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_autoconf.h"
  66 #include "mlx5_defs.h"
  67
  68 /**
  69  * Manage TX completions.
  70  *
  71  * When sending a burst, mlx5_tx_burst() posts several WRs.
  72  * To improve performance, a completion event is only required once every
  73  * MLX5_PMD_TX_PER_COMP_REQ sends. Doing so discards completion information
  74  * for other WRs, but this information would not be used anyway.
  75  *
  76  * @param txq
  77  *   Pointer to TX queue structure.
  78  *
  79  * @return
  80  *   0 on success, -1 on failure.
  81  */
  82 static int
  83 txq_complete(struct txq *txq)
  84 {
  85         unsigned int elts_comp = txq->elts_comp;
  86         unsigned int elts_tail = txq->elts_tail;
  87         const unsigned int elts_n = txq->elts_n;
  88         int wcs_n;
  89
  90         if (unlikely(elts_comp == 0))
  91                 return 0;
  92 #ifdef DEBUG_SEND
  93         DEBUG("%p: processing %u work requests completions",
  94               (void *)txq, elts_comp);
  95 #endif
  96         wcs_n = txq->poll_cnt(txq->cq, elts_comp);
  97         if (unlikely(wcs_n == 0))
  98                 return 0;
  99         if (unlikely(wcs_n < 0)) {
 100                 DEBUG("%p: ibv_poll_cq() failed (wcs_n=%d)",
 101                       (void *)txq, wcs_n);
 102                 return -1;
 103         }
 104         elts_comp -= wcs_n;
 105         assert(elts_comp <= txq->elts_comp);
 106         /*
 107          * Assume WC status is successful as nothing can be done about it
 108          * anyway.
 109          */
 110         elts_tail += wcs_n * txq->elts_comp_cd_init;
 111         if (elts_tail >= elts_n)
 112                 elts_tail -= elts_n;
 113         txq->elts_tail = elts_tail;
 114         txq->elts_comp = elts_comp;
 115         return 0;
 116 }
 117
 118 /**
 119  * Get Memory Pool (MP) from mbuf. If mbuf is indirect, the pool from which
 120  * the cloned mbuf is allocated is returned instead.
 121  *
 122  * @param buf
 123  *   Pointer to mbuf.
 124  *
 125  * @return
 126  *   Memory pool where data is located for given mbuf.
 127  */
 128 static struct rte_mempool *
 129 txq_mb2mp(struct rte_mbuf *buf)
 130 {
 131         if (unlikely(RTE_MBUF_INDIRECT(buf)))
 132                 return rte_mbuf_from_indirect(buf)->pool;
 133         return buf->pool;
 134 }
 135
 136 /**
 137  * Get Memory Region (MR) <-> Memory Pool (MP) association from txq->mp2mr[].
 138  * Add MP to txq->mp2mr[] if it's not registered yet. If mp2mr[] is full,
 139  * remove an entry first.
 140  *
 141  * @param txq
 142  *   Pointer to TX queue structure.
 143  * @param[in] mp
 144  *   Memory Pool for which a Memory Region lkey must be returned.
 145  *
 146  * @return
 147  *   mr->lkey on success, (uint32_t)-1 on failure.
 148  */
 149 static uint32_t
 150 txq_mp2mr(struct txq *txq, const struct rte_mempool *mp)
 151 {
 152         unsigned int i;
 153         struct ibv_mr *mr;
 154
 155         for (i = 0; (i != RTE_DIM(txq->mp2mr)); ++i) {
 156                 if (unlikely(txq->mp2mr[i].mp == NULL)) {
 157                         /* Unknown MP, add a new MR for it. */
 158                         break;
 159                 }
 160                 if (txq->mp2mr[i].mp == mp) {
 161                         assert(txq->mp2mr[i].lkey != (uint32_t)-1);
 162                         assert(txq->mp2mr[i].mr->lkey == txq->mp2mr[i].lkey);
 163                         return txq->mp2mr[i].lkey;
 164                 }
 165         }
 166         /* Add a new entry, register MR first. */
 167         DEBUG("%p: discovered new memory pool \"%s\" (%p)",
 168               (void *)txq, mp->name, (const void *)mp);
 169         mr = ibv_reg_mr(txq->priv->pd,
 170                         (void *)mp->elt_va_start,
 171                         (mp->elt_va_end - mp->elt_va_start),
 172                         (IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE));
 173         if (unlikely(mr == NULL)) {
 174                 DEBUG("%p: unable to configure MR, ibv_reg_mr() failed.",
 175                       (void *)txq);
 176                 return (uint32_t)-1;
 177         }
 178         if (unlikely(i == RTE_DIM(txq->mp2mr))) {
 179                 /* Table is full, remove oldest entry. */
 180                 DEBUG("%p: MR <-> MP table full, dropping oldest entry.",
 181                       (void *)txq);
 182                 --i;
 183                 claim_zero(ibv_dereg_mr(txq->mp2mr[0].mr));
 184                 memmove(&txq->mp2mr[0], &txq->mp2mr[1],
 185                         (sizeof(txq->mp2mr) - sizeof(txq->mp2mr[0])));
 186         }
 187         /* Store the new entry. */
 188         txq->mp2mr[i].mp = mp;
 189         txq->mp2mr[i].mr = mr;
 190         txq->mp2mr[i].lkey = mr->lkey;
 191         DEBUG("%p: new MR lkey for MP \"%s\" (%p): 0x%08" PRIu32,
 192               (void *)txq, mp->name, (const void *)mp, txq->mp2mr[i].lkey);
 193         return txq->mp2mr[i].lkey;
 194 }
 195
 196 struct txq_mp2mr_mbuf_check_data {
 197         const struct rte_mempool *mp;
 198         int ret;
 199 };
 200
 201 /**
 202  * Callback function for rte_mempool_obj_iter() to check whether a given
 203  * mempool object looks like a mbuf.
