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 rxq *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 rxq *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         /* L3 */
 101         (*p)[0x01] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 102                      RTE_PTYPE_L4_NONFRAG;
 103         (*p)[0x02] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 104                      RTE_PTYPE_L4_NONFRAG;
 105         /* Fragmented */
 106         (*p)[0x21] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 107                      RTE_PTYPE_L4_FRAG;
 108         (*p)[0x22] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 109                      RTE_PTYPE_L4_FRAG;
 110         /* TCP */
 111         (*p)[0x05] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 112                      RTE_PTYPE_L4_TCP;
 113         (*p)[0x06] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 114                      RTE_PTYPE_L4_TCP;
 115         /* UDP */
 116         (*p)[0x09] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 117                      RTE_PTYPE_L4_UDP;
 118         (*p)[0x0a] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 119                      RTE_PTYPE_L4_UDP;
 120         /* Repeat with outer_l3_type being set. Just in case. */
 121         (*p)[0x81] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 122                      RTE_PTYPE_L4_NONFRAG;
 123         (*p)[0x82] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 124                      RTE_PTYPE_L4_NONFRAG;
 125         (*p)[0xa1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 126                      RTE_PTYPE_L4_FRAG;
 127         (*p)[0xa2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 128                      RTE_PTYPE_L4_FRAG;
 129         (*p)[0x85] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 130                      RTE_PTYPE_L4_TCP;
 131         (*p)[0x86] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 132                      RTE_PTYPE_L4_TCP;
 133         (*p)[0x89] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 134                      RTE_PTYPE_L4_UDP;
 135         (*p)[0x8a] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 136                      RTE_PTYPE_L4_UDP;
 137         /* Tunneled - L3 */
 138         (*p)[0x41] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 139                      RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
 140                      RTE_PTYPE_INNER_L4_NONFRAG;
 141         (*p)[0x42] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 142                      RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 143                      RTE_PTYPE_INNER_L4_NONFRAG;
 144         (*p)[0xc1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 145                      RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
 146                      RTE_PTYPE_INNER_L4_NONFRAG;
 147         (*p)[0xc2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 148                      RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 149                      RTE_PTYPE_INNER_L4_NONFRAG;
 150         /* Tunneled - Fragmented */
 151         (*p)[0x61] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 152                      RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
 153                      RTE_PTYPE_INNER_L4_FRAG;
 154         (*p)[0x62] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 155                      RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 156                      RTE_PTYPE_INNER_L4_FRAG;
 157         (*p)[0xe1] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 158                      RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
 159                      RTE_PTYPE_INNER_L4_FRAG;
 160         (*p)[0xe2] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 161                      RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 162                      RTE_PTYPE_INNER_L4_FRAG;
 163         /* Tunneled - TCP */
 164         (*p)[0x45] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 165                      RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
 166                      RTE_PTYPE_L4_TCP;
 167         (*p)[0x46] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 168                      RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 169                      RTE_PTYPE_L4_TCP;
 170         (*p)[0xc5] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 171                      RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
 172                      RTE_PTYPE_L4_TCP;
 173         (*p)[0xc6] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 174                      RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 175                      RTE_PTYPE_L4_TCP;
 176         /* Tunneled - UDP */
 177         (*p)[0x49] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 178                      RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
 179                      RTE_PTYPE_L4_UDP;
 180         (*p)[0x4a] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV4_EXT_UNKNOWN |
 181                      RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 182                      RTE_PTYPE_L4_UDP;
 183         (*p)[0xc9] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 184                      RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN |
 185                      RTE_PTYPE_L4_UDP;
 186         (*p)[0xca] = RTE_PTYPE_L2_ETHER | RTE_PTYPE_L3_IPV6_EXT_UNKNOWN |
 187                      RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN |
 188                      RTE_PTYPE_L4_UDP;
 189 }
 190
 191 /**
 192  * Return the size of tailroom of WQ.
 193  *
 194  * @param txq
 195  *   Pointer to TX queue structure.
 196  * @param addr
 197  *   Pointer to tail of WQ.
 198  *
 199  * @return
 200  *   Size of tailroom.
 201  */
 202 static inline size_t
 203 tx_mlx5_wq_tailroom(struct txq *txq, void *addr)
 204 {
 205         size_t tailroom;
 206         tailroom = (uintptr_t)(txq->wqes) +
 207                    (1 << txq->wqe_n) * MLX5_WQE_SIZE -
 208                    (uintptr_t)addr;
 209         return tailroom;
 210 }
 211
 212 /**
 213  * Copy data to tailroom of circular queue.
 214  *
 215  * @param dst
 216  *   Pointer to destination.
 217  * @param src
 218  *   Pointer to source.
 219  * @param n
 220  *   Number of bytes to copy.
 221  * @param base
 222  *   Pointer to head of queue.
 223  * @param tailroom
 224  *   Size of tailroom from dst.
 225  *
 226  * @return
 227  *   Pointer after copied data.
 228  */
 229 static inline void *
 230 mlx5_copy_to_wq(void *dst, const void *src, size_t n,
 231                 void *base, size_t tailroom)
 232 {
 233         void *ret;
 234
 235         if (n > tailroom) {
 236                 rte_memcpy(dst, src, tailroom);
 237                 rte_memcpy(base, (void *)((uintptr_t)src + tailroom),
 238                            n - tailroom);
 239                 ret = (uint8_t *)base + n - tailroom;
 240         } else {
 241                 rte_memcpy(dst, src, n);
 242                 ret = (n == tailroom) ? base : (uint8_t *)dst + n;
 243         }
 244         return ret;
 245 }
 246
 247 /**
 248  * DPDK callback to check the status of a tx descriptor.
 249  *
 250  * @param tx_queue
 251  *   The tx queue.
 252  * @param[in] offset
 253  *   The index of the descriptor in the ring.
 254  *
 255  * @return
 256  *   The status of the tx descriptor.
 257  */
 258 int
 259 mlx5_tx_descriptor_status(void *tx_queue, uint16_t offset)
 260 {
 261         struct txq *txq = tx_queue;
 262         uint16_t used;
 263
 264         mlx5_tx_complete(txq);
 265         used = txq->elts_head - txq->elts_tail;
 266         if (offset < used)
 267                 return RTE_ETH_TX_DESC_FULL;
 268         return RTE_ETH_TX_DESC_DONE;
 269 }
 270
 271 /**
 272  * DPDK callback to check the status of a rx descriptor.
 273  *
 274  * @param rx_queue
 275  *   The rx queue.
 276  * @param[in] offset
 277  *   The index of the descriptor in the ring.
 278  *
 279  * @return
 280  *   The status of the tx descriptor.
 281  */
 282 int
 283 mlx5_rx_descriptor_status(void *rx_queue, uint16_t offset)
 284 {
 285         struct rxq *rxq = rx_queue;
 286         struct rxq_zip *zip = &rxq->zip;
 287         volatile struct mlx5_cqe *cqe;
 288         const unsigned int cqe_n = (1 << rxq->cqe_n);
 289         const unsigned int cqe_cnt = cqe_n - 1;
 290         unsigned int cq_ci;
 291         unsigned int used;
 292
 293         /* if we are processing a compressed cqe */
 294         if (zip->ai) {
 295                 used = zip->cqe_cnt - zip->ca;
 296                 cq_ci = zip->cq_ci;
 297         } else {
 298                 used = 0;
 299                 cq_ci = rxq->cq_ci;
 300         }
 301         cqe = &(*rxq->cqes)[cq_ci & cqe_cnt];
 302         while (check_cqe(cqe, cqe_n, cq_ci) == 0) {
 303                 int8_t op_own;
 304                 unsigned int n;
 305
 306                 op_own = cqe->op_own;
 307                 if (MLX5_CQE_FORMAT(op_own) == MLX5_COMPRESSED)
 308                         n = rte_be_to_cpu_32(cqe->byte_cnt);
 309                 else
 310                         n = 1;
 311                 cq_ci += n;
 312                 used += n;
 313                 cqe = &(*rxq->cqes)[cq_ci & cqe_cnt];
 314         }
 315         used = RTE_MIN(used, (1U << rxq->elts_n) - 1);
 316         if (offset < used)
 317                 return RTE_ETH_RX_DESC_DONE;
 318         return RTE_ETH_RX_DESC_AVAIL;
 319 }
 320
 321 /**
 322  * DPDK callback for TX.
 323  *
 324  * @param dpdk_txq
 325  *   Generic pointer to TX queue structure.
 326  * @param[in] pkts
 327  *   Packets to transmit.
 328  * @param pkts_n
 329  *   Number of packets in array.
 330  *
 331  * @return
 332  *   Number of packets successfully transmitted (<= pkts_n).
 333  */
 334 uint16_t
 335 mlx5_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
 336 {
 337         struct txq *txq = (struct txq *)dpdk_txq;
 338         uint16_t elts_head = txq->elts_head;
 339         const uint16_t elts_n = 1 << txq->elts_n;
 340         const uint16_t elts_m = elts_n - 1;
 341         unsigned int i = 0;
 342         unsigned int j = 0;
 343         unsigned int k = 0;
 344         uint16_t max_elts;
 345         unsigned int max_inline = txq->max_inline;
 346         const unsigned int inline_en = !!max_inline && txq->inline_en;
 347         uint16_t max_wqe;
 348         unsigned int comp;
 349         volatile struct mlx5_wqe_v *wqe = NULL;
 350         volatile struct mlx5_wqe_ctrl *last_wqe = NULL;
 351         unsigned int segs_n = 0;
 352         struct rte_mbuf *buf = NULL;
 353         uint8_t *raw;
 354
 355         if (unlikely(!pkts_n))
 356                 return 0;
 357         /* Prefetch first packet cacheline. */
 358         rte_prefetch0(*pkts);
 359         /* Start processing. */
 360         mlx5_tx_complete(txq);
 361         max_elts = (elts_n - (elts_head - txq->elts_tail));
 362         max_wqe = (1u << txq->wqe_n) - (txq->wqe_ci - txq->wqe_pi);
 363         if (unlikely(!max_wqe))
 364                 return 0;
 365         do {
 366                 volatile rte_v128u32_t *dseg = NULL;
 367                 uint32_t length;
 368                 unsigned int ds = 0;
 369                 unsigned int sg = 0; /* counter of additional segs attached. */
 370                 uintptr_t addr;
 371                 uint64_t naddr;
 372                 uint16_t pkt_inline_sz = MLX5_WQE_DWORD_SIZE + 2;
 373                 uint16_t tso_header_sz = 0;
 374                 uint16_t ehdr;
 375                 uint8_t cs_flags = 0;
 376                 uint64_t tso = 0;
 377                 uint16_t tso_segsz = 0;
 378 #ifdef MLX5_PMD_SOFT_COUNTERS
 379                 uint32_t total_length = 0;
 380 #endif
 381
 382                 /* first_seg */
 383                 buf = *pkts;
 384                 segs_n = buf->nb_segs;
 385                 /*
 386                  * Make sure there is enough room to store this packet and
 387                  * that one ring entry remains unused.
