drivers/net/mlx4/mlx4_rxtx.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright 2017 6WIND S.A.
   5  *   Copyright 2017 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 /**
  35  * @file
  36  * Data plane functions for mlx4 driver.
  37  */
  38
  39 #include <assert.h>
  40 #include <stdint.h>
  41 #include <string.h>
  42
  43 /* Verbs headers do not support -pedantic. */
  44 #ifdef PEDANTIC
  45 #pragma GCC diagnostic ignored "-Wpedantic"
  46 #endif
  47 #include <infiniband/verbs.h>
  48 #ifdef PEDANTIC
  49 #pragma GCC diagnostic error "-Wpedantic"
  50 #endif
  51
  52 #include <rte_branch_prediction.h>
  53 #include <rte_common.h>
  54 #include <rte_io.h>
  55 #include <rte_mbuf.h>
  56 #include <rte_mempool.h>
  57 #include <rte_prefetch.h>
  58
  59 #include "mlx4.h"
  60 #include "mlx4_prm.h"
  61 #include "mlx4_rxtx.h"
  62 #include "mlx4_utils.h"
  63
  64 /**
  65  * Pointer-value pair structure used in tx_post_send for saving the first
  66  * DWORD (32 byte) of a TXBB.
  67  */
  68 struct pv {
  69         volatile struct mlx4_wqe_data_seg *dseg;
  70         uint32_t val;
  71 };
  72
  73 /** A table to translate Rx completion flags to packet type. */
  74 uint32_t mlx4_ptype_table[0x100] __rte_cache_aligned = {
  75         /*
  76          * The index to the array should have:
  77          *  bit[7] - MLX4_CQE_L2_TUNNEL
  78          *  bit[6] - MLX4_CQE_L2_TUNNEL_IPV4
  79          *  bit[5] - MLX4_CQE_STATUS_UDP
  80          *  bit[4] - MLX4_CQE_STATUS_TCP
  81          *  bit[3] - MLX4_CQE_STATUS_IPV4OPT
  82          *  bit[2] - MLX4_CQE_STATUS_IPV6
  83          *  bit[1] - MLX4_CQE_STATUS_IPV4F
  84          *  bit[0] - MLX4_CQE_STATUS_IPV4
  85          * giving a total of up to 256 entries.
  86          */
  87         [0x00] = RTE_PTYPE_L2_ETHER,
  88         [0x01] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
  89         [0x02] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
  90                      RTE_PTYPE_L4_FRAG,
  91         [0x03] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
  92                      RTE_PTYPE_L4_FRAG,
  93         [0x04] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
  94         [0x09] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT,
  95         [0x0a] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT |
  96                      RTE_PTYPE_L4_FRAG,
  97         [0x11] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
  98                      RTE_PTYPE_L4_TCP,
  99         [0x12] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 100                      RTE_PTYPE_L4_TCP,
 101         [0x14] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 102                      RTE_PTYPE_L4_TCP,
 103         [0x18] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT |
 104                      RTE_PTYPE_L4_TCP,
 105         [0x19] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT |
 106                      RTE_PTYPE_L4_TCP,
 107         [0x1a] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT |
 108                      RTE_PTYPE_L4_TCP,
 109         [0x21] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 110                      RTE_PTYPE_L4_UDP,
 111         [0x22] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 112                      RTE_PTYPE_L4_UDP,
 113         [0x24] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 114                      RTE_PTYPE_L4_UDP,
 115         [0x28] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT |
 116                      RTE_PTYPE_L4_UDP,
 117         [0x29] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT |
 118                      RTE_PTYPE_L4_UDP,
 119         [0x2a] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT |
 120                      RTE_PTYPE_L4_UDP,
 121         /* Tunneled - L3 IPV6 */
 122         [0x80] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN,
 123         [0x81] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 124                      RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
 125         [0x82] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 126                      RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 127                      RTE_PTYPE_INNER_L4_FRAG,
 128         [0x83] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 129                      RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 130                      RTE_PTYPE_INNER_L4_FRAG,
 131         [0x84] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 132                      RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
 133         [0x88] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 134                      RTE_PTYPE_INNER_L3_IPV4_EXT,
 135         [0x89] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 136                      RTE_PTYPE_INNER_L3_IPV4_EXT,
 137         [0x8a] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 138                      RTE_PTYPE_INNER_L3_IPV4_EXT | RTE_PTYPE_INNER_L4_FRAG,
 139         /* Tunneled - L3 IPV6, TCP */
 140         [0x91] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 141                      RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 142                      RTE_PTYPE_INNER_L4_TCP,
 143         [0x92] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 144                      RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 145                      RTE_PTYPE_INNER_L4_FRAG |
 146                      RTE_PTYPE_INNER_L4_TCP,
 147         [0x93] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 148                      RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 149                      RTE_PTYPE_INNER_L4_FRAG |
 150                      RTE_PTYPE_INNER_L4_TCP,
 151         [0x94] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 152                      RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
 153                      RTE_PTYPE_INNER_L4_TCP,
 154         [0x98] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 155                      RTE_PTYPE_INNER_L3_IPV4_EXT |
 156                      RTE_PTYPE_INNER_L4_TCP,
 157         [0x99] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 158                      RTE_PTYPE_INNER_L3_IPV4_EXT |
 159                      RTE_PTYPE_INNER_L4_TCP,
 160         [0x9a] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 161                      RTE_PTYPE_INNER_L3_IPV4_EXT | RTE_PTYPE_INNER_L4_FRAG |
 162                      RTE_PTYPE_INNER_L4_TCP,
 163         /* Tunneled - L3 IPV6, UDP */
 164         [0xa1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 165                      RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 166                      RTE_PTYPE_INNER_L4_UDP,
 167         [0xa2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 168                      RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 169                      RTE_PTYPE_INNER_L4_FRAG |
 170                      RTE_PTYPE_INNER_L4_UDP,
 171         [0xa3] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 172                      RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 173                      RTE_PTYPE_INNER_L4_FRAG |