 204  *
 205  * @param[in, out] arg
 206  *   Context data (struct txq_mp2mr_mbuf_check_data). Contains mempool pointer
 207  *   and return value.
 208  * @param[in] start
 209  *   Object start address.
 210  * @param[in] end
 211  *   Object end address.
 212  * @param index
 213  *   Unused.
 214  *
 215  * @return
 216  *   Nonzero value when object is not a mbuf.
 217  */
 218 static void
 219 txq_mp2mr_mbuf_check(void *arg, void *start, void *end,
 220                      uint32_t index __rte_unused)
 221 {
 222         struct txq_mp2mr_mbuf_check_data *data = arg;
 223         struct rte_mbuf *buf =
 224                 (void *)((uintptr_t)start + data->mp->header_size);
 225
 226         (void)index;
 227         /* Check whether mbuf structure fits element size and whether mempool
 228          * pointer is valid. */
 229         if (((uintptr_t)end >= (uintptr_t)(buf + 1)) &&
 230             (buf->pool == data->mp))
 231                 data->ret = 0;
 232         else
 233                 data->ret = -1;
 234 }
 235
 236 /**
 237  * Iterator function for rte_mempool_walk() to register existing mempools and
 238  * fill the MP to MR cache of a TX queue.
 239  *
 240  * @param[in] mp
 241  *   Memory Pool to register.
 242  * @param *arg
 243  *   Pointer to TX queue structure.
 244  */
 245 void
 246 txq_mp2mr_iter(const struct rte_mempool *mp, void *arg)
 247 {
 248         struct txq *txq = arg;
 249         struct txq_mp2mr_mbuf_check_data data = {
 250                 .mp = mp,
 251                 .ret = -1,
 252         };
 253
 254         /* Discard empty mempools. */
 255         if (mp->size == 0)
 256                 return;
 257         /* Register mempool only if the first element looks like a mbuf. */
 258         rte_mempool_obj_iter((void *)mp->elt_va_start,
 259                              1,
 260                              mp->header_size + mp->elt_size + mp->trailer_size,
 261                              1,
 262                              mp->elt_pa,
 263                              mp->pg_num,
 264                              mp->pg_shift,
 265                              txq_mp2mr_mbuf_check,
 266                              &data);
 267         if (data.ret)
 268                 return;
 269         txq_mp2mr(txq, mp);
 270 }
 271
 272 #if MLX5_PMD_SGE_WR_N > 1
 273
 274 /**
 275  * Copy scattered mbuf contents to a single linear buffer.
 276  *
 277  * @param[out] linear
 278  *   Linear output buffer.
 279  * @param[in] buf
 280  *   Scattered input buffer.
 281  *
 282  * @return
 283  *   Number of bytes copied to the output buffer or 0 if not large enough.
 284  */
 285 static unsigned int
 286 linearize_mbuf(linear_t *linear, struct rte_mbuf *buf)
 287 {
 288         unsigned int size = 0;
 289         unsigned int offset;
 290
 291         do {
 292                 unsigned int len = DATA_LEN(buf);
 293
 294                 offset = size;
 295                 size += len;
 296                 if (unlikely(size > sizeof(*linear)))
 297                         return 0;
 298                 memcpy(&(*linear)[offset],
 299                        rte_pktmbuf_mtod(buf, uint8_t *),
 300                        len);
 301                 buf = NEXT(buf);
 302         } while (buf != NULL);
 303         return size;
 304 }
 305
 306 /**
 307  * Handle scattered buffers for mlx5_tx_burst().
 308  *
 309  * @param txq
 310  *   TX queue structure.
 311  * @param segs
 312  *   Number of segments in buf.
 313  * @param elt
 314  *   TX queue element to fill.
 315  * @param[in] buf
 316  *   Buffer to process.
 317  * @param elts_head
 318  *   Index of the linear buffer to use if necessary (normally txq->elts_head).
 319  * @param[out] sges
 320  *   Array filled with SGEs on success.
 321  *
 322  * @return
 323  *   A structure containing the processed packet size in bytes and the
 324  *   number of SGEs. Both fields are set to (unsigned int)-1 in case of
 325  *   failure.
