drivers/net/mlx5/mlx5_rxtx.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright 2015 6WIND S.A.
   5  *   Copyright 2015 Mellanox.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of 6WIND S.A. nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  */
  33
  34 #include <assert.h>
  35 #include <stdint.h>
  36 #include <string.h>
  37 #include <stdlib.h>
  38
  39 /* Verbs header. */
  40 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
  41 #ifdef PEDANTIC
  42 #pragma GCC diagnostic ignored "-Wpedantic"
  43 #endif
  44 #include <infiniband/verbs.h>
  45 #include <infiniband/mlx5dv.h>
  46 #ifdef PEDANTIC
  47 #pragma GCC diagnostic error "-Wpedantic"
  48 #endif
  49
  50 #include <rte_mbuf.h>
  51 #include <rte_mempool.h>
  52 #include <rte_prefetch.h>
  53 #include <rte_common.h>
  54 #include <rte_branch_prediction.h>
  55 #include <rte_ether.h>
  56
  57 #include "mlx5.h"
  58 #include "mlx5_utils.h"
  59 #include "mlx5_rxtx.h"
  60 #include "mlx5_autoconf.h"
  61 #include "mlx5_defs.h"
  62 #include "mlx5_prm.h"
  63
  64 static __rte_always_inline uint32_t
  65 rxq_cq_to_pkt_type(volatile struct mlx5_cqe *cqe);
  66
  67 static __rte_always_inline int
  68 mlx5_rx_poll_len(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe,
  69                  uint16_t cqe_cnt, uint32_t *rss_hash);
  70
  71 static __rte_always_inline uint32_t
  72 rxq_cq_to_ol_flags(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe);
  73
  74 uint32_t mlx5_ptype_table[] __rte_cache_aligned = {
  75         [0xff] = RTE_PTYPE_ALL_MASK, /* Last entry for errored packet. */
  76 };
  77
  78 /**
  79  * Build a table to translate Rx completion flags to packet type.
  80  *
  81  * @note: fix mlx5_dev_supported_ptypes_get() if any change here.
  82  */
  83 void
  84 mlx5_set_ptype_table(void)
  85 {
  86         unsigned int i;
  87         uint32_t (*p)[RTE_DIM(mlx5_ptype_table)] = &mlx5_ptype_table;
  88
  89         /* Last entry must not be overwritten, reserved for errored packet. */
  90         for (i = 0; i < RTE_DIM(mlx5_ptype_table) - 1; ++i)
  91                 (*p)[i] = RTE_PTYPE_UNKNOWN;
  92         /*
  93          * The index to the array should have:
  94          * bit[1:0] = l3_hdr_type
  95          * bit[4:2] = l4_hdr_type
  96          * bit[5] = ip_frag
  97          * bit[6] = tunneled
  98          * bit[7] = outer_l3_type
  99          */
 100         /* L2 */
 101         (*p)[0x00] = RTE_PTYPE_L2_ETHER;
 102         /* L3 */
 103         (*p)[0x01] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 104                      RTE_PTYPE_L4_NONFRAG;
 105         (*p)[0x02] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 106                      RTE_PTYPE_L4_NONFRAG;
 107         /* Fragmented */
 108         (*p)[0x21] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 109                      RTE_PTYPE_L4_FRAG;
 110         (*p)[0x22] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 111                      RTE_PTYPE_L4_FRAG;
 112         /* TCP */
 113         (*p)[0x05] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 114                      RTE_PTYPE_L4_TCP;
 115         (*p)[0x06] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 116                      RTE_PTYPE_L4_TCP;
 117         /* UDP */
 118         (*p)[0x09] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 119                      RTE_PTYPE_L4_UDP;
 120         (*p)[0x0a] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 121                      RTE_PTYPE_L4_UDP;
 122         /* Repeat with outer_l3_type being set. Just in case. */
 123         (*p)[0x81] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 124                      RTE_PTYPE_L4_NONFRAG;
 125         (*p)[0x82] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 126                      RTE_PTYPE_L4_NONFRAG;
 127         (*p)[0xa1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 128                      RTE_PTYPE_L4_FRAG;
 129         (*p)[0xa2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 130                      RTE_PTYPE_L4_FRAG;
 131         (*p)[0x85] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 132                      RTE_PTYPE_L4_TCP;
 133         (*p)[0x86] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 134                      RTE_PTYPE_L4_TCP;
 135         (*p)[0x89] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 136                      RTE_PTYPE_L4_UDP;
 137         (*p)[0x8a] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 138                      RTE_PTYPE_L4_UDP;
 139         /* Tunneled - L3 */
 140         (*p)[0x41] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 141                      RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
 142                      RTE_PTYPE_INNER_L4_NONFRAG;
 143         (*p)[0x42] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 144                      RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 145                      RTE_PTYPE_INNER_L4_NONFRAG;
 146         (*p)[0xc1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 147                      RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
 148                      RTE_PTYPE_INNER_L4_NONFRAG;
 149         (*p)[0xc2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 150                      RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 151                      RTE_PTYPE_INNER_L4_NONFRAG;
 152         /* Tunneled - Fragmented */
 153         (*p)[0x61] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 154                      RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
 155                      RTE_PTYPE_INNER_L4_FRAG;
 156         (*p)[0x62] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 157                      RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 158                      RTE_PTYPE_INNER_L4_FRAG;
 159         (*p)[0xe1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 160                      RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
 161                      RTE_PTYPE_INNER_L4_FRAG;
 162         (*p)[0xe2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 163                      RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 164                      RTE_PTYPE_INNER_L4_FRAG;
 165         /* Tunneled - TCP */
 166         (*p)[0x45] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 167                      RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
 168                      RTE_PTYPE_INNER_L4_TCP;
 169         (*p)[0x46] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 170                      RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 171                      RTE_PTYPE_INNER_L4_TCP;
 172         (*p)[0xc5] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 173                      RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
 174                      RTE_PTYPE_INNER_L4_TCP;
 175         (*p)[0xc6] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 176                      RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 177                      RTE_PTYPE_INNER_L4_TCP;
 178         /* Tunneled - UDP */
 179         (*p)[0x49] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 180                      RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
 181                      RTE_PTYPE_INNER_L4_UDP;
 182         (*p)[0x4a] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 183                      RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 184                      RTE_PTYPE_INNER_L4_UDP;
 185         (*p)[0xc9] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 186                      RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
 187                      RTE_PTYPE_INNER_L4_UDP;
 188         (*p)[0xca] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 189                      RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 190                      RTE_PTYPE_INNER_L4_UDP;
 191 }
 192
 193 /**
 194  * Return the size of tailroom of WQ.
 195  *
 196  * @param txq
 197  *   Pointer to TX queue structure.
 198  * @param addr
 199  *   Pointer to tail of WQ.
 200  *
 201  * @return
 202  *   Size of tailroom.
 203  */
 204 static inline size_t
 205 tx_mlx5_wq_tailroom(struct mlx5_txq_data *txq, void *addr)
 206 {
 207         size_t tailroom;
 208         tailroom = (uintptr_t)(txq->wqes) +
 209                    (1 << txq->wqe_n) * MLX5_WQE_SIZE -
 210                    (uintptr_t)addr;
 211         return tailroom;
 212 }
 213
 214 /**
 215  * Copy data to tailroom of circular queue.
 216  *
 217  * @param dst
 218  *   Pointer to destination.
 219  * @param src
 220  *   Pointer to source.
 221  * @param n
 222  *   Number of bytes to copy.
 223  * @param base
 224  *   Pointer to head of queue.
 225  * @param tailroom
 226  *   Size of tailroom from dst.
 227  *
 228  * @return
 229  *   Pointer after copied data.
 230  */
 231 static inline void *
 232 mlx5_copy_to_wq(void *dst, const void *src, size_t n,
 233                 void *base, size_t tailroom)
 234 {
 235         void *ret;
 236
 237         if (n > tailroom) {
 238                 rte_memcpy(dst, src, tailroom);
 239                 rte_memcpy(base, (void *)((uintptr_t)src + tailroom),
 240                            n - tailroom);
 241                 ret = (uint8_t *)base + n - tailroom;
 242         } else {
 243                 rte_memcpy(dst, src, n);
 244                 ret = (n == tailroom) ? base : (uint8_t *)dst + n;
 245         }
 246         return ret;
 247 }
 248
 249 /**
 250  * DPDK callback to check the status of a tx descriptor.
 251  *
 252  * @param tx_queue
 253  *   The tx queue.
 254  * @param[in] offset
 255  *   The index of the descriptor in the ring.
 256  *
 257  * @return
 258  *   The status of the tx descriptor.
 259  */
 260 int
 261 mlx5_tx_descriptor_status(void *tx_queue, uint16_t offset)
 262 {
 263         struct mlx5_txq_data *txq = tx_queue;
 264         uint16_t used;
 265
 266         mlx5_tx_complete(txq);
 267         used = txq->elts_head - txq->elts_tail;
 268         if (offset < used)
 269                 return RTE_ETH_TX_DESC_FULL;
 270         return RTE_ETH_TX_DESC_DONE;
 271 }
 272
 273 /**
 274  * DPDK callback to check the status of a rx descriptor.
 275  *
 276  * @param rx_queue
 277  *   The rx queue.
 278  * @param[in] offset
 279  *   The index of the descriptor in the ring.
 280  *
 281  * @return
 282  *   The status of the tx descriptor.
 283  */
 284 int
 285 mlx5_rx_descriptor_status(void *rx_queue, uint16_t offset)
 286 {
 287         struct mlx5_rxq_data *rxq = rx_queue;
 288         struct rxq_zip *zip = &rxq->zip;
 289         volatile struct mlx5_cqe *cqe;
 290         const unsigned int cqe_n = (1 << rxq->cqe_n);
 291         const unsigned int cqe_cnt = cqe_n - 1;
 292         unsigned int cq_ci;
 293         unsigned int used;
 294
 295         /* if we are processing a compressed cqe */
 296         if (zip->ai) {
 297                 used = zip->cqe_cnt - zip->ca;
 298                 cq_ci = zip->cq_ci;
 299         } else {
 300                 used = 0;
 301                 cq_ci = rxq->cq_ci;
 302         }
 303         cqe = &(*rxq->cqes)[cq_ci & cqe_cnt];
 304         while (check_cqe(cqe, cqe_n, cq_ci) == 0) {
 305                 int8_t op_own;
 306                 unsigned int n;
 307
 308                 op_own = cqe->op_own;
 309                 if (MLX5_CQE_FORMAT(op_own) == MLX5_COMPRESSED)
 310                         n = rte_be_to_cpu_32(cqe->byte_cnt);
 311                 else
 312                         n = 1;
 313                 cq_ci += n;
 314                 used += n;
 315                 cqe = &(*rxq->cqes)[cq_ci & cqe_cnt];
 316         }
 317         used = RTE_MIN(used, (1U << rxq->elts_n) - 1);
 318         if (offset < used)
 319                 return RTE_ETH_RX_DESC_DONE;
 320         return RTE_ETH_RX_DESC_AVAIL;
 321 }
 322
 323 /**
 324  * DPDK callback for TX.