 388                  */
 389                 assert(segs_n);
 390                 if (max_elts < segs_n)
 391                         break;
 392                 max_elts -= segs_n;
 393                 --segs_n;
 394                 if (unlikely(--max_wqe == 0))
 395                         break;
 396                 wqe = (volatile struct mlx5_wqe_v *)
 397                         tx_mlx5_wqe(txq, txq->wqe_ci);
 398                 rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci + 1));
 399                 if (pkts_n - i > 1)
 400                         rte_prefetch0(*(pkts + 1));
 401                 addr = rte_pktmbuf_mtod(buf, uintptr_t);
 402                 length = DATA_LEN(buf);
 403                 ehdr = (((uint8_t *)addr)[1] << 8) |
 404                        ((uint8_t *)addr)[0];
 405 #ifdef MLX5_PMD_SOFT_COUNTERS
 406                 total_length = length;
 407 #endif
 408                 if (length < (MLX5_WQE_DWORD_SIZE + 2)) {
 409                         txq->stats.oerrors++;
 410                         break;
 411                 }
 412                 /* Update element. */
 413                 (*txq->elts)[elts_head & elts_m] = buf;
 414                 /* Prefetch next buffer data. */
 415                 if (pkts_n - i > 1)
 416                         rte_prefetch0(
 417                             rte_pktmbuf_mtod(*(pkts + 1), volatile void *));
 418                 /* Should we enable HW CKSUM offload */
 419                 if (buf->ol_flags &
 420                     (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM)) {
 421                         const uint64_t is_tunneled = buf->ol_flags &
 422                                                      (PKT_TX_TUNNEL_GRE |
 423                                                       PKT_TX_TUNNEL_VXLAN);
 424
 425                         if (is_tunneled && txq->tunnel_en) {
 426                                 cs_flags = MLX5_ETH_WQE_L3_INNER_CSUM |
 427                                            MLX5_ETH_WQE_L4_INNER_CSUM;
 428                                 if (buf->ol_flags & PKT_TX_OUTER_IP_CKSUM)
 429                                         cs_flags |= MLX5_ETH_WQE_L3_CSUM;
 430                         } else {
 431                                 cs_flags = MLX5_ETH_WQE_L3_CSUM |
 432                                            MLX5_ETH_WQE_L4_CSUM;
 433                         }
 434                 }
 435                 raw = ((uint8_t *)(uintptr_t)wqe) + 2 * MLX5_WQE_DWORD_SIZE;
 436                 /* Replace the Ethernet type by the VLAN if necessary. */
 437                 if (buf->ol_flags & PKT_TX_VLAN_PKT) {
 438                         uint32_t vlan = rte_cpu_to_be_32(0x81000000 |
 439                                                          buf->vlan_tci);
 440                         unsigned int len = 2 * ETHER_ADDR_LEN - 2;
 441
 442                         addr += 2;
 443                         length -= 2;
 444                         /* Copy Destination and source mac address. */
 445                         memcpy((uint8_t *)raw, ((uint8_t *)addr), len);
 446                         /* Copy VLAN. */
 447                         memcpy((uint8_t *)raw + len, &vlan, sizeof(vlan));
 448                         /* Copy missing two bytes to end the DSeg. */
 449                         memcpy((uint8_t *)raw + len + sizeof(vlan),
 450                                ((uint8_t *)addr) + len, 2);
 451                         addr += len + 2;
 452                         length -= (len + 2);
 453                 } else {
 454                         memcpy((uint8_t *)raw, ((uint8_t *)addr) + 2,
 455                                MLX5_WQE_DWORD_SIZE);
 456                         length -= pkt_inline_sz;
 457                         addr += pkt_inline_sz;
 458                 }
 459                 raw += MLX5_WQE_DWORD_SIZE;
 460                 if (txq->tso_en) {
 461                         tso = buf->ol_flags & PKT_TX_TCP_SEG;
 462                         if (tso) {
 463                                 uintptr_t end = (uintptr_t)
 464                                                 (((uintptr_t)txq->wqes) +
 465                                                 (1 << txq->wqe_n) *
 466                                                 MLX5_WQE_SIZE);
 467                                 unsigned int copy_b;
 468                                 uint8_t vlan_sz = (buf->ol_flags &
 469                                                   PKT_TX_VLAN_PKT) ? 4 : 0;
 470                                 const uint64_t is_tunneled =
 471                                                         buf->ol_flags &
 472                                                         (PKT_TX_TUNNEL_GRE |
 473                                                          PKT_TX_TUNNEL_VXLAN);
 474
 475                                 tso_header_sz = buf->l2_len + vlan_sz +
 476                                                 buf->l3_len + buf->l4_len;
 477                                 tso_segsz = buf->tso_segsz;
 478                                 if (unlikely(tso_segsz == 0)) {
 479                                         txq->stats.oerrors++;
 480                                         break;
 481                                 }
 482                                 if (is_tunneled && txq->tunnel_en) {
 483                                         tso_header_sz += buf->outer_l2_len +
 484                                                          buf->outer_l3_len;
 485                                         cs_flags |= MLX5_ETH_WQE_L4_INNER_CSUM;
 486                                 } else {
 487                                         cs_flags |= MLX5_ETH_WQE_L4_CSUM;
 488                                 }
 489                                 if (unlikely(tso_header_sz >
 490                                              MLX5_MAX_TSO_HEADER)) {
 491                                         txq->stats.oerrors++;
 492                                         break;
 493                                 }
 494                                 copy_b = tso_header_sz - pkt_inline_sz;
 495                                 /* First seg must contain all headers. */
 496                                 assert(copy_b <= length);
 497                                 if (copy_b &&
 498                                    ((end - (uintptr_t)raw) > copy_b)) {
 499                                         uint16_t n = (MLX5_WQE_DS(copy_b) -
 500                                                       1 + 3) / 4;
 501
 502                                         if (unlikely(max_wqe < n))
 503                                                 break;
 504                                         max_wqe -= n;
 505                                         rte_memcpy((void *)raw,
 506                                                    (void *)addr, copy_b);
 507                                         addr += copy_b;
 508                                         length -= copy_b;
 509                                         /* Include padding for TSO header. */
 510                                         copy_b = MLX5_WQE_DS(copy_b) *
 511                                                  MLX5_WQE_DWORD_SIZE;
 512                                         pkt_inline_sz += copy_b;
 513                                         raw += copy_b;
 514                                 } else {
 515                                         /* NOP WQE. */
 516                                         wqe->ctrl = (rte_v128u32_t){
 517                                                      rte_cpu_to_be_32(
 518                                                         txq->wqe_ci << 8),
 519                                                      rte_cpu_to_be_32(
 520                                                         txq->qp_num_8s | 1),
 521                                                      0,
 522                                                      0,
 523                                         };
 524                                         ds = 1;
 525                                         total_length = 0;
 526                                         k++;
 527                                         goto next_wqe;
 528                                 }
 529                         }
 530                 }
 531                 /* Inline if enough room. */
 532                 if (inline_en || tso) {
 533                         uint32_t inl;
 534                         uintptr_t end = (uintptr_t)
 535                                 (((uintptr_t)txq->wqes) +
 536                                  (1 << txq->wqe_n) * MLX5_WQE_SIZE);
 537                         unsigned int inline_room = max_inline *
 538                                                    RTE_CACHE_LINE_SIZE -
 539                                                    (pkt_inline_sz - 2) -
 540                                                    !!tso * sizeof(inl);
 541                         uintptr_t addr_end = (addr + inline_room) &
 542                                              ~(RTE_CACHE_LINE_SIZE - 1);
 543                         unsigned int copy_b = (addr_end > addr) ?
 544                                 RTE_MIN((addr_end - addr), length) :
 545                                 0;
 546
 547                         if (copy_b && ((end - (uintptr_t)raw) > copy_b)) {
 548                                 /*
 549                                  * One Dseg remains in the current WQE.  To
 550                                  * keep the computation positive, it is
 551                                  * removed after the bytes to Dseg conversion.
 552                                  */
 553                                 uint16_t n = (MLX5_WQE_DS(copy_b) - 1 + 3) / 4;
 554
 555                                 if (unlikely(max_wqe < n))
 556                                         break;
 557                                 max_wqe -= n;
 558                                 if (tso) {
 559                                         uint32_t inl =
 560                                         rte_cpu_to_be_32(copy_b |
 561                                                          MLX5_INLINE_SEG);
 562
 563                                         pkt_inline_sz =
 564                                                 MLX5_WQE_DS(tso_header_sz) *
 565                                                 MLX5_WQE_DWORD_SIZE;
 566
 567                                         rte_memcpy((void *)raw,
 568                                                    (void *)&inl, sizeof(inl));
 569                                         raw += sizeof(inl);
 570                                         pkt_inline_sz += sizeof(inl);
 571                                 }
 572                                 rte_memcpy((void *)raw, (void *)addr, copy_b);
 573                                 addr += copy_b;
 574                                 length -= copy_b;
 575                                 pkt_inline_sz += copy_b;
 576                         }
 577                         /*
 578                          * 2 DWORDs consumed by the WQE header + ETH segment +
 579                          * the size of the inline part of the packet.
 580                          */
 581                         ds = 2 + MLX5_WQE_DS(pkt_inline_sz - 2);
 582                         if (length > 0) {
 583                                 if (ds % (MLX5_WQE_SIZE /
 584                                           MLX5_WQE_DWORD_SIZE) == 0) {
 585                                         if (unlikely(--max_wqe == 0))
 586                                                 break;
 587                                         dseg = (volatile rte_v128u32_t *)
 588                                                tx_mlx5_wqe(txq, txq->wqe_ci +
 589                                                            ds / 4);
 590                                 } else {
 591                                         dseg = (volatile rte_v128u32_t *)
 592                                                 ((uintptr_t)wqe +
 593                                                  (ds * MLX5_WQE_DWORD_SIZE));
 594                                 }
 595                                 goto use_dseg;
 596                         } else if (!segs_n) {
 597                                 goto next_pkt;
 598                         } else {
 599                                 /* dseg will be advance as part of next_seg */
 600                                 dseg = (volatile rte_v128u32_t *)
 601                                         ((uintptr_t)wqe +
 602                                          ((ds - 1) * MLX5_WQE_DWORD_SIZE));
 603                                 goto next_seg;
 604                         }
 605                 } else {
 606                         /*
 607                          * No inline has been done in the packet, only the
 608                          * Ethernet Header as been stored.