 326  */
 327 static struct tx_burst_sg_ret {
 328         unsigned int length;
 329         unsigned int num;
 330 }
 331 tx_burst_sg(struct txq *txq, unsigned int segs, struct txq_elt *elt,
 332             struct rte_mbuf *buf, unsigned int elts_head,
 333             struct ibv_sge (*sges)[MLX5_PMD_SGE_WR_N])
 334 {
 335         unsigned int sent_size = 0;
 336         unsigned int j;
 337         int linearize = 0;
 338
 339         /* When there are too many segments, extra segments are
 340          * linearized in the last SGE. */
 341         if (unlikely(segs > RTE_DIM(*sges))) {
 342                 segs = (RTE_DIM(*sges) - 1);
 343                 linearize = 1;
 344         }
 345         /* Update element. */
 346         elt->buf = buf;
 347         /* Register segments as SGEs. */
 348         for (j = 0; (j != segs); ++j) {
 349                 struct ibv_sge *sge = &(*sges)[j];
 350                 uint32_t lkey;
 351
 352                 /* Retrieve Memory Region key for this memory pool. */
 353                 lkey = txq_mp2mr(txq, txq_mb2mp(buf));
 354                 if (unlikely(lkey == (uint32_t)-1)) {
 355                         /* MR does not exist. */
 356                         DEBUG("%p: unable to get MP <-> MR association",
 357                               (void *)txq);
 358                         /* Clean up TX element. */
 359                         elt->buf = NULL;
 360                         goto stop;
 361                 }
 362                 /* Update SGE. */
 363                 sge->addr = rte_pktmbuf_mtod(buf, uintptr_t);
 364                 if (txq->priv->vf)
 365                         rte_prefetch0((volatile void *)
 366                                       (uintptr_t)sge->addr);
 367                 sge->length = DATA_LEN(buf);
 368                 sge->lkey = lkey;
 369                 sent_size += sge->length;
 370                 buf = NEXT(buf);
 371         }
 372         /* If buf is not NULL here and is not going to be linearized,
 373          * nb_segs is not valid. */
 374         assert(j == segs);
 375         assert((buf == NULL) || (linearize));
 376         /* Linearize extra segments. */
 377         if (linearize) {
 378                 struct ibv_sge *sge = &(*sges)[segs];
 379                 linear_t *linear = &(*txq->elts_linear)[elts_head];
 380                 unsigned int size = linearize_mbuf(linear, buf);
 381
 382                 assert(segs == (RTE_DIM(*sges) - 1));
 383                 if (size == 0) {
 384                         /* Invalid packet. */
 385                         DEBUG("%p: packet too large to be linearized.",
 386                               (void *)txq);
 387                         /* Clean up TX element. */
 388                         elt->buf = NULL;
 389                         goto stop;
 390                 }
 391                 /* If MLX5_PMD_SGE_WR_N is 1, free mbuf immediately. */
 392                 if (RTE_DIM(*sges) == 1) {
 393                         do {
 394                                 struct rte_mbuf *next = NEXT(buf);
 395
 396                                 rte_pktmbuf_free_seg(buf);
 397                                 buf = next;
 398                         } while (buf != NULL);
 399                         elt->buf = NULL;
 400                 }
 401                 /* Update SGE. */
 402                 sge->addr = (uintptr_t)&(*linear)[0];
 403                 sge->length = size;
 404                 sge->lkey = txq->mr_linear->lkey;
 405                 sent_size += size;
 406                 /* Include last segment. */
 407                 segs++;
 408         }
 409         return (struct tx_burst_sg_ret){
 410                 .length = sent_size,
 411                 .num = segs,
 412         };
 413 stop:
 414         return (struct tx_burst_sg_ret){
 415                 .length = -1,
 416                 .num = -1,
 417         };
 418 }
 419
 420 #endif /* MLX5_PMD_SGE_WR_N > 1 */
 421
 422 /**
 423  * DPDK callback for TX.
 424  *
 425  * @param dpdk_txq
 426  *   Generic pointer to TX queue structure.
 427  * @param[in] pkts
 428  *   Packets to transmit.
 429  * @param pkts_n
 430  *   Number of packets in array.
 431  *
 432  * @return
 433  *   Number of packets successfully transmitted (<= pkts_n).
 434  */
 435 uint16_t
 436 mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
 437 {
 438         struct txq *txq = (struct txq *)dpdk_txq;
 439         unsigned int elts_head = txq->elts_head;
 440         const unsigned int elts_n = txq->elts_n;
 441         unsigned int elts_comp_cd = txq->elts_comp_cd;
 442         unsigned int elts_comp = 0;
 443         unsigned int i;
 444         unsigned int max;
 445         int err;
 446         struct rte_mbuf *buf = pkts[0];
 447
 448         assert(elts_comp_cd != 0);
 449         /* Prefetch first packet cacheline. */
 450         rte_prefetch0(buf);
 451         txq_complete(txq);
 452         max = (elts_n - (elts_head - txq->elts_tail));
 453         if (max > elts_n)
 454                 max -= elts_n;
 455         assert(max >= 1);
 456         assert(max <= elts_n);
 457         /* Always leave one free entry in the ring. */
 458         --max;
 459         if (max == 0)
 460                 return 0;
 461         if (max > pkts_n)
 462                 max = pkts_n;
 463         for (i = 0; (i != max); ++i) {
 464                 struct rte_mbuf *buf_next = pkts[i + 1];
 465                 unsigned int elts_head_next =
 466                         (((elts_head + 1) == elts_n) ? 0 : elts_head + 1);
 467                 struct txq_elt *elt_next = &(*txq->elts)[elts_head_next];
 468                 struct txq_elt *elt = &(*txq->elts)[elts_head];
 469                 unsigned int segs = NB_SEGS(buf);
 470 #ifdef MLX5_PMD_SOFT_COUNTERS
 471                 unsigned int sent_size = 0;
 472 #endif
 473                 uint32_t send_flags = 0;
 474
 475                 /* Clean up old buffer. */
 476                 if (likely(elt->buf != NULL)) {
 477                         struct rte_mbuf *tmp = elt->buf;
 478
 479                         /* Faster than rte_pktmbuf_free(). */
 480                         do {
 481                                 struct rte_mbuf *next = NEXT(tmp);
 482
 483                                 rte_pktmbuf_free_seg(tmp);