 174                      RTE_PTYPE_INNER_L4_UDP,
 175         [0xa4] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 176                      RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
 177                      RTE_PTYPE_INNER_L4_UDP,
 178         [0xa8] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 179                      RTE_PTYPE_INNER_L3_IPV4_EXT |
 180                      RTE_PTYPE_INNER_L4_UDP,
 181         [0xa9] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 182                      RTE_PTYPE_INNER_L3_IPV4_EXT |
 183                      RTE_PTYPE_INNER_L4_UDP,
 184         [0xaa] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 185                      RTE_PTYPE_INNER_L3_IPV4_EXT | RTE_PTYPE_INNER_L4_FRAG |
 186                      RTE_PTYPE_INNER_L4_UDP,
 187         /* Tunneled - L3 IPV4 */
 188         [0xc0] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN,
 189         [0xc1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 190                      RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN,
 191         [0xc2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 192                      RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 193                      RTE_PTYPE_INNER_L4_FRAG,
 194         [0xc3] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 195                      RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 196                      RTE_PTYPE_INNER_L4_FRAG,
 197         [0xc4] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 198                      RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN,
 199         [0xc8] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 200                      RTE_PTYPE_INNER_L3_IPV4_EXT,
 201         [0xc9] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 202                      RTE_PTYPE_INNER_L3_IPV4_EXT,
 203         [0xca] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 204                      RTE_PTYPE_INNER_L3_IPV4_EXT |
 205                      RTE_PTYPE_INNER_L4_FRAG,
 206         /* Tunneled - L3 IPV4, TCP */
 207         [0xd0] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 208                      RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
 209                      RTE_PTYPE_INNER_L4_TCP,
 210         [0xd1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 211                      RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 212                      RTE_PTYPE_INNER_L4_TCP,
 213         [0xd2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 214                      RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 215                      RTE_PTYPE_INNER_L4_FRAG |
 216                      RTE_PTYPE_INNER_L4_TCP,
 217         [0xd3] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 218                      RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 219                      RTE_PTYPE_INNER_L4_FRAG |
 220                      RTE_PTYPE_INNER_L4_TCP,
 221         [0xd4] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 222                      RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
 223                      RTE_PTYPE_INNER_L4_TCP,
 224         [0xd8] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 225                      RTE_PTYPE_INNER_L3_IPV4_EXT |
 226                      RTE_PTYPE_INNER_L4_TCP,
 227         [0xd9] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 228                      RTE_PTYPE_INNER_L3_IPV4_EXT |
 229                      RTE_PTYPE_INNER_L4_TCP,
 230         [0xda] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 231                      RTE_PTYPE_INNER_L3_IPV4_EXT | RTE_PTYPE_INNER_L4_FRAG |
 232                      RTE_PTYPE_INNER_L4_TCP,
 233         /* Tunneled - L3 IPV4, UDP */
 234         [0xe0] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 235                      RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
 236                      RTE_PTYPE_INNER_L4_UDP,
 237         [0xe1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 238                      RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 239                      RTE_PTYPE_INNER_L4_UDP,
 240         [0xe2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 241                      RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 242                      RTE_PTYPE_INNER_L4_FRAG |
 243                      RTE_PTYPE_INNER_L4_UDP,
 244         [0xe3] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 245                      RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 246                      RTE_PTYPE_INNER_L4_FRAG |
 247                      RTE_PTYPE_INNER_L4_UDP,
 248         [0xe4] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 249                      RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
 250                      RTE_PTYPE_INNER_L4_UDP,
 251         [0xe8] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 252                      RTE_PTYPE_INNER_L3_IPV4_EXT | RTE_PTYPE_INNER_L4_UDP,
 253         [0xe9] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 254                      RTE_PTYPE_INNER_L3_IPV4_EXT | RTE_PTYPE_INNER_L4_UDP,
 255         [0xea] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 256                      RTE_PTYPE_INNER_L3_IPV4_EXT | RTE_PTYPE_INNER_L4_FRAG |
 257                      RTE_PTYPE_INNER_L4_UDP,
 258 };
 259
 260 /**
 261  * Stamp TXBB burst so it won't be reused by the HW.
 262  *
 263  * Routine is used when freeing WQE used by the chip or when failing
 264  * building an WQ entry has failed leaving partial information on the queue.
 265  *
 266  * @param sq
 267  *   Pointer to the SQ structure.
 268  * @param start
 269  *   Pointer to the first TXBB to stamp.
 270  * @param end
 271  *   Pointer to the followed end TXBB to stamp.
 272  *
 273  * @return
 274  *   Stamping burst size in byte units.
 275  */
 276 static uint32_t
 277 mlx4_txq_stamp_freed_wqe(struct mlx4_sq *sq, volatile uint32_t *start,
 278                          volatile uint32_t *end)
 279 {
 280         uint32_t stamp = sq->stamp;
 281         int32_t size = (intptr_t)end - (intptr_t)start;
 282
 283         assert(start != end);
 284         /* Hold SQ ring wrap around. */
 285         if (size < 0) {
 286                 size = (int32_t)sq->size + size;
 287                 do {
 288                         *start = stamp;
 289                         start += MLX4_SQ_STAMP_DWORDS;
 290                 } while (start != (volatile uint32_t *)sq->eob);
 291                 start = (volatile uint32_t *)sq->buf;
 292                 /* Flip invalid stamping ownership. */
 293                 stamp ^= RTE_BE32(0x1 << MLX4_SQ_OWNER_BIT);
 294                 sq->stamp = stamp;
 295                 if (start == end)
 296                         return size;
 297         }
 298         do {
 299                 *start = stamp;
 300                 start += MLX4_SQ_STAMP_DWORDS;
 301         } while (start != end);
 302         return (uint32_t)size;
 303 }
 304
 305 /**
 306  * Manage Tx completions.
