drivers/net/mlx5/mlx5_rx.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright 2021 6WIND S.A.
   3  * Copyright 2021 Mellanox Technologies, Ltd
   4  */
   5
   6 #include <stdint.h>
   7 #include <string.h>
   8 #include <stdlib.h>
   9
  10 #include <rte_mbuf.h>
  11 #include <rte_mempool.h>
  12 #include <rte_prefetch.h>
  13 #include <rte_common.h>
  14 #include <rte_branch_prediction.h>
  15 #include <rte_ether.h>
  16 #include <rte_cycles.h>
  17 #include <rte_flow.h>
  18
  19 #include <mlx5_prm.h>
  20 #include <mlx5_common.h>
  21 #include <mlx5_common_mr.h>
  22
  23 #include "mlx5_autoconf.h"
  24 #include "mlx5_defs.h"
  25 #include "mlx5.h"
  26 #include "mlx5_utils.h"
  27 #include "mlx5_rxtx.h"
  28 #include "mlx5_rx.h"
  29
  30
  31 static __rte_always_inline uint32_t
  32 rxq_cq_to_pkt_type(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe,
  33                    volatile struct mlx5_mini_cqe8 *mcqe);
  34
  35 static __rte_always_inline int
  36 mlx5_rx_poll_len(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe,
  37                  uint16_t cqe_cnt, volatile struct mlx5_mini_cqe8 **mcqe);
  38
  39 static __rte_always_inline uint32_t
  40 rxq_cq_to_ol_flags(volatile struct mlx5_cqe *cqe);
  41
  42 static __rte_always_inline void
  43 rxq_cq_to_mbuf(struct mlx5_rxq_data *rxq, struct rte_mbuf *pkt,
  44                volatile struct mlx5_cqe *cqe,
  45                volatile struct mlx5_mini_cqe8 *mcqe);
  46
  47 static inline void
  48 mlx5_lro_update_tcp_hdr(struct rte_tcp_hdr *__rte_restrict tcp,
  49                         volatile struct mlx5_cqe *__rte_restrict cqe,
  50                         uint32_t phcsum, uint8_t l4_type);
  51
  52 static inline void
  53 mlx5_lro_update_hdr(uint8_t *__rte_restrict padd,
  54                     volatile struct mlx5_cqe *__rte_restrict cqe,
  55                     volatile struct mlx5_mini_cqe8 *mcqe,
  56                     struct mlx5_rxq_data *rxq, uint32_t len);
  57
  58
  59 /**
  60  * Internal function to compute the number of used descriptors in an RX queue.
  61  *
  62  * @param rxq
  63  *   The Rx queue.
  64  *
  65  * @return
  66  *   The number of used Rx descriptor.
  67  */
  68 static uint32_t
  69 rx_queue_count(struct mlx5_rxq_data *rxq)
  70 {
  71         struct rxq_zip *zip = &rxq->zip;
  72         volatile struct mlx5_cqe *cqe;
  73         const unsigned int cqe_n = (1 << rxq->cqe_n);
  74         const unsigned int sges_n = (1 << rxq->sges_n);
  75         const unsigned int elts_n = (1 << rxq->elts_n);
  76         const unsigned int strd_n = (1 << rxq->strd_num_n);
  77         const unsigned int cqe_cnt = cqe_n - 1;
  78         unsigned int cq_ci, used;
  79
  80         /* if we are processing a compressed cqe */
  81         if (zip->ai) {
  82                 used = zip->cqe_cnt - zip->ai;
  83                 cq_ci = zip->cq_ci;
  84         } else {
  85                 used = 0;
  86                 cq_ci = rxq->cq_ci;
  87         }
  88         cqe = &(*rxq->cqes)[cq_ci & cqe_cnt];
  89         while (check_cqe(cqe, cqe_n, cq_ci) != MLX5_CQE_STATUS_HW_OWN) {
  90                 int8_t op_own;
  91                 unsigned int n;
  92
  93                 op_own = cqe->op_own;
  94                 if (MLX5_CQE_FORMAT(op_own) == MLX5_COMPRESSED)
  95                         n = rte_be_to_cpu_32(cqe->byte_cnt);
  96                 else
  97                         n = 1;
  98                 cq_ci += n;
  99                 used += n;
 100                 cqe = &(*rxq->cqes)[cq_ci & cqe_cnt];
 101         }
 102         used = RTE_MIN(used * sges_n, elts_n * strd_n);
 103         return used;
 104 }
 105
 106 /**
 107  * DPDK callback to check the status of a Rx descriptor.
 108  *
 109  * @param rx_queue
 110  *   The Rx queue.
 111  * @param[in] offset
 112  *   The index of the descriptor in the ring.
 113  *
 114  * @return
 115  *   The status of the Rx descriptor.
 116  */
 117 int
 118 mlx5_rx_descriptor_status(void *rx_queue, uint16_t offset)
 119 {
 120         struct mlx5_rxq_data *rxq = rx_queue;
 121
 122         if (offset >= (1 << rxq->cqe_n)) {
 123                 rte_errno = EINVAL;
 124                 return -rte_errno;
 125         }
 126         if (offset < rx_queue_count(rxq))
 127                 return RTE_ETH_RX_DESC_DONE;
 128         return RTE_ETH_RX_DESC_AVAIL;
 129 }
 130
 131 /**
 132  * DPDK callback to get the RX queue information.
 133  *
 134  * @param dev
 135  *   Pointer to the device structure.
 136  *
 137  * @param rx_queue_id
 138  *   Rx queue identificator.
 139  *
 140  * @param qinfo
 141  *   Pointer to the RX queue information structure.
 142  *
 143  * @return
 144  *   None.
 145  */
 146
 147 void
 148 mlx5_rxq_info_get(struct rte_eth_dev *dev, uint16_t rx_queue_id,
 149                   struct rte_eth_rxq_info *qinfo)
 150 {
 151         struct mlx5_rxq_ctrl *rxq_ctrl = mlx5_rxq_ctrl_get(dev, rx_queue_id);
 152         struct mlx5_rxq_data *rxq = mlx5_rxq_data_get(dev, rx_queue_id);
 153
 154         if (!rxq)
 155                 return;
 156         qinfo->mp = mlx5_rxq_mprq_enabled(rxq) ?
 157                                         rxq->mprq_mp : rxq->mp;
 158         qinfo->conf.rx_thresh.pthresh = 0;
 159         qinfo->conf.rx_thresh.hthresh = 0;
 160         qinfo->conf.rx_thresh.wthresh = 0;
 161         qinfo->conf.rx_free_thresh = rxq->rq_repl_thresh;
 162         qinfo->conf.rx_drop_en = 1;
 163         if (rxq_ctrl == NULL || rxq_ctrl->obj == NULL)
 164                 qinfo->conf.rx_deferred_start = 0;
 165         else
 166                 qinfo->conf.rx_deferred_start = 1;
 167         qinfo->conf.offloads = dev->data->dev_conf.rxmode.offloads;
 168         qinfo->scattered_rx = dev->data->scattered_rx;
 169         qinfo->nb_desc = mlx5_rxq_mprq_enabled(rxq) ?
