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 #include <errno.h>
  39
  40 /* Verbs header. */
  41 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
  42 #ifdef PEDANTIC
  43 #pragma GCC diagnostic ignored "-pedantic"
  44 #endif
  45 #include <infiniband/verbs.h>
  46 #ifdef PEDANTIC
  47 #pragma GCC diagnostic error "-pedantic"
  48 #endif
  49
  50 /* DPDK headers don't like -pedantic. */
  51 #ifdef PEDANTIC
  52 #pragma GCC diagnostic ignored "-pedantic"
  53 #endif
  54 #include <rte_mbuf.h>
  55 #include <rte_mempool.h>
  56 #include <rte_prefetch.h>
  57 #include <rte_common.h>
  58 #include <rte_branch_prediction.h>
  59 #ifdef PEDANTIC
  60 #pragma GCC diagnostic error "-pedantic"
  61 #endif
  62
  63 #include "mlx5.h"
  64 #include "mlx5_autoconf.h"
  65 #include "mlx5_utils.h"
  66 #include "mlx5_rxtx.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         const unsigned int elts_n = txq->elts_n;
  89         int wcs_n;
  90
  91         if (unlikely(elts_comp == 0))
  92                 return 0;
  93 #ifdef DEBUG_SEND
  94         DEBUG("%p: processing %u work requests completions",
  95               (void *)txq, elts_comp);
  96 #endif
  97         wcs_n = txq->if_cq->poll_cnt(txq->cq, elts_comp);
  98         if (unlikely(wcs_n == 0))
  99                 return 0;
 100         if (unlikely(wcs_n < 0)) {
 101                 DEBUG("%p: ibv_poll_cq() failed (wcs_n=%d)",
 102                       (void *)txq, wcs_n);
 103                 return -1;
 104         }
 105         elts_comp -= wcs_n;
 106         assert(elts_comp <= txq->elts_comp);
 107         /*
 108          * Assume WC status is successful as nothing can be done about it
 109          * anyway.
 110          */
 111         elts_tail += wcs_n * txq->elts_comp_cd_init;
 112         if (elts_tail >= elts_n)
 113                 elts_tail -= elts_n;
 114         txq->elts_tail = elts_tail;
 115         txq->elts_comp = elts_comp;
 116         return 0;
 117 }
 118
 119 /**
 120  * Get Memory Region (MR) <-> Memory Pool (MP) association from txq->mp2mr[].
 121  * Add MP to txq->mp2mr[] if it's not registered yet. If mp2mr[] is full,
 122  * remove an entry first.
 123  *
 124  * @param txq
 125  *   Pointer to TX queue structure.
 126  * @param[in] mp
 127  *   Memory Pool for which a Memory Region lkey must be returned.
 128  *
 129  * @return
 130  *   mr->lkey on success, (uint32_t)-1 on failure.
 131  */
 132 static uint32_t
 133 txq_mp2mr(struct txq *txq, struct rte_mempool *mp)
 134 {
 135         unsigned int i;
 136         struct ibv_mr *mr;
 137
 138         for (i = 0; (i != RTE_DIM(txq->mp2mr)); ++i) {
 139                 if (unlikely(txq->mp2mr[i].mp == NULL)) {
 140                         /* Unknown MP, add a new MR for it. */
 141                         break;
 142                 }
 143                 if (txq->mp2mr[i].mp == mp) {
 144                         assert(txq->mp2mr[i].lkey != (uint32_t)-1);
 145                         assert(txq->mp2mr[i].mr->lkey == txq->mp2mr[i].lkey);
 146                         return txq->mp2mr[i].lkey;
 147                 }
 148         }
 149         /* Add a new entry, register MR first. */
 150         DEBUG("%p: discovered new memory pool %p", (void *)txq, (void *)mp);
 151         mr = ibv_reg_mr(txq->priv->pd,
 152                         (void *)mp->elt_va_start,
 153                         (mp->elt_va_end - mp->elt_va_start),
 154                         (IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE));
 155         if (unlikely(mr == NULL)) {
 156                 DEBUG("%p: unable to configure MR, ibv_reg_mr() failed.",
 157                       (void *)txq);
 158                 return (uint32_t)-1;
 159         }
 160         if (unlikely(i == RTE_DIM(txq->mp2mr))) {
 161                 /* Table is full, remove oldest entry. */
 162                 DEBUG("%p: MR <-> MP table full, dropping oldest entry.",
 163                       (void *)txq);
 164                 --i;
 165                 claim_zero(ibv_dereg_mr(txq->mp2mr[i].mr));
 166                 memmove(&txq->mp2mr[0], &txq->mp2mr[1],
 167                         (sizeof(txq->mp2mr) - sizeof(txq->mp2mr[0])));
 168         }
 169         /* Store the new entry. */
 170         txq->mp2mr[i].mp = mp;
 171         txq->mp2mr[i].mr = mr;
 172         txq->mp2mr[i].lkey = mr->lkey;
 173         DEBUG("%p: new MR lkey for MP %p: 0x%08" PRIu32,
 174               (void *)txq, (void *)mp, txq->mp2mr[i].lkey);
 175         return txq->mp2mr[i].lkey;
 176 }
 177
 178 #if MLX5_PMD_SGE_WR_N > 1
 179
 180 /**
 181  * Copy scattered mbuf contents to a single linear buffer.
