket("sfc-efx-rxq-sw_desc",
 417                                          info->rxq_entries,
 418                                          sizeof(*rxq->sw_desc),
 419                                          RTE_CACHE_LINE_SIZE, socket_id);
 420         if (rxq->sw_desc == NULL)
 421                 goto fail_desc_alloc;
 422
 423         /* efx datapath is bound to efx control path */
 424         rxq->evq = sfc_rxq_by_dp_rxq(&rxq->dp)->evq;
 425         if (info->flags & SFC_RXQ_FLAG_RSS_HASH)
 426                 rxq->flags |= SFC_EFX_RXQ_FLAG_RSS_HASH;
 427         rxq->ptr_mask = info->rxq_entries - 1;
 428         rxq->batch_max = info->batch_max;
 429         rxq->prefix_size = info->prefix_size;
 430         rxq->max_fill_level = info->max_fill_level;
 431         rxq->refill_threshold = info->refill_threshold;
 432         rxq->buf_size = info->buf_size;
 433         rxq->refill_mb_pool = info->refill_mb_pool;
 434
 435         *dp_rxqp = &rxq->dp;
 436         return 0;
 437
 438 fail_desc_alloc:
 439         rte_free(rxq);
 440
 441 fail_rxq_alloc:
 442         return rc;
 443 }
 444
 445 static sfc_dp_rx_qdestroy_t sfc_efx_rx_qdestroy;
 446 static void
 447 sfc_efx_rx_qdestroy(struct sfc_dp_rxq *dp_rxq)
 448 {
 449         struct sfc_efx_rxq *rxq = sfc_efx_rxq_by_dp_rxq(dp_rxq);
 450
 451         rte_free(rxq->sw_desc);
 452         rte_free(rxq);
 453 }
 454
 455 static sfc_dp_rx_qstart_t sfc_efx_rx_qstart;
 456 static int
 457 sfc_efx_rx_qstart(struct sfc_dp_rxq *dp_rxq,
 458                   __rte_unused unsigned int evq_read_ptr)
 459 {
 460         /* libefx-based datapath is specific to libefx-based PMD */
 461         struct sfc_efx_rxq *rxq = sfc_efx_rxq_by_dp_rxq(dp_rxq);
 462         struct sfc_rxq *crxq = sfc_rxq_by_dp_rxq(dp_rxq);
 463
 464         rxq->common = crxq->common;
 465
 466         rxq->pending = rxq->completed = rxq->added = rxq->pushed = 0;
 467
 468         sfc_efx_rx_qrefill(rxq);
 469
 470         rxq->flags |= (SFC_EFX_RXQ_FLAG_STARTED | SFC_EFX_RXQ_FLAG_RUNNING);
 471
 472         return 0;
 473 }
 474
 475 static sfc_dp_rx_qstop_t sfc_efx_rx_qstop;
 476 static void
 477 sfc_efx_rx_qstop(struct sfc_dp_rxq *dp_rxq,
 478                  __rte_unused unsigned int *evq_read_ptr)
 479 {
 480         struct sfc_efx_rxq *rxq = sfc_efx_rxq_by_dp_rxq(dp_rxq);
 481
 482         rxq->flags &= ~SFC_EFX_RXQ_FLAG_RUNNING;
 483
 484         /* libefx-based datapath is bound to libefx-based PMD and uses
 485          * event queue structure directly. So, there is no necessity to
 486          * return EvQ read pointer.
 487          */
 488 }
 489
 490 static sfc_dp_rx_qpurge_t sfc_efx_rx_qpurge;
 491 static void
 492 sfc_efx_rx_qpurge(struct sfc_dp_rxq *dp_rxq)
 493 {
 494         struct sfc_efx_rxq *rxq = sfc_efx_rxq_by_dp_rxq(dp_rxq);
 495         unsigned int i;
 496         struct sfc_efx_rx_sw_desc *rxd;
 497
 498         for (i = rxq->completed; i != rxq->added; ++i) {
 499                 rxd = &rxq->sw_desc[i & rxq->ptr_mask];
 500                 rte_mbuf_raw_free(rxd->mbuf);
 501                 rxd->mbuf = NULL;
 502                 /* Packed stream relies on 0 in inactive SW desc.
 503                  * Rx queue stop is not performance critical, so
 504                  * there is no harm to do it always.
 505                  */
 506                 rxd->flags = 0;
 507                 rxd->size = 0;
 508         }
 509
 510         rxq->flags &= ~SFC_EFX_RXQ_FLAG_STARTED;
 511 }
 512
 513 struct sfc_dp_rx sfc_efx_rx = {
 514         .dp = {
 515                 .name           = SFC_KVARG_DATAPATH_EFX,
 516                 .type           = SFC_DP_RX,
 517                 .hw_fw_caps     = 0,
 518         },
 519         .features               = SFC_DP_RX_FEAT_SCATTER |
 520                                   SFC_DP_RX_FEAT_CHECKSUM,
 521         .qsize_up_rings         = sfc_efx_rx_qsize_up_rings,
 522         .qcreate                = sfc_efx_rx_qcreate,
 523         .qdestroy               = sfc_efx_rx_qdestroy,
 524         .qstart                 = sfc_efx_rx_qstart,
 525         .qstop                  = sfc_efx_rx_qstop,
 526         .qpurge                 = sfc_efx_rx_qpurge,
 527         .supported_ptypes_get   = sfc_efx_supported_ptypes_get,
 528         .qdesc_npending         = sfc_efx_rx_qdesc_npending,
 529         .qdesc_status           = sfc_efx_rx_qdesc_status,
 530         .pkt_burst              = sfc_efx_recv_pkts,
 531 };
 532
 533 static void
 534 sfc_rx_qflush(struct sfc_adapter *sa, unsigned int sw_index)
 535 {
 536         struct sfc_rxq *rxq;
 537         unsigned int retry_count;
 538         unsigned int wait_count;
 539         int rc;
 540
 541         rxq = sa->rxq_info[sw_index].rxq;
 542         SFC_ASSERT(rxq->state & SFC_RXQ_STARTED);
 543
 544         /*
 545          * Retry Rx queue flushing in the case of flush failed or
 546          * timeout. In the worst case it can delay for 6 seconds.