 170                 (1 << rxq->elts_n) * (1 << rxq->strd_num_n) :
 171                 (1 << rxq->elts_n);
 172 }
 173
 174 /**
 175  * DPDK callback to get the RX packet burst mode information.
 176  *
 177  * @param dev
 178  *   Pointer to the device structure.
 179  *
 180  * @param rx_queue_id
 181  *   Rx queue identificatior.
 182  *
 183  * @param mode
 184  *   Pointer to the burts mode information.
 185  *
 186  * @return
 187  *   0 as success, -EINVAL as failure.
 188  */
 189 int
 190 mlx5_rx_burst_mode_get(struct rte_eth_dev *dev,
 191                        uint16_t rx_queue_id __rte_unused,
 192                        struct rte_eth_burst_mode *mode)
 193 {
 194         eth_rx_burst_t pkt_burst = dev->rx_pkt_burst;
 195         struct mlx5_rxq_priv *rxq = mlx5_rxq_get(dev, rx_queue_id);
 196
 197         if (!rxq) {
 198                 rte_errno = EINVAL;
 199                 return -rte_errno;
 200         }
 201         if (pkt_burst == mlx5_rx_burst) {
 202                 snprintf(mode->info, sizeof(mode->info), "%s", "Scalar");
 203         } else if (pkt_burst == mlx5_rx_burst_mprq) {
 204                 snprintf(mode->info, sizeof(mode->info), "%s", "Multi-Packet RQ");
 205         } else if (pkt_burst == mlx5_rx_burst_vec) {
 206 #if defined RTE_ARCH_X86_64
 207                 snprintf(mode->info, sizeof(mode->info), "%s", "Vector SSE");
 208 #elif defined RTE_ARCH_ARM64
 209                 snprintf(mode->info, sizeof(mode->info), "%s", "Vector Neon");
 210 #elif defined RTE_ARCH_PPC_64
 211                 snprintf(mode->info, sizeof(mode->info), "%s", "Vector AltiVec");
 212 #else
 213                 return -EINVAL;
 214 #endif
 215         } else if (pkt_burst == mlx5_rx_burst_mprq_vec) {
 216 #if defined RTE_ARCH_X86_64
 217                 snprintf(mode->info, sizeof(mode->info), "%s", "MPRQ Vector SSE");
 218 #elif defined RTE_ARCH_ARM64
 219                 snprintf(mode->info, sizeof(mode->info), "%s", "MPRQ Vector Neon");
 220 #elif defined RTE_ARCH_PPC_64
 221                 snprintf(mode->info, sizeof(mode->info), "%s", "MPRQ Vector AltiVec");
 222 #else
 223                 return -EINVAL;
 224 #endif
 225         } else {
 226                 return -EINVAL;
 227         }
 228         return 0;
 229 }
 230
 231 /**
 232  * DPDK callback to get the number of used descriptors in a RX queue.
 233  *
 234  * @param rx_queue
 235  *   The Rx queue pointer.
 236  *
 237  * @return
 238  *   The number of used rx descriptor.
 239  *   -EINVAL if the queue is invalid
 240  */
 241 uint32_t
 242 mlx5_rx_queue_count(void *rx_queue)
 243 {
 244         struct mlx5_rxq_data *rxq = rx_queue;
 245         struct rte_eth_dev *dev;
 246
 247         if (!rxq) {
 248                 rte_errno = EINVAL;
 249                 return -rte_errno;
 250         }
 251
 252         dev = &rte_eth_devices[rxq->port_id];
 253
 254         if (dev->rx_pkt_burst == NULL ||
 255             dev->rx_pkt_burst == removed_rx_burst) {
 256                 rte_errno = ENOTSUP;
 257                 return -rte_errno;
 258         }
 259
 260         return rx_queue_count(rxq);
 261 }
 262
 263 #define CLB_VAL_IDX 0
 264 #define CLB_MSK_IDX 1
 265 static int
 266 mlx5_monitor_callback(const uint64_t value,
 267                 const uint64_t opaque[RTE_POWER_MONITOR_OPAQUE_SZ])
 268 {
 269         const uint64_t m = opaque[CLB_MSK_IDX];
 270         const uint64_t v = opaque[CLB_VAL_IDX];
 271
 272         return (value & m) == v ? -1 : 0;
 273 }
 274
 275 int mlx5_get_monitor_addr(void *rx_queue, struct rte_power_monitor_cond *pmc)
 276 {
 277         struct mlx5_rxq_data *rxq = rx_queue;
 278         const unsigned int cqe_num = 1 << rxq->cqe_n;
 279         const unsigned int cqe_mask = cqe_num - 1;
 280         const uint16_t idx = rxq->cq_ci & cqe_num;
 281         volatile struct mlx5_cqe *cqe = &(*rxq->cqes)[rxq->cq_ci & cqe_mask];
 282
 283         if (unlikely(rxq->cqes == NULL)) {
 284                 rte_errno = EINVAL;
 285                 return -rte_errno;
 286         }
 287         pmc->addr = &cqe->op_own;
 288         pmc->opaque[CLB_VAL_IDX] = !!idx;
 289         pmc->opaque[CLB_MSK_IDX] = MLX5_CQE_OWNER_MASK;
 290         pmc->fn = mlx5_monitor_callback;
 291         pmc->size = sizeof(uint8_t);
 292         return 0;
 293 }
 294
 295 /**
 296  * Translate RX completion flags to packet type.
 297  *
 298  * @param[in] rxq
 299  *   Pointer to RX queue structure.
 300  * @param[in] cqe
 301  *   Pointer to CQE.
 302  *
 303  * @note: fix mlx5_dev_supported_ptypes_get() if any change here.
 304  *
 305  * @return
 306  *   Packet type for struct rte_mbuf.
 307  */
 308 static inline uint32_t
 309 rxq_cq_to_pkt_type(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe,
 310                                    volatile struct mlx5_mini_cqe8 *mcqe)
 311 {
 312         uint8_t idx;
 313         uint8_t ptype;
 314         uint8_t pinfo = (cqe->pkt_info & 0x3) << 6;
 315
 316         /* Get l3/l4 header from mini-CQE in case L3/L4 format*/
 317         if (mcqe == NULL ||
 318             rxq->mcqe_format != MLX5_CQE_RESP_FORMAT_L34H_STRIDX)
 319                 ptype = (cqe->hdr_type_etc & 0xfc00) >> 10;
 320         else
 321                 ptype = mcqe->hdr_type >> 2;
 322         /*
 323          * The index to the array should have:
 324          * bit[1:0] = l3_hdr_type
 325          * bit[4:2] = l4_hdr_type
 326          * bit[5] = ip_frag
 327          * bit[6] = tunneled
 328          * bit[7] = outer_l3_type
 329          */
 330         idx = pinfo | ptype;
 331         return mlx5_ptype_table[idx] | rxq->tunnel * !!(idx & (1 << 6));
 332 }
 333
 334 /**
 335  * Initialize Rx WQ and indexes.