 307  *
 308  * When sending a burst, mlx4_tx_burst() posts several WRs.
 309  * To improve performance, a completion event is only required once every
 310  * MLX4_PMD_TX_PER_COMP_REQ sends. Doing so discards completion information
 311  * for other WRs, but this information would not be used anyway.
 312  *
 313  * @param txq
 314  *   Pointer to Tx queue structure.
 315  * @param elts_m
 316  *   Tx elements number mask.
 317  * @param sq
 318  *   Pointer to the SQ structure.
 319  */
 320 static void
 321 mlx4_txq_complete(struct txq *txq, const unsigned int elts_m,
 322                   struct mlx4_sq *sq)
 323 {
 324         unsigned int elts_tail = txq->elts_tail;
 325         struct mlx4_cq *cq = &txq->mcq;
 326         volatile struct mlx4_cqe *cqe;
 327         uint32_t completed;
 328         uint32_t cons_index = cq->cons_index;
 329         volatile uint32_t *first_txbb;
 330
 331         /*
 332          * Traverse over all CQ entries reported and handle each WQ entry
 333          * reported by them.
 334          */
 335         do {
 336                 cqe = (volatile struct mlx4_cqe *)mlx4_get_cqe(cq, cons_index);
 337                 if (unlikely(!!(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK) ^
 338                     !!(cons_index & cq->cqe_cnt)))
 339                         break;
 340 #ifndef NDEBUG
 341                 /*
 342                  * Make sure we read the CQE after we read the ownership bit.
 343                  */
 344                 rte_io_rmb();
 345                 if (unlikely((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) ==
 346                              MLX4_CQE_OPCODE_ERROR)) {
 347                         volatile struct mlx4_err_cqe *cqe_err =
 348                                 (volatile struct mlx4_err_cqe *)cqe;
 349                         ERROR("%p CQE error - vendor syndrome: 0x%x"
 350                               " syndrome: 0x%x\n",
 351                               (void *)txq, cqe_err->vendor_err,
 352                               cqe_err->syndrome);
 353                         break;
 354                 }
 355 #endif /* NDEBUG */
 356                 cons_index++;
 357         } while (1);
 358         completed = (cons_index - cq->cons_index) * txq->elts_comp_cd_init;
 359         if (unlikely(!completed))
 360                 return;
 361         /* First stamping address is the end of the last one. */
 362         first_txbb = (&(*txq->elts)[elts_tail & elts_m])->eocb;
 363         elts_tail += completed;
 364         /* The new tail element holds the end address. */
 365         sq->remain_size += mlx4_txq_stamp_freed_wqe(sq, first_txbb,
 366                 (&(*txq->elts)[elts_tail & elts_m])->eocb);
 367         /* Update CQ consumer index. */
 368         cq->cons_index = cons_index;
 369         *cq->set_ci_db = rte_cpu_to_be_32(cons_index & MLX4_CQ_DB_CI_MASK);
 370         txq->elts_tail = elts_tail;
 371 }
 372
 373 /**
 374  * Get memory pool (MP) from mbuf. If mbuf is indirect, the pool from which
 375  * the cloned mbuf is allocated is returned instead.
 376  *
 377  * @param buf
 378  *   Pointer to mbuf.
 379  *
 380  * @return
 381  *   Memory pool where data is located for given mbuf.
 382  */
 383 static struct rte_mempool *
 384 mlx4_txq_mb2mp(struct rte_mbuf *buf)
 385 {
 386         if (unlikely(RTE_MBUF_INDIRECT(buf)))
 387                 return rte_mbuf_from_indirect(buf)->pool;
 388         return buf->pool;
 389 }
 390
 391 /**
 392  * Write Tx data segment to the SQ.
 393  *
 394  * @param dseg
 395  *   Pointer to data segment in SQ.
 396  * @param lkey
 397  *   Memory region lkey.
 398  * @param addr
 399  *   Data address.
 400  * @param byte_count
 401  *   Big endian bytes count of the data to send.
 402  */
 403 static inline void
 404 mlx4_fill_tx_data_seg(volatile struct mlx4_wqe_data_seg *dseg,
 405                        uint32_t lkey, uintptr_t addr, rte_be32_t  byte_count)
 406 {
 407         dseg->addr = rte_cpu_to_be_64(addr);
 408         dseg->lkey = rte_cpu_to_be_32(lkey);
 409 #if RTE_CACHE_LINE_SIZE < 64
 410         /*
 411          * Need a barrier here before writing the byte_count
 412          * fields to make sure that all the data is visible
 413          * before the byte_count field is set.
 414          * Otherwise, if the segment begins a new cacheline,
 415          * the HCA prefetcher could grab the 64-byte chunk and
 416          * get a valid (!= 0xffffffff) byte count but stale
 417          * data, and end up sending the wrong data.
 418          */
 419         rte_io_wmb();
 420 #endif /* RTE_CACHE_LINE_SIZE */
 421         dseg->byte_count = byte_count;
 422 }
 423
 424 /**
 425  * Write data segments of multi-segment packet.
 426  *
 427  * @param buf
 428  *   Pointer to the first packet mbuf.
