drivers/net/mlx5/mlx5_flow.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright 2016 6WIND S.A.
   3  * Copyright 2016 Mellanox Technologies, Ltd
   4  */
   5
   6 #include <sys/queue.h>
   7 #include <string.h>
   8
   9 /* Verbs header. */
  10 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
  11 #ifdef PEDANTIC
  12 #pragma GCC diagnostic ignored "-Wpedantic"
  13 #endif
  14 #include <infiniband/verbs.h>
  15 #ifdef PEDANTIC
  16 #pragma GCC diagnostic error "-Wpedantic"
  17 #endif
  18
  19 #include <rte_ethdev_driver.h>
  20 #include <rte_flow.h>
  21 #include <rte_flow_driver.h>
  22 #include <rte_malloc.h>
  23 #include <rte_ip.h>
  24
  25 #include "mlx5.h"
  26 #include "mlx5_defs.h"
  27 #include "mlx5_prm.h"
  28 #include "mlx5_glue.h"
  29
  30 /* Define minimal priority for control plane flows. */
  31 #define MLX5_CTRL_FLOW_PRIORITY 4
  32
  33 /* Internet Protocol versions. */
  34 #define MLX5_IPV4 4
  35 #define MLX5_IPV6 6
  36
  37 #ifndef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
  38 struct ibv_flow_spec_counter_action {
  39         int dummy;
  40 };
  41 #endif
  42
  43 /* Dev ops structure defined in mlx5.c */
  44 extern const struct eth_dev_ops mlx5_dev_ops;
  45 extern const struct eth_dev_ops mlx5_dev_ops_isolate;
  46
  47 /** Structure give to the conversion functions. */
  48 struct mlx5_flow_data {
  49         struct mlx5_flow_parse *parser; /** Parser context. */
  50         struct rte_flow_error *error; /** Error context. */
  51 };
  52
  53 static int
  54 mlx5_flow_create_eth(const struct rte_flow_item *item,
  55                      const void *default_mask,
  56                      struct mlx5_flow_data *data);
  57
  58 static int
  59 mlx5_flow_create_vlan(const struct rte_flow_item *item,
  60                       const void *default_mask,
  61                       struct mlx5_flow_data *data);
  62
  63 static int
  64 mlx5_flow_create_ipv4(const struct rte_flow_item *item,
  65                       const void *default_mask,
  66                       struct mlx5_flow_data *data);
  67
  68 static int
  69 mlx5_flow_create_ipv6(const struct rte_flow_item *item,
  70                       const void *default_mask,
  71                       struct mlx5_flow_data *data);
  72
  73 static int
  74 mlx5_flow_create_udp(const struct rte_flow_item *item,
  75                      const void *default_mask,
  76                      struct mlx5_flow_data *data);
  77
  78 static int
  79 mlx5_flow_create_tcp(const struct rte_flow_item *item,
  80                      const void *default_mask,
  81                      struct mlx5_flow_data *data);
  82
  83 static int
  84 mlx5_flow_create_vxlan(const struct rte_flow_item *item,
  85                        const void *default_mask,
  86                        struct mlx5_flow_data *data);
  87
  88 struct mlx5_flow_parse;
  89
  90 static void
  91 mlx5_flow_create_copy(struct mlx5_flow_parse *parser, void *src,
  92                       unsigned int size);
  93
  94 static int
  95 mlx5_flow_create_flag_mark(struct mlx5_flow_parse *parser, uint32_t mark_id);
  96
  97 static int
  98 mlx5_flow_create_count(struct rte_eth_dev *dev, struct mlx5_flow_parse *parser);
  99
 100 /* Hash RX queue types. */
 101 enum hash_rxq_type {
 102         HASH_RXQ_TCPV4,
 103         HASH_RXQ_UDPV4,
 104         HASH_RXQ_IPV4,
 105         HASH_RXQ_TCPV6,
 106         HASH_RXQ_UDPV6,
 107         HASH_RXQ_IPV6,
 108         HASH_RXQ_ETH,
 109 };
 110
 111 /* Initialization data for hash RX queue. */
 112 struct hash_rxq_init {
 113         uint64_t hash_fields; /* Fields that participate in the hash. */
 114         uint64_t dpdk_rss_hf; /* Matching DPDK RSS hash fields. */
 115         unsigned int flow_priority; /* Flow priority to use. */
 116         unsigned int ip_version; /* Internet protocol. */
 117 };
 118
 119 /* Initialization data for hash RX queues. */
 120 const struct hash_rxq_init hash_rxq_init[] = {
 121         [HASH_RXQ_TCPV4] = {
 122                 .hash_fields = (IBV_RX_HASH_SRC_IPV4 |
 123                                 IBV_RX_HASH_DST_IPV4 |
 124                                 IBV_RX_HASH_SRC_PORT_TCP |
 125                                 IBV_RX_HASH_DST_PORT_TCP),
 126                 .dpdk_rss_hf = ETH_RSS_NONFRAG_IPV4_TCP,
 127                 .flow_priority = 1,
 128                 .ip_version = MLX5_IPV4,
 129         },
 130         [HASH_RXQ_UDPV4] = {
 131                 .hash_fields = (IBV_RX_HASH_SRC_IPV4 |
 132                                 IBV_RX_HASH_DST_IPV4 |
 133                                 IBV_RX_HASH_SRC_PORT_UDP |
 134                                 IBV_RX_HASH_DST_PORT_UDP),
 135                 .dpdk_rss_hf = ETH_RSS_NONFRAG_IPV4_UDP,
 136                 .flow_priority = 1,
 137                 .ip_version = MLX5_IPV4,
 138         },
 139         [HASH_RXQ_IPV4] = {
 140                 .hash_fields = (IBV_RX_HASH_SRC_IPV4 |
 141                                 IBV_RX_HASH_DST_IPV4),
 142                 .dpdk_rss_hf = (ETH_RSS_IPV4 |
 143                                 ETH_RSS_FRAG_IPV4),
 144                 .flow_priority = 2,
 145                 .ip_version = MLX5_IPV4,
 146         },
 147         [HASH_RXQ_TCPV6] = {
 148                 .hash_fields = (IBV_RX_HASH_SRC_IPV6 |
 149                                 IBV_RX_HASH_DST_IPV6 |
 150                                 IBV_RX_HASH_SRC_PORT_TCP |
 151                                 IBV_RX_HASH_DST_PORT_TCP),
 152                 .dpdk_rss_hf = ETH_RSS_NONFRAG_IPV6_TCP,
 153                 .flow_priority = 1,
 154                 .ip_version = MLX5_IPV6,
 155         },
 156         [HASH_RXQ_UDPV6] = {
 157                 .hash_fields = (IBV_RX_HASH_SRC_IPV6 |
 158                                 IBV_RX_HASH_DST_IPV6 |
 159                                 IBV_RX_HASH_SRC_PORT_UDP |
 160                                 IBV_RX_HASH_DST_PORT_UDP),
 161                 .dpdk_rss_hf = ETH_RSS_NONFRAG_IPV6_UDP,
 162                 .flow_priority = 1,
 163                 .ip_version = MLX5_IPV6,
 164         },
 165         [HASH_RXQ_IPV6] = {
 166                 .hash_fields = (IBV_RX_HASH_SRC_IPV6 |
 167                                 IBV_RX_HASH_DST_IPV6),
 168                 .dpdk_rss_hf = (ETH_RSS_IPV6 |
 169                                 ETH_RSS_FRAG_IPV6),
 170                 .flow_priority = 2,
 171                 .ip_version = MLX5_IPV6,
 172         },
 173         [HASH_RXQ_ETH] = {
 174                 .hash_fields = 0,
 175                 .dpdk_rss_hf = 0,
 176                 .flow_priority = 3,
 177         },
 178 };
 179
 180 /* Number of entries in hash_rxq_init[]. */
 181 const unsigned int hash_rxq_init_n = RTE_DIM(hash_rxq_init);
 182
 183 /** Structure for holding counter stats. */
 184 struct mlx5_flow_counter_stats {
 185         uint64_t hits; /**< Number of packets matched by the rule. */
 186         uint64_t bytes; /**< Number of bytes matched by the rule. */
 187 };
 188
 189 /** Structure for Drop queue. */
 190 struct mlx5_hrxq_drop {
 191         struct ibv_rwq_ind_table *ind_table; /**< Indirection table. */
 192         struct ibv_qp *qp; /**< Verbs queue pair. */
 193         struct ibv_wq *wq; /**< Verbs work queue. */
 194         struct ibv_cq *cq; /**< Verbs completion queue. */
 195 };
 196
 197 /* Flows structures. */
 198 struct mlx5_flow {
 199         uint64_t hash_fields; /**< Fields that participate in the hash. */
 200         struct ibv_flow_attr *ibv_attr; /**< Pointer to Verbs attributes. */
 201         struct ibv_flow *ibv_flow; /**< Verbs flow. */
 202         struct mlx5_hrxq *hrxq; /**< Hash Rx queues. */
 203 };
 204
 205 /* Drop flows structures. */
 206 struct mlx5_flow_drop {
 207         struct ibv_flow_attr *ibv_attr; /**< Pointer to Verbs attributes. */
 208         struct ibv_flow *ibv_flow; /**< Verbs flow. */
 209 };
 210
 211 struct rte_flow {
 212         TAILQ_ENTRY(rte_flow) next; /**< Pointer to the next flow structure. */
 213         uint32_t mark:1; /**< Set if the flow is marked. */
 214         uint32_t drop:1; /**< Drop queue. */
 215         uint16_t queues_n; /**< Number of entries in queue[]. */
 216         uint16_t (*queues)[]; /**< Queues indexes to use. */
 217         struct rte_eth_rss_conf rss_conf; /**< RSS configuration */
 218         uint8_t rss_key[40]; /**< copy of the RSS key. */
 219         struct ibv_counter_set *cs; /**< Holds the counters for the rule. */
 220         struct mlx5_flow_counter_stats counter_stats;/**<The counter stats. */
 221         struct mlx5_flow frxq[RTE_DIM(hash_rxq_init)];
 222         /**< Flow with Rx queue. */
 223 };
 224
 225 /** Static initializer for items. */
 226 #define ITEMS(...) \
 227         (const enum rte_flow_item_type []){ \
 228                 __VA_ARGS__, RTE_FLOW_ITEM_TYPE_END, \
 229         }
 230
 231 /** Structure to generate a simple graph of layers supported by the NIC. */
 232 struct mlx5_flow_items {
 233         /** List of possible actions for these items. */
 234         const enum rte_flow_action_type *const actions;
 235         /** Bit-masks corresponding to the possibilities for the item. */
 236         const void *mask;
 237         /**
 238          * Default bit-masks to use when item->mask is not provided. When
 239          * \default_mask is also NULL, the full supported bit-mask (\mask) is
 240          * used instead.
 241          */
 242         const void *default_mask;
 243         /** Bit-masks size in bytes. */
 244         const unsigned int mask_sz;
 245         /**
 246          * Conversion function from rte_flow to NIC specific flow.
 247          *
 248          * @param item
 249          *   rte_flow item to convert.
 250          * @param default_mask
 251          *   Default bit-masks to use when item->mask is not provided.
 252          * @param data
 253          *   Internal structure to store the conversion.
 254          *
 255          * @return
 256          *   0 on success, a negative errno value otherwise and rte_errno is
 257          *   set.
 258          */
 259         int (*convert)(const struct rte_flow_item *item,
 260                        const void *default_mask,
 261                        struct mlx5_flow_data *data);
 262         /** Size in bytes of the destination structure. */
 263         const unsigned int dst_sz;
 264         /** List of possible following items.  */
 265         const enum rte_flow_item_type *const items;
 266 };
 267
 268 /** Valid action for this PMD. */
 269 static const enum rte_flow_action_type valid_actions[] = {
 270         RTE_FLOW_ACTION_TYPE_DROP,
 271         RTE_FLOW_ACTION_TYPE_QUEUE,
 272         RTE_FLOW_ACTION_TYPE_MARK,
 273         RTE_FLOW_ACTION_TYPE_FLAG,
 274 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
 275         RTE_FLOW_ACTION_TYPE_COUNT,
 276 #endif
 277         RTE_FLOW_ACTION_TYPE_END,
 278 };
 279
 280 /** Graph of supported items and associated actions. */
 281 static const struct mlx5_flow_items mlx5_flow_items[] = {
 282         [RTE_FLOW_ITEM_TYPE_END] = {
 283                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_ETH,
 284                                RTE_FLOW_ITEM_TYPE_VXLAN),
 285         },
 286         [RTE_FLOW_ITEM_TYPE_ETH] = {
 287                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_VLAN,
 288                                RTE_FLOW_ITEM_TYPE_IPV4,
 289                                RTE_FLOW_ITEM_TYPE_IPV6),
 290                 .actions = valid_actions,
 291                 .mask = &(const struct rte_flow_item_eth){
 292                         .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 293                         .src.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 294                         .type = -1,
 295                 },
 296                 .default_mask = &rte_flow_item_eth_mask,
 297                 .mask_sz = sizeof(struct rte_flow_item_eth),
 298                 .convert = mlx5_flow_create_eth,
 299                 .dst_sz = sizeof(struct ibv_flow_spec_eth),
 300         },
 301         [RTE_FLOW_ITEM_TYPE_VLAN] = {
 302                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_IPV4,
 303                                RTE_FLOW_ITEM_TYPE_IPV6),
 304                 .actions = valid_actions,
 305                 .mask = &(const struct rte_flow_item_vlan){
 306                         .tci = -1,
 307                 },
 308                 .default_mask = &rte_flow_item_vlan_mask,
 309                 .mask_sz = sizeof(struct rte_flow_item_vlan),
 310                 .convert = mlx5_flow_create_vlan,
 311                 .dst_sz = 0,
 312         },
 313         [RTE_FLOW_ITEM_TYPE_IPV4] = {
 314                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP,
 315                                RTE_FLOW_ITEM_TYPE_TCP),
 316                 .actions = valid_actions,
 317                 .mask = &(const struct rte_flow_item_ipv4){
 318                         .hdr = {
 319                                 .src_addr = -1,
 320                                 .dst_addr = -1,
 321                                 .type_of_service = -1,
 322                                 .next_proto_id = -1,
 323                         },
 324                 },
 325                 .default_mask = &rte_flow_item_ipv4_mask,
 326                 .mask_sz = sizeof(struct rte_flow_item_ipv4),
 327                 .convert = mlx5_flow_create_ipv4,
 328                 .dst_sz = sizeof(struct ibv_flow_spec_ipv4_ext),
 329         },
 330         [RTE_FLOW_ITEM_TYPE_IPV6] = {
 331                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP,
 332                                RTE_FLOW_ITEM_TYPE_TCP),
 333                 .actions = valid_actions,
 334                 .mask = &(const struct rte_flow_item_ipv6){
 335                         .hdr = {
 336                                 .src_addr = {
 337                                         0xff, 0xff, 0xff, 0xff,
 338                                         0xff, 0xff, 0xff, 0xff,
 339                                         0xff, 0xff, 0xff, 0xff,
 340                                         0xff, 0xff, 0xff, 0xff,
 341                                 },
 342                                 .dst_addr = {
 343                                         0xff, 0xff, 0xff, 0xff,
 344                                         0xff, 0xff, 0xff, 0xff,
 345                                         0xff, 0xff, 0xff, 0xff,
 346                                         0xff, 0xff, 0xff, 0xff,
 347                                 },
 348                                 .vtc_flow = -1,
 349                                 .proto = -1,
 350                                 .hop_limits = -1,
 351                         },
 352                 },
 353                 .default_mask = &rte_flow_item_ipv6_mask,
 354                 .mask_sz = sizeof(struct rte_flow_item_ipv6),
 355                 .convert = mlx5_flow_create_ipv6,
 356                 .dst_sz = sizeof(struct ibv_flow_spec_ipv6),
 357         },
 358         [RTE_FLOW_ITEM_TYPE_UDP] = {
 359                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_VXLAN),
 360                 .actions = valid_actions,
 361                 .mask = &(const struct rte_flow_item_udp){
 362                         .hdr = {
 363                                 .src_port = -1,
 364                                 .dst_port = -1,
 365                         },
 366                 },
 367                 .default_mask = &rte_flow_item_udp_mask,
 368                 .mask_sz = sizeof(struct rte_flow_item_udp),
 369                 .convert = mlx5_flow_create_udp,
 370                 .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
 371         },
 372         [RTE_FLOW_ITEM_TYPE_TCP] = {
 373                 .actions = valid_actions,
 374                 .mask = &(const struct rte_flow_item_tcp){
 375                         .hdr = {
 376                                 .src_port = -1,
 377                                 .dst_port = -1,
 378                         },
 379                 },
 380                 .default_mask = &rte_flow_item_tcp_mask,
 381                 .mask_sz = sizeof(struct rte_flow_item_tcp),
 382                 .convert = mlx5_flow_create_tcp,
 383                 .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
 384         },
 385         [RTE_FLOW_ITEM_TYPE_VXLAN] = {
 386                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_ETH),
 387                 .actions = valid_actions,
 388                 .mask = &(const struct rte_flow_item_vxlan){
 389                         .vni = "\xff\xff\xff",
 390                 },
 391                 .default_mask = &rte_flow_item_vxlan_mask,
 392                 .mask_sz = sizeof(struct rte_flow_item_vxlan),
 393                 .convert = mlx5_flow_create_vxlan,
 394                 .dst_sz = sizeof(struct ibv_flow_spec_tunnel),
 395         },
 396 };
 397
 398 /** Structure to pass to the conversion function. */
 399 struct mlx5_flow_parse {
 400         uint32_t inner; /**< Set once VXLAN is encountered. */
 401         uint32_t create:1;
 402         /**< Whether resources should remain after a validate. */
 403         uint32_t drop:1; /**< Target is a drop queue. */
 404         uint32_t mark:1; /**< Mark is present in the flow. */
 405         uint32_t count:1; /**< Count is present in the flow. */
 406         uint32_t mark_id; /**< Mark identifier. */
 407         uint16_t queues[RTE_MAX_QUEUES_PER_PORT]; /**< Queues indexes to use. */
 408         uint16_t queues_n; /**< Number of entries in queue[]. */
 409         struct rte_eth_rss_conf rss_conf; /**< RSS configuration */
 410         uint8_t rss_key[40]; /**< copy of the RSS key. */
 411         enum hash_rxq_type layer; /**< Last pattern layer detected. */
 412         struct ibv_counter_set *cs; /**< Holds the counter set for the rule */
 413         struct {
 414                 struct ibv_flow_attr *ibv_attr;
 415                 /**< Pointer to Verbs attributes. */
 416                 unsigned int offset;
 417                 /**< Current position or total size of the attribute. */
 418         } queue[RTE_DIM(hash_rxq_init)];
 419 };
 420
 421 static const struct rte_flow_ops mlx5_flow_ops = {
 422         .validate = mlx5_flow_validate,
 423         .create = mlx5_flow_create,
 424         .destroy = mlx5_flow_destroy,
 425         .flush = mlx5_flow_flush,
 426 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
 427         .query = mlx5_flow_query,
 428 #else
 429         .query = NULL,
 430 #endif
 431         .isolate = mlx5_flow_isolate,
 432 };
 433
 434 /* Convert FDIR request to Generic flow. */
 435 struct mlx5_fdir {
 436         struct rte_flow_attr attr;
 437         struct rte_flow_action actions[2];
 438         struct rte_flow_item items[4];
 439         struct rte_flow_item_eth l2;
 440         struct rte_flow_item_eth l2_mask;
 441         union {
 442                 struct rte_flow_item_ipv4 ipv4;
 443                 struct rte_flow_item_ipv6 ipv6;
 444         } l3;
 445         union {
 446                 struct rte_flow_item_udp udp;
 447                 struct rte_flow_item_tcp tcp;
 448         } l4;
 449         struct rte_flow_action_queue queue;
 450 };
 451
 452 /* Verbs specification header. */
 453 struct ibv_spec_header {
 454         enum ibv_flow_spec_type type;
 455         uint16_t size;
 456 };
 457
 458 /**
 459  * Check support for a given item.