 336  *
 337  * @param[in] rxq
 338  *   Pointer to RX queue structure.
 339  */
 340 void
 341 mlx5_rxq_initialize(struct mlx5_rxq_data *rxq)
 342 {
 343         const unsigned int wqe_n = 1 << rxq->elts_n;
 344         unsigned int i;
 345
 346         for (i = 0; (i != wqe_n); ++i) {
 347                 volatile struct mlx5_wqe_data_seg *scat;
 348                 uintptr_t addr;
 349                 uint32_t byte_count;
 350
 351                 if (mlx5_rxq_mprq_enabled(rxq)) {
 352                         struct mlx5_mprq_buf *buf = (*rxq->mprq_bufs)[i];
 353
 354                         scat = &((volatile struct mlx5_wqe_mprq *)
 355                                 rxq->wqes)[i].dseg;
 356                         addr = (uintptr_t)mlx5_mprq_buf_addr(buf,
 357                                                          1 << rxq->strd_num_n);
 358                         byte_count = (1 << rxq->strd_sz_n) *
 359                                         (1 << rxq->strd_num_n);
 360                 } else {
 361                         struct rte_mbuf *buf = (*rxq->elts)[i];
 362
 363                         scat = &((volatile struct mlx5_wqe_data_seg *)
 364                                         rxq->wqes)[i];
 365                         addr = rte_pktmbuf_mtod(buf, uintptr_t);
 366                         byte_count = DATA_LEN(buf);
 367                 }
 368                 /* scat->addr must be able to store a pointer. */
 369                 MLX5_ASSERT(sizeof(scat->addr) >= sizeof(uintptr_t));
 370                 *scat = (struct mlx5_wqe_data_seg){
 371                         .addr = rte_cpu_to_be_64(addr),
 372                         .byte_count = rte_cpu_to_be_32(byte_count),
 373                         .lkey = mlx5_rx_addr2mr(rxq, addr),
 374                 };
 375         }
 376         rxq->consumed_strd = 0;
 377         rxq->decompressed = 0;
 378         rxq->rq_pi = 0;
 379         rxq->zip = (struct rxq_zip){
 380                 .ai = 0,
 381         };
 382         rxq->elts_ci = mlx5_rxq_mprq_enabled(rxq) ?
 383                 (wqe_n >> rxq->sges_n) * (1 << rxq->strd_num_n) : 0;
 384         /* Update doorbell counter. */
 385         rxq->rq_ci = wqe_n >> rxq->sges_n;
 386         rte_io_wmb();
 387         *rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
 388 }
 389
 390 /**
 391  * Handle a Rx error.
 392  * The function inserts the RQ state to reset when the first error CQE is
 393  * shown, then drains the CQ by the caller function loop. When the CQ is empty,
 394  * it moves the RQ state to ready and initializes the RQ.
 395  * Next CQE identification and error counting are in the caller responsibility.
 396  *
 397  * @param[in] rxq
 398  *   Pointer to RX queue structure.
 399  * @param[in] vec
 400  *   1 when called from vectorized Rx burst, need to prepare mbufs for the RQ.
 401  *   0 when called from non-vectorized Rx burst.
 402  *
 403  * @return
 404  *   -1 in case of recovery error, otherwise the CQE status.
 405  */
 406 int
 407 mlx5_rx_err_handle(struct mlx5_rxq_data *rxq, uint8_t vec)
 408 {
 409         const uint16_t cqe_n = 1 << rxq->cqe_n;
 410         const uint16_t cqe_mask = cqe_n - 1;
 411         const uint16_t wqe_n = 1 << rxq->elts_n;
 412         const uint16_t strd_n = 1 << rxq->strd_num_n;
 413         struct mlx5_rxq_ctrl *rxq_ctrl =
 414                         container_of(rxq, struct mlx5_rxq_ctrl, rxq);
 415         union {
 416                 volatile struct mlx5_cqe *cqe;
 417                 volatile struct mlx5_err_cqe *err_cqe;
 418         } u = {
 419                 .cqe = &(*rxq->cqes)[rxq->cq_ci & cqe_mask],
 420         };
 421         struct mlx5_mp_arg_queue_state_modify sm;
 422         int ret;
 423
 424         switch (rxq->err_state) {
 425         case MLX5_RXQ_ERR_STATE_NO_ERROR:
 426                 rxq->err_state = MLX5_RXQ_ERR_STATE_NEED_RESET;
 427                 /* Fall-through */
 428         case MLX5_RXQ_ERR_STATE_NEED_RESET:
 429                 sm.is_wq = 1;
 430                 sm.queue_id = rxq->idx;
 431                 sm.state = IBV_WQS_RESET;
 432                 if (mlx5_queue_state_modify(RXQ_DEV(rxq_ctrl), &sm))
 433                         return -1;
 434                 if (rxq_ctrl->dump_file_n <
 435                     RXQ_PORT(rxq_ctrl)->config.max_dump_files_num) {
 436                         MKSTR(err_str, "Unexpected CQE error syndrome "
 437                               "0x%02x CQN = %u RQN = %u wqe_counter = %u"
 438                               " rq_ci = %u cq_ci = %u", u.err_cqe->syndrome,
 439                               rxq->cqn, rxq_ctrl->wqn,
 440                               rte_be_to_cpu_16(u.err_cqe->wqe_counter),
 441                               rxq->rq_ci << rxq->sges_n, rxq->cq_ci);
 442                         MKSTR(name, "dpdk_mlx5_port_%u_rxq_%u_%u",
 443                               rxq->port_id, rxq->idx, (uint32_t)rte_rdtsc());
 444                         mlx5_dump_debug_information(name, NULL, err_str, 0);
 445                         mlx5_dump_debug_information(name, "MLX5 Error CQ:",
 446                                                     (const void *)((uintptr_t)
 447                                                                     rxq->cqes),
 448                                                     sizeof(*u.cqe) * cqe_n);
 449                         mlx5_dump_debug_information(name, "MLX5 Error RQ:",
 450                                                     (const void *)((uintptr_t)
 451                                                                     rxq->wqes),
 452                                                     16 * wqe_n);
 453                         rxq_ctrl->dump_file_n++;
 454                 }
 455                 rxq->err_state = MLX5_RXQ_ERR_STATE_NEED_READY;
 456                 /* Fall-through */
 457         case MLX5_RXQ_ERR_STATE_NEED_READY:
 458                 ret = check_cqe(u.cqe, cqe_n, rxq->cq_ci);
 459                 if (ret == MLX5_CQE_STATUS_HW_OWN) {
 460                         rte_io_wmb();
 461                         *rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
 462                         rte_io_wmb();
 463                         /*
 464                          * The RQ consumer index must be zeroed while moving
 465                          * from RESET state to RDY state.