 182  *
 183  * @param[out] linear
 184  *   Linear output buffer.
 185  * @param[in] buf
 186  *   Scattered input buffer.
 187  *
 188  * @return
 189  *   Number of bytes copied to the output buffer or 0 if not large enough.
 190  */
 191 static unsigned int
 192 linearize_mbuf(linear_t *linear, struct rte_mbuf *buf)
 193 {
 194         unsigned int size = 0;
 195         unsigned int offset;
 196
 197         do {
 198                 unsigned int len = DATA_LEN(buf);
 199
 200                 offset = size;
 201                 size += len;
 202                 if (unlikely(size > sizeof(*linear)))
 203                         return 0;
 204                 memcpy(&(*linear)[offset],
 205                        rte_pktmbuf_mtod(buf, uint8_t *),
 206                        len);
 207                 buf = NEXT(buf);
 208         } while (buf != NULL);
 209         return size;
 210 }
 211
 212 /**
 213  * Handle scattered buffers for mlx5_tx_burst().
 214  *
 215  * @param txq
 216  *   TX queue structure.
 217  * @param segs
 218  *   Number of segments in buf.
 219  * @param elt
 220  *   TX queue element to fill.
 221  * @param[in] buf
 222  *   Buffer to process.
 223  * @param elts_head
 224  *   Index of the linear buffer to use if necessary (normally txq->elts_head).
 225  * @param[out] sges
 226  *   Array filled with SGEs on success.
 227  *
 228  * @return
 229  *   A structure containing the processed packet size in bytes and the
 230  *   number of SGEs. Both fields are set to (unsigned int)-1 in case of
 231  *   failure.
 232  */
 233 static struct tx_burst_sg_ret {
 234         unsigned int length;
 235         unsigned int num;
 236 }
 237 tx_burst_sg(struct txq *txq, unsigned int segs, struct txq_elt *elt,
 238             struct rte_mbuf *buf, unsigned int elts_head,
 239             struct ibv_sge (*sges)[MLX5_PMD_SGE_WR_N])
 240 {
 241         unsigned int sent_size = 0;
 242         unsigned int j;
 243         int linearize = 0;
 244
 245         /* When there are too many segments, extra segments are
 246          * linearized in the last SGE. */
 247         if (unlikely(segs > RTE_DIM(*sges))) {
 248                 segs = (RTE_DIM(*sges) - 1);
 249                 linearize = 1;
 250         }
 251         /* Update element. */
 252         elt->buf = buf;
 253         /* Register segments as SGEs. */
 254         for (j = 0; (j != segs); ++j) {
 255                 struct ibv_sge *sge = &(*sges)[j];
 256                 uint32_t lkey;
 257
 258                 /* Retrieve Memory Region key for this memory pool. */
 259                 lkey = txq_mp2mr(txq, buf->pool);
 260                 if (unlikely(lkey == (uint32_t)-1)) {
 261                         /* MR does not exist. */
 262                         DEBUG("%p: unable to get MP <-> MR association",
 263                               (void *)txq);
 264                         /* Clean up TX element. */
 265                         elt->buf = NULL;
 266                         goto stop;
 267                 }
 268                 /* Update SGE. */
 269                 sge->addr = rte_pktmbuf_mtod(buf, uintptr_t);
 270                 if (txq->priv->vf)
 271                         rte_prefetch0((volatile void *)
 272                                       (uintptr_t)sge->addr);
 273                 sge->length = DATA_LEN(buf);
 274                 sge->lkey = lkey;
 275                 sent_size += sge->length;
 276                 buf = NEXT(buf);
 277         }
 278         /* If buf is not NULL here and is not going to be linearized,
 279          * nb_segs is not valid. */
 280         assert(j == segs);
 281         assert((buf == NULL) || (linearize));
 282         /* Linearize extra segments. */
 283         if (linearize) {
 284                 struct ibv_sge *sge = &(*sges)[segs];
 285                 linear_t *linear = &(*txq->elts_linear)[elts_head];
 286                 unsigned int size = linearize_mbuf(linear, buf);
 287
 288                 assert(segs == (RTE_DIM(*sges) - 1));
 289                 if (size == 0) {
 290                         /* Invalid packet. */
 291                         DEBUG("%p: packet too large to be linearized.",
 292                               (void *)txq);
 293                         /* Clean up TX element. */
 294                         elt->buf = NULL;
 295                         goto stop;
 296                 }
 297                 /* If MLX5_PMD_SGE_WR_N is 1, free mbuf immediately. */
 298                 if (RTE_DIM(*sges) == 1) {
 299                         do {
 300                                 struct rte_mbuf *next = NEXT(buf);
 301
 302                                 rte_pktmbuf_free_seg(buf);
 303                                 buf = next;
 304                         } while (buf != NULL);
 305                         elt->buf = NULL;
 306                 }
 307                 /* Update SGE. */
 308                 sge->addr = (uintptr_t)&(*linear)[0];
 309                 sge->length = size;
 310                 sge->lkey = txq->mr_linear->lkey;
 311                 sent_size += size;
 312         }
 313         return (struct tx_burst_sg_ret){
 314                 .length = sent_size,
 315                 .num = segs,
 316         };
 317 stop:
 318         return (struct tx_burst_sg_ret){
 319                 .length = -1,
 320                 .num = -1,
 321         };
 322 }
 323
 324 #endif /* MLX5_PMD_SGE_WR_N > 1 */
 325
 326 /**
 327  * DPDK callback for TX.