 460  *
 461  * @param item[in]
 462  *   Item specification.
 463  * @param mask[in]
 464  *   Bit-masks covering supported fields to compare with spec, last and mask in
 465  *   \item.
 466  * @param size
 467  *   Bit-Mask size in bytes.
 468  *
 469  * @return
 470  *   0 on success, a negative errno value otherwise and rte_errno is set.
 471  */
 472 static int
 473 mlx5_flow_item_validate(const struct rte_flow_item *item,
 474                         const uint8_t *mask, unsigned int size)
 475 {
 476         if (!item->spec && (item->mask || item->last)) {
 477                 rte_errno = EINVAL;
 478                 return -rte_errno;
 479         }
 480         if (item->spec && !item->mask) {
 481                 unsigned int i;
 482                 const uint8_t *spec = item->spec;
 483
 484                 for (i = 0; i < size; ++i)
 485                         if ((spec[i] | mask[i]) != mask[i]) {
 486                                 rte_errno = EINVAL;
 487                                 return -rte_errno;
 488                         }
 489         }
 490         if (item->last && !item->mask) {
 491                 unsigned int i;
 492                 const uint8_t *spec = item->last;
 493
 494                 for (i = 0; i < size; ++i)
 495                         if ((spec[i] | mask[i]) != mask[i]) {
 496                                 rte_errno = EINVAL;
 497                                 return -rte_errno;
 498                         }
 499         }
 500         if (item->mask) {
 501                 unsigned int i;
 502                 const uint8_t *spec = item->spec;
 503
 504                 for (i = 0; i < size; ++i)
 505                         if ((spec[i] | mask[i]) != mask[i]) {
 506                                 rte_errno = EINVAL;
 507                                 return -rte_errno;
 508                         }
 509         }
 510         if (item->spec && item->last) {
 511                 uint8_t spec[size];
 512                 uint8_t last[size];
 513                 const uint8_t *apply = mask;
 514                 unsigned int i;
 515                 int ret;
 516
 517                 if (item->mask)
 518                         apply = item->mask;
 519                 for (i = 0; i < size; ++i) {
 520                         spec[i] = ((const uint8_t *)item->spec)[i] & apply[i];
 521                         last[i] = ((const uint8_t *)item->last)[i] & apply[i];
 522                 }
 523                 ret = memcmp(spec, last, size);
 524                 if (ret != 0) {
 525                         rte_errno = EINVAL;
 526                         return -rte_errno;
 527                 }
 528         }
 529         return 0;
 530 }
 531
 532 /**
 533  * Copy the RSS configuration from the user ones, of the rss_conf is null,
 534  * uses the driver one.
 535  *
 536  * @param parser
 537  *   Internal parser structure.
 538  * @param rss_conf
 539  *   User RSS configuration to save.
 540  *
 541  * @return
 542  *   0 on success, a negative errno value otherwise and rte_errno is set.
 543  */
 544 static int
 545 mlx5_flow_convert_rss_conf(struct mlx5_flow_parse *parser,
 546                            const struct rte_eth_rss_conf *rss_conf)
 547 {
 548         /*
 549          * This function is also called at the beginning of
 550          * mlx5_flow_convert_actions() to initialize the parser with the
 551          * device default RSS configuration.
 552          */
 553         if (rss_conf) {
 554                 if (rss_conf->rss_hf & MLX5_RSS_HF_MASK) {
 555                         rte_errno = EINVAL;
 556                         return -rte_errno;
 557                 }
 558                 if (rss_conf->rss_key_len != 40) {
 559                         rte_errno = EINVAL;
 560                         return -rte_errno;
 561                 }
 562                 if (rss_conf->rss_key_len && rss_conf->rss_key) {
 563                         parser->rss_conf.rss_key_len = rss_conf->rss_key_len;
 564                         memcpy(parser->rss_key, rss_conf->rss_key,
 565                                rss_conf->rss_key_len);
 566                         parser->rss_conf.rss_key = parser->rss_key;
 567                 }
 568                 parser->rss_conf.rss_hf = rss_conf->rss_hf;
 569         }
 570         return 0;
 571 }
 572
 573 /**
 574  * Extract attribute to the parser.
 575  *
 576  * @param[in] attr
 577  *   Flow rule attributes.
 578  * @param[out] error
 579  *   Perform verbose error reporting if not NULL.
 580  *
 581  * @return
 582  *   0 on success, a negative errno value otherwise and rte_errno is set.
 583  */
 584 static int
 585 mlx5_flow_convert_attributes(const struct rte_flow_attr *attr,
 586                              struct rte_flow_error *error)
 587 {
 588         if (attr->group) {
 589                 rte_flow_error_set(error, ENOTSUP,
 590                                    RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
 591                                    NULL,
 592                                    "groups are not supported");
 593                 return -rte_errno;
 594         }
 595         if (attr->priority && attr->priority != MLX5_CTRL_FLOW_PRIORITY) {
 596                 rte_flow_error_set(error, ENOTSUP,
 597                                    RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
 598                                    NULL,
 599                                    "priorities are not supported");
 600                 return -rte_errno;
 601         }
 602         if (attr->egress) {
 603                 rte_flow_error_set(error, ENOTSUP,
 604                                    RTE_FLOW_ERROR_TYPE_ATTR_EGRESS,
 605                                    NULL,
 606                                    "egress is not supported");
 607                 return -rte_errno;
 608         }
 609         if (!attr->ingress) {
 610                 rte_flow_error_set(error, ENOTSUP,
 611                                    RTE_FLOW_ERROR_TYPE_ATTR_INGRESS,
 612                                    NULL,
 613                                    "only ingress is supported");
 614                 return -rte_errno;
 615         }
 616         return 0;
 617 }
 618
 619 /**
 620  * Extract actions request to the parser.
 621  *
 622  * @param dev
 623  *   Pointer to Ethernet device.
 624  * @param[in] actions
 625  *   Associated actions (list terminated by the END action).
 626  * @param[out] error
 627  *   Perform verbose error reporting if not NULL.
 628  * @param[in, out] parser
 629  *   Internal parser structure.
 630  *
 631  * @return
 632  *   0 on success, a negative errno value otherwise and rte_errno is set.
 633  */
 634 static int
 635 mlx5_flow_convert_actions(struct rte_eth_dev *dev,
 636                           const struct rte_flow_action actions[],
 637                           struct rte_flow_error *error,
 638                           struct mlx5_flow_parse *parser)
 639 {
 640         struct priv *priv = dev->data->dev_private;
 641         int ret;
 642
 643         /*
 644          * Add default RSS configuration necessary for Verbs to create QP even
 645          * if no RSS is necessary.
 646          */
 647         ret = mlx5_flow_convert_rss_conf(parser,
 648                                          (const struct rte_eth_rss_conf *)
 649                                          &priv->rss_conf);
 650         if (ret)
 651                 return ret;
 652         for (; actions->type != RTE_FLOW_ACTION_TYPE_END; ++actions) {
 653                 if (actions->type == RTE_FLOW_ACTION_TYPE_VOID) {
 654                         continue;
 655                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_DROP) {
 656                         parser->drop = 1;
 657                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_QUEUE) {
 658                         const struct rte_flow_action_queue *queue =
 659                                 (const struct rte_flow_action_queue *)
 660                                 actions->conf;
 661                         uint16_t n;
 662                         uint16_t found = 0;
 663
 664                         if (!queue || (queue->index > (priv->rxqs_n - 1)))
 665                                 goto exit_action_not_supported;
 666                         for (n = 0; n < parser->queues_n; ++n) {
 667                                 if (parser->queues[n] == queue->index) {
 668                                         found = 1;
 669                                         break;
 670                                 }
 671                         }
 672                         if (parser->queues_n > 1 && !found) {
 673                                 rte_flow_error_set(error, ENOTSUP,
 674                                            RTE_FLOW_ERROR_TYPE_ACTION,
 675                                            actions,
 676                                            "queue action not in RSS queues");
 677                                 return -rte_errno;
 678                         }
 679                         if (!found) {
 680                                 parser->queues_n = 1;
 681                                 parser->queues[0] = queue->index;
 682                         }
 683                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_RSS) {
 684                         const struct rte_flow_action_rss *rss =
 685                                 (const struct rte_flow_action_rss *)
 686                                 actions->conf;
 687                         uint16_t n;
 688
 689                         if (!rss || !rss->num) {
 690                                 rte_flow_error_set(error, EINVAL,
 691                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 692                                                    actions,
 693                                                    "no valid queues");
 694                                 return -rte_errno;
 695                         }
 696                         if (parser->queues_n == 1) {
 697                                 uint16_t found = 0;
 698
 699                                 assert(parser->queues_n);
 700                                 for (n = 0; n < rss->num; ++n) {
 701                                         if (parser->queues[0] ==
 702                                             rss->queue[n]) {
 703                                                 found = 1;
 704                                                 break;
 705                                         }
 706                                 }
 707                                 if (!found) {
 708                                         rte_flow_error_set(error, ENOTSUP,
 709                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 710                                                    actions,
 711                                                    "queue action not in RSS"
 712                                                    " queues");
 713                                         return -rte_errno;
 714                                 }
 715                         }
 716                         for (n = 0; n < rss->num; ++n) {
 717                                 if (rss->queue[n] >= priv->rxqs_n) {
 718                                         rte_flow_error_set(error, EINVAL,
 719                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 720                                                    actions,
 721                                                    "queue id > number of"
 722                                                    " queues");
 723                                         return -rte_errno;
 724                                 }
 725                         }
 726                         for (n = 0; n < rss->num; ++n)
 727                                 parser->queues[n] = rss->queue[n];
 728                         parser->queues_n = rss->num;
 729                         if (mlx5_flow_convert_rss_conf(parser, rss->rss_conf)) {
 730                                 rte_flow_error_set(error, EINVAL,
 731                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 732                                                    actions,
 733                                                    "wrong RSS configuration");
 734                                 return -rte_errno;
 735                         }
 736                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_MARK) {
 737                         const struct rte_flow_action_mark *mark =
 738                                 (const struct rte_flow_action_mark *)
 739                                 actions->conf;
 740
 741                         if (!mark) {
 742                                 rte_flow_error_set(error, EINVAL,
 743                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 744                                                    actions,
 745                                                    "mark must be defined");
 746                                 return -rte_errno;
 747                         } else if (mark->id >= MLX5_FLOW_MARK_MAX) {
 748                                 rte_flow_error_set(error, ENOTSUP,
 749                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 750                                                    actions,
 751                                                    "mark must be between 0"
 752                                                    " and 16777199");
 753                                 return -rte_errno;
 754                         }
 755                         parser->mark = 1;
 756                         parser->mark_id = mark->id;
 757                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_FLAG) {
 758                         parser->mark = 1;
 759                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_COUNT &&
 760                            priv->config.flow_counter_en) {
 761                         parser->count = 1;
 762                 } else {
 763                         goto exit_action_not_supported;
 764                 }
 765         }
 766         if (parser->drop && parser->mark)
 767                 parser->mark = 0;
 768         if (!parser->queues_n && !parser->drop) {
 769                 rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_HANDLE,
 770                                    NULL, "no valid action");
 771                 return -rte_errno;
 772         }
 773         return 0;
 774 exit_action_not_supported:
 775         rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
 776                            actions, "action not supported");
 777         return -rte_errno;
 778 }
 779
 780 /**
 781  * Validate items.
 782  *
 783  * @param[in] items
 784  *   Pattern specification (list terminated by the END pattern item).
 785  * @param[out] error
 786  *   Perform verbose error reporting if not NULL.
 787  * @param[in, out] parser
 788  *   Internal parser structure.
 789  *
 790  * @return
 791  *   0 on success, a negative errno value otherwise and rte_errno is set.
