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 with scattered packets support.
 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_sp(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
 521 {
 522         struct rxq *rxq = (struct rxq *)dpdk_rxq;
 523         struct rxq_elt_sp (*elts)[rxq->elts_n] = rxq->elts.sp;
 524         const unsigned int elts_n = rxq->elts_n;
 525         unsigned int elts_head = rxq->elts_head;
 526         unsigned int i;
 527         unsigned int pkts_ret = 0;
 528         int ret;
 529
 530         if (unlikely(!rxq->sp))
 531                 return mlx5_rx_burst(dpdk_rxq, pkts, pkts_n);
 532         if (unlikely(elts == NULL)) /* See RTE_DEV_CMD_SET_MTU. */
 533                 return 0;
 534         for (i = 0; (i != pkts_n); ++i) {
 535                 struct rxq_elt_sp *elt = &(*elts)[elts_head];
 536                 unsigned int len;
 537                 unsigned int pkt_buf_len;
 538                 struct rte_mbuf *pkt_buf = NULL; /* Buffer returned in pkts. */
 539                 struct rte_mbuf **pkt_buf_next = &pkt_buf;
 540                 unsigned int seg_headroom = RTE_PKTMBUF_HEADROOM;
 541                 unsigned int j = 0;
 542                 uint32_t flags;
 543                 uint16_t vlan_tci;
 544
 545                 /* Sanity checks. */
 546                 assert(elts_head < rxq->elts_n);
 547                 assert(rxq->elts_head < rxq->elts_n);
 548                 ret = rxq->poll(rxq->cq, NULL, NULL, &flags, &vlan_tci);
 549                 if (unlikely(ret < 0)) {
 550                         struct ibv_wc wc;
 551                         int wcs_n;
 552
 553                         DEBUG("rxq=%p, poll_length() failed (ret=%d)",
 554                               (void *)rxq, ret);
 555                         /* ibv_poll_cq() must be used in case of failure. */
 556                         wcs_n = ibv_poll_cq(rxq->cq, 1, &wc);
 557                         if (unlikely(wcs_n == 0))
 558                                 break;
 559                         if (unlikely(wcs_n < 0)) {
 560                                 DEBUG("rxq=%p, ibv_poll_cq() failed (wcs_n=%d)",
 561                                       (void *)rxq, wcs_n);
 562                                 break;
 563                         }
 564                         assert(wcs_n == 1);
 565                         if (unlikely(wc.status != IBV_WC_SUCCESS)) {
 566                                 /* Whatever, just repost the offending WR. */
 567                                 DEBUG("rxq=%p, wr_id=%" PRIu64 ": bad work"
 568                                       " completion status (%d): %s",
 569                                       (void *)rxq, wc.wr_id, wc.status,
 570                                       ibv_wc_status_str(wc.status));
 571 #ifdef MLX5_PMD_SOFT_COUNTERS
 572                                 /* Increment dropped packets counter. */
 573                                 ++rxq->stats.idropped;
 574 #endif
 575                                 goto repost;
 576                         }
 577                         ret = wc.byte_len;
 578                 }
 579                 if (ret == 0)
 580                         break;
 581                 assert(ret >= (rxq->crc_present << 2));
 582                 len = ret - (rxq->crc_present << 2);
 583                 pkt_buf_len = len;
 584                 /*
 585                  * Replace spent segments with new ones, concatenate and
 586                  * return them as pkt_buf.
 587                  */
 588                 while (1) {
 589                         struct ibv_sge *sge = &elt->sges[j];
 590                         struct rte_mbuf *seg = elt->bufs[j];
 591                         struct rte_mbuf *rep;
 592                         unsigned int seg_tailroom;
 593
 594                         assert(seg != NULL);
 595                         /*
 596                          * Fetch initial bytes of packet descriptor into a
 597                          * cacheline while allocating rep.
 598                          */
 599                         rte_prefetch0(seg);
 600                         rep = rte_mbuf_raw_alloc(rxq->mp);
 601                         if (unlikely(rep == NULL)) {
 602                                 /*
 603                                  * Unable to allocate a replacement mbuf,
 604                                  * repost WR.