 325  *
 326  * @param dpdk_txq
 327  *   Generic pointer to TX queue structure.
 328  * @param[in] pkts
 329  *   Packets to transmit.
 330  * @param pkts_n
 331  *   Number of packets in array.
 332  *
 333  * @return
 334  *   Number of packets successfully transmitted (<= pkts_n).
 335  */
 336 uint16_t
 337 mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
 338 {
 339         struct mlx5_txq_data *txq = (struct mlx5_txq_data *)dpdk_txq;
 340         uint16_t elts_head = txq->elts_head;
 341         const uint16_t elts_n = 1 << txq->elts_n;
 342         const uint16_t elts_m = elts_n - 1;
 343         unsigned int i = 0;
 344         unsigned int j = 0;
 345         unsigned int k = 0;
 346         uint16_t max_elts;
 347         uint16_t max_wqe;
 348         unsigned int comp;
 349         volatile struct mlx5_wqe_ctrl *last_wqe = NULL;
 350         unsigned int segs_n = 0;
 351         const unsigned int max_inline = txq->max_inline;
 352
 353         if (unlikely(!pkts_n))
 354                 return 0;
 355         /* Prefetch first packet cacheline. */
 356         rte_prefetch0(*pkts);
 357         /* Start processing. */
 358         mlx5_tx_complete(txq);
 359         max_elts = (elts_n - (elts_head - txq->elts_tail));
 360         /* A CQE slot must always be available. */
 361         assert((1u << txq->cqe_n) - (txq->cq_pi - txq->cq_ci));
 362         max_wqe = (1u << txq->wqe_n) - (txq->wqe_ci - txq->wqe_pi);
 363         if (unlikely(!max_wqe))
 364                 return 0;
 365         do {
 366                 struct rte_mbuf *buf = NULL;
 367                 uint8_t *raw;
 368                 volatile struct mlx5_wqe_v *wqe = NULL;
 369                 volatile rte_v128u32_t *dseg = NULL;
 370                 uint32_t length;
 371                 unsigned int ds = 0;
 372                 unsigned int sg = 0; /* counter of additional segs attached. */
 373                 uintptr_t addr;
 374                 uint16_t pkt_inline_sz = MLX5_WQE_DWORD_SIZE + 2;
 375                 uint16_t tso_header_sz = 0;
 376                 uint16_t ehdr;
 377                 uint8_t cs_flags;
 378                 uint64_t tso = 0;
 379                 uint16_t tso_segsz = 0;
 380 #ifdef MLX5_PMD_SOFT_COUNTERS
 381                 uint32_t total_length = 0;
 382 #endif
 383
 384                 /* first_seg */
 385                 buf = *pkts;
 386                 segs_n = buf->nb_segs;
 387                 /*
 388                  * Make sure there is enough room to store this packet and
 389                  * that one ring entry remains unused.
 390                  */
 391                 assert(segs_n);
 392                 if (max_elts < segs_n)
 393                         break;
 394                 max_elts -= segs_n;
 395                 sg = --segs_n;
 396                 if (unlikely(--max_wqe == 0))
 397                         break;
 398                 wqe = (volatile struct mlx5_wqe_v *)
 399                         tx_mlx5_wqe(txq, txq->wqe_ci);
 400                 rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci + 1));
 401                 if (pkts_n - i > 1)
 402                         rte_prefetch0(*(pkts + 1));
 403                 addr = rte_pktmbuf_mtod(buf, uintptr_t);
 404                 length = DATA_LEN(buf);
 405                 ehdr = (((uint8_t *)addr)[1] << 8) |
 406                        ((uint8_t *)addr)[0];
 407 #ifdef MLX5_PMD_SOFT_COUNTERS
 408                 total_length = length;
 409 #endif
 410                 if (length < (MLX5_WQE_DWORD_SIZE + 2)) {
 411                         txq->stats.oerrors++;
 412                         break;
 413                 }
 414                 /* Update element. */
 415                 (*txq->elts)[elts_head & elts_m] = buf;
 416                 /* Prefetch next buffer data. */
 417                 if (pkts_n - i > 1)
 418                         rte_prefetch0(
 419                             rte_pktmbuf_mtod(*(pkts + 1), volatile void *));
 420                 cs_flags = txq_ol_cksum_to_cs(txq, buf);
 421                 raw = ((uint8_t *)(uintptr_t)wqe) + 2 * MLX5_WQE_DWORD_SIZE;
 422                 /* Replace the Ethernet type by the VLAN if necessary. */
 423                 if (buf->ol_flags & PKT_TX_VLAN_PKT) {
 424                         uint32_t vlan = rte_cpu_to_be_32(0x81000000 |
 425                                                          buf->vlan_tci);
 426                         unsigned int len = 2 * ETHER_ADDR_LEN - 2;
 427
 428                         addr += 2;
 429                         length -= 2;
 430                         /* Copy Destination and source mac address. */
 431                         memcpy((uint8_t *)raw, ((uint8_t *)addr), len);
 432                         /* Copy VLAN. */
 433                         memcpy((uint8_t *)raw + len, &vlan, sizeof(vlan));
 434                         /* Copy missing two bytes to end the DSeg. */
 435                         memcpy((uint8_t *)raw + len + sizeof(vlan),
 436                                ((uint8_t *)addr) + len, 2);
 437                         addr += len + 2;
 438                         length -= (len + 2);
 439                 } else {
 440                         memcpy((uint8_t *)raw, ((uint8_t *)addr) + 2,
 441                                MLX5_WQE_DWORD_SIZE);
 442                         length -= pkt_inline_sz;
 443                         addr += pkt_inline_sz;
 444                 }
 445                 raw += MLX5_WQE_DWORD_SIZE;
 446                 tso = txq->tso_en && (buf->ol_flags & PKT_TX_TCP_SEG);
 447                 if (tso) {
 448                         uintptr_t end =
 449                                 (uintptr_t)(((uintptr_t)txq->wqes) +
 450                                             (1 << txq->wqe_n) * MLX5_WQE_SIZE);
 451                         unsigned int copy_b;
 452                         uint8_t vlan_sz =
 453                                 (buf->ol_flags & PKT_TX_VLAN_PKT) ? 4 : 0;
 454                         const uint64_t is_tunneled =
 455                                 buf->ol_flags & (PKT_TX_TUNNEL_GRE |
 456                                                  PKT_TX_TUNNEL_VXLAN);
 457
 458                         tso_header_sz = buf->l2_len + vlan_sz +
 459                                         buf->l3_len + buf->l4_len;
 460                         tso_segsz = buf->tso_segsz;
 461                         if (unlikely(tso_segsz == 0)) {
 462                                 txq->stats.oerrors++;
 463                                 break;
 464                         }
 465                         if (is_tunneled && txq->tunnel_en) {
 466                                 tso_header_sz += buf->outer_l2_len +
 467                                                  buf->outer_l3_len;
 468                                 cs_flags |= MLX5_ETH_WQE_L4_INNER_CSUM;
 469                         } else {
 470                                 cs_flags |= MLX5_ETH_WQE_L4_CSUM;
 471                         }
 472                         if (unlikely(tso_header_sz > MLX5_MAX_TSO_HEADER)) {
 473                                 txq->stats.oerrors++;
 474                                 break;
 475                         }
 476                         copy_b = tso_header_sz - pkt_inline_sz;
 477                         /* First seg must contain all headers. */
 478                         assert(copy_b <= length);
 479                         if (copy_b && ((end - (uintptr_t)raw) > copy_b)) {
 480                                 uint16_t n = (MLX5_WQE_DS(copy_b) - 1 + 3) / 4;
 481
 482                                 if (unlikely(max_wqe < n))
 483                                         break;
 484                                 max_wqe -= n;
 485                                 rte_memcpy((void *)raw, (void *)addr, copy_b);
 486                                 addr += copy_b;
 487                                 length -= copy_b;
 488                                 /* Include padding for TSO header. */
 489                                 copy_b = MLX5_WQE_DS(copy_b) *
 490                                          MLX5_WQE_DWORD_SIZE;
 491                                 pkt_inline_sz += copy_b;
 492                                 raw += copy_b;
 493                         } else {
 494                                 /* NOP WQE. */
 495                                 wqe->ctrl = (rte_v128u32_t){
 496                                         rte_cpu_to_be_32(txq->wqe_ci << 8),
 497                                         rte_cpu_to_be_32(txq->qp_num_8s | 1),
 498                                         0,
 499                                         0,
 500                                 };
 501                                 ds = 1;
 502 #ifdef MLX5_PMD_SOFT_COUNTERS
 503                                 total_length = 0;
 504 #endif
 505                                 k++;
 506                                 goto next_wqe;
 507                         }
 508                 }
 509                 /* Inline if enough room. */
 510                 if (max_inline || tso) {
 511                         uint32_t inl = 0;
 512                         uintptr_t end = (uintptr_t)
 513                                 (((uintptr_t)txq->wqes) +
 514                                  (1 << txq->wqe_n) * MLX5_WQE_SIZE);
 515                         unsigned int inline_room = max_inline *
 516                                                    RTE_CACHE_LINE_SIZE -
 517                                                    (pkt_inline_sz - 2) -
 518                                                    !!tso * sizeof(inl);
 519                         uintptr_t addr_end;
 520                         unsigned int copy_b;
 521
 522 pkt_inline:
 523                         addr_end = RTE_ALIGN_FLOOR(addr + inline_room,
 524                                                    RTE_CACHE_LINE_SIZE);
 525                         copy_b = (addr_end > addr) ?
 526                                  RTE_MIN((addr_end - addr), length) : 0;
 527                         if (copy_b && ((end - (uintptr_t)raw) > copy_b)) {
 528                                 /*
 529                                  * One Dseg remains in the current WQE.  To
 530                                  * keep the computation positive, it is
 531                                  * removed after the bytes to Dseg conversion.
 532                                  */
 533                                 uint16_t n = (MLX5_WQE_DS(copy_b) - 1 + 3) / 4;
 534
 535                                 if (unlikely(max_wqe < n))
 536                                         break;
 537                                 max_wqe -= n;
 538                                 if (tso && !inl) {
 539                                         inl = rte_cpu_to_be_32(copy_b |
 540                                                                MLX5_INLINE_SEG);
 541                                         rte_memcpy((void *)raw,
 542                                                    (void *)&inl, sizeof(inl));
 543                                         raw += sizeof(inl);
 544                                         pkt_inline_sz += sizeof(inl);
 545                                 }
 546                                 rte_memcpy((void *)raw, (void *)addr, copy_b);
 547                                 addr += copy_b;
 548                                 length -= copy_b;
 549                                 pkt_inline_sz += copy_b;
 550                         }
 551                         /*
 552                          * 2 DWORDs consumed by the WQE header + ETH segment +
 553                          * the size of the inline part of the packet.