 792  */
 793 static int
 794 mlx5_flow_convert_items_validate(const struct rte_flow_item items[],
 795                                  struct rte_flow_error *error,
 796                                  struct mlx5_flow_parse *parser)
 797 {
 798         const struct mlx5_flow_items *cur_item = mlx5_flow_items;
 799         unsigned int i;
 800         int ret = 0;
 801
 802         /* Initialise the offsets to start after verbs attribute. */
 803         for (i = 0; i != hash_rxq_init_n; ++i)
 804                 parser->queue[i].offset = sizeof(struct ibv_flow_attr);
 805         for (; items->type != RTE_FLOW_ITEM_TYPE_END; ++items) {
 806                 const struct mlx5_flow_items *token = NULL;
 807                 unsigned int n;
 808
 809                 if (items->type == RTE_FLOW_ITEM_TYPE_VOID)
 810                         continue;
 811                 for (i = 0;
 812                      cur_item->items &&
 813                      cur_item->items[i] != RTE_FLOW_ITEM_TYPE_END;
 814                      ++i) {
 815                         if (cur_item->items[i] == items->type) {
 816                                 token = &mlx5_flow_items[items->type];
 817                                 break;
 818                         }
 819                 }
 820                 if (!token)
 821                         goto exit_item_not_supported;
 822                 cur_item = token;
 823                 ret = mlx5_flow_item_validate(items,
 824                                               (const uint8_t *)cur_item->mask,
 825                                               cur_item->mask_sz);
 826                 if (ret)
 827                         goto exit_item_not_supported;
 828                 if (items->type == RTE_FLOW_ITEM_TYPE_VXLAN) {
 829                         if (parser->inner) {
 830                                 rte_flow_error_set(error, ENOTSUP,
 831                                                    RTE_FLOW_ERROR_TYPE_ITEM,
 832                                                    items,
 833                                                    "cannot recognize multiple"
 834                                                    " VXLAN encapsulations");
 835                                 return -rte_errno;
 836                         }
 837                         parser->inner = IBV_FLOW_SPEC_INNER;
 838                 }
 839                 if (parser->drop) {
 840                         parser->queue[HASH_RXQ_ETH].offset += cur_item->dst_sz;
 841                 } else {
 842                         for (n = 0; n != hash_rxq_init_n; ++n)
 843                                 parser->queue[n].offset += cur_item->dst_sz;
 844                 }
 845         }
 846         if (parser->drop) {
 847                 parser->queue[HASH_RXQ_ETH].offset +=
 848                         sizeof(struct ibv_flow_spec_action_drop);
 849         }
 850         if (parser->mark) {
 851                 for (i = 0; i != hash_rxq_init_n; ++i)
 852                         parser->queue[i].offset +=
 853                                 sizeof(struct ibv_flow_spec_action_tag);
 854         }
 855         if (parser->count) {
 856                 unsigned int size = sizeof(struct ibv_flow_spec_counter_action);
 857
 858                 for (i = 0; i != hash_rxq_init_n; ++i)
 859                         parser->queue[i].offset += size;
 860         }
 861         return 0;
 862 exit_item_not_supported:
 863         return rte_flow_error_set(error, -ret, RTE_FLOW_ERROR_TYPE_ITEM,
 864                                   items, "item not supported");
 865 }
 866
 867 /**
 868  * Allocate memory space to store verbs flow attributes.
 869  *
 870  * @param[in] size
 871  *   Amount of byte to allocate.
 872  * @param[out] error
 873  *   Perform verbose error reporting if not NULL.
 874  *
 875  * @return
 876  *   A verbs flow attribute on success, NULL otherwise and rte_errno is set.
 877  */
 878 static struct ibv_flow_attr *
 879 mlx5_flow_convert_allocate(unsigned int size, struct rte_flow_error *error)
 880 {
 881         struct ibv_flow_attr *ibv_attr;
 882
 883         ibv_attr = rte_calloc(__func__, 1, size, 0);
 884         if (!ibv_attr) {
 885                 rte_flow_error_set(error, ENOMEM,
 886                                    RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
 887                                    NULL,
 888                                    "cannot allocate verbs spec attributes");
 889                 return NULL;
 890         }
 891         return ibv_attr;
 892 }
 893
 894 /**
 895  * Make inner packet matching with an higher priority from the non Inner
 896  * matching.
 897  *
 898  * @param[in, out] parser
 899  *   Internal parser structure.
 900  * @param attr
 901  *   User flow attribute.
 902  */
 903 static void
 904 mlx5_flow_update_priority(struct mlx5_flow_parse *parser,
 905                           const struct rte_flow_attr *attr)
 906 {
 907         unsigned int i;
 908
 909         if (parser->drop) {
 910                 parser->queue[HASH_RXQ_ETH].ibv_attr->priority =
 911                         attr->priority +
 912                         hash_rxq_init[HASH_RXQ_ETH].flow_priority;
 913                 return;
 914         }
 915         for (i = 0; i != hash_rxq_init_n; ++i) {
 916                 if (parser->queue[i].ibv_attr) {
 917                         parser->queue[i].ibv_attr->priority =
 918                                 attr->priority +
 919                                 hash_rxq_init[i].flow_priority -
 920                                 (parser->inner ? 1 : 0);
 921                 }
 922         }
 923 }
 924
 925 /**
 926  * Finalise verbs flow attributes.
 927  *
 928  * @param[in, out] parser
 929  *   Internal parser structure.
 930  */
 931 static void
 932 mlx5_flow_convert_finalise(struct mlx5_flow_parse *parser)
 933 {
 934         const unsigned int ipv4 =
 935                 hash_rxq_init[parser->layer].ip_version == MLX5_IPV4;
 936         const enum hash_rxq_type hmin = ipv4 ? HASH_RXQ_TCPV4 : HASH_RXQ_TCPV6;
 937         const enum hash_rxq_type hmax = ipv4 ? HASH_RXQ_IPV4 : HASH_RXQ_IPV6;
 938         const enum hash_rxq_type ohmin = ipv4 ? HASH_RXQ_TCPV6 : HASH_RXQ_TCPV4;
 939         const enum hash_rxq_type ohmax = ipv4 ? HASH_RXQ_IPV6 : HASH_RXQ_IPV4;
 940         const enum hash_rxq_type ip = ipv4 ? HASH_RXQ_IPV4 : HASH_RXQ_IPV6;
 941         unsigned int i;
 942
 943         /* Remove any other flow not matching the pattern. */
 944         if (parser->queues_n == 1 && !parser->rss_conf.rss_hf) {
 945                 for (i = 0; i != hash_rxq_init_n; ++i) {
 946                         if (i == HASH_RXQ_ETH)
 947                                 continue;
 948                         rte_free(parser->queue[i].ibv_attr);
 949                         parser->queue[i].ibv_attr = NULL;
 950                 }
 951                 return;
 952         }
 953         if (parser->layer == HASH_RXQ_ETH) {
 954                 goto fill;
 955         } else {
 956                 /*
 957                  * This layer becomes useless as the pattern define under
 958                  * layers.
 959                  */
 960                 rte_free(parser->queue[HASH_RXQ_ETH].ibv_attr);
 961                 parser->queue[HASH_RXQ_ETH].ibv_attr = NULL;
 962         }
 963         /* Remove opposite kind of layer e.g. IPv6 if the pattern is IPv4. */
 964         for (i = ohmin; i != (ohmax + 1); ++i) {
 965                 if (!parser->queue[i].ibv_attr)
 966                         continue;
 967                 rte_free(parser->queue[i].ibv_attr);
 968                 parser->queue[i].ibv_attr = NULL;
 969         }
 970         /* Remove impossible flow according to the RSS configuration. */
 971         if (hash_rxq_init[parser->layer].dpdk_rss_hf &
 972             parser->rss_conf.rss_hf) {
 973                 /* Remove any other flow. */
 974                 for (i = hmin; i != (hmax + 1); ++i) {
 975                         if ((i == parser->layer) ||
 976                              (!parser->queue[i].ibv_attr))
 977                                 continue;
 978                         rte_free(parser->queue[i].ibv_attr);
 979                         parser->queue[i].ibv_attr = NULL;
 980                 }
 981         } else  if (!parser->queue[ip].ibv_attr) {
 982                 /* no RSS possible with the current configuration. */
 983                 parser->queues_n = 1;
 984                 return;
 985         }
 986 fill:
 987         /*
 988          * Fill missing layers in verbs specifications, or compute the correct
 989          * offset to allocate the memory space for the attributes and
 990          * specifications.
 991          */
 992         for (i = 0; i != hash_rxq_init_n - 1; ++i) {
 993                 union {
 994                         struct ibv_flow_spec_ipv4_ext ipv4;
 995                         struct ibv_flow_spec_ipv6 ipv6;
 996                         struct ibv_flow_spec_tcp_udp udp_tcp;
 997                 } specs;
 998                 void *dst;
 999                 uint16_t size;
1000
1001                 if (i == parser->layer)
1002                         continue;
1003                 if (parser->layer == HASH_RXQ_ETH) {
1004                         if (hash_rxq_init[i].ip_version == MLX5_IPV4) {
1005                                 size = sizeof(struct ibv_flow_spec_ipv4_ext);
1006                                 specs.ipv4 = (struct ibv_flow_spec_ipv4_ext){
1007                                         .type = IBV_FLOW_SPEC_IPV4_EXT,
1008                                         .size = size,
1009                                 };
1010                         } else {
1011                                 size = sizeof(struct ibv_flow_spec_ipv6);
1012                                 specs.ipv6 = (struct ibv_flow_spec_ipv6){
1013                                         .type = IBV_FLOW_SPEC_IPV6,
1014                                         .size = size,
1015                                 };
1016                         }
1017                         if (parser->queue[i].ibv_attr) {
1018                                 dst = (void *)((uintptr_t)
1019                                                parser->queue[i].ibv_attr +
1020                                                parser->queue[i].offset);
1021                                 memcpy(dst, &specs, size);
1022                                 ++parser->queue[i].ibv_attr->num_of_specs;
1023                         }
1024                         parser->queue[i].offset += size;
1025                 }
1026                 if ((i == HASH_RXQ_UDPV4) || (i == HASH_RXQ_TCPV4) ||
1027                     (i == HASH_RXQ_UDPV6) || (i == HASH_RXQ_TCPV6)) {
1028                         size = sizeof(struct ibv_flow_spec_tcp_udp);
1029                         specs.udp_tcp = (struct ibv_flow_spec_tcp_udp) {
1030                                 .type = ((i == HASH_RXQ_UDPV4 ||
1031                                           i == HASH_RXQ_UDPV6) ?
1032                                          IBV_FLOW_SPEC_UDP :
1033                                          IBV_FLOW_SPEC_TCP),
1034                                 .size = size,
1035                         };
1036                         if (parser->queue[i].ibv_attr) {
1037                                 dst = (void *)((uintptr_t)
1038                                                parser->queue[i].ibv_attr +
1039                                                parser->queue[i].offset);
1040                                 memcpy(dst, &specs, size);
1041                                 ++parser->queue[i].ibv_attr->num_of_specs;
1042                         }
1043                         parser->queue[i].offset += size;
1044                 }
1045         }
1046 }
1047
1048 /**
1049  * Validate and convert a flow supported by the NIC.
1050  *
1051  * @param dev
1052  *   Pointer to Ethernet device.
1053  * @param[in] attr
1054  *   Flow rule attributes.
1055  * @param[in] pattern
1056  *   Pattern specification (list terminated by the END pattern item).
1057  * @param[in] actions
1058  *   Associated actions (list terminated by the END action).
1059  * @param[out] error
1060  *   Perform verbose error reporting if not NULL.
1061  * @param[in, out] parser
1062  *   Internal parser structure.
1063  *
1064  * @return
1065  *   0 on success, a negative errno value otherwise and rte_errno is set.
1066  */
1067 static int
1068 mlx5_flow_convert(struct rte_eth_dev *dev,
1069                   const struct rte_flow_attr *attr,
1070                   const struct rte_flow_item items[],
1071                   const struct rte_flow_action actions[],
1072                   struct rte_flow_error *error,
1073                   struct mlx5_flow_parse *parser)
1074 {
1075         const struct mlx5_flow_items *cur_item = mlx5_flow_items;
1076         unsigned int i;
1077         int ret;
1078
1079         /* First step. Validate the attributes, items and actions. */
1080         *parser = (struct mlx5_flow_parse){
1081                 .create = parser->create,
1082                 .layer = HASH_RXQ_ETH,
1083                 .mark_id = MLX5_FLOW_MARK_DEFAULT,
1084         };
1085         ret = mlx5_flow_convert_attributes(attr, error);
1086         if (ret)
1087                 return ret;
1088         ret = mlx5_flow_convert_actions(dev, actions, error, parser);
1089         if (ret)
1090                 return ret;
1091         ret = mlx5_flow_convert_items_validate(items, error, parser);
1092         if (ret)
1093                 return ret;
1094         mlx5_flow_convert_finalise(parser);
1095         /*
1096          * Second step.
1097          * Allocate the memory space to store verbs specifications.
1098          */
1099         if (parser->drop) {
1100                 unsigned int offset = parser->queue[HASH_RXQ_ETH].offset;
1101
1102                 parser->queue[HASH_RXQ_ETH].ibv_attr =
1103                         mlx5_flow_convert_allocate(offset, error);
1104                 if (!parser->queue[HASH_RXQ_ETH].ibv_attr)
1105                         goto exit_enomem;
1106                 parser->queue[HASH_RXQ_ETH].offset =
1107                         sizeof(struct ibv_flow_attr);
1108         } else {
1109                 for (i = 0; i != hash_rxq_init_n; ++i) {
1110                         unsigned int offset;
1111
1112                         if (!(parser->rss_conf.rss_hf &
1113                               hash_rxq_init[i].dpdk_rss_hf) &&
1114                             (i != HASH_RXQ_ETH))
1115                                 continue;
1116                         offset = parser->queue[i].offset;
1117                         parser->queue[i].ibv_attr =
1118                                 mlx5_flow_convert_allocate(offset, error);
1119                         if (!parser->queue[i].ibv_attr)
1120                                 goto exit_enomem;
1121                         parser->queue[i].offset = sizeof(struct ibv_flow_attr);
1122                 }
1123         }
1124         /* Third step. Conversion parse, fill the specifications. */
1125         parser->inner = 0;
1126         for (; items->type != RTE_FLOW_ITEM_TYPE_END; ++items) {
1127                 struct mlx5_flow_data data = {
1128                         .parser = parser,
1129                         .error = error,
1130                 };
1131
1132                 if (items->type == RTE_FLOW_ITEM_TYPE_VOID)
1133                         continue;
1134                 cur_item = &mlx5_flow_items[items->type];
1135                 ret = cur_item->convert(items,
1136                                         (cur_item->default_mask ?
1137                                          cur_item->default_mask :
1138                                          cur_item->mask),
1139                                          &data);
1140                 if (ret)
1141                         goto exit_free;
1142         }
1143         if (parser->mark)
1144                 mlx5_flow_create_flag_mark(parser, parser->mark_id);
1145         if (parser->count && parser->create) {
1146                 mlx5_flow_create_count(dev, parser);
1147                 if (!parser->cs)
1148                         goto exit_count_error;
1149         }
1150         /*
1151          * Last step. Complete missing specification to reach the RSS
1152          * configuration.
1153          */
1154         if (!parser->drop)
1155                 mlx5_flow_convert_finalise(parser);
1156         mlx5_flow_update_priority(parser, attr);
1157 exit_free:
1158         /* Only verification is expected, all resources should be released. */
1159         if (!parser->create) {
1160                 for (i = 0; i != hash_rxq_init_n; ++i) {
1161                         if (parser->queue[i].ibv_attr) {
1162                                 rte_free(parser->queue[i].ibv_attr);
1163                                 parser->queue[i].ibv_attr = NULL;
1164                         }
1165                 }
1166         }
1167         return ret;
1168 exit_enomem:
1169         for (i = 0; i != hash_rxq_init_n; ++i) {
1170                 if (parser->queue[i].ibv_attr) {
1171                         rte_free(parser->queue[i].ibv_attr);
1172                         parser->queue[i].ibv_attr = NULL;
1173                 }
1174         }
1175         rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
1176                            NULL, "cannot allocate verbs spec attributes");
1177         return -rte_errno;
1178 exit_count_error:
1179         rte_flow_error_set(error, EINVAL, RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
1180                            NULL, "cannot create counter");
1181         return -rte_errno;
1182 }
1183
1184 /**
1185  * Copy the specification created into the flow.