 609                          */
 610                         dseg = (volatile rte_v128u32_t *)
 611                                 ((uintptr_t)wqe + (3 * MLX5_WQE_DWORD_SIZE));
 612                         ds = 3;
 613 use_dseg:
 614                         /* Add the remaining packet as a simple ds. */
 615                         naddr = rte_cpu_to_be_64(addr);
 616                         *dseg = (rte_v128u32_t){
 617                                 rte_cpu_to_be_32(length),
 618                                 mlx5_tx_mb2mr(txq, buf),
 619                                 naddr,
 620                                 naddr >> 32,
 621                         };
 622                         ++ds;
 623                         if (!segs_n)
 624                                 goto next_pkt;
 625                 }
 626 next_seg:
 627                 assert(buf);
 628                 assert(ds);
 629                 assert(wqe);
 630                 /*
 631                  * Spill on next WQE when the current one does not have
 632                  * enough room left. Size of WQE must a be a multiple
 633                  * of data segment size.
 634                  */
 635                 assert(!(MLX5_WQE_SIZE % MLX5_WQE_DWORD_SIZE));
 636                 if (!(ds % (MLX5_WQE_SIZE / MLX5_WQE_DWORD_SIZE))) {
 637                         if (unlikely(--max_wqe == 0))
 638                                 break;
 639                         dseg = (volatile rte_v128u32_t *)
 640                                tx_mlx5_wqe(txq, txq->wqe_ci + ds / 4);
 641                         rte_prefetch0(tx_mlx5_wqe(txq,
 642                                                   txq->wqe_ci + ds / 4 + 1));
 643                 } else {
 644                         ++dseg;
 645                 }
 646                 ++ds;
 647                 buf = buf->next;
 648                 assert(buf);
 649                 length = DATA_LEN(buf);
 650 #ifdef MLX5_PMD_SOFT_COUNTERS
 651                 total_length += length;
 652 #endif
 653                 /* Store segment information. */
 654                 naddr = rte_cpu_to_be_64(rte_pktmbuf_mtod(buf, uintptr_t));
 655                 *dseg = (rte_v128u32_t){
 656                         rte_cpu_to_be_32(length),
 657                         mlx5_tx_mb2mr(txq, buf),
 658                         naddr,
 659                         naddr >> 32,
 660                 };
 661                 (*txq->elts)[++elts_head & elts_m] = buf;
 662                 ++sg;
 663                 /* Advance counter only if all segs are successfully posted. */
 664                 if (sg < segs_n)
 665                         goto next_seg;
 666                 else
 667                         j += sg;
 668 next_pkt:
 669                 if (ds > MLX5_DSEG_MAX) {
 670                         txq->stats.oerrors++;
 671                         break;
 672                 }
 673                 ++elts_head;
 674                 ++pkts;
 675                 ++i;
 676                 /* Initialize known and common part of the WQE structure. */
 677                 if (tso) {
 678                         wqe->ctrl = (rte_v128u32_t){
 679                                 rte_cpu_to_be_32((txq->wqe_ci << 8) |
 680                                                  MLX5_OPCODE_TSO),
 681                                 rte_cpu_to_be_32(txq->qp_num_8s | ds),
 682                                 0,
 683                                 0,
 684                         };
 685                         wqe->eseg = (rte_v128u32_t){
 686                                 0,
 687                                 cs_flags | (rte_cpu_to_be_16(tso_segsz) << 16),
 688                                 0,
 689                                 (ehdr << 16) | rte_cpu_to_be_16(tso_header_sz),
 690                         };
 691                 } else {
 692                         wqe->ctrl = (rte_v128u32_t){
 693                                 rte_cpu_to_be_32((txq->wqe_ci << 8) |
 694                                                  MLX5_OPCODE_SEND),
 695                                 rte_cpu_to_be_32(txq->qp_num_8s | ds),
 696                                 0,
 697                                 0,
 698                         };
 699                         wqe->eseg = (rte_v128u32_t){
 700                                 0,
 701                                 cs_flags,
 702                                 0,
 703                                 (ehdr << 16) | rte_cpu_to_be_16(pkt_inline_sz),
 704                         };
 705                 }
 706 next_wqe:
 707                 txq->wqe_ci += (ds + 3) / 4;
 708                 /* Save the last successful WQE for completion request */
 709                 last_wqe = (volatile struct mlx5_wqe_ctrl *)wqe;
 710 #ifdef MLX5_PMD_SOFT_COUNTERS
 711                 /* Increment sent bytes counter. */
 712                 txq->stats.obytes += total_length;
 713 #endif
 714         } while (i < pkts_n);
 715         /* Take a shortcut if nothing must be sent. */
 716         if (unlikely((i + k) == 0))
 717                 return 0;
 718         txq->elts_head += (i + j);
 719         /* Check whether completion threshold has been reached. */
 720         comp = txq->elts_comp + i + j + k;
 721         if (comp >= MLX5_TX_COMP_THRESH) {
 722                 /* Request completion on last WQE. */
 723                 last_wqe->ctrl2 = rte_cpu_to_be_32(8);
 724                 /* Save elts_head in unused "immediate" field of WQE. */
 725                 last_wqe->ctrl3 = txq->elts_head;
 726                 txq->elts_comp = 0;
 727         } else {
 728                 txq->elts_comp = comp;
 729         }
 730 #ifdef MLX5_PMD_SOFT_COUNTERS
 731         /* Increment sent packets counter. */
 732         txq->stats.opackets += i;
 733 #endif
 734         /* Ring QP doorbell. */
 735         mlx5_tx_dbrec(txq, (volatile struct mlx5_wqe *)last_wqe);
 736         return i;
 737 }
 738
 739 /**
 740  * Open a MPW session.
 741  *
 742  * @param txq
 743  *   Pointer to TX queue structure.
 744  * @param mpw
 745  *   Pointer to MPW session structure.
 746  * @param length
 747  *   Packet length.
 748  */
 749 static inline void
 750 mlx5_mpw_new(struct txq *txq, struct mlx5_mpw *mpw, uint32_t length)
 751 {
 752         uint16_t idx = txq->wqe_ci & ((1 << txq->wqe_n) - 1);
 753         volatile struct mlx5_wqe_data_seg (*dseg)[MLX5_MPW_DSEG_MAX] =
 754                 (volatile struct mlx5_wqe_data_seg (*)[])
 755                 tx_mlx5_wqe(txq, idx + 1);
 756
 757         mpw->state = MLX5_MPW_STATE_OPENED;
 758         mpw->pkts_n = 0;
 759         mpw->len = length;
 760         mpw->total_len = 0;
 761         mpw->wqe = (volatile struct mlx5_wqe *)tx_mlx5_wqe(txq, idx);
 762         mpw->wqe->eseg.mss = rte_cpu_to_be_16(length);
 763         mpw->wqe->eseg.inline_hdr_sz = 0;
 764         mpw->wqe->eseg.rsvd0 = 0;
 765         mpw->wqe->eseg.rsvd1 = 0;
 766         mpw->wqe->eseg.rsvd2 = 0;
 767         mpw->wqe->ctrl[0] = rte_cpu_to_be_32((MLX5_OPC_MOD_MPW << 24) |
 768                                              (txq->wqe_ci << 8) |
 769                                              MLX5_OPCODE_TSO);
 770         mpw->wqe->ctrl[2] = 0;
 771         mpw->wqe->ctrl[3] = 0;
 772         mpw->data.dseg[0] = (volatile struct mlx5_wqe_data_seg *)
 773                 (((uintptr_t)mpw->wqe) + (2 * MLX5_WQE_DWORD_SIZE));
 774         mpw->data.dseg[1] = (volatile struct mlx5_wqe_data_seg *)
 775                 (((uintptr_t)mpw->wqe) + (3 * MLX5_WQE_DWORD_SIZE));
 776         mpw->data.dseg[2] = &(*dseg)[0];
 777         mpw->data.dseg[3] = &(*dseg)[1];
 778         mpw->data.dseg[4] = &(*dseg)[2];
 779 }
 780
 781 /**
 782  * Close a MPW session.
 783  *
 784  * @param txq
 785  *   Pointer to TX queue structure.
 786  * @param mpw
 787  *   Pointer to MPW session structure.
 788  */
 789 static inline void
 790 mlx5_mpw_close(struct txq *txq, struct mlx5_mpw *mpw)
 791 {
 792         unsigned int num = mpw->pkts_n;
 793
 794         /*
 795          * Store size in multiple of 16 bytes. Control and Ethernet segments
 796          * count as 2.
 797          */
 798         mpw->wqe->ctrl[1] = rte_cpu_to_be_32(txq->qp_num_8s | (2 + num));
 799         mpw->state = MLX5_MPW_STATE_CLOSED;
 800         if (num < 3)
 801                 ++txq->wqe_ci;
 802         else
 803                 txq->wqe_ci += 2;
 804         rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci));
 805         rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci + 1));
 806 }
 807
 808 /**
 809  * DPDK callback for TX with MPW support.
 810  *
 811  * @param dpdk_txq
 812  *   Generic pointer to TX queue structure.
 813  * @param[in] pkts
 814  *   Packets to transmit.
 815  * @param pkts_n
 816  *   Number of packets in array.
 817  *
 818  * @return
 819  *   Number of packets successfully transmitted (<= pkts_n).
 820  */
 821 uint16_t
 822 mlx5_tx_burst_mpw(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
 823 {
 824         struct txq *txq = (struct txq *)dpdk_txq;
 825         uint16_t elts_head = txq->elts_head;
 826         const uint16_t elts_n = 1 << txq->elts_n;
 827         const uint16_t elts_m = elts_n - 1;
 828         unsigned int i = 0;
 829         unsigned int j = 0;
 830         uint16_t max_elts;
 831         uint16_t max_wqe;
 832         unsigned int comp;
 833         struct mlx5_mpw mpw = {
 834                 .state = MLX5_MPW_STATE_CLOSED,
 835         };
 836
 837         if (unlikely(!pkts_n))
 838                 return 0;
 839         /* Prefetch first packet cacheline. */
 840         rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci));
 841         rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci + 1));
 842         /* Start processing. */
 843         mlx5_tx_complete(txq);
 844         max_elts = (elts_n - (elts_head - txq->elts_tail));
 845         max_wqe = (1u << txq->wqe_n) - (txq->wqe_ci - txq->wqe_pi);
 846         if (unlikely(!max_wqe))
 847                 return 0;
 848         do {
 849                 struct rte_mbuf *buf = *(pkts++);
 850                 uint32_t length;
 851                 unsigned int segs_n = buf->nb_segs;
 852                 uint32_t cs_flags = 0;
 853
 854                 /*
 855                  * Make sure there is enough room to store this packet and
 856                  * that one ring entry remains unused.