 466                          */
 467                         *rxq->rq_db = rte_cpu_to_be_32(0);
 468                         rte_io_wmb();
 469                         sm.is_wq = 1;
 470                         sm.queue_id = rxq->idx;
 471                         sm.state = IBV_WQS_RDY;
 472                         if (mlx5_queue_state_modify(RXQ_DEV(rxq_ctrl), &sm))
 473                                 return -1;
 474                         if (vec) {
 475                                 const uint32_t elts_n =
 476                                         mlx5_rxq_mprq_enabled(rxq) ?
 477                                         wqe_n * strd_n : wqe_n;
 478                                 const uint32_t e_mask = elts_n - 1;
 479                                 uint32_t elts_ci =
 480                                         mlx5_rxq_mprq_enabled(rxq) ?
 481                                         rxq->elts_ci : rxq->rq_ci;
 482                                 uint32_t elt_idx;
 483                                 struct rte_mbuf **elt;
 484                                 int i;
 485                                 unsigned int n = elts_n - (elts_ci -
 486                                                           rxq->rq_pi);
 487
 488                                 for (i = 0; i < (int)n; ++i) {
 489                                         elt_idx = (elts_ci + i) & e_mask;
 490                                         elt = &(*rxq->elts)[elt_idx];
 491                                         *elt = rte_mbuf_raw_alloc(rxq->mp);
 492                                         if (!*elt) {
 493                                                 for (i--; i >= 0; --i) {
 494                                                         elt_idx = (elts_ci +
 495                                                                    i) & elts_n;
 496                                                         elt = &(*rxq->elts)
 497                                                                 [elt_idx];
 498                                                         rte_pktmbuf_free_seg
 499                                                                 (*elt);
 500                                                 }
 501                                                 return -1;
 502                                         }
 503                                 }
 504                                 for (i = 0; i < (int)elts_n; ++i) {
 505                                         elt = &(*rxq->elts)[i];
 506                                         DATA_LEN(*elt) =
 507                                                 (uint16_t)((*elt)->buf_len -
 508                                                 rte_pktmbuf_headroom(*elt));
 509                                 }
 510                                 /* Padding with a fake mbuf for vec Rx. */
 511                                 for (i = 0; i < MLX5_VPMD_DESCS_PER_LOOP; ++i)
 512                                         (*rxq->elts)[elts_n + i] =
 513                                                                 &rxq->fake_mbuf;
 514                         }
 515                         mlx5_rxq_initialize(rxq);
 516                         rxq->err_state = MLX5_RXQ_ERR_STATE_NO_ERROR;
 517                 }
 518                 return ret;
 519         default:
 520                 return -1;
 521         }
 522 }
 523
 524 /**
 525  * Get size of the next packet for a given CQE. For compressed CQEs, the
 526  * consumer index is updated only once all packets of the current one have
 527  * been processed.
 528  *
 529  * @param rxq
 530  *   Pointer to RX queue.
 531  * @param cqe
 532  *   CQE to process.
 533  * @param[out] mcqe
 534  *   Store pointer to mini-CQE if compressed. Otherwise, the pointer is not
 535  *   written.
 536  *
 537  * @return
 538  *   0 in case of empty CQE, otherwise the packet size in bytes.
 539  */
 540 static inline int
 541 mlx5_rx_poll_len(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe,
 542                  uint16_t cqe_cnt, volatile struct mlx5_mini_cqe8 **mcqe)
 543 {
 544         struct rxq_zip *zip = &rxq->zip;
 545         uint16_t cqe_n = cqe_cnt + 1;
 546         int len;
 547         uint16_t idx, end;
 548
 549         do {
 550                 len = 0;
 551                 /* Process compressed data in the CQE and mini arrays. */
 552                 if (zip->ai) {
 553                         volatile struct mlx5_mini_cqe8 (*mc)[8] =
 554                                 (volatile struct mlx5_mini_cqe8 (*)[8])
 555                                 (uintptr_t)(&(*rxq->cqes)[zip->ca &
 556                                                           cqe_cnt].pkt_info);
 557                         len = rte_be_to_cpu_32((*mc)[zip->ai & 7].byte_cnt &
 558                                                rxq->byte_mask);
 559                         *mcqe = &(*mc)[zip->ai & 7];
 560                         if ((++zip->ai & 7) == 0) {
 561                                 /* Invalidate consumed CQEs */
 562                                 idx = zip->ca;
 563                                 end = zip->na;
 564                                 while (idx != end) {
 565                                         (*rxq->cqes)[idx & cqe_cnt].op_own =
 566                                                 MLX5_CQE_INVALIDATE;
 567                                         ++idx;
 568                                 }
 569                                 /*
 570                                  * Increment consumer index to skip the number
 571                                  * of CQEs consumed. Hardware leaves holes in
 572                                  * the CQ ring for software use.
 573                                  */
 574                                 zip->ca = zip->na;
 575                                 zip->na += 8;
 576                         }
 577                         if (unlikely(rxq->zip.ai == rxq->zip.cqe_cnt)) {
 578                                 /* Invalidate the rest */
 579                                 idx = zip->ca;
 580                                 end = zip->cq_ci;
 581
 582                                 while (idx != end) {
 583                                         (*rxq->cqes)[idx & cqe_cnt].op_own =
 584                                                 MLX5_CQE_INVALIDATE;
 585                                         ++idx;
 586                                 }
 587                                 rxq->cq_ci = zip->cq_ci;
 588                                 zip->ai = 0;
 589                         }
 590                 /*
 591                  * No compressed data, get next CQE and verify if it is
 592                  * compressed.
 593                  */
 594                 } else {
 595                         int ret;
 596                         int8_t op_own;
 597                         uint32_t cq_ci;
 598
 599                         ret = check_cqe(cqe, cqe_n, rxq->cq_ci);
 600                         if (unlikely(ret != MLX5_CQE_STATUS_SW_OWN)) {
 601                                 if (unlikely(ret == MLX5_CQE_STATUS_ERR ||
 602                                              rxq->err_state)) {
 603                                         ret = mlx5_rx_err_handle(rxq, 0);
 604                                         if (ret == MLX5_CQE_STATUS_HW_OWN ||
 605                                             ret == -1)
 606                                                 return 0;
 607                                 } else {
 608                                         return 0;
 609                                 }
 610                         }
 611                         /*
 612                          * Introduce the local variable to have queue cq_ci
 613                          * index in queue structure always consistent with
 614                          * actual CQE boundary (not pointing to the middle
 615                          * of compressed CQE session).
 616                          */
 617                         cq_ci = rxq->cq_ci + 1;
 618                         op_own = cqe->op_own;
 619                         if (MLX5_CQE_FORMAT(op_own) == MLX5_COMPRESSED) {
 620                                 volatile struct mlx5_mini_cqe8 (*mc)[8] =
 621                                         (volatile struct mlx5_mini_cqe8 (*)[8])
 622                                         (uintptr_t)(&(*rxq->cqes)
 623                                                 [cq_ci & cqe_cnt].pkt_info);
 624
 625                                 /* Fix endianness. */
 626                                 zip->cqe_cnt = rte_be_to_cpu_32(cqe->byte_cnt);
 627                                 /*
 628                                  * Current mini array position is the one
 629                                  * returned by check_cqe64().