 547          */
 548         for (retry_count = 0;
 549              ((rxq->state & SFC_RXQ_FLUSHED) == 0) &&
 550              (retry_count < SFC_RX_QFLUSH_ATTEMPTS);
 551              ++retry_count) {
 552                 rc = efx_rx_qflush(rxq->common);
 553                 if (rc != 0) {
 554                         rxq->state |= (rc == EALREADY) ?
 555                                 SFC_RXQ_FLUSHED : SFC_RXQ_FLUSH_FAILED;
 556                         break;
 557                 }
 558                 rxq->state &= ~SFC_RXQ_FLUSH_FAILED;
 559                 rxq->state |= SFC_RXQ_FLUSHING;
 560
 561                 /*
 562                  * Wait for Rx queue flush done or failed event at least
 563                  * SFC_RX_QFLUSH_POLL_WAIT_MS milliseconds and not more
 564                  * than 2 seconds (SFC_RX_QFLUSH_POLL_WAIT_MS multiplied
 565                  * by SFC_RX_QFLUSH_POLL_ATTEMPTS).
 566                  */
 567                 wait_count = 0;
 568                 do {
 569                         rte_delay_ms(SFC_RX_QFLUSH_POLL_WAIT_MS);
 570                         sfc_ev_qpoll(rxq->evq);
 571                 } while ((rxq->state & SFC_RXQ_FLUSHING) &&
 572                          (wait_count++ < SFC_RX_QFLUSH_POLL_ATTEMPTS));
 573
 574                 if (rxq->state & SFC_RXQ_FLUSHING)
 575                         sfc_err(sa, "RxQ %u flush timed out", sw_index);
 576
 577                 if (rxq->state & SFC_RXQ_FLUSH_FAILED)
 578                         sfc_err(sa, "RxQ %u flush failed", sw_index);
 579
 580                 if (rxq->state & SFC_RXQ_FLUSHED)
 581                         sfc_notice(sa, "RxQ %u flushed", sw_index);
 582         }
 583
 584         sa->priv.dp_rx->qpurge(rxq->dp);
 585 }
 586
 587 static int
 588 sfc_rx_default_rxq_set_filter(struct sfc_adapter *sa, struct sfc_rxq *rxq)
 589 {
 590         struct sfc_rss *rss = &sa->rss;
 591         boolean_t need_rss = (rss->channels > 0) ? B_TRUE : B_FALSE;
 592         struct sfc_port *port = &sa->port;
 593         int rc;
 594
 595         /*
 596          * If promiscuous or all-multicast mode has been requested, setting
 597          * filter for the default Rx queue might fail, in particular, while
 598          * running over PCI function which is not a member of corresponding
 599          * privilege groups; if this occurs, few iterations will be made to
 600          * repeat this step without promiscuous and all-multicast flags set
 601          */
 602 retry:
 603         rc = efx_mac_filter_default_rxq_set(sa->nic, rxq->common, need_rss);
 604         if (rc == 0)
 605                 return 0;
 606         else if (rc != EOPNOTSUPP)
 607                 return rc;
 608
 609         if (port->promisc) {
 610                 sfc_warn(sa, "promiscuous mode has been requested, "
 611                              "but the HW rejects it");
 612                 sfc_warn(sa, "promiscuous mode will be disabled");
 613
 614                 port->promisc = B_FALSE;
 615                 rc = sfc_set_rx_mode(sa);
 616                 if (rc != 0)
 617                         return rc;
 618
 619                 goto retry;
 620         }
 621
 622         if (port->allmulti) {
 623                 sfc_warn(sa, "all-multicast mode has been requested, "
 624                              "but the HW rejects it");
 625                 sfc_warn(sa, "all-multicast mode will be disabled");
 626
 627                 port->allmulti = B_FALSE;
 628                 rc = sfc_set_rx_mode(sa);
 629                 if (rc != 0)
 630                         return rc;
 631
 632                 goto retry;
 633         }
 634
 635         return rc;
 636 }
 637
 638 int
 639 sfc_rx_qstart(struct sfc_adapter *sa, unsigned int sw_index)
 640 {
 641         struct sfc_port *port = &sa->port;
 642         struct sfc_rxq_info *rxq_info;
 643         struct sfc_rxq *rxq;
 644         struct sfc_evq *evq;
 645         int rc;
 646
 647         sfc_log_init(sa, "sw_index=%u", sw_index);
 648
 649         SFC_ASSERT(sw_index < sa->rxq_count);
 650
 651         rxq_info = &sa->rxq_info[sw_index];
 652         rxq = rxq_info->rxq;
 653         SFC_ASSERT(rxq != NULL);
 654         SFC_ASSERT(rxq->state == SFC_RXQ_INITIALIZED);
 655
 656         evq = rxq->evq;
 657
 658         rc = sfc_ev_qstart(evq, sfc_evq_index_by_rxq_sw_index(sa, sw_index));
 659         if (rc != 0)
 660                 goto fail_ev_qstart;
 661
 662         switch (rxq_info->type) {
 663         case EFX_RXQ_TYPE_DEFAULT:
 664                 rc = efx_rx_qcreate(sa->nic, rxq->hw_index, 0, rxq_info->type,
 665                         &rxq->mem, rxq_info->entries, 0 /* not used on EF10 */,
 666                         rxq_info->type_flags, evq->common, &rxq->common);
 667                 break;
 668         case EFX_RXQ_TYPE_ES_SUPER_BUFFER: {
 669                 struct rte_mempool *mp = rxq_info->refill_mb_pool;
 670                 struct rte_mempool_info mp_info;
 671
 672                 rc = rte_mempool_ops_get_info(mp, &mp_info);
 673                 if (rc != 0) {
 674                         /* Positive errno is used in the driver */
 675                         rc = -rc;
 676                         goto fail_mp_get_info;
 677                 }
 678                 if (mp_info.contig_block_size <= 0) {
 679                         rc = EINVAL;
 680                         goto fail_bad_contig_block_size;
 681                 }
 682                 rc = efx_rx_qcreate_es_super_buffer(sa->nic, rxq->hw_index, 0,
 683                         mp_info.contig_block_size, rxq->buf_size,
 684                         mp->header_size + mp->elt_size + mp->trailer_size,
 685                         sa->rxd_wait_timeout_ns,
 686                         &rxq->mem, rxq_info->entries, rxq_info->type_flags,
 687                         evq->common, &rxq->common);