 857                  */
 858                 assert(segs_n);
 859                 if (max_elts < segs_n)
 860                         break;
 861                 /* Do not bother with large packets MPW cannot handle. */
 862                 if (segs_n > MLX5_MPW_DSEG_MAX) {
 863                         txq->stats.oerrors++;
 864                         break;
 865                 }
 866                 max_elts -= segs_n;
 867                 --pkts_n;
 868                 /* Should we enable HW CKSUM offload */
 869                 if (buf->ol_flags &
 870                     (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM))
 871                         cs_flags = MLX5_ETH_WQE_L3_CSUM | MLX5_ETH_WQE_L4_CSUM;
 872                 /* Retrieve packet information. */
 873                 length = PKT_LEN(buf);
 874                 assert(length);
 875                 /* Start new session if packet differs. */
 876                 if ((mpw.state == MLX5_MPW_STATE_OPENED) &&
 877                     ((mpw.len != length) ||
 878                      (segs_n != 1) ||
 879                      (mpw.wqe->eseg.cs_flags != cs_flags)))
 880                         mlx5_mpw_close(txq, &mpw);
 881                 if (mpw.state == MLX5_MPW_STATE_CLOSED) {
 882                         /*
 883                          * Multi-Packet WQE consumes at most two WQE.
 884                          * mlx5_mpw_new() expects to be able to use such
 885                          * resources.
 886                          */
 887                         if (unlikely(max_wqe < 2))
 888                                 break;
 889                         max_wqe -= 2;
 890                         mlx5_mpw_new(txq, &mpw, length);
 891                         mpw.wqe->eseg.cs_flags = cs_flags;
 892                 }
 893                 /* Multi-segment packets must be alone in their MPW. */
 894                 assert((segs_n == 1) || (mpw.pkts_n == 0));
 895 #if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
 896                 length = 0;
 897 #endif
 898                 do {
 899                         volatile struct mlx5_wqe_data_seg *dseg;
 900                         uintptr_t addr;
 901
 902                         assert(buf);
 903                         (*txq->elts)[elts_head++ & elts_m] = buf;
 904                         dseg = mpw.data.dseg[mpw.pkts_n];
 905                         addr = rte_pktmbuf_mtod(buf, uintptr_t);
 906                         *dseg = (struct mlx5_wqe_data_seg){
 907                                 .byte_count = rte_cpu_to_be_32(DATA_LEN(buf)),
 908                                 .lkey = mlx5_tx_mb2mr(txq, buf),
 909                                 .addr = rte_cpu_to_be_64(addr),
 910                         };
 911 #if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
 912                         length += DATA_LEN(buf);
 913 #endif
 914                         buf = buf->next;
 915                         ++mpw.pkts_n;
 916                         ++j;
 917                 } while (--segs_n);
 918                 assert(length == mpw.len);
 919                 if (mpw.pkts_n == MLX5_MPW_DSEG_MAX)
 920                         mlx5_mpw_close(txq, &mpw);
 921 #ifdef MLX5_PMD_SOFT_COUNTERS
 922                 /* Increment sent bytes counter. */
 923                 txq->stats.obytes += length;
 924 #endif
 925                 ++i;
 926         } while (pkts_n);
 927         /* Take a shortcut if nothing must be sent. */
 928         if (unlikely(i == 0))
 929                 return 0;
 930         /* Check whether completion threshold has been reached. */
 931         /* "j" includes both packets and segments. */
 932         comp = txq->elts_comp + j;
 933         if (comp >= MLX5_TX_COMP_THRESH) {
 934                 volatile struct mlx5_wqe *wqe = mpw.wqe;
 935
 936                 /* Request completion on last WQE. */
 937                 wqe->ctrl[2] = rte_cpu_to_be_32(8);
 938                 /* Save elts_head in unused "immediate" field of WQE. */
 939                 wqe->ctrl[3] = elts_head;
 940                 txq->elts_comp = 0;
 941         } else {
 942                 txq->elts_comp = comp;
 943         }
 944 #ifdef MLX5_PMD_SOFT_COUNTERS
 945         /* Increment sent packets counter. */
 946         txq->stats.opackets += i;
 947 #endif
 948         /* Ring QP doorbell. */
 949         if (mpw.state == MLX5_MPW_STATE_OPENED)
 950                 mlx5_mpw_close(txq, &mpw);
 951         mlx5_tx_dbrec(txq, mpw.wqe);
 952         txq->elts_head = elts_head;
 953         return i;
 954 }
 955
 956 /**
 957  * Open a MPW inline session.
 958  *
 959  * @param txq
 960  *   Pointer to TX queue structure.
 961  * @param mpw
 962  *   Pointer to MPW session structure.
 963  * @param length
 964  *   Packet length.
 965  */
 966 static inline void
 967 mlx5_mpw_inline_new(struct txq *txq, struct mlx5_mpw *mpw, uint32_t length)
 968 {
 969         uint16_t idx = txq->wqe_ci & ((1 << txq->wqe_n) - 1);
 970         struct mlx5_wqe_inl_small *inl;
 971
 972         mpw->state = MLX5_MPW_INL_STATE_OPENED;
 973         mpw->pkts_n = 0;
 974         mpw->len = length;
 975         mpw->total_len = 0;
 976         mpw->wqe = (volatile struct mlx5_wqe *)tx_mlx5_wqe(txq, idx);
 977         mpw->wqe->ctrl[0] = rte_cpu_to_be_32((MLX5_OPC_MOD_MPW << 24) |
 978                                              (txq->wqe_ci << 8) |
 979                                              MLX5_OPCODE_TSO);
 980         mpw->wqe->ctrl[2] = 0;
 981         mpw->wqe->ctrl[3] = 0;
 982         mpw->wqe->eseg.mss = rte_cpu_to_be_16(length);
 983         mpw->wqe->eseg.inline_hdr_sz = 0;
 984         mpw->wqe->eseg.cs_flags = 0;
 985         mpw->wqe->eseg.rsvd0 = 0;
 986         mpw->wqe->eseg.rsvd1 = 0;
 987         mpw->wqe->eseg.rsvd2 = 0;
 988         inl = (struct mlx5_wqe_inl_small *)
 989                 (((uintptr_t)mpw->wqe) + 2 * MLX5_WQE_DWORD_SIZE);
 990         mpw->data.raw = (uint8_t *)&inl->raw;
 991 }
 992
 993 /**
 994  * Close a MPW inline session.
 995  *
 996  * @param txq
 997  *   Pointer to TX queue structure.
 998  * @param mpw
 999  *   Pointer to MPW session structure.
1000  */
1001 static inline void
1002 mlx5_mpw_inline_close(struct txq *txq, struct mlx5_mpw *mpw)
1003 {
1004         unsigned int size;
1005         struct mlx5_wqe_inl_small *inl = (struct mlx5_wqe_inl_small *)
1006                 (((uintptr_t)mpw->wqe) + (2 * MLX5_WQE_DWORD_SIZE));
1007
1008         size = MLX5_WQE_SIZE - MLX5_MWQE64_INL_DATA + mpw->total_len;
1009         /*
1010          * Store size in multiple of 16 bytes. Control and Ethernet segments
1011          * count as 2.
1012          */
1013         mpw->wqe->ctrl[1] = rte_cpu_to_be_32(txq->qp_num_8s |
1014                                              MLX5_WQE_DS(size));
1015         mpw->state = MLX5_MPW_STATE_CLOSED;
1016         inl->byte_cnt = rte_cpu_to_be_32(mpw->total_len | MLX5_INLINE_SEG);
1017         txq->wqe_ci += (size + (MLX5_WQE_SIZE - 1)) / MLX5_WQE_SIZE;
1018 }
1019
1020 /**
1021  * DPDK callback for TX with MPW inline support.
1022  *
1023  * @param dpdk_txq
1024  *   Generic pointer to TX queue structure.
1025  * @param[in] pkts
1026  *   Packets to transmit.
1027  * @param pkts_n
1028  *   Number of packets in array.
1029  *
1030  * @return
1031  *   Number of packets successfully transmitted (<= pkts_n).
1032  */
1033 uint16_t
1034 mlx5_tx_burst_mpw_inline(void *dpdk_txq, struct rte_mbuf **pkts,
1035                          uint16_t pkts_n)
1036 {
1037         struct txq *txq = (struct txq *)dpdk_txq;
1038         uint16_t elts_head = txq->elts_head;
1039         const uint16_t elts_n = 1 << txq->elts_n;
1040         const uint16_t elts_m = elts_n - 1;
1041         unsigned int i = 0;
1042         unsigned int j = 0;
1043         uint16_t max_elts;
1044         uint16_t max_wqe;
1045         unsigned int comp;
1046         unsigned int inline_room = txq->max_inline * RTE_CACHE_LINE_SIZE;
1047         struct mlx5_mpw mpw = {
1048                 .state = MLX5_MPW_STATE_CLOSED,
1049         };
1050         /*
1051          * Compute the maximum number of WQE which can be consumed by inline
1052          * code.
1053          * - 2 DSEG for:
1054          *   - 1 control segment,
1055          *   - 1 Ethernet segment,
1056          * - N Dseg from the inline request.
1057          */
1058         const unsigned int wqe_inl_n =
1059                 ((2 * MLX5_WQE_DWORD_SIZE +
1060                   txq->max_inline * RTE_CACHE_LINE_SIZE) +
1061                  RTE_CACHE_LINE_SIZE - 1) / RTE_CACHE_LINE_SIZE;
1062
1063         if (unlikely(!pkts_n))
1064                 return 0;
1065         /* Prefetch first packet cacheline. */
1066         rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci));
1067         rte_prefetch0(tx_mlx5_wqe(txq, txq->wqe_ci + 1));
1068         /* Start processing. */
1069         mlx5_tx_complete(txq);
1070         max_elts = (elts_n - (elts_head - txq->elts_tail));
1071         do {
1072                 struct rte_mbuf *buf = *(pkts++);
1073                 uintptr_t addr;
1074                 uint32_t length;
1075                 unsigned int segs_n = buf->nb_segs;
1076                 uint32_t cs_flags = 0;
1077
1078                 /*
1079                  * Make sure there is enough room to store this packet and
1080                  * that one ring entry remains unused.
1081                  */
1082                 assert(segs_n);
1083                 if (max_elts < segs_n)
1084                         break;
1085                 /* Do not bother with large packets MPW cannot handle. */
1086                 if (segs_n > MLX5_MPW_DSEG_MAX) {
1087                         txq->stats.oerrors++;
1088                         break;
1089                 }
1090                 max_elts -= segs_n;
1091                 --pkts_n;
1092                 /*
1093                  * Compute max_wqe in case less WQE were consumed in previous
1094                  * iteration.