 484                                 tmp = next;
 485                         } while (tmp != NULL);
 486                 }
 487                 if (i + 1 < max)
 488                         rte_prefetch0(buf_next);
 489                 /* Request TX completion. */
 490                 if (unlikely(--elts_comp_cd == 0)) {
 491                         elts_comp_cd = txq->elts_comp_cd_init;
 492                         ++elts_comp;
 493                         send_flags |= IBV_EXP_QP_BURST_SIGNALED;
 494                 }
 495                 /* Should we enable HW CKSUM offload */
 496                 if (buf->ol_flags &
 497                     (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM)) {
 498                         send_flags |= IBV_EXP_QP_BURST_IP_CSUM;
 499                         /* HW does not support checksum offloads at arbitrary
 500                          * offsets but automatically recognizes the packet
 501                          * type. For inner L3/L4 checksums, only VXLAN (UDP)
 502                          * tunnels are currently supported. */
 503                         if (RTE_ETH_IS_TUNNEL_PKT(buf->packet_type))
 504                                 send_flags |= IBV_EXP_QP_BURST_TUNNEL;
 505                 }
 506                 if (likely(segs == 1)) {
 507                         uintptr_t addr;
 508                         uint32_t length;
 509                         uint32_t lkey;
 510                         uintptr_t buf_next_addr;
 511
 512                         /* Retrieve buffer information. */
 513                         addr = rte_pktmbuf_mtod(buf, uintptr_t);
 514                         length = DATA_LEN(buf);
 515                         /* Update element. */
 516                         elt->buf = buf;
 517                         if (txq->priv->vf)
 518                                 rte_prefetch0((volatile void *)
 519                                               (uintptr_t)addr);
 520                         RTE_MBUF_PREFETCH_TO_FREE(elt_next->buf);
 521                         /* Prefetch next buffer data. */
 522                         if (i + 1 < max) {
 523                                 buf_next_addr =
 524                                         rte_pktmbuf_mtod(buf_next, uintptr_t);
 525                                 rte_prefetch0((volatile void *)
 526                                               (uintptr_t)buf_next_addr);
 527                         }
 528                         /* Put packet into send queue. */
 529 #if MLX5_PMD_MAX_INLINE > 0
 530                         if (length <= txq->max_inline)
 531                                 err = txq->send_pending_inline
 532                                         (txq->qp,
 533                                          (void *)addr,
 534                                          length,
 535                                          send_flags);
 536                         else
 537 #endif
 538                         {
 539                                 /* Retrieve Memory Region key for this
 540                                  * memory pool. */
 541                                 lkey = txq_mp2mr(txq, txq_mb2mp(buf));
 542                                 if (unlikely(lkey == (uint32_t)-1)) {
 543                                         /* MR does not exist. */
 544                                         DEBUG("%p: unable to get MP <-> MR"
 545                                               " association", (void *)txq);
 546                                         /* Clean up TX element. */
 547                                         elt->buf = NULL;
 548                                         goto stop;
 549                                 }
 550                                 err = txq->send_pending
 551                                         (txq->qp,
 552                                          addr,
 553                                          length,
 554                                          lkey,
 555                                          send_flags);
 556                         }
 557                         if (unlikely(err))
 558                                 goto stop;
 559 #ifdef MLX5_PMD_SOFT_COUNTERS
 560                         sent_size += length;
 561 #endif
 562                 } else {
 563 #if MLX5_PMD_SGE_WR_N > 1
 564                         struct ibv_sge sges[MLX5_PMD_SGE_WR_N];
 565                         struct tx_burst_sg_ret ret;
 566
 567                         ret = tx_burst_sg(txq, segs, elt, buf, elts_head,
 568                                           &sges);
 569                         if (ret.length == (unsigned int)-1)
 570                                 goto stop;
 571                         RTE_MBUF_PREFETCH_TO_FREE(elt_next->buf);
 572                         /* Put SG list into send queue. */
 573                         err = txq->send_pending_sg_list
 574                                 (txq->qp,
 575                                  sges,
 576                                  ret.num,
 577                                  send_flags);
 578                         if (unlikely(err))
 579                                 goto stop;
 580 #ifdef MLX5_PMD_SOFT_COUNTERS
 581                         sent_size += ret.length;
 582 #endif
 583 #else /* MLX5_PMD_SGE_WR_N > 1 */
 584                         DEBUG("%p: TX scattered buffers support not"
 585                               " compiled in", (void *)txq);
 586                         goto stop;
 587 #endif /* MLX5_PMD_SGE_WR_N > 1 */
 588                 }
 589                 elts_head = elts_head_next;
 590                 buf = buf_next;
 591 #ifdef MLX5_PMD_SOFT_COUNTERS
 592                 /* Increment sent bytes counter. */
 593                 txq->stats.obytes += sent_size;
 594 #endif
 595         }
 596 stop:
 597         /* Take a shortcut if nothing must be sent. */
 598         if (unlikely(i == 0))
 599                 return 0;
 600 #ifdef MLX5_PMD_SOFT_COUNTERS
 601         /* Increment sent packets counter. */
 602         txq->stats.opackets += i;
 603 #endif
 604         /* Ring QP doorbell. */
 605         err = txq->send_flush(txq->qp);
 606         if (unlikely(err)) {
 607                 /* A nonzero value is not supposed to be returned.