 630                                  *
 631                                  * If completion comprises several mini arrays,
 632                                  * as a special case the second one is located
 633                                  * 7 CQEs after the initial CQE instead of 8
 634                                  * for subsequent ones.
 635                                  */
 636                                 zip->ca = cq_ci;
 637                                 zip->na = zip->ca + 7;
 638                                 /* Compute the next non compressed CQE. */
 639                                 zip->cq_ci = rxq->cq_ci + zip->cqe_cnt;
 640                                 /* Get packet size to return. */
 641                                 len = rte_be_to_cpu_32((*mc)[0].byte_cnt &
 642                                                        rxq->byte_mask);
 643                                 *mcqe = &(*mc)[0];
 644                                 zip->ai = 1;
 645                                 /* Prefetch all to be invalidated */
 646                                 idx = zip->ca;
 647                                 end = zip->cq_ci;
 648                                 while (idx != end) {
 649                                         rte_prefetch0(&(*rxq->cqes)[(idx) &
 650                                                                     cqe_cnt]);
 651                                         ++idx;
 652                                 }
 653                         } else {
 654                                 rxq->cq_ci = cq_ci;
 655                                 len = rte_be_to_cpu_32(cqe->byte_cnt);
 656                         }
 657                 }
 658                 if (unlikely(rxq->err_state)) {
 659                         cqe = &(*rxq->cqes)[rxq->cq_ci & cqe_cnt];
 660                         ++rxq->stats.idropped;
 661                 } else {
 662                         return len;
 663                 }
 664         } while (1);
 665 }
 666
 667 /**
 668  * Translate RX completion flags to offload flags.
 669  *
 670  * @param[in] cqe
 671  *   Pointer to CQE.
 672  *
 673  * @return
 674  *   Offload flags (ol_flags) for struct rte_mbuf.
 675  */
 676 static inline uint32_t
 677 rxq_cq_to_ol_flags(volatile struct mlx5_cqe *cqe)
 678 {
 679         uint32_t ol_flags = 0;
 680         uint16_t flags = rte_be_to_cpu_16(cqe->hdr_type_etc);
 681
 682         ol_flags =
 683                 TRANSPOSE(flags,
 684                           MLX5_CQE_RX_L3_HDR_VALID,
 685                           RTE_MBUF_F_RX_IP_CKSUM_GOOD) |
 686                 TRANSPOSE(flags,
 687                           MLX5_CQE_RX_L4_HDR_VALID,
 688                           RTE_MBUF_F_RX_L4_CKSUM_GOOD);
 689         return ol_flags;
 690 }
 691
 692 /**
 693  * Fill in mbuf fields from RX completion flags.
 694  * Note that pkt->ol_flags should be initialized outside of this function.
 695  *
 696  * @param rxq
 697  *   Pointer to RX queue.
 698  * @param pkt
 699  *   mbuf to fill.
 700  * @param cqe
 701  *   CQE to process.
 702  * @param rss_hash_res
 703  *   Packet RSS Hash result.
 704  */
 705 static inline void
 706 rxq_cq_to_mbuf(struct mlx5_rxq_data *rxq, struct rte_mbuf *pkt,
 707                volatile struct mlx5_cqe *cqe,
 708                volatile struct mlx5_mini_cqe8 *mcqe)
 709 {
 710         /* Update packet information. */
 711         pkt->packet_type = rxq_cq_to_pkt_type(rxq, cqe, mcqe);
 712         pkt->port = unlikely(rxq->shared) ? cqe->user_index_low : rxq->port_id;
 713
 714         if (rxq->rss_hash) {
 715                 uint32_t rss_hash_res = 0;
 716
 717                 /* If compressed, take hash result from mini-CQE. */
 718                 if (mcqe == NULL ||
 719                     rxq->mcqe_format != MLX5_CQE_RESP_FORMAT_HASH)
 720                         rss_hash_res = rte_be_to_cpu_32(cqe->rx_hash_res);
 721                 else
 722                         rss_hash_res = rte_be_to_cpu_32(mcqe->rx_hash_result);
 723                 if (rss_hash_res) {
 724                         pkt->hash.rss = rss_hash_res;
 725                         pkt->ol_flags |= RTE_MBUF_F_RX_RSS_HASH;
 726                 }
 727         }
 728         if (rxq->mark) {
 729                 uint32_t mark = 0;
 730
 731                 /* If compressed, take flow tag from mini-CQE. */
 732                 if (mcqe == NULL ||
 733                     rxq->mcqe_format != MLX5_CQE_RESP_FORMAT_FTAG_STRIDX)
 734                         mark = cqe->sop_drop_qpn;
 735                 else
 736                         mark = ((mcqe->byte_cnt_flow & 0xff) << 8) |
 737                                 (mcqe->flow_tag_high << 16);
 738                 if (MLX5_FLOW_MARK_IS_VALID(mark)) {
 739                         pkt->ol_flags |= RTE_MBUF_F_RX_FDIR;
 740                         if (mark != RTE_BE32(MLX5_FLOW_MARK_DEFAULT)) {
 741                                 pkt->ol_flags |= RTE_MBUF_F_RX_FDIR_ID;
 742                                 pkt->hash.fdir.hi = mlx5_flow_mark_get(mark);
 743                         }
 744                 }
 745         }
 746         if (rxq->dynf_meta) {
 747                 uint32_t meta = rte_be_to_cpu_32(cqe->flow_table_metadata) &
 748                         rxq->flow_meta_port_mask;
 749
 750                 if (meta) {
 751                         pkt->ol_flags |= rxq->flow_meta_mask;
 752                         *RTE_MBUF_DYNFIELD(pkt, rxq->flow_meta_offset,
 753                                                 uint32_t *) = meta;
 754                 }
 755         }
 756         if (rxq->csum)
 757                 pkt->ol_flags |= rxq_cq_to_ol_flags(cqe);
 758         if (rxq->vlan_strip) {
 759                 bool vlan_strip;
 760
 761                 if (mcqe == NULL ||
 762                     rxq->mcqe_format != MLX5_CQE_RESP_FORMAT_L34H_STRIDX)
 763                         vlan_strip = cqe->hdr_type_etc &
 764                                      RTE_BE16(MLX5_CQE_VLAN_STRIPPED);
 765                 else
 766                         vlan_strip = mcqe->hdr_type &
 767                                      RTE_BE16(MLX5_CQE_VLAN_STRIPPED);
 768                 if (vlan_strip) {
 769                         pkt->ol_flags |= RTE_MBUF_F_RX_VLAN | RTE_MBUF_F_RX_VLAN_STRIPPED;
 770                         pkt->vlan_tci = rte_be_to_cpu_16(cqe->vlan_info);
 771                 }
 772         }
 773         if (rxq->hw_timestamp) {
 774                 uint64_t ts = rte_be_to_cpu_64(cqe->timestamp);
 775
 776                 if (rxq->rt_timestamp)
 777                         ts = mlx5_txpp_convert_rx_ts(rxq->sh, ts);
 778                 mlx5_timestamp_set(pkt, rxq->timestamp_offset, ts);
 779                 pkt->ol_flags |= rxq->timestamp_rx_flag;
 780         }
 781 }
 782
 783 /**
 784  * DPDK callback for RX.