 328  *
 329  * @param dpdk_txq
 330  *   Generic pointer to TX queue structure.
 331  * @param[in] pkts
 332  *   Packets to transmit.
 333  * @param pkts_n
 334  *   Number of packets in array.
 335  *
 336  * @return
 337  *   Number of packets successfully transmitted (<= pkts_n).
 338  */
 339 uint16_t
 340 mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
 341 {
 342         struct txq *txq = (struct txq *)dpdk_txq;
 343         unsigned int elts_head = txq->elts_head;
 344         const unsigned int elts_tail = txq->elts_tail;
 345         const unsigned int elts_n = txq->elts_n;
 346         unsigned int elts_comp_cd = txq->elts_comp_cd;
 347         unsigned int elts_comp = 0;
 348         unsigned int i;
 349         unsigned int max;
 350         int err;
 351
 352         assert(elts_comp_cd != 0);
 353         txq_complete(txq);
 354         max = (elts_n - (elts_head - elts_tail));
 355         if (max > elts_n)
 356                 max -= elts_n;
 357         assert(max >= 1);
 358         assert(max <= elts_n);
 359         /* Always leave one free entry in the ring. */
 360         --max;
 361         if (max == 0)
 362                 return 0;
 363         if (max > pkts_n)
 364                 max = pkts_n;
 365         for (i = 0; (i != max); ++i) {
 366                 struct rte_mbuf *buf = pkts[i];
 367                 unsigned int elts_head_next =
 368                         (((elts_head + 1) == elts_n) ? 0 : elts_head + 1);
 369                 struct txq_elt *elt_next = &(*txq->elts)[elts_head_next];
 370                 struct txq_elt *elt = &(*txq->elts)[elts_head];
 371                 unsigned int segs = NB_SEGS(buf);
 372 #ifdef MLX5_PMD_SOFT_COUNTERS
 373                 unsigned int sent_size = 0;
 374 #endif
 375                 uint32_t send_flags = 0;
 376
 377                 /* Clean up old buffer. */
 378                 if (likely(elt->buf != NULL)) {
 379                         struct rte_mbuf *tmp = elt->buf;
 380
 381                         /* Faster than rte_pktmbuf_free(). */
 382                         do {
 383                                 struct rte_mbuf *next = NEXT(tmp);
 384
 385                                 rte_pktmbuf_free_seg(tmp);
 386                                 tmp = next;
 387                         } while (tmp != NULL);
 388                 }
 389                 /* Request TX completion. */
 390                 if (unlikely(--elts_comp_cd == 0)) {
 391                         elts_comp_cd = txq->elts_comp_cd_init;
 392                         ++elts_comp;
 393                         send_flags |= IBV_EXP_QP_BURST_SIGNALED;
 394                 }
 395                 /* Should we enable HW CKSUM offload */
 396                 if (buf->ol_flags &
 397                     (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM)) {
 398                         send_flags |= IBV_EXP_QP_BURST_IP_CSUM;
 399                         /* HW does not support checksum offloads at arbitrary
 400                          * offsets but automatically recognizes the packet
 401                          * type. For inner L3/L4 checksums, only VXLAN (UDP)
 402                          * tunnels are currently supported. */
 403                         if (RTE_ETH_IS_TUNNEL_PKT(buf->packet_type))
 404                                 send_flags |= IBV_EXP_QP_BURST_TUNNEL;
 405                 }
 406                 if (likely(segs == 1)) {
 407                         uintptr_t addr;
 408                         uint32_t length;
 409                         uint32_t lkey;
 410
 411                         /* Retrieve buffer information. */
 412                         addr = rte_pktmbuf_mtod(buf, uintptr_t);
 413                         length = DATA_LEN(buf);
 414                         /* Retrieve Memory Region key for this memory pool. */
 415                         lkey = txq_mp2mr(txq, buf->pool);
 416                         if (unlikely(lkey == (uint32_t)-1)) {
 417                                 /* MR does not exist. */
 418                                 DEBUG("%p: unable to get MP <-> MR"
 419                                       " association", (void *)txq);
