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         }
 329         return (struct tx_burst_sg_ret){
 330                 .length = sent_size,
 331                 .num = segs,
 332         };
 333 stop:
 334         return (struct tx_burst_sg_ret){
 335                 .length = -1,
 336                 .num = -1,
 337         };
 338 }
 339
 340 #endif /* MLX5_PMD_SGE_WR_N > 1 */
 341
 342 /**
 343  * DPDK callback for TX.
 344  *
 345  * @param dpdk_txq
 346  *   Generic pointer to TX queue structure.
 347  * @param[in] pkts
 348  *   Packets to transmit.
 349  * @param pkts_n
 350  *   Number of packets in array.
 351  *
 352  * @return
 353  *   Number of packets successfully transmitted (<= pkts_n).
 354  */
 355 uint16_t
 356 mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
 357 {
 358         struct txq *txq = (struct txq *)dpdk_txq;
 359         unsigned int elts_head = txq->elts_head;
 360         const unsigned int elts_tail = txq->elts_tail;
 361         const unsigned int elts_n = txq->elts_n;
 362         unsigned int elts_comp_cd = txq->elts_comp_cd;
 363         unsigned int elts_comp = 0;
 364         unsigned int i;
 365         unsigned int max;
 366         int err;
 367
 368         assert(elts_comp_cd != 0);
 369         txq_complete(txq);
 370         max = (elts_n - (elts_head - elts_tail));
 371         if (max > elts_n)
 372                 max -= elts_n;
 373         assert(max >= 1);
 374         assert(max <= elts_n);
 375         /* Always leave one free entry in the ring. */
 376         --max;
 377         if (max == 0)
 378                 return 0;
 379         if (max > pkts_n)
 380                 max = pkts_n;
 381         for (i = 0; (i != max); ++i) {
 382                 struct rte_mbuf *buf = pkts[i];
 383                 unsigned int elts_head_next =
 384                         (((elts_head + 1) == elts_n) ? 0 : elts_head + 1);
 385                 struct txq_elt *elt_next = &(*txq->elts)[elts_head_next];
 386                 struct txq_elt *elt = &(*txq->elts)[elts_head];
 387                 unsigned int segs = NB_SEGS(buf);
 388 #ifdef MLX5_PMD_SOFT_COUNTERS
 389                 unsigned int sent_size = 0;
 390 #endif
 391                 uint32_t send_flags = 0;
 392
 393                 /* Clean up old buffer. */
 394                 if (likely(elt->buf != NULL)) {
 395                         struct rte_mbuf *tmp = elt->buf;
 396
 397                         /* Faster than rte_pktmbuf_free(). */
 398                         do {
 399                                 struct rte_mbuf *next = NEXT(tmp);
 400
 401                                 rte_pktmbuf_free_seg(tmp);
 402                                 tmp = next;
 403                         } while (tmp != NULL);
 404                 }
 405                 /* Request TX completion. */
 406                 if (unlikely(--elts_comp_cd == 0)) {
 407                         elts_comp_cd = txq->elts_comp_cd_init;
 408                         ++elts_comp;
 409                         send_flags |= IBV_EXP_QP_BURST_SIGNALED;
 410                 }
 411                 /* Should we enable HW CKSUM offload */
 412                 if (buf->ol_flags &
 413                     (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM)) {
 414                         send_flags |= IBV_EXP_QP_BURST_IP_CSUM;
 415                         /* HW does not support checksum offloads at arbitrary
 416                          * offsets but automatically recognizes the packet
 417                          * type. For inner L3/L4 checksums, only VXLAN (UDP)
 418                          * tunnels are currently supported. */
 419                         if (RTE_ETH_IS_TUNNEL_PKT(buf->packet_type))
 420                                 send_flags |= IBV_EXP_QP_BURST_TUNNEL;
 421                 }
 422                 if (likely(segs == 1)) {
 423                         uintptr_t addr;
 424                         uint32_t length;
 425                         uint32_t lkey;
 426
 427                         /* Retrieve buffer information. */
 428                         addr = rte_pktmbuf_mtod(buf, uintptr_t);
 429                         length = DATA_LEN(buf);
 430                         /* Retrieve Memory Region key for this memory pool. */
 431                         lkey = txq_mp2mr(txq, txq_mb2mp(buf));