 429  * @param txq
 430  *   Pointer to Tx queue structure.
 431  * @param ctrl
 432  *   Pointer to the WQE control segment.
 433  *
 434  * @return
 435  *   Pointer to the next WQE control segment on success, NULL otherwise.
 436  */
 437 static volatile struct mlx4_wqe_ctrl_seg *
 438 mlx4_tx_burst_segs(struct rte_mbuf *buf, struct txq *txq,
 439                    volatile struct mlx4_wqe_ctrl_seg *ctrl)
 440 {
 441         struct pv *pv = (struct pv *)txq->bounce_buf;
 442         struct mlx4_sq *sq = &txq->msq;
 443         struct rte_mbuf *sbuf = buf;
 444         uint32_t lkey;
 445         int pv_counter = 0;
 446         int nb_segs = buf->nb_segs;
 447         uint32_t wqe_size;
 448         volatile struct mlx4_wqe_data_seg *dseg =
 449                 (volatile struct mlx4_wqe_data_seg *)(ctrl + 1);
 450
 451         ctrl->fence_size = 1 + nb_segs;
 452         wqe_size = RTE_ALIGN((uint32_t)(ctrl->fence_size << MLX4_SEG_SHIFT),
 453                              MLX4_TXBB_SIZE);
 454         /* Validate WQE size and WQE space in the send queue. */
 455         if (sq->remain_size < wqe_size ||
 456             wqe_size > MLX4_MAX_WQE_SIZE)
 457                 return NULL;
 458         /*
 459          * Fill the data segments with buffer information.
 460          * First WQE TXBB head segment is always control segment,
 461          * so jump to tail TXBB data segments code for the first
 462          * WQE data segments filling.
 463          */
 464         goto txbb_tail_segs;
 465 txbb_head_seg:
 466         /* Memory region key (big endian) for this memory pool. */
 467         lkey = mlx4_txq_mp2mr(txq, mlx4_txq_mb2mp(sbuf));
 468         if (unlikely(lkey == (uint32_t)-1)) {
 469                 DEBUG("%p: unable to get MP <-> MR association",
 470                       (void *)txq);
 471                 return NULL;
 472         }
 473         /* Handle WQE wraparound. */
 474         if (dseg >=
 475                 (volatile struct mlx4_wqe_data_seg *)sq->eob)
 476                 dseg = (volatile struct mlx4_wqe_data_seg *)
 477                         sq->buf;
 478         dseg->addr = rte_cpu_to_be_64(rte_pktmbuf_mtod(sbuf, uintptr_t));
 479         dseg->lkey = rte_cpu_to_be_32(lkey);
 480         /*
 481          * This data segment starts at the beginning of a new
 482          * TXBB, so we need to postpone its byte_count writing
 483          * for later.
 484          */
 485         pv[pv_counter].dseg = dseg;
 486         /*
 487          * Zero length segment is treated as inline segment
 488          * with zero data.
 489          */
 490         pv[pv_counter++].val = rte_cpu_to_be_32(sbuf->data_len ?
 491                                                 sbuf->data_len : 0x80000000);
 492         sbuf = sbuf->next;
 493         dseg++;
 494         nb_segs--;
 495 txbb_tail_segs:
 496         /* Jump to default if there are more than two segments remaining. */
 497         switch (nb_segs) {
 498         default:
 499                 lkey = mlx4_txq_mp2mr(txq, mlx4_txq_mb2mp(sbuf));
 500                 if (unlikely(lkey == (uint32_t)-1)) {
 501                         DEBUG("%p: unable to get MP <-> MR association",
 502                               (void *)txq);
 503                         return NULL;
 504                 }
 505                 mlx4_fill_tx_data_seg(dseg, lkey,
 506                                       rte_pktmbuf_mtod(sbuf, uintptr_t),
 507                                       rte_cpu_to_be_32(sbuf->data_len ?
 508                                                        sbuf->data_len :
 509                                                        0x80000000));
 510                 sbuf = sbuf->next;
 511                 dseg++;
 512                 nb_segs--;
 513                 /* fallthrough */
 514         case 2:
 515                 lkey = mlx4_txq_mp2mr(txq, mlx4_txq_mb2mp(sbuf));
 516                 if (unlikely(lkey == (uint32_t)-1)) {
 517                         DEBUG("%p: unable to get MP <-> MR association",
 518                               (void *)txq);
 519                         return NULL;
 520                 }
 521                 mlx4_fill_tx_data_seg(dseg, lkey,
 522                                       rte_pktmbuf_mtod(sbuf, uintptr_t),
 523                                       rte_cpu_to_be_32(sbuf->data_len ?
 524                                                        sbuf->data_len :
 525                                                        0x80000000));
 526                 sbuf = sbuf->next;
 527                 dseg++;
 528                 nb_segs--;
 529                 /* fallthrough */
 530         case 1:
 531                 lkey = mlx4_txq_mp2mr(txq, mlx4_txq_mb2mp(sbuf));
 532                 if (unlikely(lkey == (uint32_t)-1)) {
 533                         DEBUG("%p: unable to get MP <-> MR association",
 534                               (void *)txq);
 535                         return NULL;
 536                 }
 537                 mlx4_fill_tx_data_seg(dseg, lkey,
 538                                       rte_pktmbuf_mtod(sbuf, uintptr_t),
 539                                       rte_cpu_to_be_32(sbuf->data_len ?