1186  *
1187  * @param parser
1188  *   Internal parser structure.
1189  * @param src
1190  *   Create specification.
1191  * @param size
1192  *   Size in bytes of the specification to copy.
1193  */
1194 static void
1195 mlx5_flow_create_copy(struct mlx5_flow_parse *parser, void *src,
1196                       unsigned int size)
1197 {
1198         unsigned int i;
1199         void *dst;
1200
1201         for (i = 0; i != hash_rxq_init_n; ++i) {
1202                 if (!parser->queue[i].ibv_attr)
1203                         continue;
1204                 /* Specification must be the same l3 type or none. */
1205                 if (parser->layer == HASH_RXQ_ETH ||
1206                     (hash_rxq_init[parser->layer].ip_version ==
1207                      hash_rxq_init[i].ip_version) ||
1208                     (hash_rxq_init[i].ip_version == 0)) {
1209                         dst = (void *)((uintptr_t)parser->queue[i].ibv_attr +
1210                                         parser->queue[i].offset);
1211                         memcpy(dst, src, size);
1212                         ++parser->queue[i].ibv_attr->num_of_specs;
1213                         parser->queue[i].offset += size;
1214                 }
1215         }
1216 }
1217
1218 /**
1219  * Convert Ethernet item to Verbs specification.
1220  *
1221  * @param item[in]
1222  *   Item specification.
1223  * @param default_mask[in]
1224  *   Default bit-masks to use when item->mask is not provided.
1225  * @param data[in, out]
1226  *   User structure.
1227  *
1228  * @return
1229  *   0 on success, a negative errno value otherwise and rte_errno is set.
1230  */
1231 static int
1232 mlx5_flow_create_eth(const struct rte_flow_item *item,
1233                      const void *default_mask,
1234                      struct mlx5_flow_data *data)
1235 {
1236         const struct rte_flow_item_eth *spec = item->spec;
1237         const struct rte_flow_item_eth *mask = item->mask;
1238         struct mlx5_flow_parse *parser = data->parser;
1239         const unsigned int eth_size = sizeof(struct ibv_flow_spec_eth);
1240         struct ibv_flow_spec_eth eth = {
1241                 .type = parser->inner | IBV_FLOW_SPEC_ETH,
1242                 .size = eth_size,
1243         };
1244
1245         /* Don't update layer for the inner pattern. */
1246         if (!parser->inner)
1247                 parser->layer = HASH_RXQ_ETH;
1248         if (spec) {
1249                 unsigned int i;
1250
1251                 if (!mask)
1252                         mask = default_mask;
1253                 memcpy(&eth.val.dst_mac, spec->dst.addr_bytes, ETHER_ADDR_LEN);
1254                 memcpy(&eth.val.src_mac, spec->src.addr_bytes, ETHER_ADDR_LEN);
1255                 eth.val.ether_type = spec->type;
1256                 memcpy(&eth.mask.dst_mac, mask->dst.addr_bytes, ETHER_ADDR_LEN);
1257                 memcpy(&eth.mask.src_mac, mask->src.addr_bytes, ETHER_ADDR_LEN);
1258                 eth.mask.ether_type = mask->type;
1259                 /* Remove unwanted bits from values. */
1260                 for (i = 0; i < ETHER_ADDR_LEN; ++i) {
1261                         eth.val.dst_mac[i] &= eth.mask.dst_mac[i];
1262                         eth.val.src_mac[i] &= eth.mask.src_mac[i];
1263                 }
1264                 eth.val.ether_type &= eth.mask.ether_type;
1265         }
1266         mlx5_flow_create_copy(parser, &eth, eth_size);
1267         return 0;
1268 }
1269
1270 /**
1271  * Convert VLAN item to Verbs specification.
1272  *
1273  * @param item[in]
1274  *   Item specification.
1275  * @param default_mask[in]
1276  *   Default bit-masks to use when item->mask is not provided.
1277  * @param data[in, out]
1278  *   User structure.
1279  *
1280  * @return
1281  *   0 on success, a negative errno value otherwise and rte_errno is set.
1282  */
1283 static int
1284 mlx5_flow_create_vlan(const struct rte_flow_item *item,
1285                       const void *default_mask,
1286                       struct mlx5_flow_data *data)
1287 {
1288         const struct rte_flow_item_vlan *spec = item->spec;
1289         const struct rte_flow_item_vlan *mask = item->mask;
1290         struct mlx5_flow_parse *parser = data->parser;
1291         struct ibv_flow_spec_eth *eth;
1292         const unsigned int eth_size = sizeof(struct ibv_flow_spec_eth);
1293
1294         if (spec) {
1295                 unsigned int i;
1296                 if (!mask)
1297                         mask = default_mask;
1298
1299                 for (i = 0; i != hash_rxq_init_n; ++i) {
1300                         if (!parser->queue[i].ibv_attr)
1301                                 continue;
1302
1303                         eth = (void *)((uintptr_t)parser->queue[i].ibv_attr +
1304                                        parser->queue[i].offset - eth_size);
1305                         eth->val.vlan_tag = spec->tci;
1306                         eth->mask.vlan_tag = mask->tci;
1307                         eth->val.vlan_tag &= eth->mask.vlan_tag;
1308                         /*
1309                          * From verbs perspective an empty VLAN is equivalent
1310                          * to a packet without VLAN layer.
1311                          */
1312                         if (!eth->mask.vlan_tag)
1313                                 goto error;
1314                 }
1315                 return 0;
1316         }
1317 error:
1318         return rte_flow_error_set(data->error, EINVAL, RTE_FLOW_ERROR_TYPE_ITEM,
1319                                   item, "VLAN cannot be empty");
1320 }
1321
1322 /**
1323  * Convert IPv4 item to Verbs specification.
1324  *
1325  * @param item[in]
1326  *   Item specification.
1327  * @param default_mask[in]
1328  *   Default bit-masks to use when item->mask is not provided.
1329  * @param data[in, out]
1330  *   User structure.
1331  *
1332  * @return
1333  *   0 on success, a negative errno value otherwise and rte_errno is set.
1334  */
1335 static int
1336 mlx5_flow_create_ipv4(const struct rte_flow_item *item,
1337                       const void *default_mask,
1338                       struct mlx5_flow_data *data)
1339 {
1340         const struct rte_flow_item_ipv4 *spec = item->spec;
1341         const struct rte_flow_item_ipv4 *mask = item->mask;
1342         struct mlx5_flow_parse *parser = data->parser;
1343         unsigned int ipv4_size = sizeof(struct ibv_flow_spec_ipv4_ext);
1344         struct ibv_flow_spec_ipv4_ext ipv4 = {
1345                 .type = parser->inner | IBV_FLOW_SPEC_IPV4_EXT,
1346                 .size = ipv4_size,
1347         };
1348
1349         /* Don't update layer for the inner pattern. */
1350         if (!parser->inner)
1351                 parser->layer = HASH_RXQ_IPV4;
1352         if (spec) {
1353                 if (!mask)
1354                         mask = default_mask;
1355                 ipv4.val = (struct ibv_flow_ipv4_ext_filter){
1356                         .src_ip = spec->hdr.src_addr,
1357                         .dst_ip = spec->hdr.dst_addr,
1358                         .proto = spec->hdr.next_proto_id,
1359                         .tos = spec->hdr.type_of_service,
1360                 };
1361                 ipv4.mask = (struct ibv_flow_ipv4_ext_filter){
1362                         .src_ip = mask->hdr.src_addr,
1363                         .dst_ip = mask->hdr.dst_addr,
1364                         .proto = mask->hdr.next_proto_id,
1365                         .tos = mask->hdr.type_of_service,
1366                 };
1367                 /* Remove unwanted bits from values. */
1368                 ipv4.val.src_ip &= ipv4.mask.src_ip;
1369                 ipv4.val.dst_ip &= ipv4.mask.dst_ip;
1370                 ipv4.val.proto &= ipv4.mask.proto;
1371                 ipv4.val.tos &= ipv4.mask.tos;
1372         }
1373         mlx5_flow_create_copy(parser, &ipv4, ipv4_size);
1374         return 0;
1375 }
1376
1377 /**
1378  * Convert IPv6 item to Verbs specification.
1379  *
1380  * @param item[in]
1381  *   Item specification.
1382  * @param default_mask[in]
1383  *   Default bit-masks to use when item->mask is not provided.
1384  * @param data[in, out]
1385  *   User structure.
1386  *
1387  * @return
1388  *   0 on success, a negative errno value otherwise and rte_errno is set.
1389  */
1390 static int
1391 mlx5_flow_create_ipv6(const struct rte_flow_item *item,
1392                       const void *default_mask,
1393                       struct mlx5_flow_data *data)
1394 {
1395         const struct rte_flow_item_ipv6 *spec = item->spec;
1396         const struct rte_flow_item_ipv6 *mask = item->mask;
1397         struct mlx5_flow_parse *parser = data->parser;
1398         unsigned int ipv6_size = sizeof(struct ibv_flow_spec_ipv6);
1399         struct ibv_flow_spec_ipv6 ipv6 = {
1400                 .type = parser->inner | IBV_FLOW_SPEC_IPV6,
1401                 .size = ipv6_size,
1402         };
1403
1404         /* Don't update layer for the inner pattern. */
1405         if (!parser->inner)
1406                 parser->layer = HASH_RXQ_IPV6;
1407         if (spec) {
1408                 unsigned int i;
1409                 uint32_t vtc_flow_val;
1410                 uint32_t vtc_flow_mask;
1411
1412                 if (!mask)
1413                         mask = default_mask;
1414                 memcpy(&ipv6.val.src_ip, spec->hdr.src_addr,
1415                        RTE_DIM(ipv6.val.src_ip));
1416                 memcpy(&ipv6.val.dst_ip, spec->hdr.dst_addr,
1417                        RTE_DIM(ipv6.val.dst_ip));
1418                 memcpy(&ipv6.mask.src_ip, mask->hdr.src_addr,
1419                        RTE_DIM(ipv6.mask.src_ip));
1420                 memcpy(&ipv6.mask.dst_ip, mask->hdr.dst_addr,
1421                        RTE_DIM(ipv6.mask.dst_ip));
1422                 vtc_flow_val = rte_be_to_cpu_32(spec->hdr.vtc_flow);
1423                 vtc_flow_mask = rte_be_to_cpu_32(mask->hdr.vtc_flow);
1424                 ipv6.val.flow_label =
1425                         rte_cpu_to_be_32((vtc_flow_val & IPV6_HDR_FL_MASK) >>
1426                                          IPV6_HDR_FL_SHIFT);
1427                 ipv6.val.traffic_class = (vtc_flow_val & IPV6_HDR_TC_MASK) >>
1428                                          IPV6_HDR_TC_SHIFT;
1429                 ipv6.val.next_hdr = spec->hdr.proto;
1430                 ipv6.val.hop_limit = spec->hdr.hop_limits;
1431                 ipv6.mask.flow_label =
1432                         rte_cpu_to_be_32((vtc_flow_mask & IPV6_HDR_FL_MASK) >>
1433                                          IPV6_HDR_FL_SHIFT);
1434                 ipv6.mask.traffic_class = (vtc_flow_mask & IPV6_HDR_TC_MASK) >>
1435                                           IPV6_HDR_TC_SHIFT;
1436                 ipv6.mask.next_hdr = mask->hdr.proto;
1437                 ipv6.mask.hop_limit = mask->hdr.hop_limits;
1438                 /* Remove unwanted bits from values. */
1439                 for (i = 0; i < RTE_DIM(ipv6.val.src_ip); ++i) {
1440                         ipv6.val.src_ip[i] &= ipv6.mask.src_ip[i];
1441                         ipv6.val.dst_ip[i] &= ipv6.mask.dst_ip[i];
1442                 }
1443                 ipv6.val.flow_label &= ipv6.mask.flow_label;
1444                 ipv6.val.traffic_class &= ipv6.mask.traffic_class;
1445                 ipv6.val.next_hdr &= ipv6.mask.next_hdr;
1446                 ipv6.val.hop_limit &= ipv6.mask.hop_limit;
1447         }
1448         mlx5_flow_create_copy(parser, &ipv6, ipv6_size);
1449         return 0;
1450 }
1451
1452 /**
1453  * Convert UDP item to Verbs specification.
1454  *
1455  * @param item[in]
1456  *   Item specification.
1457  * @param default_mask[in]
1458  *   Default bit-masks to use when item->mask is not provided.
1459  * @param data[in, out]
1460  *   User structure.
1461  *
1462  * @return
1463  *   0 on success, a negative errno value otherwise and rte_errno is set.
1464  */
1465 static int
1466 mlx5_flow_create_udp(const struct rte_flow_item *item,
1467                      const void *default_mask,
1468                      struct mlx5_flow_data *data)
1469 {
1470         const struct rte_flow_item_udp *spec = item->spec;
1471         const struct rte_flow_item_udp *mask = item->mask;
1472         struct mlx5_flow_parse *parser = data->parser;
1473         unsigned int udp_size = sizeof(struct ibv_flow_spec_tcp_udp);
1474         struct ibv_flow_spec_tcp_udp udp = {
1475                 .type = parser->inner | IBV_FLOW_SPEC_UDP,
1476                 .size = udp_size,
1477         };
1478
1479         /* Don't update layer for the inner pattern. */
1480         if (!parser->inner) {
1481                 if (parser->layer == HASH_RXQ_IPV4)
1482                         parser->layer = HASH_RXQ_UDPV4;
1483                 else
1484                         parser->layer = HASH_RXQ_UDPV6;
1485         }
1486         if (spec) {
1487                 if (!mask)
1488                         mask = default_mask;
1489                 udp.val.dst_port = spec->hdr.dst_port;
1490                 udp.val.src_port = spec->hdr.src_port;
1491                 udp.mask.dst_port = mask->hdr.dst_port;
1492                 udp.mask.src_port = mask->hdr.src_port;
1493                 /* Remove unwanted bits from values. */
1494                 udp.val.src_port &= udp.mask.src_port;
1495                 udp.val.dst_port &= udp.mask.dst_port;
1496         }
1497         mlx5_flow_create_copy(parser, &udp, udp_size);
1498         return 0;
1499 }
1500
1501 /**
1502  * Convert TCP item to Verbs specification.
1503  *
1504  * @param item[in]
1505  *   Item specification.
1506  * @param default_mask[in]
1507  *   Default bit-masks to use when item->mask is not provided.
1508  * @param data[in, out]
1509  *   User structure.
1510  *
1511  * @return
1512  *   0 on success, a negative errno value otherwise and rte_errno is set.
1513  */
1514 static int
1515 mlx5_flow_create_tcp(const struct rte_flow_item *item,
1516                      const void *default_mask,
1517                      struct mlx5_flow_data *data)
1518 {
1519         const struct rte_flow_item_tcp *spec = item->spec;
1520         const struct rte_flow_item_tcp *mask = item->mask;
1521         struct mlx5_flow_parse *parser = data->parser;
1522         unsigned int tcp_size = sizeof(struct ibv_flow_spec_tcp_udp);
1523         struct ibv_flow_spec_tcp_udp tcp = {
1524                 .type = parser->inner | IBV_FLOW_SPEC_TCP,
1525                 .size = tcp_size,
1526         };
1527
1528         /* Don't update layer for the inner pattern. */
1529         if (!parser->inner) {
1530                 if (parser->layer == HASH_RXQ_IPV4)
1531                         parser->layer = HASH_RXQ_TCPV4;
1532                 else
1533                         parser->layer = HASH_RXQ_TCPV6;
1534         }
1535         if (spec) {
1536                 if (!mask)
1537                         mask = default_mask;
1538                 tcp.val.dst_port = spec->hdr.dst_port;
1539                 tcp.val.src_port = spec->hdr.src_port;
1540                 tcp.mask.dst_port = mask->hdr.dst_port;
1541                 tcp.mask.src_port = mask->hdr.src_port;
1542                 /* Remove unwanted bits from values. */
1543                 tcp.val.src_port &= tcp.mask.src_port;
1544                 tcp.val.dst_port &= tcp.mask.dst_port;
1545         }
1546         mlx5_flow_create_copy(parser, &tcp, tcp_size);
1547         return 0;
1548 }
1549
1550 /**
1551  * Convert VXLAN item to Verbs specification.