 432                         if (unlikely(lkey == (uint32_t)-1)) {
 433                                 /* MR does not exist. */
 434                                 DEBUG("%p: unable to get MP <-> MR"
 435                                       " association", (void *)txq);
 436                                 /* Clean up TX element. */
 437                                 elt->buf = NULL;
 438                                 goto stop;
 439                         }
 440                         /* Update element. */
 441                         elt->buf = buf;
 442                         if (txq->priv->vf)
 443                                 rte_prefetch0((volatile void *)
 444                                               (uintptr_t)addr);
 445                         RTE_MBUF_PREFETCH_TO_FREE(elt_next->buf);
 446                         /* Put packet into send queue. */
 447 #if MLX5_PMD_MAX_INLINE > 0
 448                         if (length <= txq->max_inline)
 449                                 err = txq->if_qp->send_pending_inline
 450                                         (txq->qp,
 451                                          (void *)addr,
 452                                          length,
 453                                          send_flags);
 454                         else
 455 #endif
 456                                 err = txq->if_qp->send_pending
 457                                         (txq->qp,
 458                                          addr,
 459                                          length,
 460                                          lkey,
 461                                          send_flags);
 462                         if (unlikely(err))
 463                                 goto stop;
 464 #ifdef MLX5_PMD_SOFT_COUNTERS
 465                         sent_size += length;
 466 #endif
 467                 } else {
 468 #if MLX5_PMD_SGE_WR_N > 1
 469                         struct ibv_sge sges[MLX5_PMD_SGE_WR_N];
 470                         struct tx_burst_sg_ret ret;
 471
 472                         ret = tx_burst_sg(txq, segs, elt, buf, elts_head,
 473                                           &sges);
 474                         if (ret.length == (unsigned int)-1)
 475                                 goto stop;
 476                         RTE_MBUF_PREFETCH_TO_FREE(elt_next->buf);
 477                         /* Put SG list into send queue. */
 478                         err = txq->if_qp->send_pending_sg_list
 479                                 (txq->qp,
 480                                  sges,
 481                                  ret.num,
 482                                  send_flags);
 483                         if (unlikely(err))
 484                                 goto stop;
 485 #ifdef MLX5_PMD_SOFT_COUNTERS
 486                         sent_size += ret.length;
 487 #endif
 488 #else /* MLX5_PMD_SGE_WR_N > 1 */
 489                         DEBUG("%p: TX scattered buffers support not"
 490                               " compiled in", (void *)txq);
 491                         goto stop;
 492 #endif /* MLX5_PMD_SGE_WR_N > 1 */
 493                 }
 494                 elts_head = elts_head_next;
 495 #ifdef MLX5_PMD_SOFT_COUNTERS
 496                 /* Increment sent bytes counter. */
 497                 txq->stats.obytes += sent_size;
 498 #endif
 499         }
 500 stop:
 501         /* Take a shortcut if nothing must be sent. */
 502         if (unlikely(i == 0))
 503                 return 0;
 504 #ifdef MLX5_PMD_SOFT_COUNTERS
 505         /* Increment sent packets counter. */
 506         txq->stats.opackets += i;
 507 #endif
 508         /* Ring QP doorbell. */
 509         err = txq->if_qp->send_flush(txq->qp);
 510         if (unlikely(err)) {
 511                 /* A nonzero value is not supposed to be returned.
 512                  * Nothing can be done about it. */
 513                 DEBUG("%p: send_flush() failed with error %d",
 514                       (void *)txq, err);
 515         }
 516         txq->elts_head = elts_head;
 517         txq->elts_comp += elts_comp;
 518         txq->elts_comp_cd = elts_comp_cd;
 519         return i;
 520 }
 521
 522 /**
 523  * Translate RX completion flags to packet type.
 524  *
 525  * @param flags
 526  *   RX completion flags returned by poll_length_flags().
 527  *
 528  * @return
 529  *   Packet type for struct rte_mbuf.
