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 #ifdef PEDANTIC
  59 #pragma GCC diagnostic error "-pedantic"
  60 #endif
  61
  62 #include "mlx5.h"
  63 #include "mlx5_utils.h"
  64 #include "mlx5_rxtx.h"
  65 #include "mlx5_defs.h"
  66
  67 /**
  68  * Manage TX completions.
  69  *
  70  * When sending a burst, mlx5_tx_burst() posts several WRs.
  71  * To improve performance, a completion event is only required once every
  72  * MLX5_PMD_TX_PER_COMP_REQ sends. Doing so discards completion information
  73  * for other WRs, but this information would not be used anyway.
  74  *
  75  * @param txq
  76  *   Pointer to TX queue structure.
  77  *
  78  * @return
  79  *   0 on success, -1 on failure.
  80  */
  81 static int
  82 txq_complete(struct txq *txq)
  83 {
  84         unsigned int elts_comp = txq->elts_comp;
  85         unsigned int elts_tail = txq->elts_tail;
  86         const unsigned int elts_n = txq->elts_n;
  87         int wcs_n;
  88
  89         if (unlikely(elts_comp == 0))
  90                 return 0;
  91 #ifdef DEBUG_SEND
  92         DEBUG("%p: processing %u work requests completions",
  93               (void *)txq, elts_comp);
  94 #endif
  95         wcs_n = txq->if_cq->poll_cnt(txq->cq, elts_comp);
  96         if (unlikely(wcs_n == 0))
  97                 return 0;
  98         if (unlikely(wcs_n < 0)) {
  99                 DEBUG("%p: ibv_poll_cq() failed (wcs_n=%d)",
 100                       (void *)txq, wcs_n);
 101                 return -1;
 102         }
 103         elts_comp -= wcs_n;
 104         assert(elts_comp <= txq->elts_comp);
 105         /*
 106          * Assume WC status is successful as nothing can be done about it
 107          * anyway.
 108          */
 109         elts_tail += wcs_n * txq->elts_comp_cd_init;
 110         if (elts_tail >= elts_n)
 111                 elts_tail -= elts_n;
 112         txq->elts_tail = elts_tail;
 113         txq->elts_comp = elts_comp;
 114         return 0;
 115 }
 116
 117 /**
 118  * Get Memory Pool (MP) from mbuf. If mbuf is indirect, the pool from which
 119  * the cloned mbuf is allocated is returned instead.
 120  *
 121  * @param buf
 122  *   Pointer to mbuf.
 123  *
 124  * @return
 125  *   Memory pool where data is located for given mbuf.
 126  */
 127 static struct rte_mempool *
 128 txq_mb2mp(struct rte_mbuf *buf)
 129 {
 130         if (unlikely(RTE_MBUF_INDIRECT(buf)))
 131                 return rte_mbuf_from_indirect(buf)->pool;
 132         return buf->pool;
 133 }
 134
 135 /**
 136  * Get Memory Region (MR) <-> Memory Pool (MP) association from txq->mp2mr[].
 137  * Add MP to txq->mp2mr[] if it's not registered yet. If mp2mr[] is full,
 138  * remove an entry first.
 139  *
 140  * @param txq
 141  *   Pointer to TX queue structure.
 142  * @param[in] mp
 143  *   Memory Pool for which a Memory Region lkey must be returned.
 144  *
 145  * @return
 146  *   mr->lkey on success, (uint32_t)-1 on failure.
 147  */
 148 static uint32_t
 149 txq_mp2mr(struct txq *txq, struct rte_mempool *mp)
 150 {
 151         unsigned int i;
 152         struct ibv_mr *mr;
 153
 154         for (i = 0; (i != RTE_DIM(txq->mp2mr)); ++i) {
 155                 if (unlikely(txq->mp2mr[i].mp == NULL)) {
 156                         /* Unknown MP, add a new MR for it. */
 157                         break;
 158                 }
 159                 if (txq->mp2mr[i].mp == mp) {
 160                         assert(txq->mp2mr[i].lkey != (uint32_t)-1);
 161                         assert(txq->mp2mr[i].mr->lkey == txq->mp2mr[i].lkey);
 162                         return txq->mp2mr[i].lkey;
 163                 }
 164         }
 165         /* Add a new entry, register MR first. */
 166         DEBUG("%p: discovered new memory pool %p", (void *)txq, (void *)mp);
 167         mr = ibv_reg_mr(txq->priv->pd,
 168                         (void *)mp->elt_va_start,
 169                         (mp->elt_va_end - mp->elt_va_start),
 170                         (IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE));
 171         if (unlikely(mr == NULL)) {
 172                 DEBUG("%p: unable to configure MR, ibv_reg_mr() failed.",
 173                       (void *)txq);
 174                 return (uint32_t)-1;
 175         }
 176         if (unlikely(i == RTE_DIM(txq->mp2mr))) {
 177                 /* Table is full, remove oldest entry. */
 178                 DEBUG("%p: MR <-> MP table full, dropping oldest entry.",
 179                       (void *)txq);
 180                 --i;
 181                 claim_zero(ibv_dereg_mr(txq->mp2mr[i].mr));
 182                 memmove(&txq->mp2mr[0], &txq->mp2mr[1],
 183                         (sizeof(txq->mp2mr) - sizeof(txq->mp2mr[0])));
 184         }
 185         /* Store the new entry. */
 186         txq->mp2mr[i].mp = mp;
 187         txq->mp2mr[i].mr = mr;
 188         txq->mp2mr[i].lkey = mr->lkey;
 189         DEBUG("%p: new MR lkey for MP %p: 0x%08" PRIu32,
 190               (void *)txq, (void *)mp, txq->mp2mr[i].lkey);
 191         return txq->mp2mr[i].lkey;
 192 }
 193
 194 #if MLX5_PMD_SGE_WR_N > 1
 195
 196 /**
 197  * Copy scattered mbuf contents to a single linear buffer.