1552  *
1553  * @param item[in]
1554  *   Item specification.
1555  * @param default_mask[in]
1556  *   Default bit-masks to use when item->mask is not provided.
1557  * @param data[in, out]
1558  *   User structure.
1559  *
1560  * @return
1561  *   0 on success, a negative errno value otherwise and rte_errno is set.
1562  */
1563 static int
1564 mlx5_flow_create_vxlan(const struct rte_flow_item *item,
1565                        const void *default_mask,
1566                        struct mlx5_flow_data *data)
1567 {
1568         const struct rte_flow_item_vxlan *spec = item->spec;
1569         const struct rte_flow_item_vxlan *mask = item->mask;
1570         struct mlx5_flow_parse *parser = data->parser;
1571         unsigned int size = sizeof(struct ibv_flow_spec_tunnel);
1572         struct ibv_flow_spec_tunnel vxlan = {
1573                 .type = parser->inner | IBV_FLOW_SPEC_VXLAN_TUNNEL,
1574                 .size = size,
1575         };
1576         union vni {
1577                 uint32_t vlan_id;
1578                 uint8_t vni[4];
1579         } id;
1580
1581         id.vni[0] = 0;
1582         parser->inner = IBV_FLOW_SPEC_INNER;
1583         if (spec) {
1584                 if (!mask)
1585                         mask = default_mask;
1586                 memcpy(&id.vni[1], spec->vni, 3);
1587                 vxlan.val.tunnel_id = id.vlan_id;
1588                 memcpy(&id.vni[1], mask->vni, 3);
1589                 vxlan.mask.tunnel_id = id.vlan_id;
1590                 /* Remove unwanted bits from values. */
1591                 vxlan.val.tunnel_id &= vxlan.mask.tunnel_id;
1592         }
1593         /*
1594          * Tunnel id 0 is equivalent as not adding a VXLAN layer, if only this
1595          * layer is defined in the Verbs specification it is interpreted as
1596          * wildcard and all packets will match this rule, if it follows a full
1597          * stack layer (ex: eth / ipv4 / udp), all packets matching the layers
1598          * before will also match this rule.
1599          * To avoid such situation, VNI 0 is currently refused.
1600          */
1601         if (!vxlan.val.tunnel_id)
1602                 return rte_flow_error_set(data->error, EINVAL,
1603                                           RTE_FLOW_ERROR_TYPE_ITEM,
1604                                           item,
1605                                           "VxLAN vni cannot be 0");
1606         mlx5_flow_create_copy(parser, &vxlan, size);
1607         return 0;
1608 }
1609
1610 /**
1611  * Convert mark/flag action to Verbs specification.
1612  *
1613  * @param parser
1614  *   Internal parser structure.
1615  * @param mark_id
1616  *   Mark identifier.
1617  *
1618  * @return
1619  *   0 on success, a negative errno value otherwise and rte_errno is set.
1620  */
1621 static int
1622 mlx5_flow_create_flag_mark(struct mlx5_flow_parse *parser, uint32_t mark_id)
1623 {
1624         unsigned int size = sizeof(struct ibv_flow_spec_action_tag);
1625         struct ibv_flow_spec_action_tag tag = {
1626                 .type = IBV_FLOW_SPEC_ACTION_TAG,
1627                 .size = size,
1628                 .tag_id = mlx5_flow_mark_set(mark_id),
1629         };
1630
1631         assert(parser->mark);
1632         mlx5_flow_create_copy(parser, &tag, size);
1633         return 0;
1634 }
1635
1636 /**
1637  * Convert count action to Verbs specification.
1638  *
1639  * @param dev
1640  *   Pointer to Ethernet device.
1641  * @param parser
1642  *   Pointer to MLX5 flow parser structure.
1643  *
1644  * @return
1645  *   0 on success, a negative errno value otherwise and rte_errno is set.
1646  */
1647 static int
1648 mlx5_flow_create_count(struct rte_eth_dev *dev __rte_unused,
1649                        struct mlx5_flow_parse *parser __rte_unused)
1650 {
1651 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
1652         struct priv *priv = dev->data->dev_private;
1653         unsigned int size = sizeof(struct ibv_flow_spec_counter_action);
1654         struct ibv_counter_set_init_attr init_attr = {0};
1655         struct ibv_flow_spec_counter_action counter = {
1656                 .type = IBV_FLOW_SPEC_ACTION_COUNT,
1657                 .size = size,
1658                 .counter_set_handle = 0,
1659         };
1660
1661         init_attr.counter_set_id = 0;
1662         parser->cs = mlx5_glue->create_counter_set(priv->ctx, &init_attr);
1663         if (!parser->cs) {
1664                 rte_errno = EINVAL;
1665                 return -rte_errno;
1666         }
1667         counter.counter_set_handle = parser->cs->handle;
1668         mlx5_flow_create_copy(parser, &counter, size);
1669 #endif
1670         return 0;
1671 }
1672
1673 /**
1674  * Complete flow rule creation with a drop queue.
1675  *
1676  * @param dev
1677  *   Pointer to Ethernet device.
1678  * @param parser
1679  *   Internal parser structure.
1680  * @param flow
1681  *   Pointer to the rte_flow.
1682  * @param[out] error
1683  *   Perform verbose error reporting if not NULL.
1684  *
1685  * @return
1686  *   0 on success, a negative errno value otherwise and rte_errno is set.
1687  */
1688 static int
1689 mlx5_flow_create_action_queue_drop(struct rte_eth_dev *dev,
1690                                    struct mlx5_flow_parse *parser,
1691                                    struct rte_flow *flow,
1692                                    struct rte_flow_error *error)
1693 {
1694         struct priv *priv = dev->data->dev_private;
1695         struct ibv_flow_spec_action_drop *drop;
1696         unsigned int size = sizeof(struct ibv_flow_spec_action_drop);
1697
1698         assert(priv->pd);
1699         assert(priv->ctx);
1700         flow->drop = 1;
1701         drop = (void *)((uintptr_t)parser->queue[HASH_RXQ_ETH].ibv_attr +
1702                         parser->queue[HASH_RXQ_ETH].offset);
1703         *drop = (struct ibv_flow_spec_action_drop){
1704                         .type = IBV_FLOW_SPEC_ACTION_DROP,
1705                         .size = size,
1706         };
1707         ++parser->queue[HASH_RXQ_ETH].ibv_attr->num_of_specs;
1708         parser->queue[HASH_RXQ_ETH].offset += size;
1709         flow->frxq[HASH_RXQ_ETH].ibv_attr =
1710                 parser->queue[HASH_RXQ_ETH].ibv_attr;
1711         if (parser->count)
1712                 flow->cs = parser->cs;
1713         if (!priv->dev->data->dev_started)
1714                 return 0;
1715         parser->queue[HASH_RXQ_ETH].ibv_attr = NULL;
1716         flow->frxq[HASH_RXQ_ETH].ibv_flow =
1717                 mlx5_glue->create_flow(priv->flow_drop_queue->qp,
1718                                        flow->frxq[HASH_RXQ_ETH].ibv_attr);
1719         if (!flow->frxq[HASH_RXQ_ETH].ibv_flow) {
1720                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1721                                    NULL, "flow rule creation failure");
1722                 goto error;
1723         }
1724         return 0;
1725 error:
1726         assert(flow);
1727         if (flow->frxq[HASH_RXQ_ETH].ibv_flow) {
1728                 claim_zero(mlx5_glue->destroy_flow
1729                            (flow->frxq[HASH_RXQ_ETH].ibv_flow));
1730                 flow->frxq[HASH_RXQ_ETH].ibv_flow = NULL;
1731         }
1732         if (flow->frxq[HASH_RXQ_ETH].ibv_attr) {
1733                 rte_free(flow->frxq[HASH_RXQ_ETH].ibv_attr);
1734                 flow->frxq[HASH_RXQ_ETH].ibv_attr = NULL;
1735         }
1736         if (flow->cs) {
1737                 claim_zero(mlx5_glue->destroy_counter_set(flow->cs));
1738                 flow->cs = NULL;
1739                 parser->cs = NULL;
1740         }
1741         return -rte_errno;
1742 }
1743
1744 /**
1745  * Create hash Rx queues when RSS is enabled.
1746  *
1747  * @param dev
1748  *   Pointer to Ethernet device.
1749  * @param parser
1750  *   Internal parser structure.
1751  * @param flow
1752  *   Pointer to the rte_flow.
1753  * @param[out] error
1754  *   Perform verbose error reporting if not NULL.
1755  *
1756  * @return
1757  *   0 on success, a negative errno value otherwise and rte_errno is set.
1758  */
1759 static int
1760 mlx5_flow_create_action_queue_rss(struct rte_eth_dev *dev,
1761                                   struct mlx5_flow_parse *parser,
1762                                   struct rte_flow *flow,
1763                                   struct rte_flow_error *error)
1764 {
1765         struct priv *priv = dev->data->dev_private;
1766         unsigned int i;
1767
1768         for (i = 0; i != hash_rxq_init_n; ++i) {
1769                 uint64_t hash_fields;
1770
1771                 if (!parser->queue[i].ibv_attr)
1772                         continue;
1773                 flow->frxq[i].ibv_attr = parser->queue[i].ibv_attr;
1774                 parser->queue[i].ibv_attr = NULL;
1775                 hash_fields = hash_rxq_init[i].hash_fields;
1776                 if (!priv->dev->data->dev_started)
1777                         continue;
1778                 flow->frxq[i].hrxq =
1779                         mlx5_hrxq_get(dev,
1780                                       parser->rss_conf.rss_key,
1781                                       parser->rss_conf.rss_key_len,
1782                                       hash_fields,
1783                                       parser->queues,
1784                                       parser->queues_n);
1785                 if (flow->frxq[i].hrxq)
1786                         continue;
1787                 flow->frxq[i].hrxq =
1788                         mlx5_hrxq_new(dev,
1789                                       parser->rss_conf.rss_key,
1790                                       parser->rss_conf.rss_key_len,
1791                                       hash_fields,
1792                                       parser->queues,
1793                                       parser->queues_n);
1794                 if (!flow->frxq[i].hrxq) {
1795                         return rte_flow_error_set(error, ENOMEM,
1796                                                   RTE_FLOW_ERROR_TYPE_HANDLE,
1797                                                   NULL,
1798                                                   "cannot create hash rxq");
1799                 }
1800         }
1801         return 0;
1802 }
1803
1804 /**
1805  * Complete flow rule creation.
1806  *
1807  * @param dev
1808  *   Pointer to Ethernet device.
1809  * @param parser
1810  *   Internal parser structure.
1811  * @param flow
1812  *   Pointer to the rte_flow.
1813  * @param[out] error
1814  *   Perform verbose error reporting if not NULL.
1815  *
1816  * @return
1817  *   0 on success, a negative errno value otherwise and rte_errno is set.
1818  */
1819 static int
1820 mlx5_flow_create_action_queue(struct rte_eth_dev *dev,
1821                               struct mlx5_flow_parse *parser,
1822                               struct rte_flow *flow,
1823                               struct rte_flow_error *error)
1824 {
1825         struct priv *priv = dev->data->dev_private;
1826         int ret;
1827         unsigned int i;
1828         unsigned int flows_n = 0;
1829
1830         assert(priv->pd);
1831         assert(priv->ctx);
1832         assert(!parser->drop);
1833         ret = mlx5_flow_create_action_queue_rss(dev, parser, flow, error);
1834         if (ret)
1835                 goto error;
1836         if (parser->count)
1837                 flow->cs = parser->cs;
1838         if (!priv->dev->data->dev_started)
1839                 return 0;
1840         for (i = 0; i != hash_rxq_init_n; ++i) {
1841                 if (!flow->frxq[i].hrxq)
1842                         continue;
1843                 flow->frxq[i].ibv_flow =
1844                         mlx5_glue->create_flow(flow->frxq[i].hrxq->qp,
1845                                                flow->frxq[i].ibv_attr);
1846                 if (!flow->frxq[i].ibv_flow) {
1847                         rte_flow_error_set(error, ENOMEM,
1848                                            RTE_FLOW_ERROR_TYPE_HANDLE,
1849                                            NULL, "flow rule creation failure");
1850                         goto error;
1851                 }
1852                 ++flows_n;
1853                 DRV_LOG(DEBUG, "port %u %p type %d QP %p ibv_flow %p",
1854                         dev->data->port_id,
1855                         (void *)flow, i,
1856                         (void *)flow->frxq[i].hrxq,
1857                         (void *)flow->frxq[i].ibv_flow);
1858         }
1859         if (!flows_n) {
1860                 rte_flow_error_set(error, EINVAL, RTE_FLOW_ERROR_TYPE_HANDLE,
1861                                    NULL, "internal error in flow creation");
1862                 goto error;
1863         }
1864         for (i = 0; i != parser->queues_n; ++i) {
1865                 struct mlx5_rxq_data *q =
1866                         (*priv->rxqs)[parser->queues[i]];
1867
1868                 q->mark |= parser->mark;
1869         }
1870         return 0;
1871 error:
1872         ret = rte_errno; /* Save rte_errno before cleanup. */
1873         assert(flow);
1874         for (i = 0; i != hash_rxq_init_n; ++i) {
1875                 if (flow->frxq[i].ibv_flow) {
1876                         struct ibv_flow *ibv_flow = flow->frxq[i].ibv_flow;
1877
1878                         claim_zero(mlx5_glue->destroy_flow(ibv_flow));
1879                 }
1880                 if (flow->frxq[i].hrxq)
1881                         mlx5_hrxq_release(dev, flow->frxq[i].hrxq);
1882                 if (flow->frxq[i].ibv_attr)
1883                         rte_free(flow->frxq[i].ibv_attr);
1884         }
1885         if (flow->cs) {
1886                 claim_zero(mlx5_glue->destroy_counter_set(flow->cs));
1887                 flow->cs = NULL;
1888                 parser->cs = NULL;
1889         }
1890         rte_errno = ret; /* Restore rte_errno. */
1891         return -rte_errno;
1892 }
1893
1894 /**
1895  * Convert a flow.
1896  *
1897  * @param dev
1898  *   Pointer to Ethernet device.
1899  * @param list
1900  *   Pointer to a TAILQ flow list.
1901  * @param[in] attr
1902  *   Flow rule attributes.
1903  * @param[in] pattern
1904  *   Pattern specification (list terminated by the END pattern item).
1905  * @param[in] actions
1906  *   Associated actions (list terminated by the END action).
1907  * @param[out] error
1908  *   Perform verbose error reporting if not NULL.
1909  *
1910  * @return
1911  *   A flow on success, NULL otherwise and rte_errno is set.
