drivers/net/mlx5/mlx5_flow.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright 2016 6WIND S.A.
   3  * Copyright 2016 Mellanox Technologies, Ltd
   4  */
   5
   6 #include <sys/queue.h>
   7 #include <stdint.h>
   8 #include <string.h>
   9
  10 /* Verbs header. */
  11 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
  12 #ifdef PEDANTIC
  13 #pragma GCC diagnostic ignored "-Wpedantic"
  14 #endif
  15 #include <infiniband/verbs.h>
  16 #ifdef PEDANTIC
  17 #pragma GCC diagnostic error "-Wpedantic"
  18 #endif
  19
  20 #include <rte_common.h>
  21 #include <rte_eth_ctrl.h>
  22 #include <rte_ethdev_driver.h>
  23 #include <rte_flow.h>
  24 #include <rte_flow_driver.h>
  25 #include <rte_malloc.h>
  26 #include <rte_ip.h>
  27
  28 #include "mlx5.h"
  29 #include "mlx5_defs.h"
  30 #include "mlx5_prm.h"
  31 #include "mlx5_glue.h"
  32
  33 /* Define minimal priority for control plane flows. */
  34 #define MLX5_CTRL_FLOW_PRIORITY 4
  35
  36 /* Internet Protocol versions. */
  37 #define MLX5_IPV4 4
  38 #define MLX5_IPV6 6
  39
  40 #ifndef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
  41 struct ibv_flow_spec_counter_action {
  42         int dummy;
  43 };
  44 #endif
  45
  46 /* Dev ops structure defined in mlx5.c */
  47 extern const struct eth_dev_ops mlx5_dev_ops;
  48 extern const struct eth_dev_ops mlx5_dev_ops_isolate;
  49
  50 /** Structure give to the conversion functions. */
  51 struct mlx5_flow_data {
  52         struct mlx5_flow_parse *parser; /** Parser context. */
  53         struct rte_flow_error *error; /** Error context. */
  54 };
  55
  56 static int
  57 mlx5_flow_create_eth(const struct rte_flow_item *item,
  58                      const void *default_mask,
  59                      struct mlx5_flow_data *data);
  60
  61 static int
  62 mlx5_flow_create_vlan(const struct rte_flow_item *item,
  63                       const void *default_mask,
  64                       struct mlx5_flow_data *data);
  65
  66 static int
  67 mlx5_flow_create_ipv4(const struct rte_flow_item *item,
  68                       const void *default_mask,
  69                       struct mlx5_flow_data *data);
  70
  71 static int
  72 mlx5_flow_create_ipv6(const struct rte_flow_item *item,
  73                       const void *default_mask,
  74                       struct mlx5_flow_data *data);
  75
  76 static int
  77 mlx5_flow_create_udp(const struct rte_flow_item *item,
  78                      const void *default_mask,
  79                      struct mlx5_flow_data *data);
  80
  81 static int
  82 mlx5_flow_create_tcp(const struct rte_flow_item *item,
  83                      const void *default_mask,
  84                      struct mlx5_flow_data *data);
  85
  86 static int
  87 mlx5_flow_create_vxlan(const struct rte_flow_item *item,
  88                        const void *default_mask,
  89                        struct mlx5_flow_data *data);
  90
  91 struct mlx5_flow_parse;
  92
  93 static void
  94 mlx5_flow_create_copy(struct mlx5_flow_parse *parser, void *src,
  95                       unsigned int size);
  96
  97 static int
  98 mlx5_flow_create_flag_mark(struct mlx5_flow_parse *parser, uint32_t mark_id);
  99
 100 static int
 101 mlx5_flow_create_count(struct rte_eth_dev *dev, struct mlx5_flow_parse *parser);
 102
 103 /* Hash RX queue types. */
 104 enum hash_rxq_type {
 105         HASH_RXQ_TCPV4,
 106         HASH_RXQ_UDPV4,
 107         HASH_RXQ_IPV4,
 108         HASH_RXQ_TCPV6,
 109         HASH_RXQ_UDPV6,
 110         HASH_RXQ_IPV6,
 111         HASH_RXQ_ETH,
 112 };
 113
 114 /* Initialization data for hash RX queue. */
 115 struct hash_rxq_init {
 116         uint64_t hash_fields; /* Fields that participate in the hash. */
 117         uint64_t dpdk_rss_hf; /* Matching DPDK RSS hash fields. */
 118         unsigned int flow_priority; /* Flow priority to use. */
 119         unsigned int ip_version; /* Internet protocol. */
 120 };
 121
 122 /* Initialization data for hash RX queues. */
 123 const struct hash_rxq_init hash_rxq_init[] = {
 124         [HASH_RXQ_TCPV4] = {
 125                 .hash_fields = (IBV_RX_HASH_SRC_IPV4 |
 126                                 IBV_RX_HASH_DST_IPV4 |
 127                                 IBV_RX_HASH_SRC_PORT_TCP |
 128                                 IBV_RX_HASH_DST_PORT_TCP),
 129                 .dpdk_rss_hf = ETH_RSS_NONFRAG_IPV4_TCP,
 130                 .flow_priority = 1,
 131                 .ip_version = MLX5_IPV4,
 132         },
 133         [HASH_RXQ_UDPV4] = {
 134                 .hash_fields = (IBV_RX_HASH_SRC_IPV4 |
 135                                 IBV_RX_HASH_DST_IPV4 |
 136                                 IBV_RX_HASH_SRC_PORT_UDP |
 137                                 IBV_RX_HASH_DST_PORT_UDP),
 138                 .dpdk_rss_hf = ETH_RSS_NONFRAG_IPV4_UDP,
 139                 .flow_priority = 1,
 140                 .ip_version = MLX5_IPV4,
 141         },
 142         [HASH_RXQ_IPV4] = {
 143                 .hash_fields = (IBV_RX_HASH_SRC_IPV4 |
 144                                 IBV_RX_HASH_DST_IPV4),
 145                 .dpdk_rss_hf = (ETH_RSS_IPV4 |
 146                                 ETH_RSS_FRAG_IPV4),
 147                 .flow_priority = 2,
 148                 .ip_version = MLX5_IPV4,
 149         },
 150         [HASH_RXQ_TCPV6] = {
 151                 .hash_fields = (IBV_RX_HASH_SRC_IPV6 |
 152                                 IBV_RX_HASH_DST_IPV6 |
 153                                 IBV_RX_HASH_SRC_PORT_TCP |
 154                                 IBV_RX_HASH_DST_PORT_TCP),
 155                 .dpdk_rss_hf = ETH_RSS_NONFRAG_IPV6_TCP,
 156                 .flow_priority = 1,
 157                 .ip_version = MLX5_IPV6,
 158         },
 159         [HASH_RXQ_UDPV6] = {
 160                 .hash_fields = (IBV_RX_HASH_SRC_IPV6 |
 161                                 IBV_RX_HASH_DST_IPV6 |
 162                                 IBV_RX_HASH_SRC_PORT_UDP |
 163                                 IBV_RX_HASH_DST_PORT_UDP),
 164                 .dpdk_rss_hf = ETH_RSS_NONFRAG_IPV6_UDP,
 165                 .flow_priority = 1,
 166                 .ip_version = MLX5_IPV6,
 167         },
 168         [HASH_RXQ_IPV6] = {
 169                 .hash_fields = (IBV_RX_HASH_SRC_IPV6 |
 170                                 IBV_RX_HASH_DST_IPV6),
 171                 .dpdk_rss_hf = (ETH_RSS_IPV6 |
 172                                 ETH_RSS_FRAG_IPV6),
 173                 .flow_priority = 2,
 174                 .ip_version = MLX5_IPV6,
 175         },
 176         [HASH_RXQ_ETH] = {
 177                 .hash_fields = 0,
 178                 .dpdk_rss_hf = 0,
 179                 .flow_priority = 3,
 180         },
 181 };
 182
 183 /* Number of entries in hash_rxq_init[]. */
 184 const unsigned int hash_rxq_init_n = RTE_DIM(hash_rxq_init);
 185
 186 /** Structure for holding counter stats. */
 187 struct mlx5_flow_counter_stats {
 188         uint64_t hits; /**< Number of packets matched by the rule. */
 189         uint64_t bytes; /**< Number of bytes matched by the rule. */
 190 };
 191
 192 /** Structure for Drop queue. */
 193 struct mlx5_hrxq_drop {
 194         struct ibv_rwq_ind_table *ind_table; /**< Indirection table. */
 195         struct ibv_qp *qp; /**< Verbs queue pair. */
 196         struct ibv_wq *wq; /**< Verbs work queue. */
 197         struct ibv_cq *cq; /**< Verbs completion queue. */
 198 };
 199
 200 /* Flows structures. */
 201 struct mlx5_flow {
 202         uint64_t hash_fields; /**< Fields that participate in the hash. */
 203         struct ibv_flow_attr *ibv_attr; /**< Pointer to Verbs attributes. */
 204         struct ibv_flow *ibv_flow; /**< Verbs flow. */
 205         struct mlx5_hrxq *hrxq; /**< Hash Rx queues. */
 206 };
 207
 208 /* Drop flows structures. */
 209 struct mlx5_flow_drop {
 210         struct ibv_flow_attr *ibv_attr; /**< Pointer to Verbs attributes. */
 211         struct ibv_flow *ibv_flow; /**< Verbs flow. */
 212 };
 213
 214 struct rte_flow {
 215         TAILQ_ENTRY(rte_flow) next; /**< Pointer to the next flow structure. */
 216         uint32_t mark:1; /**< Set if the flow is marked. */
 217         uint32_t drop:1; /**< Drop queue. */
 218         struct rte_flow_action_rss rss_conf; /**< RSS configuration */
 219         uint16_t (*queues)[]; /**< Queues indexes to use. */
 220         uint8_t rss_key[40]; /**< copy of the RSS key. */
 221         struct ibv_counter_set *cs; /**< Holds the counters for the rule. */
 222         struct mlx5_flow_counter_stats counter_stats;/**<The counter stats. */
 223         struct mlx5_flow frxq[RTE_DIM(hash_rxq_init)];
 224         /**< Flow with Rx queue. */
 225 };
 226
 227 /** Static initializer for items. */
 228 #define ITEMS(...) \
 229         (const enum rte_flow_item_type []){ \
 230                 __VA_ARGS__, RTE_FLOW_ITEM_TYPE_END, \
 231         }
 232
 233 /** Structure to generate a simple graph of layers supported by the NIC. */
 234 struct mlx5_flow_items {
 235         /** List of possible actions for these items. */
 236         const enum rte_flow_action_type *const actions;
 237         /** Bit-masks corresponding to the possibilities for the item. */
 238         const void *mask;
 239         /**
 240          * Default bit-masks to use when item->mask is not provided. When
 241          * \default_mask is also NULL, the full supported bit-mask (\mask) is
 242          * used instead.
 243          */
 244         const void *default_mask;
 245         /** Bit-masks size in bytes. */
 246         const unsigned int mask_sz;
 247         /**
 248          * Conversion function from rte_flow to NIC specific flow.
 249          *
 250          * @param item
 251          *   rte_flow item to convert.
 252          * @param default_mask
 253          *   Default bit-masks to use when item->mask is not provided.
 254          * @param data
 255          *   Internal structure to store the conversion.
 256          *
 257          * @return
 258          *   0 on success, a negative errno value otherwise and rte_errno is
 259          *   set.