 198  *
 199  * @param[out] linear
 200  *   Linear output buffer.
 201  * @param[in] buf
 202  *   Scattered input buffer.
 203  *
 204  * @return
 205  *   Number of bytes copied to the output buffer or 0 if not large enough.
 206  */
 207 static unsigned int
 208 linearize_mbuf(linear_t *linear, struct rte_mbuf *buf)
 209 {
 210         unsigned int size = 0;
 211         unsigned int offset;
 212
 213         do {
 214                 unsigned int len = DATA_LEN(buf);
 215
 216                 offset = size;
 217                 size += len;
 218                 if (unlikely(size > sizeof(*linear)))
 219                         return 0;
 220                 memcpy(&(*linear)[offset],
 221                        rte_pktmbuf_mtod(buf, uint8_t *),
 222                        len);
 223                 buf = NEXT(buf);
 224         } while (buf != NULL);
 225         return size;
 226 }
 227
 228 /**
 229  * Handle scattered buffers for mlx5_tx_burst().
 230  *
 231  * @param txq
 232  *   TX queue structure.
 233  * @param segs
 234  *   Number of segments in buf.
 235  * @param elt
 236  *   TX queue element to fill.
 237  * @param[in] buf
 238  *   Buffer to process.
 239  * @param elts_head
 240  *   Index of the linear buffer to use if necessary (normally txq->elts_head).
 241  * @param[out] sges
 242  *   Array filled with SGEs on success.
 243  *
 244  * @return
 245  *   A structure containing the processed packet size in bytes and the
 246  *   number of SGEs. Both fields are set to (unsigned int)-1 in case of
 247  *   failure.
 248  */
 249 static struct tx_burst_sg_ret {
 250         unsigned int length;
 251         unsigned int num;
 252 }
 253 tx_burst_sg(struct txq *txq, unsigned int segs, struct txq_elt *elt,
 254             struct rte_mbuf *buf, unsigned int elts_head,
 255             struct ibv_sge (*sges)[MLX5_PMD_SGE_WR_N])
 256 {
 257         unsigned int sent_size = 0;
 258         unsigned int j;
 259         int linearize = 0;
 260
 261         /* When there are too many segments, extra segments are
 262          * linearized in the last SGE. */
 263         if (unlikely(segs > RTE_DIM(*sges))) {
 264                 segs = (RTE_DIM(*sges) - 1);
 265                 linearize = 1;
 266         }
 267         /* Update element. */
 268         elt->buf = buf;
 269         /* Register segments as SGEs. */
 270         for (j = 0; (j != segs); ++j) {
 271                 struct ibv_sge *sge = &(*sges)[j];
 272                 uint32_t lkey;
 273
 274                 /* Retrieve Memory Region key for this memory pool. */
 275                 lkey = txq_mp2mr(txq, txq_mb2mp(buf));
 276                 if (unlikely(lkey == (uint32_t)-1)) {
 277                         /* MR does not exist. */
 278                         DEBUG("%p: unable to get MP <-> MR association",
 279                               (void *)txq);
 280                         /* Clean up TX element. */
 281                         elt->buf = NULL;
 282                         goto stop;
 283                 }
 284                 /* Update SGE. */
 285                 sge->addr = rte_pktmbuf_mtod(buf, uintptr_t);
 286                 if (txq->priv->vf)
 287                         rte_prefetch0((volatile void *)
 288                                       (uintptr_t)sge->addr);
 289                 sge->length = DATA_LEN(buf);
 290                 sge->lkey = lkey;
 291                 sent_size += sge->length;
 292                 buf = NEXT(buf);
 293         }
 294         /* If buf is not NULL here and is not going to be linearized,
 295          * nb_segs is not valid. */
 296         assert(j == segs);
 297         assert((buf == NULL) || (linearize));
 298         /* Linearize extra segments. */
 299         if (linearize) {
 300                 struct ibv_sge *sge = &(*sges)[segs];
 301                 linear_t *linear = &(*txq->elts_linear)[elts_head];
 302                 unsigned int size = linearize_mbuf(linear, buf);
 303
 304                 assert(segs == (RTE_DIM(*sges) - 1));
 305                 if (size == 0) {
 306                         /* Invalid packet. */
 307                         DEBUG("%p: packet too large to be linearized.",
 308                               (void *)txq);
 309                         /* Clean up TX element. */
 310                         elt->buf = NULL;
 311                         goto stop;
 312                 }
 313                 /* If MLX5_PMD_SGE_WR_N is 1, free mbuf immediately. */
 314                 if (RTE_DIM(*sges) == 1) {
 315                         do {
 316                                 struct rte_mbuf *next = NEXT(buf);