1912  */
1913 static struct rte_flow *
1914 mlx5_flow_list_create(struct rte_eth_dev *dev,
1915                       struct mlx5_flows *list,
1916                       const struct rte_flow_attr *attr,
1917                       const struct rte_flow_item items[],
1918                       const struct rte_flow_action actions[],
1919                       struct rte_flow_error *error)
1920 {
1921         struct mlx5_flow_parse parser = { .create = 1, };
1922         struct rte_flow *flow = NULL;
1923         unsigned int i;
1924         int ret;
1925
1926         ret = mlx5_flow_convert(dev, attr, items, actions, error, &parser);
1927         if (ret)
1928                 goto exit;
1929         flow = rte_calloc(__func__, 1,
1930                           sizeof(*flow) + parser.queues_n * sizeof(uint16_t),
1931                           0);
1932         if (!flow) {
1933                 rte_flow_error_set(error, ENOMEM,
1934                                    RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
1935                                    NULL,
1936                                    "cannot allocate flow memory");
1937                 return NULL;
1938         }
1939         /* Copy queues configuration. */
1940         flow->queues = (uint16_t (*)[])(flow + 1);
1941         memcpy(flow->queues, parser.queues, parser.queues_n * sizeof(uint16_t));
1942         flow->queues_n = parser.queues_n;
1943         flow->mark = parser.mark;
1944         /* Copy RSS configuration. */
1945         flow->rss_conf = parser.rss_conf;
1946         flow->rss_conf.rss_key = flow->rss_key;
1947         memcpy(flow->rss_key, parser.rss_key, parser.rss_conf.rss_key_len);
1948         /* finalise the flow. */
1949         if (parser.drop)
1950                 ret = mlx5_flow_create_action_queue_drop(dev, &parser, flow,
1951                                                          error);
1952         else
1953                 ret = mlx5_flow_create_action_queue(dev, &parser, flow, error);
1954         if (ret)
1955                 goto exit;
1956         TAILQ_INSERT_TAIL(list, flow, next);
1957         DRV_LOG(DEBUG, "port %u flow created %p", dev->data->port_id,
1958                 (void *)flow);
1959         return flow;
1960 exit:
1961         DRV_LOG(ERR, "port %u flow creation error: %s", dev->data->port_id,
1962                 error->message);
1963         for (i = 0; i != hash_rxq_init_n; ++i) {
1964                 if (parser.queue[i].ibv_attr)
1965                         rte_free(parser.queue[i].ibv_attr);
1966         }
1967         rte_free(flow);
1968         return NULL;
1969 }
1970
1971 /**
1972  * Validate a flow supported by the NIC.
1973  *
1974  * @see rte_flow_validate()
1975  * @see rte_flow_ops
1976  */
1977 int
1978 mlx5_flow_validate(struct rte_eth_dev *dev,
1979                    const struct rte_flow_attr *attr,
1980                    const struct rte_flow_item items[],
1981                    const struct rte_flow_action actions[],
1982                    struct rte_flow_error *error)
1983 {
1984         struct mlx5_flow_parse parser = { .create = 0, };
1985
1986         return mlx5_flow_convert(dev, attr, items, actions, error, &parser);
1987 }
1988
1989 /**
1990  * Create a flow.
1991  *
1992  * @see rte_flow_create()
1993  * @see rte_flow_ops
1994  */
1995 struct rte_flow *
1996 mlx5_flow_create(struct rte_eth_dev *dev,
1997                  const struct rte_flow_attr *attr,
1998                  const struct rte_flow_item items[],
1999                  const struct rte_flow_action actions[],
2000                  struct rte_flow_error *error)
2001 {
2002         struct priv *priv = dev->data->dev_private;
2003
2004         return mlx5_flow_list_create(dev, &priv->flows, attr, items, actions,
2005                                      error);
2006 }
2007
2008 /**
2009  * Destroy a flow in a list.
2010  *
2011  * @param dev
2012  *   Pointer to Ethernet device.
2013  * @param list
2014  *   Pointer to a TAILQ flow list.
2015  * @param[in] flow
2016  *   Flow to destroy.
2017  */
2018 static void
2019 mlx5_flow_list_destroy(struct rte_eth_dev *dev, struct mlx5_flows *list,
2020                        struct rte_flow *flow)
2021 {
2022         struct priv *priv = dev->data->dev_private;
2023         unsigned int i;
2024
2025         if (flow->drop || !flow->mark)
2026                 goto free;
2027         for (i = 0; i != flow->queues_n; ++i) {
2028                 struct rte_flow *tmp;
2029                 int mark = 0;
2030
2031                 /*
2032                  * To remove the mark from the queue, the queue must not be
2033                  * present in any other marked flow (RSS or not).
2034                  */
2035                 TAILQ_FOREACH(tmp, list, next) {
2036                         unsigned int j;
2037                         uint16_t *tqs = NULL;
2038                         uint16_t tq_n = 0;
2039
2040                         if (!tmp->mark)
2041                                 continue;
2042                         for (j = 0; j != hash_rxq_init_n; ++j) {
2043                                 if (!tmp->frxq[j].hrxq)
2044                                         continue;
2045                                 tqs = tmp->frxq[j].hrxq->ind_table->queues;
2046                                 tq_n = tmp->frxq[j].hrxq->ind_table->queues_n;
2047                         }
2048                         if (!tq_n)
2049                                 continue;
2050                         for (j = 0; (j != tq_n) && !mark; j++)
2051                                 if (tqs[j] == (*flow->queues)[i])
2052                                         mark = 1;
2053                 }
2054                 (*priv->rxqs)[(*flow->queues)[i]]->mark = mark;
2055         }
2056 free:
2057         if (flow->drop) {
2058                 if (flow->frxq[HASH_RXQ_ETH].ibv_flow)
2059                         claim_zero(mlx5_glue->destroy_flow
2060                                    (flow->frxq[HASH_RXQ_ETH].ibv_flow));
2061                 rte_free(flow->frxq[HASH_RXQ_ETH].ibv_attr);
2062         } else {
2063                 for (i = 0; i != hash_rxq_init_n; ++i) {
2064                         struct mlx5_flow *frxq = &flow->frxq[i];
2065
2066                         if (frxq->ibv_flow)
2067                                 claim_zero(mlx5_glue->destroy_flow
2068                                            (frxq->ibv_flow));
2069                         if (frxq->hrxq)
2070                                 mlx5_hrxq_release(dev, frxq->hrxq);
2071                         if (frxq->ibv_attr)
2072                                 rte_free(frxq->ibv_attr);
2073                 }
2074         }
2075         if (flow->cs) {
2076                 claim_zero(mlx5_glue->destroy_counter_set(flow->cs));
2077                 flow->cs = NULL;
2078         }
2079         TAILQ_REMOVE(list, flow, next);
2080         DRV_LOG(DEBUG, "port %u flow destroyed %p", dev->data->port_id,
2081                 (void *)flow);
2082         rte_free(flow);
2083 }
2084
2085 /**
2086  * Destroy all flows.
2087  *
2088  * @param dev
2089  *   Pointer to Ethernet device.
2090  * @param list
2091  *   Pointer to a TAILQ flow list.
2092  */
2093 void
2094 mlx5_flow_list_flush(struct rte_eth_dev *dev, struct mlx5_flows *list)
2095 {
2096         while (!TAILQ_EMPTY(list)) {
2097                 struct rte_flow *flow;
2098
2099                 flow = TAILQ_FIRST(list);
2100                 mlx5_flow_list_destroy(dev, list, flow);
2101         }
2102 }
2103
2104 /**
2105  * Create drop queue.
2106  *
2107  * @param dev
2108  *   Pointer to Ethernet device.
2109  *
2110  * @return
2111  *   0 on success, a negative errno value otherwise and rte_errno is set.
2112  */
2113 int
2114 mlx5_flow_create_drop_queue(struct rte_eth_dev *dev)
2115 {
2116         struct priv *priv = dev->data->dev_private;
2117         struct mlx5_hrxq_drop *fdq = NULL;
2118
2119         assert(priv->pd);
2120         assert(priv->ctx);
2121         fdq = rte_calloc(__func__, 1, sizeof(*fdq), 0);
2122         if (!fdq) {
2123                 DRV_LOG(WARNING,
2124                         "port %u cannot allocate memory for drop queue",
2125                         dev->data->port_id);
2126                 rte_errno = ENOMEM;
2127                 return -rte_errno;
2128         }
2129         fdq->cq = mlx5_glue->create_cq(priv->ctx, 1, NULL, NULL, 0);
2130         if (!fdq->cq) {
2131                 DRV_LOG(WARNING, "port %u cannot allocate CQ for drop queue",
2132                         dev->data->port_id);
2133                 rte_errno = errno;
2134                 goto error;
2135         }
2136         fdq->wq = mlx5_glue->create_wq
2137                 (priv->ctx,
2138                  &(struct ibv_wq_init_attr){
2139                         .wq_type = IBV_WQT_RQ,
2140                         .max_wr = 1,
2141                         .max_sge = 1,
2142                         .pd = priv->pd,
2143                         .cq = fdq->cq,
2144                  });
2145         if (!fdq->wq) {
2146                 DRV_LOG(WARNING, "port %u cannot allocate WQ for drop queue",
2147                         dev->data->port_id);
2148                 rte_errno = errno;
2149                 goto error;
2150         }
2151         fdq->ind_table = mlx5_glue->create_rwq_ind_table
2152                 (priv->ctx,
2153                  &(struct ibv_rwq_ind_table_init_attr){
2154                         .log_ind_tbl_size = 0,
2155                         .ind_tbl = &fdq->wq,
2156                         .comp_mask = 0,
2157                  });
2158         if (!fdq->ind_table) {
2159                 DRV_LOG(WARNING,
2160                         "port %u cannot allocate indirection table for drop"
2161                         " queue",
2162                         dev->data->port_id);
2163                 rte_errno = errno;
2164                 goto error;
2165         }
2166         fdq->qp = mlx5_glue->create_qp_ex
2167                 (priv->ctx,
2168                  &(struct ibv_qp_init_attr_ex){
2169                         .qp_type = IBV_QPT_RAW_PACKET,
2170                         .comp_mask =
2171                                 IBV_QP_INIT_ATTR_PD |
2172                                 IBV_QP_INIT_ATTR_IND_TABLE |
2173                                 IBV_QP_INIT_ATTR_RX_HASH,
2174                         .rx_hash_conf = (struct ibv_rx_hash_conf){
2175                                 .rx_hash_function =
2176                                         IBV_RX_HASH_FUNC_TOEPLITZ,
2177                                 .rx_hash_key_len = rss_hash_default_key_len,
2178                                 .rx_hash_key = rss_hash_default_key,
2179                                 .rx_hash_fields_mask = 0,
2180                                 },
2181                         .rwq_ind_tbl = fdq->ind_table,
2182                         .pd = priv->pd
2183                  });
2184         if (!fdq->qp) {
2185                 DRV_LOG(WARNING, "port %u cannot allocate QP for drop queue",
2186                         dev->data->port_id);
2187                 rte_errno = errno;
2188                 goto error;
2189         }
2190         priv->flow_drop_queue = fdq;
2191         return 0;
2192 error:
2193         if (fdq->qp)
2194                 claim_zero(mlx5_glue->destroy_qp(fdq->qp));
2195         if (fdq->ind_table)
2196                 claim_zero(mlx5_glue->destroy_rwq_ind_table(fdq->ind_table));
2197         if (fdq->wq)
2198                 claim_zero(mlx5_glue->destroy_wq(fdq->wq));
2199         if (fdq->cq)
2200                 claim_zero(mlx5_glue->destroy_cq(fdq->cq));
2201         if (fdq)
2202                 rte_free(fdq);
2203         priv->flow_drop_queue = NULL;
2204         return -rte_errno;
2205 }
2206
2207 /**
2208  * Delete drop queue.
2209  *
2210  * @param dev
2211  *   Pointer to Ethernet device.
2212  */
2213 void
2214 mlx5_flow_delete_drop_queue(struct rte_eth_dev *dev)
2215 {
2216         struct priv *priv = dev->data->dev_private;
2217         struct mlx5_hrxq_drop *fdq = priv->flow_drop_queue;
2218
2219         if (!fdq)
2220                 return;
2221         if (fdq->qp)
2222                 claim_zero(mlx5_glue->destroy_qp(fdq->qp));
2223         if (fdq->ind_table)
2224                 claim_zero(mlx5_glue->destroy_rwq_ind_table(fdq->ind_table));
2225         if (fdq->wq)
2226                 claim_zero(mlx5_glue->destroy_wq(fdq->wq));
2227         if (fdq->cq)
2228                 claim_zero(mlx5_glue->destroy_cq(fdq->cq));
2229         rte_free(fdq);
2230         priv->flow_drop_queue = NULL;
2231 }
2232
2233 /**
2234  * Remove all flows.
2235  *
2236  * @param dev
2237  *   Pointer to Ethernet device.
2238  * @param list
2239  *   Pointer to a TAILQ flow list.
2240  */
2241 void
2242 mlx5_flow_stop(struct rte_eth_dev *dev, struct mlx5_flows *list)
2243 {
2244         struct priv *priv = dev->data->dev_private;
2245         struct rte_flow *flow;
2246
2247         TAILQ_FOREACH_REVERSE(flow, list, mlx5_flows, next) {
2248                 unsigned int i;
2249                 struct mlx5_ind_table_ibv *ind_tbl = NULL;
2250
2251                 if (flow->drop) {
2252                         if (!flow->frxq[HASH_RXQ_ETH].ibv_flow)
2253                                 continue;
2254                         claim_zero(mlx5_glue->destroy_flow
2255                                    (flow->frxq[HASH_RXQ_ETH].ibv_flow));
2256                         flow->frxq[HASH_RXQ_ETH].ibv_flow = NULL;
2257                         DRV_LOG(DEBUG, "port %u flow %p removed",
2258                                 dev->data->port_id, (void *)flow);
2259                         /* Next flow. */
2260                         continue;
2261                 }
2262                 /* Verify the flow has not already been cleaned. */
2263                 for (i = 0; i != hash_rxq_init_n; ++i) {
2264                         if (!flow->frxq[i].ibv_flow)
2265                                 continue;
2266                         /*
2267                          * Indirection table may be necessary to remove the
2268                          * flags in the Rx queues.
2269                          * This helps to speed-up the process by avoiding
2270                          * another loop.
2271                          */
2272                         ind_tbl = flow->frxq[i].hrxq->ind_table;
2273                         break;
2274                 }
2275                 if (i == hash_rxq_init_n)
2276                         return;
2277                 if (flow->mark) {
2278                         assert(ind_tbl);
2279                         for (i = 0; i != ind_tbl->queues_n; ++i)
2280                                 (*priv->rxqs)[ind_tbl->queues[i]]->mark = 0;
2281                 }
2282                 for (i = 0; i != hash_rxq_init_n; ++i) {
2283                         if (!flow->frxq[i].ibv_flow)
2284                                 continue;
2285                         claim_zero(mlx5_glue->destroy_flow
2286                                    (flow->frxq[i].ibv_flow));
2287                         flow->frxq[i].ibv_flow = NULL;
2288                         mlx5_hrxq_release(dev, flow->frxq[i].hrxq);
2289                         flow->frxq[i].hrxq = NULL;
2290                 }
2291                 DRV_LOG(DEBUG, "port %u flow %p removed", dev->data->port_id,
2292                         (void *)flow);
2293         }
2294 }
2295
2296 /**
2297  * Add all flows.
2298  *
2299  * @param dev
2300  *   Pointer to Ethernet device.
2301  * @param list
2302  *   Pointer to a TAILQ flow list.
2303  *
2304  * @return
2305  *   0 on success, a negative errno value otherwise and rte_errno is set.