 688                 break;
 689         }
 690         default:
 691                 rc = ENOTSUP;
 692         }
 693         if (rc != 0)
 694                 goto fail_rx_qcreate;
 695
 696         efx_rx_qenable(rxq->common);
 697
 698         rc = sa->priv.dp_rx->qstart(rxq->dp, evq->read_ptr);
 699         if (rc != 0)
 700                 goto fail_dp_qstart;
 701
 702         rxq->state |= SFC_RXQ_STARTED;
 703
 704         if ((sw_index == 0) && !port->isolated) {
 705                 rc = sfc_rx_default_rxq_set_filter(sa, rxq);
 706                 if (rc != 0)
 707                         goto fail_mac_filter_default_rxq_set;
 708         }
 709
 710         /* It seems to be used by DPDK for debug purposes only ('rte_ether') */
 711         sa->eth_dev->data->rx_queue_state[sw_index] =
 712                 RTE_ETH_QUEUE_STATE_STARTED;
 713
 714         return 0;
 715
 716 fail_mac_filter_default_rxq_set:
 717         sa->priv.dp_rx->qstop(rxq->dp, &rxq->evq->read_ptr);
 718
 719 fail_dp_qstart:
 720         sfc_rx_qflush(sa, sw_index);
 721
 722 fail_rx_qcreate:
 723 fail_bad_contig_block_size:
 724 fail_mp_get_info:
 725         sfc_ev_qstop(evq);
 726
 727 fail_ev_qstart:
 728         return rc;
 729 }
 730
 731 void
 732 sfc_rx_qstop(struct sfc_adapter *sa, unsigned int sw_index)
 733 {
 734         struct sfc_rxq_info *rxq_info;
 735         struct sfc_rxq *rxq;
 736
 737         sfc_log_init(sa, "sw_index=%u", sw_index);
 738
 739         SFC_ASSERT(sw_index < sa->rxq_count);
 740
 741         rxq_info = &sa->rxq_info[sw_index];
 742         rxq = rxq_info->rxq;
 743
 744         if (rxq == NULL || rxq->state == SFC_RXQ_INITIALIZED)
 745                 return;
 746         SFC_ASSERT(rxq->state & SFC_RXQ_STARTED);
 747
 748         /* It seems to be used by DPDK for debug purposes only ('rte_ether') */
 749         sa->eth_dev->data->rx_queue_state[sw_index] =
 750                 RTE_ETH_QUEUE_STATE_STOPPED;
 751
 752         sa->priv.dp_rx->qstop(rxq->dp, &rxq->evq->read_ptr);
 753
 754         if (sw_index == 0)
 755                 efx_mac_filter_default_rxq_clear(sa->nic);
 756
 757         sfc_rx_qflush(sa, sw_index);
 758
 759         rxq->state = SFC_RXQ_INITIALIZED;
 760
 761         efx_rx_qdestroy(rxq->common);
 762
 763         sfc_ev_qstop(rxq->evq);
 764 }
 765
 766 uint64_t
 767 sfc_rx_get_dev_offload_caps(struct sfc_adapter *sa)
 768 {
 769         const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
 770         uint64_t caps = 0;
 771
 772         caps |= DEV_RX_OFFLOAD_JUMBO_FRAME;
 773
 774         if (sa->priv.dp_rx->features & SFC_DP_RX_FEAT_CHECKSUM) {
 775                 caps |= DEV_RX_OFFLOAD_IPV4_CKSUM;
 776                 caps |= DEV_RX_OFFLOAD_UDP_CKSUM;
 777                 caps |= DEV_RX_OFFLOAD_TCP_CKSUM;
 778         }
 779
 780         if (encp->enc_tunnel_encapsulations_supported &&
 781             (sa->priv.dp_rx->features & SFC_DP_RX_FEAT_TUNNELS))
 782                 caps |= DEV_RX_OFFLOAD_OUTER_IPV4_CKSUM;
 783
 784         return caps;
 785 }
 786
 787 uint64_t
 788 sfc_rx_get_queue_offload_caps(struct sfc_adapter *sa)
 789 {
 790         uint64_t caps = 0;
 791
 792         if (sa->priv.dp_rx->features & SFC_DP_RX_FEAT_SCATTER)
 793                 caps |= DEV_RX_OFFLOAD_SCATTER;
 794
 795         return caps;
 796 }
 797
 798 static int
 799 sfc_rx_qcheck_conf(struct sfc_adapter *sa, unsigned int rxq_max_fill_level,
 800                    const struct rte_eth_rxconf *rx_conf,
 801                    __rte_unused uint64_t offloads)
 802 {
 803         int rc = 0;
 804
 805         if (rx_conf->rx_thresh.pthresh != 0 ||
 806             rx_conf->rx_thresh.hthresh != 0 ||
 807             rx_conf->rx_thresh.wthresh != 0) {
 808                 sfc_warn(sa,
 809                         "RxQ prefetch/host/writeback thresholds are not supported");
 810         }
 811
 812         if (rx_conf->rx_free_thresh > rxq_max_fill_level) {
 813                 sfc_err(sa,
 814                         "RxQ free threshold too large: %u vs maximum %u",
 815                         rx_conf->rx_free_thresh, rxq_max_fill_level);
 816                 rc = EINVAL;
 817         }
 818
 819         if (rx_conf->rx_drop_en == 0) {
 820                 sfc_err(sa, "RxQ drop disable is not supported");
 821                 rc = EINVAL;
 822         }
 823
 824         return rc;
 825 }
 826
 827 static unsigned int
 828 sfc_rx_mbuf_data_alignment(struct rte_mempool *mb_pool)
 829 {
 830         uint32_t data_off;
 831         uint32_t order;
 832
 833         /* The mbuf object itself is always cache line aligned */
 834         order = rte_bsf32(RTE_CACHE_LINE_SIZE);
 835
 836         /* Data offset from mbuf object start */
 837         data_off = sizeof(struct rte_mbuf) + rte_pktmbuf_priv_size(mb_pool) +
 838                 RTE_PKTMBUF_HEADROOM;
 839
 840         order = MIN(order, rte_bsf32(data_off));
 841
 842         return 1u << order;
 843 }
 844
 845 static uint16_t
 846 sfc_rx_mb_pool_buf_size(struct sfc_adapter *sa, struct rte_mempool *mb_pool)
 847 {
 848         const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
 849         const uint32_t nic_align_start = MAX(1, encp->enc_rx_buf_align_start);
 850         const uint32_t nic_align_end = MAX(1, encp->enc_rx_buf_align_end);
 851         uint16_t buf_size;
 852         unsigned int buf_aligned;
 853         unsigned int start_alignment;
 854         unsigned int end_padding_alignment;
 855
 856         /* Below it is assumed that both alignments are power of 2 */
 857         SFC_ASSERT(rte_is_power_of_2(nic_align_start));
 858         SFC_ASSERT(rte_is_power_of_2(nic_align_end));
 859
 860         /*
 861          * mbuf is always cache line aligned, double-check
 862          * that it meets rx buffer start alignment requirements.