 260          */
 261         int (*convert)(const struct rte_flow_item *item,
 262                        const void *default_mask,
 263                        struct mlx5_flow_data *data);
 264         /** Size in bytes of the destination structure. */
 265         const unsigned int dst_sz;
 266         /** List of possible following items.  */
 267         const enum rte_flow_item_type *const items;
 268 };
 269
 270 /** Valid action for this PMD. */
 271 static const enum rte_flow_action_type valid_actions[] = {
 272         RTE_FLOW_ACTION_TYPE_DROP,
 273         RTE_FLOW_ACTION_TYPE_QUEUE,
 274         RTE_FLOW_ACTION_TYPE_MARK,
 275         RTE_FLOW_ACTION_TYPE_FLAG,
 276 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
 277         RTE_FLOW_ACTION_TYPE_COUNT,
 278 #endif
 279         RTE_FLOW_ACTION_TYPE_END,
 280 };
 281
 282 /** Graph of supported items and associated actions. */
 283 static const struct mlx5_flow_items mlx5_flow_items[] = {
 284         [RTE_FLOW_ITEM_TYPE_END] = {
 285                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_ETH,
 286                                RTE_FLOW_ITEM_TYPE_VXLAN),
 287         },
 288         [RTE_FLOW_ITEM_TYPE_ETH] = {
 289                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_VLAN,
 290                                RTE_FLOW_ITEM_TYPE_IPV4,
 291                                RTE_FLOW_ITEM_TYPE_IPV6),
 292                 .actions = valid_actions,
 293                 .mask = &(const struct rte_flow_item_eth){
 294                         .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 295                         .src.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 296                         .type = -1,
 297                 },
 298                 .default_mask = &rte_flow_item_eth_mask,
 299                 .mask_sz = sizeof(struct rte_flow_item_eth),
 300                 .convert = mlx5_flow_create_eth,
 301                 .dst_sz = sizeof(struct ibv_flow_spec_eth),
 302         },
 303         [RTE_FLOW_ITEM_TYPE_VLAN] = {
 304                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_IPV4,
 305                                RTE_FLOW_ITEM_TYPE_IPV6),
 306                 .actions = valid_actions,
 307                 .mask = &(const struct rte_flow_item_vlan){
 308                         .tci = -1,
 309                 },
 310                 .default_mask = &rte_flow_item_vlan_mask,
 311                 .mask_sz = sizeof(struct rte_flow_item_vlan),
 312                 .convert = mlx5_flow_create_vlan,
 313                 .dst_sz = 0,
 314         },
 315         [RTE_FLOW_ITEM_TYPE_IPV4] = {
 316                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP,
 317                                RTE_FLOW_ITEM_TYPE_TCP),
 318                 .actions = valid_actions,
 319                 .mask = &(const struct rte_flow_item_ipv4){
 320                         .hdr = {
 321                                 .src_addr = -1,
 322                                 .dst_addr = -1,
 323                                 .type_of_service = -1,
 324                                 .next_proto_id = -1,
 325                         },
 326                 },
 327                 .default_mask = &rte_flow_item_ipv4_mask,
 328                 .mask_sz = sizeof(struct rte_flow_item_ipv4),
 329                 .convert = mlx5_flow_create_ipv4,
 330                 .dst_sz = sizeof(struct ibv_flow_spec_ipv4_ext),
 331         },
 332         [RTE_FLOW_ITEM_TYPE_IPV6] = {
 333                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP,
 334                                RTE_FLOW_ITEM_TYPE_TCP),
 335                 .actions = valid_actions,
 336                 .mask = &(const struct rte_flow_item_ipv6){
 337                         .hdr = {
 338                                 .src_addr = {
 339                                         0xff, 0xff, 0xff, 0xff,
 340                                         0xff, 0xff, 0xff, 0xff,
 341                                         0xff, 0xff, 0xff, 0xff,
 342                                         0xff, 0xff, 0xff, 0xff,
 343                                 },
 344                                 .dst_addr = {
 345                                         0xff, 0xff, 0xff, 0xff,
 346                                         0xff, 0xff, 0xff, 0xff,
 347                                         0xff, 0xff, 0xff, 0xff,
 348                                         0xff, 0xff, 0xff, 0xff,
 349                                 },
 350                                 .vtc_flow = -1,
 351                                 .proto = -1,
 352                                 .hop_limits = -1,
 353                         },
 354                 },
 355                 .default_mask = &rte_flow_item_ipv6_mask,
 356                 .mask_sz = sizeof(struct rte_flow_item_ipv6),
 357                 .convert = mlx5_flow_create_ipv6,
 358                 .dst_sz = sizeof(struct ibv_flow_spec_ipv6),
 359         },
 360         [RTE_FLOW_ITEM_TYPE_UDP] = {
 361                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_VXLAN),
 362                 .actions = valid_actions,
 363                 .mask = &(const struct rte_flow_item_udp){
 364                         .hdr = {
 365                                 .src_port = -1,
 366                                 .dst_port = -1,
 367                         },
 368                 },
 369                 .default_mask = &rte_flow_item_udp_mask,
 370                 .mask_sz = sizeof(struct rte_flow_item_udp),
 371                 .convert = mlx5_flow_create_udp,
 372                 .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
 373         },
 374         [RTE_FLOW_ITEM_TYPE_TCP] = {
 375                 .actions = valid_actions,
 376                 .mask = &(const struct rte_flow_item_tcp){
 377                         .hdr = {
 378                                 .src_port = -1,
 379                                 .dst_port = -1,
 380                         },
 381                 },
 382                 .default_mask = &rte_flow_item_tcp_mask,
 383                 .mask_sz = sizeof(struct rte_flow_item_tcp),
 384                 .convert = mlx5_flow_create_tcp,
 385                 .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
 386         },
 387         [RTE_FLOW_ITEM_TYPE_VXLAN] = {
 388                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_ETH),
 389                 .actions = valid_actions,
 390                 .mask = &(const struct rte_flow_item_vxlan){
 391                         .vni = "\xff\xff\xff",
 392                 },
 393                 .default_mask = &rte_flow_item_vxlan_mask,
 394                 .mask_sz = sizeof(struct rte_flow_item_vxlan),
 395                 .convert = mlx5_flow_create_vxlan,
 396                 .dst_sz = sizeof(struct ibv_flow_spec_tunnel),
 397         },
 398 };
 399
 400 /** Structure to pass to the conversion function. */
 401 struct mlx5_flow_parse {
 402         uint32_t inner; /**< Set once VXLAN is encountered. */
 403         uint32_t create:1;
 404         /**< Whether resources should remain after a validate. */
 405         uint32_t drop:1; /**< Target is a drop queue. */
 406         uint32_t mark:1; /**< Mark is present in the flow. */
 407         uint32_t count:1; /**< Count is present in the flow. */
 408         uint32_t mark_id; /**< Mark identifier. */
 409         struct rte_flow_action_rss rss_conf; /**< RSS configuration */
 410         uint16_t queues[RTE_MAX_QUEUES_PER_PORT]; /**< Queues indexes to use. */
 411         uint8_t rss_key[40]; /**< copy of the RSS key. */
 412         enum hash_rxq_type layer; /**< Last pattern layer detected. */
 413         struct ibv_counter_set *cs; /**< Holds the counter set for the rule */
 414         struct {
 415                 struct ibv_flow_attr *ibv_attr;
 416                 /**< Pointer to Verbs attributes. */
 417                 unsigned int offset;
 418                 /**< Current position or total size of the attribute. */
 419         } queue[RTE_DIM(hash_rxq_init)];
 420 };
 421
 422 static const struct rte_flow_ops mlx5_flow_ops = {
 423         .validate = mlx5_flow_validate,
 424         .create = mlx5_flow_create,
 425         .destroy = mlx5_flow_destroy,
 426         .flush = mlx5_flow_flush,
 427 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
 428         .query = mlx5_flow_query,
 429 #else
 430         .query = NULL,
 431 #endif
 432         .isolate = mlx5_flow_isolate,
 433 };
 434
 435 /* Convert FDIR request to Generic flow. */
 436 struct mlx5_fdir {
 437         struct rte_flow_attr attr;
 438         struct rte_flow_action actions[2];
 439         struct rte_flow_item items[4];
 440         struct rte_flow_item_eth l2;
 441         struct rte_flow_item_eth l2_mask;
 442         union {
 443                 struct rte_flow_item_ipv4 ipv4;
 444                 struct rte_flow_item_ipv6 ipv6;
 445         } l3;
 446         union {
 447                 struct rte_flow_item_ipv4 ipv4;
 448                 struct rte_flow_item_ipv6 ipv6;
 449         } l3_mask;
 450         union {
 451                 struct rte_flow_item_udp udp;
 452                 struct rte_flow_item_tcp tcp;
 453         } l4;
 454         union {
 455                 struct rte_flow_item_udp udp;
 456                 struct rte_flow_item_tcp tcp;
 457         } l4_mask;
 458         struct rte_flow_action_queue queue;
 459 };
 460
 461 /* Verbs specification header. */
 462 struct ibv_spec_header {
 463         enum ibv_flow_spec_type type;
 464         uint16_t size;
 465 };
 466
 467 /**
 468  * Check support for a given item.
 469  *
 470  * @param item[in]
 471  *   Item specification.
 472  * @param mask[in]
 473  *   Bit-masks covering supported fields to compare with spec, last and mask in
 474  *   \item.
 475  * @param size
 476  *   Bit-Mask size in bytes.
 477  *
 478  * @return
 479  *   0 on success, a negative errno value otherwise and rte_errno is set.
 480  */
 481 static int
 482 mlx5_flow_item_validate(const struct rte_flow_item *item,
 483                         const uint8_t *mask, unsigned int size)
 484 {
 485         if (!item->spec && (item->mask || item->last)) {
 486                 rte_errno = EINVAL;
 487                 return -rte_errno;
 488         }
 489         if (item->spec && !item->mask) {
 490                 unsigned int i;
 491                 const uint8_t *spec = item->spec;
 492
 493                 for (i = 0; i < size; ++i)
 494                         if ((spec[i] | mask[i]) != mask[i]) {
 495                                 rte_errno = EINVAL;
 496                                 return -rte_errno;
 497                         }
 498         }
 499         if (item->last && !item->mask) {
 500                 unsigned int i;
 501                 const uint8_t *spec = item->last;
 502
 503                 for (i = 0; i < size; ++i)
 504                         if ((spec[i] | mask[i]) != mask[i]) {
 505                                 rte_errno = EINVAL;
 506                                 return -rte_errno;
 507                         }
 508         }
 509         if (item->mask) {
 510                 unsigned int i;
 511                 const uint8_t *spec = item->spec;
 512
 513                 for (i = 0; i < size; ++i)
 514                         if ((spec[i] | mask[i]) != mask[i]) {
 515                                 rte_errno = EINVAL;
 516                                 return -rte_errno;
 517                         }
 518         }
 519         if (item->spec && item->last) {
 520                 uint8_t spec[size];
 521                 uint8_t last[size];
 522                 const uint8_t *apply = mask;
 523                 unsigned int i;
 524                 int ret;
 525
 526                 if (item->mask)
 527                         apply = item->mask;
 528                 for (i = 0; i < size; ++i) {
 529                         spec[i] = ((const uint8_t *)item->spec)[i] & apply[i];
 530                         last[i] = ((const uint8_t *)item->last)[i] & apply[i];
 531                 }
 532                 ret = memcmp(spec, last, size);
 533                 if (ret != 0) {
 534                         rte_errno = EINVAL;
 535                         return -rte_errno;
 536                 }
 537         }
 538         return 0;
 539 }
 540
 541 /**
 542  * Extract attribute to the parser.
 543  *
 544  * @param[in] attr
 545  *   Flow rule attributes.
 546  * @param[out] error
 547  *   Perform verbose error reporting if not NULL.
 548  *
 549  * @return
 550  *   0 on success, a negative errno value otherwise and rte_errno is set.
 551  */
 552 static int
 553 mlx5_flow_convert_attributes(const struct rte_flow_attr *attr,
 554                              struct rte_flow_error *error)
 555 {
 556         if (attr->group) {
 557                 rte_flow_error_set(error, ENOTSUP,
 558                                    RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
 559                                    NULL,
 560                                    "groups are not supported");
 561                 return -rte_errno;
 562         }
 563         if (attr->priority && attr->priority != MLX5_CTRL_FLOW_PRIORITY) {
 564                 rte_flow_error_set(error, ENOTSUP,
 565                                    RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
 566                                    NULL,
 567                                    "priorities are not supported");
 568                 return -rte_errno;
 569         }
 570         if (attr->egress) {
 571                 rte_flow_error_set(error, ENOTSUP,
 572                                    RTE_FLOW_ERROR_TYPE_ATTR_EGRESS,
 573                                    NULL,
 574                                    "egress is not supported");
 575                 return -rte_errno;
 576         }
 577         if (!attr->ingress) {
 578                 rte_flow_error_set(error, ENOTSUP,
 579                                    RTE_FLOW_ERROR_TYPE_ATTR_INGRESS,
 580                                    NULL,
 581                                    "only ingress is supported");
 582                 return -rte_errno;
 583         }
 584         return 0;
 585 }
 586
 587 /**
 588  * Extract actions request to the parser.
 589  *
 590  * @param dev
 591  *   Pointer to Ethernet device.
 592  * @param[in] actions
 593  *   Associated actions (list terminated by the END action).
 594  * @param[out] error
 595  *   Perform verbose error reporting if not NULL.
 596  * @param[in, out] parser
 597  *   Internal parser structure.
 598  *
 599  * @return
 600  *   0 on success, a negative errno value otherwise and rte_errno is set.
 601  */
 602 static int
 603 mlx5_flow_convert_actions(struct rte_eth_dev *dev,
 604                           const struct rte_flow_action actions[],
 605                           struct rte_flow_error *error,
 606                           struct mlx5_flow_parse *parser)
 607 {
 608         enum { FATE = 1, MARK = 2, COUNT = 4, };
 609         uint32_t overlap = 0;
 610         struct priv *priv = dev->data->dev_private;
 611
 612         for (; actions->type != RTE_FLOW_ACTION_TYPE_END; ++actions) {
 613                 if (actions->type == RTE_FLOW_ACTION_TYPE_VOID) {
 614                         continue;
 615                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_DROP) {
 616                         if (overlap & FATE)
 617                                 goto exit_action_overlap;
 618                         overlap |= FATE;
 619                         parser->drop = 1;
 620                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_QUEUE) {
 621                         const struct rte_flow_action_queue *queue =
 622                                 (const struct rte_flow_action_queue *)
 623                                 actions->conf;
 624
 625                         if (overlap & FATE)
 626                                 goto exit_action_overlap;
 627                         overlap |= FATE;
 628                         if (!queue || (queue->index > (priv->rxqs_n - 1)))
 629                                 goto exit_action_not_supported;
 630                         parser->queues[0] = queue->index;
 631                         parser->rss_conf = (struct rte_flow_action_rss){
 632                                 .queue_num = 1,
 633                                 .queue = parser->queues,
 634                         };
 635                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_RSS) {
 636                         const struct rte_flow_action_rss *rss =
 637                                 (const struct rte_flow_action_rss *)
 638                                 actions->conf;
 639                         const uint8_t *rss_key;
 640                         uint32_t rss_key_len;
 641                         uint16_t n;
 642
 643                         if (overlap & FATE)
 644                                 goto exit_action_overlap;
 645                         overlap |= FATE;
 646                         if (rss->func &&
 647                             rss->func != RTE_ETH_HASH_FUNCTION_TOEPLITZ) {
 648                                 rte_flow_error_set(error, EINVAL,
 649                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 650                                                    actions,
 651                                                    "the only supported RSS hash"
 652                                                    " function is Toeplitz");
 653                                 return -rte_errno;
 654                         }
 655                         if (rss->level) {
 656                                 rte_flow_error_set(error, EINVAL,
 657                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 658                                                    actions,
 659                                                    "a nonzero RSS encapsulation"
 660                                                    " level is not supported");
 661                                 return -rte_errno;
 662                         }
 663                         if (rss->types & MLX5_RSS_HF_MASK) {
 664                                 rte_flow_error_set(error, EINVAL,
 665                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 666                                                    actions,
 667                                                    "unsupported RSS type"
 668                                                    " requested");
 669                                 return -rte_errno;
 670                         }
 671                         if (rss->key_len) {
 672                                 rss_key_len = rss->key_len;
 673                                 rss_key = rss->key;
 674                         } else {
 675                                 rss_key_len = rss_hash_default_key_len;
 676                                 rss_key = rss_hash_default_key;
 677                         }
 678                         if (rss_key_len != RTE_DIM(parser->rss_key)) {
 679                                 rte_flow_error_set(error, EINVAL,
 680                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 681                                                    actions,
 682                                                    "RSS hash key must be"
 683                                                    " exactly 40 bytes long");
 684                                 return -rte_errno;
 685                         }
 686                         if (!rss->queue_num) {
 687                                 rte_flow_error_set(error, EINVAL,
 688                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 689                                                    actions,
 690                                                    "no valid queues");
 691                                 return -rte_errno;
 692                         }
 693                         if (rss->queue_num > RTE_DIM(parser->queues)) {
 694                                 rte_flow_error_set(error, EINVAL,
 695                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 696                                                    actions,
 697                                                    "too many queues for RSS"
 698                                                    " context");
 699                                 return -rte_errno;
 700                         }
 701                         for (n = 0; n < rss->queue_num; ++n) {
 702                                 if (rss->queue[n] >= priv->rxqs_n) {
 703                                         rte_flow_error_set(error, EINVAL,
 704                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 705                                                    actions,
 706                                                    "queue id > number of"
 707                                                    " queues");
 708                                         return -rte_errno;
 709                                 }
 710                         }
 711                         parser->rss_conf = (struct rte_flow_action_rss){
 712                                 .func = RTE_ETH_HASH_FUNCTION_DEFAULT,
 713                                 .level = 0,
 714                                 .types = rss->types,
 715                                 .key_len = rss_key_len,
 716                                 .queue_num = rss->queue_num,
 717                                 .key = memcpy(parser->rss_key, rss_key,
 718                                               sizeof(*rss_key) * rss_key_len),
 719                                 .queue = memcpy(parser->queues, rss->queue,
 720                                                 sizeof(*rss->queue) *
 721                                                 rss->queue_num),
 722                         };
 723                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_MARK) {
 724                         const struct rte_flow_action_mark *mark =
 725                                 (const struct rte_flow_action_mark *)
 726                                 actions->conf;
 727
 728                         if (overlap & MARK)
 729                                 goto exit_action_overlap;
 730                         overlap |= MARK;
 731                         if (!mark) {
 732                                 rte_flow_error_set(error, EINVAL,
 733                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 734                                                    actions,
 735                                                    "mark must be defined");
 736                                 return -rte_errno;
 737                         } else if (mark->id >= MLX5_FLOW_MARK_MAX) {
 738                                 rte_flow_error_set(error, ENOTSUP,
 739                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 740                                                    actions,
 741                                                    "mark must be between 0"
 742                                                    " and 16777199");
 743                                 return -rte_errno;
 744                         }
 745                         parser->mark = 1;
 746                         parser->mark_id = mark->id;
 747                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_FLAG) {
 748                         if (overlap & MARK)
 749                                 goto exit_action_overlap;
 750                         overlap |= MARK;
 751                         parser->mark = 1;
 752                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_COUNT &&
 753                            priv->config.flow_counter_en) {
 754                         if (overlap & COUNT)
 755                                 goto exit_action_overlap;
 756                         overlap |= COUNT;
 757                         parser->count = 1;
 758                 } else {
 759                         goto exit_action_not_supported;
 760                 }
 761         }
 762         /* When fate is unknown, drop traffic. */
 763         if (!(overlap & FATE))
 764                 parser->drop = 1;
 765         if (parser->drop && parser->mark)
 766                 parser->mark = 0;
 767         if (!parser->rss_conf.queue_num && !parser->drop) {
 768                 rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_HANDLE,
 769                                    NULL, "no valid action");
 770                 return -rte_errno;
 771         }
 772         return 0;
 773 exit_action_not_supported:
 774         rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
 775                            actions, "action not supported");
 776         return -rte_errno;
 777 exit_action_overlap:
 778         rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
 779                            actions, "overlapping actions are not supported");
 780         return -rte_errno;
 781 }
 782
 783 /**
 784  * Validate items.