 540                                                        sbuf->data_len :
 541                                                        0x80000000));
 542                 nb_segs--;
 543                 if (nb_segs) {
 544                         sbuf = sbuf->next;
 545                         dseg++;
 546                         goto txbb_head_seg;
 547                 }
 548                 /* fallthrough */
 549         case 0:
 550                 break;
 551         }
 552         /* Write the first DWORD of each TXBB save earlier. */
 553         if (pv_counter) {
 554                 /* Need a barrier here before writing the byte_count. */
 555                 rte_io_wmb();
 556                 for (--pv_counter; pv_counter  >= 0; pv_counter--)
 557                         pv[pv_counter].dseg->byte_count = pv[pv_counter].val;
 558         }
 559         sq->remain_size -= wqe_size;
 560         /* Align next WQE address to the next TXBB. */
 561         return (volatile struct mlx4_wqe_ctrl_seg *)
 562                 ((volatile uint8_t *)ctrl + wqe_size);
 563 }
 564
 565 /**
 566  * DPDK callback for Tx.
 567  *
 568  * @param dpdk_txq
 569  *   Generic pointer to Tx queue structure.
 570  * @param[in] pkts
 571  *   Packets to transmit.
 572  * @param pkts_n
 573  *   Number of packets in array.
 574  *
 575  * @return
 576  *   Number of packets successfully transmitted (<= pkts_n).
 577  */
 578 uint16_t
 579 mlx4_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
 580 {
 581         struct txq *txq = (struct txq *)dpdk_txq;
 582         unsigned int elts_head = txq->elts_head;
 583         const unsigned int elts_n = txq->elts_n;
 584         const unsigned int elts_m = elts_n - 1;
 585         unsigned int bytes_sent = 0;
 586         unsigned int i;
 587         unsigned int max = elts_head - txq->elts_tail;
 588         struct mlx4_sq *sq = &txq->msq;
 589         volatile struct mlx4_wqe_ctrl_seg *ctrl;
 590         struct txq_elt *elt;
 591
 592         assert(txq->elts_comp_cd != 0);
 593         if (likely(max >= txq->elts_comp_cd_init))
 594                 mlx4_txq_complete(txq, elts_m, sq);
 595         max = elts_n - max;
 596         assert(max >= 1);
 597         assert(max <= elts_n);
 598         /* Always leave one free entry in the ring. */
 599         --max;
 600         if (max > pkts_n)
 601                 max = pkts_n;
 602         elt = &(*txq->elts)[elts_head & elts_m];
 603         /* First Tx burst element saves the next WQE control segment. */
 604         ctrl = elt->wqe;
 605         for (i = 0; (i != max); ++i) {
 606                 struct rte_mbuf *buf = pkts[i];
 607                 struct txq_elt *elt_next = &(*txq->elts)[++elts_head & elts_m];
 608                 uint32_t owner_opcode = sq->owner_opcode;
 609                 volatile struct mlx4_wqe_data_seg *dseg =
 610                                 (volatile struct mlx4_wqe_data_seg *)(ctrl + 1);
 611                 volatile struct mlx4_wqe_ctrl_seg *ctrl_next;
 612                 union {
 613                         uint32_t flags;
 614                         uint16_t flags16[2];
 615                 } srcrb;
 616                 uint32_t lkey;
 617
 618                 /* Clean up old buffer. */
 619                 if (likely(elt->buf != NULL)) {
 620                         struct rte_mbuf *tmp = elt->buf;
 621
 622 #ifndef NDEBUG
 623                         /* Poisoning. */
 624                         memset(&elt->buf, 0x66, sizeof(struct rte_mbuf *));
 625 #endif
 626                         /* Faster than rte_pktmbuf_free(). */
 627                         do {
 628                                 struct rte_mbuf *next = tmp->next;
 629
 630                                 rte_pktmbuf_free_seg(tmp);
 631                                 tmp = next;
 632                         } while (tmp != NULL);
 633                 }
 634                 RTE_MBUF_PREFETCH_TO_FREE(elt_next->buf);
 635                 if (buf->nb_segs == 1) {
 636                         /* Validate WQE space in the send queue. */
 637                         if (sq->remain_size < MLX4_TXBB_SIZE) {
 638                                 elt->buf = NULL;
 639                                 break;
 640                         }
 641                         lkey = mlx4_txq_mp2mr(txq, mlx4_txq_mb2mp(buf));
 642                         if (unlikely(lkey == (uint32_t)-1)) {
 643                                 /* MR does not exist. */
 644                                 DEBUG("%p: unable to get MP <-> MR association",
 645                                       (void *)txq);
 646                                 elt->buf = NULL;
 647                                 break;
 648                         }
 649                         mlx4_fill_tx_data_seg(dseg++, lkey,
 650                                               rte_pktmbuf_mtod(buf, uintptr_t),
 651                                               rte_cpu_to_be_32(buf->data_len));
 652                         /* Set WQE size in 16-byte units. */
 653                         ctrl->fence_size = 0x2;
 654                         sq->remain_size -= MLX4_TXBB_SIZE;
 655                         /* Align next WQE address to the next TXBB. */
 656                         ctrl_next = ctrl + 0x4;
 657                 } else {
 658                         ctrl_next = mlx4_tx_burst_segs(buf, txq, ctrl);
 659                         if (!ctrl_next) {
 660                                 elt->buf = NULL;
 661                                 break;
 662                         }
 663                 }
 664                 /* Hold SQ ring wrap around. */
 665                 if ((volatile uint8_t *)ctrl_next >= sq->eob) {
 666                         ctrl_next = (volatile struct mlx4_wqe_ctrl_seg *)
 667                                 ((volatile uint8_t *)ctrl_next - sq->size);
 668                         /* Flip HW valid ownership. */
 669                         sq->owner_opcode ^= 0x1 << MLX4_SQ_OWNER_BIT;
 670                 }
 671                 /*
 672                  * For raw Ethernet, the SOLICIT flag is used to indicate
 673                  * that no ICRC should be calculated.