 608                  * Nothing can be done about it. */
 609                 DEBUG("%p: send_flush() failed with error %d",
 610                       (void *)txq, err);
 611         }
 612         txq->elts_head = elts_head;
 613         txq->elts_comp += elts_comp;
 614         txq->elts_comp_cd = elts_comp_cd;
 615         return i;
 616 }
 617
 618 /**
 619  * Translate RX completion flags to packet type.
 620  *
 621  * @param flags
 622  *   RX completion flags returned by poll_length_flags().
 623  *
 624  * @return
 625  *   Packet type for struct rte_mbuf.
 626  */
 627 static inline uint32_t
 628 rxq_cq_to_pkt_type(uint32_t flags)
 629 {
 630         uint32_t pkt_type;
 631
 632         if (flags & IBV_EXP_CQ_RX_TUNNEL_PACKET)
 633                 pkt_type =
 634                         TRANSPOSE(flags,
 635                                   IBV_EXP_CQ_RX_OUTER_IPV4_PACKET,
 636                                   RTE_PTYPE_L3_IPV4) |
 637                         TRANSPOSE(flags,
 638                                   IBV_EXP_CQ_RX_OUTER_IPV6_PACKET,
 639                                   RTE_PTYPE_L3_IPV6) |
 640                         TRANSPOSE(flags,
 641                                   IBV_EXP_CQ_RX_IPV4_PACKET,
 642                                   RTE_PTYPE_INNER_L3_IPV4) |
 643                         TRANSPOSE(flags,
 644                                   IBV_EXP_CQ_RX_IPV6_PACKET,
 645                                   RTE_PTYPE_INNER_L3_IPV6);
 646         else
 647                 pkt_type =
 648                         TRANSPOSE(flags,
 649                                   IBV_EXP_CQ_RX_IPV4_PACKET,
 650                                   RTE_PTYPE_L3_IPV4) |
 651                         TRANSPOSE(flags,
 652                                   IBV_EXP_CQ_RX_IPV6_PACKET,
 653                                   RTE_PTYPE_L3_IPV6);
 654         return pkt_type;
 655 }
 656
 657 /**
 658  * Translate RX completion flags to offload flags.
 659  *
 660  * @param[in] rxq
 661  *   Pointer to RX queue structure.
 662  * @param flags
 663  *   RX completion flags returned by poll_length_flags().
 664  *
 665  * @return
 666  *   Offload flags (ol_flags) for struct rte_mbuf.
 667  */
 668 static inline uint32_t
 669 rxq_cq_to_ol_flags(const struct rxq *rxq, uint32_t flags)
 670 {
 671         uint32_t ol_flags = 0;
 672
 673         if (rxq->csum)
 674                 ol_flags |=
 675                         TRANSPOSE(~flags,
 676                                   IBV_EXP_CQ_RX_IP_CSUM_OK,
 677                                   PKT_RX_IP_CKSUM_BAD) |
 678                         TRANSPOSE(~flags,
 679                                   IBV_EXP_CQ_RX_TCP_UDP_CSUM_OK,
 680                                   PKT_RX_L4_CKSUM_BAD);
 681         /*
 682          * PKT_RX_IP_CKSUM_BAD and PKT_RX_L4_CKSUM_BAD are used in place
 683          * of PKT_RX_EIP_CKSUM_BAD because the latter is not functional
 684          * (its value is 0).
 685          */
 686         if ((flags & IBV_EXP_CQ_RX_TUNNEL_PACKET) && (rxq->csum_l2tun))
 687                 ol_flags |=
 688                         TRANSPOSE(~flags,
 689                                   IBV_EXP_CQ_RX_OUTER_IP_CSUM_OK,
 690                                   PKT_RX_IP_CKSUM_BAD) |
 691                         TRANSPOSE(~flags,
 692                                   IBV_EXP_CQ_RX_OUTER_TCP_UDP_CSUM_OK,
 693                                   PKT_RX_L4_CKSUM_BAD);
 694         return ol_flags;
 695 }
 696
 697 /**
 698  * DPDK callback for RX with scattered packets support.
 699  *
 700  * @param dpdk_rxq
 701  *   Generic pointer to RX queue structure.
 702  * @param[out] pkts
 703  *   Array to store received packets.
 704  * @param pkts_n
 705  *   Maximum number of packets in array.
 706  *
 707  * @return
 708  *   Number of packets successfully received (<= pkts_n).