 530  */
 531 static inline uint32_t
 532 rxq_cq_to_pkt_type(uint32_t flags)
 533 {
 534         uint32_t pkt_type;
 535
 536         if (flags & IBV_EXP_CQ_RX_TUNNEL_PACKET)
 537                 pkt_type =
 538                         TRANSPOSE(flags,
 539                                   IBV_EXP_CQ_RX_OUTER_IPV4_PACKET,
 540                                   RTE_PTYPE_L3_IPV4) |
 541                         TRANSPOSE(flags,
 542                                   IBV_EXP_CQ_RX_OUTER_IPV6_PACKET,
 543                                   RTE_PTYPE_L3_IPV6) |
 544                         TRANSPOSE(flags,
 545                                   IBV_EXP_CQ_RX_IPV4_PACKET,
 546                                   RTE_PTYPE_INNER_L3_IPV4) |
 547                         TRANSPOSE(flags,
 548                                   IBV_EXP_CQ_RX_IPV6_PACKET,
 549                                   RTE_PTYPE_INNER_L3_IPV6);
 550         else
 551                 pkt_type =
 552                         TRANSPOSE(flags,
 553                                   IBV_EXP_CQ_RX_IPV4_PACKET,
 554                                   RTE_PTYPE_L3_IPV4) |
 555                         TRANSPOSE(flags,
 556                                   IBV_EXP_CQ_RX_IPV6_PACKET,
 557                                   RTE_PTYPE_L3_IPV6);
 558         return pkt_type;
 559 }
 560
 561 /**
 562  * Translate RX completion flags to offload flags.
 563  *
 564  * @param[in] rxq
 565  *   Pointer to RX queue structure.
 566  * @param flags
 567  *   RX completion flags returned by poll_length_flags().
 568  *
 569  * @return
 570  *   Offload flags (ol_flags) for struct rte_mbuf.
 571  */
 572 static inline uint32_t
 573 rxq_cq_to_ol_flags(const struct rxq *rxq, uint32_t flags)
 574 {
 575         uint32_t ol_flags = 0;
 576
 577         if (rxq->csum)
 578                 ol_flags |=
 579                         TRANSPOSE(~flags,
 580                                   IBV_EXP_CQ_RX_IP_CSUM_OK,
 581                                   PKT_RX_IP_CKSUM_BAD) |
 582                         TRANSPOSE(~flags,
 583                                   IBV_EXP_CQ_RX_TCP_UDP_CSUM_OK,
 584                                   PKT_RX_L4_CKSUM_BAD);
 585         /*
 586          * PKT_RX_IP_CKSUM_BAD and PKT_RX_L4_CKSUM_BAD are used in place
 587          * of PKT_RX_EIP_CKSUM_BAD because the latter is not functional
 588          * (its value is 0).
 589          */
 590         if ((flags & IBV_EXP_CQ_RX_TUNNEL_PACKET) && (rxq->csum_l2tun))
 591                 ol_flags |=
 592                         TRANSPOSE(~flags,
 593                                   IBV_EXP_CQ_RX_OUTER_IP_CSUM_OK,
 594                                   PKT_RX_IP_CKSUM_BAD) |
 595                         TRANSPOSE(~flags,
 596                                   IBV_EXP_CQ_RX_OUTER_TCP_UDP_CSUM_OK,
 597                                   PKT_RX_L4_CKSUM_BAD);
 598         return ol_flags;
 599 }
 600
 601 /**
 602  * DPDK callback for RX with scattered packets support.
 603  *
 604  * @param dpdk_rxq
 605  *   Generic pointer to RX queue structure.
 606  * @param[out] pkts
 607  *   Array to store received packets.
 608  * @param pkts_n
 609  *   Maximum number of packets in array.
 610  *
 611  * @return
 612  *   Number of packets successfully received (<= pkts_n).