 554                          */
 555                         ds = 2 + MLX5_WQE_DS(pkt_inline_sz - 2);
 556                         if (length > 0) {
 557                                 if (ds % (MLX5_WQE_SIZE /
 558                                           MLX5_WQE_DWORD_SIZE) == 0) {
 559                                         if (unlikely(--max_wqe == 0))
 560                                                 break;
 561                                         dseg = (volatile rte_v128u32_t *)
 562                                                tx_mlx5_wqe(txq, txq->wqe_ci +
 563                                                            ds / 4);
 564                                 } else {
 565                                         dseg = (volatile rte_v128u32_t *)
 566                                                 ((uintptr_t)wqe +
 567                                                  (ds * MLX5_WQE_DWORD_SIZE));
 568                                 }
 569                                 goto use_dseg;
 570                         } else if (!segs_n) {
 571                                 goto next_pkt;
 572                         } else {
 573                                 raw += copy_b;
 574                                 inline_room -= copy_b;
 575                                 --segs_n;
 576                                 buf = buf->next;
 577                                 assert(buf);
 578                                 addr = rte_pktmbuf_mtod(buf, uintptr_t);
 579                                 length = DATA_LEN(buf);
 580 #ifdef MLX5_PMD_SOFT_COUNTERS
 581                                 total_length += length;
 582 #endif
 583                                 (*txq->elts)[++elts_head & elts_m] = buf;
 584                                 goto pkt_inline;
 585                         }
 586                 } else {
 587                         /*
 588                          * No inline has been done in the packet, only the
 589                          * Ethernet Header as been stored.
 590                          */
 591                         dseg = (volatile rte_v128u32_t *)
 592                                 ((uintptr_t)wqe + (3 * MLX5_WQE_DWORD_SIZE));
 593                         ds = 3;
 594 use_dseg:
 595                         /* Add the remaining packet as a simple ds. */
 596                         addr = rte_cpu_to_be_64(addr);
 597                         *dseg = (rte_v128u32_t){
 598                                 rte_cpu_to_be_32(length),
 599                                 mlx5_tx_mb2mr(txq, buf),
 600                                 addr,
 601                                 addr >> 32,
 602                         };
 603                         ++ds;
 604                         if (!segs_n)
 605                                 goto next_pkt;
 606                 }
 607 next_seg:
 608                 assert(buf);
 609                 assert(ds);
 610                 assert(wqe);
 611                 /*
 612                  * Spill on next WQE when the current one does not have
 613                  * enough room left. Size of WQE must a be a multiple
 614                  * of data segment size.
 615                  */
 616                 assert(!(MLX5_WQE_SIZE % MLX5_WQE_DWORD_SIZE));
 617                 if (!(ds % (MLX5_WQE_SIZE / MLX5_WQE_DWORD_SIZE))) {
 618                         if (unlikely(--max_wqe == 0))
 619                                 break;
 620                         dseg = (volatile rte_v128u32_t *)
 621                                tx_mlx5_wqe(txq, txq->wqe_ci + ds / 4);
 622                         rte_prefetch0(tx_mlx5_wqe(txq,
 623                                                   txq->wqe_ci + ds / 4 + 1));
 624                 } else {
 625                         ++dseg;
 626                 }
 627                 ++ds;
 628                 buf = buf->next;
 629                 assert(buf);
 630                 length = DATA_LEN(buf);
 631 #ifdef MLX5_PMD_SOFT_COUNTERS
 632                 total_length += length;
 633 #endif
 634                 /* Store segment information. */
 635                 addr = rte_cpu_to_be_64(rte_pktmbuf_mtod(buf, uintptr_t));
 636                 *dseg = (rte_v128u32_t){
 637                         rte_cpu_to_be_32(length),
 638                         mlx5_tx_mb2mr(txq, buf),
 639                         addr,
 640                         addr >> 32,
 641                 };
 642                 (*txq->elts)[++elts_head & elts_m] = buf;
 643                 if (--segs_n)
 644                         goto next_seg;
 645 next_pkt:
 646                 if (ds > MLX5_DSEG_MAX) {
 647                         txq->stats.oerrors++;
 648                         break;
 649                 }
 650                 ++elts_head;
 651                 ++pkts;
 652                 ++i;
 653                 j += sg;
 654                 /* Initialize known and common part of the WQE structure. */
 655                 if (tso) {
 656                         wqe->ctrl = (rte_v128u32_t){
 657                                 rte_cpu_to_be_32((txq->wqe_ci << 8) |
 658                                                  MLX5_OPCODE_TSO),
 659                                 rte_cpu_to_be_32(txq->qp_num_8s | ds),
 660                                 0,
 661                                 0,
 662                         };
 663                         wqe->eseg = (rte_v128u32_t){
 664                                 0,
 665                                 cs_flags | (rte_cpu_to_be_16(tso_segsz) << 16),
 666                                 0,
 667                                 (ehdr << 16) | rte_cpu_to_be_16(tso_header_sz),
 668                         };
 669                 } else {
 670                         wqe->ctrl = (rte_v128u32_t){
 671                                 rte_cpu_to_be_32((txq->wqe_ci << 8) |
 672                                                  MLX5_OPCODE_SEND),
 673                                 rte_cpu_to_be_32(txq->qp_num_8s | ds),
 674                                 0,
 675                                 0,
 676                         };
 677                         wqe->eseg = (rte_v128u32_t){
 678                                 0,
 679                                 cs_flags,
 680                                 0,
 681                                 (ehdr << 16) | rte_cpu_to_be_16(pkt_inline_sz),
 682                         };
 683                 }
 684 next_wqe:
 685                 txq->wqe_ci += (ds + 3) / 4;
 686                 /* Save the last successful WQE for completion request */
 687                 last_wqe = (volatile struct mlx5_wqe_ctrl *)wqe;
 688 #ifdef MLX5_PMD_SOFT_COUNTERS
 689                 /* Increment sent bytes counter. */
 690                 txq->stats.obytes += total_length;
 691 #endif
 692         } while (i < pkts_n);
 693         /* Take a shortcut if nothing must be sent. */
 694         if (unlikely((i + k) == 0))
 695                 return 0;
 696         txq->elts_head += (i + j);
 697         /* Check whether completion threshold has been reached. */
 698         comp = txq->elts_comp + i + j + k;
 699         if (comp >= MLX5_TX_COMP_THRESH) {
 700                 /* Request completion on last WQE. */
 701                 last_wqe->ctrl2 = rte_cpu_to_be_32(8);
 702                 /* Save elts_head in unused "immediate" field of WQE. */
 703                 last_wqe->ctrl3 = txq->elts_head;
 704                 txq->elts_comp = 0;
 705 #ifndef NDEBUG
 706                 ++txq->cq_pi;
 707 #endif
 708         } else {
 709                 txq->elts_comp = comp;
 710         }
 711 #ifdef MLX5_PMD_SOFT_COUNTERS
 712         /* Increment sent packets counter. */
 713         txq->stats.opackets += i;
 714 #endif
 715         /* Ring QP doorbell. */
 716         mlx5_tx_dbrec(txq, (volatile struct mlx5_wqe *)last_wqe);
 717         return i;
 718 }
 719
 720 /**
 721  * Open a MPW session.
 722  *
 723  * @param txq
 724  *   Pointer to TX queue structure.
 725  * @param mpw
 726  *   Pointer to MPW session structure.
 727  * @param length
 728  *   Packet length.
 729  */
 730 static inline void
 731 mlx5_mpw_new(struct mlx5_txq_data *txq, struct mlx5_mpw *mpw, uint32_t length)
 732 {
 733         uint16_t idx = txq->wqe_ci & ((1 << txq->wqe_n) - 1);
 734         volatile struct mlx5_wqe_data_seg (*dseg)[MLX5_MPW_DSEG_MAX] =
 735                 (volatile struct mlx5_wqe_data_seg (*)[])
 736                 tx_mlx5_wqe(txq, idx + 1);
 737
 738         mpw->state = MLX5_MPW_STATE_OPENED;
 739         mpw->pkts_n = 0;
 740         mpw->len = length;
 741         mpw->total_len = 0;
 742         mpw->wqe = (volatile struct mlx5_wqe *)tx_mlx5_wqe(txq, idx);
 743         mpw->wqe->eseg.mss = rte_cpu_to_be_16(length);
 744         mpw->wqe->eseg.inline_hdr_sz = 0;
 745         mpw->wqe->eseg.rsvd0 = 0;
 746         mpw->wqe->eseg.rsvd1 = 0;
 747         mpw->wqe->eseg.rsvd2 = 0;
 748         mpw->wqe->ctrl[0] = rte_cpu_to_be_32((MLX5_OPC_MOD_MPW << 24) |
 749                                              (txq->wqe_ci << 8) |
 750                                              MLX5_OPCODE_TSO);
 751         mpw->wqe->ctrl[2] = 0;
 752         mpw->wqe->ctrl[3] = 0;
 753         mpw->data.dseg[0] = (volatile struct mlx5_wqe_data_seg *)
 754                 (((uintptr_t)mpw->wqe) + (2 * MLX5_WQE_DWORD_SIZE));
 755         mpw->data.dseg[1] = (volatile struct mlx5_wqe_data_seg *)
 756                 (((uintptr_t)mpw->wqe) + (3 * MLX5_WQE_DWORD_SIZE));
 757         mpw->data.dseg[2] = &(*dseg)[0];
 758         mpw->data.dseg[3] = &(*dseg)[1];
 759         mpw->data.dseg[4] = &(*dseg)[2];
 760 }
 761
 762 /**
 763  * Close a MPW session.
 764  *
 765  * @param txq
 766  *   Pointer to TX queue structure.
 767  * @param mpw
 768  *   Pointer to MPW session structure.
 769  */
 770 static inline void
 771 mlx5_mpw_close(struct mlx5_txq_data *txq, struct mlx5_mpw *mpw)
 772 {
 773         unsigned int num = mpw->pkts_n;
 774
 775         /*
 776          * Store size in multiple of 16 bytes. Control and Ethernet segments
 777          * count as 2.
 778          */
 779         mpw->wqe->ctrl[1] = rte_cpu_to_be_32(txq->qp_num_8s | (2 + num));
 780         mpw->state = MLX5_MPW_STATE_CLOSED;
 781         if (num < 3)
 782                 ++txq->wqe_ci;
 783         else
 784                 txq->wqe_ci += 2;
 785         rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci));
 786         rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci + 1));
 787 }
 788
 789 /**
 790  * DPDK callback for TX with MPW support.