 674                  */
 675                 if (--txq->elts_comp_cd == 0) {
 676                         /* Save the completion burst end address. */
 677                         elt_next->eocb = (volatile uint32_t *)ctrl_next;
 678                         txq->elts_comp_cd = txq->elts_comp_cd_init;
 679                         srcrb.flags = RTE_BE32(MLX4_WQE_CTRL_SOLICIT |
 680                                                MLX4_WQE_CTRL_CQ_UPDATE);
 681                 } else {
 682                         srcrb.flags = RTE_BE32(MLX4_WQE_CTRL_SOLICIT);
 683                 }
 684                 /* Enable HW checksum offload if requested */
 685                 if (txq->csum &&
 686                     (buf->ol_flags &
 687                      (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM))) {
 688                         const uint64_t is_tunneled = (buf->ol_flags &
 689                                                       (PKT_TX_TUNNEL_GRE |
 690                                                        PKT_TX_TUNNEL_VXLAN));
 691
 692                         if (is_tunneled && txq->csum_l2tun) {
 693                                 owner_opcode |= MLX4_WQE_CTRL_IIP_HDR_CSUM |
 694                                                 MLX4_WQE_CTRL_IL4_HDR_CSUM;
 695                                 if (buf->ol_flags & PKT_TX_OUTER_IP_CKSUM)
 696                                         srcrb.flags |=
 697                                             RTE_BE32(MLX4_WQE_CTRL_IP_HDR_CSUM);
 698                         } else {
 699                                 srcrb.flags |=
 700                                         RTE_BE32(MLX4_WQE_CTRL_IP_HDR_CSUM |
 701                                                 MLX4_WQE_CTRL_TCP_UDP_CSUM);
 702                         }
 703                 }
 704                 if (txq->lb) {
 705                         /*
 706                          * Copy destination MAC address to the WQE, this allows
 707                          * loopback in eSwitch, so that VFs and PF can
 708                          * communicate with each other.
 709                          */
 710                         srcrb.flags16[0] = *(rte_pktmbuf_mtod(buf, uint16_t *));
 711                         ctrl->imm = *(rte_pktmbuf_mtod_offset(buf, uint32_t *,
 712                                               sizeof(uint16_t)));
 713                 } else {
 714                         ctrl->imm = 0;
 715                 }
 716                 ctrl->srcrb_flags = srcrb.flags;
 717                 /*
 718                  * Make sure descriptor is fully written before
 719                  * setting ownership bit (because HW can start
 720                  * executing as soon as we do).
 721                  */
 722                 rte_io_wmb();
 723                 ctrl->owner_opcode = rte_cpu_to_be_32(owner_opcode);
 724                 elt->buf = buf;
 725                 bytes_sent += buf->pkt_len;
 726                 ctrl = ctrl_next;
 727                 elt = elt_next;
 728         }
 729         /* Take a shortcut if nothing must be sent. */
 730         if (unlikely(i == 0))
 731                 return 0;
 732         /* Save WQE address of the next Tx burst element. */
 733         elt->wqe = ctrl;
 734         /* Increment send statistics counters. */
 735         txq->stats.opackets += i;
 736         txq->stats.obytes += bytes_sent;
 737         /* Make sure that descriptors are written before doorbell record. */
 738         rte_wmb();
 739         /* Ring QP doorbell. */
 740         rte_write32(txq->msq.doorbell_qpn, txq->msq.db);
 741         txq->elts_head += i;
 742         return i;
 743 }
 744
 745 /**
 746  * Translate Rx completion flags to packet type.
 747  *
 748  * @param[in] cqe
 749  *   Pointer to CQE.
 750  *
 751  * @return
 752  *   Packet type for struct rte_mbuf.
 753  */
 754 static inline uint32_t
 755 rxq_cq_to_pkt_type(volatile struct mlx4_cqe *cqe,
 756                    uint32_t l2tun_offload)
 757 {
 758         uint8_t idx = 0;
 759         uint32_t pinfo = rte_be_to_cpu_32(cqe->vlan_my_qpn);
 760         uint32_t status = rte_be_to_cpu_32(cqe->status);
 761
 762         /*
 763          * The index to the array should have:
 764          *  bit[7] - MLX4_CQE_L2_TUNNEL
 765          *  bit[6] - MLX4_CQE_L2_TUNNEL_IPV4
 766          */
 767         if (l2tun_offload && (pinfo & MLX4_CQE_L2_TUNNEL))
 768                 idx |= ((pinfo & MLX4_CQE_L2_TUNNEL) >> 20) |
 769                        ((pinfo & MLX4_CQE_L2_TUNNEL_IPV4) >> 19);
 770         /*
 771          * The index to the array should have:
 772          *  bit[5] - MLX4_CQE_STATUS_UDP
 773          *  bit[4] - MLX4_CQE_STATUS_TCP
 774          *  bit[3] - MLX4_CQE_STATUS_IPV4OPT
 775          *  bit[2] - MLX4_CQE_STATUS_IPV6
 776          *  bit[1] - MLX4_CQE_STATUS_IPV4F
 777          *  bit[0] - MLX4_CQE_STATUS_IPV4
 778          * giving a total of up to 256 entries.
 779          */
 780         idx |= ((status & MLX4_CQE_STATUS_PTYPE_MASK) >> 22);
 781         return mlx4_ptype_table[idx];
 782 }
 783
 784 /**
 785  * Translate Rx completion flags to offload flags.
 786  *
 787  * @param flags
 788  *   Rx completion flags returned by mlx4_cqe_flags().
 789  * @param csum
 790  *   Whether Rx checksums are enabled.