 605                                  */
 606                                 DEBUG("rxq=%p: can't allocate a new mbuf",
 607                                       (void *)rxq);
 608                                 if (pkt_buf != NULL) {
 609                                         *pkt_buf_next = NULL;
 610                                         rte_pktmbuf_free(pkt_buf);
 611                                 }
 612                                 /* Increment out of memory counters. */
 613                                 ++rxq->stats.rx_nombuf;
 614                                 ++rxq->priv->dev->data->rx_mbuf_alloc_failed;
 615                                 goto repost;
 616                         }
 617 #ifndef NDEBUG
 618                         /* Poison user-modifiable fields in rep. */
 619                         NEXT(rep) = (void *)((uintptr_t)-1);
 620                         SET_DATA_OFF(rep, 0xdead);
 621                         DATA_LEN(rep) = 0xd00d;
 622                         PKT_LEN(rep) = 0xdeadd00d;
 623                         NB_SEGS(rep) = 0x2a;
 624                         PORT(rep) = 0x2a;
 625                         rep->ol_flags = -1;
 626 #endif
 627                         assert(rep->buf_len == seg->buf_len);
 628                         /* Reconfigure sge to use rep instead of seg. */
 629                         assert(sge->lkey == rxq->mr->lkey);
 630                         sge->addr = ((uintptr_t)rep->buf_addr + seg_headroom);
 631                         elt->bufs[j] = rep;
 632                         ++j;
 633                         /* Update pkt_buf if it's the first segment, or link
 634                          * seg to the previous one and update pkt_buf_next. */
 635                         *pkt_buf_next = seg;
 636                         pkt_buf_next = &NEXT(seg);
 637                         /* Update seg information. */
 638                         seg_tailroom = (seg->buf_len - seg_headroom);
 639                         assert(sge->length == seg_tailroom);
 640                         SET_DATA_OFF(seg, seg_headroom);
 641                         if (likely(len <= seg_tailroom)) {
 642                                 /* Last segment. */
 643                                 DATA_LEN(seg) = len;
 644                                 PKT_LEN(seg) = len;
 645                                 /* Sanity check. */
 646                                 assert(rte_pktmbuf_headroom(seg) ==
 647                                        seg_headroom);
 648                                 assert(rte_pktmbuf_tailroom(seg) ==
 649                                        (seg_tailroom - len));
 650                                 break;
 651                         }
 652                         DATA_LEN(seg) = seg_tailroom;
 653                         PKT_LEN(seg) = seg_tailroom;
 654                         /* Sanity check. */
 655                         assert(rte_pktmbuf_headroom(seg) == seg_headroom);
 656                         assert(rte_pktmbuf_tailroom(seg) == 0);
 657                         /* Fix len and clear headroom for next segments. */
 658                         len -= seg_tailroom;
 659                         seg_headroom = 0;
 660                 }
 661                 /* Update head and tail segments. */
 662                 *pkt_buf_next = NULL;
 663                 assert(pkt_buf != NULL);
 664                 assert(j != 0);
 665                 NB_SEGS(pkt_buf) = j;
 666                 PORT(pkt_buf) = rxq->port_id;
 667                 PKT_LEN(pkt_buf) = pkt_buf_len;
 668                 if (rxq->csum | rxq->csum_l2tun | rxq->vlan_strip) {
 669                         pkt_buf->packet_type = rxq_cq_to_pkt_type(flags);
 670                         pkt_buf->ol_flags = rxq_cq_to_ol_flags(rxq, flags);
 671 #ifdef HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS
 672                         if (flags & IBV_EXP_CQ_RX_CVLAN_STRIPPED_V1) {
 673                                 pkt_buf->ol_flags |= PKT_RX_VLAN_PKT |
 674                                         PKT_RX_VLAN_STRIPPED;
 675                                 pkt_buf->vlan_tci = vlan_tci;
 676                         }
 677 #endif /* HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS */
 678                 }
 679
 680                 /* Return packet. */
 681                 *(pkts++) = pkt_buf;
 682                 ++pkts_ret;
 683 #ifdef MLX5_PMD_SOFT_COUNTERS
 684                 /* Increment bytes counter. */
 685                 rxq->stats.ibytes += pkt_buf_len;
 686 #endif
 687 repost:
 688                 ret = rxq->recv(rxq->wq, elt->sges, RTE_DIM(elt->sges));
 689                 if (unlikely(ret)) {
 690                         /* Inability to repost WRs is fatal. */
 691                         DEBUG("%p: recv_sg_list(): failed (ret=%d)",
 692                               (void *)rxq->priv,
 693                               ret);
 694                         abort();
 695                 }
 696                 if (++elts_head >= elts_n)
 697                         elts_head = 0;
 698                 continue;
 699         }
 700         if (unlikely(i == 0))
 701                 return 0;
 702         rxq->elts_head = elts_head;
 703 #ifdef MLX5_PMD_SOFT_COUNTERS
 704         /* Increment packets counter. */
 705         rxq->stats.ipackets += pkts_ret;
 706 #endif
 707         return pkts_ret;
 708 }
 709
 710 /**
 711  * DPDK callback for RX.
