drivers/net/mlx5/mlx5_rxtx.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright 2015 6WIND S.A.
   5  *   Copyright 2015 Mellanox.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of 6WIND S.A. nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  */
  33
  34 #include <assert.h>
  35 #include <stdint.h>
  36 #include <string.h>
  37 #include <stdlib.h>
  38
  39 /* Verbs header. */
  40 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
  41 #ifdef PEDANTIC
  42 #pragma GCC diagnostic ignored "-pedantic"
  43 #endif
  44 #include <infiniband/verbs.h>
  45 #ifdef PEDANTIC
  46 #pragma GCC diagnostic error "-pedantic"
  47 #endif
  48
  49 /* DPDK headers don't like -pedantic. */
  50 #ifdef PEDANTIC
  51 #pragma GCC diagnostic ignored "-pedantic"
  52 #endif
  53 #include <rte_mbuf.h>
  54 #include <rte_mempool.h>
  55 #include <rte_prefetch.h>
  56 #include <rte_common.h>
  57 #include <rte_branch_prediction.h>
  58 #include <rte_memory.h>
  59 #ifdef PEDANTIC
  60 #pragma GCC diagnostic error "-pedantic"
  61 #endif
  62
  63 #include "mlx5.h"
  64 #include "mlx5_utils.h"
  65 #include "mlx5_rxtx.h"
  66 #include "mlx5_autoconf.h"
  67 #include "mlx5_defs.h"
  68
  69 /**
  70  * Manage TX completions.
  71  *
  72  * When sending a burst, mlx5_tx_burst() posts several WRs.
  73  * To improve performance, a completion event is only required once every
  74  * MLX5_PMD_TX_PER_COMP_REQ sends. Doing so discards completion information
  75  * for other WRs, but this information would not be used anyway.
  76  *
  77  * @param txq
  78  *   Pointer to TX queue structure.
  79  *
  80  * @return
  81  *   0 on success, -1 on failure.
  82  */
  83 static int
  84 txq_complete(struct txq *txq)
  85 {
  86         unsigned int elts_comp = txq->elts_comp;
  87         unsigned int elts_tail = txq->elts_tail;
  88         unsigned int elts_free = txq->elts_tail;
  89         const unsigned int elts_n = txq->elts_n;
  90         int wcs_n;
  91
  92         if (unlikely(elts_comp == 0))
  93                 return 0;
  94 #ifdef DEBUG_SEND
  95         DEBUG("%p: processing %u work requests completions",
  96               (void *)txq, elts_comp);
  97 #endif
  98         wcs_n = txq->poll_cnt(txq->cq, elts_comp);
  99         if (unlikely(wcs_n == 0))
 100                 return 0;
 101         if (unlikely(wcs_n < 0)) {
 102                 DEBUG("%p: ibv_poll_cq() failed (wcs_n=%d)",
 103                       (void *)txq, wcs_n);
 104                 return -1;
 105         }
 106         elts_comp -= wcs_n;
 107         assert(elts_comp <= txq->elts_comp);
 108         /*
 109          * Assume WC status is successful as nothing can be done about it
 110          * anyway.
 111          */
 112         elts_tail += wcs_n * txq->elts_comp_cd_init;
 113         if (elts_tail >= elts_n)
 114                 elts_tail -= elts_n;
 115
 116         while (elts_free != elts_tail) {
 117                 struct txq_elt *elt = &(*txq->elts)[elts_free];
 118                 unsigned int elts_free_next =
 119                         (((elts_free + 1) == elts_n) ? 0 : elts_free + 1);
 120                 struct rte_mbuf *tmp = elt->buf;
 121                 struct txq_elt *elt_next = &(*txq->elts)[elts_free_next];
 122
 123 #ifndef NDEBUG
 124                 /* Poisoning. */
 125                 memset(elt, 0x66, sizeof(*elt));
 126 #endif
 127                 RTE_MBUF_PREFETCH_TO_FREE(elt_next->buf);
 128                 /* Faster than rte_pktmbuf_free(). */
 129                 do {
 130                         struct rte_mbuf *next = NEXT(tmp);
 131
 132                         rte_pktmbuf_free_seg(tmp);
 133                         tmp = next;
 134                 } while (tmp != NULL);
 135                 elts_free = elts_free_next;
 136         }
 137
 138         txq->elts_tail = elts_tail;
 139         txq->elts_comp = elts_comp;
 140         return 0;
 141 }
 142
 143 /**
 144  * Get Memory Pool (MP) from mbuf. If mbuf is indirect, the pool from which
 145  * the cloned mbuf is allocated is returned instead.