 791  *
 792  * @param dpdk_txq
 793  *   Generic pointer to TX queue structure.
 794  * @param[in] pkts
 795  *   Packets to transmit.
 796  * @param pkts_n
 797  *   Number of packets in array.
 798  *
 799  * @return
 800  *   Number of packets successfully transmitted (<= pkts_n).
 801  */
 802 uint16_t
 803 mlx5_tx_burst_mpw(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
 804 {
 805         struct mlx5_txq_data *txq = (struct mlx5_txq_data *)dpdk_txq;
 806         uint16_t elts_head = txq->elts_head;
 807         const uint16_t elts_n = 1 << txq->elts_n;
 808         const uint16_t elts_m = elts_n - 1;
 809         unsigned int i = 0;
 810         unsigned int j = 0;
 811         uint16_t max_elts;
 812         uint16_t max_wqe;
 813         unsigned int comp;
 814         struct mlx5_mpw mpw = {
 815                 .state = MLX5_MPW_STATE_CLOSED,
 816         };
 817
 818         if (unlikely(!pkts_n))
 819                 return 0;
 820         /* Prefetch first packet cacheline. */
 821         rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci));
 822         rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci + 1));
 823         /* Start processing. */
 824         mlx5_tx_complete(txq);
 825         max_elts = (elts_n - (elts_head - txq->elts_tail));
 826         /* A CQE slot must always be available. */
 827         assert((1u << txq->cqe_n) - (txq->cq_pi - txq->cq_ci));
 828         max_wqe = (1u << txq->wqe_n) - (txq->wqe_ci - txq->wqe_pi);
 829         if (unlikely(!max_wqe))
 830                 return 0;
 831         do {
 832                 struct rte_mbuf *buf = *(pkts++);
 833                 uint32_t length;
 834                 unsigned int segs_n = buf->nb_segs;
 835                 uint32_t cs_flags;
 836
 837                 /*
 838                  * Make sure there is enough room to store this packet and
 839                  * that one ring entry remains unused.
 840                  */
 841                 assert(segs_n);
 842                 if (max_elts < segs_n)
 843                         break;
 844                 /* Do not bother with large packets MPW cannot handle. */
 845                 if (segs_n > MLX5_MPW_DSEG_MAX) {
 846                         txq->stats.oerrors++;
 847                         break;
 848                 }
 849                 max_elts -= segs_n;
 850                 --pkts_n;
 851                 cs_flags = txq_ol_cksum_to_cs(txq, buf);
 852                 /* Retrieve packet information. */
 853                 length = PKT_LEN(buf);
 854                 assert(length);
 855                 /* Start new session if packet differs. */
 856                 if ((mpw.state == MLX5_MPW_STATE_OPENED) &&
 857                     ((mpw.len != length) ||
 858                      (segs_n != 1) ||
 859                      (mpw.wqe->eseg.cs_flags != cs_flags)))
 860                         mlx5_mpw_close(txq, &mpw);
 861                 if (mpw.state == MLX5_MPW_STATE_CLOSED) {
 862                         /*
 863                          * Multi-Packet WQE consumes at most two WQE.
 864                          * mlx5_mpw_new() expects to be able to use such
 865                          * resources.
 866                          */
 867                         if (unlikely(max_wqe < 2))
 868                                 break;
 869                         max_wqe -= 2;
 870                         mlx5_mpw_new(txq, &mpw, length);
 871                         mpw.wqe->eseg.cs_flags = cs_flags;
 872                 }
 873                 /* Multi-segment packets must be alone in their MPW. */
 874                 assert((segs_n == 1) || (mpw.pkts_n == 0));
 875 #if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
 876                 length = 0;
 877 #endif
 878                 do {
 879                         volatile struct mlx5_wqe_data_seg *dseg;
 880                         uintptr_t addr;
 881
 882                         assert(buf);
 883                         (*txq->elts)[elts_head++ & elts_m] = buf;
 884                         dseg = mpw.data.dseg[mpw.pkts_n];
 885                         addr = rte_pktmbuf_mtod(buf, uintptr_t);
 886                         *dseg = (struct mlx5_wqe_data_seg){
 887                                 .byte_count = rte_cpu_to_be_32(DATA_LEN(buf)),
 888                                 .lkey = mlx5_tx_mb2mr(txq, buf),
 889                                 .addr = rte_cpu_to_be_64(addr),
 890                         };
 891 #if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
 892                         length += DATA_LEN(buf);
 893 #endif
 894                         buf = buf->next;
 895                         ++mpw.pkts_n;
 896                         ++j;
 897                 } while (--segs_n);
 898                 assert(length == mpw.len);
 899                 if (mpw.pkts_n == MLX5_MPW_DSEG_MAX)
 900                         mlx5_mpw_close(txq, &mpw);
 901 #ifdef MLX5_PMD_SOFT_COUNTERS
 902                 /* Increment sent bytes counter. */
 903                 txq->stats.obytes += length;
 904 #endif
 905                 ++i;
 906         } while (pkts_n);
 907         /* Take a shortcut if nothing must be sent. */
 908         if (unlikely(i == 0))
 909                 return 0;
 910         /* Check whether completion threshold has been reached. */
 911         /* "j" includes both packets and segments. */
 912         comp = txq->elts_comp + j;
 913         if (comp >= MLX5_TX_COMP_THRESH) {
 914                 volatile struct mlx5_wqe *wqe = mpw.wqe;
 915
 916                 /* Request completion on last WQE. */
 917                 wqe->ctrl[2] = rte_cpu_to_be_32(8);
 918                 /* Save elts_head in unused "immediate" field of WQE. */
 919                 wqe->ctrl[3] = elts_head;
 920                 txq->elts_comp = 0;
 921 #ifndef NDEBUG
 922                 ++txq->cq_pi;
 923 #endif
 924         } else {
 925                 txq->elts_comp = comp;
 926         }
 927 #ifdef MLX5_PMD_SOFT_COUNTERS
 928         /* Increment sent packets counter. */
 929         txq->stats.opackets += i;
 930 #endif
 931         /* Ring QP doorbell. */
 932         if (mpw.state == MLX5_MPW_STATE_OPENED)
 933                 mlx5_mpw_close(txq, &mpw);
 934         mlx5_tx_dbrec(txq, mpw.wqe);
 935         txq->elts_head = elts_head;
 936         return i;
 937 }
 938
 939 /**
 940  * Open a MPW inline session.
 941  *
 942  * @param txq
 943  *   Pointer to TX queue structure.
 944  * @param mpw
 945  *   Pointer to MPW session structure.
 946  * @param length
 947  *   Packet length.
 948  */
 949 static inline void
 950 mlx5_mpw_inline_new(struct mlx5_txq_data *txq, struct mlx5_mpw *mpw,
 951                     uint32_t length)
 952 {
 953         uint16_t idx = txq->wqe_ci & ((1 << txq->wqe_n) - 1);
 954         struct mlx5_wqe_inl_small *inl;
 955
 956         mpw->state = MLX5_MPW_INL_STATE_OPENED;
 957         mpw->pkts_n = 0;
 958         mpw->len = length;
 959         mpw->total_len = 0;
 960         mpw->wqe = (volatile struct mlx5_wqe *)tx_mlx5_wqe(txq, idx);
 961         mpw->wqe->ctrl[0] = rte_cpu_to_be_32((MLX5_OPC_MOD_MPW << 24) |
 962                                              (txq->wqe_ci << 8) |
 963                                              MLX5_OPCODE_TSO);
 964         mpw->wqe->ctrl[2] = 0;
 965         mpw->wqe->ctrl[3] = 0;
 966         mpw->wqe->eseg.mss = rte_cpu_to_be_16(length);
 967         mpw->wqe->eseg.inline_hdr_sz = 0;
 968         mpw->wqe->eseg.cs_flags = 0;
 969         mpw->wqe->eseg.rsvd0 = 0;
 970         mpw->wqe->eseg.rsvd1 = 0;
 971         mpw->wqe->eseg.rsvd2 = 0;
 972         inl = (struct mlx5_wqe_inl_small *)
 973                 (((uintptr_t)mpw->wqe) + 2 * MLX5_WQE_DWORD_SIZE);
 974         mpw->data.raw = (uint8_t *)&inl->raw;
 975 }
 976
 977 /**
 978  * Close a MPW inline session.
 979  *
 980  * @param txq
 981  *   Pointer to TX queue structure.
 982  * @param mpw
 983  *   Pointer to MPW session structure.
 984  */
 985 static inline void
 986 mlx5_mpw_inline_close(struct mlx5_txq_data *txq, struct mlx5_mpw *mpw)
 987 {
 988         unsigned int size;
 989         struct mlx5_wqe_inl_small *inl = (struct mlx5_wqe_inl_small *)
 990                 (((uintptr_t)mpw->wqe) + (2 * MLX5_WQE_DWORD_SIZE));
 991
 992         size = MLX5_WQE_SIZE - MLX5_MWQE64_INL_DATA + mpw->total_len;
 993         /*
 994          * Store size in multiple of 16 bytes. Control and Ethernet segments
 995          * count as 2.
 996          */
 997         mpw->wqe->ctrl[1] = rte_cpu_to_be_32(txq->qp_num_8s |
 998                                              MLX5_WQE_DS(size));
 999         mpw->state = MLX5_MPW_STATE_CLOSED;
1000         inl->byte_cnt = rte_cpu_to_be_32(mpw->total_len | MLX5_INLINE_SEG);
1001         txq->wqe_ci += (size + (MLX5_WQE_SIZE - 1)) / MLX5_WQE_SIZE;
1002 }
1003
1004 /**
1005  * DPDK callback for TX with MPW inline support.
1006  *
1007  * @param dpdk_txq
1008  *   Generic pointer to TX queue structure.
1009  * @param[in] pkts
1010  *   Packets to transmit.
1011  * @param pkts_n
1012  *   Number of packets in array.
1013  *
1014  * @return
1015  *   Number of packets successfully transmitted (<= pkts_n).
1016  */
1017 uint16_t
1018 mlx5_tx_burst_mpw_inline(void *dpdk_txq, struct rte_mbuf **pkts,
1019                          uint16_t pkts_n)
1020 {
1021         struct mlx5_txq_data *txq = (struct mlx5_txq_data *)dpdk_txq;
1022         uint16_t elts_head = txq->elts_head;
1023         const uint16_t elts_n = 1 << txq->elts_n;
1024         const uint16_t elts_m = elts_n - 1;
1025         unsigned int i = 0;
1026         unsigned int j = 0;
1027         uint16_t max_elts;
1028         uint16_t max_wqe;
1029         unsigned int comp;
1030         unsigned int inline_room = txq->max_inline * RTE_CACHE_LINE_SIZE;
1031         struct mlx5_mpw mpw = {
1032                 .state = MLX5_MPW_STATE_CLOSED,
1033         };
1034         /*
1035          * Compute the maximum number of WQE which can be consumed by inline
1036          * code.