 613  */
 614 uint16_t
 615 mlx5_rx_burst_sp(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
 616 {
 617         struct rxq *rxq = (struct rxq *)dpdk_rxq;
 618         struct rxq_elt_sp (*elts)[rxq->elts_n] = rxq->elts.sp;
 619         const unsigned int elts_n = rxq->elts_n;
 620         unsigned int elts_head = rxq->elts_head;
 621         unsigned int i;
 622         unsigned int pkts_ret = 0;
 623         int ret;
 624
 625         if (unlikely(!rxq->sp))
 626                 return mlx5_rx_burst(dpdk_rxq, pkts, pkts_n);
 627         if (unlikely(elts == NULL)) /* See RTE_DEV_CMD_SET_MTU. */
 628                 return 0;
 629         for (i = 0; (i != pkts_n); ++i) {
 630                 struct rxq_elt_sp *elt = &(*elts)[elts_head];
 631                 unsigned int len;
 632                 unsigned int pkt_buf_len;
 633                 struct rte_mbuf *pkt_buf = NULL; /* Buffer returned in pkts. */
 634                 struct rte_mbuf **pkt_buf_next = &pkt_buf;
 635                 unsigned int seg_headroom = RTE_PKTMBUF_HEADROOM;
 636                 unsigned int j = 0;
 637                 uint32_t flags;
 638
 639                 /* Sanity checks. */
 640                 assert(elts_head < rxq->elts_n);
 641                 assert(rxq->elts_head < rxq->elts_n);
 642                 ret = rxq->if_cq->poll_length_flags(rxq->cq, NULL, NULL,
 643                                                     &flags);
 644                 if (unlikely(ret < 0)) {
 645                         struct ibv_wc wc;
 646                         int wcs_n;
 647
 648                         DEBUG("rxq=%p, poll_length() failed (ret=%d)",
 649                               (void *)rxq, ret);
 650                         /* ibv_poll_cq() must be used in case of failure. */
 651                         wcs_n = ibv_poll_cq(rxq->cq, 1, &wc);
 652                         if (unlikely(wcs_n == 0))
 653                                 break;
 654                         if (unlikely(wcs_n < 0)) {
 655                                 DEBUG("rxq=%p, ibv_poll_cq() failed (wcs_n=%d)",
 656                                       (void *)rxq, wcs_n);
 657                                 break;
 658                         }
 659                         assert(wcs_n == 1);
 660                         if (unlikely(wc.status != IBV_WC_SUCCESS)) {
 661                                 /* Whatever, just repost the offending WR. */
 662                                 DEBUG("rxq=%p, wr_id=%" PRIu64 ": bad work"
 663                                       " completion status (%d): %s",
 664                                       (void *)rxq, wc.wr_id, wc.status,
 665                                       ibv_wc_status_str(wc.status));
 666 #ifdef MLX5_PMD_SOFT_COUNTERS
 667                                 /* Increment dropped packets counter. */
 668                                 ++rxq->stats.idropped;
 669 #endif
 670                                 goto repost;
 671                         }
 672                         ret = wc.byte_len;
 673                 }
 674                 if (ret == 0)
 675                         break;
 676                 len = ret;
 677                 pkt_buf_len = len;
 678                 /*
 679                  * Replace spent segments with new ones, concatenate and
 680                  * return them as pkt_buf.
 681                  */
 682                 while (1) {
 683                         struct ibv_sge *sge = &elt->sges[j];
 684                         struct rte_mbuf *seg = elt->bufs[j];
 685                         struct rte_mbuf *rep;
 686                         unsigned int seg_tailroom;
 687
 688                         assert(seg != NULL);
 689                         /*
 690                          * Fetch initial bytes of packet descriptor into a
 691                          * cacheline while allocating rep.
 692                          */
 693                         rte_prefetch0(seg);
 694                         rep = __rte_mbuf_raw_alloc(rxq->mp);
 695                         if (unlikely(rep == NULL)) {
 696                                 /*
 697                                  * Unable to allocate a replacement mbuf,
 698                                  * repost WR.