 146  *
 147  * @param buf
 148  *   Pointer to mbuf.
 149  *
 150  * @return
 151  *   Memory pool where data is located for given mbuf.
 152  */
 153 static struct rte_mempool *
 154 txq_mb2mp(struct rte_mbuf *buf)
 155 {
 156         if (unlikely(RTE_MBUF_INDIRECT(buf)))
 157                 return rte_mbuf_from_indirect(buf)->pool;
 158         return buf->pool;
 159 }
 160
 161 static inline uint32_t
 162 txq_mp2mr(struct txq *txq, struct rte_mempool *mp)
 163         __attribute__((always_inline));
 164
 165 /**
 166  * Get Memory Region (MR) <-> Memory Pool (MP) association from txq->mp2mr[].
 167  * Add MP to txq->mp2mr[] if it's not registered yet. If mp2mr[] is full,
 168  * remove an entry first.
 169  *
 170  * @param txq
 171  *   Pointer to TX queue structure.
 172  * @param[in] mp
 173  *   Memory Pool for which a Memory Region lkey must be returned.
 174  *
 175  * @return
 176  *   mr->lkey on success, (uint32_t)-1 on failure.
 177  */
 178 static inline uint32_t
 179 txq_mp2mr(struct txq *txq, struct rte_mempool *mp)
 180 {
 181         unsigned int i;
 182         uint32_t lkey = (uint32_t)-1;
 183
 184         for (i = 0; (i != RTE_DIM(txq->mp2mr)); ++i) {
 185                 if (unlikely(txq->mp2mr[i].mp == NULL)) {
 186                         /* Unknown MP, add a new MR for it. */
 187                         break;
 188                 }
 189                 if (txq->mp2mr[i].mp == mp) {
 190                         assert(txq->mp2mr[i].lkey != (uint32_t)-1);
 191                         assert(txq->mp2mr[i].mr->lkey == txq->mp2mr[i].lkey);
 192                         lkey = txq->mp2mr[i].lkey;
 193                         break;
 194                 }
 195         }
 196         if (unlikely(lkey == (uint32_t)-1))
 197                 lkey = txq_mp2mr_reg(txq, mp, i);
 198         return lkey;
 199 }
 200
 201 /**
 202  * Insert VLAN using mbuf headroom space.
 203  *
 204  * @param buf
 205  *   Buffer for VLAN insertion.
 206  *
 207  * @return
 208  *   0 on success, errno value on failure.
 209  */
 210 static inline int
 211 insert_vlan_sw(struct rte_mbuf *buf)
 212 {
 213         uintptr_t addr;
 214         uint32_t vlan;
 215         uint16_t head_room_len = rte_pktmbuf_headroom(buf);
 216
 217         if (head_room_len < 4)
 218                 return EINVAL;
 219
 220         addr = rte_pktmbuf_mtod(buf, uintptr_t);
 221         vlan = htonl(0x81000000 | buf->vlan_tci);
 222         memmove((void *)(addr - 4), (void *)addr, 12);
 223         memcpy((void *)(addr + 8), &vlan, sizeof(vlan));
 224
 225         SET_DATA_OFF(buf, head_room_len - 4);
 226         DATA_LEN(buf) += 4;
 227
 228         return 0;
 229 }
 230
 231 /**
 232  * DPDK callback for TX.
 233  *
 234  * @param dpdk_txq
 235  *   Generic pointer to TX queue structure.