 785  *
 786  * @param[in] items
 787  *   Pattern specification (list terminated by the END pattern item).
 788  * @param[out] error
 789  *   Perform verbose error reporting if not NULL.
 790  * @param[in, out] parser
 791  *   Internal parser structure.
 792  *
 793  * @return
 794  *   0 on success, a negative errno value otherwise and rte_errno is set.
 795  */
 796 static int
 797 mlx5_flow_convert_items_validate(const struct rte_flow_item items[],
 798                                  struct rte_flow_error *error,
 799                                  struct mlx5_flow_parse *parser)
 800 {
 801         const struct mlx5_flow_items *cur_item = mlx5_flow_items;
 802         unsigned int i;
 803         int ret = 0;
 804
 805         /* Initialise the offsets to start after verbs attribute. */
 806         for (i = 0; i != hash_rxq_init_n; ++i)
 807                 parser->queue[i].offset = sizeof(struct ibv_flow_attr);
 808         for (; items->type != RTE_FLOW_ITEM_TYPE_END; ++items) {
 809                 const struct mlx5_flow_items *token = NULL;
 810                 unsigned int n;
 811
 812                 if (items->type == RTE_FLOW_ITEM_TYPE_VOID)
 813                         continue;
 814                 for (i = 0;
 815                      cur_item->items &&
 816                      cur_item->items[i] != RTE_FLOW_ITEM_TYPE_END;
 817                      ++i) {
 818                         if (cur_item->items[i] == items->type) {
 819                                 token = &mlx5_flow_items[items->type];
 820                                 break;
 821                         }
 822                 }
 823                 if (!token) {
 824                         ret = -ENOTSUP;
 825                         goto exit_item_not_supported;
 826                 }
 827                 cur_item = token;
 828                 ret = mlx5_flow_item_validate(items,
 829                                               (const uint8_t *)cur_item->mask,
 830                                               cur_item->mask_sz);
 831                 if (ret)
 832                         goto exit_item_not_supported;
 833                 if (items->type == RTE_FLOW_ITEM_TYPE_VXLAN) {
 834                         if (parser->inner) {
 835                                 rte_flow_error_set(error, ENOTSUP,
 836                                                    RTE_FLOW_ERROR_TYPE_ITEM,
 837                                                    items,
 838                                                    "cannot recognize multiple"
 839                                                    " VXLAN encapsulations");
 840                                 return -rte_errno;
 841                         }
 842                         parser->inner = IBV_FLOW_SPEC_INNER;
 843                 }
 844                 if (parser->drop) {
 845                         parser->queue[HASH_RXQ_ETH].offset += cur_item->dst_sz;
 846                 } else {
 847                         for (n = 0; n != hash_rxq_init_n; ++n)
 848                                 parser->queue[n].offset += cur_item->dst_sz;
 849                 }
 850         }
 851         if (parser->drop) {
 852                 parser->queue[HASH_RXQ_ETH].offset +=
 853                         sizeof(struct ibv_flow_spec_action_drop);
 854         }
 855         if (parser->mark) {
 856                 for (i = 0; i != hash_rxq_init_n; ++i)
 857                         parser->queue[i].offset +=
 858                                 sizeof(struct ibv_flow_spec_action_tag);
 859         }
 860         if (parser->count) {
 861                 unsigned int size = sizeof(struct ibv_flow_spec_counter_action);
 862
 863                 for (i = 0; i != hash_rxq_init_n; ++i)
 864                         parser->queue[i].offset += size;
 865         }
 866         return 0;
 867 exit_item_not_supported:
 868         return rte_flow_error_set(error, -ret, RTE_FLOW_ERROR_TYPE_ITEM,
 869                                   items, "item not supported");
 870 }
 871
 872 /**
 873  * Allocate memory space to store verbs flow attributes.
 874  *
 875  * @param[in] size
 876  *   Amount of byte to allocate.
 877  * @param[out] error
 878  *   Perform verbose error reporting if not NULL.
 879  *
 880  * @return
 881  *   A verbs flow attribute on success, NULL otherwise and rte_errno is set.
 882  */
 883 static struct ibv_flow_attr *
 884 mlx5_flow_convert_allocate(unsigned int size, struct rte_flow_error *error)
 885 {
 886         struct ibv_flow_attr *ibv_attr;
 887
 888         ibv_attr = rte_calloc(__func__, 1, size, 0);
 889         if (!ibv_attr) {
 890                 rte_flow_error_set(error, ENOMEM,
 891                                    RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
 892                                    NULL,
 893                                    "cannot allocate verbs spec attributes");
 894                 return NULL;
 895         }
 896         return ibv_attr;
 897 }
 898
 899 /**
 900  * Make inner packet matching with an higher priority from the non Inner
 901  * matching.
 902  *
 903  * @param[in, out] parser
 904  *   Internal parser structure.
 905  * @param attr
 906  *   User flow attribute.
 907  */
 908 static void
 909 mlx5_flow_update_priority(struct mlx5_flow_parse *parser,
 910                           const struct rte_flow_attr *attr)
 911 {
 912         unsigned int i;
 913
 914         if (parser->drop) {
 915                 parser->queue[HASH_RXQ_ETH].ibv_attr->priority =
 916                         attr->priority +
 917                         hash_rxq_init[HASH_RXQ_ETH].flow_priority;
 918                 return;
 919         }
 920         for (i = 0; i != hash_rxq_init_n; ++i) {
 921                 if (parser->queue[i].ibv_attr) {
 922                         parser->queue[i].ibv_attr->priority =
 923                                 attr->priority +
 924                                 hash_rxq_init[i].flow_priority -
 925                                 (parser->inner ? 1 : 0);
 926                 }
 927         }
 928 }
 929
 930 /**
 931  * Finalise verbs flow attributes.
 932  *
 933  * @param[in, out] parser
 934  *   Internal parser structure.
 935  */
 936 static void
 937 mlx5_flow_convert_finalise(struct mlx5_flow_parse *parser)
 938 {
 939         const unsigned int ipv4 =
 940                 hash_rxq_init[parser->layer].ip_version == MLX5_IPV4;
 941         const enum hash_rxq_type hmin = ipv4 ? HASH_RXQ_TCPV4 : HASH_RXQ_TCPV6;
 942         const enum hash_rxq_type hmax = ipv4 ? HASH_RXQ_IPV4 : HASH_RXQ_IPV6;
 943         const enum hash_rxq_type ohmin = ipv4 ? HASH_RXQ_TCPV6 : HASH_RXQ_TCPV4;
 944         const enum hash_rxq_type ohmax = ipv4 ? HASH_RXQ_IPV6 : HASH_RXQ_IPV4;
 945         const enum hash_rxq_type ip = ipv4 ? HASH_RXQ_IPV4 : HASH_RXQ_IPV6;
 946         unsigned int i;
 947
 948         /* Remove any other flow not matching the pattern. */
 949         if (parser->rss_conf.queue_num == 1 && !parser->rss_conf.types) {
 950                 for (i = 0; i != hash_rxq_init_n; ++i) {
 951                         if (i == HASH_RXQ_ETH)
 952                                 continue;
 953                         rte_free(parser->queue[i].ibv_attr);
 954                         parser->queue[i].ibv_attr = NULL;
 955                 }
 956                 return;
 957         }
 958         if (parser->layer == HASH_RXQ_ETH) {
 959                 goto fill;
 960         } else {
 961                 /*
 962                  * This layer becomes useless as the pattern define under
 963                  * layers.
 964                  */
 965                 rte_free(parser->queue[HASH_RXQ_ETH].ibv_attr);
 966                 parser->queue[HASH_RXQ_ETH].ibv_attr = NULL;
 967         }
 968         /* Remove opposite kind of layer e.g. IPv6 if the pattern is IPv4. */
 969         for (i = ohmin; i != (ohmax + 1); ++i) {
 970                 if (!parser->queue[i].ibv_attr)
 971                         continue;
 972                 rte_free(parser->queue[i].ibv_attr);
 973                 parser->queue[i].ibv_attr = NULL;
 974         }
 975         /* Remove impossible flow according to the RSS configuration. */
 976         if (hash_rxq_init[parser->layer].dpdk_rss_hf &
 977             parser->rss_conf.types) {
 978                 /* Remove any other flow. */
 979                 for (i = hmin; i != (hmax + 1); ++i) {
 980                         if ((i == parser->layer) ||
 981                              (!parser->queue[i].ibv_attr))
 982                                 continue;
 983                         rte_free(parser->queue[i].ibv_attr);
 984                         parser->queue[i].ibv_attr = NULL;
 985                 }
 986         } else  if (!parser->queue[ip].ibv_attr) {
 987                 /* no RSS possible with the current configuration. */
 988                 parser->rss_conf.queue_num = 1;
 989                 return;
 990         }
 991 fill:
 992         /*
 993          * Fill missing layers in verbs specifications, or compute the correct
 994          * offset to allocate the memory space for the attributes and
 995          * specifications.
 996          */
 997         for (i = 0; i != hash_rxq_init_n - 1; ++i) {
 998                 union {
 999                         struct ibv_flow_spec_ipv4_ext ipv4;
1000                         struct ibv_flow_spec_ipv6 ipv6;
1001                         struct ibv_flow_spec_tcp_udp udp_tcp;
1002                 } specs;
1003                 void *dst;
1004                 uint16_t size;
1005
1006                 if (i == parser->layer)
1007                         continue;
1008                 if (parser->layer == HASH_RXQ_ETH) {
1009                         if (hash_rxq_init[i].ip_version == MLX5_IPV4) {
1010                                 size = sizeof(struct ibv_flow_spec_ipv4_ext);
1011                                 specs.ipv4 = (struct ibv_flow_spec_ipv4_ext){
1012                                         .type = IBV_FLOW_SPEC_IPV4_EXT,
1013                                         .size = size,
1014                                 };
1015                         } else {
1016                                 size = sizeof(struct ibv_flow_spec_ipv6);
1017                                 specs.ipv6 = (struct ibv_flow_spec_ipv6){
1018                                         .type = IBV_FLOW_SPEC_IPV6,
1019                                         .size = size,
1020                                 };
1021                         }
1022                         if (parser->queue[i].ibv_attr) {
1023                                 dst = (void *)((uintptr_t)
1024                                                parser->queue[i].ibv_attr +
1025                                                parser->queue[i].offset);
1026                                 memcpy(dst, &specs, size);
1027                                 ++parser->queue[i].ibv_attr->num_of_specs;
1028                         }
1029                         parser->queue[i].offset += size;
1030                 }
1031                 if ((i == HASH_RXQ_UDPV4) || (i == HASH_RXQ_TCPV4) ||
1032                     (i == HASH_RXQ_UDPV6) || (i == HASH_RXQ_TCPV6)) {
1033                         size = sizeof(struct ibv_flow_spec_tcp_udp);
1034                         specs.udp_tcp = (struct ibv_flow_spec_tcp_udp) {
1035                                 .type = ((i == HASH_RXQ_UDPV4 ||
1036                                           i == HASH_RXQ_UDPV6) ?
1037                                          IBV_FLOW_SPEC_UDP :
1038                                          IBV_FLOW_SPEC_TCP),
1039                                 .size = size,
1040                         };
1041                         if (parser->queue[i].ibv_attr) {
1042                                 dst = (void *)((uintptr_t)
1043                                                parser->queue[i].ibv_attr +
1044                                                parser->queue[i].offset);
1045                                 memcpy(dst, &specs, size);
1046                                 ++parser->queue[i].ibv_attr->num_of_specs;
1047                         }
1048                         parser->queue[i].offset += size;
1049                 }
1050         }
1051 }
1052
1053 /**
1054  * Validate and convert a flow supported by the NIC.
1055  *
1056  * @param dev
1057  *   Pointer to Ethernet device.
1058  * @param[in] attr
1059  *   Flow rule attributes.
1060  * @param[in] pattern
1061  *   Pattern specification (list terminated by the END pattern item).
1062  * @param[in] actions
1063  *   Associated actions (list terminated by the END action).
1064  * @param[out] error
1065  *   Perform verbose error reporting if not NULL.
1066  * @param[in, out] parser
1067  *   Internal parser structure.
1068  *
1069  * @return
1070  *   0 on success, a negative errno value otherwise and rte_errno is set.