 699                                  */
 700                                 DEBUG("rxq=%p: can't allocate a new mbuf",
 701                                       (void *)rxq);
 702                                 if (pkt_buf != NULL) {
 703                                         *pkt_buf_next = NULL;
 704                                         rte_pktmbuf_free(pkt_buf);
 705                                 }
 706                                 /* Increment out of memory counters. */
 707                                 ++rxq->stats.rx_nombuf;
 708                                 ++rxq->priv->dev->data->rx_mbuf_alloc_failed;
 709                                 goto repost;
 710                         }
 711 #ifndef NDEBUG
 712                         /* Poison user-modifiable fields in rep. */
 713                         NEXT(rep) = (void *)((uintptr_t)-1);
 714                         SET_DATA_OFF(rep, 0xdead);
 715                         DATA_LEN(rep) = 0xd00d;
 716                         PKT_LEN(rep) = 0xdeadd00d;
 717                         NB_SEGS(rep) = 0x2a;
 718                         PORT(rep) = 0x2a;
 719                         rep->ol_flags = -1;
 720 #endif
 721                         assert(rep->buf_len == seg->buf_len);
 722                         assert(rep->buf_len == rxq->mb_len);
 723                         /* Reconfigure sge to use rep instead of seg. */
 724                         assert(sge->lkey == rxq->mr->lkey);
 725                         sge->addr = ((uintptr_t)rep->buf_addr + seg_headroom);
 726                         elt->bufs[j] = rep;
 727                         ++j;
 728                         /* Update pkt_buf if it's the first segment, or link
 729                          * seg to the previous one and update pkt_buf_next. */
 730                         *pkt_buf_next = seg;
 731                         pkt_buf_next = &NEXT(seg);
 732                         /* Update seg information. */
 733                         seg_tailroom = (seg->buf_len - seg_headroom);
 734                         assert(sge->length == seg_tailroom);
 735                         SET_DATA_OFF(seg, seg_headroom);
 736                         if (likely(len <= seg_tailroom)) {
 737                                 /* Last segment. */
 738                                 DATA_LEN(seg) = len;
 739                                 PKT_LEN(seg) = len;
 740                                 /* Sanity check. */
 741                                 assert(rte_pktmbuf_headroom(seg) ==
 742                                        seg_headroom);
 743                                 assert(rte_pktmbuf_tailroom(seg) ==
 744                                        (seg_tailroom - len));
 745                                 break;
 746                         }
 747                         DATA_LEN(seg) = seg_tailroom;
 748                         PKT_LEN(seg) = seg_tailroom;
 749                         /* Sanity check. */
 750                         assert(rte_pktmbuf_headroom(seg) == seg_headroom);
 751                         assert(rte_pktmbuf_tailroom(seg) == 0);
 752                         /* Fix len and clear headroom for next segments. */
 753                         len -= seg_tailroom;
 754                         seg_headroom = 0;
 755                 }
 756                 /* Update head and tail segments. */
 757                 *pkt_buf_next = NULL;
 758                 assert(pkt_buf != NULL);
 759                 assert(j != 0);
 760                 NB_SEGS(pkt_buf) = j;
 761                 PORT(pkt_buf) = rxq->port_id;
 762                 PKT_LEN(pkt_buf) = pkt_buf_len;
 763                 pkt_buf->packet_type = rxq_cq_to_pkt_type(flags);
 764                 pkt_buf->ol_flags = rxq_cq_to_ol_flags(rxq, flags);
 765
 766                 /* Return packet. */
 767                 *(pkts++) = pkt_buf;
 768                 ++pkts_ret;
 769 #ifdef MLX5_PMD_SOFT_COUNTERS
 770                 /* Increment bytes counter. */
 771                 rxq->stats.ibytes += pkt_buf_len;
 772 #endif
 773 repost:
 774                 ret = rxq->if_wq->recv_sg_list(rxq->wq,
 775                                                elt->sges,
 776                                                RTE_DIM(elt->sges));
 777                 if (unlikely(ret)) {
 778                         /* Inability to repost WRs is fatal. */
 779                         DEBUG("%p: recv_sg_list(): failed (ret=%d)",
 780                               (void *)rxq->priv,
 781                               ret);
 782                         abort();
 783                 }
 784                 if (++elts_head >= elts_n)
 785                         elts_head = 0;
 786                 continue;
 787         }
 788         if (unlikely(i == 0))
 789                 return 0;
 790         rxq->elts_head = elts_head;
 791 #ifdef MLX5_PMD_SOFT_COUNTERS
 792         /* Increment packets counter. */
 793         rxq->stats.ipackets += pkts_ret;
 794 #endif
 795         return pkts_ret;
 796 }
 797
 798 /**
 799  * DPDK callback for RX.