 317
 318                                 rte_pktmbuf_free_seg(buf);
 319                                 buf = next;
 320                         } while (buf != NULL);
 321                         elt->buf = NULL;
 322                 }
 323                 /* Update SGE. */
 324                 sge->addr = (uintptr_t)&(*linear)[0];
 325                 sge->length = size;
 326                 sge->lkey = txq->mr_linear->lkey;
 327                 sent_size += size;
 328                 /* Include last segment. */
 329                 segs++;
 330         }
 331         return (struct tx_burst_sg_ret){
 332                 .length = sent_size,
 333                 .num = segs,
 334         };
 335 stop:
 336         return (struct tx_burst_sg_ret){
 337                 .length = -1,
 338                 .num = -1,
 339         };
 340 }
 341
 342 #endif /* MLX5_PMD_SGE_WR_N > 1 */
 343
 344 /**
 345  * DPDK callback for TX.
 346  *
 347  * @param dpdk_txq
 348  *   Generic pointer to TX queue structure.
 349  * @param[in] pkts
 350  *   Packets to transmit.
 351  * @param pkts_n
 352  *   Number of packets in array.
 353  *
 354  * @return
 355  *   Number of packets successfully transmitted (<= pkts_n).
 356  */
 357 uint16_t
 358 mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
 359 {
 360         struct txq *txq = (struct txq *)dpdk_txq;
 361         unsigned int elts_head = txq->elts_head;
 362         const unsigned int elts_tail = txq->elts_tail;
 363         const unsigned int elts_n = txq->elts_n;
 364         unsigned int elts_comp_cd = txq->elts_comp_cd;
 365         unsigned int elts_comp = 0;
 366         unsigned int i;
 367         unsigned int max;
 368         int err;
 369
 370         assert(elts_comp_cd != 0);
 371         txq_complete(txq);
 372         max = (elts_n - (elts_head - elts_tail));
 373         if (max > elts_n)
 374                 max -= elts_n;
 375         assert(max >= 1);
 376         assert(max <= elts_n);
 377         /* Always leave one free entry in the ring. */
 378         --max;
 379         if (max == 0)
 380                 return 0;
 381         if (max > pkts_n)
 382                 max = pkts_n;
 383         for (i = 0; (i != max); ++i) {
 384                 struct rte_mbuf *buf = pkts[i];
 385                 unsigned int elts_head_next =
 386                         (((elts_head + 1) == elts_n) ? 0 : elts_head + 1);
 387                 struct txq_elt *elt_next = &(*txq->elts)[elts_head_next];
 388                 struct txq_elt *elt = &(*txq->elts)[elts_head];
 389                 unsigned int segs = NB_SEGS(buf);
 390 #ifdef MLX5_PMD_SOFT_COUNTERS
 391                 unsigned int sent_size = 0;
 392 #endif
 393                 uint32_t send_flags = 0;
 394
 395                 /* Clean up old buffer. */
 396                 if (likely(elt->buf != NULL)) {
 397                         struct rte_mbuf *tmp = elt->buf;
 398
 399                         /* Faster than rte_pktmbuf_free(). */
 400                         do {
 401                                 struct rte_mbuf *next = NEXT(tmp);
 402
 403                                 rte_pktmbuf_free_seg(tmp);
 404                                 tmp = next;
 405                         } while (tmp != NULL);
 406                 }
 407                 /* Request TX completion. */
 408                 if (unlikely(--elts_comp_cd == 0)) {
 409                         elts_comp_cd = txq->elts_comp_cd_init;
 410                         ++elts_comp;
 411                         send_flags |= IBV_EXP_QP_BURST_SIGNALED;
 412                 }
 413                 /* Should we enable HW CKSUM offload */
 414                 if (buf->ol_flags &
 415                     (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM)) {
 416                         send_flags |= IBV_EXP_QP_BURST_IP_CSUM;
 417                         /* HW does not support checksum offloads at arbitrary
 418                          * offsets but automatically recognizes the packet
 419                          * type. For inner L3/L4 checksums, only VXLAN (UDP)
 420                          * tunnels are currently supported. */
 421                         if (RTE_ETH_IS_TUNNEL_PKT(buf->packet_type))
 422                                 send_flags |= IBV_EXP_QP_BURST_TUNNEL;
 423                 }
 424                 if (likely(segs == 1)) {
 425                         uintptr_t addr;
 426                         uint32_t length;
 427                         uint32_t lkey;
 428
 429                         /* Retrieve buffer information. */
 430                         addr = rte_pktmbuf_mtod(buf, uintptr_t);