2306  */
2307 int
2308 mlx5_flow_start(struct rte_eth_dev *dev, struct mlx5_flows *list)
2309 {
2310         struct priv *priv = dev->data->dev_private;
2311         struct rte_flow *flow;
2312
2313         TAILQ_FOREACH(flow, list, next) {
2314                 unsigned int i;
2315
2316                 if (flow->drop) {
2317                         flow->frxq[HASH_RXQ_ETH].ibv_flow =
2318                                 mlx5_glue->create_flow
2319                                 (priv->flow_drop_queue->qp,
2320                                  flow->frxq[HASH_RXQ_ETH].ibv_attr);
2321                         if (!flow->frxq[HASH_RXQ_ETH].ibv_flow) {
2322                                 DRV_LOG(DEBUG,
2323                                         "port %u flow %p cannot be applied",
2324                                         dev->data->port_id, (void *)flow);
2325                                 rte_errno = EINVAL;
2326                                 return -rte_errno;
2327                         }
2328                         DRV_LOG(DEBUG, "port %u flow %p applied",
2329                                 dev->data->port_id, (void *)flow);
2330                         /* Next flow. */
2331                         continue;
2332                 }
2333                 for (i = 0; i != hash_rxq_init_n; ++i) {
2334                         if (!flow->frxq[i].ibv_attr)
2335                                 continue;
2336                         flow->frxq[i].hrxq =
2337                                 mlx5_hrxq_get(dev, flow->rss_conf.rss_key,
2338                                               flow->rss_conf.rss_key_len,
2339                                               hash_rxq_init[i].hash_fields,
2340                                               (*flow->queues),
2341                                               flow->queues_n);
2342                         if (flow->frxq[i].hrxq)
2343                                 goto flow_create;
2344                         flow->frxq[i].hrxq =
2345                                 mlx5_hrxq_new(dev, flow->rss_conf.rss_key,
2346                                               flow->rss_conf.rss_key_len,
2347                                               hash_rxq_init[i].hash_fields,
2348                                               (*flow->queues),
2349                                               flow->queues_n);
2350                         if (!flow->frxq[i].hrxq) {
2351                                 DRV_LOG(DEBUG,
2352                                         "port %u flow %p cannot be applied",
2353                                         dev->data->port_id, (void *)flow);
2354                                 rte_errno = EINVAL;
2355                                 return -rte_errno;
2356                         }
2357 flow_create:
2358                         flow->frxq[i].ibv_flow =
2359                                 mlx5_glue->create_flow(flow->frxq[i].hrxq->qp,
2360                                                        flow->frxq[i].ibv_attr);
2361                         if (!flow->frxq[i].ibv_flow) {
2362                                 DRV_LOG(DEBUG,
2363                                         "port %u flow %p cannot be applied",
2364                                         dev->data->port_id, (void *)flow);
2365                                 rte_errno = EINVAL;
2366                                 return -rte_errno;
2367                         }
2368                         DRV_LOG(DEBUG, "port %u flow %p applied",
2369                                 dev->data->port_id, (void *)flow);
2370                 }
2371                 if (!flow->mark)
2372                         continue;
2373                 for (i = 0; i != flow->queues_n; ++i)
2374                         (*priv->rxqs)[(*flow->queues)[i]]->mark = 1;
2375         }
2376         return 0;
2377 }
2378
2379 /**
2380  * Verify the flow list is empty
2381  *
2382  * @param dev
2383  *  Pointer to Ethernet device.
2384  *
2385  * @return the number of flows not released.
2386  */
2387 int
2388 mlx5_flow_verify(struct rte_eth_dev *dev)
2389 {
2390         struct priv *priv = dev->data->dev_private;
2391         struct rte_flow *flow;
2392         int ret = 0;
2393
2394         TAILQ_FOREACH(flow, &priv->flows, next) {
2395                 DRV_LOG(DEBUG, "port %u flow %p still referenced",
2396                         dev->data->port_id, (void *)flow);
2397                 ++ret;
2398         }
2399         return ret;
2400 }
2401
2402 /**
2403  * Enable a control flow configured from the control plane.
2404  *
2405  * @param dev
2406  *   Pointer to Ethernet device.
2407  * @param eth_spec
2408  *   An Ethernet flow spec to apply.
2409  * @param eth_mask
2410  *   An Ethernet flow mask to apply.
2411  * @param vlan_spec
2412  *   A VLAN flow spec to apply.
2413  * @param vlan_mask
2414  *   A VLAN flow mask to apply.
2415  *
2416  * @return
2417  *   0 on success, a negative errno value otherwise and rte_errno is set.
2418  */
2419 int
2420 mlx5_ctrl_flow_vlan(struct rte_eth_dev *dev,
2421                     struct rte_flow_item_eth *eth_spec,
2422                     struct rte_flow_item_eth *eth_mask,
2423                     struct rte_flow_item_vlan *vlan_spec,
2424                     struct rte_flow_item_vlan *vlan_mask)
2425 {
2426         struct priv *priv = dev->data->dev_private;
2427         const struct rte_flow_attr attr = {
2428                 .ingress = 1,
2429                 .priority = MLX5_CTRL_FLOW_PRIORITY,
2430         };
2431         struct rte_flow_item items[] = {
2432                 {
2433                         .type = RTE_FLOW_ITEM_TYPE_ETH,
2434                         .spec = eth_spec,
2435                         .last = NULL,
2436                         .mask = eth_mask,
2437                 },
2438                 {
2439                         .type = (vlan_spec) ? RTE_FLOW_ITEM_TYPE_VLAN :
2440                                 RTE_FLOW_ITEM_TYPE_END,
2441                         .spec = vlan_spec,
2442                         .last = NULL,
2443                         .mask = vlan_mask,
2444                 },
2445                 {
2446                         .type = RTE_FLOW_ITEM_TYPE_END,
2447                 },
2448         };
2449         struct rte_flow_action actions[] = {
2450                 {
2451                         .type = RTE_FLOW_ACTION_TYPE_RSS,
2452                 },
2453                 {
2454                         .type = RTE_FLOW_ACTION_TYPE_END,
2455                 },
2456         };
2457         struct rte_flow *flow;
2458         struct rte_flow_error error;
2459         unsigned int i;
2460         union {
2461                 struct rte_flow_action_rss rss;
2462                 struct {
2463                         const struct rte_eth_rss_conf *rss_conf;
2464                         uint16_t num;
2465                         uint16_t queue[RTE_MAX_QUEUES_PER_PORT];
2466                 } local;
2467         } action_rss;
2468
2469         if (!priv->reta_idx_n) {
2470                 rte_errno = EINVAL;
2471                 return -rte_errno;
2472         }
2473         for (i = 0; i != priv->reta_idx_n; ++i)
2474                 action_rss.local.queue[i] = (*priv->reta_idx)[i];
2475         action_rss.local.rss_conf = &priv->rss_conf;
2476         action_rss.local.num = priv->reta_idx_n;
2477         actions[0].conf = (const void *)&action_rss.rss;
2478         flow = mlx5_flow_list_create(dev, &priv->ctrl_flows, &attr, items,
2479                                      actions, &error);
2480         if (!flow)
2481                 return -rte_errno;
2482         return 0;
2483 }
2484
2485 /**
2486  * Enable a flow control configured from the control plane.
2487  *
2488  * @param dev
2489  *   Pointer to Ethernet device.
2490  * @param eth_spec
2491  *   An Ethernet flow spec to apply.
2492  * @param eth_mask
2493  *   An Ethernet flow mask to apply.
2494  *
2495  * @return
2496  *   0 on success, a negative errno value otherwise and rte_errno is set.
2497  */
2498 int
2499 mlx5_ctrl_flow(struct rte_eth_dev *dev,
2500                struct rte_flow_item_eth *eth_spec,
2501                struct rte_flow_item_eth *eth_mask)
2502 {
2503         return mlx5_ctrl_flow_vlan(dev, eth_spec, eth_mask, NULL, NULL);
2504 }
2505
2506 /**
2507  * Destroy a flow.
2508  *
2509  * @see rte_flow_destroy()
2510  * @see rte_flow_ops
2511  */
2512 int
2513 mlx5_flow_destroy(struct rte_eth_dev *dev,
2514                   struct rte_flow *flow,
2515                   struct rte_flow_error *error __rte_unused)
2516 {
2517         struct priv *priv = dev->data->dev_private;
2518
2519         mlx5_flow_list_destroy(dev, &priv->flows, flow);
2520         return 0;
2521 }
2522
2523 /**
2524  * Destroy all flows.
2525  *
2526  * @see rte_flow_flush()
2527  * @see rte_flow_ops
2528  */
2529 int
2530 mlx5_flow_flush(struct rte_eth_dev *dev,
2531                 struct rte_flow_error *error __rte_unused)
2532 {
2533         struct priv *priv = dev->data->dev_private;
2534
2535         mlx5_flow_list_flush(dev, &priv->flows);
2536         return 0;
2537 }
2538
2539 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
2540 /**
2541  * Query flow counter.
2542  *
2543  * @param cs
2544  *   the counter set.
2545  * @param counter_value
2546  *   returned data from the counter.
2547  *
2548  * @return
2549  *   0 on success, a negative errno value otherwise and rte_errno is set.
2550  */
2551 static int
2552 mlx5_flow_query_count(struct ibv_counter_set *cs,
2553                       struct mlx5_flow_counter_stats *counter_stats,
2554                       struct rte_flow_query_count *query_count,
2555                       struct rte_flow_error *error)
2556 {
2557         uint64_t counters[2];
2558         struct ibv_query_counter_set_attr query_cs_attr = {
2559                 .cs = cs,
2560                 .query_flags = IBV_COUNTER_SET_FORCE_UPDATE,
2561         };
2562         struct ibv_counter_set_data query_out = {
2563                 .out = counters,
2564                 .outlen = 2 * sizeof(uint64_t),
2565         };
2566         int err = mlx5_glue->query_counter_set(&query_cs_attr, &query_out);
2567
2568         if (err)
2569                 return rte_flow_error_set(error, err,
2570                                           RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
2571                                           NULL,
2572                                           "cannot read counter");
2573         query_count->hits_set = 1;
2574         query_count->bytes_set = 1;
2575         query_count->hits = counters[0] - counter_stats->hits;
2576         query_count->bytes = counters[1] - counter_stats->bytes;
2577         if (query_count->reset) {
2578                 counter_stats->hits = counters[0];
2579                 counter_stats->bytes = counters[1];
2580         }
2581         return 0;
2582 }
2583
2584 /**
2585  * Query a flows.
2586  *
2587  * @see rte_flow_query()
2588  * @see rte_flow_ops
2589  */
2590 int
2591 mlx5_flow_query(struct rte_eth_dev *dev __rte_unused,
2592                 struct rte_flow *flow,
2593                 enum rte_flow_action_type action __rte_unused,
2594                 void *data,
2595                 struct rte_flow_error *error)
2596 {
2597         if (flow->cs) {
2598                 int ret;
2599
2600                 ret = mlx5_flow_query_count(flow->cs,
2601                                             &flow->counter_stats,
2602                                             (struct rte_flow_query_count *)data,
2603                                             error);
2604                 if (ret)
2605                         return ret;
2606         } else {
2607                 return rte_flow_error_set(error, EINVAL,
2608                                           RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
2609                                           NULL,
2610                                           "no counter found for flow");
2611         }
2612         return 0;
2613 }
2614 #endif
2615
2616 /**
2617  * Isolated mode.
2618  *
2619  * @see rte_flow_isolate()
2620  * @see rte_flow_ops
2621  */
2622 int
2623 mlx5_flow_isolate(struct rte_eth_dev *dev,
2624                   int enable,
2625                   struct rte_flow_error *error)
2626 {
2627         struct priv *priv = dev->data->dev_private;
2628
2629         if (dev->data->dev_started) {
2630                 rte_flow_error_set(error, EBUSY,
2631                                    RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
2632                                    NULL,
2633                                    "port must be stopped first");
2634                 return -rte_errno;
2635         }
2636         priv->isolated = !!enable;
2637         if (enable)
2638                 priv->dev->dev_ops = &mlx5_dev_ops_isolate;
2639         else
2640                 priv->dev->dev_ops = &mlx5_dev_ops;
2641         return 0;
2642 }
2643
2644 /**
2645  * Convert a flow director filter to a generic flow.
2646  *
2647  * @param dev
2648  *   Pointer to Ethernet device.
2649  * @param fdir_filter
2650  *   Flow director filter to add.
2651  * @param attributes
2652  *   Generic flow parameters structure.
2653  *
2654  * @return
2655  *   0 on success, a negative errno value otherwise and rte_errno is set.
2656  */
2657 static int
2658 mlx5_fdir_filter_convert(struct rte_eth_dev *dev,
2659                          const struct rte_eth_fdir_filter *fdir_filter,
2660                          struct mlx5_fdir *attributes)
2661 {
2662         struct priv *priv = dev->data->dev_private;
2663         const struct rte_eth_fdir_input *input = &fdir_filter->input;
2664
2665         /* Validate queue number. */
2666         if (fdir_filter->action.rx_queue >= priv->rxqs_n) {
2667                 DRV_LOG(ERR, "port %u invalid queue number %d",
2668                         dev->data->port_id, fdir_filter->action.rx_queue);
2669                 rte_errno = EINVAL;
2670                 return -rte_errno;
2671         }
2672         attributes->attr.ingress = 1;
2673         attributes->items[0] = (struct rte_flow_item) {
2674                 .type = RTE_FLOW_ITEM_TYPE_ETH,
2675                 .spec = &attributes->l2,
2676                 .mask = &attributes->l2_mask,
2677         };
2678         switch (fdir_filter->action.behavior) {
2679         case RTE_ETH_FDIR_ACCEPT:
2680                 attributes->actions[0] = (struct rte_flow_action){
2681                         .type = RTE_FLOW_ACTION_TYPE_QUEUE,
2682                         .conf = &attributes->queue,
2683                 };
2684                 break;
2685         case RTE_ETH_FDIR_REJECT:
2686                 attributes->actions[0] = (struct rte_flow_action){
2687                         .type = RTE_FLOW_ACTION_TYPE_DROP,
2688                 };
2689                 break;
2690         default:
2691                 DRV_LOG(ERR, "port %u invalid behavior %d",
2692                         dev->data->port_id,
2693                         fdir_filter->action.behavior);
2694                 rte_errno = ENOTSUP;
2695                 return -rte_errno;
2696         }
2697         attributes->queue.index = fdir_filter->action.rx_queue;
2698         switch (fdir_filter->input.flow_type) {
2699         case RTE_ETH_FLOW_NONFRAG_IPV4_UDP:
2700                 attributes->l3.ipv4.hdr = (struct ipv4_hdr){
2701                         .src_addr = input->flow.udp4_flow.ip.src_ip,
2702                         .dst_addr = input->flow.udp4_flow.ip.dst_ip,
2703                         .time_to_live = input->flow.udp4_flow.ip.ttl,
2704                         .type_of_service = input->flow.udp4_flow.ip.tos,
2705                         .next_proto_id = input->flow.udp4_flow.ip.proto,
2706                 };
2707                 attributes->l4.udp.hdr = (struct udp_hdr){
2708                         .src_port = input->flow.udp4_flow.src_port,
2709                         .dst_port = input->flow.udp4_flow.dst_port,
2710                 };
2711                 attributes->items[1] = (struct rte_flow_item){
2712                         .type = RTE_FLOW_ITEM_TYPE_IPV4,
2713                         .spec = &attributes->l3,
2714                         .mask = &attributes->l3,
2715                 };
2716                 attributes->items[2] = (struct rte_flow_item){
2717                         .type = RTE_FLOW_ITEM_TYPE_UDP,
2718                         .spec = &attributes->l4,
2719                         .mask = &attributes->l4,
2720                 };
2721                 break;
2722         case RTE_ETH_FLOW_NONFRAG_IPV4_TCP:
2723                 attributes->l3.ipv4.hdr = (struct ipv4_hdr){
2724                         .src_addr = input->flow.tcp4_flow.ip.src_ip,
2725                         .dst_addr = input->flow.tcp4_flow.ip.dst_ip,
2726                         .time_to_live = input->flow.tcp4_flow.ip.ttl,
2727                         .type_of_service = input->flow.tcp4_flow.ip.tos,
2728                         .next_proto_id = input->flow.tcp4_flow.ip.proto,
2729                 };
2730                 attributes->l4.tcp.hdr = (struct tcp_hdr){
2731                         .src_port = input->flow.tcp4_flow.src_port,
2732                         .dst_port = input->flow.tcp4_flow.dst_port,
2733                 };
2734                 attributes->items[1] = (struct rte_flow_item){
2735                         .type = RTE_FLOW_ITEM_TYPE_IPV4,
2736                         .spec = &attributes->l3,
2737                         .mask = &attributes->l3,
2738                 };
2739                 attributes->items[2] = (struct rte_flow_item){
2740                         .type = RTE_FLOW_ITEM_TYPE_TCP,
2741                         .spec = &attributes->l4,
2742                         .mask = &attributes->l4,
2743                 };
2744                 break;
2745         case RTE_ETH_FLOW_NONFRAG_IPV4_OTHER:
2746                 attributes->l3.ipv4.hdr = (struct ipv4_hdr){
2747                         .src_addr = input->flow.ip4_flow.src_ip,
2748                         .dst_addr = input->flow.ip4_flow.dst_ip,
2749                         .time_to_live = input->flow.ip4_flow.ttl,
2750                         .type_of_service = input->flow.ip4_flow.tos,
2751                         .next_proto_id = input->flow.ip4_flow.proto,
2752                 };
2753                 attributes->items[1] = (struct rte_flow_item){
2754                         .type = RTE_FLOW_ITEM_TYPE_IPV4,
2755                         .spec = &attributes->l3,
2756                         .mask = &attributes->l3,
2757                 };
2758                 break;
2759         case RTE_ETH_FLOW_NONFRAG_IPV6_UDP:
2760                 attributes->l3.ipv6.hdr = (struct ipv6_hdr){
2761                         .hop_limits = input->flow.udp6_flow.ip.hop_limits,
2762                         .proto = input->flow.udp6_flow.ip.proto,
2763                 };
2764                 memcpy(attributes->l3.ipv6.hdr.src_addr,
2765                        input->flow.udp6_flow.ip.src_ip,
2766                        RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
2767                 memcpy(attributes->l3.ipv6.hdr.dst_addr,
2768                        input->flow.udp6_flow.ip.dst_ip,
2769                        RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
2770                 attributes->l4.udp.hdr = (struct udp_hdr){
2771                         .src_port = input->flow.udp6_flow.src_port,
2772                         .dst_port = input->flow.udp6_flow.dst_port,
2773                 };
2774                 attributes->items[1] = (struct rte_flow_item){
2775                         .type = RTE_FLOW_ITEM_TYPE_IPV6,
2776                         .spec = &attributes->l3,
2777                         .mask = &attributes->l3,
2778                 };
2779                 attributes->items[2] = (struct rte_flow_item){
2780                         .type = RTE_FLOW_ITEM_TYPE_UDP,
2781                         .spec = &attributes->l4,
2782                         .mask = &attributes->l4,
2783                 };
2784                 break;
2785         case RTE_ETH_FLOW_NONFRAG_IPV6_TCP:
2786                 attributes->l3.ipv6.hdr = (struct ipv6_hdr){
2787                         .hop_limits = input->flow.tcp6_flow.ip.hop_limits,
2788                         .proto = input->flow.tcp6_flow.ip.proto,
2789                 };
2790                 memcpy(attributes->l3.ipv6.hdr.src_addr,
2791                        input->flow.tcp6_flow.ip.src_ip,
2792                        RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
2793                 memcpy(attributes->l3.ipv6.hdr.dst_addr,
2794                        input->flow.tcp6_flow.ip.dst_ip,
2795                        RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
2796                 attributes->l4.tcp.hdr = (struct tcp_hdr){
2797                         .src_port = input->flow.tcp6_flow.src_port,
2798                         .dst_port = input->flow.tcp6_flow.dst_port,
2799                 };
2800                 attributes->items[1] = (struct rte_flow_item){
2801                         .type = RTE_FLOW_ITEM_TYPE_IPV6,
2802                         .spec = &attributes->l3,
2803                         .mask = &attributes->l3,
2804                 };
2805                 attributes->items[2] = (struct rte_flow_item){
2806                         .type = RTE_FLOW_ITEM_TYPE_TCP,
2807                         .spec = &attributes->l4,
2808                         .mask = &attributes->l4,
2809                 };
2810                 break;
2811         case RTE_ETH_FLOW_NONFRAG_IPV6_OTHER:
2812                 attributes->l3.ipv6.hdr = (struct ipv6_hdr){
2813                         .hop_limits = input->flow.ipv6_flow.hop_limits,
2814                         .proto = input->flow.ipv6_flow.proto,
2815                 };
2816                 memcpy(attributes->l3.ipv6.hdr.src_addr,
2817                        input->flow.ipv6_flow.src_ip,
2818                        RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
2819                 memcpy(attributes->l3.ipv6.hdr.dst_addr,
2820                        input->flow.ipv6_flow.dst_ip,
2821                        RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
2822                 attributes->items[1] = (struct rte_flow_item){
2823                         .type = RTE_FLOW_ITEM_TYPE_IPV6,
2824                         .spec = &attributes->l3,
2825                         .mask = &attributes->l3,
2826                 };
2827                 break;
2828         default:
2829                 DRV_LOG(ERR, "port %u invalid flow type%d",
2830                         dev->data->port_id, fdir_filter->input.flow_type);
2831                 rte_errno = ENOTSUP;
2832                 return -rte_errno;
2833         }
2834         return 0;
2835 }
2836
2837 /**
2838  * Add new flow director filter and store it in list.