 785  *
 786  * @param dpdk_rxq
 787  *   Generic pointer to RX queue structure.
 788  * @param[out] pkts
 789  *   Array to store received packets.
 790  * @param pkts_n
 791  *   Maximum number of packets in array.
 792  *
 793  * @return
 794  *   Number of packets successfully received (<= pkts_n).
 795  */
 796 uint16_t
 797 mlx5_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
 798 {
 799         struct mlx5_rxq_data *rxq = dpdk_rxq;
 800         const unsigned int wqe_cnt = (1 << rxq->elts_n) - 1;
 801         const unsigned int cqe_cnt = (1 << rxq->cqe_n) - 1;
 802         const unsigned int sges_n = rxq->sges_n;
 803         struct rte_mbuf *pkt = NULL;
 804         struct rte_mbuf *seg = NULL;
 805         volatile struct mlx5_cqe *cqe =
 806                 &(*rxq->cqes)[rxq->cq_ci & cqe_cnt];
 807         unsigned int i = 0;
 808         unsigned int rq_ci = rxq->rq_ci << sges_n;
 809         int len = 0; /* keep its value across iterations. */
 810
 811         while (pkts_n) {
 812                 unsigned int idx = rq_ci & wqe_cnt;
 813                 volatile struct mlx5_wqe_data_seg *wqe =
 814                         &((volatile struct mlx5_wqe_data_seg *)rxq->wqes)[idx];
 815                 struct rte_mbuf *rep = (*rxq->elts)[idx];
 816                 volatile struct mlx5_mini_cqe8 *mcqe = NULL;
 817
 818                 if (pkt)
 819                         NEXT(seg) = rep;
 820                 seg = rep;
 821                 rte_prefetch0(seg);
 822                 rte_prefetch0(cqe);
 823                 rte_prefetch0(wqe);
 824                 /* Allocate the buf from the same pool. */
 825                 rep = rte_mbuf_raw_alloc(seg->pool);
 826                 if (unlikely(rep == NULL)) {
 827                         ++rxq->stats.rx_nombuf;
 828                         if (!pkt) {
 829                                 /*
 830                                  * no buffers before we even started,
 831                                  * bail out silently.
 832                                  */
 833                                 break;
 834                         }
 835                         while (pkt != seg) {
 836                                 MLX5_ASSERT(pkt != (*rxq->elts)[idx]);
 837                                 rep = NEXT(pkt);
 838                                 NEXT(pkt) = NULL;
 839                                 NB_SEGS(pkt) = 1;
 840                                 rte_mbuf_raw_free(pkt);
 841                                 pkt = rep;
 842                         }
 843                         rq_ci >>= sges_n;
 844                         ++rq_ci;
 845                         rq_ci <<= sges_n;
 846                         break;
 847                 }
 848                 if (!pkt) {
 849                         cqe = &(*rxq->cqes)[rxq->cq_ci & cqe_cnt];
 850                         len = mlx5_rx_poll_len(rxq, cqe, cqe_cnt, &mcqe);
 851                         if (!len) {
 852                                 rte_mbuf_raw_free(rep);
 853                                 break;
 854                         }
 855                         pkt = seg;
 856                         MLX5_ASSERT(len >= (rxq->crc_present << 2));
 857                         pkt->ol_flags &= RTE_MBUF_F_EXTERNAL;
 858                         rxq_cq_to_mbuf(rxq, pkt, cqe, mcqe);
 859                         if (rxq->crc_present)
 860                                 len -= RTE_ETHER_CRC_LEN;
 861                         PKT_LEN(pkt) = len;
 862                         if (cqe->lro_num_seg > 1) {
 863                                 mlx5_lro_update_hdr
 864                                         (rte_pktmbuf_mtod(pkt, uint8_t *), cqe,
 865                                          mcqe, rxq, len);
 866                                 pkt->ol_flags |= RTE_MBUF_F_RX_LRO;
 867                                 pkt->tso_segsz = len / cqe->lro_num_seg;
 868                         }
 869                 }
 870                 DATA_LEN(rep) = DATA_LEN(seg);
 871                 PKT_LEN(rep) = PKT_LEN(seg);
 872                 SET_DATA_OFF(rep, DATA_OFF(seg));
 873                 PORT(rep) = PORT(seg);
 874                 (*rxq->elts)[idx] = rep;
 875                 /*
 876                  * Fill NIC descriptor with the new buffer. The lkey and size
 877                  * of the buffers are already known, only the buffer address
 878                  * changes.
 879                  */
 880                 wqe->addr = rte_cpu_to_be_64(rte_pktmbuf_mtod(rep, uintptr_t));
 881                 /* If there's only one MR, no need to replace LKey in WQE. */
 882                 if (unlikely(mlx5_mr_btree_len(&rxq->mr_ctrl.cache_bh) > 1))
 883                         wqe->lkey = mlx5_rx_mb2mr(rxq, rep);
 884                 if (len > DATA_LEN(seg)) {
 885                         len -= DATA_LEN(seg);
 886                         ++NB_SEGS(pkt);
 887                         ++rq_ci;
 888                         continue;
 889                 }
 890                 DATA_LEN(seg) = len;
 891 #ifdef MLX5_PMD_SOFT_COUNTERS
 892                 /* Increment bytes counter. */
 893                 rxq->stats.ibytes += PKT_LEN(pkt);
 894 #endif
 895                 /* Return packet. */
 896                 *(pkts++) = pkt;
 897                 pkt = NULL;
 898                 --pkts_n;
 899                 ++i;
 900                 /* Align consumer index to the next stride. */
 901                 rq_ci >>= sges_n;
 902                 ++rq_ci;
 903                 rq_ci <<= sges_n;
 904         }
 905         if (unlikely(i == 0 && ((rq_ci >> sges_n) == rxq->rq_ci)))
 906                 return 0;
 907         /* Update the consumer index. */
 908         rxq->rq_ci = rq_ci >> sges_n;
 909         rte_io_wmb();
 910         *rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
 911         rte_io_wmb();
 912         *rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
 913 #ifdef MLX5_PMD_SOFT_COUNTERS
 914         /* Increment packets counter. */
 915         rxq->stats.ipackets += i;
 916 #endif
 917         return i;
 918 }
 919
 920 /**
 921  * Update LRO packet TCP header.