 236  * @param[in] pkts
 237  *   Packets to transmit.
 238  * @param pkts_n
 239  *   Number of packets in array.
 240  *
 241  * @return
 242  *   Number of packets successfully transmitted (<= pkts_n).
 243  */
 244 uint16_t
 245 mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
 246 {
 247         struct txq *txq = (struct txq *)dpdk_txq;
 248         unsigned int elts_head = txq->elts_head;
 249         const unsigned int elts_n = txq->elts_n;
 250         unsigned int elts_comp_cd = txq->elts_comp_cd;
 251         unsigned int elts_comp = 0;
 252         unsigned int i;
 253         unsigned int max;
 254         int err;
 255         struct rte_mbuf *buf = pkts[0];
 256
 257         assert(elts_comp_cd != 0);
 258         /* Prefetch first packet cacheline. */
 259         rte_prefetch0(buf);
 260         txq_complete(txq);
 261         max = (elts_n - (elts_head - txq->elts_tail));
 262         if (max > elts_n)
 263                 max -= elts_n;
 264         assert(max >= 1);
 265         assert(max <= elts_n);
 266         /* Always leave one free entry in the ring. */
 267         --max;
 268         if (max == 0)
 269                 return 0;
 270         if (max > pkts_n)
 271                 max = pkts_n;
 272         for (i = 0; (i != max); ++i) {
 273                 struct rte_mbuf *buf_next = pkts[i + 1];
 274                 unsigned int elts_head_next =
 275                         (((elts_head + 1) == elts_n) ? 0 : elts_head + 1);
 276                 struct txq_elt *elt = &(*txq->elts)[elts_head];
 277                 uint32_t send_flags = 0;
 278 #ifdef HAVE_VERBS_VLAN_INSERTION
 279                 int insert_vlan = 0;
 280 #endif /* HAVE_VERBS_VLAN_INSERTION */
 281                 uintptr_t addr;
 282                 uint32_t length;
 283                 uint32_t lkey;
 284                 uintptr_t buf_next_addr;
 285
 286                 if (i + 1 < max)
 287                         rte_prefetch0(buf_next);
 288                 /* Request TX completion. */
 289                 if (unlikely(--elts_comp_cd == 0)) {
 290                         elts_comp_cd = txq->elts_comp_cd_init;
 291                         ++elts_comp;
 292                         send_flags |= IBV_EXP_QP_BURST_SIGNALED;
 293                 }
 294                 /* Should we enable HW CKSUM offload */
 295                 if (buf->ol_flags &
 296                     (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM)) {
 297                         send_flags |= IBV_EXP_QP_BURST_IP_CSUM;
 298                         /* HW does not support checksum offloads at arbitrary
 299                          * offsets but automatically recognizes the packet
 300                          * type. For inner L3/L4 checksums, only VXLAN (UDP)
 301                          * tunnels are currently supported. */
 302                         if (RTE_ETH_IS_TUNNEL_PKT(buf->packet_type))
 303                                 send_flags |= IBV_EXP_QP_BURST_TUNNEL;
 304                 }
 305                 if (buf->ol_flags & PKT_TX_VLAN_PKT) {
 306 #ifdef HAVE_VERBS_VLAN_INSERTION
 307                         if (!txq->priv->mps)
 308                                 insert_vlan = 1;
 309                         else
 310 #endif /* HAVE_VERBS_VLAN_INSERTION */
 311                         {
 312                                 err = insert_vlan_sw(buf);
 313                                 if (unlikely(err))
 314                                         goto stop;
 315                         }
 316                 }
 317                 /* Retrieve buffer information. */
 318                 addr = rte_pktmbuf_mtod(buf, uintptr_t);
 319                 length = DATA_LEN(buf);
 320                 /* Update element. */
 321                 elt->buf = buf;
 322                 if (txq->priv->sriov)
 323                         rte_prefetch0((volatile void *)
 324                                       (uintptr_t)addr);
 325                 /* Prefetch next buffer data. */
 326                 if (i + 1 < max) {
 327                         buf_next_addr =
 328                                 rte_pktmbuf_mtod(buf_next, uintptr_t);
 329                         rte_prefetch0((volatile void *)
 330                                       (uintptr_t)buf_next_addr);
 331                 }
 332                 /* Put packet into send queue. */
 333 #if MLX5_PMD_MAX_INLINE > 0
 334                 if (length <= txq->max_inline) {
 335 #ifdef HAVE_VERBS_VLAN_INSERTION
 336                         if (insert_vlan)
 337                                 err = txq->send_pending_inline_vlan
 338                                         (txq->qp,
 339                                          (void *)addr,
 340                                          length,
 341                                          send_flags,
 342                                          &buf->vlan_tci);
 343                         else
 344 #endif /* HAVE_VERBS_VLAN_INSERTION */
 345                                 err = txq->send_pending_inline
 346                                         (txq->qp,
 347                                          (void *)addr,
 348                                          length,
 349                                          send_flags);
 350                 } else
 351 #endif
 352                 {
 353                         /*
 354                          * Retrieve Memory Region key for this
 355                          * memory pool.