 791  * @param csum_l2tun
 792  *   Whether Rx L2 tunnel checksums are enabled.
 793  *
 794  * @return
 795  *   Offload flags (ol_flags) in mbuf format.
 796  */
 797 static inline uint32_t
 798 rxq_cq_to_ol_flags(uint32_t flags, int csum, int csum_l2tun)
 799 {
 800         uint32_t ol_flags = 0;
 801
 802         if (csum)
 803                 ol_flags |=
 804                         mlx4_transpose(flags,
 805                                        MLX4_CQE_STATUS_IP_HDR_CSUM_OK,
 806                                        PKT_RX_IP_CKSUM_GOOD) |
 807                         mlx4_transpose(flags,
 808                                        MLX4_CQE_STATUS_TCP_UDP_CSUM_OK,
 809                                        PKT_RX_L4_CKSUM_GOOD);
 810         if ((flags & MLX4_CQE_L2_TUNNEL) && csum_l2tun)
 811                 ol_flags |=
 812                         mlx4_transpose(flags,
 813                                        MLX4_CQE_L2_TUNNEL_IPOK,
 814                                        PKT_RX_IP_CKSUM_GOOD) |
 815                         mlx4_transpose(flags,
 816                                        MLX4_CQE_L2_TUNNEL_L4_CSUM,
 817                                        PKT_RX_L4_CKSUM_GOOD);
 818         return ol_flags;
 819 }
 820
 821 /**
 822  * Extract checksum information from CQE flags.
 823  *
 824  * @param cqe
 825  *   Pointer to CQE structure.
 826  * @param csum
 827  *   Whether Rx checksums are enabled.
 828  * @param csum_l2tun
 829  *   Whether Rx L2 tunnel checksums are enabled.
 830  *
 831  * @return
 832  *   CQE checksum information.
 833  */
 834 static inline uint32_t
 835 mlx4_cqe_flags(volatile struct mlx4_cqe *cqe, int csum, int csum_l2tun)
 836 {
 837         uint32_t flags = 0;
 838
 839         /*
 840          * The relevant bits are in different locations on their
 841          * CQE fields therefore we can join them in one 32bit
 842          * variable.
 843          */
 844         if (csum)
 845                 flags = (rte_be_to_cpu_32(cqe->status) &
 846                          MLX4_CQE_STATUS_IPV4_CSUM_OK);
 847         if (csum_l2tun)
 848                 flags |= (rte_be_to_cpu_32(cqe->vlan_my_qpn) &
 849                           (MLX4_CQE_L2_TUNNEL |
 850                            MLX4_CQE_L2_TUNNEL_IPOK |
 851                            MLX4_CQE_L2_TUNNEL_L4_CSUM |
 852                            MLX4_CQE_L2_TUNNEL_IPV4));
 853         return flags;
 854 }
 855
 856 /**
 857  * Poll one CQE from CQ.
 858  *
 859  * @param rxq
 860  *   Pointer to the receive queue structure.
 861  * @param[out] out
 862  *   Just polled CQE.
 863  *
 864  * @return
 865  *   Number of bytes of the CQE, 0 in case there is no completion.
 866  */
 867 static unsigned int
 868 mlx4_cq_poll_one(struct rxq *rxq, volatile struct mlx4_cqe **out)
 869 {
 870         int ret = 0;
 871         volatile struct mlx4_cqe *cqe = NULL;
 872         struct mlx4_cq *cq = &rxq->mcq;
 873
 874         cqe = (volatile struct mlx4_cqe *)mlx4_get_cqe(cq, cq->cons_index);
 875         if (!!(cqe->owner_sr_opcode & MLX4_CQE_OWNER_MASK) ^
 876             !!(cq->cons_index & cq->cqe_cnt))
 877                 goto out;
 878         /*
 879          * Make sure we read CQ entry contents after we've checked the
 880          * ownership bit.
 881          */
 882         rte_rmb();
 883         assert(!(cqe->owner_sr_opcode & MLX4_CQE_IS_SEND_MASK));
 884         assert((cqe->owner_sr_opcode & MLX4_CQE_OPCODE_MASK) !=
 885                MLX4_CQE_OPCODE_ERROR);
 886         ret = rte_be_to_cpu_32(cqe->byte_cnt);
 887         ++cq->cons_index;
 888 out:
 889         *out = cqe;
 890         return ret;
 891 }
 892
 893 /**
 894  * DPDK callback for Rx with scattered packets support.
 895  *
 896  * @param dpdk_rxq
 897  *   Generic pointer to Rx queue structure.
 898  * @param[out] pkts
 899  *   Array to store received packets.
 900  * @param pkts_n
 901  *   Maximum number of packets in array.
 902  *
 903  * @return
 904  *   Number of packets successfully received (<= pkts_n).
 905  */
 906 uint16_t
 907 mlx4_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
 908 {
 909         struct rxq *rxq = dpdk_rxq;
 910         const uint32_t wr_cnt = (1 << rxq->elts_n) - 1;
 911         const uint16_t sges_n = rxq->sges_n;
 912         struct rte_mbuf *pkt = NULL;
 913         struct rte_mbuf *seg = NULL;
 914         unsigned int i = 0;
 915         uint32_t rq_ci = rxq->rq_ci << sges_n;
 916         int len = 0;
 917
 918         while (pkts_n) {
 919                 volatile struct mlx4_cqe *cqe;
 920                 uint32_t idx = rq_ci & wr_cnt;
 921                 struct rte_mbuf *rep = (*rxq->elts)[idx];
 922                 volatile struct mlx4_wqe_data_seg *scat = &(*rxq->wqes)[idx];
 923
 924                 /* Update the 'next' pointer of the previous segment. */
 925                 if (pkt)
 926                         seg->next = rep;
 927                 seg = rep;
 928                 rte_prefetch0(seg);
 929                 rte_prefetch0(scat);
 930                 rep = rte_mbuf_raw_alloc(rxq->mp);
 931                 if (unlikely(rep == NULL)) {
 932                         ++rxq->stats.rx_nombuf;
 933                         if (!pkt) {
 934                                 /*
 935                                  * No buffers before we even started,
 936                                  * bail out silently.