1037          * - 2 DSEG for:
1038          *   - 1 control segment,
1039          *   - 1 Ethernet segment,
1040          * - N Dseg from the inline request.
1041          */
1042         const unsigned int wqe_inl_n =
1043                 ((2 * MLX5_WQE_DWORD_SIZE +
1044                   txq->max_inline * RTE_CACHE_LINE_SIZE) +
1045                  RTE_CACHE_LINE_SIZE - 1) / RTE_CACHE_LINE_SIZE;
1046
1047         if (unlikely(!pkts_n))
1048                 return 0;
1049         /* Prefetch first packet cacheline. */
1050         rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci));
1051         rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci + 1));
1052         /* Start processing. */
1053         mlx5_tx_complete(txq);
1054         max_elts = (elts_n - (elts_head - txq->elts_tail));
1055         /* A CQE slot must always be available. */
1056         assert((1u << txq->cqe_n) - (txq->cq_pi - txq->cq_ci));
1057         do {
1058                 struct rte_mbuf *buf = *(pkts++);
1059                 uintptr_t addr;
1060                 uint32_t length;
1061                 unsigned int segs_n = buf->nb_segs;
1062                 uint8_t cs_flags;
1063
1064                 /*
1065                  * Make sure there is enough room to store this packet and
1066                  * that one ring entry remains unused.
1067                  */
1068                 assert(segs_n);
1069                 if (max_elts < segs_n)
1070                         break;
1071                 /* Do not bother with large packets MPW cannot handle. */
1072                 if (segs_n > MLX5_MPW_DSEG_MAX) {
1073                         txq->stats.oerrors++;
1074                         break;
1075                 }
1076                 max_elts -= segs_n;
1077                 --pkts_n;
1078                 /*
1079                  * Compute max_wqe in case less WQE were consumed in previous
1080                  * iteration.
1081                  */
1082                 max_wqe = (1u << txq->wqe_n) - (txq->wqe_ci - txq->wqe_pi);
1083                 cs_flags = txq_ol_cksum_to_cs(txq, buf);
1084                 /* Retrieve packet information. */
1085                 length = PKT_LEN(buf);
1086                 /* Start new session if packet differs. */
1087                 if (mpw.state == MLX5_MPW_STATE_OPENED) {
1088                         if ((mpw.len != length) ||
1089                             (segs_n != 1) ||
1090                             (mpw.wqe->eseg.cs_flags != cs_flags))
1091                                 mlx5_mpw_close(txq, &mpw);
1092                 } else if (mpw.state == MLX5_MPW_INL_STATE_OPENED) {
1093                         if ((mpw.len != length) ||
1094                             (segs_n != 1) ||
1095                             (length > inline_room) ||
1096                             (mpw.wqe->eseg.cs_flags != cs_flags)) {
1097                                 mlx5_mpw_inline_close(txq, &mpw);
1098                                 inline_room =
1099                                         txq->max_inline * RTE_CACHE_LINE_SIZE;
1100                         }
1101                 }
1102                 if (mpw.state == MLX5_MPW_STATE_CLOSED) {
1103                         if ((segs_n != 1) ||
1104                             (length > inline_room)) {
1105                                 /*
1106                                  * Multi-Packet WQE consumes at most two WQE.
1107                                  * mlx5_mpw_new() expects to be able to use
1108                                  * such resources.
1109                                  */
1110                                 if (unlikely(max_wqe < 2))
1111                                         break;
1112                                 max_wqe -= 2;
1113                                 mlx5_mpw_new(txq, &mpw, length);
1114                                 mpw.wqe->eseg.cs_flags = cs_flags;
1115                         } else {
1116                                 if (unlikely(max_wqe < wqe_inl_n))
1117                                         break;
1118                                 max_wqe -= wqe_inl_n;
1119                                 mlx5_mpw_inline_new(txq, &mpw, length);
1120                                 mpw.wqe->eseg.cs_flags = cs_flags;
1121                         }
1122                 }
1123                 /* Multi-segment packets must be alone in their MPW. */
1124                 assert((segs_n == 1) || (mpw.pkts_n == 0));
1125                 if (mpw.state == MLX5_MPW_STATE_OPENED) {
1126                         assert(inline_room ==
1127                                txq->max_inline * RTE_CACHE_LINE_SIZE);
1128 #if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
1129                         length = 0;
1130 #endif
1131                         do {
1132                                 volatile struct mlx5_wqe_data_seg *dseg;
1133
1134                                 assert(buf);
1135                                 (*txq->elts)[elts_head++ & elts_m] = buf;
1136                                 dseg = mpw.data.dseg[mpw.pkts_n];
1137                                 addr = rte_pktmbuf_mtod(buf, uintptr_t);
1138                                 *dseg = (struct mlx5_wqe_data_seg){
1139                                         .byte_count =
1140                                                rte_cpu_to_be_32(DATA_LEN(buf)),
1141                                         .lkey = mlx5_tx_mb2mr(txq, buf),
1142                                         .addr = rte_cpu_to_be_64(addr),
1143                                 };
1144 #if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
1145                                 length += DATA_LEN(buf);
1146 #endif
1147                                 buf = buf->next;
1148                                 ++mpw.pkts_n;
1149                                 ++j;
1150                         } while (--segs_n);
1151                         assert(length == mpw.len);
1152                         if (mpw.pkts_n == MLX5_MPW_DSEG_MAX)
1153                                 mlx5_mpw_close(txq, &mpw);
1154                 } else {
1155                         unsigned int max;
1156
1157                         assert(mpw.state == MLX5_MPW_INL_STATE_OPENED);
1158                         assert(length <= inline_room);
1159                         assert(length == DATA_LEN(buf));
1160                         addr = rte_pktmbuf_mtod(buf, uintptr_t);
1161                         (*txq->elts)[elts_head++ & elts_m] = buf;
1162                         /* Maximum number of bytes before wrapping. */
1163                         max = ((((uintptr_t)(txq->wqes)) +
1164                                 (1 << txq->wqe_n) *
1165                                 MLX5_WQE_SIZE) -
1166                                (uintptr_t)mpw.data.raw);
1167                         if (length > max) {
1168                                 rte_memcpy((void *)(uintptr_t)mpw.data.raw,
1169                                            (void *)addr,
1170                                            max);
1171                                 mpw.data.raw = (volatile void *)txq->wqes;
1172                                 rte_memcpy((void *)(uintptr_t)mpw.data.raw,
1173                                            (void *)(addr + max),
1174                                            length - max);
1175                                 mpw.data.raw += length - max;
1176                         } else {
1177                                 rte_memcpy((void *)(uintptr_t)mpw.data.raw,
1178                                            (void *)addr,
1179                                            length);
1180
1181                                 if (length == max)
1182                                         mpw.data.raw =
1183                                                 (volatile void *)txq->wqes;
1184                                 else
1185                                         mpw.data.raw += length;
1186                         }
1187                         ++mpw.pkts_n;
1188                         mpw.total_len += length;
1189                         ++j;
1190                         if (mpw.pkts_n == MLX5_MPW_DSEG_MAX) {
1191                                 mlx5_mpw_inline_close(txq, &mpw);
1192                                 inline_room =
1193                                         txq->max_inline * RTE_CACHE_LINE_SIZE;
1194                         } else {
1195                                 inline_room -= length;
1196                         }
1197                 }
1198 #ifdef MLX5_PMD_SOFT_COUNTERS
1199                 /* Increment sent bytes counter. */
1200                 txq->stats.obytes += length;
1201 #endif
1202                 ++i;
1203         } while (pkts_n);
1204         /* Take a shortcut if nothing must be sent. */
1205         if (unlikely(i == 0))
1206                 return 0;
1207         /* Check whether completion threshold has been reached. */
1208         /* "j" includes both packets and segments. */
1209         comp = txq->elts_comp + j;
1210         if (comp >= MLX5_TX_COMP_THRESH) {
1211                 volatile struct mlx5_wqe *wqe = mpw.wqe;
1212
1213                 /* Request completion on last WQE. */
1214                 wqe->ctrl[2] = rte_cpu_to_be_32(8);
1215                 /* Save elts_head in unused "immediate" field of WQE. */
1216                 wqe->ctrl[3] = elts_head;
1217                 txq->elts_comp = 0;
1218 #ifndef NDEBUG
1219                 ++txq->cq_pi;
1220 #endif
1221         } else {
1222                 txq->elts_comp = comp;
1223         }
1224 #ifdef MLX5_PMD_SOFT_COUNTERS
1225         /* Increment sent packets counter. */
1226         txq->stats.opackets += i;
1227 #endif
1228         /* Ring QP doorbell. */
1229         if (mpw.state == MLX5_MPW_INL_STATE_OPENED)
1230                 mlx5_mpw_inline_close(txq, &mpw);
1231         else if (mpw.state == MLX5_MPW_STATE_OPENED)
1232                 mlx5_mpw_close(txq, &mpw);
1233         mlx5_tx_dbrec(txq, mpw.wqe);
1234         txq->elts_head = elts_head;
1235         return i;
1236 }
1237
1238 /**
1239  * Open an Enhanced MPW session.
1240  *
1241  * @param txq
1242  *   Pointer to TX queue structure.
1243  * @param mpw
1244  *   Pointer to MPW session structure.
1245  * @param length
1246  *   Packet length.
1247  */
1248 static inline void
1249 mlx5_empw_new(struct mlx5_txq_data *txq, struct mlx5_mpw *mpw, int padding)
1250 {
1251         uint16_t idx = txq->wqe_ci & ((1 << txq->wqe_n) - 1);
1252
1253         mpw->state = MLX5_MPW_ENHANCED_STATE_OPENED;
1254         mpw->pkts_n = 0;
1255         mpw->total_len = sizeof(struct mlx5_wqe);
1256         mpw->wqe = (volatile struct mlx5_wqe *)tx_mlx5_wqe(txq, idx);
1257         mpw->wqe->ctrl[0] =
1258                 rte_cpu_to_be_32((MLX5_OPC_MOD_ENHANCED_MPSW << 24) |
1259                                  (txq->wqe_ci << 8) |
1260                                  MLX5_OPCODE_ENHANCED_MPSW);
1261         mpw->wqe->ctrl[2] = 0;
1262         mpw->wqe->ctrl[3] = 0;
1263         memset((void *)(uintptr_t)&mpw->wqe->eseg, 0, MLX5_WQE_DWORD_SIZE);
1264         if (unlikely(padding)) {
1265                 uintptr_t addr = (uintptr_t)(mpw->wqe + 1);
1266
1267                 /* Pad the first 2 DWORDs with zero-length inline header. */
1268                 *(volatile uint32_t *)addr = rte_cpu_to_be_32(MLX5_INLINE_SEG);
1269                 *(volatile uint32_t *)(addr + MLX5_WQE_DWORD_SIZE) =
1270                         rte_cpu_to_be_32(MLX5_INLINE_SEG);
1271                 mpw->total_len += 2 * MLX5_WQE_DWORD_SIZE;
1272                 /* Start from the next WQEBB. */
1273                 mpw->data.raw = (volatile void *)(tx_mlx5_wqe(txq, idx + 1));
1274         } else {
1275                 mpw->data.raw = (volatile void *)(mpw->wqe + 1);
1276         }
1277 }
1278
1279 /**
1280  * Close an Enhanced MPW session.