 863          */
 864
 865         /* Start from mbuf pool data room size */
 866         buf_size = rte_pktmbuf_data_room_size(mb_pool);
 867
 868         /* Remove headroom */
 869         if (buf_size <= RTE_PKTMBUF_HEADROOM) {
 870                 sfc_err(sa,
 871                         "RxQ mbuf pool %s object data room size %u is smaller than headroom %u",
 872                         mb_pool->name, buf_size, RTE_PKTMBUF_HEADROOM);
 873                 return 0;
 874         }
 875         buf_size -= RTE_PKTMBUF_HEADROOM;
 876
 877         /* Calculate guaranteed data start alignment */
 878         buf_aligned = sfc_rx_mbuf_data_alignment(mb_pool);
 879
 880         /* Reserve space for start alignment */
 881         if (buf_aligned < nic_align_start) {
 882                 start_alignment = nic_align_start - buf_aligned;
 883                 if (buf_size <= start_alignment) {
 884                         sfc_err(sa,
 885                                 "RxQ mbuf pool %s object data room size %u is insufficient for headroom %u and buffer start alignment %u required by NIC",
 886                                 mb_pool->name,
 887                                 rte_pktmbuf_data_room_size(mb_pool),
 888                                 RTE_PKTMBUF_HEADROOM, start_alignment);
 889                         return 0;
 890                 }
 891                 buf_aligned = nic_align_start;
 892                 buf_size -= start_alignment;
 893         } else {
 894                 start_alignment = 0;
 895         }
 896
 897         /* Make sure that end padding does not write beyond the buffer */
 898         if (buf_aligned < nic_align_end) {
 899                 /*
 900                  * Estimate space which can be lost. If guarnteed buffer
 901                  * size is odd, lost space is (nic_align_end - 1). More
 902                  * accurate formula is below.
 903                  */
 904                 end_padding_alignment = nic_align_end -
 905                         MIN(buf_aligned, 1u << (rte_bsf32(buf_size) - 1));
 906                 if (buf_size <= end_padding_alignment) {
 907                         sfc_err(sa,
 908                                 "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",
 909                                 mb_pool->name,
 910                                 rte_pktmbuf_data_room_size(mb_pool),
 911                                 RTE_PKTMBUF_HEADROOM, start_alignment,
 912                                 end_padding_alignment);
 913                         return 0;
 914                 }
 915                 buf_size -= end_padding_alignment;
 916         } else {
 917                 /*
 918                  * Start is aligned the same or better than end,
 919                  * just align length.
 920                  */
 921                 buf_size = P2ALIGN(buf_size, nic_align_end);
 922         }
 923
 924         return buf_size;
 925 }
 926
 927 int
 928 sfc_rx_qinit(struct sfc_adapter *sa, unsigned int sw_index,
 929              uint16_t nb_rx_desc, unsigned int socket_id,
 930              const struct rte_eth_rxconf *rx_conf,
 931              struct rte_mempool *mb_pool)
 932 {
 933         const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
 934         struct sfc_rss *rss = &sa->rss;
 935         int rc;
 936         unsigned int rxq_entries;
 937         unsigned int evq_entries;
 938         unsigned int rxq_max_fill_level;
 939         uint64_t offloads;
 940         uint16_t buf_size;
 941         struct sfc_rxq_info *rxq_info;
 942         struct sfc_evq *evq;
 943         struct sfc_rxq *rxq;
 944         struct sfc_dp_rx_qcreate_info info;
 945
 946         rc = sa->priv.dp_rx->qsize_up_rings(nb_rx_desc, mb_pool, &rxq_entries,
 947                                             &evq_entries, &rxq_max_fill_level);
 948         if (rc != 0)
 949                 goto fail_size_up_rings;
 950         SFC_ASSERT(rxq_entries >= EFX_RXQ_MINNDESCS);
 951         SFC_ASSERT(rxq_entries <= EFX_RXQ_MAXNDESCS);
 952         SFC_ASSERT(rxq_max_fill_level <= nb_rx_desc);
 953
 954         offloads = rx_conf->offloads |
 955                 sa->eth_dev->data->dev_conf.rxmode.offloads;
 956         rc = sfc_rx_qcheck_conf(sa, rxq_max_fill_level, rx_conf, offloads);
 957         if (rc != 0)
 958                 goto fail_bad_conf;
 959
 960         buf_size = sfc_rx_mb_pool_buf_size(sa, mb_pool);
 961         if (buf_size == 0) {
 962                 sfc_err(sa, "RxQ %u mbuf pool object size is too small",
 963                         sw_index);
 964                 rc = EINVAL;
 965                 goto fail_bad_conf;
 966         }
 967
 968         if ((buf_size < sa->port.pdu + encp->enc_rx_prefix_size) &&
 969             (~offloads & DEV_RX_OFFLOAD_SCATTER)) {
 970                 sfc_err(sa, "Rx scatter is disabled and RxQ %u mbuf pool "
 971                         "object size is too small", sw_index);
 972                 sfc_err(sa, "RxQ %u calculated Rx buffer size is %u vs "
 973                         "PDU size %u plus Rx prefix %u bytes",
 974                         sw_index, buf_size, (unsigned int)sa->port.pdu,
 975                         encp->enc_rx_prefix_size);
 976                 rc = EINVAL;
 977                 goto fail_bad_conf;
 978         }
 979
 980         SFC_ASSERT(sw_index < sa->rxq_count);
 981         rxq_info = &sa->rxq_info[sw_index];
 982
 983         SFC_ASSERT(rxq_entries <= rxq_info->max_entries);
 984         rxq_info->entries = rxq_entries;
 985
 986         if (sa->priv.dp_rx->dp.hw_fw_caps & SFC_DP_HW_FW_CAP_RX_ES_SUPER_BUFFER)
 987                 rxq_info->type = EFX_RXQ_TYPE_ES_SUPER_BUFFER;
 988         else
 989                 rxq_info->type = EFX_RXQ_TYPE_DEFAULT;
 990
 991         rxq_info->type_flags =
 992                 (offloads & DEV_RX_OFFLOAD_SCATTER) ?