 937                                  */
 938                                 break;
 939                         }
 940                         while (pkt != seg) {
 941                                 assert(pkt != (*rxq->elts)[idx]);
 942                                 rep = pkt->next;
 943                                 pkt->next = NULL;
 944                                 pkt->nb_segs = 1;
 945                                 rte_mbuf_raw_free(pkt);
 946                                 pkt = rep;
 947                         }
 948                         break;
 949                 }
 950                 if (!pkt) {
 951                         /* Looking for the new packet. */
 952                         len = mlx4_cq_poll_one(rxq, &cqe);
 953                         if (!len) {
 954                                 rte_mbuf_raw_free(rep);
 955                                 break;
 956                         }
 957                         if (unlikely(len < 0)) {
 958                                 /* Rx error, packet is likely too large. */
 959                                 rte_mbuf_raw_free(rep);
 960                                 ++rxq->stats.idropped;
 961                                 goto skip;
 962                         }
 963                         pkt = seg;
 964                         /* Update packet information. */
 965                         pkt->packet_type =
 966                                 rxq_cq_to_pkt_type(cqe, rxq->l2tun_offload);
 967                         pkt->ol_flags = PKT_RX_RSS_HASH;
 968                         pkt->hash.rss = cqe->immed_rss_invalid;
 969                         pkt->pkt_len = len;
 970                         if (rxq->csum | rxq->csum_l2tun) {
 971                                 uint32_t flags =
 972                                         mlx4_cqe_flags(cqe,
 973                                                        rxq->csum,
 974                                                        rxq->csum_l2tun);
 975
 976                                 pkt->ol_flags =
 977                                         rxq_cq_to_ol_flags(flags,
 978                                                            rxq->csum,
 979                                                            rxq->csum_l2tun);
 980                         }
 981                 }
 982                 rep->nb_segs = 1;
 983                 rep->port = rxq->port_id;
 984                 rep->data_len = seg->data_len;
 985                 rep->data_off = seg->data_off;
 986                 (*rxq->elts)[idx] = rep;
 987                 /*
 988                  * Fill NIC descriptor with the new buffer. The lkey and size
 989                  * of the buffers are already known, only the buffer address
 990                  * changes.
 991                  */
 992                 scat->addr = rte_cpu_to_be_64(rte_pktmbuf_mtod(rep, uintptr_t));
 993                 if (len > seg->data_len) {
 994                         len -= seg->data_len;
 995                         ++pkt->nb_segs;
 996                         ++rq_ci;
 997                         continue;
 998                 }
 999                 /* The last segment. */
1000                 seg->data_len = len;
1001                 /* Increment bytes counter. */
1002                 rxq->stats.ibytes += pkt->pkt_len;
1003                 /* Return packet. */
1004                 *(pkts++) = pkt;
1005                 pkt = NULL;
1006                 --pkts_n;
1007                 ++i;
1008 skip:
1009                 /* Align consumer index to the next stride. */
1010                 rq_ci >>= sges_n;
1011                 ++rq_ci;
1012                 rq_ci <<= sges_n;
1013         }
1014         if (unlikely(i == 0 && (rq_ci >> sges_n) == rxq->rq_ci))
1015                 return 0;
1016         /* Update the consumer index. */
1017         rxq->rq_ci = rq_ci >> sges_n;
1018         rte_wmb();
1019         *rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
1020         *rxq->mcq.set_ci_db =
1021                 rte_cpu_to_be_32(rxq->mcq.cons_index & MLX4_CQ_DB_CI_MASK);
1022         /* Increment packets counter. */
1023         rxq->stats.ipackets += i;
1024         return i;
1025 }
1026
1027 /**
1028  * Dummy DPDK callback for Tx.
1029  *
1030  * This function is used to temporarily replace the real callback during
1031  * unsafe control operations on the queue, or in case of error.
1032  *
1033  * @param dpdk_txq
1034  *   Generic pointer to Tx queue structure.
1035  * @param[in] pkts
1036  *   Packets to transmit.
1037  * @param pkts_n
1038  *   Number of packets in array.
1039  *
1040  * @return
1041  *   Number of packets successfully transmitted (<= pkts_n).
1042  */
1043 uint16_t
1044 mlx4_tx_burst_removed(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
1045 {
1046         (void)dpdk_txq;
1047         (void)pkts;
1048         (void)pkts_n;
1049         return 0;
1050 }
1051
1052 /**
1053  * Dummy DPDK callback for Rx.
1054  *
1055  * This function is used to temporarily replace the real callback during
1056  * unsafe control operations on the queue, or in case of error.
1057  *
1058  * @param dpdk_rxq
1059  *   Generic pointer to Rx queue structure.
1060  * @param[out] pkts
1061  *   Array to store received packets.
1062  * @param pkts_n
1063  *   Maximum number of packets in array.
1064  *
1065  * @return
1066  *   Number of packets successfully received (<= pkts_n).
1067  */
1068 uint16_t
1069 mlx4_rx_burst_removed(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
1070 {
1071         (void)dpdk_rxq;
1072         (void)pkts;
1073         (void)pkts_n;
1074         return 0;
1075 }