1071  */
1072 static int
1073 mlx5_flow_convert(struct rte_eth_dev *dev,
1074                   const struct rte_flow_attr *attr,
1075                   const struct rte_flow_item items[],
1076                   const struct rte_flow_action actions[],
1077                   struct rte_flow_error *error,
1078                   struct mlx5_flow_parse *parser)
1079 {
1080         const struct mlx5_flow_items *cur_item = mlx5_flow_items;
1081         unsigned int i;
1082         int ret;
1083
1084         /* First step. Validate the attributes, items and actions. */
1085         *parser = (struct mlx5_flow_parse){
1086                 .create = parser->create,
1087                 .layer = HASH_RXQ_ETH,
1088                 .mark_id = MLX5_FLOW_MARK_DEFAULT,
1089         };
1090         ret = mlx5_flow_convert_attributes(attr, error);
1091         if (ret)
1092                 return ret;
1093         ret = mlx5_flow_convert_actions(dev, actions, error, parser);
1094         if (ret)
1095                 return ret;
1096         ret = mlx5_flow_convert_items_validate(items, error, parser);
1097         if (ret)
1098                 return ret;
1099         mlx5_flow_convert_finalise(parser);
1100         /*
1101          * Second step.
1102          * Allocate the memory space to store verbs specifications.
1103          */
1104         if (parser->drop) {
1105                 unsigned int offset = parser->queue[HASH_RXQ_ETH].offset;
1106
1107                 parser->queue[HASH_RXQ_ETH].ibv_attr =
1108                         mlx5_flow_convert_allocate(offset, error);
1109                 if (!parser->queue[HASH_RXQ_ETH].ibv_attr)
1110                         goto exit_enomem;
1111                 parser->queue[HASH_RXQ_ETH].offset =
1112                         sizeof(struct ibv_flow_attr);
1113         } else {
1114                 for (i = 0; i != hash_rxq_init_n; ++i) {
1115                         unsigned int offset;
1116
1117                         if (!(parser->rss_conf.types &
1118                               hash_rxq_init[i].dpdk_rss_hf) &&
1119                             (i != HASH_RXQ_ETH))
1120                                 continue;
1121                         offset = parser->queue[i].offset;
1122                         parser->queue[i].ibv_attr =
1123                                 mlx5_flow_convert_allocate(offset, error);
1124                         if (!parser->queue[i].ibv_attr)
1125                                 goto exit_enomem;
1126                         parser->queue[i].offset = sizeof(struct ibv_flow_attr);
1127                 }
1128         }
1129         /* Third step. Conversion parse, fill the specifications. */
1130         parser->inner = 0;
1131         for (; items->type != RTE_FLOW_ITEM_TYPE_END; ++items) {
1132                 struct mlx5_flow_data data = {
1133                         .parser = parser,
1134                         .error = error,
1135                 };
1136
1137                 if (items->type == RTE_FLOW_ITEM_TYPE_VOID)
1138                         continue;
1139                 cur_item = &mlx5_flow_items[items->type];
1140                 ret = cur_item->convert(items,
1141                                         (cur_item->default_mask ?
1142                                          cur_item->default_mask :
1143                                          cur_item->mask),
1144                                          &data);
1145                 if (ret)
1146                         goto exit_free;
1147         }
1148         if (parser->mark)
1149                 mlx5_flow_create_flag_mark(parser, parser->mark_id);
1150         if (parser->count && parser->create) {
1151                 mlx5_flow_create_count(dev, parser);
1152                 if (!parser->cs)
1153                         goto exit_count_error;
1154         }
1155         /*
1156          * Last step. Complete missing specification to reach the RSS
1157          * configuration.
1158          */
1159         if (!parser->drop)
1160                 mlx5_flow_convert_finalise(parser);
1161         mlx5_flow_update_priority(parser, attr);
1162 exit_free:
1163         /* Only verification is expected, all resources should be released. */
1164         if (!parser->create) {
1165                 for (i = 0; i != hash_rxq_init_n; ++i) {
1166                         if (parser->queue[i].ibv_attr) {
1167                                 rte_free(parser->queue[i].ibv_attr);
1168                                 parser->queue[i].ibv_attr = NULL;
1169                         }
1170                 }
1171         }
1172         return ret;
1173 exit_enomem:
1174         for (i = 0; i != hash_rxq_init_n; ++i) {
1175                 if (parser->queue[i].ibv_attr) {
1176                         rte_free(parser->queue[i].ibv_attr);
1177                         parser->queue[i].ibv_attr = NULL;
1178                 }
1179         }
1180         rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
1181                            NULL, "cannot allocate verbs spec attributes");
1182         return -rte_errno;
1183 exit_count_error:
1184         rte_flow_error_set(error, EINVAL, RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
1185                            NULL, "cannot create counter");
1186         return -rte_errno;
1187 }
1188
1189 /**
1190  * Copy the specification created into the flow.
1191  *
1192  * @param parser
1193  *   Internal parser structure.
1194  * @param src
1195  *   Create specification.
1196  * @param size
1197  *   Size in bytes of the specification to copy.
1198  */
1199 static void
1200 mlx5_flow_create_copy(struct mlx5_flow_parse *parser, void *src,
1201                       unsigned int size)
1202 {
1203         unsigned int i;
1204         void *dst;
1205
1206         for (i = 0; i != hash_rxq_init_n; ++i) {
1207                 if (!parser->queue[i].ibv_attr)
1208                         continue;
1209                 /* Specification must be the same l3 type or none. */
1210                 if (parser->layer == HASH_RXQ_ETH ||
1211                     (hash_rxq_init[parser->layer].ip_version ==
1212                      hash_rxq_init[i].ip_version) ||
1213                     (hash_rxq_init[i].ip_version == 0)) {
1214                         dst = (void *)((uintptr_t)parser->queue[i].ibv_attr +
1215                                         parser->queue[i].offset);
1216                         memcpy(dst, src, size);
1217                         ++parser->queue[i].ibv_attr->num_of_specs;
1218                         parser->queue[i].offset += size;
1219                 }
1220         }
1221 }
1222
1223 /**
1224  * Convert Ethernet item to Verbs specification.
1225  *
1226  * @param item[in]
1227  *   Item specification.
1228  * @param default_mask[in]
1229  *   Default bit-masks to use when item->mask is not provided.
1230  * @param data[in, out]
1231  *   User structure.
1232  *
1233  * @return
1234  *   0 on success, a negative errno value otherwise and rte_errno is set.
1235  */
1236 static int
1237 mlx5_flow_create_eth(const struct rte_flow_item *item,
1238                      const void *default_mask,
1239                      struct mlx5_flow_data *data)
1240 {
1241         const struct rte_flow_item_eth *spec = item->spec;
1242         const struct rte_flow_item_eth *mask = item->mask;
1243         struct mlx5_flow_parse *parser = data->parser;
1244         const unsigned int eth_size = sizeof(struct ibv_flow_spec_eth);
1245         struct ibv_flow_spec_eth eth = {
1246                 .type = parser->inner | IBV_FLOW_SPEC_ETH,
1247                 .size = eth_size,
1248         };
1249
1250         /* Don't update layer for the inner pattern. */
1251         if (!parser->inner)
1252                 parser->layer = HASH_RXQ_ETH;
1253         if (spec) {
1254                 unsigned int i;
1255
1256                 if (!mask)
1257                         mask = default_mask;
1258                 memcpy(&eth.val.dst_mac, spec->dst.addr_bytes, ETHER_ADDR_LEN);
1259                 memcpy(&eth.val.src_mac, spec->src.addr_bytes, ETHER_ADDR_LEN);
1260                 eth.val.ether_type = spec->type;
1261                 memcpy(&eth.mask.dst_mac, mask->dst.addr_bytes, ETHER_ADDR_LEN);
1262                 memcpy(&eth.mask.src_mac, mask->src.addr_bytes, ETHER_ADDR_LEN);
1263                 eth.mask.ether_type = mask->type;
1264                 /* Remove unwanted bits from values. */
1265                 for (i = 0; i < ETHER_ADDR_LEN; ++i) {
1266                         eth.val.dst_mac[i] &= eth.mask.dst_mac[i];
1267                         eth.val.src_mac[i] &= eth.mask.src_mac[i];
1268                 }
1269                 eth.val.ether_type &= eth.mask.ether_type;
1270         }
1271         mlx5_flow_create_copy(parser, &eth, eth_size);
1272         return 0;
1273 }
1274
1275 /**
1276  * Convert VLAN item to Verbs specification.
1277  *
1278  * @param item[in]
1279  *   Item specification.
1280  * @param default_mask[in]
1281  *   Default bit-masks to use when item->mask is not provided.
1282  * @param data[in, out]
1283  *   User structure.
1284  *
1285  * @return
1286  *   0 on success, a negative errno value otherwise and rte_errno is set.
1287  */
1288 static int
1289 mlx5_flow_create_vlan(const struct rte_flow_item *item,
1290                       const void *default_mask,
1291                       struct mlx5_flow_data *data)
1292 {
1293         const struct rte_flow_item_vlan *spec = item->spec;
1294         const struct rte_flow_item_vlan *mask = item->mask;
1295         struct mlx5_flow_parse *parser = data->parser;
1296         struct ibv_flow_spec_eth *eth;
1297         const unsigned int eth_size = sizeof(struct ibv_flow_spec_eth);
1298
1299         if (spec) {
1300                 unsigned int i;
1301                 if (!mask)
1302                         mask = default_mask;
1303
1304                 for (i = 0; i != hash_rxq_init_n; ++i) {
1305                         if (!parser->queue[i].ibv_attr)
1306                                 continue;
1307
1308                         eth = (void *)((uintptr_t)parser->queue[i].ibv_attr +
1309                                        parser->queue[i].offset - eth_size);
1310                         eth->val.vlan_tag = spec->tci;
1311                         eth->mask.vlan_tag = mask->tci;
1312                         eth->val.vlan_tag &= eth->mask.vlan_tag;
1313                         /*
1314                          * From verbs perspective an empty VLAN is equivalent
1315                          * to a packet without VLAN layer.
1316                          */
1317                         if (!eth->mask.vlan_tag)
1318                                 goto error;
1319                 }
1320                 return 0;
1321         }
1322 error:
1323         return rte_flow_error_set(data->error, EINVAL, RTE_FLOW_ERROR_TYPE_ITEM,
1324                                   item, "VLAN cannot be empty");
1325 }
1326
1327 /**
1328  * Convert IPv4 item to Verbs specification.
1329  *
1330  * @param item[in]
1331  *   Item specification.
1332  * @param default_mask[in]
1333  *   Default bit-masks to use when item->mask is not provided.
1334  * @param data[in, out]
1335  *   User structure.
1336  *
1337  * @return
1338  *   0 on success, a negative errno value otherwise and rte_errno is set.
1339  */
1340 static int
1341 mlx5_flow_create_ipv4(const struct rte_flow_item *item,
1342                       const void *default_mask,
1343                       struct mlx5_flow_data *data)
1344 {
1345         const struct rte_flow_item_ipv4 *spec = item->spec;
1346         const struct rte_flow_item_ipv4 *mask = item->mask;
1347         struct mlx5_flow_parse *parser = data->parser;
1348         unsigned int ipv4_size = sizeof(struct ibv_flow_spec_ipv4_ext);
1349         struct ibv_flow_spec_ipv4_ext ipv4 = {
1350                 .type = parser->inner | IBV_FLOW_SPEC_IPV4_EXT,
1351                 .size = ipv4_size,
1352         };
1353
1354         /* Don't update layer for the inner pattern. */
1355         if (!parser->inner)
1356                 parser->layer = HASH_RXQ_IPV4;
1357         if (spec) {
1358                 if (!mask)
1359                         mask = default_mask;
1360                 ipv4.val = (struct ibv_flow_ipv4_ext_filter){
1361                         .src_ip = spec->hdr.src_addr,
1362                         .dst_ip = spec->hdr.dst_addr,
1363                         .proto = spec->hdr.next_proto_id,
1364                         .tos = spec->hdr.type_of_service,
1365                 };
1366                 ipv4.mask = (struct ibv_flow_ipv4_ext_filter){
1367                         .src_ip = mask->hdr.src_addr,
1368                         .dst_ip = mask->hdr.dst_addr,
1369                         .proto = mask->hdr.next_proto_id,
1370                         .tos = mask->hdr.type_of_service,
1371                 };
1372                 /* Remove unwanted bits from values. */
1373                 ipv4.val.src_ip &= ipv4.mask.src_ip;
1374                 ipv4.val.dst_ip &= ipv4.mask.dst_ip;
1375                 ipv4.val.proto &= ipv4.mask.proto;
1376                 ipv4.val.tos &= ipv4.mask.tos;
1377         }
1378         mlx5_flow_create_copy(parser, &ipv4, ipv4_size);
1379         return 0;
1380 }
1381
1382 /**
1383  * Convert IPv6 item to Verbs specification.
1384  *
1385  * @param item[in]
1386  *   Item specification.
1387  * @param default_mask[in]
1388  *   Default bit-masks to use when item->mask is not provided.
1389  * @param data[in, out]
1390  *   User structure.
1391  *
1392  * @return
1393  *   0 on success, a negative errno value otherwise and rte_errno is set.
1394  */
1395 static int
1396 mlx5_flow_create_ipv6(const struct rte_flow_item *item,
1397                       const void *default_mask,
1398                       struct mlx5_flow_data *data)
1399 {
1400         const struct rte_flow_item_ipv6 *spec = item->spec;
1401         const struct rte_flow_item_ipv6 *mask = item->mask;
1402         struct mlx5_flow_parse *parser = data->parser;
1403         unsigned int ipv6_size = sizeof(struct ibv_flow_spec_ipv6);
1404         struct ibv_flow_spec_ipv6 ipv6 = {
1405                 .type = parser->inner | IBV_FLOW_SPEC_IPV6,
1406                 .size = ipv6_size,
1407         };
1408
1409         /* Don't update layer for the inner pattern. */
1410         if (!parser->inner)
1411                 parser->layer = HASH_RXQ_IPV6;
1412         if (spec) {
1413                 unsigned int i;
1414                 uint32_t vtc_flow_val;
1415                 uint32_t vtc_flow_mask;
1416
1417                 if (!mask)
1418                         mask = default_mask;
1419                 memcpy(&ipv6.val.src_ip, spec->hdr.src_addr,
1420                        RTE_DIM(ipv6.val.src_ip));
1421                 memcpy(&ipv6.val.dst_ip, spec->hdr.dst_addr,
1422                        RTE_DIM(ipv6.val.dst_ip));
1423                 memcpy(&ipv6.mask.src_ip, mask->hdr.src_addr,
1424                        RTE_DIM(ipv6.mask.src_ip));
1425                 memcpy(&ipv6.mask.dst_ip, mask->hdr.dst_addr,
1426                        RTE_DIM(ipv6.mask.dst_ip));
1427                 vtc_flow_val = rte_be_to_cpu_32(spec->hdr.vtc_flow);
1428                 vtc_flow_mask = rte_be_to_cpu_32(mask->hdr.vtc_flow);
1429                 ipv6.val.flow_label =
1430                         rte_cpu_to_be_32((vtc_flow_val & IPV6_HDR_FL_MASK) >>
1431                                          IPV6_HDR_FL_SHIFT);
1432                 ipv6.val.traffic_class = (vtc_flow_val & IPV6_HDR_TC_MASK) >>
1433                                          IPV6_HDR_TC_SHIFT;
1434                 ipv6.val.next_hdr = spec->hdr.proto;
1435                 ipv6.val.hop_limit = spec->hdr.hop_limits;
1436                 ipv6.mask.flow_label =
1437                         rte_cpu_to_be_32((vtc_flow_mask & IPV6_HDR_FL_MASK) >>
1438                                          IPV6_HDR_FL_SHIFT);
1439                 ipv6.mask.traffic_class = (vtc_flow_mask & IPV6_HDR_TC_MASK) >>
1440                                           IPV6_HDR_TC_SHIFT;
1441                 ipv6.mask.next_hdr = mask->hdr.proto;
1442                 ipv6.mask.hop_limit = mask->hdr.hop_limits;
1443                 /* Remove unwanted bits from values. */
1444                 for (i = 0; i < RTE_DIM(ipv6.val.src_ip); ++i) {
1445                         ipv6.val.src_ip[i] &= ipv6.mask.src_ip[i];
1446                         ipv6.val.dst_ip[i] &= ipv6.mask.dst_ip[i];
1447                 }
1448                 ipv6.val.flow_label &= ipv6.mask.flow_label;
1449                 ipv6.val.traffic_class &= ipv6.mask.traffic_class;
1450                 ipv6.val.next_hdr &= ipv6.mask.next_hdr;
1451                 ipv6.val.hop_limit &= ipv6.mask.hop_limit;
1452         }
1453         mlx5_flow_create_copy(parser, &ipv6, ipv6_size);
1454         return 0;
1455 }
1456
1457 /**
1458  * Convert UDP item to Verbs specification.