 709  */
 710 uint16_t
 711 mlx5_rx_burst_sp(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
 712 {
 713         struct rxq *rxq = (struct rxq *)dpdk_rxq;
 714         struct rxq_elt_sp (*elts)[rxq->elts_n] = rxq->elts.sp;
 715         const unsigned int elts_n = rxq->elts_n;
 716         unsigned int elts_head = rxq->elts_head;
 717         unsigned int i;
 718         unsigned int pkts_ret = 0;
 719         int ret;
 720
 721         if (unlikely(!rxq->sp))
 722                 return mlx5_rx_burst(dpdk_rxq, pkts, pkts_n);
 723         if (unlikely(elts == NULL)) /* See RTE_DEV_CMD_SET_MTU. */
 724                 return 0;
 725         for (i = 0; (i != pkts_n); ++i) {
 726                 struct rxq_elt_sp *elt = &(*elts)[elts_head];
 727                 unsigned int len;
 728                 unsigned int pkt_buf_len;
 729                 struct rte_mbuf *pkt_buf = NULL; /* Buffer returned in pkts. */
 730                 struct rte_mbuf **pkt_buf_next = &pkt_buf;
 731                 unsigned int seg_headroom = RTE_PKTMBUF_HEADROOM;
 732                 unsigned int j = 0;
 733                 uint32_t flags;
 734                 uint16_t vlan_tci;
 735
 736                 /* Sanity checks. */
 737                 assert(elts_head < rxq->elts_n);
 738                 assert(rxq->elts_head < rxq->elts_n);
 739                 ret = rxq->poll(rxq->cq, NULL, NULL, &flags, &vlan_tci);
 740                 if (unlikely(ret < 0)) {
 741                         struct ibv_wc wc;
 742                         int wcs_n;
 743
 744                         DEBUG("rxq=%p, poll_length() failed (ret=%d)",
 745                               (void *)rxq, ret);
 746                         /* ibv_poll_cq() must be used in case of failure. */
 747                         wcs_n = ibv_poll_cq(rxq->cq, 1, &wc);
 748                         if (unlikely(wcs_n == 0))
 749                                 break;
 750                         if (unlikely(wcs_n < 0)) {
 751                                 DEBUG("rxq=%p, ibv_poll_cq() failed (wcs_n=%d)",
 752                                       (void *)rxq, wcs_n);
 753                                 break;
 754                         }
 755                         assert(wcs_n == 1);
 756                         if (unlikely(wc.status != IBV_WC_SUCCESS)) {
 757                                 /* Whatever, just repost the offending WR. */
 758                                 DEBUG("rxq=%p, wr_id=%" PRIu64 ": bad work"
 759                                       " completion status (%d): %s",
 760                                       (void *)rxq, wc.wr_id, wc.status,
 761                                       ibv_wc_status_str(wc.status));
 762 #ifdef MLX5_PMD_SOFT_COUNTERS
 763                                 /* Increment dropped packets counter. */
 764                                 ++rxq->stats.idropped;
 765 #endif
 766                                 goto repost;
 767                         }
 768                         ret = wc.byte_len;
 769                 }
 770                 if (ret == 0)
 771                         break;
 772                 len = ret;
 773                 pkt_buf_len = len;
 774                 /*
 775                  * Replace spent segments with new ones, concatenate and
 776                  * return them as pkt_buf.
 777                  */
 778                 while (1) {
 779                         struct ibv_sge *sge = &elt->sges[j];
 780                         struct rte_mbuf *seg = elt->bufs[j];
 781                         struct rte_mbuf *rep;
 782                         unsigned int seg_tailroom;
 783
 784                         assert(seg != NULL);
 785                         /*
 786                          * Fetch initial bytes of packet descriptor into a
 787                          * cacheline while allocating rep.
 788                          */
 789                         rte_prefetch0(seg);
 790                         rep = __rte_mbuf_raw_alloc(rxq->mp);
 791                         if (unlikely(rep == NULL)) {
 792                                 /*
 793                                  * Unable to allocate a replacement mbuf,
 794                                  * repost WR.
 795                                  */
 796                                 DEBUG("rxq=%p: can't allocate a new mbuf",
 797                                       (void *)rxq);
 798                                 if (pkt_buf != NULL) {
 799                                         *pkt_buf_next = NULL;
 800                                         rte_pktmbuf_free(pkt_buf);
 801                                 }
 802                                 /* Increment out of memory counters. */
 803                                 ++rxq->stats.rx_nombuf;
 804                                 ++rxq->priv->dev->data->rx_mbuf_alloc_failed;
 805                                 goto repost;
 806                         }
 807 #ifndef NDEBUG
 808                         /* Poison user-modifiable fields in rep. */
 809                         NEXT(rep) = (void *)((uintptr_t)-1);
 810                         SET_DATA_OFF(rep, 0xdead);
 811                         DATA_LEN(rep) = 0xd00d;
 812                         PKT_LEN(rep) = 0xdeadd00d;
 813                         NB_SEGS(rep) = 0x2a;
 814                         PORT(rep) = 0x2a;
 815                         rep->ol_flags = -1;
 816 #endif
 817                         assert(rep->buf_len == seg->buf_len);
 818                         assert(rep->buf_len == rxq->mb_len);
 819                         /* Reconfigure sge to use rep instead of seg. */
 820                         assert(sge->lkey == rxq->mr->lkey);
 821                         sge->addr = ((uintptr_t)rep->buf_addr + seg_headroom);
 822                         elt->bufs[j] = rep;
 823                         ++j;
 824                         /* Update pkt_buf if it's the first segment, or link
 825                          * seg to the previous one and update pkt_buf_next. */
 826                         *pkt_buf_next = seg;
 827                         pkt_buf_next = &NEXT(seg);
 828                         /* Update seg information. */
 829                         seg_tailroom = (seg->buf_len - seg_headroom);
 830                         assert(sge->length == seg_tailroom);
 831                         SET_DATA_OFF(seg, seg_headroom);
 832                         if (likely(len <= seg_tailroom)) {
 833                                 /* Last segment. */
 834                                 DATA_LEN(seg) = len;
 835                                 PKT_LEN(seg) = len;
 836                                 /* Sanity check. */
 837                                 assert(rte_pktmbuf_headroom(seg) ==
 838                                        seg_headroom);
 839                                 assert(rte_pktmbuf_tailroom(seg) ==