 800  *
 801  * The following function is the same as mlx5_rx_burst_sp(), except it doesn't
 802  * manage scattered packets. Improves performance when MRU is lower than the
 803  * size of the first segment.
 804  *
 805  * @param dpdk_rxq
 806  *   Generic pointer to RX queue structure.
 807  * @param[out] pkts
 808  *   Array to store received packets.
 809  * @param pkts_n
 810  *   Maximum number of packets in array.
 811  *
 812  * @return
 813  *   Number of packets successfully received (<= pkts_n).
 814  */
 815 uint16_t
 816 mlx5_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
 817 {
 818         struct rxq *rxq = (struct rxq *)dpdk_rxq;
 819         struct rxq_elt (*elts)[rxq->elts_n] = rxq->elts.no_sp;
 820         const unsigned int elts_n = rxq->elts_n;
 821         unsigned int elts_head = rxq->elts_head;
 822         struct ibv_sge sges[pkts_n];
 823         unsigned int i;
 824         unsigned int pkts_ret = 0;
 825         int ret;
 826
 827         if (unlikely(rxq->sp))
 828                 return mlx5_rx_burst_sp(dpdk_rxq, pkts, pkts_n);
 829         for (i = 0; (i != pkts_n); ++i) {
 830                 struct rxq_elt *elt = &(*elts)[elts_head];
 831                 unsigned int len;
 832                 struct rte_mbuf *seg = elt->buf;
 833                 struct rte_mbuf *rep;
 834                 uint32_t flags;
 835
 836                 /* Sanity checks. */
 837                 assert(seg != NULL);
 838                 assert(elts_head < rxq->elts_n);
 839                 assert(rxq->elts_head < rxq->elts_n);
 840                 /*
 841                  * Fetch initial bytes of packet descriptor into a
 842                  * cacheline while allocating rep.
 843                  */
 844                 rte_prefetch0(seg);
 845                 rte_prefetch0(&seg->cacheline1);
 846                 ret = rxq->if_cq->poll_length_flags(rxq->cq, NULL, NULL,
 847                                                     &flags);
 848                 if (unlikely(ret < 0)) {
 849                         struct ibv_wc wc;
 850                         int wcs_n;
 851
 852                         DEBUG("rxq=%p, poll_length() failed (ret=%d)",
 853                               (void *)rxq, ret);
 854                         /* ibv_poll_cq() must be used in case of failure. */
 855                         wcs_n = ibv_poll_cq(rxq->cq, 1, &wc);
 856                         if (unlikely(wcs_n == 0))
 857                                 break;
 858                         if (unlikely(wcs_n < 0)) {
 859                                 DEBUG("rxq=%p, ibv_poll_cq() failed (wcs_n=%d)",
 860                                       (void *)rxq, wcs_n);
 861                                 break;
 862                         }
 863                         assert(wcs_n == 1);
 864                         if (unlikely(wc.status != IBV_WC_SUCCESS)) {
 865                                 /* Whatever, just repost the offending WR. */
 866                                 DEBUG("rxq=%p, wr_id=%" PRIu64 ": bad work"
 867                                       " completion status (%d): %s",
 868                                       (void *)rxq, wc.wr_id, wc.status,
 869                                       ibv_wc_status_str(wc.status));
 870 #ifdef MLX5_PMD_SOFT_COUNTERS
 871                                 /* Increment dropped packets counter. */
 872                                 ++rxq->stats.idropped;
 873 #endif
 874                                 /* Add SGE to array for repost. */
 875                                 sges[i] = elt->sge;
 876                                 goto repost;
 877                         }
 878                         ret = wc.byte_len;
 879                 }
 880                 if (ret == 0)
 881                         break;
 882                 len = ret;
 883                 rep = __rte_mbuf_raw_alloc(rxq->mp);
 884                 if (unlikely(rep == NULL)) {
 885                         /*
 886                          * Unable to allocate a replacement mbuf,
 887                          * repost WR.