 431                         length = DATA_LEN(buf);
 432                         /* Retrieve Memory Region key for this memory pool. */
 433                         lkey = txq_mp2mr(txq, txq_mb2mp(buf));
 434                         if (unlikely(lkey == (uint32_t)-1)) {
 435                                 /* MR does not exist. */
 436                                 DEBUG("%p: unable to get MP <-> MR"
 437                                       " association", (void *)txq);
 438                                 /* Clean up TX element. */
 439                                 elt->buf = NULL;
 440                                 goto stop;
 441                         }
 442                         /* Update element. */
 443                         elt->buf = buf;
 444                         if (txq->priv->vf)
 445                                 rte_prefetch0((volatile void *)
 446                                               (uintptr_t)addr);
 447                         RTE_MBUF_PREFETCH_TO_FREE(elt_next->buf);
 448                         /* Put packet into send queue. */
 449 #if MLX5_PMD_MAX_INLINE > 0
 450                         if (length <= txq->max_inline)
 451                                 err = txq->if_qp->send_pending_inline
 452                                         (txq->qp,
 453                                          (void *)addr,
 454                                          length,
 455                                          send_flags);
 456                         else
 457 #endif
 458                                 err = txq->if_qp->send_pending
 459                                         (txq->qp,
 460                                          addr,
 461                                          length,
 462                                          lkey,
 463                                          send_flags);
 464                         if (unlikely(err))
 465                                 goto stop;
 466 #ifdef MLX5_PMD_SOFT_COUNTERS
 467                         sent_size += length;
 468 #endif
 469                 } else {
 470 #if MLX5_PMD_SGE_WR_N > 1
 471                         struct ibv_sge sges[MLX5_PMD_SGE_WR_N];
 472                         struct tx_burst_sg_ret ret;
 473
 474                         ret = tx_burst_sg(txq, segs, elt, buf, elts_head,
 475                                           &sges);
 476                         if (ret.length == (unsigned int)-1)
 477                                 goto stop;
 478                         RTE_MBUF_PREFETCH_TO_FREE(elt_next->buf);
 479                         /* Put SG list into send queue. */
 480                         err = txq->if_qp->send_pending_sg_list
 481                                 (txq->qp,
 482                                  sges,
 483                                  ret.num,
 484                                  send_flags);
 485                         if (unlikely(err))
 486                                 goto stop;
 487 #ifdef MLX5_PMD_SOFT_COUNTERS
 488                         sent_size += ret.length;
 489 #endif
 490 #else /* MLX5_PMD_SGE_WR_N > 1 */
 491                         DEBUG("%p: TX scattered buffers support not"
 492                               " compiled in", (void *)txq);
 493                         goto stop;
 494 #endif /* MLX5_PMD_SGE_WR_N > 1 */
 495                 }
 496                 elts_head = elts_head_next;
 497 #ifdef MLX5_PMD_SOFT_COUNTERS
 498                 /* Increment sent bytes counter. */
 499                 txq->stats.obytes += sent_size;
 500 #endif
 501         }
 502 stop:
 503         /* Take a shortcut if nothing must be sent. */
 504         if (unlikely(i == 0))
 505                 return 0;
 506 #ifdef MLX5_PMD_SOFT_COUNTERS
 507         /* Increment sent packets counter. */
 508         txq->stats.opackets += i;
 509 #endif
 510         /* Ring QP doorbell. */
 511         err = txq->if_qp->send_flush(txq->qp);
 512         if (unlikely(err)) {
 513                 /* A nonzero value is not supposed to be returned.
 514                  * Nothing can be done about it. */
 515                 DEBUG("%p: send_flush() failed with error %d",
 516                       (void *)txq, err);
 517         }
 518         txq->elts_head = elts_head;
 519         txq->elts_comp += elts_comp;
 520         txq->elts_comp_cd = elts_comp_cd;
 521         return i;
 522 }
 523
 524 /**
 525  * Translate RX completion flags to packet type.
 526  *
 527  * @param flags
 528  *   RX completion flags returned by poll_length_flags().
 529  *
 530  * @return
 531  *   Packet type for struct rte_mbuf.