 840                                        (seg_tailroom - len));
 841                                 break;
 842                         }
 843                         DATA_LEN(seg) = seg_tailroom;
 844                         PKT_LEN(seg) = seg_tailroom;
 845                         /* Sanity check. */
 846                         assert(rte_pktmbuf_headroom(seg) == seg_headroom);
 847                         assert(rte_pktmbuf_tailroom(seg) == 0);
 848                         /* Fix len and clear headroom for next segments. */
 849                         len -= seg_tailroom;
 850                         seg_headroom = 0;
 851                 }
 852                 /* Update head and tail segments. */
 853                 *pkt_buf_next = NULL;
 854                 assert(pkt_buf != NULL);
 855                 assert(j != 0);
 856                 NB_SEGS(pkt_buf) = j;
 857                 PORT(pkt_buf) = rxq->port_id;
 858                 PKT_LEN(pkt_buf) = pkt_buf_len;
 859                 if (rxq->csum | rxq->csum_l2tun | rxq->vlan_strip) {
 860                         pkt_buf->packet_type = rxq_cq_to_pkt_type(flags);
 861                         pkt_buf->ol_flags = rxq_cq_to_ol_flags(rxq, flags);
 862 #ifdef HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS
 863                         if (flags & IBV_EXP_CQ_RX_CVLAN_STRIPPED_V1) {
 864                                 pkt_buf->ol_flags |= PKT_RX_VLAN_PKT;
 865                                 pkt_buf->vlan_tci = vlan_tci;
 866                         }
 867 #endif /* HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS */
 868                 }
 869
 870                 /* Return packet. */
 871                 *(pkts++) = pkt_buf;
 872                 ++pkts_ret;
 873 #ifdef MLX5_PMD_SOFT_COUNTERS
 874                 /* Increment bytes counter. */
 875                 rxq->stats.ibytes += pkt_buf_len;
 876 #endif
 877 repost:
 878                 ret = rxq->recv(rxq->wq, elt->sges, RTE_DIM(elt->sges));
 879                 if (unlikely(ret)) {
 880                         /* Inability to repost WRs is fatal. */
 881                         DEBUG("%p: recv_sg_list(): failed (ret=%d)",
 882                               (void *)rxq->priv,
 883                               ret);
 884                         abort();
 885                 }
 886                 if (++elts_head >= elts_n)
 887                         elts_head = 0;
 888                 continue;
 889         }
 890         if (unlikely(i == 0))
 891                 return 0;
 892         rxq->elts_head = elts_head;
 893 #ifdef MLX5_PMD_SOFT_COUNTERS
 894         /* Increment packets counter. */
 895         rxq->stats.ipackets += pkts_ret;
 896 #endif
 897         return pkts_ret;
 898 }
 899
 900 /**
 901  * DPDK callback for RX.
 902  *
 903  * The following function is the same as mlx5_rx_burst_sp(), except it doesn't
 904  * manage scattered packets. Improves performance when MRU is lower than the
 905  * size of the first segment.
 906  *
 907  * @param dpdk_rxq
 908  *   Generic pointer to RX queue structure.
 909  * @param[out] pkts
 910  *   Array to store received packets.
 911  * @param pkts_n
 912  *   Maximum number of packets in array.
 913  *
 914  * @return
 915  *   Number of packets successfully received (<= pkts_n).
 916  */
 917 uint16_t
 918 mlx5_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
 919 {
 920         struct rxq *rxq = (struct rxq *)dpdk_rxq;
 921         struct rxq_elt (*elts)[rxq->elts_n] = rxq->elts.no_sp;
 922         const unsigned int elts_n = rxq->elts_n;
 923         unsigned int elts_head = rxq->elts_head;
 924         struct ibv_sge sges[pkts_n];
 925         unsigned int i;
 926         unsigned int pkts_ret = 0;
 927         int ret;
 928
 929         if (unlikely(rxq->sp))
 930                 return mlx5_rx_burst_sp(dpdk_rxq, pkts, pkts_n);
 931         for (i = 0; (i != pkts_n); ++i) {
 932                 struct rxq_elt *elt = &(*elts)[elts_head];
 933                 unsigned int len;
 934                 struct rte_mbuf *seg = elt->buf;
 935                 struct rte_mbuf *rep;
 936                 uint32_t flags;
 937                 uint16_t vlan_tci;
 938
 939                 /* Sanity checks. */
 940                 assert(seg != NULL);
 941                 assert(elts_head < rxq->elts_n);
 942                 assert(rxq->elts_head < rxq->elts_n);
 943                 /*
 944                  * Fetch initial bytes of packet descriptor into a
 945                  * cacheline while allocating rep.
 946                  */
 947                 rte_prefetch0(seg);
 948                 rte_prefetch0(&seg->cacheline1);
 949                 ret = rxq->poll(rxq->cq, NULL, NULL, &flags, &vlan_tci);
 950                 if (unlikely(ret < 0)) {
 951                         struct ibv_wc wc;
 952                         int wcs_n;
 953
 954                         DEBUG("rxq=%p, poll_length() failed (ret=%d)",
 955                               (void *)rxq, ret);
 956                         /* ibv_poll_cq() must be used in case of failure. */
 957                         wcs_n = ibv_poll_cq(rxq->cq, 1, &wc);
 958                         if (unlikely(wcs_n == 0))
 959                                 break;
 960                         if (unlikely(wcs_n < 0)) {
 961                                 DEBUG("rxq=%p, ibv_poll_cq() failed (wcs_n=%d)",
 962                                       (void *)rxq, wcs_n);
 963                                 break;
 964                         }
 965                         assert(wcs_n == 1);
 966                         if (unlikely(wc.status != IBV_WC_SUCCESS)) {
 967                                 /* Whatever, just repost the offending WR. */
 968                                 DEBUG("rxq=%p, wr_id=%" PRIu64 ": bad work"
 969                                       " completion status (%d): %s",
 970                                       (void *)rxq, wc.wr_id, wc.status,
 971                                       ibv_wc_status_str(wc.status));
 972 #ifdef MLX5_PMD_SOFT_COUNTERS
 973                                 /* Increment dropped packets counter. */
 974                                 ++rxq->stats.idropped;
 975 #endif
 976                                 /* Add SGE to array for repost. */
 977                                 sges[i] = elt->sge;
 978                                 goto repost;
 979                         }
 980                         ret = wc.byte_len;
 981                 }
 982                 if (ret == 0)
 983                         break;
 984                 len = ret;
 985                 rep = __rte_mbuf_raw_alloc(rxq->mp);
 986                 if (unlikely(rep == NULL)) {
 987                         /*
 988                          * Unable to allocate a replacement mbuf,
 989                          * repost WR.