1459  *
1460  * @param item[in]
1461  *   Item specification.
1462  * @param default_mask[in]
1463  *   Default bit-masks to use when item->mask is not provided.
1464  * @param data[in, out]
1465  *   User structure.
1466  *
1467  * @return
1468  *   0 on success, a negative errno value otherwise and rte_errno is set.
1469  */
1470 static int
1471 mlx5_flow_create_udp(const struct rte_flow_item *item,
1472                      const void *default_mask,
1473                      struct mlx5_flow_data *data)
1474 {
1475         const struct rte_flow_item_udp *spec = item->spec;
1476         const struct rte_flow_item_udp *mask = item->mask;
1477         struct mlx5_flow_parse *parser = data->parser;
1478         unsigned int udp_size = sizeof(struct ibv_flow_spec_tcp_udp);
1479         struct ibv_flow_spec_tcp_udp udp = {
1480                 .type = parser->inner | IBV_FLOW_SPEC_UDP,
1481                 .size = udp_size,
1482         };
1483
1484         /* Don't update layer for the inner pattern. */
1485         if (!parser->inner) {
1486                 if (parser->layer == HASH_RXQ_IPV4)
1487                         parser->layer = HASH_RXQ_UDPV4;
1488                 else
1489                         parser->layer = HASH_RXQ_UDPV6;
1490         }
1491         if (spec) {
1492                 if (!mask)
1493                         mask = default_mask;
1494                 udp.val.dst_port = spec->hdr.dst_port;
1495                 udp.val.src_port = spec->hdr.src_port;
1496                 udp.mask.dst_port = mask->hdr.dst_port;
1497                 udp.mask.src_port = mask->hdr.src_port;
1498                 /* Remove unwanted bits from values. */
1499                 udp.val.src_port &= udp.mask.src_port;
1500                 udp.val.dst_port &= udp.mask.dst_port;
1501         }
1502         mlx5_flow_create_copy(parser, &udp, udp_size);
1503         return 0;
1504 }
1505
1506 /**
1507  * Convert TCP item to Verbs specification.
1508  *
1509  * @param item[in]
1510  *   Item specification.
1511  * @param default_mask[in]
1512  *   Default bit-masks to use when item->mask is not provided.
1513  * @param data[in, out]
1514  *   User structure.
1515  *
1516  * @return
1517  *   0 on success, a negative errno value otherwise and rte_errno is set.
1518  */
1519 static int
1520 mlx5_flow_create_tcp(const struct rte_flow_item *item,
1521                      const void *default_mask,
1522                      struct mlx5_flow_data *data)
1523 {
1524         const struct rte_flow_item_tcp *spec = item->spec;
1525         const struct rte_flow_item_tcp *mask = item->mask;
1526         struct mlx5_flow_parse *parser = data->parser;
1527         unsigned int tcp_size = sizeof(struct ibv_flow_spec_tcp_udp);
1528         struct ibv_flow_spec_tcp_udp tcp = {
1529                 .type = parser->inner | IBV_FLOW_SPEC_TCP,
1530                 .size = tcp_size,
1531         };
1532
1533         /* Don't update layer for the inner pattern. */
1534         if (!parser->inner) {
1535                 if (parser->layer == HASH_RXQ_IPV4)
1536                         parser->layer = HASH_RXQ_TCPV4;
1537                 else
1538                         parser->layer = HASH_RXQ_TCPV6;
1539         }
1540         if (spec) {
1541                 if (!mask)
1542                         mask = default_mask;
1543                 tcp.val.dst_port = spec->hdr.dst_port;
1544                 tcp.val.src_port = spec->hdr.src_port;
1545                 tcp.mask.dst_port = mask->hdr.dst_port;
1546                 tcp.mask.src_port = mask->hdr.src_port;
1547                 /* Remove unwanted bits from values. */
1548                 tcp.val.src_port &= tcp.mask.src_port;
1549                 tcp.val.dst_port &= tcp.mask.dst_port;
1550         }
1551         mlx5_flow_create_copy(parser, &tcp, tcp_size);
1552         return 0;
1553 }
1554
1555 /**
1556  * Convert VXLAN item to Verbs specification.
1557  *
1558  * @param item[in]
1559  *   Item specification.
1560  * @param default_mask[in]
1561  *   Default bit-masks to use when item->mask is not provided.
1562  * @param data[in, out]
1563  *   User structure.
1564  *
1565  * @return
1566  *   0 on success, a negative errno value otherwise and rte_errno is set.
1567  */
1568 static int
1569 mlx5_flow_create_vxlan(const struct rte_flow_item *item,
1570                        const void *default_mask,
1571                        struct mlx5_flow_data *data)
1572 {
1573         const struct rte_flow_item_vxlan *spec = item->spec;
1574         const struct rte_flow_item_vxlan *mask = item->mask;
1575         struct mlx5_flow_parse *parser = data->parser;
1576         unsigned int size = sizeof(struct ibv_flow_spec_tunnel);
1577         struct ibv_flow_spec_tunnel vxlan = {
1578                 .type = parser->inner | IBV_FLOW_SPEC_VXLAN_TUNNEL,
1579                 .size = size,
1580         };
1581         union vni {
1582                 uint32_t vlan_id;
1583                 uint8_t vni[4];
1584         } id;
1585
1586         id.vni[0] = 0;
1587         parser->inner = IBV_FLOW_SPEC_INNER;
1588         if (spec) {
1589                 if (!mask)
1590                         mask = default_mask;
1591                 memcpy(&id.vni[1], spec->vni, 3);
1592                 vxlan.val.tunnel_id = id.vlan_id;
1593                 memcpy(&id.vni[1], mask->vni, 3);
1594                 vxlan.mask.tunnel_id = id.vlan_id;
1595                 /* Remove unwanted bits from values. */
1596                 vxlan.val.tunnel_id &= vxlan.mask.tunnel_id;
1597         }
1598         /*
1599          * Tunnel id 0 is equivalent as not adding a VXLAN layer, if only this
1600          * layer is defined in the Verbs specification it is interpreted as
1601          * wildcard and all packets will match this rule, if it follows a full
1602          * stack layer (ex: eth / ipv4 / udp), all packets matching the layers
1603          * before will also match this rule.
1604          * To avoid such situation, VNI 0 is currently refused.
1605          */
1606         if (!vxlan.val.tunnel_id)
1607                 return rte_flow_error_set(data->error, EINVAL,
1608                                           RTE_FLOW_ERROR_TYPE_ITEM,
1609                                           item,
1610                                           "VxLAN vni cannot be 0");
1611         mlx5_flow_create_copy(parser, &vxlan, size);
1612         return 0;
1613 }
1614
1615 /**
1616  * Convert mark/flag action to Verbs specification.
1617  *
1618  * @param parser
1619  *   Internal parser structure.
1620  * @param mark_id
1621  *   Mark identifier.
1622  *
1623  * @return
1624  *   0 on success, a negative errno value otherwise and rte_errno is set.
1625  */
1626 static int
1627 mlx5_flow_create_flag_mark(struct mlx5_flow_parse *parser, uint32_t mark_id)
1628 {
1629         unsigned int size = sizeof(struct ibv_flow_spec_action_tag);
1630         struct ibv_flow_spec_action_tag tag = {
1631                 .type = IBV_FLOW_SPEC_ACTION_TAG,
1632                 .size = size,
1633                 .tag_id = mlx5_flow_mark_set(mark_id),
1634         };
1635
1636         assert(parser->mark);
1637         mlx5_flow_create_copy(parser, &tag, size);
1638         return 0;
1639 }
1640
1641 /**
1642  * Convert count action to Verbs specification.
1643  *
1644  * @param dev
1645  *   Pointer to Ethernet device.
1646  * @param parser
1647  *   Pointer to MLX5 flow parser structure.
1648  *
1649  * @return
1650  *   0 on success, a negative errno value otherwise and rte_errno is set.
1651  */
1652 static int
1653 mlx5_flow_create_count(struct rte_eth_dev *dev __rte_unused,
1654                        struct mlx5_flow_parse *parser __rte_unused)
1655 {
1656 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
1657         struct priv *priv = dev->data->dev_private;
1658         unsigned int size = sizeof(struct ibv_flow_spec_counter_action);
1659         struct ibv_counter_set_init_attr init_attr = {0};
1660         struct ibv_flow_spec_counter_action counter = {
1661                 .type = IBV_FLOW_SPEC_ACTION_COUNT,
1662                 .size = size,
1663                 .counter_set_handle = 0,
1664         };
1665
1666         init_attr.counter_set_id = 0;
1667         parser->cs = mlx5_glue->create_counter_set(priv->ctx, &init_attr);
1668         if (!parser->cs) {
1669                 rte_errno = EINVAL;
1670                 return -rte_errno;
1671         }
1672         counter.counter_set_handle = parser->cs->handle;
1673         mlx5_flow_create_copy(parser, &counter, size);
1674 #endif
1675         return 0;
1676 }
1677
1678 /**
1679  * Complete flow rule creation with a drop queue.
1680  *
1681  * @param dev
1682  *   Pointer to Ethernet device.
1683  * @param parser
1684  *   Internal parser structure.
1685  * @param flow
1686  *   Pointer to the rte_flow.
1687  * @param[out] error
1688  *   Perform verbose error reporting if not NULL.
1689  *
1690  * @return
1691  *   0 on success, a negative errno value otherwise and rte_errno is set.
1692  */
1693 static int
1694 mlx5_flow_create_action_queue_drop(struct rte_eth_dev *dev,
1695                                    struct mlx5_flow_parse *parser,
1696                                    struct rte_flow *flow,
1697                                    struct rte_flow_error *error)
1698 {
1699         struct priv *priv = dev->data->dev_private;
1700         struct ibv_flow_spec_action_drop *drop;
1701         unsigned int size = sizeof(struct ibv_flow_spec_action_drop);
1702
1703         assert(priv->pd);
1704         assert(priv->ctx);
1705         flow->drop = 1;
1706         drop = (void *)((uintptr_t)parser->queue[HASH_RXQ_ETH].ibv_attr +
1707                         parser->queue[HASH_RXQ_ETH].offset);
1708         *drop = (struct ibv_flow_spec_action_drop){
1709                         .type = IBV_FLOW_SPEC_ACTION_DROP,
1710                         .size = size,
1711         };
1712         ++parser->queue[HASH_RXQ_ETH].ibv_attr->num_of_specs;
1713         parser->queue[HASH_RXQ_ETH].offset += size;
1714         flow->frxq[HASH_RXQ_ETH].ibv_attr =
1715                 parser->queue[HASH_RXQ_ETH].ibv_attr;
1716         if (parser->count)
1717                 flow->cs = parser->cs;
1718         if (!priv->dev->data->dev_started)
1719                 return 0;
1720         parser->queue[HASH_RXQ_ETH].ibv_attr = NULL;
1721         flow->frxq[HASH_RXQ_ETH].ibv_flow =
1722                 mlx5_glue->create_flow(priv->flow_drop_queue->qp,
1723                                        flow->frxq[HASH_RXQ_ETH].ibv_attr);
1724         if (!flow->frxq[HASH_RXQ_ETH].ibv_flow) {
1725                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1726                                    NULL, "flow rule creation failure");
1727                 goto error;
1728         }
1729         return 0;
1730 error:
1731         assert(flow);
1732         if (flow->frxq[HASH_RXQ_ETH].ibv_flow) {
1733                 claim_zero(mlx5_glue->destroy_flow
1734                            (flow->frxq[HASH_RXQ_ETH].ibv_flow));
1735                 flow->frxq[HASH_RXQ_ETH].ibv_flow = NULL;
1736         }
1737         if (flow->frxq[HASH_RXQ_ETH].ibv_attr) {
1738                 rte_free(flow->frxq[HASH_RXQ_ETH].ibv_attr);
1739                 flow->frxq[HASH_RXQ_ETH].ibv_attr = NULL;
1740         }
1741         if (flow->cs) {
1742                 claim_zero(mlx5_glue->destroy_counter_set(flow->cs));
1743                 flow->cs = NULL;
1744                 parser->cs = NULL;
1745         }
1746         return -rte_errno;
1747 }
1748
1749 /**
1750  * Create hash Rx queues when RSS is enabled.
1751  *
1752  * @param dev
1753  *   Pointer to Ethernet device.
1754  * @param parser
1755  *   Internal parser structure.
1756  * @param flow
1757  *   Pointer to the rte_flow.
1758  * @param[out] error
1759  *   Perform verbose error reporting if not NULL.
1760  *
1761  * @return
1762  *   0 on success, a negative errno value otherwise and rte_errno is set.