 532  */
 533 static inline uint32_t
 534 rxq_cq_to_pkt_type(uint32_t flags)
 535 {
 536         uint32_t pkt_type;
 537
 538         if (flags & IBV_EXP_CQ_RX_TUNNEL_PACKET)
 539                 pkt_type =
 540                         TRANSPOSE(flags,
 541                                   IBV_EXP_CQ_RX_OUTER_IPV4_PACKET,
 542                                   RTE_PTYPE_L3_IPV4) |
 543                         TRANSPOSE(flags,
 544                                   IBV_EXP_CQ_RX_OUTER_IPV6_PACKET,
 545                                   RTE_PTYPE_L3_IPV6) |
 546                         TRANSPOSE(flags,
 547                                   IBV_EXP_CQ_RX_IPV4_PACKET,
 548                                   RTE_PTYPE_INNER_L3_IPV4) |
 549                         TRANSPOSE(flags,
 550                                   IBV_EXP_CQ_RX_IPV6_PACKET,
 551                                   RTE_PTYPE_INNER_L3_IPV6);
 552         else
 553                 pkt_type =
 554                         TRANSPOSE(flags,
 555                                   IBV_EXP_CQ_RX_IPV4_PACKET,
 556                                   RTE_PTYPE_L3_IPV4) |
 557                         TRANSPOSE(flags,
 558                                   IBV_EXP_CQ_RX_IPV6_PACKET,
 559                                   RTE_PTYPE_L3_IPV6);
 560         return pkt_type;
 561 }
 562
 563 /**
 564  * Translate RX completion flags to offload flags.
 565  *
 566  * @param[in] rxq
 567  *   Pointer to RX queue structure.
 568  * @param flags
 569  *   RX completion flags returned by poll_length_flags().
 570  *
 571  * @return
 572  *   Offload flags (ol_flags) for struct rte_mbuf.
 573  */
 574 static inline uint32_t
 575 rxq_cq_to_ol_flags(const struct rxq *rxq, uint32_t flags)
 576 {
 577         uint32_t ol_flags = 0;
 578
 579         if (rxq->csum)
 580                 ol_flags |=
 581                         TRANSPOSE(~flags,
 582                                   IBV_EXP_CQ_RX_IP_CSUM_OK,
 583                                   PKT_RX_IP_CKSUM_BAD) |
 584                         TRANSPOSE(~flags,
 585                                   IBV_EXP_CQ_RX_TCP_UDP_CSUM_OK,
 586                                   PKT_RX_L4_CKSUM_BAD);
 587         /*
 588          * PKT_RX_IP_CKSUM_BAD and PKT_RX_L4_CKSUM_BAD are used in place
 589          * of PKT_RX_EIP_CKSUM_BAD because the latter is not functional
 590          * (its value is 0).
 591          */
 592         if ((flags & IBV_EXP_CQ_RX_TUNNEL_PACKET) && (rxq->csum_l2tun))
 593                 ol_flags |=
 594                         TRANSPOSE(~flags,
 595                                   IBV_EXP_CQ_RX_OUTER_IP_CSUM_OK,
 596                                   PKT_RX_IP_CKSUM_BAD) |
 597                         TRANSPOSE(~flags,
 598                                   IBV_EXP_CQ_RX_OUTER_TCP_UDP_CSUM_OK,
 599                                   PKT_RX_L4_CKSUM_BAD);
 600         return ol_flags;
 601 }
 602
 603 /**
 604  * DPDK callback for RX with scattered packets support.
 605  *
 606  * @param dpdk_rxq
 607  *   Generic pointer to RX queue structure.
 608  * @param[out] pkts
 609  *   Array to store received packets.
 610  * @param pkts_n
 611  *   Maximum number of packets in array.
 612  *
 613  * @return
 614  *   Number of packets successfully received (<= pkts_n).
 615  */
 616 uint16_t
 617 mlx5_rx_burst_sp(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
 618 {
 619         struct rxq *rxq = (struct rxq *)dpdk_rxq;
 620         struct rxq_elt_sp (*elts)[rxq->elts_n] = rxq->elts.sp;
 621         const unsigned int elts_n = rxq->elts_n;
 622         unsigned int elts_head = rxq->elts_head;
 623         unsigned int i;
 624         unsigned int pkts_ret = 0;
 625         int ret;
 626
 627         if (unlikely(!rxq->sp))
 628                 return mlx5_rx_burst(dpdk_rxq, pkts, pkts_n);
 629         if (unlikely(elts == NULL)) /* See RTE_DEV_CMD_SET_MTU. */
 630                 return 0;
 631         for (i = 0; (i != pkts_n); ++i) {
 632                 struct rxq_elt_sp *elt = &(*elts)[elts_head];
 633                 unsigned int len;
 634                 unsigned int pkt_buf_len;
 635                 struct rte_mbuf *pkt_buf = NULL; /* Buffer returned in pkts. */
 636                 struct rte_mbuf **pkt_buf_next = &pkt_buf;
 637                 unsigned int seg_headroom = RTE_PKTMBUF_HEADROOM;
 638                 unsigned int j = 0;
 639                 uint32_t flags;
 640
 641                 /* Sanity checks. */
 642                 assert(elts_head < rxq->elts_n);
 643                 assert(rxq->elts_head < rxq->elts_n);
 644                 ret = rxq->if_cq->poll_length_flags(rxq->cq, NULL, NULL,
 645                                                     &flags);
 646                 if (unlikely(ret < 0)) {
 647                         struct ibv_wc wc;
 648                         int wcs_n;
 649
 650                         DEBUG("rxq=%p, poll_length() failed (ret=%d)",
 651                               (void *)rxq, ret);
 652                         /* ibv_poll_cq() must be used in case of failure. */
 653                         wcs_n = ibv_poll_cq(rxq->cq, 1, &wc);
 654                         if (unlikely(wcs_n == 0))
 655                                 break;
 656                         if (unlikely(wcs_n < 0)) {
 657                                 DEBUG("rxq=%p, ibv_poll_cq() failed (wcs_n=%d)",
 658                                       (void *)rxq, wcs_n);
 659                                 break;
 660                         }
 661                         assert(wcs_n == 1);