 990                          */
 991                         DEBUG("rxq=%p: can't allocate a new mbuf",
 992                               (void *)rxq);
 993                         /* Increment out of memory counters. */
 994                         ++rxq->stats.rx_nombuf;
 995                         ++rxq->priv->dev->data->rx_mbuf_alloc_failed;
 996                         goto repost;
 997                 }
 998
 999                 /* Reconfigure sge to use rep instead of seg. */
1000                 elt->sge.addr = (uintptr_t)rep->buf_addr + RTE_PKTMBUF_HEADROOM;
1001                 assert(elt->sge.lkey == rxq->mr->lkey);
1002                 elt->buf = rep;
1003
1004                 /* Add SGE to array for repost. */
1005                 sges[i] = elt->sge;
1006
1007                 /* Update seg information. */
1008                 SET_DATA_OFF(seg, RTE_PKTMBUF_HEADROOM);
1009                 NB_SEGS(seg) = 1;
1010                 PORT(seg) = rxq->port_id;
1011                 NEXT(seg) = NULL;
1012                 PKT_LEN(seg) = len;
1013                 DATA_LEN(seg) = len;
1014                 if (rxq->csum | rxq->csum_l2tun | rxq->vlan_strip) {
1015                         seg->packet_type = rxq_cq_to_pkt_type(flags);
1016                         seg->ol_flags = rxq_cq_to_ol_flags(rxq, flags);
1017 #ifdef HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS
1018                         if (flags & IBV_EXP_CQ_RX_CVLAN_STRIPPED_V1) {
1019                                 seg->ol_flags |= PKT_RX_VLAN_PKT;
1020                                 seg->vlan_tci = vlan_tci;
1021                         }
1022 #endif /* HAVE_EXP_DEVICE_ATTR_VLAN_OFFLOADS */
1023                 }
1024                 /* Return packet. */
1025                 *(pkts++) = seg;
1026                 ++pkts_ret;
1027 #ifdef MLX5_PMD_SOFT_COUNTERS
1028                 /* Increment bytes counter. */
1029                 rxq->stats.ibytes += len;
1030 #endif
1031 repost:
1032                 if (++elts_head >= elts_n)
1033                         elts_head = 0;
1034                 continue;
1035         }
1036         if (unlikely(i == 0))
1037                 return 0;
1038         /* Repost WRs. */
1039 #ifdef DEBUG_RECV
1040         DEBUG("%p: reposting %u WRs", (void *)rxq, i);
1041 #endif
1042         ret = rxq->recv(rxq->wq, sges, i);
1043         if (unlikely(ret)) {
1044                 /* Inability to repost WRs is fatal. */
1045                 DEBUG("%p: recv_burst(): failed (ret=%d)",
1046                       (void *)rxq->priv,
1047                       ret);
1048                 abort();
1049         }
1050         rxq->elts_head = elts_head;
1051 #ifdef MLX5_PMD_SOFT_COUNTERS
1052         /* Increment packets counter. */
1053         rxq->stats.ipackets += pkts_ret;
1054 #endif
1055         return pkts_ret;
1056 }
1057
1058 /**
1059  * Dummy DPDK callback for TX.
1060  *
1061  * This function is used to temporarily replace the real callback during
1062  * unsafe control operations on the queue, or in case of error.
1063  *
1064  * @param dpdk_txq
1065  *   Generic pointer to TX queue structure.
1066  * @param[in] pkts
1067  *   Packets to transmit.
1068  * @param pkts_n
1069  *   Number of packets in array.
1070  *
1071  * @return
1072  *   Number of packets successfully transmitted (<= pkts_n).
1073  */
1074 uint16_t
1075 removed_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
1076 {
1077         (void)dpdk_txq;
1078         (void)pkts;
1079         (void)pkts_n;
1080         return 0;
1081 }
1082
1083 /**
1084  * Dummy DPDK callback for RX.
1085  *
1086  * This function is used to temporarily replace the real callback during
1087  * unsafe control operations on the queue, or in case of error.
1088  *
1089  * @param dpdk_rxq
1090  *   Generic pointer to RX queue structure.
1091  * @param[out] pkts
1092  *   Array to store received packets.
1093  * @param pkts_n
1094  *   Maximum number of packets in array.
1095  *
1096  * @return
1097  *   Number of packets successfully received (<= pkts_n).
1098  */
1099 uint16_t
1100 removed_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
1101 {
1102         (void)dpdk_rxq;
1103         (void)pkts;
1104         (void)pkts_n;
1105         return 0;
1106 }