1281  *
1282  * @param txq
1283  *   Pointer to TX queue structure.
1284  * @param mpw
1285  *   Pointer to MPW session structure.
1286  *
1287  * @return
1288  *   Number of consumed WQEs.
1289  */
1290 static inline uint16_t
1291 mlx5_empw_close(struct mlx5_txq_data *txq, struct mlx5_mpw *mpw)
1292 {
1293         uint16_t ret;
1294
1295         /* Store size in multiple of 16 bytes. Control and Ethernet segments
1296          * count as 2.
1297          */
1298         mpw->wqe->ctrl[1] = rte_cpu_to_be_32(txq->qp_num_8s |
1299                                              MLX5_WQE_DS(mpw->total_len));
1300         mpw->state = MLX5_MPW_STATE_CLOSED;
1301         ret = (mpw->total_len + (MLX5_WQE_SIZE - 1)) / MLX5_WQE_SIZE;
1302         txq->wqe_ci += ret;
1303         return ret;
1304 }
1305
1306 /**
1307  * TX with Enhanced MPW support.
1308  *
1309  * @param txq
1310  *   Pointer to TX queue structure.
1311  * @param[in] pkts
1312  *   Packets to transmit.
1313  * @param pkts_n
1314  *   Number of packets in array.
1315  *
1316  * @return
1317  *   Number of packets successfully transmitted (<= pkts_n).
1318  */
1319 static inline uint16_t
1320 txq_burst_empw(struct mlx5_txq_data *txq, struct rte_mbuf **pkts,
1321                uint16_t pkts_n)
1322 {
1323         uint16_t elts_head = txq->elts_head;
1324         const uint16_t elts_n = 1 << txq->elts_n;
1325         const uint16_t elts_m = elts_n - 1;
1326         unsigned int i = 0;
1327         unsigned int j = 0;
1328         uint16_t max_elts;
1329         uint16_t max_wqe;
1330         unsigned int max_inline = txq->max_inline * RTE_CACHE_LINE_SIZE;
1331         unsigned int mpw_room = 0;
1332         unsigned int inl_pad = 0;
1333         uint32_t inl_hdr;
1334         struct mlx5_mpw mpw = {
1335                 .state = MLX5_MPW_STATE_CLOSED,
1336         };
1337
1338         if (unlikely(!pkts_n))
1339                 return 0;
1340         /* Start processing. */
1341         mlx5_tx_complete(txq);
1342         max_elts = (elts_n - (elts_head - txq->elts_tail));
1343         /* A CQE slot must always be available. */
1344         assert((1u << txq->cqe_n) - (txq->cq_pi - txq->cq_ci));
1345         max_wqe = (1u << txq->wqe_n) - (txq->wqe_ci - txq->wqe_pi);
1346         if (unlikely(!max_wqe))
1347                 return 0;
1348         do {
1349                 struct rte_mbuf *buf = *(pkts++);
1350                 uintptr_t addr;
1351                 unsigned int n;
1352                 unsigned int do_inline = 0; /* Whether inline is possible. */
1353                 uint32_t length;
1354                 uint8_t cs_flags;
1355
1356                 /* Multi-segmented packet is handled in slow-path outside. */
1357                 assert(NB_SEGS(buf) == 1);
1358                 /* Make sure there is enough room to store this packet. */
1359                 if (max_elts - j == 0)
1360                         break;
1361                 cs_flags = txq_ol_cksum_to_cs(txq, buf);
1362                 /* Retrieve packet information. */
1363                 length = PKT_LEN(buf);
1364                 /* Start new session if:
1365                  * - multi-segment packet
1366                  * - no space left even for a dseg
1367                  * - next packet can be inlined with a new WQE
1368                  * - cs_flag differs
1369                  */
1370                 if (mpw.state == MLX5_MPW_ENHANCED_STATE_OPENED) {
1371                         if ((inl_pad + sizeof(struct mlx5_wqe_data_seg) >
1372                              mpw_room) ||
1373                             (length <= txq->inline_max_packet_sz &&
1374                              inl_pad + sizeof(inl_hdr) + length >
1375                              mpw_room) ||
1376                             (mpw.wqe->eseg.cs_flags != cs_flags))
1377                                 max_wqe -= mlx5_empw_close(txq, &mpw);
1378                 }
1379                 if (unlikely(mpw.state == MLX5_MPW_STATE_CLOSED)) {
1380                         /* In Enhanced MPW, inline as much as the budget is
1381                          * allowed. The remaining space is to be filled with
1382                          * dsegs. If the title WQEBB isn't padded, it will have
1383                          * 2 dsegs there.
1384                          */
1385                         mpw_room = RTE_MIN(MLX5_WQE_SIZE_MAX,
1386                                            (max_inline ? max_inline :
1387                                             pkts_n * MLX5_WQE_DWORD_SIZE) +
1388                                            MLX5_WQE_SIZE);
1389                         if (unlikely(max_wqe * MLX5_WQE_SIZE < mpw_room))
1390                                 break;
1391                         /* Don't pad the title WQEBB to not waste WQ. */
1392                         mlx5_empw_new(txq, &mpw, 0);
1393                         mpw_room -= mpw.total_len;
1394                         inl_pad = 0;
1395                         do_inline = length <= txq->inline_max_packet_sz &&
1396                                     sizeof(inl_hdr) + length <= mpw_room &&
1397                                     !txq->mpw_hdr_dseg;
1398                         mpw.wqe->eseg.cs_flags = cs_flags;
1399                 } else {
1400                         /* Evaluate whether the next packet can be inlined.
1401                          * Inlininig is possible when:
1402                          * - length is less than configured value
1403                          * - length fits for remaining space
1404                          * - not required to fill the title WQEBB with dsegs
1405                          */
1406                         do_inline =
1407                                 length <= txq->inline_max_packet_sz &&
1408                                 inl_pad + sizeof(inl_hdr) + length <=
1409                                  mpw_room &&
1410                                 (!txq->mpw_hdr_dseg ||
1411                                  mpw.total_len >= MLX5_WQE_SIZE);
1412                 }
1413                 if (do_inline) {
1414                         /* Inline packet into WQE. */
1415                         unsigned int max;
1416
1417                         assert(mpw.state == MLX5_MPW_ENHANCED_STATE_OPENED);
1418                         assert(length == DATA_LEN(buf));
1419                         inl_hdr = rte_cpu_to_be_32(length | MLX5_INLINE_SEG);
1420                         addr = rte_pktmbuf_mtod(buf, uintptr_t);
1421                         mpw.data.raw = (volatile void *)
1422                                 ((uintptr_t)mpw.data.raw + inl_pad);
1423                         max = tx_mlx5_wq_tailroom(txq,
1424                                         (void *)(uintptr_t)mpw.data.raw);
1425                         /* Copy inline header. */
1426                         mpw.data.raw = (volatile void *)
1427                                 mlx5_copy_to_wq(
1428                                           (void *)(uintptr_t)mpw.data.raw,
1429                                           &inl_hdr,
1430                                           sizeof(inl_hdr),
1431                                           (void *)(uintptr_t)txq->wqes,
1432                                           max);
1433                         max = tx_mlx5_wq_tailroom(txq,
1434                                         (void *)(uintptr_t)mpw.data.raw);
1435                         /* Copy packet data. */
1436                         mpw.data.raw = (volatile void *)
1437                                 mlx5_copy_to_wq(
1438                                           (void *)(uintptr_t)mpw.data.raw,
1439                                           (void *)addr,
1440                                           length,
1441                                           (void *)(uintptr_t)txq->wqes,
1442                                           max);
1443                         ++mpw.pkts_n;
1444                         mpw.total_len += (inl_pad + sizeof(inl_hdr) + length);
1445                         /* No need to get completion as the entire packet is
1446                          * copied to WQ. Free the buf right away.
1447                          */
1448                         rte_pktmbuf_free_seg(buf);
1449                         mpw_room -= (inl_pad + sizeof(inl_hdr) + length);
1450                         /* Add pad in the next packet if any. */
1451                         inl_pad = (((uintptr_t)mpw.data.raw +
1452                                         (MLX5_WQE_DWORD_SIZE - 1)) &
1453                                         ~(MLX5_WQE_DWORD_SIZE - 1)) -
1454                                   (uintptr_t)mpw.data.raw;
1455                 } else {
1456                         /* No inline. Load a dseg of packet pointer. */
1457                         volatile rte_v128u32_t *dseg;
1458
1459                         assert(mpw.state == MLX5_MPW_ENHANCED_STATE_OPENED);
1460                         assert((inl_pad + sizeof(*dseg)) <= mpw_room);
1461                         assert(length == DATA_LEN(buf));
1462                         if (!tx_mlx5_wq_tailroom(txq,
1463                                         (void *)((uintptr_t)mpw.data.raw
1464                                                 + inl_pad)))
1465                                 dseg = (volatile void *)txq->wqes;
1466                         else
1467                                 dseg = (volatile void *)
1468                                         ((uintptr_t)mpw.data.raw +
1469                                          inl_pad);
1470                         (*txq->elts)[elts_head++ & elts_m] = buf;
1471                         addr = rte_pktmbuf_mtod(buf, uintptr_t);
1472                         for (n = 0; n * RTE_CACHE_LINE_SIZE < length; n++)
1473                                 rte_prefetch2((void *)(addr +
1474                                                 n * RTE_CACHE_LINE_SIZE));
1475                         addr = rte_cpu_to_be_64(addr);
1476                         *dseg = (rte_v128u32_t) {
1477                                 rte_cpu_to_be_32(length),
1478                                 mlx5_tx_mb2mr(txq, buf),
1479                                 addr,
1480                                 addr >> 32,
1481                         };
1482                         mpw.data.raw = (volatile void *)(dseg + 1);
1483                         mpw.total_len += (inl_pad + sizeof(*dseg));
1484                         ++j;
1485                         ++mpw.pkts_n;
1486                         mpw_room -= (inl_pad + sizeof(*dseg));
1487                         inl_pad = 0;
1488                 }
1489 #ifdef MLX5_PMD_SOFT_COUNTERS
1490                 /* Increment sent bytes counter. */
1491                 txq->stats.obytes += length;
1492 #endif
1493                 ++i;
1494         } while (i < pkts_n);
1495         /* Take a shortcut if nothing must be sent. */
1496         if (unlikely(i == 0))
1497                 return 0;
1498         /* Check whether completion threshold has been reached. */
1499         if (txq->elts_comp + j >= MLX5_TX_COMP_THRESH ||
1500                         (uint16_t)(txq->wqe_ci - txq->mpw_comp) >=
1501                          (1 << txq->wqe_n) / MLX5_TX_COMP_THRESH_INLINE_DIV) {
1502                 volatile struct mlx5_wqe *wqe = mpw.wqe;
1503
1504                 /* Request completion on last WQE. */
1505                 wqe->ctrl[2] = rte_cpu_to_be_32(8);
1506                 /* Save elts_head in unused "immediate" field of WQE. */
1507                 wqe->ctrl[3] = elts_head;
1508                 txq->elts_comp = 0;
1509                 txq->mpw_comp = txq->wqe_ci;
1510 #ifndef NDEBUG
1511                 ++txq->cq_pi;
1512 #endif
1513         } else {
1514                 txq->elts_comp += j;
1515         }
1516 #ifdef MLX5_PMD_SOFT_COUNTERS
1517         /* Increment sent packets counter. */
1518         txq->stats.opackets += i;
1519 #endif
1520         if (mpw.state == MLX5_MPW_ENHANCED_STATE_OPENED)
1521                 mlx5_empw_close(txq, &mpw);
1522         /* Ring QP doorbell. */
1523         mlx5_tx_dbrec(txq, mpw.wqe);
1524         txq->elts_head = elts_head;
1525         return i;
1526 }
1527
1528 /**
1529  * DPDK callback for TX with Enhanced MPW support.