 993                 EFX_RXQ_FLAG_SCATTER : EFX_RXQ_FLAG_NONE;
 994
 995         if ((encp->enc_tunnel_encapsulations_supported != 0) &&
 996             (sa->priv.dp_rx->features & SFC_DP_RX_FEAT_TUNNELS))
 997                 rxq_info->type_flags |= EFX_RXQ_FLAG_INNER_CLASSES;
 998
 999         rc = sfc_ev_qinit(sa, SFC_EVQ_TYPE_RX, sw_index,
1000                           evq_entries, socket_id, &evq);
1001         if (rc != 0)
1002                 goto fail_ev_qinit;
1003
1004         rc = ENOMEM;
1005         rxq = rte_zmalloc_socket("sfc-rxq", sizeof(*rxq), RTE_CACHE_LINE_SIZE,
1006                                  socket_id);
1007         if (rxq == NULL)
1008                 goto fail_rxq_alloc;
1009
1010         rxq_info->rxq = rxq;
1011
1012         rxq->evq = evq;
1013         rxq->hw_index = sw_index;
1014         rxq_info->refill_threshold =
1015                 RTE_MAX(rx_conf->rx_free_thresh, SFC_RX_REFILL_BULK);
1016         rxq_info->refill_mb_pool = mb_pool;
1017         rxq->buf_size = buf_size;
1018
1019         rc = sfc_dma_alloc(sa, "rxq", sw_index, EFX_RXQ_SIZE(rxq_info->entries),
1020                            socket_id, &rxq->mem);
1021         if (rc != 0)
1022                 goto fail_dma_alloc;
1023
1024         memset(&info, 0, sizeof(info));
1025         info.refill_mb_pool = rxq_info->refill_mb_pool;
1026         info.max_fill_level = rxq_max_fill_level;
1027         info.refill_threshold = rxq_info->refill_threshold;
1028         info.buf_size = buf_size;
1029         info.batch_max = encp->enc_rx_batch_max;
1030         info.prefix_size = encp->enc_rx_prefix_size;
1031
1032         if (rss->hash_support == EFX_RX_HASH_AVAILABLE && rss->channels > 0)
1033                 info.flags |= SFC_RXQ_FLAG_RSS_HASH;
1034
1035         info.rxq_entries = rxq_info->entries;
1036         info.rxq_hw_ring = rxq->mem.esm_base;
1037         info.evq_entries = evq_entries;
1038         info.evq_hw_ring = evq->mem.esm_base;
1039         info.hw_index = rxq->hw_index;
1040         info.mem_bar = sa->mem_bar.esb_base;
1041         info.vi_window_shift = encp->enc_vi_window_shift;
1042
1043         rc = sa->priv.dp_rx->qcreate(sa->eth_dev->data->port_id, sw_index,
1044                                      &RTE_ETH_DEV_TO_PCI(sa->eth_dev)->addr,
1045                                      socket_id, &info, &rxq->dp);
1046         if (rc != 0)
1047                 goto fail_dp_rx_qcreate;
1048
1049         evq->dp_rxq = rxq->dp;
1050
1051         rxq->state = SFC_RXQ_INITIALIZED;
1052
1053         rxq_info->deferred_start = (rx_conf->rx_deferred_start != 0);
1054
1055         return 0;
1056
1057 fail_dp_rx_qcreate:
1058         sfc_dma_free(sa, &rxq->mem);
1059
1060 fail_dma_alloc:
1061         rxq_info->rxq = NULL;
1062         rte_free(rxq);
1063
1064 fail_rxq_alloc:
1065         sfc_ev_qfini(evq);
1066
1067 fail_ev_qinit:
1068         rxq_info->entries = 0;
1069
1070 fail_bad_conf:
1071 fail_size_up_rings:
1072         sfc_log_init(sa, "failed %d", rc);
1073         return rc;
1074 }
1075
1076 void
1077 sfc_rx_qfini(struct sfc_adapter *sa, unsigned int sw_index)
1078 {
1079         struct sfc_rxq_info *rxq_info;
1080         struct sfc_rxq *rxq;
1081
1082         SFC_ASSERT(sw_index < sa->rxq_count);
1083         sa->eth_dev->data->rx_queues[sw_index] = NULL;
1084
1085         rxq_info = &sa->rxq_info[sw_index];
1086
1087         rxq = rxq_info->rxq;
1088         SFC_ASSERT(rxq->state == SFC_RXQ_INITIALIZED);
1089
1090         sa->priv.dp_rx->qdestroy(rxq->dp);
1091         rxq->dp = NULL;
1092
1093         rxq_info->rxq = NULL;
1094         rxq_info->entries = 0;
1095
1096         sfc_dma_free(sa, &rxq->mem);
1097
1098         sfc_ev_qfini(rxq->evq);
1099         rxq->evq = NULL;
1100
1101         rte_free(rxq);
1102 }
1103
1104 /*
1105  * Mapping between RTE RSS hash functions and their EFX counterparts.