 356                          */
 357                         lkey = txq_mp2mr(txq, txq_mb2mp(buf));
 358                         if (unlikely(lkey == (uint32_t)-1)) {
 359                                 /* MR does not exist. */
 360                                 DEBUG("%p: unable to get MP <-> MR"
 361                                       " association", (void *)txq);
 362                                 /* Clean up TX element. */
 363                                 elt->buf = NULL;
 364                                 goto stop;
 365                         }
 366 #ifdef HAVE_VERBS_VLAN_INSERTION
 367                         if (insert_vlan)
 368                                 err = txq->send_pending_vlan
 369                                         (txq->qp,
 370                                          addr,
 371                                          length,
 372                                          lkey,
 373                                          send_flags,
 374                                          &buf->vlan_tci);
 375                         else
 376 #endif /* HAVE_VERBS_VLAN_INSERTION */
 377                                 err = txq->send_pending
 378                                         (txq->qp,
 379                                          addr,
 380                                          length,
 381                                          lkey,
 382                                          send_flags);
 383                 }
 384                 if (unlikely(err))
 385                         goto stop;
 386 #ifdef MLX5_PMD_SOFT_COUNTERS
 387                 /* Increment sent bytes counter. */
 388                 txq->stats.obytes += length;
 389 #endif
 390 stop:
 391                 elts_head = elts_head_next;
 392                 buf = buf_next;
 393         }
 394         /* Take a shortcut if nothing must be sent. */
 395         if (unlikely(i == 0))
 396                 return 0;
 397 #ifdef MLX5_PMD_SOFT_COUNTERS
 398         /* Increment sent packets counter. */
 399         txq->stats.opackets += i;
 400 #endif
 401         /* Ring QP doorbell. */
 402         err = txq->send_flush(txq->qp);
 403         if (unlikely(err)) {
 404                 /* A nonzero value is not supposed to be returned.
 405                  * Nothing can be done about it. */
 406                 DEBUG("%p: send_flush() failed with error %d",
 407                       (void *)txq, err);
 408         }
 409         txq->elts_head = elts_head;
 410         txq->elts_comp += elts_comp;
 411         txq->elts_comp_cd = elts_comp_cd;
 412         return i;
 413 }
 414
 415 /**
 416  * Translate RX completion flags to packet type.
 417  *
 418  * @param flags
 419  *   RX completion flags returned by poll_length_flags().
 420  *
 421  * @note: fix mlx5_dev_supported_ptypes_get() if any change here.
 422  *
 423  * @return
 424  *   Packet type for struct rte_mbuf.