1530  *
1531  * @param dpdk_txq
1532  *   Generic pointer to TX queue structure.
1533  * @param[in] pkts
1534  *   Packets to transmit.
1535  * @param pkts_n
1536  *   Number of packets in array.
1537  *
1538  * @return
1539  *   Number of packets successfully transmitted (<= pkts_n).
1540  */
1541 uint16_t
1542 mlx5_tx_burst_empw(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
1543 {
1544         struct mlx5_txq_data *txq = (struct mlx5_txq_data *)dpdk_txq;
1545         uint16_t nb_tx = 0;
1546
1547         while (pkts_n > nb_tx) {
1548                 uint16_t n;
1549                 uint16_t ret;
1550
1551                 n = txq_count_contig_multi_seg(&pkts[nb_tx], pkts_n - nb_tx);
1552                 if (n) {
1553                         ret = mlx5_tx_burst(dpdk_txq, &pkts[nb_tx], n);
1554                         if (!ret)
1555                                 break;
1556                         nb_tx += ret;
1557                 }
1558                 n = txq_count_contig_single_seg(&pkts[nb_tx], pkts_n - nb_tx);
1559                 if (n) {
1560                         ret = txq_burst_empw(txq, &pkts[nb_tx], n);
1561                         if (!ret)
1562                                 break;
1563                         nb_tx += ret;
1564                 }
1565         }
1566         return nb_tx;
1567 }
1568
1569 /**
1570  * Translate RX completion flags to packet type.
1571  *
1572  * @param[in] cqe
1573  *   Pointer to CQE.
1574  *
1575  * @note: fix mlx5_dev_supported_ptypes_get() if any change here.
1576  *
1577  * @return
1578  *   Packet type for struct rte_mbuf.
1579  */
1580 static inline uint32_t
1581 rxq_cq_to_pkt_type(volatile struct mlx5_cqe *cqe)
1582 {
1583         uint8_t idx;
1584         uint8_t pinfo = cqe->pkt_info;
1585         uint16_t ptype = cqe->hdr_type_etc;
1586
1587         /*
1588          * The index to the array should have:
1589          * bit[1:0] = l3_hdr_type
1590          * bit[4:2] = l4_hdr_type
1591          * bit[5] = ip_frag
1592          * bit[6] = tunneled
1593          * bit[7] = outer_l3_type
1594          */
1595         idx = ((pinfo & 0x3) << 6) | ((ptype & 0xfc00) >> 10);
1596         return mlx5_ptype_table[idx];
1597 }
1598
1599 /**
1600  * Get size of the next packet for a given CQE. For compressed CQEs, the
1601  * consumer index is updated only once all packets of the current one have
1602  * been processed.
1603  *
1604  * @param rxq
1605  *   Pointer to RX queue.
1606  * @param cqe
1607  *   CQE to process.
1608  * @param[out] rss_hash
1609  *   Packet RSS Hash result.
1610  *
1611  * @return
1612  *   Packet size in bytes (0 if there is none), -1 in case of completion
1613  *   with error.
1614  */
1615 static inline int
1616 mlx5_rx_poll_len(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe,
1617                  uint16_t cqe_cnt, uint32_t *rss_hash)
1618 {
1619         struct rxq_zip *zip = &rxq->zip;
1620         uint16_t cqe_n = cqe_cnt + 1;
1621         int len = 0;
1622         uint16_t idx, end;
1623
1624         /* Process compressed data in the CQE and mini arrays. */
1625         if (zip->ai) {
1626                 volatile struct mlx5_mini_cqe8 (*mc)[8] =
1627                         (volatile struct mlx5_mini_cqe8 (*)[8])
1628                         (uintptr_t)(&(*rxq->cqes)[zip->ca & cqe_cnt].pkt_info);
1629
1630                 len = rte_be_to_cpu_32((*mc)[zip->ai & 7].byte_cnt);
1631                 *rss_hash = rte_be_to_cpu_32((*mc)[zip->ai & 7].rx_hash_result);
1632                 if ((++zip->ai & 7) == 0) {
1633                         /* Invalidate consumed CQEs */
1634                         idx = zip->ca;
1635                         end = zip->na;
1636                         while (idx != end) {
1637                                 (*rxq->cqes)[idx & cqe_cnt].op_own =
1638                                         MLX5_CQE_INVALIDATE;
1639                                 ++idx;
1640                         }
1641                         /*
1642                          * Increment consumer index to skip the number of
1643                          * CQEs consumed. Hardware leaves holes in the CQ
1644                          * ring for software use.
1645                          */
1646                         zip->ca = zip->na;
1647                         zip->na += 8;
1648                 }
1649                 if (unlikely(rxq->zip.ai == rxq->zip.cqe_cnt)) {
1650                         /* Invalidate the rest */
1651                         idx = zip->ca;
1652                         end = zip->cq_ci;
1653
1654                         while (idx != end) {
1655                                 (*rxq->cqes)[idx & cqe_cnt].op_own =
1656                                         MLX5_CQE_INVALIDATE;
1657                                 ++idx;
1658                         }
1659                         rxq->cq_ci = zip->cq_ci;
1660                         zip->ai = 0;
1661                 }
1662         /* No compressed data, get next CQE and verify if it is compressed. */
1663         } else {
1664                 int ret;
1665                 int8_t op_own;
1666
1667                 ret = check_cqe(cqe, cqe_n, rxq->cq_ci);
1668                 if (unlikely(ret == 1))
1669                         return 0;
1670                 ++rxq->cq_ci;
1671                 op_own = cqe->op_own;
1672                 rte_cio_rmb();
1673                 if (MLX5_CQE_FORMAT(op_own) == MLX5_COMPRESSED) {
1674                         volatile struct mlx5_mini_cqe8 (*mc)[8] =
1675                                 (volatile struct mlx5_mini_cqe8 (*)[8])
1676                                 (uintptr_t)(&(*rxq->cqes)[rxq->cq_ci &
1677                                                           cqe_cnt].pkt_info);
1678
1679                         /* Fix endianness. */
1680                         zip->cqe_cnt = rte_be_to_cpu_32(cqe->byte_cnt);
1681                         /*
1682                          * Current mini array position is the one returned by
1683                          * check_cqe64().
1684                          *
1685                          * If completion comprises several mini arrays, as a
1686                          * special case the second one is located 7 CQEs after
1687                          * the initial CQE instead of 8 for subsequent ones.
1688                          */
1689                         zip->ca = rxq->cq_ci;
1690                         zip->na = zip->ca + 7;
1691                         /* Compute the next non compressed CQE. */
1692                         --rxq->cq_ci;
1693                         zip->cq_ci = rxq->cq_ci + zip->cqe_cnt;
1694                         /* Get packet size to return. */
1695                         len = rte_be_to_cpu_32((*mc)[0].byte_cnt);
1696                         *rss_hash = rte_be_to_cpu_32((*mc)[0].rx_hash_result);
1697                         zip->ai = 1;
1698                         /* Prefetch all the entries to be invalidated */
1699                         idx = zip->ca;
1700                         end = zip->cq_ci;
1701                         while (idx != end) {
1702                                 rte_prefetch0(&(*rxq->cqes)[(idx) & cqe_cnt]);
1703                                 ++idx;
1704                         }
1705                 } else {
1706                         len = rte_be_to_cpu_32(cqe->byte_cnt);
1707                         *rss_hash = rte_be_to_cpu_32(cqe->rx_hash_res);
1708                 }
1709                 /* Error while receiving packet. */
1710                 if (unlikely(MLX5_CQE_OPCODE(op_own) == MLX5_CQE_RESP_ERR))
1711                         return -1;
1712         }
1713         return len;
1714 }
1715
1716 /**
1717  * Translate RX completion flags to offload flags.
1718  *
1719  * @param[in] rxq
1720  *   Pointer to RX queue structure.
1721  * @param[in] cqe
1722  *   Pointer to CQE.
1723  *
1724  * @return
1725  *   Offload flags (ol_flags) for struct rte_mbuf.
1726  */
1727 static inline uint32_t
1728 rxq_cq_to_ol_flags(struct mlx5_rxq_data *rxq, volatile struct mlx5_cqe *cqe)
1729 {
1730         uint32_t ol_flags = 0;
1731         uint16_t flags = rte_be_to_cpu_16(cqe->hdr_type_etc);
1732
1733         ol_flags =
1734                 TRANSPOSE(flags,
1735                           MLX5_CQE_RX_L3_HDR_VALID,
1736                           PKT_RX_IP_CKSUM_GOOD) |
1737                 TRANSPOSE(flags,
1738                           MLX5_CQE_RX_L4_HDR_VALID,
1739                           PKT_RX_L4_CKSUM_GOOD);
1740         if ((cqe->pkt_info & MLX5_CQE_RX_TUNNEL_PACKET) && (rxq->csum_l2tun))
1741                 ol_flags |=
1742                         TRANSPOSE(flags,
1743                                   MLX5_CQE_RX_L3_HDR_VALID,
1744                                   PKT_RX_IP_CKSUM_GOOD) |
1745                         TRANSPOSE(flags,
1746                                   MLX5_CQE_RX_L4_HDR_VALID,
1747                                   PKT_RX_L4_CKSUM_GOOD);
1748         return ol_flags;
1749 }
1750
1751 /**
1752  * DPDK callback for RX.