 888                          */
 889                         DEBUG("rxq=%p: can't allocate a new mbuf",
 890                               (void *)rxq);
 891                         /* Increment out of memory counters. */
 892                         ++rxq->stats.rx_nombuf;
 893                         ++rxq->priv->dev->data->rx_mbuf_alloc_failed;
 894                         goto repost;
 895                 }
 896
 897                 /* Reconfigure sge to use rep instead of seg. */
 898                 elt->sge.addr = (uintptr_t)rep->buf_addr + RTE_PKTMBUF_HEADROOM;
 899                 assert(elt->sge.lkey == rxq->mr->lkey);
 900                 elt->buf = rep;
 901
 902                 /* Add SGE to array for repost. */
 903                 sges[i] = elt->sge;
 904
 905                 /* Update seg information. */
 906                 SET_DATA_OFF(seg, RTE_PKTMBUF_HEADROOM);
 907                 NB_SEGS(seg) = 1;
 908                 PORT(seg) = rxq->port_id;
 909                 NEXT(seg) = NULL;
 910                 PKT_LEN(seg) = len;
 911                 DATA_LEN(seg) = len;
 912                 seg->packet_type = rxq_cq_to_pkt_type(flags);
 913                 seg->ol_flags = rxq_cq_to_ol_flags(rxq, flags);
 914
 915                 /* Return packet. */
 916                 *(pkts++) = seg;
 917                 ++pkts_ret;
 918 #ifdef MLX5_PMD_SOFT_COUNTERS
 919                 /* Increment bytes counter. */
 920                 rxq->stats.ibytes += len;
 921 #endif
 922 repost:
 923                 if (++elts_head >= elts_n)
 924                         elts_head = 0;
 925                 continue;
 926         }
 927         if (unlikely(i == 0))
 928                 return 0;
 929         /* Repost WRs. */
 930 #ifdef DEBUG_RECV
 931         DEBUG("%p: reposting %u WRs", (void *)rxq, i);
 932 #endif
 933         ret = rxq->if_wq->recv_burst(rxq->wq, sges, i);
 934         if (unlikely(ret)) {
 935                 /* Inability to repost WRs is fatal. */
 936                 DEBUG("%p: recv_burst(): failed (ret=%d)",
 937                       (void *)rxq->priv,
 938                       ret);
 939                 abort();
 940         }
 941         rxq->elts_head = elts_head;
 942 #ifdef MLX5_PMD_SOFT_COUNTERS
 943         /* Increment packets counter. */
 944         rxq->stats.ipackets += pkts_ret;
 945 #endif
 946         return pkts_ret;
 947 }
 948
 949 /**
 950  * Dummy DPDK callback for TX.
 951  *
 952  * This function is used to temporarily replace the real callback during
 953  * unsafe control operations on the queue, or in case of error.
 954  *
 955  * @param dpdk_txq
 956  *   Generic pointer to TX queue structure.
 957  * @param[in] pkts
 958  *   Packets to transmit.
 959  * @param pkts_n
 960  *   Number of packets in array.
 961  *
 962  * @return
 963  *   Number of packets successfully transmitted (<= pkts_n).
 964  */
 965 uint16_t
 966 removed_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
 967 {
 968         (void)dpdk_txq;
 969         (void)pkts;
 970         (void)pkts_n;
 971         return 0;
 972 }
 973
 974 /**
 975  * Dummy DPDK callback for RX.
 976  *
 977  * This function is used to temporarily replace the real callback during
 978  * unsafe control operations on the queue, or in case of error.
 979  *
 980  * @param dpdk_rxq
 981  *   Generic pointer to RX queue structure.
 982  * @param[out] pkts
 983  *   Array to store received packets.
 984  * @param pkts_n
 985  *   Maximum number of packets in array.
 986  *
 987  * @return
 988  *   Number of packets successfully received (<= pkts_n).
 989  */
 990 uint16_t
 991 removed_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
 992 {
 993         (void)dpdk_rxq;
 994         (void)pkts;
 995         (void)pkts_n;
 996         return 0;
 997 }