 425  */
 426 static inline uint32_t
 427 rxq_cq_to_pkt_type(uint32_t flags)
 428 {
 429         uint32_t pkt_type;
 430
 431         if (flags & IBV_EXP_CQ_RX_TUNNEL_PACKET)
 432                 pkt_type =
 433                         TRANSPOSE(flags,
 434                                   IBV_EXP_CQ_RX_OUTER_IPV4_PACKET,
 435                                   RTE_PTYPE_L3_IPV4) |
 436                         TRANSPOSE(flags,
 437                                   IBV_EXP_CQ_RX_OUTER_IPV6_PACKET,
 438                                   RTE_PTYPE_L3_IPV6) |
 439                         TRANSPOSE(flags,
 440                                   IBV_EXP_CQ_RX_IPV4_PACKET,
 441                                   RTE_PTYPE_INNER_L3_IPV4) |
 442                         TRANSPOSE(flags,
 443                                   IBV_EXP_CQ_RX_IPV6_PACKET,
 444                                   RTE_PTYPE_INNER_L3_IPV6);
 445         else
 446                 pkt_type =
 447                         TRANSPOSE(flags,
 448                                   IBV_EXP_CQ_RX_IPV4_PACKET,
 449                                   RTE_PTYPE_L3_IPV4) |
 450                         TRANSPOSE(flags,
 451                                   IBV_EXP_CQ_RX_IPV6_PACKET,
 452                                   RTE_PTYPE_L3_IPV6);
 453         return pkt_type;
 454 }
 455
 456 /**
 457  * Translate RX completion flags to offload flags.
 458  *
 459  * @param[in] rxq
 460  *   Pointer to RX queue structure.
 461  * @param flags
 462  *   RX completion flags returned by poll_length_flags().
 463  *
 464  * @return
 465  *   Offload flags (ol_flags) for struct rte_mbuf.
 466  */
 467 static inline uint32_t
 468 rxq_cq_to_ol_flags(const struct rxq *rxq, uint32_t flags)
 469 {
 470         uint32_t ol_flags = 0;
 471
 472         if (rxq->csum) {
 473                 /* Set IP checksum flag only for IPv4/IPv6 packets. */
 474                 if (flags &
 475                     (IBV_EXP_CQ_RX_IPV4_PACKET | IBV_EXP_CQ_RX_IPV6_PACKET))
 476                         ol_flags |=
 477                                 TRANSPOSE(~flags,
 478                                         IBV_EXP_CQ_RX_IP_CSUM_OK,
 479                                         PKT_RX_IP_CKSUM_BAD);
 480 #ifdef HAVE_EXP_CQ_RX_TCP_PACKET
 481                 /* Set L4 checksum flag only for TCP/UDP packets. */
 482                 if (flags &
 483                     (IBV_EXP_CQ_RX_TCP_PACKET | IBV_EXP_CQ_RX_UDP_PACKET))
 484 #endif /* HAVE_EXP_CQ_RX_TCP_PACKET */
 485                         ol_flags |=
 486                                 TRANSPOSE(~flags,
 487                                         IBV_EXP_CQ_RX_TCP_UDP_CSUM_OK,
 488                                         PKT_RX_L4_CKSUM_BAD);
 489         }
 490         /*
 491          * PKT_RX_IP_CKSUM_BAD and PKT_RX_L4_CKSUM_BAD are used in place
 492          * of PKT_RX_EIP_CKSUM_BAD because the latter is not functional
 493          * (its value is 0).
 494          */
 495         if ((flags & IBV_EXP_CQ_RX_TUNNEL_PACKET) && (rxq->csum_l2tun))
 496                 ol_flags |=
 497                         TRANSPOSE(~flags,
 498                                   IBV_EXP_CQ_RX_OUTER_IP_CSUM_OK,
 499                                   PKT_RX_IP_CKSUM_BAD) |
 500                         TRANSPOSE(~flags,
 501                                   IBV_EXP_CQ_RX_OUTER_TCP_UDP_CSUM_OK,
 502                                   PKT_RX_L4_CKSUM_BAD);
 503         return ol_flags;
 504 }
 505
 506 /**
 507  * DPDK callback for RX.
 508  *
 509  * @param dpdk_rxq
 510  *   Generic pointer to RX queue structure.
 511  * @param[out] pkts
 512  *   Array to store received packets.
 513  * @param pkts_n
 514  *   Maximum number of packets in array.