 420                                 /* Clean up TX element. */
 421                                 elt->buf = NULL;
 422                                 goto stop;
 423                         }
 424                         /* Update element. */
 425                         elt->buf = buf;
 426                         if (txq->priv->vf)
 427                                 rte_prefetch0((volatile void *)
 428                                               (uintptr_t)addr);
 429                         RTE_MBUF_PREFETCH_TO_FREE(elt_next->buf);
 430                         /* Put packet into send queue. */
 431 #if MLX5_PMD_MAX_INLINE > 0
 432                         if (length <= txq->max_inline)
 433                                 err = txq->if_qp->send_pending_inline
 434                                         (txq->qp,
 435                                          (void *)addr,
 436                                          length,
 437                                          send_flags);
 438                         else
 439 #endif
 440                                 err = txq->if_qp->send_pending
 441                                         (txq->qp,
 442                                          addr,
 443                                          length,
 444                                          lkey,
 445                                          send_flags);
 446                         if (unlikely(err))
 447                                 goto stop;
 448 #ifdef MLX5_PMD_SOFT_COUNTERS
 449                         sent_size += length;
 450 #endif
 451                 } else {
 452 #if MLX5_PMD_SGE_WR_N > 1
 453                         struct ibv_sge sges[MLX5_PMD_SGE_WR_N];
 454                         struct tx_burst_sg_ret ret;
 455
 456                         ret = tx_burst_sg(txq, segs, elt, buf, elts_head,
 457                                           &sges);
 458                         if (ret.length == (unsigned int)-1)
 459                                 goto stop;
 460                         RTE_MBUF_PREFETCH_TO_FREE(elt_next->buf);
 461                         /* Put SG list into send queue. */
 462                         err = txq->if_qp->send_pending_sg_list
 463                                 (txq->qp,
 464                                  sges,
 465                                  ret.num,
 466                                  send_flags);
 467                         if (unlikely(err))
 468                                 goto stop;
 469 #ifdef MLX5_PMD_SOFT_COUNTERS
 470                         sent_size += ret.length;
 471 #endif
 472 #else /* MLX5_PMD_SGE_WR_N > 1 */
 473                         DEBUG("%p: TX scattered buffers support not"
 474                               " compiled in", (void *)txq);
 475                         goto stop;
 476 #endif /* MLX5_PMD_SGE_WR_N > 1 */
 477                 }
 478                 elts_head = elts_head_next;
 479 #ifdef MLX5_PMD_SOFT_COUNTERS
 480                 /* Increment sent bytes counter. */
 481                 txq->stats.obytes += sent_size;
 482 #endif
 483         }
 484 stop:
 485         /* Take a shortcut if nothing must be sent. */
 486         if (unlikely(i == 0))
 487                 return 0;
 488 #ifdef MLX5_PMD_SOFT_COUNTERS
 489         /* Increment sent packets counter. */
 490         txq->stats.opackets += i;
 491 #endif
 492         /* Ring QP doorbell. */
 493         err = txq->if_qp->send_flush(txq->qp);
 494         if (unlikely(err)) {
 495                 /* A nonzero value is not supposed to be returned.
 496                  * Nothing can be done about it. */
 497                 DEBUG("%p: send_flush() failed with error %d",
 498                       (void *)txq, err);
 499         }
 500         txq->elts_head = elts_head;
 501         txq->elts_comp += elts_comp;
 502         txq->elts_comp_cd = elts_comp_cd;
 503         return i;
 504 }
 505
 506 /**
 507  * Translate RX completion flags to packet type.
 508  *
 509  * @param flags
 510  *   RX completion flags returned by poll_length_flags().
 511  *
 512  * @return
 513  *   Packet type for struct rte_mbuf.