1095                  */
1096                 max_wqe = (1u << txq->wqe_n) - (txq->wqe_ci - txq->wqe_pi);
1097                 /* Should we enable HW CKSUM offload */
1098                 if (buf->ol_flags &
1099                     (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM))
1100                         cs_flags = MLX5_ETH_WQE_L3_CSUM | MLX5_ETH_WQE_L4_CSUM;
1101                 /* Retrieve packet information. */
1102                 length = PKT_LEN(buf);
1103                 /* Start new session if packet differs. */
1104                 if (mpw.state == MLX5_MPW_STATE_OPENED) {
1105                         if ((mpw.len != length) ||
1106                             (segs_n != 1) ||
1107                             (mpw.wqe->eseg.cs_flags != cs_flags))
1108                                 mlx5_mpw_close(txq, &mpw);
1109                 } else if (mpw.state == MLX5_MPW_INL_STATE_OPENED) {
1110                         if ((mpw.len != length) ||
1111                             (segs_n != 1) ||
1112                             (length > inline_room) ||
1113                             (mpw.wqe->eseg.cs_flags != cs_flags)) {
1114                                 mlx5_mpw_inline_close(txq, &mpw);
1115                                 inline_room =
1116                                         txq->max_inline * RTE_CACHE_LINE_SIZE;
1117                         }
1118                 }
1119                 if (mpw.state == MLX5_MPW_STATE_CLOSED) {
1120                         if ((segs_n != 1) ||
1121                             (length > inline_room)) {
1122                                 /*
1123                                  * Multi-Packet WQE consumes at most two WQE.
1124                                  * mlx5_mpw_new() expects to be able to use
1125                                  * such resources.
1126                                  */
1127                                 if (unlikely(max_wqe < 2))
1128                                         break;
1129                                 max_wqe -= 2;
1130                                 mlx5_mpw_new(txq, &mpw, length);
1131                                 mpw.wqe->eseg.cs_flags = cs_flags;
1132                         } else {
1133                                 if (unlikely(max_wqe < wqe_inl_n))
1134                                         break;
1135                                 max_wqe -= wqe_inl_n;
1136                                 mlx5_mpw_inline_new(txq, &mpw, length);
1137                                 mpw.wqe->eseg.cs_flags = cs_flags;
1138                         }
1139                 }
1140                 /* Multi-segment packets must be alone in their MPW. */
1141                 assert((segs_n == 1) || (mpw.pkts_n == 0));
1142                 if (mpw.state == MLX5_MPW_STATE_OPENED) {
1143                         assert(inline_room ==
1144                                txq->max_inline * RTE_CACHE_LINE_SIZE);
1145 #if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
1146                         length = 0;
1147 #endif
1148                         do {
1149                                 volatile struct mlx5_wqe_data_seg *dseg;
1150
1151                                 assert(buf);
1152                                 (*txq->elts)[elts_head++ & elts_m] = buf;
1153                                 dseg = mpw.data.dseg[mpw.pkts_n];
1154                                 addr = rte_pktmbuf_mtod(buf, uintptr_t);
1155                                 *dseg = (struct mlx5_wqe_data_seg){
1156                                         .byte_count =
1157                                                rte_cpu_to_be_32(DATA_LEN(buf)),
1158                                         .lkey = mlx5_tx_mb2mr(txq, buf),
1159                                         .addr = rte_cpu_to_be_64(addr),
1160                                 };
1161 #if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
1162                                 length += DATA_LEN(buf);
1163 #endif
1164                                 buf = buf->next;
1165                                 ++mpw.pkts_n;
1166                                 ++j;
1167                         } while (--segs_n);
1168                         assert(length == mpw.len);
1169                         if (mpw.pkts_n == MLX5_MPW_DSEG_MAX)
1170                                 mlx5_mpw_close(txq, &mpw);
1171                 } else {
1172                         unsigned int max;
1173
1174                         assert(mpw.state == MLX5_MPW_INL_STATE_OPENED);
1175                         assert(length <= inline_room);
1176                         assert(length == DATA_LEN(buf));
1177                         addr = rte_pktmbuf_mtod(buf, uintptr_t);
1178                         (*txq->elts)[elts_head++ & elts_m] = buf;
1179                         /* Maximum number of bytes before wrapping. */
1180                         max = ((((uintptr_t)(txq->wqes)) +
1181                                 (1 << txq->wqe_n) *
1182                                 MLX5_WQE_SIZE) -
1183                                (uintptr_t)mpw.data.raw);
1184                         if (length > max) {
1185                                 rte_memcpy((void *)(uintptr_t)mpw.data.raw,
1186                                            (void *)addr,
1187                                            max);
1188                                 mpw.data.raw = (volatile void *)txq->wqes;
1189                                 rte_memcpy((void *)(uintptr_t)mpw.data.raw,
1190                                            (void *)(addr + max),
1191                                            length - max);
1192                                 mpw.data.raw += length - max;
1193                         } else {
1194                                 rte_memcpy((void *)(uintptr_t)mpw.data.raw,
1195                                            (void *)addr,
1196                                            length);
1197
1198                                 if (length == max)
1199                                         mpw.data.raw =
1200                                                 (volatile void *)txq->wqes;
1201                                 else
1202                                         mpw.data.raw += length;
1203                         }
1204                         ++mpw.pkts_n;
1205                         mpw.total_len += length;
1206                         ++j;
1207                         if (mpw.pkts_n == MLX5_MPW_DSEG_MAX) {
1208                                 mlx5_mpw_inline_close(txq, &mpw);
1209                                 inline_room =
1210                                         txq->max_inline * RTE_CACHE_LINE_SIZE;
1211                         } else {
1212                                 inline_room -= length;
1213                         }
1214                 }
1215 #ifdef MLX5_PMD_SOFT_COUNTERS
1216                 /* Increment sent bytes counter. */
1217                 txq->stats.obytes += length;
1218 #endif
1219                 ++i;
1220         } while (pkts_n);
1221         /* Take a shortcut if nothing must be sent. */
1222         if (unlikely(i == 0))
1223                 return 0;
1224         /* Check whether completion threshold has been reached. */
1225         /* "j" includes both packets and segments. */
1226         comp = txq->elts_comp + j;
1227         if (comp >= MLX5_TX_COMP_THRESH) {
1228                 volatile struct mlx5_wqe *wqe = mpw.wqe;
1229
1230                 /* Request completion on last WQE. */
1231                 wqe->ctrl[2] = rte_cpu_to_be_32(8);
1232                 /* Save elts_head in unused "immediate" field of WQE. */
1233                 wqe->ctrl[3] = elts_head;
1234                 txq->elts_comp = 0;
1235         } else {
1236                 txq->elts_comp = comp;
1237         }
1238 #ifdef MLX5_PMD_SOFT_COUNTERS
1239         /* Increment sent packets counter. */
1240         txq->stats.opackets += i;
1241 #endif
1242         /* Ring QP doorbell. */
1243         if (mpw.state == MLX5_MPW_INL_STATE_OPENED)
1244                 mlx5_mpw_inline_close(txq, &mpw);
1245         else if (mpw.state == MLX5_MPW_STATE_OPENED)
1246                 mlx5_mpw_close(txq, &mpw);
1247         mlx5_tx_dbrec(txq, mpw.wqe);
1248         txq->elts_head = elts_head;
1249         return i;
1250 }
1251
1252 /**
1253  * Open an Enhanced MPW session.
1254  *
1255  * @param txq
1256  *   Pointer to TX queue structure.
1257  * @param mpw
1258  *   Pointer to MPW session structure.
1259  * @param length
1260  *   Packet length.
1261  */
1262 static inline void
1263 mlx5_empw_new(struct txq *txq, struct mlx5_mpw *mpw, int padding)
1264 {
1265         uint16_t idx = txq->wqe_ci & ((1 << txq->wqe_n) - 1);
1266
1267         mpw->state = MLX5_MPW_ENHANCED_STATE_OPENED;
1268         mpw->pkts_n = 0;
1269         mpw->total_len = sizeof(struct mlx5_wqe);
1270         mpw->wqe = (volatile struct mlx5_wqe *)tx_mlx5_wqe(txq, idx);
1271         mpw->wqe->ctrl[0] =
1272                 rte_cpu_to_be_32((MLX5_OPC_MOD_ENHANCED_MPSW << 24) |
1273                                  (txq->wqe_ci << 8) |
1274                                  MLX5_OPCODE_ENHANCED_MPSW);
1275         mpw->wqe->ctrl[2] = 0;
1276         mpw->wqe->ctrl[3] = 0;
1277         memset((void *)(uintptr_t)&mpw->wqe->eseg, 0, MLX5_WQE_DWORD_SIZE);
1278         if (unlikely(padding)) {
1279                 uintptr_t addr = (uintptr_t)(mpw->wqe + 1);
1280
1281                 /* Pad the first 2 DWORDs with zero-length inline header. */
1282                 *(volatile uint32_t *)addr = rte_cpu_to_be_32(MLX5_INLINE_SEG);
1283                 *(volatile uint32_t *)(addr + MLX5_WQE_DWORD_SIZE) =
1284                         rte_cpu_to_be_32(MLX5_INLINE_SEG);
1285                 mpw->total_len += 2 * MLX5_WQE_DWORD_SIZE;
1286                 /* Start from the next WQEBB. */
1287                 mpw->data.raw = (volatile void *)(tx_mlx5_wqe(txq, idx + 1));
1288         } else {
1289                 mpw->data.raw = (volatile void *)(mpw->wqe + 1);
1290         }
1291 }
1292
1293 /**
1294  * Close an Enhanced MPW session.
1295  *
1296  * @param txq
1297  *   Pointer to TX queue structure.
1298  * @param mpw
1299  *   Pointer to MPW session structure.
1300  *
1301  * @return
1302  *   Number of consumed WQEs.
1303  */
1304 static inline uint16_t
1305 mlx5_empw_close(struct txq *txq, struct mlx5_mpw *mpw)
1306 {
1307         uint16_t ret;
1308
1309         /* Store size in multiple of 16 bytes. Control and Ethernet segments
1310          * count as 2.
1311          */
1312         mpw->wqe->ctrl[1] = rte_cpu_to_be_32(txq->qp_num_8s |
1313                                              MLX5_WQE_DS(mpw->total_len));
1314         mpw->state = MLX5_MPW_STATE_CLOSED;
1315         ret = (mpw->total_len + (MLX5_WQE_SIZE - 1)) / MLX5_WQE_SIZE;
1316         txq->wqe_ci += ret;
1317         return ret;
1318 }
1319
1320 /**
1321  * DPDK callback for TX with Enhanced MPW support.