1106  */
1107 static const struct sfc_rss_hf_rte_to_efx sfc_rss_hf_map[] = {
1108         { ETH_RSS_NONFRAG_IPV4_TCP,
1109           EFX_RX_HASH(IPV4_TCP, 4TUPLE) },
1110         { ETH_RSS_NONFRAG_IPV4_UDP,
1111           EFX_RX_HASH(IPV4_UDP, 4TUPLE) },
1112         { ETH_RSS_NONFRAG_IPV6_TCP | ETH_RSS_IPV6_TCP_EX,
1113           EFX_RX_HASH(IPV6_TCP, 4TUPLE) },
1114         { ETH_RSS_NONFRAG_IPV6_UDP | ETH_RSS_IPV6_UDP_EX,
1115           EFX_RX_HASH(IPV6_UDP, 4TUPLE) },
1116         { ETH_RSS_IPV4 | ETH_RSS_FRAG_IPV4 | ETH_RSS_NONFRAG_IPV4_OTHER,
1117           EFX_RX_HASH(IPV4_TCP, 2TUPLE) | EFX_RX_HASH(IPV4_UDP, 2TUPLE) |
1118           EFX_RX_HASH(IPV4, 2TUPLE) },
1119         { ETH_RSS_IPV6 | ETH_RSS_FRAG_IPV6 | ETH_RSS_NONFRAG_IPV6_OTHER |
1120           ETH_RSS_IPV6_EX,
1121           EFX_RX_HASH(IPV6_TCP, 2TUPLE) | EFX_RX_HASH(IPV6_UDP, 2TUPLE) |
1122           EFX_RX_HASH(IPV6, 2TUPLE) }
1123 };
1124
1125 static efx_rx_hash_type_t
1126 sfc_rx_hash_types_mask_supp(efx_rx_hash_type_t hash_type,
1127                             unsigned int *hash_type_flags_supported,
1128                             unsigned int nb_hash_type_flags_supported)
1129 {
1130         efx_rx_hash_type_t hash_type_masked = 0;
1131         unsigned int i, j;
1132
1133         for (i = 0; i < nb_hash_type_flags_supported; ++i) {
1134                 unsigned int class_tuple_lbn[] = {
1135                         EFX_RX_CLASS_IPV4_TCP_LBN,
1136                         EFX_RX_CLASS_IPV4_UDP_LBN,
1137                         EFX_RX_CLASS_IPV4_LBN,
1138                         EFX_RX_CLASS_IPV6_TCP_LBN,
1139                         EFX_RX_CLASS_IPV6_UDP_LBN,
1140                         EFX_RX_CLASS_IPV6_LBN
1141                 };
1142
1143                 for (j = 0; j < RTE_DIM(class_tuple_lbn); ++j) {
1144                         unsigned int tuple_mask = EFX_RX_CLASS_HASH_4TUPLE;
1145                         unsigned int flag;
1146
1147                         tuple_mask <<= class_tuple_lbn[j];
1148                         flag = hash_type & tuple_mask;
1149
1150                         if (flag == hash_type_flags_supported[i])
1151                                 hash_type_masked |= flag;
1152                 }
1153         }
1154
1155         return hash_type_masked;
1156 }
1157
1158 int
1159 sfc_rx_hash_init(struct sfc_adapter *sa)
1160 {
1161         struct sfc_rss *rss = &sa->rss;
1162         const efx_nic_cfg_t *encp = efx_nic_cfg_get(sa->nic);
1163         uint32_t alg_mask = encp->enc_rx_scale_hash_alg_mask;
1164         efx_rx_hash_alg_t alg;
1165         unsigned int flags_supp[EFX_RX_HASH_NFLAGS];
1166         unsigned int nb_flags_supp;
1167         struct sfc_rss_hf_rte_to_efx *hf_map;
1168         struct sfc_rss_hf_rte_to_efx *entry;
1169         efx_rx_hash_type_t efx_hash_types;
1170         unsigned int i;
1171         int rc;
1172
1173         if (alg_mask & (1U << EFX_RX_HASHALG_TOEPLITZ))
1174                 alg = EFX_RX_HASHALG_TOEPLITZ;
1175         else if (alg_mask & (1U << EFX_RX_HASHALG_PACKED_STREAM))
1176                 alg = EFX_RX_HASHALG_PACKED_STREAM;
1177         else
1178                 return EINVAL;
1179
1180         rc = efx_rx_scale_hash_flags_get(sa->nic, alg, flags_supp,
1181                                          RTE_DIM(flags_supp), &nb_flags_supp);
1182         if (rc != 0)
1183                 return rc;
1184
1185         hf_map = rte_calloc_socket("sfc-rss-hf-map",
1186                                    RTE_DIM(sfc_rss_hf_map),
1187                                    sizeof(*hf_map), 0, sa->socket_id);
1188         if (hf_map == NULL)
1189                 return ENOMEM;
1190
1191         entry = hf_map;
1192         efx_hash_types = 0;
1193         for (i = 0; i < RTE_DIM(sfc_rss_hf_map); ++i) {
1194                 efx_rx_hash_type_t ht;
1195
1196                 ht = sfc_rx_hash_types_mask_supp(sfc_rss_hf_map[i].efx,
1197                                                  flags_supp, nb_flags_supp);
1198                 if (ht != 0) {
1199                         entry->rte = sfc_rss_hf_map[i].rte;
1200                         entry->efx = ht;
1201                         efx_hash_types |= ht;
1202                         ++entry;
1203                 }
1204         }
1205
1206         rss->hash_alg = alg;
1207         rss->hf_map_nb_entries = (unsigned int)(entry - hf_map);
1208         rss->hf_map = hf_map;
1209         rss->hash_types = efx_hash_types;
1210
1211         return 0;
1212 }
1213
1214 void
1215 sfc_rx_hash_fini(struct sfc_adapter *sa)
1216 {
1217         struct sfc_rss *rss = &sa->rss;
1218
1219         rte_free(rss->hf_map);
1220 }
1221
1222 int
1223 sfc_rx_hf_rte_to_efx(struct sfc_adapter *sa, uint64_t rte,
1224                      efx_rx_hash_type_t *efx)
1225 {
1226         struct sfc_rss *rss = &sa->rss;
1227         efx_rx_hash_type_t hash_types = 0;
1228         unsigned int i;
1229
1230         for (i = 0; i < rss->hf_map_nb_entries; ++i) {
1231                 uint64_t rte_mask = rss->hf_map[i].rte;
1232
1233                 if ((rte & rte_mask) != 0) {
1234                         rte &= ~rte_mask;
1235                         hash_types |= rss->hf_map[i].efx;
1236                 }
1237         }
1238
1239         if (rte != 0) {
1240                 sfc_err(sa, "unsupported hash functions requested");
1241                 return EINVAL;
1242         }
1243
1244         *efx = hash_types;
1245
1246         return 0;
1247 }
1248
1249 uint64_t
1250 sfc_rx_hf_efx_to_rte(struct sfc_adapter *sa, efx_rx_hash_type_t efx)
1251 {
1252         struct sfc_rss *rss = &sa->rss;
1253         uint64_t rte = 0;
1254         unsigned int i;
1255