 514  */
 515 static inline uint32_t
 516 rxq_cq_to_pkt_type(uint32_t flags)
 517 {
 518         uint32_t pkt_type;
 519
 520         if (flags & IBV_EXP_CQ_RX_TUNNEL_PACKET)
 521                 pkt_type =
 522                         TRANSPOSE(flags,
 523                                   IBV_EXP_CQ_RX_OUTER_IPV4_PACKET,
 524                                   RTE_PTYPE_L3_IPV4) |
 525                         TRANSPOSE(flags,
 526                                   IBV_EXP_CQ_RX_OUTER_IPV6_PACKET,
 527                                   RTE_PTYPE_L3_IPV6) |
 528                         TRANSPOSE(flags,
 529                                   IBV_EXP_CQ_RX_IPV4_PACKET,
 530                                   RTE_PTYPE_INNER_L3_IPV4) |
 531                         TRANSPOSE(flags,
 532                                   IBV_EXP_CQ_RX_IPV6_PACKET,
 533                                   RTE_PTYPE_INNER_L3_IPV6);
 534         else
 535                 pkt_type =
 536                         TRANSPOSE(flags,
 537                                   IBV_EXP_CQ_RX_IPV4_PACKET,
 538                                   RTE_PTYPE_L3_IPV4) |
 539                         TRANSPOSE(flags,
 540                                   IBV_EXP_CQ_RX_IPV6_PACKET,
 541                                   RTE_PTYPE_L3_IPV6);
 542         return pkt_type;
 543 }
 544
 545 /**
 546  * Translate RX completion flags to offload flags.
 547  *
 548  * @param[in] rxq
 549  *   Pointer to RX queue structure.
 550  * @param flags
 551  *   RX completion flags returned by poll_length_flags().
 552  *
 553  * @return
 554  *   Offload flags (ol_flags) for struct rte_mbuf.
 555  */
 556 static inline uint32_t
 557 rxq_cq_to_ol_flags(const struct rxq *rxq, uint32_t flags)
 558 {
 559         uint32_t ol_flags = 0;
 560
 561         if (rxq->csum)
 562                 ol_flags |=
 563                         TRANSPOSE(~flags,
 564                                   IBV_EXP_CQ_RX_IP_CSUM_OK,
 565                                   PKT_RX_IP_CKSUM_BAD) |
 566                         TRANSPOSE(~flags,
 567                                   IBV_EXP_CQ_RX_TCP_UDP_CSUM_OK,
 568                                   PKT_RX_L4_CKSUM_BAD);
 569         /*
 570          * PKT_RX_IP_CKSUM_BAD and PKT_RX_L4_CKSUM_BAD are used in place
 571          * of PKT_RX_EIP_CKSUM_BAD because the latter is not functional
 572          * (its value is 0).
 573          */
 574         if ((flags & IBV_EXP_CQ_RX_TUNNEL_PACKET) && (rxq->csum_l2tun))
 575                 ol_flags |=
 576                         TRANSPOSE(~flags,
 577                                   IBV_EXP_CQ_RX_OUTER_IP_CSUM_OK,
 578                                   PKT_RX_IP_CKSUM_BAD) |
 579                         TRANSPOSE(~flags,
 580                                   IBV_EXP_CQ_RX_OUTER_TCP_UDP_CSUM_OK,
 581                                   PKT_RX_L4_CKSUM_BAD);
 582         return ol_flags;
 583 }
 584
 585 /**
 586  * DPDK callback for RX with scattered packets support.
 587  *
 588  * @param dpdk_rxq
 589  *   Generic pointer to RX queue structure.
 590  * @param[out] pkts
 591  *   Array to store received packets.
 592  * @param pkts_n
 593  *   Maximum number of packets in array.
 594  *
 595  * @return
 596  *   Number of packets successfully received (<= pkts_n).
 597  */
 598 uint16_t
 599 mlx5_rx_burst_sp(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
 600 {
 601         struct rxq *rxq = (struct rxq *)dpdk_rxq;
 602         struct rxq_elt_sp (*elts)[rxq->elts_n] = rxq->elts.sp;
 603         const unsigned int elts_n = rxq->elts_n;
 604         unsigned int elts_head = rxq->elts_head;
 605         unsigned int i;
 606         unsigned int pkts_ret = 0;
 607         int ret;
 608
 609         if (unlikely(!rxq->sp))
 610                 return mlx5_rx_burst(dpdk_rxq, pkts, pkts_n);
 611         if (unlikely(elts == NULL)) /* See RTE_DEV_CMD_SET_MTU. */
 612                 return 0;
 613         for (i = 0; (i != pkts_n); ++i) {
 614                 struct rxq_elt_sp *elt = &(*elts)[elts_head];
 615                 struct ibv_recv_wr *wr = &elt->wr;
 616                 uint64_t wr_id = wr->wr_id;
 617                 unsigned int len;
 618                 unsigned int pkt_buf_len;
 619                 struct rte_mbuf *pkt_buf = NULL; /* Buffer returned in pkts. */
 620                 struct rte_mbuf **pkt_buf_next = &pkt_buf;
 621                 unsigned int seg_headroom = RTE_PKTMBUF_HEADROOM;
 622                 unsigned int j = 0;
 623                 uint32_t flags;
 624
 625                 /* Sanity checks. */