1322  *
1323  * @param dpdk_txq
1324  *   Generic pointer to TX queue structure.
1325  * @param[in] pkts
1326  *   Packets to transmit.
1327  * @param pkts_n
1328  *   Number of packets in array.
1329  *
1330  * @return
1331  *   Number of packets successfully transmitted (<= pkts_n).
1332  */
1333 uint16_t
1334 mlx5_tx_burst_empw(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
1335 {
1336         struct txq *txq = (struct txq *)dpdk_txq;
1337         uint16_t elts_head = txq->elts_head;
1338         const uint16_t elts_n = 1 << txq->elts_n;
1339         const uint16_t elts_m = elts_n - 1;
1340         unsigned int i = 0;
1341         unsigned int j = 0;
1342         uint16_t max_elts;
1343         uint16_t max_wqe;
1344         unsigned int max_inline = txq->max_inline * RTE_CACHE_LINE_SIZE;
1345         unsigned int mpw_room = 0;
1346         unsigned int inl_pad = 0;
1347         uint32_t inl_hdr;
1348         struct mlx5_mpw mpw = {
1349                 .state = MLX5_MPW_STATE_CLOSED,
1350         };
1351
1352         if (unlikely(!pkts_n))
1353                 return 0;
1354         /* Start processing. */
1355         mlx5_tx_complete(txq);
1356         max_elts = (elts_n - (elts_head - txq->elts_tail));
1357         /* A CQE slot must always be available. */
1358         assert((1u << txq->cqe_n) - (txq->cq_pi - txq->cq_ci));
1359         max_wqe = (1u << txq->wqe_n) - (txq->wqe_ci - txq->wqe_pi);
1360         if (unlikely(!max_wqe))
1361                 return 0;
1362         do {
1363                 struct rte_mbuf *buf = *(pkts++);
1364                 uintptr_t addr;
1365                 uint64_t naddr;
1366                 unsigned int n;
1367                 unsigned int do_inline = 0; /* Whether inline is possible. */
1368                 uint32_t length;
1369                 unsigned int segs_n = buf->nb_segs;
1370                 uint32_t cs_flags = 0;
1371
1372                 /*
1373                  * Make sure there is enough room to store this packet and
1374                  * that one ring entry remains unused.
1375                  */
1376                 assert(segs_n);
1377                 if (max_elts - j < segs_n)
1378                         break;
1379                 /* Do not bother with large packets MPW cannot handle. */
1380                 if (segs_n > MLX5_MPW_DSEG_MAX) {
1381                         txq->stats.oerrors++;
1382                         break;
1383                 }
1384                 /* Should we enable HW CKSUM offload. */
1385                 if (buf->ol_flags &
1386                     (PKT_TX_IP_CKSUM | PKT_TX_TCP_CKSUM | PKT_TX_UDP_CKSUM))
1387                         cs_flags = MLX5_ETH_WQE_L3_CSUM | MLX5_ETH_WQE_L4_CSUM;
1388                 /* Retrieve packet information. */
1389                 length = PKT_LEN(buf);
1390                 /* Start new session if:
1391                  * - multi-segment packet
1392                  * - no space left even for a dseg
1393                  * - next packet can be inlined with a new WQE
1394                  * - cs_flag differs
1395                  * It can't be MLX5_MPW_STATE_OPENED as always have a single
1396                  * segmented packet.
1397                  */
1398                 if (mpw.state == MLX5_MPW_ENHANCED_STATE_OPENED) {
1399                         if ((segs_n != 1) ||
1400                             (inl_pad + sizeof(struct mlx5_wqe_data_seg) >
1401                               mpw_room) ||
1402                             (length <= txq->inline_max_packet_sz &&
1403                              inl_pad + sizeof(inl_hdr) + length >
1404                               mpw_room) ||
1405                             (mpw.wqe->eseg.cs_flags != cs_flags))
1406                                 max_wqe -= mlx5_empw_close(txq, &mpw);
1407                 }
1408                 if (unlikely(mpw.state == MLX5_MPW_STATE_CLOSED)) {
1409                         if (unlikely(segs_n != 1)) {
1410                                 /* Fall back to legacy MPW.
1411                                  * A MPW session consumes 2 WQEs at most to
1412                                  * include MLX5_MPW_DSEG_MAX pointers.
1413                                  */
1414                                 if (unlikely(max_wqe < 2))
1415                                         break;
1416                                 mlx5_mpw_new(txq, &mpw, length);
1417                         } else {
1418                                 /* In Enhanced MPW, inline as much as the budget
1419                                  * is allowed. The remaining space is to be
1420                                  * filled with dsegs. If the title WQEBB isn't
1421                                  * padded, it will have 2 dsegs there.
1422                                  */
1423                                 mpw_room = RTE_MIN(MLX5_WQE_SIZE_MAX,
1424                                             (max_inline ? max_inline :
1425                                              pkts_n * MLX5_WQE_DWORD_SIZE) +
1426                                             MLX5_WQE_SIZE);
1427                                 if (unlikely(max_wqe * MLX5_WQE_SIZE <
1428                                               mpw_room))
1429                                         break;
1430                                 /* Don't pad the title WQEBB to not waste WQ. */
1431                                 mlx5_empw_new(txq, &mpw, 0);
1432                                 mpw_room -= mpw.total_len;
1433                                 inl_pad = 0;
1434                                 do_inline =
1435                                         length <= txq->inline_max_packet_sz &&
1436                                         sizeof(inl_hdr) + length <= mpw_room &&
1437                                         !txq->mpw_hdr_dseg;
1438                         }
1439                         mpw.wqe->eseg.cs_flags = cs_flags;
1440                 } else {
1441                         /* Evaluate whether the next packet can be inlined.
1442                          * Inlininig is possible when:
1443                          * - length is less than configured value
1444                          * - length fits for remaining space
1445                          * - not required to fill the title WQEBB with dsegs
1446                          */
1447                         do_inline =
1448                                 length <= txq->inline_max_packet_sz &&
1449                                 inl_pad + sizeof(inl_hdr) + length <=
1450                                  mpw_room &&
1451                                 (!txq->mpw_hdr_dseg ||
1452                                  mpw.total_len >= MLX5_WQE_SIZE);
1453                 }
1454                 /* Multi-segment packets must be alone in their MPW. */
1455                 assert((segs_n == 1) || (mpw.pkts_n == 0));
1456                 if (unlikely(mpw.state == MLX5_MPW_STATE_OPENED)) {
1457 #if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
1458                         length = 0;
1459 #endif
1460                         do {
1461                                 volatile struct mlx5_wqe_data_seg *dseg;
1462
1463                                 assert(buf);
1464                                 (*txq->elts)[elts_head++ & elts_m] = buf;
1465                                 dseg = mpw.data.dseg[mpw.pkts_n];
1466                                 addr = rte_pktmbuf_mtod(buf, uintptr_t);
1467                                 *dseg = (struct mlx5_wqe_data_seg){
1468                                         .byte_count = rte_cpu_to_be_32(
1469                                                                 DATA_LEN(buf)),
1470                                         .lkey = mlx5_tx_mb2mr(txq, buf),
1471                                         .addr = rte_cpu_to_be_64(addr),
1472                                 };
1473 #if defined(MLX5_PMD_SOFT_COUNTERS) || !defined(NDEBUG)
1474                                 length += DATA_LEN(buf);
1475 #endif
1476                                 buf = buf->next;
1477                                 ++j;
1478                                 ++mpw.pkts_n;
1479                         } while (--segs_n);
1480                         /* A multi-segmented packet takes one MPW session.
1481                          * TODO: Pack more multi-segmented packets if possible.
1482                          */
1483                         mlx5_mpw_close(txq, &mpw);
1484                         if (mpw.pkts_n < 3)
1485                                 max_wqe--;
1486                         else
1487                                 max_wqe -= 2;
1488                 } else if (do_inline) {
1489                         /* Inline packet into WQE. */
1490                         unsigned int max;
1491
1492                         assert(mpw.state == MLX5_MPW_ENHANCED_STATE_OPENED);
1493                         assert(length == DATA_LEN(buf));
1494                         inl_hdr = rte_cpu_to_be_32(length | MLX5_INLINE_SEG);
1495                         addr = rte_pktmbuf_mtod(buf, uintptr_t);
1496                         mpw.data.raw = (volatile void *)
1497                                 ((uintptr_t)mpw.data.raw + inl_pad);
1498                         max = tx_mlx5_wq_tailroom(txq,
1499                                         (void *)(uintptr_t)mpw.data.raw);
1500                         /* Copy inline header. */
1501                         mpw.data.raw = (volatile void *)
1502                                 mlx5_copy_to_wq(
1503                                           (void *)(uintptr_t)mpw.data.raw,
1504                                           &inl_hdr,
1505                                           sizeof(inl_hdr),
1506                                           (void *)(uintptr_t)txq->wqes,
1507                                           max);
1508                         max = tx_mlx5_wq_tailroom(txq,
1509                                         (void *)(uintptr_t)mpw.data.raw);
1510                         /* Copy packet data. */
1511                         mpw.data.raw = (volatile void *)
1512                                 mlx5_copy_to_wq(
1513                                           (void *)(uintptr_t)mpw.data.raw,
1514                                           (void *)addr,
1515                                           length,
1516                                           (void *)(uintptr_t)txq->wqes,
1517                                           max);
1518                         ++mpw.pkts_n;
1519                         mpw.total_len += (inl_pad + sizeof(inl_hdr) + length);
1520                         /* No need to get completion as the entire packet is
1521                          * copied to WQ. Free the buf right away.