1763  */
1764 static int
1765 mlx5_flow_create_action_queue_rss(struct rte_eth_dev *dev,
1766                                   struct mlx5_flow_parse *parser,
1767                                   struct rte_flow *flow,
1768                                   struct rte_flow_error *error)
1769 {
1770         struct priv *priv = dev->data->dev_private;
1771         unsigned int i;
1772
1773         for (i = 0; i != hash_rxq_init_n; ++i) {
1774                 uint64_t hash_fields;
1775
1776                 if (!parser->queue[i].ibv_attr)
1777                         continue;
1778                 flow->frxq[i].ibv_attr = parser->queue[i].ibv_attr;
1779                 parser->queue[i].ibv_attr = NULL;
1780                 hash_fields = hash_rxq_init[i].hash_fields;
1781                 if (!priv->dev->data->dev_started)
1782                         continue;
1783                 flow->frxq[i].hrxq =
1784                         mlx5_hrxq_get(dev,
1785                                       parser->rss_conf.key,
1786                                       parser->rss_conf.key_len,
1787                                       hash_fields,
1788                                       parser->rss_conf.queue,
1789                                       parser->rss_conf.queue_num);
1790                 if (flow->frxq[i].hrxq)
1791                         continue;
1792                 flow->frxq[i].hrxq =
1793                         mlx5_hrxq_new(dev,
1794                                       parser->rss_conf.key,
1795                                       parser->rss_conf.key_len,
1796                                       hash_fields,
1797                                       parser->rss_conf.queue,
1798                                       parser->rss_conf.queue_num);
1799                 if (!flow->frxq[i].hrxq) {
1800                         return rte_flow_error_set(error, ENOMEM,
1801                                                   RTE_FLOW_ERROR_TYPE_HANDLE,
1802                                                   NULL,
1803                                                   "cannot create hash rxq");
1804                 }
1805         }
1806         return 0;
1807 }
1808
1809 /**
1810  * Complete flow rule creation.
1811  *
1812  * @param dev
1813  *   Pointer to Ethernet device.
1814  * @param parser
1815  *   Internal parser structure.
1816  * @param flow
1817  *   Pointer to the rte_flow.
1818  * @param[out] error
1819  *   Perform verbose error reporting if not NULL.
1820  *
1821  * @return
1822  *   0 on success, a negative errno value otherwise and rte_errno is set.
1823  */
1824 static int
1825 mlx5_flow_create_action_queue(struct rte_eth_dev *dev,
1826                               struct mlx5_flow_parse *parser,
1827                               struct rte_flow *flow,
1828                               struct rte_flow_error *error)
1829 {
1830         struct priv *priv = dev->data->dev_private;
1831         int ret;
1832         unsigned int i;
1833         unsigned int flows_n = 0;
1834
1835         assert(priv->pd);
1836         assert(priv->ctx);
1837         assert(!parser->drop);
1838         ret = mlx5_flow_create_action_queue_rss(dev, parser, flow, error);
1839         if (ret)
1840                 goto error;
1841         if (parser->count)
1842                 flow->cs = parser->cs;
1843         if (!priv->dev->data->dev_started)
1844                 return 0;
1845         for (i = 0; i != hash_rxq_init_n; ++i) {
1846                 if (!flow->frxq[i].hrxq)
1847                         continue;
1848                 flow->frxq[i].ibv_flow =
1849                         mlx5_glue->create_flow(flow->frxq[i].hrxq->qp,
1850                                                flow->frxq[i].ibv_attr);
1851                 if (!flow->frxq[i].ibv_flow) {
1852                         rte_flow_error_set(error, ENOMEM,
1853                                            RTE_FLOW_ERROR_TYPE_HANDLE,
1854                                            NULL, "flow rule creation failure");
1855                         goto error;
1856                 }
1857                 ++flows_n;
1858                 DRV_LOG(DEBUG, "port %u %p type %d QP %p ibv_flow %p",
1859                         dev->data->port_id,
1860                         (void *)flow, i,
1861                         (void *)flow->frxq[i].hrxq,
1862                         (void *)flow->frxq[i].ibv_flow);
1863         }
1864         if (!flows_n) {
1865                 rte_flow_error_set(error, EINVAL, RTE_FLOW_ERROR_TYPE_HANDLE,
1866                                    NULL, "internal error in flow creation");
1867                 goto error;
1868         }
1869         for (i = 0; i != parser->rss_conf.queue_num; ++i) {
1870                 struct mlx5_rxq_data *q =
1871                         (*priv->rxqs)[parser->rss_conf.queue[i]];
1872
1873                 q->mark |= parser->mark;
1874         }
1875         return 0;
1876 error:
1877         ret = rte_errno; /* Save rte_errno before cleanup. */
1878         assert(flow);
1879         for (i = 0; i != hash_rxq_init_n; ++i) {
1880                 if (flow->frxq[i].ibv_flow) {
1881                         struct ibv_flow *ibv_flow = flow->frxq[i].ibv_flow;
1882
1883                         claim_zero(mlx5_glue->destroy_flow(ibv_flow));
1884                 }
1885                 if (flow->frxq[i].hrxq)
1886                         mlx5_hrxq_release(dev, flow->frxq[i].hrxq);
1887                 if (flow->frxq[i].ibv_attr)
1888                         rte_free(flow->frxq[i].ibv_attr);
1889         }
1890         if (flow->cs) {
1891                 claim_zero(mlx5_glue->destroy_counter_set(flow->cs));
1892                 flow->cs = NULL;
1893                 parser->cs = NULL;
1894         }
1895         rte_errno = ret; /* Restore rte_errno. */
1896         return -rte_errno;
1897 }
1898
1899 /**
1900  * Convert a flow.
1901  *
1902  * @param dev
1903  *   Pointer to Ethernet device.
1904  * @param list
1905  *   Pointer to a TAILQ flow list.
1906  * @param[in] attr
1907  *   Flow rule attributes.
1908  * @param[in] pattern
1909  *   Pattern specification (list terminated by the END pattern item).
1910  * @param[in] actions
1911  *   Associated actions (list terminated by the END action).
1912  * @param[out] error
1913  *   Perform verbose error reporting if not NULL.
1914  *
1915  * @return
1916  *   A flow on success, NULL otherwise and rte_errno is set.
1917  */
1918 static struct rte_flow *
1919 mlx5_flow_list_create(struct rte_eth_dev *dev,
1920                       struct mlx5_flows *list,
1921                       const struct rte_flow_attr *attr,
1922                       const struct rte_flow_item items[],
1923                       const struct rte_flow_action actions[],
1924                       struct rte_flow_error *error)
1925 {
1926         struct mlx5_flow_parse parser = { .create = 1, };
1927         struct rte_flow *flow = NULL;
1928         unsigned int i;
1929         int ret;
1930
1931         ret = mlx5_flow_convert(dev, attr, items, actions, error, &parser);
1932         if (ret)
1933                 goto exit;
1934         flow = rte_calloc(__func__, 1,
1935                           sizeof(*flow) +
1936                           parser.rss_conf.queue_num * sizeof(uint16_t),
1937                           0);
1938         if (!flow) {
1939                 rte_flow_error_set(error, ENOMEM,
1940                                    RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
1941                                    NULL,
1942                                    "cannot allocate flow memory");
1943                 return NULL;
1944         }
1945         /* Copy configuration. */
1946         flow->queues = (uint16_t (*)[])(flow + 1);
1947         flow->rss_conf = (struct rte_flow_action_rss){
1948                 .func = RTE_ETH_HASH_FUNCTION_DEFAULT,
1949                 .level = 0,
1950                 .types = parser.rss_conf.types,
1951                 .key_len = parser.rss_conf.key_len,
1952                 .queue_num = parser.rss_conf.queue_num,
1953                 .key = memcpy(flow->rss_key, parser.rss_conf.key,
1954                               sizeof(*parser.rss_conf.key) *
1955                               parser.rss_conf.key_len),
1956                 .queue = memcpy(flow->queues, parser.rss_conf.queue,
1957                                 sizeof(*parser.rss_conf.queue) *
1958                                 parser.rss_conf.queue_num),
1959         };
1960         flow->mark = parser.mark;
1961         /* finalise the flow. */
1962         if (parser.drop)
1963                 ret = mlx5_flow_create_action_queue_drop(dev, &parser, flow,
1964                                                          error);
1965         else
1966                 ret = mlx5_flow_create_action_queue(dev, &parser, flow, error);
1967         if (ret)
1968                 goto exit;
1969         TAILQ_INSERT_TAIL(list, flow, next);
1970         DRV_LOG(DEBUG, "port %u flow created %p", dev->data->port_id,
1971                 (void *)flow);
1972         return flow;
1973 exit:
1974         DRV_LOG(ERR, "port %u flow creation error: %s", dev->data->port_id,
1975                 error->message);
1976         for (i = 0; i != hash_rxq_init_n; ++i) {
1977                 if (parser.queue[i].ibv_attr)
1978                         rte_free(parser.queue[i].ibv_attr);
1979         }
1980         rte_free(flow);
1981         return NULL;
1982 }
1983
1984 /**
1985  * Validate a flow supported by the NIC.
1986  *
1987  * @see rte_flow_validate()
1988  * @see rte_flow_ops
1989  */
1990 int
1991 mlx5_flow_validate(struct rte_eth_dev *dev,
1992                    const struct rte_flow_attr *attr,
1993                    const struct rte_flow_item items[],
1994                    const struct rte_flow_action actions[],
1995                    struct rte_flow_error *error)
1996 {
1997         struct mlx5_flow_parse parser = { .create = 0, };
1998
1999         return mlx5_flow_convert(dev, attr, items, actions, error, &parser);
2000 }
2001
2002 /**
2003  * Create a flow.
2004  *
2005  * @see rte_flow_create()
2006  * @see rte_flow_ops
2007  */
2008 struct rte_flow *
2009 mlx5_flow_create(struct rte_eth_dev *dev,
2010                  const struct rte_flow_attr *attr,
2011                  const struct rte_flow_item items[],
2012                  const struct rte_flow_action actions[],
2013                  struct rte_flow_error *error)
2014 {
2015         struct priv *priv = dev->data->dev_private;
2016
2017         return mlx5_flow_list_create(dev, &priv->flows, attr, items, actions,
2018                                      error);
2019 }
2020
2021 /**
2022  * Destroy a flow in a list.
2023  *
2024  * @param dev
2025  *   Pointer to Ethernet device.
2026  * @param list
2027  *   Pointer to a TAILQ flow list.
2028  * @param[in] flow
2029  *   Flow to destroy.
2030  */
2031 static void
2032 mlx5_flow_list_destroy(struct rte_eth_dev *dev, struct mlx5_flows *list,
2033                        struct rte_flow *flow)
2034 {
2035         struct priv *priv = dev->data->dev_private;
2036         unsigned int i;
2037
2038         if (flow->drop || !flow->mark)
2039                 goto free;
2040         for (i = 0; i != flow->rss_conf.queue_num; ++i) {
2041                 struct rte_flow *tmp;
2042                 int mark = 0;
2043
2044                 /*
2045                  * To remove the mark from the queue, the queue must not be
2046                  * present in any other marked flow (RSS or not).
2047                  */
2048                 TAILQ_FOREACH(tmp, list, next) {
2049                         unsigned int j;
2050                         uint16_t *tqs = NULL;
2051                         uint16_t tq_n = 0;
2052
2053                         if (!tmp->mark)
2054                                 continue;
2055                         for (j = 0; j != hash_rxq_init_n; ++j) {
2056                                 if (!tmp->frxq[j].hrxq)
2057                                         continue;
2058                                 tqs = tmp->frxq[j].hrxq->ind_table->queues;
2059                                 tq_n = tmp->frxq[j].hrxq->ind_table->queues_n;
2060                         }
2061                         if (!tq_n)
2062                                 continue;
2063                         for (j = 0; (j != tq_n) && !mark; j++)
2064                                 if (tqs[j] == (*flow->queues)[i])
2065                                         mark = 1;
2066                 }
2067                 (*priv->rxqs)[(*flow->queues)[i]]->mark = mark;
2068         }
2069 free:
2070         if (flow->drop) {
2071                 if (flow->frxq[HASH_RXQ_ETH].ibv_flow)
2072                         claim_zero(mlx5_glue->destroy_flow
2073                                    (flow->frxq[HASH_RXQ_ETH].ibv_flow));
2074                 rte_free(flow->frxq[HASH_RXQ_ETH].ibv_attr);
2075         } else {
2076                 for (i = 0; i != hash_rxq_init_n; ++i) {
2077                         struct mlx5_flow *frxq = &flow->frxq[i];
2078
2079                         if (frxq->ibv_flow)
2080                                 claim_zero(mlx5_glue->destroy_flow
2081                                            (frxq->ibv_flow));
2082                         if (frxq->hrxq)
2083                                 mlx5_hrxq_release(dev, frxq->hrxq);
2084                         if (frxq->ibv_attr)
2085                                 rte_free(frxq->ibv_attr);
2086                 }
2087         }
2088         if (flow->cs) {
2089                 claim_zero(mlx5_glue->destroy_counter_set(flow->cs));
2090                 flow->cs = NULL;
2091         }
2092         TAILQ_REMOVE(list, flow, next);
2093         DRV_LOG(DEBUG, "port %u flow destroyed %p", dev->data->port_id,
2094                 (void *)flow);
2095         rte_free(flow);
2096 }
2097
2098 /**
2099  * Destroy all flows.
2100  *
2101  * @param dev
2102  *   Pointer to Ethernet device.
2103  * @param list
2104  *   Pointer to a TAILQ flow list.
2105  */
2106 void
2107 mlx5_flow_list_flush(struct rte_eth_dev *dev, struct mlx5_flows *list)
2108 {
2109         while (!TAILQ_EMPTY(list)) {
2110                 struct rte_flow *flow;
2111
2112                 flow = TAILQ_FIRST(list);
2113                 mlx5_flow_list_destroy(dev, list, flow);
2114         }
2115 }
2116
2117 /**
2118  * Create drop queue.
2119  *
2120  * @param dev
2121  *   Pointer to Ethernet device.
2122  *
2123  * @return
2124  *   0 on success, a negative errno value otherwise and rte_errno is set.