 712  *
 713  * The following function is the same as mlx5_rx_burst_sp(), except it doesn't
 714  * manage scattered packets. Improves performance when MRU is lower than the
 715  * size of the first segment.
 716  *
 717  * @param dpdk_rxq
 718  *   Generic pointer to RX queue structure.
 719  * @param[out] pkts
 720  *   Array to store received packets.
 721  * @param pkts_n
 722  *   Maximum number of packets in array.
 723  *
 724  * @return
 725  *   Number of packets successfully received (<= pkts_n).
 726  */
 727 uint16_t
 728 mlx5_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
 729 {
 730         struct rxq *rxq = (struct rxq *)dpdk_rxq;
 731         struct rxq_elt (*elts)[rxq->elts_n] = rxq->elts.no_sp;
 732         const unsigned int elts_n = rxq->elts_n;
 733         unsigned int elts_head = rxq->elts_head;
 734         struct ibv_sge sges[pkts_n];
 735         unsigned int i;
 736         unsigned int pkts_ret = 0;
 737         int ret;
 738
 739         if (unlikely(rxq->sp))
 740                 return mlx5_rx_burst_sp(dpdk_rxq, pkts, pkts_n);
 741         for (i = 0; (i != pkts_n); ++i) {
 742                 struct rxq_elt *elt = &(*elts)[elts_head];
 743                 unsigned int len;
 744                 struct rte_mbuf *seg = elt->buf;
 745                 struct rte_mbuf *rep;
 746                 uint32_t flags;
 747                 uint16_t vlan_tci;
 748
 749                 /* Sanity checks. */
 750                 assert(seg != NULL);
 751                 assert(elts_head < rxq->elts_n);
 752                 assert(rxq->elts_head < rxq->elts_n);
 753                 /*
 754                  * Fetch initial bytes of packet descriptor into a
 755                  * cacheline while allocating rep.
 756                  */
 757                 rte_mbuf_prefetch_part1(seg);
 758                 rte_mbuf_prefetch_part2(seg);
 759                 ret = rxq->poll(rxq->cq, NULL, NULL, &flags, &vlan_tci);
 760                 if (unlikely(ret < 0)) {
 761                         struct ibv_wc wc;
 762                         int wcs_n;
 763
 764                         DEBUG("rxq=%p, poll_length() failed (ret=%d)",
 765                               (void *)rxq, ret);
 766                         /* ibv_poll_cq() must be used in case of failure. */
 767                         wcs_n = ibv_poll_cq(rxq->cq, 1, &wc);
 768                         if (unlikely(wcs_n == 0))
 769                                 break;
 770                         if (unlikely(wcs_n < 0)) {
 771                                 DEBUG("rxq=%p, ibv_poll_cq() failed (wcs_n=%d)",
 772                                       (void *)rxq, wcs_n);
 773                                 break;
 774                         }
 775                         assert(wcs_n == 1);
 776                         if (unlikely(wc.status != IBV_WC_SUCCESS)) {
 777                                 /* Whatever, just repost the offending WR. */
 778                                 DEBUG("rxq=%p, wr_id=%" PRIu64 ": bad work"
 779                                       " completion status (%d): %s",
 780                                       (void *)rxq, wc.wr_id, wc.status,
 781                                       ibv_wc_status_str(wc.status));
 782 #ifdef MLX5_PMD_SOFT_COUNTERS
 783                                 /* Increment dropped packets counter. */
 784                                 ++rxq->stats.idropped;
 785 #endif
 786                                 /* Add SGE to array for repost. */
 787                                 sges[i] = elt->sge;
 788                                 goto repost;
 789                         }
 790                         ret = wc.byte_len;
 791                 }
 792                 if (ret == 0)
 793                         break;
 794                 assert(ret >= (rxq->crc_present << 2));
 795                 len = ret - (rxq->crc_present << 2);
 796                 rep = rte_mbuf_raw_alloc(rxq->mp);
 797                 if (unlikely(rep == NULL)) {
 798                         /*
 799                          * Unable to allocate a replacement mbuf,
 800                          * repost WR.
