drivers/net/mlx5/mlx5_flow.c

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