1256         for (i = 0; i < rss->hf_map_nb_entries; ++i) {
1257                 efx_rx_hash_type_t hash_type = rss->hf_map[i].efx;
1258
1259                 if ((efx & hash_type) == hash_type)
1260                         rte |= rss->hf_map[i].rte;
1261         }
1262
1263         return rte;
1264 }
1265
1266 static int
1267 sfc_rx_process_adv_conf_rss(struct sfc_adapter *sa,
1268                             struct rte_eth_rss_conf *conf)
1269 {
1270         struct sfc_rss *rss = &sa->rss;
1271         efx_rx_hash_type_t efx_hash_types = rss->hash_types;
1272         uint64_t rss_hf = sfc_rx_hf_efx_to_rte(sa, efx_hash_types);
1273         int rc;
1274
1275         if (rss->context_type != EFX_RX_SCALE_EXCLUSIVE) {
1276                 if ((conf->rss_hf != 0 && conf->rss_hf != rss_hf) ||
1277                     conf->rss_key != NULL)
1278                         return EINVAL;
1279         }
1280
1281         if (conf->rss_hf != 0) {
1282                 rc = sfc_rx_hf_rte_to_efx(sa, conf->rss_hf, &efx_hash_types);
1283                 if (rc != 0)
1284                         return rc;
1285         }
1286
1287         if (conf->rss_key != NULL) {
1288                 if (conf->rss_key_len != sizeof(rss->key)) {
1289                         sfc_err(sa, "RSS key size is wrong (should be %lu)",
1290                                 sizeof(rss->key));
1291                         return EINVAL;
1292                 }
1293                 rte_memcpy(rss->key, conf->rss_key, sizeof(rss->key));
1294         }
1295
1296         rss->hash_types = efx_hash_types;
1297
1298         return 0;
1299 }
1300
1301 static int
1302 sfc_rx_rss_config(struct sfc_adapter *sa)
1303 {
1304         struct sfc_rss *rss = &sa->rss;
1305         int rc = 0;
1306
1307         if (rss->channels > 0) {
1308                 rc = efx_rx_scale_mode_set(sa->nic, EFX_RSS_CONTEXT_DEFAULT,
1309                                            rss->hash_alg, rss->hash_types,
1310                                            B_TRUE);
1311                 if (rc != 0)
1312                         goto finish;
1313
1314                 rc = efx_rx_scale_key_set(sa->nic, EFX_RSS_CONTEXT_DEFAULT,
1315                                           rss->key, sizeof(rss->key));
1316                 if (rc != 0)
1317                         goto finish;
1318
1319                 rc = efx_rx_scale_tbl_set(sa->nic, EFX_RSS_CONTEXT_DEFAULT,
1320                                           rss->tbl, RTE_DIM(rss->tbl));
1321         }
1322
1323 finish:
1324         return rc;
1325 }
1326
1327 int
1328 sfc_rx_start(struct sfc_adapter *sa)
1329 {
1330         unsigned int sw_index;
1331         int rc;
1332
1333         sfc_log_init(sa, "rxq_count=%u", sa->rxq_count);
1334
1335         rc = efx_rx_init(sa->nic);
1336         if (rc != 0)
1337                 goto fail_rx_init;
1338
1339         rc = sfc_rx_rss_config(sa);
1340         if (rc != 0)
1341                 goto fail_rss_config;
1342
1343         for (sw_index = 0; sw_index < sa->rxq_count; ++sw_index) {
1344                 if (sa->rxq_info[sw_index].rxq != NULL &&
1345                     (!sa->rxq_info[sw_index].deferred_start ||
1346                      sa->rxq_info[sw_index].deferred_started)) {
1347                         rc = sfc_rx_qstart(sa, sw_index);
1348                         if (rc != 0)
1349                                 goto fail_rx_qstart;
1350                 }
1351         }
1352
1353         return 0;
1354
1355 fail_rx_qstart:
1356         while (sw_index-- > 0)
1357                 sfc_rx_qstop(sa, sw_index);
1358
1359 fail_rss_config:
1360         efx_rx_fini(sa->nic);
1361
1362 fail_rx_init:
1363         sfc_log_init(sa, "failed %d", rc);
1364         return rc;
1365 }
1366
1367 void
1368 sfc_rx_stop(struct sfc_adapter *sa)
1369 {
1370         unsigned int sw_index;
1371
1372         sfc_log_init(sa, "rxq_count=%u", sa->rxq_count);
1373
1374         sw_index = sa->rxq_count;
1375         while (sw_index-- > 0) {
1376                 if (sa->rxq_info[sw_index].rxq != NULL)
1377                         sfc_rx_qstop(sa, sw_index);
1378         }
1379
1380         efx_rx_fini(sa->nic);
1381 }
1382
1383 static int
1384 sfc_rx_qinit_info(struct sfc_adapter *sa, unsigned int sw_index)
1385 {
1386         struct sfc_rxq_info *rxq_info = &sa->rxq_info[sw_index];
1387         unsigned int max_entries;
1388
1389         max_entries = EFX_RXQ_MAXNDESCS;
1390         SFC_ASSERT(rte_is_power_of_2(max_entries));
1391
1392         rxq_info->max_entries = max_entries;
1393
1394         return 0;
1395 }
1396
1397 static int
1398 sfc_rx_check_mode(struct sfc_adapter *sa, struct rte_eth_rxmode *rxmode)
1399 {
1400         uint64_t offloads_supported = sfc_rx_get_dev_offload_caps(sa) |
1401                                       sfc_rx_get_queue_offload_caps(sa);
1402         struct sfc_rss *rss = &sa->rss;
1403         int rc = 0;
1404
1405         switch (rxmode->mq_mode) {
1406         case ETH_MQ_RX_NONE:
1407                 /* No special checks are required */
1408                 break;
1409         case ETH_MQ_RX_RSS:
1410                 if (rss->context_type == EFX_RX_SCALE_UNAVAILABLE) {
1411                         sfc_err(sa, "RSS is not available");
1412                         rc = EINVAL;
1413                 }
1414                 break;
1415         default:
1416                 sfc_err(sa, "Rx multi-queue mode %u not supported",
1417                         rxmode->mq_mode);
1418                 rc = EINVAL;
1419         }
1420
1421         /*
1422          * Requested offloads are validated against supported by ethdev,
1423          * so unsupported offloads cannot be added as the result of
1424          * below check.