 801                          */
 802                         DEBUG("rxq=%p: can't allocate a new mbuf",
 803                               (void *)rxq);
 804                         /* Increment out of memory counters. */
 805                         ++rxq->stats.rx_nombuf;
 806                         ++rxq->priv->dev->data->rx_mbuf_alloc_failed;
 807                         goto repost;
 808                 }
 809
 810                 /* Reconfigure sge to use rep instead of seg. */
 811                 elt->sge.addr = (uintptr_t)rep->buf_addr + RTE_PKTMBUF_HEADROOM;
 812                 assert(elt->sge.lkey == rxq->mr->lkey);
 813                 elt->buf = rep;
 814
 815                 /* Add SGE to array for repost. */
 816                 sges[i] = elt->sge;
 817
 818                 /* Update seg information. */
 819                 SET_DATA_OFF(seg, RTE_PKTMBUF_HEADROOM);
 820                 NB_SEGS(seg) = 1;
 821                 PORT(seg) = rxq->port_id;
 822                 NEXT(seg) = NULL;
 823                 PKT_LEN(seg) = len;
 824                 DATA_LEN(seg) = len;
 825                 if (rxq->csum | rxq->csum_l2tun | rxq->vlan_strip) {
 826                         seg->packet_type = rxq_cq_to_pkt_type(flags);
 827                         seg->ol_flags = rxq_cq_to_ol_flags(rxq, flags);
 828 #ifdef HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS
 829                         if (flags & IBV_EXP_CQ_RX_CVLAN_STRIPPED_V1) {
 830                                 seg->ol_flags |= PKT_RX_VLAN_PKT |
 831                                         PKT_RX_VLAN_STRIPPED;
 832                                 seg->vlan_tci = vlan_tci;
 833                         }
 834 #endif /* HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS */
 835                 }
 836                 /* Return packet. */
 837                 *(pkts++) = seg;
 838                 ++pkts_ret;
 839 #ifdef MLX5_PMD_SOFT_COUNTERS
 840                 /* Increment bytes counter. */
 841                 rxq->stats.ibytes += len;
 842 #endif
 843 repost:
 844                 if (++elts_head >= elts_n)
 845                         elts_head = 0;
 846                 continue;
 847         }
 848         if (unlikely(i == 0))
 849                 return 0;
 850         /* Repost WRs. */
 851 #ifdef DEBUG_RECV
 852         DEBUG("%p: reposting %u WRs", (void *)rxq, i);
 853 #endif
 854         ret = rxq->recv(rxq->wq, sges, i);
 855         if (unlikely(ret)) {
 856                 /* Inability to repost WRs is fatal. */
 857                 DEBUG("%p: recv_burst(): failed (ret=%d)",
 858                       (void *)rxq->priv,
 859                       ret);
 860                 abort();
 861         }
 862         rxq->elts_head = elts_head;
 863 #ifdef MLX5_PMD_SOFT_COUNTERS
 864         /* Increment packets counter. */
 865         rxq->stats.ipackets += pkts_ret;
 866 #endif
 867         return pkts_ret;
 868 }
 869
 870 /**
 871  * Dummy DPDK callback for TX.
 872  *
 873  * This function is used to temporarily replace the real callback during
 874  * unsafe control operations on the queue, or in case of error.
 875  *
 876  * @param dpdk_txq
 877  *   Generic pointer to TX queue structure.
 878  * @param[in] pkts
 879  *   Packets to transmit.
 880  * @param pkts_n
 881  *   Number of packets in array.
 882  *
 883  * @return
 884  *   Number of packets successfully transmitted (<= pkts_n).
 885  */
 886 uint16_t
 887 removed_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
 888 {
 889         (void)dpdk_txq;
 890         (void)pkts;
 891         (void)pkts_n;
 892         return 0;
 893 }
 894
 895 /**
 896  * Dummy DPDK callback for RX.
 897  *
 898  * This function is used to temporarily replace the real callback during
 899  * unsafe control operations on the queue, or in case of error.
 900  *
 901  * @param dpdk_rxq
 902  *   Generic pointer to RX queue structure.
 903  * @param[out] pkts
 904  *   Array to store received packets.
 905  * @param pkts_n
 906  *   Maximum number of packets in array.
 907  *
 908  * @return
 909  *   Number of packets successfully received (<= pkts_n).
 910  */
 911 uint16_t
 912 removed_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
 913 {
 914         (void)dpdk_rxq;
 915         (void)pkts;
 916         (void)pkts_n;
 917         return 0;
 918 }