 515  *
 516  * @return
 517  *   Number of packets successfully received (<= pkts_n).
 518  */
 519 uint16_t
 520 mlx5_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
 521 {
 522         struct rxq *rxq = (struct rxq *)dpdk_rxq;
 523         struct rxq_elt (*elts)[rxq->elts_n] = rxq->elts;
 524         const unsigned int elts_n = rxq->elts_n;
 525         unsigned int elts_head = rxq->elts_head;
 526         struct ibv_sge sges[pkts_n];
 527         unsigned int i;
 528         unsigned int pkts_ret = 0;
 529         int ret;
 530
 531         for (i = 0; (i != pkts_n); ++i) {
 532                 struct rxq_elt *elt = &(*elts)[elts_head];
 533                 unsigned int len;
 534                 struct rte_mbuf *seg = elt->buf;
 535                 struct rte_mbuf *rep;
 536                 uint32_t flags;
 537                 uint16_t vlan_tci;
 538
 539                 /* Sanity checks. */
 540                 assert(seg != NULL);
 541                 assert(elts_head < rxq->elts_n);
 542                 assert(rxq->elts_head < rxq->elts_n);
 543                 /*
 544                  * Fetch initial bytes of packet descriptor into a
 545                  * cacheline while allocating rep.
 546                  */
 547                 rte_mbuf_prefetch_part1(seg);
 548                 rte_mbuf_prefetch_part2(seg);
 549                 ret = rxq->poll(rxq->cq, NULL, NULL, &flags, &vlan_tci);
 550                 if (unlikely(ret < 0)) {
 551                         struct ibv_wc wc;
 552                         int wcs_n;
 553
 554                         DEBUG("rxq=%p, poll_length() failed (ret=%d)",
 555                               (void *)rxq, ret);
 556                         /* ibv_poll_cq() must be used in case of failure. */
 557                         wcs_n = ibv_poll_cq(rxq->cq, 1, &wc);
 558                         if (unlikely(wcs_n == 0))
 559                                 break;
 560                         if (unlikely(wcs_n < 0)) {
 561                                 DEBUG("rxq=%p, ibv_poll_cq() failed (wcs_n=%d)",
 562                                       (void *)rxq, wcs_n);
 563                                 break;
 564                         }
 565                         assert(wcs_n == 1);
 566                         if (unlikely(wc.status != IBV_WC_SUCCESS)) {
 567                                 /* Whatever, just repost the offending WR. */
 568                                 DEBUG("rxq=%p, wr_id=%" PRIu64 ": bad work"
 569                                       " completion status (%d): %s",
 570                                       (void *)rxq, wc.wr_id, wc.status,
 571                                       ibv_wc_status_str(wc.status));
 572 #ifdef MLX5_PMD_SOFT_COUNTERS
 573                                 /* Increment dropped packets counter. */
 574                                 ++rxq->stats.idropped;
 575 #endif
 576                                 /* Add SGE to array for repost. */
 577                                 sges[i] = elt->sge;
 578                                 goto repost;
 579                         }
 580                         ret = wc.byte_len;
 581                 }
 582                 if (ret == 0)
 583                         break;
 584                 assert(ret >= (rxq->crc_present << 2));
 585                 len = ret - (rxq->crc_present << 2);
 586                 rep = rte_mbuf_raw_alloc(rxq->mp);
 587                 if (unlikely(rep == NULL)) {
 588                         /*
 589                          * Unable to allocate a replacement mbuf,
 590                          * repost WR.