1425          */
1426         if ((rxmode->offloads & DEV_RX_OFFLOAD_CHECKSUM) !=
1427             (offloads_supported & DEV_RX_OFFLOAD_CHECKSUM)) {
1428                 sfc_warn(sa, "Rx checksum offloads cannot be disabled - always on (IPv4/TCP/UDP)");
1429                 rxmode->offloads |= DEV_RX_OFFLOAD_CHECKSUM;
1430         }
1431
1432         if ((offloads_supported & DEV_RX_OFFLOAD_OUTER_IPV4_CKSUM) &&
1433             (~rxmode->offloads & DEV_RX_OFFLOAD_OUTER_IPV4_CKSUM)) {
1434                 sfc_warn(sa, "Rx outer IPv4 checksum offload cannot be disabled - always on");
1435                 rxmode->offloads |= DEV_RX_OFFLOAD_OUTER_IPV4_CKSUM;
1436         }
1437
1438         return rc;
1439 }
1440
1441 /**
1442  * Destroy excess queues that are no longer needed after reconfiguration
1443  * or complete close.
1444  */
1445 static void
1446 sfc_rx_fini_queues(struct sfc_adapter *sa, unsigned int nb_rx_queues)
1447 {
1448         int sw_index;
1449
1450         SFC_ASSERT(nb_rx_queues <= sa->rxq_count);
1451
1452         sw_index = sa->rxq_count;
1453         while (--sw_index >= (int)nb_rx_queues) {
1454                 if (sa->rxq_info[sw_index].rxq != NULL)
1455                         sfc_rx_qfini(sa, sw_index);
1456         }
1457
1458         sa->rxq_count = nb_rx_queues;
1459 }
1460
1461 /**
1462  * Initialize Rx subsystem.
1463  *
1464  * Called at device (re)configuration stage when number of receive queues is
1465  * specified together with other device level receive configuration.
1466  *
1467  * It should be used to allocate NUMA-unaware resources.
1468  */
1469 int
1470 sfc_rx_configure(struct sfc_adapter *sa)
1471 {
1472         struct sfc_rss *rss = &sa->rss;
1473         struct rte_eth_conf *dev_conf = &sa->eth_dev->data->dev_conf;
1474         const unsigned int nb_rx_queues = sa->eth_dev->data->nb_rx_queues;
1475         int rc;
1476
1477         sfc_log_init(sa, "nb_rx_queues=%u (old %u)",
1478                      nb_rx_queues, sa->rxq_count);
1479
1480         rc = sfc_rx_check_mode(sa, &dev_conf->rxmode);
1481         if (rc != 0)
1482                 goto fail_check_mode;
1483
1484         if (nb_rx_queues == sa->rxq_count)
1485                 goto configure_rss;
1486
1487         if (sa->rxq_info == NULL) {
1488                 rc = ENOMEM;
1489                 sa->rxq_info = rte_calloc_socket("sfc-rxqs", nb_rx_queues,
1490                                                  sizeof(sa->rxq_info[0]), 0,
1491                                                  sa->socket_id);
1492                 if (sa->rxq_info == NULL)
1493                         goto fail_rxqs_alloc;
1494         } else {
1495                 struct sfc_rxq_info *new_rxq_info;
1496
1497                 if (nb_rx_queues < sa->rxq_count)
1498                         sfc_rx_fini_queues(sa, nb_rx_queues);
1499
1500                 rc = ENOMEM;
1501                 new_rxq_info =
1502                         rte_realloc(sa->rxq_info,
1503                                     nb_rx_queues * sizeof(sa->rxq_info[0]), 0);
1504                 if (new_rxq_info == NULL && nb_rx_queues > 0)
1505                         goto fail_rxqs_realloc;
1506
1507                 sa->rxq_info = new_rxq_info;
1508                 if (nb_rx_queues > sa->rxq_count)
1509                         memset(&sa->rxq_info[sa->rxq_count], 0,
1510                                (nb_rx_queues - sa->rxq_count) *
1511                                sizeof(sa->rxq_info[0]));
1512         }
1513
1514         while (sa->rxq_count < nb_rx_queues) {
1515                 rc = sfc_rx_qinit_info(sa, sa->rxq_count);
1516                 if (rc != 0)
1517                         goto fail_rx_qinit_info;
1518
1519                 sa->rxq_count++;
1520         }
1521
1522 configure_rss:
1523         rss->channels = (dev_conf->rxmode.mq_mode == ETH_MQ_RX_RSS) ?
1524                          MIN(sa->rxq_count, EFX_MAXRSS) : 0;
1525
1526         if (rss->channels > 0) {
1527                 struct rte_eth_rss_conf *adv_conf_rss;
1528                 unsigned int sw_index;
1529
1530                 for (sw_index = 0; sw_index < EFX_RSS_TBL_SIZE; ++sw_index)
1531                         rss->tbl[sw_index] = sw_index % rss->channels;
1532
1533                 adv_conf_rss = &dev_conf->rx_adv_conf.rss_conf;
1534                 rc = sfc_rx_process_adv_conf_rss(sa, adv_conf_rss);
1535                 if (rc != 0)
1536                         goto fail_rx_process_adv_conf_rss;
1537         }
1538
1539         return 0;
1540
1541 fail_rx_process_adv_conf_rss:
1542 fail_rx_qinit_info:
1543 fail_rxqs_realloc:
1544 fail_rxqs_alloc:
1545         sfc_rx_close(sa);
1546
1547 fail_check_mode:
1548         sfc_log_init(sa, "failed %d", rc);
1549         return rc;
1550 }
1551
1552 /**
1553  * Shutdown Rx subsystem.
1554  *
1555  * Called at device close stage, for example, before device shutdown.
1556  */
1557 void
1558 sfc_rx_close(struct sfc_adapter *sa)
1559 {
1560         struct sfc_rss *rss = &sa->rss;
1561
1562         sfc_rx_fini_queues(sa, 0);
1563
1564         rss->channels = 0;
1565
1566         rte_free(sa->rxq_info);
1567         sa->rxq_info = NULL;
1568 }