2125  */
2126 int
2127 mlx5_flow_create_drop_queue(struct rte_eth_dev *dev)
2128 {
2129         struct priv *priv = dev->data->dev_private;
2130         struct mlx5_hrxq_drop *fdq = NULL;
2131
2132         assert(priv->pd);
2133         assert(priv->ctx);
2134         fdq = rte_calloc(__func__, 1, sizeof(*fdq), 0);
2135         if (!fdq) {
2136                 DRV_LOG(WARNING,
2137                         "port %u cannot allocate memory for drop queue",
2138                         dev->data->port_id);
2139                 rte_errno = ENOMEM;
2140                 return -rte_errno;
2141         }
2142         fdq->cq = mlx5_glue->create_cq(priv->ctx, 1, NULL, NULL, 0);
2143         if (!fdq->cq) {
2144                 DRV_LOG(WARNING, "port %u cannot allocate CQ for drop queue",
2145                         dev->data->port_id);
2146                 rte_errno = errno;
2147                 goto error;
2148         }
2149         fdq->wq = mlx5_glue->create_wq
2150                 (priv->ctx,
2151                  &(struct ibv_wq_init_attr){
2152                         .wq_type = IBV_WQT_RQ,
2153                         .max_wr = 1,
2154                         .max_sge = 1,
2155                         .pd = priv->pd,
2156                         .cq = fdq->cq,
2157                  });
2158         if (!fdq->wq) {
2159                 DRV_LOG(WARNING, "port %u cannot allocate WQ for drop queue",
2160                         dev->data->port_id);
2161                 rte_errno = errno;
2162                 goto error;
2163         }
2164         fdq->ind_table = mlx5_glue->create_rwq_ind_table
2165                 (priv->ctx,
2166                  &(struct ibv_rwq_ind_table_init_attr){
2167                         .log_ind_tbl_size = 0,
2168                         .ind_tbl = &fdq->wq,
2169                         .comp_mask = 0,
2170                  });
2171         if (!fdq->ind_table) {
2172                 DRV_LOG(WARNING,
2173                         "port %u cannot allocate indirection table for drop"
2174                         " queue",
2175                         dev->data->port_id);
2176                 rte_errno = errno;
2177                 goto error;
2178         }
2179         fdq->qp = mlx5_glue->create_qp_ex
2180                 (priv->ctx,
2181                  &(struct ibv_qp_init_attr_ex){
2182                         .qp_type = IBV_QPT_RAW_PACKET,
2183                         .comp_mask =
2184                                 IBV_QP_INIT_ATTR_PD |
2185                                 IBV_QP_INIT_ATTR_IND_TABLE |
2186                                 IBV_QP_INIT_ATTR_RX_HASH,
2187                         .rx_hash_conf = (struct ibv_rx_hash_conf){
2188                                 .rx_hash_function =
2189                                         IBV_RX_HASH_FUNC_TOEPLITZ,
2190                                 .rx_hash_key_len = rss_hash_default_key_len,
2191                                 .rx_hash_key = rss_hash_default_key,
2192                                 .rx_hash_fields_mask = 0,
2193                                 },
2194                         .rwq_ind_tbl = fdq->ind_table,
2195                         .pd = priv->pd
2196                  });
2197         if (!fdq->qp) {
2198                 DRV_LOG(WARNING, "port %u cannot allocate QP for drop queue",
2199                         dev->data->port_id);
2200                 rte_errno = errno;
2201                 goto error;
2202         }
2203         priv->flow_drop_queue = fdq;
2204         return 0;
2205 error:
2206         if (fdq->qp)
2207                 claim_zero(mlx5_glue->destroy_qp(fdq->qp));
2208         if (fdq->ind_table)
2209                 claim_zero(mlx5_glue->destroy_rwq_ind_table(fdq->ind_table));
2210         if (fdq->wq)
2211                 claim_zero(mlx5_glue->destroy_wq(fdq->wq));
2212         if (fdq->cq)
2213                 claim_zero(mlx5_glue->destroy_cq(fdq->cq));
2214         if (fdq)
2215                 rte_free(fdq);
2216         priv->flow_drop_queue = NULL;
2217         return -rte_errno;
2218 }
2219
2220 /**
2221  * Delete drop queue.
2222  *
2223  * @param dev
2224  *   Pointer to Ethernet device.
2225  */
2226 void
2227 mlx5_flow_delete_drop_queue(struct rte_eth_dev *dev)
2228 {
2229         struct priv *priv = dev->data->dev_private;
2230         struct mlx5_hrxq_drop *fdq = priv->flow_drop_queue;
2231
2232         if (!fdq)
2233                 return;
2234         if (fdq->qp)
2235                 claim_zero(mlx5_glue->destroy_qp(fdq->qp));
2236         if (fdq->ind_table)
2237                 claim_zero(mlx5_glue->destroy_rwq_ind_table(fdq->ind_table));
2238         if (fdq->wq)
2239                 claim_zero(mlx5_glue->destroy_wq(fdq->wq));
2240         if (fdq->cq)
2241                 claim_zero(mlx5_glue->destroy_cq(fdq->cq));
2242         rte_free(fdq);
2243         priv->flow_drop_queue = NULL;
2244 }
2245
2246 /**
2247  * Remove all flows.
2248  *
2249  * @param dev
2250  *   Pointer to Ethernet device.
2251  * @param list
2252  *   Pointer to a TAILQ flow list.
2253  */
2254 void
2255 mlx5_flow_stop(struct rte_eth_dev *dev, struct mlx5_flows *list)
2256 {
2257         struct priv *priv = dev->data->dev_private;
2258         struct rte_flow *flow;
2259
2260         TAILQ_FOREACH_REVERSE(flow, list, mlx5_flows, next) {
2261                 unsigned int i;
2262                 struct mlx5_ind_table_ibv *ind_tbl = NULL;
2263
2264                 if (flow->drop) {
2265                         if (!flow->frxq[HASH_RXQ_ETH].ibv_flow)
2266                                 continue;
2267                         claim_zero(mlx5_glue->destroy_flow
2268                                    (flow->frxq[HASH_RXQ_ETH].ibv_flow));
2269                         flow->frxq[HASH_RXQ_ETH].ibv_flow = NULL;
2270                         DRV_LOG(DEBUG, "port %u flow %p removed",
2271                                 dev->data->port_id, (void *)flow);
2272                         /* Next flow. */
2273                         continue;
2274                 }
2275                 /* Verify the flow has not already been cleaned. */
2276                 for (i = 0; i != hash_rxq_init_n; ++i) {
2277                         if (!flow->frxq[i].ibv_flow)
2278                                 continue;
2279                         /*
2280                          * Indirection table may be necessary to remove the
2281                          * flags in the Rx queues.
2282                          * This helps to speed-up the process by avoiding
2283                          * another loop.
2284                          */
2285                         ind_tbl = flow->frxq[i].hrxq->ind_table;
2286                         break;
2287                 }
2288                 if (i == hash_rxq_init_n)
2289                         return;
2290                 if (flow->mark) {
2291                         assert(ind_tbl);
2292                         for (i = 0; i != ind_tbl->queues_n; ++i)
2293                                 (*priv->rxqs)[ind_tbl->queues[i]]->mark = 0;
2294                 }
2295                 for (i = 0; i != hash_rxq_init_n; ++i) {
2296                         if (!flow->frxq[i].ibv_flow)
2297                                 continue;
2298                         claim_zero(mlx5_glue->destroy_flow
2299                                    (flow->frxq[i].ibv_flow));
2300                         flow->frxq[i].ibv_flow = NULL;
2301                         mlx5_hrxq_release(dev, flow->frxq[i].hrxq);
2302                         flow->frxq[i].hrxq = NULL;
2303                 }
2304                 DRV_LOG(DEBUG, "port %u flow %p removed", dev->data->port_id,
2305                         (void *)flow);
2306         }
2307 }
2308
2309 /**
2310  * Add all flows.
2311  *
2312  * @param dev
2313  *   Pointer to Ethernet device.
2314  * @param list
2315  *   Pointer to a TAILQ flow list.
2316  *
2317  * @return
2318  *   0 on success, a negative errno value otherwise and rte_errno is set.
2319  */
2320 int
2321 mlx5_flow_start(struct rte_eth_dev *dev, struct mlx5_flows *list)
2322 {
2323         struct priv *priv = dev->data->dev_private;
2324         struct rte_flow *flow;
2325
2326         TAILQ_FOREACH(flow, list, next) {
2327                 unsigned int i;
2328
2329                 if (flow->drop) {
2330                         flow->frxq[HASH_RXQ_ETH].ibv_flow =
2331                                 mlx5_glue->create_flow
2332                                 (priv->flow_drop_queue->qp,
2333                                  flow->frxq[HASH_RXQ_ETH].ibv_attr);
2334                         if (!flow->frxq[HASH_RXQ_ETH].ibv_flow) {
2335                                 DRV_LOG(DEBUG,
2336                                         "port %u flow %p cannot be applied",
2337                                         dev->data->port_id, (void *)flow);
2338                                 rte_errno = EINVAL;
2339                                 return -rte_errno;
2340                         }
2341                         DRV_LOG(DEBUG, "port %u flow %p applied",
2342                                 dev->data->port_id, (void *)flow);
2343                         /* Next flow. */
2344                         continue;
2345                 }
2346                 for (i = 0; i != hash_rxq_init_n; ++i) {
2347                         if (!flow->frxq[i].ibv_attr)
2348                                 continue;
2349                         flow->frxq[i].hrxq =
2350                                 mlx5_hrxq_get(dev, flow->rss_conf.key,
2351                                               flow->rss_conf.key_len,
2352                                               hash_rxq_init[i].hash_fields,
2353                                               flow->rss_conf.queue,
2354                                               flow->rss_conf.queue_num);
2355                         if (flow->frxq[i].hrxq)
2356                                 goto flow_create;
2357                         flow->frxq[i].hrxq =
2358                                 mlx5_hrxq_new(dev, flow->rss_conf.key,
2359                                               flow->rss_conf.key_len,
2360                                               hash_rxq_init[i].hash_fields,
2361                                               flow->rss_conf.queue,
2362                                               flow->rss_conf.queue_num);
2363                         if (!flow->frxq[i].hrxq) {
2364                                 DRV_LOG(DEBUG,
2365                                         "port %u flow %p cannot be applied",
2366                                         dev->data->port_id, (void *)flow);
2367                                 rte_errno = EINVAL;
2368                                 return -rte_errno;
2369                         }
2370 flow_create:
2371                         flow->frxq[i].ibv_flow =
2372                                 mlx5_glue->create_flow(flow->frxq[i].hrxq->qp,
2373                                                        flow->frxq[i].ibv_attr);
2374                         if (!flow->frxq[i].ibv_flow) {
2375                                 DRV_LOG(DEBUG,
2376                                         "port %u flow %p cannot be applied",
2377                                         dev->data->port_id, (void *)flow);
2378                                 rte_errno = EINVAL;
2379                                 return -rte_errno;
2380                         }
2381                         DRV_LOG(DEBUG, "port %u flow %p applied",
2382                                 dev->data->port_id, (void *)flow);
2383                 }
2384                 if (!flow->mark)
2385                         continue;
2386                 for (i = 0; i != flow->rss_conf.queue_num; ++i)
2387                         (*priv->rxqs)[flow->rss_conf.queue[i]]->mark = 1;
2388         }
2389         return 0;
2390 }
2391
2392 /**
2393  * Verify the flow list is empty
2394  *
2395  * @param dev
2396  *  Pointer to Ethernet device.
2397  *
2398  * @return the number of flows not released.
2399  */
2400 int
2401 mlx5_flow_verify(struct rte_eth_dev *dev)
2402 {
2403         struct priv *priv = dev->data->dev_private;
2404         struct rte_flow *flow;
2405         int ret = 0;
2406
2407         TAILQ_FOREACH(flow, &priv->flows, next) {
2408                 DRV_LOG(DEBUG, "port %u flow %p still referenced",
2409                         dev->data->port_id, (void *)flow);
2410                 ++ret;
2411         }
2412         return ret;
2413 }
2414
2415 /**
2416  * Enable a control flow configured from the control plane.
2417  *
2418  * @param dev
2419  *   Pointer to Ethernet device.
2420  * @param eth_spec
2421  *   An Ethernet flow spec to apply.
2422  * @param eth_mask
2423  *   An Ethernet flow mask to apply.
2424  * @param vlan_spec
2425  *   A VLAN flow spec to apply.
2426  * @param vlan_mask
2427  *   A VLAN flow mask to apply.
2428  *
2429  * @return
2430  *   0 on success, a negative errno value otherwise and rte_errno is set.
2431  */
2432 int
2433 mlx5_ctrl_flow_vlan(struct rte_eth_dev *dev,
2434                     struct rte_flow_item_eth *eth_spec,
2435                     struct rte_flow_item_eth *eth_mask,
2436                     struct rte_flow_item_vlan *vlan_spec,
2437                     struct rte_flow_item_vlan *vlan_mask)
2438 {
2439         struct priv *priv = dev->data->dev_private;
2440         const struct rte_flow_attr attr = {
2441                 .ingress = 1,
2442                 .priority = MLX5_CTRL_FLOW_PRIORITY,
2443         };
2444         struct rte_flow_item items[] = {
2445                 {
2446                         .type = RTE_FLOW_ITEM_TYPE_ETH,
2447                         .spec = eth_spec,
2448                         .last = NULL,
2449                         .mask = eth_mask,
2450                 },
2451                 {
2452                         .type = (vlan_spec) ? RTE_FLOW_ITEM_TYPE_VLAN :
2453                                 RTE_FLOW_ITEM_TYPE_END,
2454                         .spec = vlan_spec,
2455                         .last = NULL,
2456                         .mask = vlan_mask,
2457                 },
2458                 {
2459                         .type = RTE_FLOW_ITEM_TYPE_END,
2460                 },
2461         };
2462         uint16_t queue[priv->reta_idx_n];
2463         struct rte_flow_action_rss action_rss = {
2464                 .func = RTE_ETH_HASH_FUNCTION_DEFAULT,
2465                 .level = 0,
2466                 .types = priv->rss_conf.rss_hf,
2467                 .key_len = priv->rss_conf.rss_key_len,
2468                 .queue_num = priv->reta_idx_n,
2469                 .key = priv->rss_conf.rss_key,
2470                 .queue = queue,
2471         };
2472         struct rte_flow_action actions[] = {
2473                 {
2474                         .type = RTE_FLOW_ACTION_TYPE_RSS,
2475                         .conf = &action_rss,
2476                 },
2477                 {
2478                         .type = RTE_FLOW_ACTION_TYPE_END,
2479                 },
2480         };
2481         struct rte_flow *flow;
2482         struct rte_flow_error error;
2483         unsigned int i;
2484
2485         if (!priv->reta_idx_n) {
2486                 rte_errno = EINVAL;
2487                 return -rte_errno;
2488         }
2489         for (i = 0; i != priv->reta_idx_n; ++i)
2490                 queue[i] = (*priv->reta_idx)[i];
2491         flow = mlx5_flow_list_create(dev, &priv->ctrl_flows, &attr, items,
2492                                      actions, &error);
2493         if (!flow)
2494                 return -rte_errno;
2495         return 0;
2496 }
2497
2498 /**
2499  * Enable a flow control configured from the control plane.
2500  *
2501  * @param dev
2502  *   Pointer to Ethernet device.
2503  * @param eth_spec
2504  *   An Ethernet flow spec to apply.
2505  * @param eth_mask
2506  *   An Ethernet flow mask to apply.
2507  *
2508  * @return
2509  *   0 on success, a negative errno value otherwise and rte_errno is set.
2510  */
2511 int
2512 mlx5_ctrl_flow(struct rte_eth_dev *dev,
2513                struct rte_flow_item_eth *eth_spec,
2514                struct rte_flow_item_eth *eth_mask)
2515 {
2516         return mlx5_ctrl_flow_vlan(dev, eth_spec, eth_mask, NULL, NULL);
2517 }
2518
2519 /**
2520  * Destroy a flow.