 591                          */
 592                         DEBUG("rxq=%p: can't allocate a new mbuf",
 593                               (void *)rxq);
 594                         /* Increment out of memory counters. */
 595                         ++rxq->stats.rx_nombuf;
 596                         ++rxq->priv->dev->data->rx_mbuf_alloc_failed;
 597                         goto repost;
 598                 }
 599
 600                 /* Reconfigure sge to use rep instead of seg. */
 601                 elt->sge.addr = (uintptr_t)rep->buf_addr + RTE_PKTMBUF_HEADROOM;
 602                 assert(elt->sge.lkey == rxq->mr->lkey);
 603                 elt->buf = rep;
 604
 605                 /* Add SGE to array for repost. */
 606                 sges[i] = elt->sge;
 607
 608                 /* Update seg information. */
 609                 SET_DATA_OFF(seg, RTE_PKTMBUF_HEADROOM);
 610                 NB_SEGS(seg) = 1;
 611                 PORT(seg) = rxq->port_id;
 612                 NEXT(seg) = NULL;
 613                 PKT_LEN(seg) = len;
 614                 DATA_LEN(seg) = len;
 615                 if (rxq->csum | rxq->csum_l2tun | rxq->vlan_strip) {
 616                         seg->packet_type = rxq_cq_to_pkt_type(flags);
 617                         seg->ol_flags = rxq_cq_to_ol_flags(rxq, flags);
 618 #ifdef HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS
 619                         if (flags & IBV_EXP_CQ_RX_CVLAN_STRIPPED_V1) {
 620                                 seg->ol_flags |= PKT_RX_VLAN_PKT |
 621                                         PKT_RX_VLAN_STRIPPED;
 622                                 seg->vlan_tci = vlan_tci;
 623                         }
 624 #endif /* HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS */
 625                 }
 626                 /* Return packet. */
 627                 *(pkts++) = seg;
 628                 ++pkts_ret;
 629 #ifdef MLX5_PMD_SOFT_COUNTERS
 630                 /* Increment bytes counter. */
 631                 rxq->stats.ibytes += len;
 632 #endif
 633 repost:
 634                 if (++elts_head >= elts_n)
 635                         elts_head = 0;
 636                 continue;
 637         }
 638         if (unlikely(i == 0))
 639                 return 0;
 640         /* Repost WRs. */
 641 #ifdef DEBUG_RECV
 642         DEBUG("%p: reposting %u WRs", (void *)rxq, i);
 643 #endif
 644         ret = rxq->recv(rxq->wq, sges, i);
 645         if (unlikely(ret)) {
 646                 /* Inability to repost WRs is fatal. */
 647                 DEBUG("%p: recv_burst(): failed (ret=%d)",
 648                       (void *)rxq->priv,
 649                       ret);
 650                 abort();
 651         }
 652         rxq->elts_head = elts_head;
 653 #ifdef MLX5_PMD_SOFT_COUNTERS
 654         /* Increment packets counter. */
 655         rxq->stats.ipackets += pkts_ret;
 656 #endif
 657         return pkts_ret;
 658 }
 659
 660 /**
 661  * Dummy DPDK callback for TX.
 662  *
 663  * This function is used to temporarily replace the real callback during
 664  * unsafe control operations on the queue, or in case of error.
 665  *
 666  * @param dpdk_txq
 667  *   Generic pointer to TX queue structure.
 668  * @param[in] pkts
 669  *   Packets to transmit.
 670  * @param pkts_n
 671  *   Number of packets in array.
 672  *
 673  * @return
 674  *   Number of packets successfully transmitted (<= pkts_n).
 675  */
 676 uint16_t
 677 removed_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
 678 {
 679         (void)dpdk_txq;
 680         (void)pkts;
 681         (void)pkts_n;
 682         return 0;
 683 }
 684
 685 /**
 686  * Dummy DPDK callback for RX.
 687  *
 688  * This function is used to temporarily replace the real callback during
 689  * unsafe control operations on the queue, or in case of error.
 690  *
 691  * @param dpdk_rxq
 692  *   Generic pointer to RX queue structure.
 693  * @param[out] pkts
 694  *   Array to store received packets.
 695  * @param pkts_n
 696  *   Maximum number of packets in array.
 697  *
 698  * @return
 699  *   Number of packets successfully received (<= pkts_n).
 700  */
 701 uint16_t
 702 removed_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
 703 {
 704         (void)dpdk_rxq;
 705         (void)pkts;
 706         (void)pkts_n;
 707         return 0;
 708 }