1522                          */
1523                         rte_pktmbuf_free_seg(buf);
1524                         mpw_room -= (inl_pad + sizeof(inl_hdr) + length);
1525                         /* Add pad in the next packet if any. */
1526                         inl_pad = (((uintptr_t)mpw.data.raw +
1527                                         (MLX5_WQE_DWORD_SIZE - 1)) &
1528                                         ~(MLX5_WQE_DWORD_SIZE - 1)) -
1529                                   (uintptr_t)mpw.data.raw;
1530                 } else {
1531                         /* No inline. Load a dseg of packet pointer. */
1532                         volatile rte_v128u32_t *dseg;
1533
1534                         assert(mpw.state == MLX5_MPW_ENHANCED_STATE_OPENED);
1535                         assert((inl_pad + sizeof(*dseg)) <= mpw_room);
1536                         assert(length == DATA_LEN(buf));
1537                         if (!tx_mlx5_wq_tailroom(txq,
1538                                         (void *)((uintptr_t)mpw.data.raw
1539                                                 + inl_pad)))
1540                                 dseg = (volatile void *)txq->wqes;
1541                         else
1542                                 dseg = (volatile void *)
1543                                         ((uintptr_t)mpw.data.raw +
1544                                          inl_pad);
1545                         (*txq->elts)[elts_head++ & elts_m] = buf;
1546                         addr = rte_pktmbuf_mtod(buf, uintptr_t);
1547                         for (n = 0; n * RTE_CACHE_LINE_SIZE < length; n++)
1548                                 rte_prefetch2((void *)(addr +
1549                                                 n * RTE_CACHE_LINE_SIZE));
1550                         naddr = rte_cpu_to_be_64(addr);
1551                         *dseg = (rte_v128u32_t) {
1552                                 rte_cpu_to_be_32(length),
1553                                 mlx5_tx_mb2mr(txq, buf),
1554                                 naddr,
1555                                 naddr >> 32,
1556                         };
1557                         mpw.data.raw = (volatile void *)(dseg + 1);
1558                         mpw.total_len += (inl_pad + sizeof(*dseg));
1559                         ++j;
1560                         ++mpw.pkts_n;
1561                         mpw_room -= (inl_pad + sizeof(*dseg));
1562                         inl_pad = 0;
1563                 }
1564 #ifdef MLX5_PMD_SOFT_COUNTERS
1565                 /* Increment sent bytes counter. */
1566                 txq->stats.obytes += length;
1567 #endif
1568                 ++i;
1569         } while (i < pkts_n);
1570         /* Take a shortcut if nothing must be sent. */
1571         if (unlikely(i == 0))
1572                 return 0;
1573         /* Check whether completion threshold has been reached. */
1574         if (txq->elts_comp + j >= MLX5_TX_COMP_THRESH ||
1575                         (uint16_t)(txq->wqe_ci - txq->mpw_comp) >=
1576                          (1 << txq->wqe_n) / MLX5_TX_COMP_THRESH_INLINE_DIV) {
1577                 volatile struct mlx5_wqe *wqe = mpw.wqe;
1578
1579                 /* Request completion on last WQE. */
1580                 wqe->ctrl[2] = rte_cpu_to_be_32(8);
1581                 /* Save elts_head in unused "immediate" field of WQE. */
1582                 wqe->ctrl[3] = elts_head;
1583                 txq->elts_comp = 0;
1584                 txq->mpw_comp = txq->wqe_ci;
1585                 txq->cq_pi++;
1586         } else {
1587                 txq->elts_comp += j;
1588         }
1589 #ifdef MLX5_PMD_SOFT_COUNTERS
1590         /* Increment sent packets counter. */
1591         txq->stats.opackets += i;
1592 #endif
1593         if (mpw.state == MLX5_MPW_ENHANCED_STATE_OPENED)
1594                 mlx5_empw_close(txq, &mpw);
1595         else if (mpw.state == MLX5_MPW_STATE_OPENED)
1596                 mlx5_mpw_close(txq, &mpw);
1597         /* Ring QP doorbell. */
1598         mlx5_tx_dbrec(txq, mpw.wqe);
1599         txq->elts_head = elts_head;
1600         return i;
1601 }
1602
1603 /**
1604  * Translate RX completion flags to packet type.
1605  *
1606  * @param[in] cqe
1607  *   Pointer to CQE.
1608  *
1609  * @note: fix mlx5_dev_supported_ptypes_get() if any change here.
1610  *
1611  * @return
1612  *   Packet type for struct rte_mbuf.
1613  */
1614 static inline uint32_t
1615 rxq_cq_to_pkt_type(volatile struct mlx5_cqe *cqe)
1616 {
1617         uint8_t idx;
1618         uint8_t pinfo = cqe->pkt_info;
1619         uint16_t ptype = cqe->hdr_type_etc;
1620
1621         /*
1622          * The index to the array should have:
1623          * bit[1:0] = l3_hdr_type
1624          * bit[4:2] = l4_hdr_type
1625          * bit[5] = ip_frag
1626          * bit[6] = tunneled
1627          * bit[7] = outer_l3_type
1628          */
1629         idx = ((pinfo & 0x3) << 6) | ((ptype & 0xfc00) >> 10);
1630         return mlx5_ptype_table[idx];
1631 }
1632
1633 /**
1634  * Get size of the next packet for a given CQE. For compressed CQEs, the
1635  * consumer index is updated only once all packets of the current one have
1636  * been processed.
1637  *
1638  * @param rxq
1639  *   Pointer to RX queue.
1640  * @param cqe
1641  *   CQE to process.
1642  * @param[out] rss_hash
1643  *   Packet RSS Hash result.
1644  *
1645  * @return
1646  *   Packet size in bytes (0 if there is none), -1 in case of completion
1647  *   with error.
1648  */
1649 static inline int
1650 mlx5_rx_poll_len(struct rxq *rxq, volatile struct mlx5_cqe *cqe,
1651                  uint16_t cqe_cnt, uint32_t *rss_hash)
1652 {
1653         struct rxq_zip *zip = &rxq->zip;
1654         uint16_t cqe_n = cqe_cnt + 1;
1655         int len = 0;
1656         uint16_t idx, end;
1657
1658         /* Process compressed data in the CQE and mini arrays. */
1659         if (zip->ai) {
1660                 volatile struct mlx5_mini_cqe8 (*mc)[8] =
1661                         (volatile struct mlx5_mini_cqe8 (*)[8])
1662                         (uintptr_t)(&(*rxq->cqes)[zip->ca & cqe_cnt].pkt_info);
1663
1664                 len = rte_be_to_cpu_32((*mc)[zip->ai & 7].byte_cnt);
1665                 *rss_hash = rte_be_to_cpu_32((*mc)[zip->ai & 7].rx_hash_result);
1666                 if ((++zip->ai & 7) == 0) {
1667                         /* Invalidate consumed CQEs */
1668                         idx = zip->ca;
1669                         end = zip->na;
1670                         while (idx != end) {
1671                                 (*rxq->cqes)[idx & cqe_cnt].op_own =
1672                                         MLX5_CQE_INVALIDATE;
1673                                 ++idx;
1674                         }
1675                         /*
1676                          * Increment consumer index to skip the number of
1677                          * CQEs consumed. Hardware leaves holes in the CQ
1678                          * ring for software use.
1679                          */
1680                         zip->ca = zip->na;
1681                         zip->na += 8;
1682                 }
1683                 if (unlikely(rxq->zip.ai == rxq->zip.cqe_cnt)) {
1684                         /* Invalidate the rest */
1685                         idx = zip->ca;
1686                         end = zip->cq_ci;
1687
1688                         while (idx != end) {
1689                                 (*rxq->cqes)[idx & cqe_cnt].op_own =
1690                                         MLX5_CQE_INVALIDATE;
1691                                 ++idx;
1692                         }
1693                         rxq->cq_ci = zip->cq_ci;
1694                         zip->ai = 0;
1695                 }
1696         /* No compressed data, get next CQE and verify if it is compressed. */
1697         } else {
1698                 int ret;
1699                 int8_t op_own;
1700
1701                 ret = check_cqe(cqe, cqe_n, rxq->cq_ci);
1702                 if (unlikely(ret == 1))
1703                         return 0;
1704                 ++rxq->cq_ci;
1705                 op_own = cqe->op_own;
1706                 if (MLX5_CQE_FORMAT(op_own) == MLX5_COMPRESSED) {
1707                         volatile struct mlx5_mini_cqe8 (*mc)[8] =
1708                                 (volatile struct mlx5_mini_cqe8 (*)[8])
1709                                 (uintptr_t)(&(*rxq->cqes)[rxq->cq_ci &
1710                                                           cqe_cnt].pkt_info);
1711
1712                         /* Fix endianness. */
1713                         zip->cqe_cnt = rte_be_to_cpu_32(cqe->byte_cnt);
1714                         /*
1715                          * Current mini array position is the one returned by
1716                          * check_cqe64().
1717                          *
1718                          * If completion comprises several mini arrays, as a
1719                          * special case the second one is located 7 CQEs after
1720                          * the initial CQE instead of 8 for subsequent ones.
1721                          */
1722                         zip->ca = rxq->cq_ci;
1723                         zip->na = zip->ca + 7;
1724                         /* Compute the next non compressed CQE. */
1725                         --rxq->cq_ci;
1726                         zip->cq_ci = rxq->cq_ci + zip->cqe_cnt;
1727                         /* Get packet size to return. */
1728                         len = rte_be_to_cpu_32((*mc)[0].byte_cnt);
1729                         *rss_hash = rte_be_to_cpu_32((*mc)[0].rx_hash_result);
1730                         zip->ai = 1;
1731                         /* Prefetch all the entries to be invalidated */
1732                         idx = zip->ca;
1733                         end = zip->cq_ci;
1734                         while (idx != end) {
1735                                 rte_prefetch0(&(*rxq->cqes)[(idx) & cqe_cnt]);
1736                                 ++idx;
1737                         }
1738                 } else {
1739                         len = rte_be_to_cpu_32(cqe->byte_cnt);
1740                         *rss_hash = rte_be_to_cpu_32(cqe->rx_hash_res);
1741                 }
1742                 /* Error while receiving packet. */
1743                 if (unlikely(MLX5_CQE_OPCODE(op_own) == MLX5_CQE_RESP_ERR))
1744                         return -1;
1745         }
1746         return len;
1747 }
1748
1749 /**
1750  * Translate RX completion flags to offload flags.
1751  *
1752  * @param[in] rxq
1753  *   Pointer to RX queue structure.
1754  * @param[in] cqe
1755  *   Pointer to CQE.
1756  *
1757  * @return
1758  *   Offload flags (ol_flags) for struct rte_mbuf.
1759  */
1760 static inline uint32_t
1761 rxq_cq_to_ol_flags(struct rxq *rxq, volatile struct mlx5_cqe *cqe)
1762 {
1763         uint32_t ol_flags = 0;
1764         uint16_t flags = rte_be_to_cpu_16(cqe->hdr_type_etc);
1765
1766         ol_flags =
1767                 TRANSPOSE(flags,
1768                           MLX5_CQE_RX_L3_HDR_VALID,
1769                           PKT_RX_IP_CKSUM_GOOD) |
1770                 TRANSPOSE(flags,
1771                           MLX5_CQE_RX_L4_HDR_VALID,
1772                           PKT_RX_L4_CKSUM_GOOD);
1773         if ((cqe->pkt_info & MLX5_CQE_RX_TUNNEL_PACKET) && (rxq->csum_l2tun))
1774                 ol_flags |=
1775                         TRANSPOSE(flags,
1776                                   MLX5_CQE_RX_L3_HDR_VALID,
1777                                   PKT_RX_IP_CKSUM_GOOD) |
1778                         TRANSPOSE(flags,
1779                                   MLX5_CQE_RX_L4_HDR_VALID,
1780                                   PKT_RX_L4_CKSUM_GOOD);
1781         return ol_flags;
1782 }
1783
1784 /**
1785  * DPDK callback for RX.