1753  *
1754  * @param dpdk_rxq
1755  *   Generic pointer to RX queue structure.
1756  * @param[out] pkts
1757  *   Array to store received packets.
1758  * @param pkts_n
1759  *   Maximum number of packets in array.
1760  *
1761  * @return
1762  *   Number of packets successfully received (<= pkts_n).
1763  */
1764 uint16_t
1765 mlx5_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
1766 {
1767         struct mlx5_rxq_data *rxq = dpdk_rxq;
1768         const unsigned int wqe_cnt = (1 << rxq->elts_n) - 1;
1769         const unsigned int cqe_cnt = (1 << rxq->cqe_n) - 1;
1770         const unsigned int sges_n = rxq->sges_n;
1771         struct rte_mbuf *pkt = NULL;
1772         struct rte_mbuf *seg = NULL;
1773         volatile struct mlx5_cqe *cqe =
1774                 &(*rxq->cqes)[rxq->cq_ci & cqe_cnt];
1775         unsigned int i = 0;
1776         unsigned int rq_ci = rxq->rq_ci << sges_n;
1777         int len = 0; /* keep its value across iterations. */
1778
1779         while (pkts_n) {
1780                 unsigned int idx = rq_ci & wqe_cnt;
1781                 volatile struct mlx5_wqe_data_seg *wqe = &(*rxq->wqes)[idx];
1782                 struct rte_mbuf *rep = (*rxq->elts)[idx];
1783                 uint32_t rss_hash_res = 0;
1784
1785                 if (pkt)
1786                         NEXT(seg) = rep;
1787                 seg = rep;
1788                 rte_prefetch0(seg);
1789                 rte_prefetch0(cqe);
1790                 rte_prefetch0(wqe);
1791                 rep = rte_mbuf_raw_alloc(rxq->mp);
1792                 if (unlikely(rep == NULL)) {
1793                         ++rxq->stats.rx_nombuf;
1794                         if (!pkt) {
1795                                 /*
1796                                  * no buffers before we even started,
1797                                  * bail out silently.
1798                                  */
1799                                 break;
1800                         }
1801                         while (pkt != seg) {
1802                                 assert(pkt != (*rxq->elts)[idx]);
1803                                 rep = NEXT(pkt);
1804                                 NEXT(pkt) = NULL;
1805                                 NB_SEGS(pkt) = 1;
1806                                 rte_mbuf_raw_free(pkt);
1807                                 pkt = rep;
1808                         }
1809                         break;
1810                 }
1811                 if (!pkt) {
1812                         cqe = &(*rxq->cqes)[rxq->cq_ci & cqe_cnt];
1813                         len = mlx5_rx_poll_len(rxq, cqe, cqe_cnt,
1814                                                &rss_hash_res);
1815                         if (!len) {
1816                                 rte_mbuf_raw_free(rep);
1817                                 break;
1818                         }
1819                         if (unlikely(len == -1)) {
1820                                 /* RX error, packet is likely too large. */
1821                                 rte_mbuf_raw_free(rep);
1822                                 ++rxq->stats.idropped;
1823                                 goto skip;
1824                         }
1825                         pkt = seg;
1826                         assert(len >= (rxq->crc_present << 2));
1827                         /* Update packet information. */
1828                         pkt->packet_type = rxq_cq_to_pkt_type(cqe);
1829                         pkt->ol_flags = 0;
1830                         if (rss_hash_res && rxq->rss_hash) {
1831                                 pkt->hash.rss = rss_hash_res;
1832                                 pkt->ol_flags = PKT_RX_RSS_HASH;
1833                         }
1834                         if (rxq->mark &&
1835                             MLX5_FLOW_MARK_IS_VALID(cqe->sop_drop_qpn)) {
1836                                 pkt->ol_flags |= PKT_RX_FDIR;
1837                                 if (cqe->sop_drop_qpn !=
1838                                     rte_cpu_to_be_32(MLX5_FLOW_MARK_DEFAULT)) {
1839                                         uint32_t mark = cqe->sop_drop_qpn;
1840
1841                                         pkt->ol_flags |= PKT_RX_FDIR_ID;
1842                                         pkt->hash.fdir.hi =
1843                                                 mlx5_flow_mark_get(mark);
1844                                 }
1845                         }
1846                         if (rxq->csum | rxq->csum_l2tun)
1847                                 pkt->ol_flags |= rxq_cq_to_ol_flags(rxq, cqe);
1848                         if (rxq->vlan_strip &&
1849                             (cqe->hdr_type_etc &
1850                              rte_cpu_to_be_16(MLX5_CQE_VLAN_STRIPPED))) {
1851                                 pkt->ol_flags |= PKT_RX_VLAN |
1852                                         PKT_RX_VLAN_STRIPPED;
1853                                 pkt->vlan_tci =
1854                                         rte_be_to_cpu_16(cqe->vlan_info);
1855                         }
1856                         if (rxq->hw_timestamp) {
1857                                 pkt->timestamp =
1858                                         rte_be_to_cpu_64(cqe->timestamp);
1859                                 pkt->ol_flags |= PKT_RX_TIMESTAMP;
1860                         }
1861                         if (rxq->crc_present)
1862                                 len -= ETHER_CRC_LEN;
1863                         PKT_LEN(pkt) = len;
1864                 }
1865                 DATA_LEN(rep) = DATA_LEN(seg);
1866                 PKT_LEN(rep) = PKT_LEN(seg);
1867                 SET_DATA_OFF(rep, DATA_OFF(seg));
1868                 PORT(rep) = PORT(seg);
1869                 (*rxq->elts)[idx] = rep;
1870                 /*
1871                  * Fill NIC descriptor with the new buffer.  The lkey and size
1872                  * of the buffers are already known, only the buffer address
1873                  * changes.
1874                  */
1875                 wqe->addr = rte_cpu_to_be_64(rte_pktmbuf_mtod(rep, uintptr_t));
1876                 if (len > DATA_LEN(seg)) {
1877                         len -= DATA_LEN(seg);
1878                         ++NB_SEGS(pkt);
1879                         ++rq_ci;
1880                         continue;
1881                 }
1882                 DATA_LEN(seg) = len;
1883 #ifdef MLX5_PMD_SOFT_COUNTERS
1884                 /* Increment bytes counter. */
1885                 rxq->stats.ibytes += PKT_LEN(pkt);
1886 #endif
1887                 /* Return packet. */
1888                 *(pkts++) = pkt;
1889                 pkt = NULL;
1890                 --pkts_n;
1891                 ++i;
1892 skip:
1893                 /* Align consumer index to the next stride. */
1894                 rq_ci >>= sges_n;
1895                 ++rq_ci;
1896                 rq_ci <<= sges_n;
1897         }
1898         if (unlikely((i == 0) && ((rq_ci >> sges_n) == rxq->rq_ci)))
1899                 return 0;
1900         /* Update the consumer index. */
1901         rxq->rq_ci = rq_ci >> sges_n;
1902         rte_cio_wmb();
1903         *rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
1904         rte_cio_wmb();
1905         *rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
1906 #ifdef MLX5_PMD_SOFT_COUNTERS
1907         /* Increment packets counter. */
1908         rxq->stats.ipackets += i;
1909 #endif
1910         return i;
1911 }
1912
1913 /**
1914  * Dummy DPDK callback for TX.
1915  *
1916  * This function is used to temporarily replace the real callback during
1917  * unsafe control operations on the queue, or in case of error.
1918  *
1919  * @param dpdk_txq
1920  *   Generic pointer to TX queue structure.
1921  * @param[in] pkts
1922  *   Packets to transmit.
1923  * @param pkts_n
1924  *   Number of packets in array.
1925  *
1926  * @return
1927  *   Number of packets successfully transmitted (<= pkts_n).
1928  */
1929 uint16_t
1930 removed_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
1931 {
1932         (void)dpdk_txq;
1933         (void)pkts;
1934         (void)pkts_n;
1935         return 0;
1936 }
1937
1938 /**
1939  * Dummy DPDK callback for RX.
1940  *
1941  * This function is used to temporarily replace the real callback during
1942  * unsafe control operations on the queue, or in case of error.
1943  *
1944  * @param dpdk_rxq
1945  *   Generic pointer to RX queue structure.
1946  * @param[out] pkts
1947  *   Array to store received packets.
1948  * @param pkts_n
1949  *   Maximum number of packets in array.
1950  *
1951  * @return
1952  *   Number of packets successfully received (<= pkts_n).
1953  */
1954 uint16_t
1955 removed_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
1956 {
1957         (void)dpdk_rxq;
1958         (void)pkts;
1959         (void)pkts_n;
1960         return 0;
1961 }
1962
1963 /*
1964  * Vectorized Rx/Tx routines are not compiled in when required vector
1965  * instructions are not supported on a target architecture. The following null
1966  * stubs are needed for linkage when those are not included outside of this file
1967  * (e.g.  mlx5_rxtx_vec_sse.c for x86).
1968  */
1969
1970 uint16_t __attribute__((weak))
1971 mlx5_tx_burst_raw_vec(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
1972 {
1973         (void)dpdk_txq;
1974         (void)pkts;
1975         (void)pkts_n;
1976         return 0;
1977 }
1978
1979 uint16_t __attribute__((weak))
1980 mlx5_tx_burst_vec(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
1981 {
1982         (void)dpdk_txq;
1983         (void)pkts;
1984         (void)pkts_n;
1985         return 0;
1986 }
1987
1988 uint16_t __attribute__((weak))
1989 mlx5_rx_burst_vec(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
1990 {
1991         (void)dpdk_rxq;
1992         (void)pkts;
1993         (void)pkts_n;
1994         return 0;
1995 }
1996
1997 int __attribute__((weak))
1998 priv_check_raw_vec_tx_support(struct priv *priv, struct rte_eth_dev *dev)
1999 {
2000         (void)priv;
2001         (void)dev;
2002         return -ENOTSUP;
2003 }
2004
2005 int __attribute__((weak))
2006 priv_check_vec_tx_support(struct priv *priv, struct rte_eth_dev *dev)
2007 {
2008         (void)priv;
2009         (void)dev;
2010         return -ENOTSUP;
2011 }
2012
2013 int __attribute__((weak))
2014 rxq_check_vec_support(struct mlx5_rxq_data *rxq)
2015 {
2016         (void)rxq;
2017         return -ENOTSUP;
2018 }
2019
2020 int __attribute__((weak))
2021 priv_check_vec_rx_support(struct priv *priv)
2022 {
2023         (void)priv;
2024         return -ENOTSUP;
2025 }