2839  *
2840  * @param dev
2841  *   Pointer to Ethernet device.
2842  * @param fdir_filter
2843  *   Flow director filter to add.
2844  *
2845  * @return
2846  *   0 on success, a negative errno value otherwise and rte_errno is set.
2847  */
2848 static int
2849 mlx5_fdir_filter_add(struct rte_eth_dev *dev,
2850                      const struct rte_eth_fdir_filter *fdir_filter)
2851 {
2852         struct priv *priv = dev->data->dev_private;
2853         struct mlx5_fdir attributes = {
2854                 .attr.group = 0,
2855                 .l2_mask = {
2856                         .dst.addr_bytes = "\x00\x00\x00\x00\x00\x00",
2857                         .src.addr_bytes = "\x00\x00\x00\x00\x00\x00",
2858                         .type = 0,
2859                 },
2860         };
2861         struct mlx5_flow_parse parser = {
2862                 .layer = HASH_RXQ_ETH,
2863         };
2864         struct rte_flow_error error;
2865         struct rte_flow *flow;
2866         int ret;
2867
2868         ret = mlx5_fdir_filter_convert(dev, fdir_filter, &attributes);
2869         if (ret)
2870                 return ret;
2871         ret = mlx5_flow_convert(dev, &attributes.attr, attributes.items,
2872                                 attributes.actions, &error, &parser);
2873         if (ret)
2874                 return ret;
2875         flow = mlx5_flow_list_create(dev, &priv->flows, &attributes.attr,
2876                                      attributes.items, attributes.actions,
2877                                      &error);
2878         if (flow) {
2879                 DRV_LOG(DEBUG, "port %u FDIR created %p", dev->data->port_id,
2880                         (void *)flow);
2881                 return 0;
2882         }
2883         return -rte_errno;
2884 }
2885
2886 /**
2887  * Delete specific filter.
2888  *
2889  * @param dev
2890  *   Pointer to Ethernet device.
2891  * @param fdir_filter
2892  *   Filter to be deleted.
2893  *
2894  * @return
2895  *   0 on success, a negative errno value otherwise and rte_errno is set.
2896  */
2897 static int
2898 mlx5_fdir_filter_delete(struct rte_eth_dev *dev,
2899                         const struct rte_eth_fdir_filter *fdir_filter)
2900 {
2901         struct priv *priv = dev->data->dev_private;
2902         struct mlx5_fdir attributes = {
2903                 .attr.group = 0,
2904         };
2905         struct mlx5_flow_parse parser = {
2906                 .create = 1,
2907                 .layer = HASH_RXQ_ETH,
2908         };
2909         struct rte_flow_error error;
2910         struct rte_flow *flow;
2911         unsigned int i;
2912         int ret;
2913
2914         ret = mlx5_fdir_filter_convert(dev, fdir_filter, &attributes);
2915         if (ret)
2916                 return ret;
2917         ret = mlx5_flow_convert(dev, &attributes.attr, attributes.items,
2918                                 attributes.actions, &error, &parser);
2919         if (ret)
2920                 goto exit;
2921         /*
2922          * Special case for drop action which is only set in the
2923          * specifications when the flow is created.  In this situation the
2924          * drop specification is missing.
2925          */
2926         if (parser.drop) {
2927                 struct ibv_flow_spec_action_drop *drop;
2928
2929                 drop = (void *)((uintptr_t)parser.queue[HASH_RXQ_ETH].ibv_attr +
2930                                 parser.queue[HASH_RXQ_ETH].offset);
2931                 *drop = (struct ibv_flow_spec_action_drop){
2932                         .type = IBV_FLOW_SPEC_ACTION_DROP,
2933                         .size = sizeof(struct ibv_flow_spec_action_drop),
2934                 };
2935                 parser.queue[HASH_RXQ_ETH].ibv_attr->num_of_specs++;
2936         }
2937         TAILQ_FOREACH(flow, &priv->flows, next) {
2938                 struct ibv_flow_attr *attr;
2939                 struct ibv_spec_header *attr_h;
2940                 void *spec;
2941                 struct ibv_flow_attr *flow_attr;
2942                 struct ibv_spec_header *flow_h;
2943                 void *flow_spec;
2944                 unsigned int specs_n;
2945
2946                 attr = parser.queue[HASH_RXQ_ETH].ibv_attr;
2947                 flow_attr = flow->frxq[HASH_RXQ_ETH].ibv_attr;
2948                 /* Compare first the attributes. */
2949                 if (memcmp(attr, flow_attr, sizeof(struct ibv_flow_attr)))
2950                         continue;
2951                 if (attr->num_of_specs == 0)
2952                         continue;
2953                 spec = (void *)((uintptr_t)attr +
2954                                 sizeof(struct ibv_flow_attr));
2955                 flow_spec = (void *)((uintptr_t)flow_attr +
2956                                      sizeof(struct ibv_flow_attr));
2957                 specs_n = RTE_MIN(attr->num_of_specs, flow_attr->num_of_specs);
2958                 for (i = 0; i != specs_n; ++i) {
2959                         attr_h = spec;
2960                         flow_h = flow_spec;
2961                         if (memcmp(spec, flow_spec,
2962                                    RTE_MIN(attr_h->size, flow_h->size)))
2963                                 goto wrong_flow;
2964                         spec = (void *)((uintptr_t)spec + attr_h->size);
2965                         flow_spec = (void *)((uintptr_t)flow_spec +
2966                                              flow_h->size);
2967                 }
2968                 /* At this point, the flow match. */
2969                 break;
2970 wrong_flow:
2971                 /* The flow does not match. */
2972                 continue;
2973         }
2974         ret = rte_errno; /* Save rte_errno before cleanup. */
2975         if (flow)
2976                 mlx5_flow_list_destroy(dev, &priv->flows, flow);
2977 exit:
2978         for (i = 0; i != hash_rxq_init_n; ++i) {
2979                 if (parser.queue[i].ibv_attr)
2980                         rte_free(parser.queue[i].ibv_attr);
2981         }
2982         rte_errno = ret; /* Restore rte_errno. */
2983         return -rte_errno;
2984 }
2985
2986 /**
2987  * Update queue for specific filter.
2988  *
2989  * @param dev
2990  *   Pointer to Ethernet device.
2991  * @param fdir_filter
2992  *   Filter to be updated.
2993  *
2994  * @return
2995  *   0 on success, a negative errno value otherwise and rte_errno is set.
2996  */
2997 static int
2998 mlx5_fdir_filter_update(struct rte_eth_dev *dev,
2999                         const struct rte_eth_fdir_filter *fdir_filter)
3000 {
3001         int ret;
3002
3003         ret = mlx5_fdir_filter_delete(dev, fdir_filter);
3004         if (ret)
3005                 return ret;
3006         return mlx5_fdir_filter_add(dev, fdir_filter);
3007 }
3008
3009 /**
3010  * Flush all filters.
3011  *
3012  * @param dev
3013  *   Pointer to Ethernet device.
3014  */
3015 static void
3016 mlx5_fdir_filter_flush(struct rte_eth_dev *dev)
3017 {
3018         struct priv *priv = dev->data->dev_private;
3019
3020         mlx5_flow_list_flush(dev, &priv->flows);
3021 }
3022
3023 /**
3024  * Get flow director information.
3025  *
3026  * @param dev
3027  *   Pointer to Ethernet device.
3028  * @param[out] fdir_info
3029  *   Resulting flow director information.
3030  */
3031 static void
3032 mlx5_fdir_info_get(struct rte_eth_dev *dev, struct rte_eth_fdir_info *fdir_info)
3033 {
3034         struct priv *priv = dev->data->dev_private;
3035         struct rte_eth_fdir_masks *mask =
3036                 &priv->dev->data->dev_conf.fdir_conf.mask;
3037
3038         fdir_info->mode = priv->dev->data->dev_conf.fdir_conf.mode;
3039         fdir_info->guarant_spc = 0;
3040         rte_memcpy(&fdir_info->mask, mask, sizeof(fdir_info->mask));
3041         fdir_info->max_flexpayload = 0;
3042         fdir_info->flow_types_mask[0] = 0;
3043         fdir_info->flex_payload_unit = 0;
3044         fdir_info->max_flex_payload_segment_num = 0;
3045         fdir_info->flex_payload_limit = 0;
3046         memset(&fdir_info->flex_conf, 0, sizeof(fdir_info->flex_conf));
3047 }
3048
3049 /**
3050  * Deal with flow director operations.
3051  *
3052  * @param dev
3053  *   Pointer to Ethernet device.
3054  * @param filter_op
3055  *   Operation to perform.
3056  * @param arg
3057  *   Pointer to operation-specific structure.
3058  *
3059  * @return
3060  *   0 on success, a negative errno value otherwise and rte_errno is set.
3061  */
3062 static int
3063 mlx5_fdir_ctrl_func(struct rte_eth_dev *dev, enum rte_filter_op filter_op,
3064                     void *arg)
3065 {
3066         struct priv *priv = dev->data->dev_private;
3067         enum rte_fdir_mode fdir_mode =
3068                 priv->dev->data->dev_conf.fdir_conf.mode;
3069
3070         if (filter_op == RTE_ETH_FILTER_NOP)
3071                 return 0;
3072         if (fdir_mode != RTE_FDIR_MODE_PERFECT &&
3073             fdir_mode != RTE_FDIR_MODE_PERFECT_MAC_VLAN) {
3074                 DRV_LOG(ERR, "port %u flow director mode %d not supported",
3075                         dev->data->port_id, fdir_mode);
3076                 rte_errno = EINVAL;
3077                 return -rte_errno;
3078         }
3079         switch (filter_op) {
3080         case RTE_ETH_FILTER_ADD:
3081                 return mlx5_fdir_filter_add(dev, arg);
3082         case RTE_ETH_FILTER_UPDATE:
3083                 return mlx5_fdir_filter_update(dev, arg);
3084         case RTE_ETH_FILTER_DELETE:
3085                 return mlx5_fdir_filter_delete(dev, arg);
3086         case RTE_ETH_FILTER_FLUSH:
3087                 mlx5_fdir_filter_flush(dev);
3088                 break;
3089         case RTE_ETH_FILTER_INFO:
3090                 mlx5_fdir_info_get(dev, arg);
3091                 break;
3092         default:
3093                 DRV_LOG(DEBUG, "port %u unknown operation %u",
3094                         dev->data->port_id, filter_op);
3095                 rte_errno = EINVAL;
3096                 return -rte_errno;
3097         }
3098         return 0;
3099 }
3100
3101 /**
3102  * Manage filter operations.
3103  *
3104  * @param dev
3105  *   Pointer to Ethernet device structure.
3106  * @param filter_type
3107  *   Filter type.
3108  * @param filter_op
3109  *   Operation to perform.
3110  * @param arg
3111  *   Pointer to operation-specific structure.
3112  *
3113  * @return
3114  *   0 on success, a negative errno value otherwise and rte_errno is set.
3115  */
3116 int
3117 mlx5_dev_filter_ctrl(struct rte_eth_dev *dev,
3118                      enum rte_filter_type filter_type,
3119                      enum rte_filter_op filter_op,
3120                      void *arg)
3121 {
3122         switch (filter_type) {
3123         case RTE_ETH_FILTER_GENERIC:
3124                 if (filter_op != RTE_ETH_FILTER_GET) {
3125                         rte_errno = EINVAL;
3126                         return -rte_errno;
3127                 }
3128                 *(const void **)arg = &mlx5_flow_ops;
3129                 return 0;
3130         case RTE_ETH_FILTER_FDIR:
3131                 return mlx5_fdir_ctrl_func(dev, filter_op, arg);
3132         default:
3133                 DRV_LOG(ERR, "port %u filter type (%d) not supported",
3134                         dev->data->port_id, filter_type);
3135                 rte_errno = ENOTSUP;
3136                 return -rte_errno;
3137         }
3138         return 0;
3139 }