 626 #ifdef NDEBUG
 627                 (void)wr_id;
 628 #endif
 629                 assert(wr_id < rxq->elts_n);
 630                 assert(wr->sg_list == elt->sges);
 631                 assert(wr->num_sge == RTE_DIM(elt->sges));
 632                 assert(elts_head < rxq->elts_n);
 633                 assert(rxq->elts_head < rxq->elts_n);
 634                 ret = rxq->if_cq->poll_length_flags(rxq->cq, NULL, NULL,
 635                                                     &flags);
 636                 if (unlikely(ret < 0)) {
 637                         struct ibv_wc wc;
 638                         int wcs_n;
 639
 640                         DEBUG("rxq=%p, poll_length() failed (ret=%d)",
 641                               (void *)rxq, ret);
 642                         /* ibv_poll_cq() must be used in case of failure. */
 643                         wcs_n = ibv_poll_cq(rxq->cq, 1, &wc);
 644                         if (unlikely(wcs_n == 0))
 645                                 break;
 646                         if (unlikely(wcs_n < 0)) {
 647                                 DEBUG("rxq=%p, ibv_poll_cq() failed (wcs_n=%d)",
 648                                       (void *)rxq, wcs_n);
 649                                 break;
 650                         }
 651                         assert(wcs_n == 1);
 652                         if (unlikely(wc.status != IBV_WC_SUCCESS)) {
 653                                 /* Whatever, just repost the offending WR. */
 654                                 DEBUG("rxq=%p, wr_id=%" PRIu64 ": bad work"
 655                                       " completion status (%d): %s",
 656                                       (void *)rxq, wc.wr_id, wc.status,
 657                                       ibv_wc_status_str(wc.status));
 658 #ifdef MLX5_PMD_SOFT_COUNTERS
 659                                 /* Increment dropped packets counter. */
 660                                 ++rxq->stats.idropped;
 661 #endif
 662                                 goto repost;
 663                         }
 664                         ret = wc.byte_len;
 665                 }
 666                 if (ret == 0)
 667                         break;
 668                 len = ret;
 669                 pkt_buf_len = len;
 670                 /*
 671                  * Replace spent segments with new ones, concatenate and
 672                  * return them as pkt_buf.
 673                  */
 674                 while (1) {
 675                         struct ibv_sge *sge = &elt->sges[j];
 676                         struct rte_mbuf *seg = elt->bufs[j];
 677                         struct rte_mbuf *rep;
 678                         unsigned int seg_tailroom;
 679
 680                         /*
 681                          * Fetch initial bytes of packet descriptor into a
 682                          * cacheline while allocating rep.
 683                          */
 684                         rte_prefetch0(seg);
 685                         rep = __rte_mbuf_raw_alloc(rxq->mp);
 686                         if (unlikely(rep == NULL)) {
 687                                 /*
 688                                  * Unable to allocate a replacement mbuf,
 689                                  * repost WR.
 690                                  */
 691                                 DEBUG("rxq=%p, wr_id=%" PRIu64 ":"
 692                                       " can't allocate a new mbuf",
 693                                       (void *)rxq, wr_id);
 694                                 if (pkt_buf != NULL) {
 695                                         *pkt_buf_next = NULL;
 696                                         rte_pktmbuf_free(pkt_buf);
 697                                 }
 698                                 /* Increment out of memory counters. */
 699                                 ++rxq->stats.rx_nombuf;
 700                                 ++rxq->priv->dev->data->rx_mbuf_alloc_failed;
 701                                 goto repost;
 702                         }
 703 #ifndef NDEBUG
 704                         /* Poison user-modifiable fields in rep. */
 705                         NEXT(rep) = (void *)((uintptr_t)-1);
 706                         SET_DATA_OFF(rep, 0xdead);
 707                         DATA_LEN(rep) = 0xd00d;
 708                         PKT_LEN(rep) = 0xdeadd00d;
 709                         NB_SEGS(rep) = 0x2a;
 710                         PORT(rep) = 0x2a;
 711                         rep->ol_flags = -1;
 712 #endif