 922  * The HW LRO feature doesn't update the TCP header after coalescing the
 923  * TCP segments but supplies information in CQE to fill it by SW.
 924  *
 925  * @param tcp
 926  *   Pointer to the TCP header.
 927  * @param cqe
 928  *   Pointer to the completion entry.
 929  * @param phcsum
 930  *   The L3 pseudo-header checksum.
 931  */
 932 static inline void
 933 mlx5_lro_update_tcp_hdr(struct rte_tcp_hdr *__rte_restrict tcp,
 934                         volatile struct mlx5_cqe *__rte_restrict cqe,
 935                         uint32_t phcsum, uint8_t l4_type)
 936 {
 937         /*
 938          * The HW calculates only the TCP payload checksum, need to complete
 939          * the TCP header checksum and the L3 pseudo-header checksum.
 940          */
 941         uint32_t csum = phcsum + cqe->csum;
 942
 943         if (l4_type == MLX5_L4_HDR_TYPE_TCP_EMPTY_ACK ||
 944             l4_type == MLX5_L4_HDR_TYPE_TCP_WITH_ACL) {
 945                 tcp->tcp_flags |= RTE_TCP_ACK_FLAG;
 946                 tcp->recv_ack = cqe->lro_ack_seq_num;
 947                 tcp->rx_win = cqe->lro_tcp_win;
 948         }
 949         if (cqe->lro_tcppsh_abort_dupack & MLX5_CQE_LRO_PUSH_MASK)
 950                 tcp->tcp_flags |= RTE_TCP_PSH_FLAG;
 951         tcp->cksum = 0;
 952         csum += rte_raw_cksum(tcp, (tcp->data_off >> 4) * 4);
 953         csum = ((csum & 0xffff0000) >> 16) + (csum & 0xffff);
 954         csum = (~csum) & 0xffff;
 955         if (csum == 0)
 956                 csum = 0xffff;
 957         tcp->cksum = csum;
 958 }
 959
 960 /**
 961  * Update LRO packet headers.
 962  * The HW LRO feature doesn't update the L3/TCP headers after coalescing the
 963  * TCP segments but supply information in CQE to fill it by SW.
 964  *
 965  * @param padd
 966  *   The packet address.
 967  * @param cqe
 968  *   Pointer to the completion entry.
 969  * @param len
 970  *   The packet length.
 971  */
 972 static inline void
 973 mlx5_lro_update_hdr(uint8_t *__rte_restrict padd,
 974                     volatile struct mlx5_cqe *__rte_restrict cqe,
 975                     volatile struct mlx5_mini_cqe8 *mcqe,
 976                     struct mlx5_rxq_data *rxq, uint32_t len)
 977 {
 978         union {
 979                 struct rte_ether_hdr *eth;
 980                 struct rte_vlan_hdr *vlan;
 981                 struct rte_ipv4_hdr *ipv4;
 982                 struct rte_ipv6_hdr *ipv6;
 983                 struct rte_tcp_hdr *tcp;
 984                 uint8_t *hdr;
 985         } h = {
 986                 .hdr = padd,
 987         };
 988         uint16_t proto = h.eth->ether_type;
 989         uint32_t phcsum;
 990         uint8_t l4_type;
 991
 992         h.eth++;
 993         while (proto == RTE_BE16(RTE_ETHER_TYPE_VLAN) ||
 994                proto == RTE_BE16(RTE_ETHER_TYPE_QINQ)) {
 995                 proto = h.vlan->eth_proto;
 996                 h.vlan++;
 997         }
 998         if (proto == RTE_BE16(RTE_ETHER_TYPE_IPV4)) {
 999                 h.ipv4->time_to_live = cqe->lro_min_ttl;
1000                 h.ipv4->total_length = rte_cpu_to_be_16(len - (h.hdr - padd));
1001                 h.ipv4->hdr_checksum = 0;
1002                 h.ipv4->hdr_checksum = rte_ipv4_cksum(h.ipv4);
1003                 phcsum = rte_ipv4_phdr_cksum(h.ipv4, 0);
1004                 h.ipv4++;
1005         } else {
1006                 h.ipv6->hop_limits = cqe->lro_min_ttl;
1007                 h.ipv6->payload_len = rte_cpu_to_be_16(len - (h.hdr - padd) -
1008                                                        sizeof(*h.ipv6));
1009                 phcsum = rte_ipv6_phdr_cksum(h.ipv6, 0);
1010                 h.ipv6++;
1011         }
1012         if (mcqe == NULL ||
1013             rxq->mcqe_format != MLX5_CQE_RESP_FORMAT_L34H_STRIDX)
1014                 l4_type = (rte_be_to_cpu_16(cqe->hdr_type_etc) &
1015                            MLX5_CQE_L4_TYPE_MASK) >> MLX5_CQE_L4_TYPE_SHIFT;
1016         else
1017                 l4_type = (rte_be_to_cpu_16(mcqe->hdr_type) &
1018                            MLX5_CQE_L4_TYPE_MASK) >> MLX5_CQE_L4_TYPE_SHIFT;
1019         mlx5_lro_update_tcp_hdr(h.tcp, cqe, phcsum, l4_type);
1020 }
1021
1022 void
1023 mlx5_mprq_buf_free(struct mlx5_mprq_buf *buf)
1024 {
1025         mlx5_mprq_buf_free_cb(NULL, buf);
1026 }
1027
1028 /**
1029  * DPDK callback for RX with Multi-Packet RQ support.
1030  *
1031  * @param dpdk_rxq
1032  *   Generic pointer to RX queue structure.
1033  * @param[out] pkts
1034  *   Array to store received packets.
1035  * @param pkts_n
1036  *   Maximum number of packets in array.
1037  *
1038  * @return
1039  *   Number of packets successfully received (<= pkts_n).