1786  *
1787  * @param dpdk_rxq
1788  *   Generic pointer to RX queue structure.
1789  * @param[out] pkts
1790  *   Array to store received packets.
1791  * @param pkts_n
1792  *   Maximum number of packets in array.
1793  *
1794  * @return
1795  *   Number of packets successfully received (<= pkts_n).
1796  */
1797 uint16_t
1798 mlx5_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
1799 {
1800         struct rxq *rxq = dpdk_rxq;
1801         const unsigned int wqe_cnt = (1 << rxq->elts_n) - 1;
1802         const unsigned int cqe_cnt = (1 << rxq->cqe_n) - 1;
1803         const unsigned int sges_n = rxq->sges_n;
1804         struct rte_mbuf *pkt = NULL;
1805         struct rte_mbuf *seg = NULL;
1806         volatile struct mlx5_cqe *cqe =
1807                 &(*rxq->cqes)[rxq->cq_ci & cqe_cnt];
1808         unsigned int i = 0;
1809         unsigned int rq_ci = rxq->rq_ci << sges_n;
1810         int len = 0; /* keep its value across iterations. */
1811
1812         while (pkts_n) {
1813                 unsigned int idx = rq_ci & wqe_cnt;
1814                 volatile struct mlx5_wqe_data_seg *wqe = &(*rxq->wqes)[idx];
1815                 struct rte_mbuf *rep = (*rxq->elts)[idx];
1816                 uint32_t rss_hash_res = 0;
1817
1818                 if (pkt)
1819                         NEXT(seg) = rep;
1820                 seg = rep;
1821                 rte_prefetch0(seg);
1822                 rte_prefetch0(cqe);
1823                 rte_prefetch0(wqe);
1824                 rep = rte_mbuf_raw_alloc(rxq->mp);
1825                 if (unlikely(rep == NULL)) {
1826                         ++rxq->stats.rx_nombuf;
1827                         if (!pkt) {
1828                                 /*
1829                                  * no buffers before we even started,
1830                                  * bail out silently.
1831                                  */
1832                                 break;
1833                         }
1834                         while (pkt != seg) {
1835                                 assert(pkt != (*rxq->elts)[idx]);
1836                                 rep = NEXT(pkt);
1837                                 NEXT(pkt) = NULL;
1838                                 NB_SEGS(pkt) = 1;
1839                                 rte_mbuf_raw_free(pkt);
1840                                 pkt = rep;
1841                         }
1842                         break;
1843                 }
1844                 if (!pkt) {
1845                         cqe = &(*rxq->cqes)[rxq->cq_ci & cqe_cnt];
1846                         len = mlx5_rx_poll_len(rxq, cqe, cqe_cnt,
1847                                                &rss_hash_res);
1848                         if (!len) {
1849                                 rte_mbuf_raw_free(rep);
1850                                 break;
1851                         }
1852                         if (unlikely(len == -1)) {
1853                                 /* RX error, packet is likely too large. */
1854                                 rte_mbuf_raw_free(rep);
1855                                 ++rxq->stats.idropped;
1856                                 goto skip;
1857                         }
1858                         pkt = seg;
1859                         assert(len >= (rxq->crc_present << 2));
1860                         /* Update packet information. */
1861                         pkt->packet_type = rxq_cq_to_pkt_type(cqe);
1862                         pkt->ol_flags = 0;
1863                         if (rss_hash_res && rxq->rss_hash) {
1864                                 pkt->hash.rss = rss_hash_res;
1865                                 pkt->ol_flags = PKT_RX_RSS_HASH;
1866                         }
1867                         if (rxq->mark &&
1868                             MLX5_FLOW_MARK_IS_VALID(cqe->sop_drop_qpn)) {
1869                                 pkt->ol_flags |= PKT_RX_FDIR;
1870                                 if (cqe->sop_drop_qpn !=
1871                                     rte_cpu_to_be_32(MLX5_FLOW_MARK_DEFAULT)) {
1872                                         uint32_t mark = cqe->sop_drop_qpn;
1873
1874                                         pkt->ol_flags |= PKT_RX_FDIR_ID;
1875                                         pkt->hash.fdir.hi =
1876                                                 mlx5_flow_mark_get(mark);
1877                                 }
1878                         }
1879                         if (rxq->csum | rxq->csum_l2tun)
1880                                 pkt->ol_flags |= rxq_cq_to_ol_flags(rxq, cqe);
1881                         if (rxq->vlan_strip &&
1882                             (cqe->hdr_type_etc &
1883                              rte_cpu_to_be_16(MLX5_CQE_VLAN_STRIPPED))) {
1884                                 pkt->ol_flags |= PKT_RX_VLAN_PKT |
1885                                         PKT_RX_VLAN_STRIPPED;
1886                                 pkt->vlan_tci =
1887                                         rte_be_to_cpu_16(cqe->vlan_info);
1888                         }
1889                         if (rxq->crc_present)
1890                                 len -= ETHER_CRC_LEN;
1891                         PKT_LEN(pkt) = len;
1892                 }
1893                 DATA_LEN(rep) = DATA_LEN(seg);
1894                 PKT_LEN(rep) = PKT_LEN(seg);
1895                 SET_DATA_OFF(rep, DATA_OFF(seg));
1896                 PORT(rep) = PORT(seg);
1897                 (*rxq->elts)[idx] = rep;
1898                 /*
1899                  * Fill NIC descriptor with the new buffer.  The lkey and size
1900                  * of the buffers are already known, only the buffer address
1901                  * changes.
1902                  */
1903                 wqe->addr = rte_cpu_to_be_64(rte_pktmbuf_mtod(rep, uintptr_t));
1904                 if (len > DATA_LEN(seg)) {
1905                         len -= DATA_LEN(seg);
1906                         ++NB_SEGS(pkt);
1907                         ++rq_ci;
1908                         continue;
1909                 }
1910                 DATA_LEN(seg) = len;
1911 #ifdef MLX5_PMD_SOFT_COUNTERS
1912                 /* Increment bytes counter. */
1913                 rxq->stats.ibytes += PKT_LEN(pkt);
1914 #endif
1915                 /* Return packet. */
1916                 *(pkts++) = pkt;
1917                 pkt = NULL;
1918                 --pkts_n;
1919                 ++i;
1920 skip:
1921                 /* Align consumer index to the next stride. */
1922                 rq_ci >>= sges_n;
1923                 ++rq_ci;
1924                 rq_ci <<= sges_n;
1925         }
1926         if (unlikely((i == 0) && ((rq_ci >> sges_n) == rxq->rq_ci)))
1927                 return 0;
1928         /* Update the consumer index. */
1929         rxq->rq_ci = rq_ci >> sges_n;
1930         rte_wmb();
1931         *rxq->cq_db = rte_cpu_to_be_32(rxq->cq_ci);
1932         rte_wmb();
1933         *rxq->rq_db = rte_cpu_to_be_32(rxq->rq_ci);
1934 #ifdef MLX5_PMD_SOFT_COUNTERS
1935         /* Increment packets counter. */
1936         rxq->stats.ipackets += i;
1937 #endif
1938         return i;
1939 }
1940
1941 /**
1942  * Dummy DPDK callback for TX.
1943  *
1944  * This function is used to temporarily replace the real callback during
1945  * unsafe control operations on the queue, or in case of error.
1946  *
1947  * @param dpdk_txq
1948  *   Generic pointer to TX queue structure.
1949  * @param[in] pkts
1950  *   Packets to transmit.
1951  * @param pkts_n
1952  *   Number of packets in array.
1953  *
1954  * @return
1955  *   Number of packets successfully transmitted (<= pkts_n).
1956  */
1957 uint16_t
1958 removed_tx_burst(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
1959 {
1960         (void)dpdk_txq;
1961         (void)pkts;
1962         (void)pkts_n;
1963         return 0;
1964 }
1965
1966 /**
1967  * Dummy DPDK callback for RX.
1968  *
1969  * This function is used to temporarily replace the real callback during
1970  * unsafe control operations on the queue, or in case of error.
1971  *
1972  * @param dpdk_rxq
1973  *   Generic pointer to RX queue structure.
1974  * @param[out] pkts
1975  *   Array to store received packets.
1976  * @param pkts_n
1977  *   Maximum number of packets in array.
1978  *
1979  * @return
1980  *   Number of packets successfully received (<= pkts_n).
1981  */
1982 uint16_t
1983 removed_rx_burst(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
1984 {
1985         (void)dpdk_rxq;
1986         (void)pkts;
1987         (void)pkts_n;
1988         return 0;
1989 }
1990
1991 /*
1992  * Vectorized Rx/Tx routines are not compiled in when required vector
1993  * instructions are not supported on a target architecture. The following null
1994  * stubs are needed for linkage when those are not included outside of this file
1995  * (e.g.  mlx5_rxtx_vec_sse.c for x86).
1996  */
1997
1998 uint16_t __attribute__((weak))
1999 mlx5_tx_burst_raw_vec(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
2000 {
2001         (void)dpdk_txq;
2002         (void)pkts;
2003         (void)pkts_n;
2004         return 0;
2005 }
2006
2007 uint16_t __attribute__((weak))
2008 mlx5_tx_burst_vec(void *dpdk_txq, struct rte_mbuf **pkts, uint16_t pkts_n)
2009 {
2010         (void)dpdk_txq;
2011         (void)pkts;
2012         (void)pkts_n;
2013         return 0;
2014 }
2015
2016 uint16_t __attribute__((weak))
2017 mlx5_rx_burst_vec(void *dpdk_rxq, struct rte_mbuf **pkts, uint16_t pkts_n)
2018 {
2019         (void)dpdk_rxq;
2020         (void)pkts;
2021         (void)pkts_n;
2022         return 0;
2023 }
2024
2025 int __attribute__((weak))
2026 priv_check_raw_vec_tx_support(struct priv *priv)
2027 {
2028         (void)priv;
2029         return -ENOTSUP;
2030 }
2031
2032 int __attribute__((weak))
2033 priv_check_vec_tx_support(struct priv *priv)
2034 {
2035         (void)priv;
2036         return -ENOTSUP;
2037 }
2038
2039 int __attribute__((weak))
2040 rxq_check_vec_support(struct rxq *rxq)
2041 {
2042         (void)rxq;
2043         return -ENOTSUP;
2044 }
2045
2046 int __attribute__((weak))
2047 priv_check_vec_rx_support(struct priv *priv)
2048 {
2049         (void)priv;
2050         return -ENOTSUP;
2051 }