 713                         assert(rep->buf_len == seg->buf_len);
 714                         assert(rep->buf_len == rxq->mb_len);
 715                         /* Reconfigure sge to use rep instead of seg. */
 716                         assert(sge->lkey == rxq->mr->lkey);
 717                         sge->addr = ((uintptr_t)rep->buf_addr + seg_headroom);
 718                         elt->bufs[j] = rep;
 719                         ++j;
 720                         /* Update pkt_buf if it's the first segment, or link
 721                          * seg to the previous one and update pkt_buf_next. */
 722                         *pkt_buf_next = seg;
 723                         pkt_buf_next = &NEXT(seg);
 724                         /* Update seg information. */
 725                         seg_tailroom = (seg->buf_len - seg_headroom);
 726                         assert(sge->length == seg_tailroom);
 727                         SET_DATA_OFF(seg, seg_headroom);
 728                         if (likely(len <= seg_tailroom)) {
 729                                 /* Last segment. */
 730                                 DATA_LEN(seg) = len;
 731                                 PKT_LEN(seg) = len;
 732                                 /* Sanity check. */
 733                                 assert(rte_pktmbuf_headroom(seg) ==
 734                                        seg_headroom);
 735                                 assert(rte_pktmbuf_tailroom(seg) ==
 736                                        (seg_tailroom - len));
 737                                 break;
 738                         }
 739                         DATA_LEN(seg) = seg_tailroom;
 740                         PKT_LEN(seg) = seg_tailroom;
 741                         /* Sanity check. */
 742                         assert(rte_pktmbuf_headroom(seg) == seg_headroom);
 743                         assert(rte_pktmbuf_tailroom(seg) == 0);
 744                         /* Fix len and clear headroom for next segments. */
 745                         len -= seg_tailroom;
 746                         seg_headroom = 0;
 747                 }
 748                 /* Update head and tail segments. */
 749                 *pkt_buf_next = NULL;
 750                 assert(pkt_buf != NULL);
 751                 assert(j != 0);
 752                 NB_SEGS(pkt_buf) = j;
 753                 PORT(pkt_buf) = rxq->port_id;
 754                 PKT_LEN(pkt_buf) = pkt_buf_len;
 755                 pkt_buf->packet_type = rxq_cq_to_pkt_type(flags);
 756                 pkt_buf->ol_flags = rxq_cq_to_ol_flags(rxq, flags);
 757
 758                 /* Return packet. */
 759                 *(pkts++) = pkt_buf;
 760                 ++pkts_ret;
 761 #ifdef MLX5_PMD_SOFT_COUNTERS
 762                 /* Increment bytes counter. */
 763                 rxq->stats.ibytes += pkt_buf_len;
 764 #endif
 765 repost:
 766 #ifdef HAVE_EXP_QP_BURST_RECV_SG_LIST
 767                 ret = rxq->if_qp->recv_sg_list(rxq->qp,
 768                                                elt->sges,
 769                                                RTE_DIM(elt->sges));
 770 #else /* HAVE_EXP_QP_BURST_RECV_SG_LIST */
 771                 errno = ENOSYS;
 772                 ret = -1;
 773 #endif /* HAVE_EXP_QP_BURST_RECV_SG_LIST */
 774                 if (unlikely(ret)) {
 775                         /* Inability to repost WRs is fatal. */
 776                         DEBUG("%p: recv_sg_list(): failed (ret=%d)",
 777                               (void *)rxq->priv,
 778                               ret);
 779                         abort();
 780                 }
 781                 if (++elts_head >= elts_n)
 782                         elts_head = 0;
 783                 continue;
 784         }
 785         if (unlikely(i == 0))
 786                 return 0;
 787         rxq->elts_head = elts_head;
 788 #ifdef MLX5_PMD_SOFT_COUNTERS
 789         /* Increment packets counter. */
 790         rxq->stats.ipackets += pkts_ret;
 791 #endif
 792         return pkts_ret;
 793 }
 794
 795 /**
 796  * DPDK callback for RX.
 797  *
 798  * The following function is the same as mlx5_rx_burst_sp(), except it doesn't
 799  * manage scattered packets. Improves performance when MRU is lower than the
 800  * size of the first segment.
 801  *
 802  * @param dpdk_rxq
 803  *   Generic pointer to RX queue structure.
 804  * @param[out] pkts
 805  *   Array to store received packets.
 806  * @param pkts_n
 807  *   Maximum number of packets in array.
 808  *
 809  * @return
 810  *   Number of packets successfully received (<= pkts_n).