1040  */
1041 uint16_t
1042 mlx5_rx_burst_mprq(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
1043 {
1044         struct mlx5_rxq_data *rxq = dpdk_rxq;
1045         const uint32_t strd_n = 1 << rxq->strd_num_n;
1046         const uint32_t strd_sz = 1 << rxq->strd_sz_n;
1047         const uint32_t cq_mask = (1 << rxq->cqe_n) - 1;
1048         const uint32_t wq_mask = (1 << rxq->elts_n) - 1;
1049         volatile struct mlx5_cqe *cqe = &(*rxq->cqes)[rxq->cq_ci & cq_mask];
1050         unsigned int i = 0;
1051         uint32_t rq_ci = rxq->rq_ci;
1052         uint16_t consumed_strd = rxq->consumed_strd;
1053         struct mlx5_mprq_buf *buf = (*rxq->mprq_bufs)[rq_ci & wq_mask];
1054
1055         while (i < pkts_n) {
1056                 struct rte_mbuf *pkt;
1057                 int ret;
1058                 uint32_t len;
1059                 uint16_t strd_cnt;
1060                 uint16_t strd_idx;
1061                 uint32_t byte_cnt;
1062                 volatile struct mlx5_mini_cqe8 *mcqe = NULL;
1063                 enum mlx5_rqx_code rxq_code;
1064
1065                 if (consumed_strd == strd_n) {
1066                         /* Replace WQE if the buffer is still in use. */
1067                         mprq_buf_replace(rxq, rq_ci & wq_mask);
1068                         /* Advance to the next WQE. */
1069                         consumed_strd = 0;
1070                         ++rq_ci;
1071                         buf = (*rxq->mprq_bufs)[rq_ci & wq_mask];
1072                 }
1073                 cqe = &(*rxq->cqes)[rxq->cq_ci & cq_mask];
1074                 ret = mlx5_rx_poll_len(rxq, cqe, cq_mask, &mcqe);
1075                 if (!ret)
1076                         break;
1077                 byte_cnt = ret;
1078                 len = (byte_cnt & MLX5_MPRQ_LEN_MASK) >> MLX5_MPRQ_LEN_SHIFT;
1079                 MLX5_ASSERT((int)len >= (rxq->crc_present << 2));
1080                 if (rxq->crc_present)
1081                         len -= RTE_ETHER_CRC_LEN;
1082                 if (mcqe &&
1083                     rxq->mcqe_format == MLX5_CQE_RESP_FORMAT_FTAG_STRIDX)
1084                         strd_cnt = (len / strd_sz) + !!(len % strd_sz);
1085                 else
1086                         strd_cnt = (byte_cnt & MLX5_MPRQ_STRIDE_NUM_MASK) >>
1087                                    MLX5_MPRQ_STRIDE_NUM_SHIFT;
1088                 MLX5_ASSERT(strd_cnt);
1089                 consumed_strd += strd_cnt;
1090                 if (byte_cnt & MLX5_MPRQ_FILLER_MASK)
1091                         continue;
1092                 strd_idx = rte_be_to_cpu_16(mcqe == NULL ?
1093                                         cqe->wqe_counter :
1094                                         mcqe->stride_idx);
1095                 MLX5_ASSERT(strd_idx < strd_n);
1096                 MLX5_ASSERT(!((rte_be_to_cpu_16(cqe->wqe_id) ^ rq_ci) &
1097                             wq_mask));
1098                 pkt = rte_pktmbuf_alloc(rxq->mp);
1099                 if (unlikely(pkt == NULL)) {
1100                         ++rxq->stats.rx_nombuf;
1101                         break;
1102                 }
1103                 len = (byte_cnt & MLX5_MPRQ_LEN_MASK) >> MLX5_MPRQ_LEN_SHIFT;
1104                 MLX5_ASSERT((int)len >= (rxq->crc_present << 2));
1105                 if (rxq->crc_present)
1106                         len -= RTE_ETHER_CRC_LEN;
1107                 rxq_code = mprq_buf_to_pkt(rxq, pkt, len, buf,
1108                                            strd_idx, strd_cnt);
1109                 if (unlikely(rxq_code != MLX5_RXQ_CODE_EXIT)) {
1110                         rte_pktmbuf_free_seg(pkt);
1111                         if (rxq_code == MLX5_RXQ_CODE_DROPPED) {
1112                                 ++rxq->stats.idropped;
1113                                 continue;
1114                         }
1115                         if (rxq_code == MLX5_RXQ_CODE_NOMBUF) {
1116                                 ++rxq->stats.rx_nombuf;
1117                                 break;
1118                         }
1119                 }
1120                 rxq_cq_to_mbuf(rxq, pkt, cqe, mcqe);
1121                 if (cqe->lro_num_seg > 1) {
1122                         mlx5_lro_update_hdr(rte_pktmbuf_mtod(pkt, uint8_t *),
1123                                             cqe, mcqe, rxq, len);
1124                         pkt->ol_flags |= RTE_MBUF_F_RX_LRO;
1125                         pkt->tso_segsz = len / cqe->lro_num_seg;
1126                 }
1127                 PKT_LEN(pkt) = len;
1128                 PORT(pkt) = rxq->port_id;
1129 #ifdef MLX5_PMD_SOFT_COUNTERS
1130                 /* Increment bytes counter. */
1131                 rxq->stats.ibytes += PKT_LEN(pkt);
1132 #endif
1133                 /* Return packet. */
1134                 *(pkts++) = pkt;
1135                 ++i;
1136         }
1137         /* Update the consumer indexes. */
1138         rxq->consumed_strd = consumed_strd;
1139         rte_io_wmb();
1140         *rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
1141         if (rq_ci != rxq->rq_ci) {
1142                 rxq->rq_ci = rq_ci;
1143                 rte_io_wmb();
1144                 *rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
1145         }
1146 #ifdef MLX5_PMD_SOFT_COUNTERS
1147         /* Increment packets counter. */
1148         rxq->stats.ipackets += i;
1149 #endif
1150         return i;
1151 }
1152
1153 /**
1154  * Dummy DPDK callback for RX.
1155  *
1156  * This function is used to temporarily replace the real callback during
1157  * unsafe control operations on the queue, or in case of error.
1158  *
1159  * @param dpdk_rxq
1160  *   Generic pointer to RX queue structure.
1161  * @param[out] pkts
1162  *   Array to store received packets.
1163  * @param pkts_n
1164  *   Maximum number of packets in array.
1165  *
1166  * @return
1167  *   Number of packets successfully received (<= pkts_n).
1168  */
1169 uint16_t
1170 removed_rx_burst(void *dpdk_rxq __rte_unused,
1171                  struct rte_mbuf **pkts __rte_unused,
1172                  uint16_t pkts_n __rte_unused)
1173 {
1174         rte_mb();
1175         return 0;
1176 }
1177
1178 /*
1179  * Vectorized Rx routines are not compiled in when required vector instructions
1180  * are not supported on a target architecture.
1181  * The following null stubs are needed for linkage when those are not included
1182  * outside of this file (e.g. mlx5_rxtx_vec_sse.c for x86).
1183  */
1184
1185 __rte_weak uint16_t
1186 mlx5_rx_burst_vec(void *dpdk_rxq __rte_unused,
1187                   struct rte_mbuf **pkts __rte_unused,
1188                   uint16_t pkts_n __rte_unused)
1189 {
1190         return 0;
1191 }
1192
1193 __rte_weak uint16_t
1194 mlx5_rx_burst_mprq_vec(void *dpdk_rxq __rte_unused,
1195                        struct rte_mbuf **pkts __rte_unused,
1196                        uint16_t pkts_n __rte_unused)
1197 {
1198         return 0;
1199 }
1200
1201 __rte_weak int
1202 mlx5_rxq_check_vec_support(struct mlx5_rxq_data *rxq __rte_unused)
1203 {
1204         return -ENOTSUP;
1205 }
1206
1207 __rte_weak int
1208 mlx5_check_vec_rx_support(struct rte_eth_dev *dev __rte_unused)
1209 {
1210         return -ENOTSUP;
1211 }
1212