 662                         if (unlikely(wc.status != IBV_WC_SUCCESS)) {
 663                                 /* Whatever, just repost the offending WR. */
 664                                 DEBUG("rxq=%p, wr_id=%" PRIu64 ": bad work"
 665                                       " completion status (%d): %s",
 666                                       (void *)rxq, wc.wr_id, wc.status,
 667                                       ibv_wc_status_str(wc.status));
 668 #ifdef MLX5_PMD_SOFT_COUNTERS
 669                                 /* Increment dropped packets counter. */
 670                                 ++rxq->stats.idropped;
 671 #endif
 672                                 goto repost;
 673                         }
 674                         ret = wc.byte_len;
 675                 }
 676                 if (ret == 0)
 677                         break;
 678                 len = ret;
 679                 pkt_buf_len = len;
 680                 /*
 681                  * Replace spent segments with new ones, concatenate and
 682                  * return them as pkt_buf.
 683                  */
 684                 while (1) {
 685                         struct ibv_sge *sge = &elt->sges[j];
 686                         struct rte_mbuf *seg = elt->bufs[j];
 687                         struct rte_mbuf *rep;
 688                         unsigned int seg_tailroom;
 689
 690                         assert(seg != NULL);
 691                         /*
 692                          * Fetch initial bytes of packet descriptor into a
 693                          * cacheline while allocating rep.
 694                          */
 695                         rte_prefetch0(seg);
 696                         rep = __rte_mbuf_raw_alloc(rxq->mp);
 697                         if (unlikely(rep == NULL)) {
 698                                 /*
 699                                  * Unable to allocate a replacement mbuf,
 700                                  * repost WR.
 701                                  */
 702                                 DEBUG("rxq=%p: can't allocate a new mbuf",
 703                                       (void *)rxq);
 704                                 if (pkt_buf != NULL) {
 705                                         *pkt_buf_next = NULL;
 706                                         rte_pktmbuf_free(pkt_buf);
 707                                 }
 708                                 /* Increment out of memory counters. */
 709                                 ++rxq->stats.rx_nombuf;
 710                                 ++rxq->priv->dev->data->rx_mbuf_alloc_failed;
 711                                 goto repost;
 712                         }
 713 #ifndef NDEBUG
 714                         /* Poison user-modifiable fields in rep. */
 715                         NEXT(rep) = (void *)((uintptr_t)-1);
 716                         SET_DATA_OFF(rep, 0xdead);
 717                         DATA_LEN(rep) = 0xd00d;
 718                         PKT_LEN(rep) = 0xdeadd00d;
 719                         NB_SEGS(rep) = 0x2a;
 720                         PORT(rep) = 0x2a;
 721                         rep->ol_flags = -1;
 722 #endif
 723                         assert(rep->buf_len == seg->buf_len);
 724                         assert(rep->buf_len == rxq->mb_len);
 725                         /* Reconfigure sge to use rep instead of seg. */
 726                         assert(sge->lkey == rxq->mr->lkey);
 727                         sge->addr = ((uintptr_t)rep->buf_addr + seg_headroom);
 728                         elt->bufs[j] = rep;
 729                         ++j;
 730                         /* Update pkt_buf if it's the first segment, or link
 731                          * seg to the previous one and update pkt_buf_next. */
 732                         *pkt_buf_next = seg;
 733                         pkt_buf_next = &NEXT(seg);
 734                         /* Update seg information. */
 735                         seg_tailroom = (seg->buf_len - seg_headroom);
 736                         assert(sge->length == seg_tailroom);
 737                         SET_DATA_OFF(seg, seg_headroom);
 738                         if (likely(len <= seg_tailroom)) {
 739                                 /* Last segment. */
 740                                 DATA_LEN(seg) = len;
 741                                 PKT_LEN(seg) = len;
 742                                 /* Sanity check. */
 743                                 assert(rte_pktmbuf_headroom(seg) ==
 744                                        seg_headroom);
 745                                 assert(rte_pktmbuf_tailroom(seg) ==
 746                                        (seg_tailroom - len));
 747                                 break;
 748                         }
 749                         DATA_LEN(seg) = seg_tailroom;
 750                         PKT_LEN(seg) = seg_tailroom;
 751                         /* Sanity check. */
 752                         assert(rte_pktmbuf_headroom(seg) == seg_headroom);
 753                         assert(rte_pktmbuf_tailroom(seg) == 0);
 754                         /* Fix len and clear headroom for next segments. */
 755                         len -= seg_tailroom;
 756                         seg_headroom = 0;
 757                 }
 758                 /* Update head and tail segments. */
 759                 *pkt_buf_next = NULL;
 760                 assert(pkt_buf != NULL);
 761                 assert(j != 0);
 762                 NB_SEGS(pkt_buf) = j;
 763                 PORT(pkt_buf) = rxq->port_id;