 811  */
 812 uint16_t
 813 mlx5_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
 814 {
 815         struct rxq *rxq = (struct rxq *)dpdk_rxq;
 816         struct rxq_elt (*elts)[rxq->elts_n] = rxq->elts.no_sp;
 817         const unsigned int elts_n = rxq->elts_n;
 818         unsigned int elts_head = rxq->elts_head;
 819         struct ibv_sge sges[pkts_n];
 820         unsigned int i;
 821         unsigned int pkts_ret = 0;
 822         int ret;
 823
 824         if (unlikely(rxq->sp))
 825                 return mlx5_rx_burst_sp(dpdk_rxq, pkts, pkts_n);
 826         for (i = 0; (i != pkts_n); ++i) {
 827                 struct rxq_elt *elt = &(*elts)[elts_head];
 828                 struct ibv_recv_wr *wr = &elt->wr;
 829                 uint64_t wr_id = wr->wr_id;
 830                 unsigned int len;
 831                 struct rte_mbuf *seg = (void *)((uintptr_t)elt->sge.addr -
 832                         WR_ID(wr_id).offset);
 833                 struct rte_mbuf *rep;
 834                 uint32_t flags;
 835
 836                 /* Sanity checks. */
 837                 assert(WR_ID(wr_id).id < rxq->elts_n);
 838                 assert(wr->sg_list == &elt->sge);
 839                 assert(wr->num_sge == 1);
 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, wr_id=%" PRIu32 ":"
 892                               " can't allocate a new mbuf",
 893                               (void *)rxq, WR_ID(wr_id).id);
 894                         /* Increment out of memory counters. */
 895                         ++rxq->stats.rx_nombuf;
 896                         ++rxq->priv->dev->data->rx_mbuf_alloc_failed;
 897                         goto repost;
 898                 }
 899
 900                 /* Reconfigure sge to use rep instead of seg. */
 901                 elt->sge.addr = (uintptr_t)rep->buf_addr + RTE_PKTMBUF_HEADROOM;
 902                 assert(elt->sge.lkey == rxq->mr->lkey);
 903                 WR_ID(wr->wr_id).offset =
 904                         (((uintptr_t)rep->buf_addr + RTE_PKTMBUF_HEADROOM) -
 905                          (uintptr_t)rep);
 906                 assert(WR_ID(wr->wr_id).id == WR_ID(wr_id).id);
 907
 908                 /* Add SGE to array for repost. */
 909                 sges[i] = elt->sge;
 910
 911                 /* Update seg information. */
 912                 SET_DATA_OFF(seg, RTE_PKTMBUF_HEADROOM);
 913                 NB_SEGS(seg) = 1;
 914                 PORT(seg) = rxq->port_id;
 915                 NEXT(seg) = NULL;
 916                 PKT_LEN(seg) = len;
 917                 DATA_LEN(seg) = len;
 918                 seg->packet_type = rxq_cq_to_pkt_type(flags);
 919                 seg->ol_flags = rxq_cq_to_ol_flags(rxq, flags);
 920
 921                 /* Return packet. */
 922                 *(pkts++) = seg;
 923                 ++pkts_ret;
 924 #ifdef MLX5_PMD_SOFT_COUNTERS
 925                 /* Increment bytes counter. */
 926                 rxq->stats.ibytes += len;
 927 #endif
 928 repost:
 929                 if (++elts_head >= elts_n)
 930                         elts_head = 0;
 931                 continue;
 932         }
 933         if (unlikely(i == 0))
 934                 return 0;
 935         /* Repost WRs. */
 936 #ifdef DEBUG_RECV
 937         DEBUG("%p: reposting %u WRs", (void *)rxq, i);
 938 #endif
 939         ret = rxq->if_qp->recv_burst(rxq->qp, sges, i);
 940         if (unlikely(ret)) {
 941                 /* Inability to repost WRs is fatal. */
 942                 DEBUG("%p: recv_burst(): failed (ret=%d)",
 943                       (void *)rxq->priv,
 944                       ret);
 945                 abort();
 946         }
 947         rxq->elts_head = elts_head;
 948 #ifdef MLX5_PMD_SOFT_COUNTERS
 949         /* Increment packets counter. */
 950         rxq->stats.ipackets += pkts_ret;
 951 #endif
 952         return pkts_ret;
 953 }
 954
 955 /**
 956  * Dummy DPDK callback for TX.
 957  *
 958  * This function is used to temporarily replace the real callback during
 959  * unsafe control operations on the queue, or in case of error.
 960  *
 961  * @param dpdk_txq
 962  *   Generic pointer to TX queue structure.
 963  * @param[in] pkts
 964  *   Packets to transmit.
 965  * @param pkts_n
 966  *   Number of packets in array.
 967  *
 968  * @return
 969  *   Number of packets successfully transmitted (<= pkts_n).
 970  */
 971 uint16_t
 972 removed_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
 973 {
 974         (void)dpdk_txq;
 975         (void)pkts;
 976         (void)pkts_n;
 977         return 0;
 978 }
 979
 980 /**
 981  * Dummy DPDK callback for RX.
 982  *
 983  * This function is used to temporarily replace the real callback during
 984  * unsafe control operations on the queue, or in case of error.
 985  *
 986  * @param dpdk_rxq
 987  *   Generic pointer to RX queue structure.
 988  * @param[out] pkts
 989  *   Array to store received packets.
 990  * @param pkts_n
 991  *   Maximum number of packets in array.
 992  *
 993  * @return
 994  *   Number of packets successfully received (<= pkts_n).
 995  */
 996 uint16_t
 997 removed_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
 998 {
 999         (void)dpdk_rxq;
1000         (void)pkts;
1001         (void)pkts_n;
1002         return 0;
1003 }