 764                 PKT_LEN(pkt_buf) = pkt_buf_len;
 765                 pkt_buf->packet_type = rxq_cq_to_pkt_type(flags);
 766                 pkt_buf->ol_flags = rxq_cq_to_ol_flags(rxq, flags);
 767
 768                 /* Return packet. */
 769                 *(pkts++) = pkt_buf;
 770                 ++pkts_ret;
 771 #ifdef MLX5_PMD_SOFT_COUNTERS
 772                 /* Increment bytes counter. */
 773                 rxq->stats.ibytes += pkt_buf_len;
 774 #endif
 775 repost:
 776                 ret = rxq->if_wq->recv_sg_list(rxq->wq,
 777                                                elt->sges,
 778                                                RTE_DIM(elt->sges));
 779                 if (unlikely(ret)) {
 780                         /* Inability to repost WRs is fatal. */
 781                         DEBUG("%p: recv_sg_list(): failed (ret=%d)",
 782                               (void *)rxq->priv,
 783                               ret);
 784                         abort();
 785                 }
 786                 if (++elts_head >= elts_n)
 787                         elts_head = 0;
 788                 continue;
 789         }
 790         if (unlikely(i == 0))
 791                 return 0;
 792         rxq->elts_head = elts_head;
 793 #ifdef MLX5_PMD_SOFT_COUNTERS
 794         /* Increment packets counter. */
 795         rxq->stats.ipackets += pkts_ret;
 796 #endif
 797         return pkts_ret;
 798 }
 799
 800 /**
 801  * DPDK callback for RX.
 802  *
 803  * The following function is the same as mlx5_rx_burst_sp(), except it doesn't
 804  * manage scattered packets. Improves performance when MRU is lower than the
 805  * size of the first segment.
 806  *
 807  * @param dpdk_rxq
 808  *   Generic pointer to RX queue structure.
 809  * @param[out] pkts
 810  *   Array to store received packets.
 811  * @param pkts_n
 812  *   Maximum number of packets in array.
 813  *
 814  * @return
 815  *   Number of packets successfully received (<= pkts_n).
 816  */
 817 uint16_t
 818 mlx5_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
 819 {
 820         struct rxq *rxq = (struct rxq *)dpdk_rxq;
 821         struct rxq_elt (*elts)[rxq->elts_n] = rxq->elts.no_sp;
 822         const unsigned int elts_n = rxq->elts_n;
 823         unsigned int elts_head = rxq->elts_head;
 824         struct ibv_sge sges[pkts_n];
 825         unsigned int i;
 826         unsigned int pkts_ret = 0;
 827         int ret;
 828
 829         if (unlikely(rxq->sp))
 830                 return mlx5_rx_burst_sp(dpdk_rxq, pkts, pkts_n);
 831         for (i = 0; (i != pkts_n); ++i) {
 832                 struct rxq_elt *elt = &(*elts)[elts_head];
 833                 unsigned int len;
 834                 struct rte_mbuf *seg = elt->buf;
 835                 struct rte_mbuf *rep;
 836                 uint32_t flags;
 837
 838                 /* Sanity checks. */
 839                 assert(seg != NULL);
 840                 assert(elts_head < rxq->elts_n);
 841                 assert(rxq->elts_head < rxq->elts_n);
 842                 /*
 843                  * Fetch initial bytes of packet descriptor into a
 844                  * cacheline while allocating rep.
 845                  */
 846                 rte_prefetch0(seg);
 847                 rte_prefetch0(&seg->cacheline1);
 848                 ret = rxq->if_cq->poll_length_flags(rxq->cq, NULL, NULL,
 849                                                     &flags);
 850                 if (unlikely(ret < 0)) {
 851                         struct ibv_wc wc;
 852                         int wcs_n;
 853
 854                         DEBUG("rxq=%p, poll_length() failed (ret=%d)",
 855                               (void *)rxq, ret);
 856                         /* ibv_poll_cq() must be used in case of failure. */
 857                         wcs_n = ibv_poll_cq(rxq->cq, 1, &wc);
 858                         if (unlikely(wcs_n == 0))
 859                                 break;
 860                         if (unlikely(wcs_n < 0)) {
 861                                 DEBUG("rxq=%p, ibv_poll_cq() failed (wcs_n=%d)",
 862                                       (void *)rxq, wcs_n);
 863                                 break;
 864                         }
 865                         assert(wcs_n == 1);
 866                         if (unlikely(wc.status != IBV_WC_SUCCESS)) {
 867                                 /* Whatever, just repost the offending WR. */
 868                                 DEBUG("rxq=%p, wr_id=%" PRIu64 ": bad work"
 869                                       " completion status (%d): %s",
 870                                       (void *)rxq, wc.wr_id, wc.status,
 871                                       ibv_wc_status_str(wc.status));
 872 #ifdef MLX5_PMD_SOFT_COUNTERS
 873                                 /* Increment dropped packets counter. */
 874                                 ++rxq->stats.idropped;
 875 #endif
 876                                 /* Add SGE to array for repost. */
 877                                 sges[i] = elt->sge;
 878                                 goto repost;
 879                         }
 880                         ret = wc.byte_len;
 881                 }
 882                 if (ret == 0)
 883                         break;
 884                 len = ret;
 885                 rep = __rte_mbuf_raw_alloc(rxq->mp);
 886                 if (unlikely(rep == NULL)) {
 887                         /*
 888                          * Unable to allocate a replacement mbuf,
 889                          * repost WR.
 890                          */
 891                         DEBUG("rxq=%p: can't allocate a new mbuf",
 892                               (void *)rxq);
 893                         /* Increment out of memory counters. */
 894                         ++rxq->stats.rx_nombuf;
 895                         ++rxq->priv->dev->data->rx_mbuf_alloc_failed;
 896                         goto repost;
 897                 }
 898
 899                 /* Reconfigure sge to use rep instead of seg. */
 900                 elt->sge.addr = (uintptr_t)rep->buf_addr + RTE_PKTMBUF_HEADROOM;
 901                 assert(elt->sge.lkey == rxq->mr->lkey);
 902                 elt->buf = rep;
 903
 904                 /* Add SGE to array for repost. */
 905                 sges[i] = elt->sge;
 906
 907                 /* Update seg information. */
 908                 SET_DATA_OFF(seg, RTE_PKTMBUF_HEADROOM);
 909                 NB_SEGS(seg) = 1;
 910                 PORT(seg) = rxq->port_id;
 911                 NEXT(seg) = NULL;
 912                 PKT_LEN(seg) = len;
 913                 DATA_LEN(seg) = len;
 914                 seg->packet_type = rxq_cq_to_pkt_type(flags);
 915                 seg->ol_flags = rxq_cq_to_ol_flags(rxq, flags);
 916
 917                 /* Return packet. */
 918                 *(pkts++) = seg;
 919                 ++pkts_ret;
 920 #ifdef MLX5_PMD_SOFT_COUNTERS
 921                 /* Increment bytes counter. */
 922                 rxq->stats.ibytes += len;
 923 #endif
 924 repost:
 925                 if (++elts_head >= elts_n)
 926                         elts_head = 0;
 927                 continue;
 928         }
 929         if (unlikely(i == 0))
 930                 return 0;
 931         /* Repost WRs. */
 932 #ifdef DEBUG_RECV
 933         DEBUG("%p: reposting %u WRs", (void *)rxq, i);
 934 #endif
 935         ret = rxq->if_wq->recv_burst(rxq->wq, sges, i);
 936         if (unlikely(ret)) {
 937                 /* Inability to repost WRs is fatal. */
 938                 DEBUG("%p: recv_burst(): failed (ret=%d)",
 939                       (void *)rxq->priv,
 940                       ret);
 941                 abort();
 942         }
 943         rxq->elts_head = elts_head;
 944 #ifdef MLX5_PMD_SOFT_COUNTERS
 945         /* Increment packets counter. */
 946         rxq->stats.ipackets += pkts_ret;
 947 #endif
 948         return pkts_ret;
 949 }
 950
 951 /**
 952  * Dummy DPDK callback for TX.
 953  *
 954  * This function is used to temporarily replace the real callback during
 955  * unsafe control operations on the queue, or in case of error.
 956  *
 957  * @param dpdk_txq
 958  *   Generic pointer to TX queue structure.
 959  * @param[in] pkts
 960  *   Packets to transmit.
 961  * @param pkts_n
 962  *   Number of packets in array.
 963  *
 964  * @return
 965  *   Number of packets successfully transmitted (<= pkts_n).
 966  */
 967 uint16_t
 968 removed_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
 969 {
 970         (void)dpdk_txq;
 971         (void)pkts;
 972         (void)pkts_n;
 973         return 0;
 974 }
 975
 976 /**
 977  * Dummy DPDK callback for RX.
 978  *
 979  * This function is used to temporarily replace the real callback during
 980  * unsafe control operations on the queue, or in case of error.
 981  *
 982  * @param dpdk_rxq
 983  *   Generic pointer to RX queue structure.
 984  * @param[out] pkts
 985  *   Array to store received packets.
 986  * @param pkts_n
 987  *   Maximum number of packets in array.
 988  *
 989  * @return
 990  *   Number of packets successfully received (<= pkts_n).
 991  */
 992 uint16_t
 993 removed_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
 994 {
 995         (void)dpdk_rxq;
 996         (void)pkts;
 997         (void)pkts_n;
 998         return 0;
 999 }