2521  *
2522  * @see rte_flow_destroy()
2523  * @see rte_flow_ops
2524  */
2525 int
2526 mlx5_flow_destroy(struct rte_eth_dev *dev,
2527                   struct rte_flow *flow,
2528                   struct rte_flow_error *error __rte_unused)
2529 {
2530         struct priv *priv = dev->data->dev_private;
2531
2532         mlx5_flow_list_destroy(dev, &priv->flows, flow);
2533         return 0;
2534 }
2535
2536 /**
2537  * Destroy all flows.
2538  *
2539  * @see rte_flow_flush()
2540  * @see rte_flow_ops
2541  */
2542 int
2543 mlx5_flow_flush(struct rte_eth_dev *dev,
2544                 struct rte_flow_error *error __rte_unused)
2545 {
2546         struct priv *priv = dev->data->dev_private;
2547
2548         mlx5_flow_list_flush(dev, &priv->flows);
2549         return 0;
2550 }
2551
2552 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
2553 /**
2554  * Query flow counter.
2555  *
2556  * @param cs
2557  *   the counter set.
2558  * @param counter_value
2559  *   returned data from the counter.
2560  *
2561  * @return
2562  *   0 on success, a negative errno value otherwise and rte_errno is set.
2563  */
2564 static int
2565 mlx5_flow_query_count(struct ibv_counter_set *cs,
2566                       struct mlx5_flow_counter_stats *counter_stats,
2567                       struct rte_flow_query_count *query_count,
2568                       struct rte_flow_error *error)
2569 {
2570         uint64_t counters[2];
2571         struct ibv_query_counter_set_attr query_cs_attr = {
2572                 .cs = cs,
2573                 .query_flags = IBV_COUNTER_SET_FORCE_UPDATE,
2574         };
2575         struct ibv_counter_set_data query_out = {
2576                 .out = counters,
2577                 .outlen = 2 * sizeof(uint64_t),
2578         };
2579         int err = mlx5_glue->query_counter_set(&query_cs_attr, &query_out);
2580
2581         if (err)
2582                 return rte_flow_error_set(error, err,
2583                                           RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
2584                                           NULL,
2585                                           "cannot read counter");
2586         query_count->hits_set = 1;
2587         query_count->bytes_set = 1;
2588         query_count->hits = counters[0] - counter_stats->hits;
2589         query_count->bytes = counters[1] - counter_stats->bytes;
2590         if (query_count->reset) {
2591                 counter_stats->hits = counters[0];
2592                 counter_stats->bytes = counters[1];
2593         }
2594         return 0;
2595 }
2596
2597 /**
2598  * Query a flows.
2599  *
2600  * @see rte_flow_query()
2601  * @see rte_flow_ops
2602  */
2603 int
2604 mlx5_flow_query(struct rte_eth_dev *dev __rte_unused,
2605                 struct rte_flow *flow,
2606                 enum rte_flow_action_type action __rte_unused,
2607                 void *data,
2608                 struct rte_flow_error *error)
2609 {
2610         if (flow->cs) {
2611                 int ret;
2612
2613                 ret = mlx5_flow_query_count(flow->cs,
2614                                             &flow->counter_stats,
2615                                             (struct rte_flow_query_count *)data,
2616                                             error);
2617                 if (ret)
2618                         return ret;
2619         } else {
2620                 return rte_flow_error_set(error, EINVAL,
2621                                           RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
2622                                           NULL,
2623                                           "no counter found for flow");
2624         }
2625         return 0;
2626 }
2627 #endif
2628
2629 /**
2630  * Isolated mode.
2631  *
2632  * @see rte_flow_isolate()
2633  * @see rte_flow_ops
2634  */
2635 int
2636 mlx5_flow_isolate(struct rte_eth_dev *dev,
2637                   int enable,
2638                   struct rte_flow_error *error)
2639 {
2640         struct priv *priv = dev->data->dev_private;
2641
2642         if (dev->data->dev_started) {
2643                 rte_flow_error_set(error, EBUSY,
2644                                    RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
2645                                    NULL,
2646                                    "port must be stopped first");
2647                 return -rte_errno;
2648         }
2649         priv->isolated = !!enable;
2650         if (enable)
2651                 priv->dev->dev_ops = &mlx5_dev_ops_isolate;
2652         else
2653                 priv->dev->dev_ops = &mlx5_dev_ops;
2654         return 0;
2655 }
2656
2657 /**
2658  * Convert a flow director filter to a generic flow.
2659  *
2660  * @param dev
2661  *   Pointer to Ethernet device.
2662  * @param fdir_filter
2663  *   Flow director filter to add.
2664  * @param attributes
2665  *   Generic flow parameters structure.
2666  *
2667  * @return
2668  *   0 on success, a negative errno value otherwise and rte_errno is set.
2669  */
2670 static int
2671 mlx5_fdir_filter_convert(struct rte_eth_dev *dev,
2672                          const struct rte_eth_fdir_filter *fdir_filter,
2673                          struct mlx5_fdir *attributes)
2674 {
2675         struct priv *priv = dev->data->dev_private;
2676         const struct rte_eth_fdir_input *input = &fdir_filter->input;
2677         const struct rte_eth_fdir_masks *mask =
2678                 &dev->data->dev_conf.fdir_conf.mask;
2679
2680         /* Validate queue number. */
2681         if (fdir_filter->action.rx_queue >= priv->rxqs_n) {
2682                 DRV_LOG(ERR, "port %u invalid queue number %d",
2683                         dev->data->port_id, fdir_filter->action.rx_queue);
2684                 rte_errno = EINVAL;
2685                 return -rte_errno;
2686         }
2687         attributes->attr.ingress = 1;
2688         attributes->items[0] = (struct rte_flow_item) {
2689                 .type = RTE_FLOW_ITEM_TYPE_ETH,
2690                 .spec = &attributes->l2,
2691                 .mask = &attributes->l2_mask,
2692         };
2693         switch (fdir_filter->action.behavior) {
2694         case RTE_ETH_FDIR_ACCEPT:
2695                 attributes->actions[0] = (struct rte_flow_action){
2696                         .type = RTE_FLOW_ACTION_TYPE_QUEUE,
2697                         .conf = &attributes->queue,
2698                 };
2699                 break;
2700         case RTE_ETH_FDIR_REJECT:
2701                 attributes->actions[0] = (struct rte_flow_action){
2702                         .type = RTE_FLOW_ACTION_TYPE_DROP,
2703                 };
2704                 break;
2705         default:
2706                 DRV_LOG(ERR, "port %u invalid behavior %d",
2707                         dev->data->port_id,
2708                         fdir_filter->action.behavior);
2709                 rte_errno = ENOTSUP;
2710                 return -rte_errno;
2711         }
2712         attributes->queue.index = fdir_filter->action.rx_queue;
2713         /* Handle L3. */
2714         switch (fdir_filter->input.flow_type) {
2715         case RTE_ETH_FLOW_NONFRAG_IPV4_UDP:
2716         case RTE_ETH_FLOW_NONFRAG_IPV4_TCP:
2717         case RTE_ETH_FLOW_NONFRAG_IPV4_OTHER:
2718                 attributes->l3.ipv4.hdr = (struct ipv4_hdr){
2719                         .src_addr = input->flow.ip4_flow.src_ip,
2720                         .dst_addr = input->flow.ip4_flow.dst_ip,
2721                         .time_to_live = input->flow.ip4_flow.ttl,
2722                         .type_of_service = input->flow.ip4_flow.tos,
2723                         .next_proto_id = input->flow.ip4_flow.proto,
2724                 };
2725                 attributes->l3_mask.ipv4.hdr = (struct ipv4_hdr){
2726                         .src_addr = mask->ipv4_mask.src_ip,
2727                         .dst_addr = mask->ipv4_mask.dst_ip,
2728                         .time_to_live = mask->ipv4_mask.ttl,
2729                         .type_of_service = mask->ipv4_mask.tos,
2730                         .next_proto_id = mask->ipv4_mask.proto,
2731                 };
2732                 attributes->items[1] = (struct rte_flow_item){
2733                         .type = RTE_FLOW_ITEM_TYPE_IPV4,
2734                         .spec = &attributes->l3,
2735                         .mask = &attributes->l3_mask,
2736                 };
2737                 break;
2738         case RTE_ETH_FLOW_NONFRAG_IPV6_UDP:
2739         case RTE_ETH_FLOW_NONFRAG_IPV6_TCP:
2740         case RTE_ETH_FLOW_NONFRAG_IPV6_OTHER:
2741                 attributes->l3.ipv6.hdr = (struct ipv6_hdr){
2742                         .hop_limits = input->flow.ipv6_flow.hop_limits,
2743                         .proto = input->flow.ipv6_flow.proto,
2744                 };
2745
2746                 memcpy(attributes->l3.ipv6.hdr.src_addr,
2747                        input->flow.ipv6_flow.src_ip,
2748                        RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
2749                 memcpy(attributes->l3.ipv6.hdr.dst_addr,
2750                        input->flow.ipv6_flow.dst_ip,
2751                        RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
2752                 memcpy(attributes->l3_mask.ipv6.hdr.src_addr,
2753                        mask->ipv6_mask.src_ip,
2754                        RTE_DIM(attributes->l3_mask.ipv6.hdr.src_addr));
2755                 memcpy(attributes->l3_mask.ipv6.hdr.dst_addr,
2756                        mask->ipv6_mask.dst_ip,
2757                        RTE_DIM(attributes->l3_mask.ipv6.hdr.src_addr));
2758                 attributes->items[1] = (struct rte_flow_item){
2759                         .type = RTE_FLOW_ITEM_TYPE_IPV6,
2760                         .spec = &attributes->l3,
2761                         .mask = &attributes->l3_mask,
2762                 };
2763                 break;
2764         default:
2765                 DRV_LOG(ERR, "port %u invalid flow type%d",
2766                         dev->data->port_id, fdir_filter->input.flow_type);
2767                 rte_errno = ENOTSUP;
2768                 return -rte_errno;
2769         }
2770         /* Handle L4. */
2771         switch (fdir_filter->input.flow_type) {
2772         case RTE_ETH_FLOW_NONFRAG_IPV4_UDP:
2773                 attributes->l4.udp.hdr = (struct udp_hdr){
2774                         .src_port = input->flow.udp4_flow.src_port,
2775                         .dst_port = input->flow.udp4_flow.dst_port,
2776                 };
2777                 attributes->l4_mask.udp.hdr = (struct udp_hdr){
2778                         .src_port = mask->src_port_mask,
2779                         .dst_port = mask->dst_port_mask,
2780                 };
2781                 attributes->items[2] = (struct rte_flow_item){
2782                         .type = RTE_FLOW_ITEM_TYPE_UDP,
2783                         .spec = &attributes->l4,
2784                         .mask = &attributes->l4_mask,
2785                 };
2786                 break;
2787         case RTE_ETH_FLOW_NONFRAG_IPV4_TCP:
2788                 attributes->l4.tcp.hdr = (struct tcp_hdr){
2789                         .src_port = input->flow.tcp4_flow.src_port,
2790                         .dst_port = input->flow.tcp4_flow.dst_port,
2791                 };
2792                 attributes->l4_mask.tcp.hdr = (struct tcp_hdr){
2793                         .src_port = mask->src_port_mask,
2794                         .dst_port = mask->dst_port_mask,
2795                 };
2796                 attributes->items[2] = (struct rte_flow_item){
2797                         .type = RTE_FLOW_ITEM_TYPE_TCP,
2798                         .spec = &attributes->l4,
2799                         .mask = &attributes->l4_mask,
2800                 };
2801                 break;
2802         case RTE_ETH_FLOW_NONFRAG_IPV6_UDP:
2803                 attributes->l4.udp.hdr = (struct udp_hdr){
2804                         .src_port = input->flow.udp6_flow.src_port,
2805                         .dst_port = input->flow.udp6_flow.dst_port,
2806                 };
2807                 attributes->l4_mask.udp.hdr = (struct udp_hdr){
2808                         .src_port = mask->src_port_mask,
2809                         .dst_port = mask->dst_port_mask,
2810                 };
2811                 attributes->items[2] = (struct rte_flow_item){
2812                         .type = RTE_FLOW_ITEM_TYPE_UDP,
2813                         .spec = &attributes->l4,
2814                         .mask = &attributes->l4_mask,
2815                 };
2816                 break;
2817         case RTE_ETH_FLOW_NONFRAG_IPV6_TCP:
2818                 attributes->l4.tcp.hdr = (struct tcp_hdr){
2819                         .src_port = input->flow.tcp6_flow.src_port,
2820                         .dst_port = input->flow.tcp6_flow.dst_port,
2821                 };
2822                 attributes->l4_mask.tcp.hdr = (struct tcp_hdr){
2823                         .src_port = mask->src_port_mask,
2824                         .dst_port = mask->dst_port_mask,
2825                 };
2826                 attributes->items[2] = (struct rte_flow_item){
2827                         .type = RTE_FLOW_ITEM_TYPE_TCP,
2828                         .spec = &attributes->l4,
2829                         .mask = &attributes->l4_mask,
2830                 };
2831                 break;
2832         case RTE_ETH_FLOW_NONFRAG_IPV4_OTHER:
2833         case RTE_ETH_FLOW_NONFRAG_IPV6_OTHER:
2834                 break;
2835         default:
2836                 DRV_LOG(ERR, "port %u invalid flow type%d",
2837                         dev->data->port_id, fdir_filter->input.flow_type);
2838                 rte_errno = ENOTSUP;
2839                 return -rte_errno;
2840         }
2841         return 0;
2842 }
2843
2844 /**
2845  * Add new flow director filter and store it in list.
2846  *
2847  * @param dev
2848  *   Pointer to Ethernet device.
2849  * @param fdir_filter
2850  *   Flow director filter to add.
2851  *
2852  * @return
2853  *   0 on success, a negative errno value otherwise and rte_errno is set.
2854  */
2855 static int
2856 mlx5_fdir_filter_add(struct rte_eth_dev *dev,
2857                      const struct rte_eth_fdir_filter *fdir_filter)
2858 {
2859         struct priv *priv = dev->data->dev_private;
2860         struct mlx5_fdir attributes = {
2861                 .attr.group = 0,
2862                 .l2_mask = {
2863                         .dst.addr_bytes = "\x00\x00\x00\x00\x00\x00",
2864                         .src.addr_bytes = "\x00\x00\x00\x00\x00\x00",
2865                         .type = 0,
2866                 },
2867         };
2868         struct mlx5_flow_parse parser = {
2869                 .layer = HASH_RXQ_ETH,
2870         };
2871         struct rte_flow_error error;
2872         struct rte_flow *flow;
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                 return ret;
2882         flow = mlx5_flow_list_create(dev, &priv->flows, &attributes.attr,
2883                                      attributes.items, attributes.actions,
2884                                      &error);
2885         if (flow) {
2886                 DRV_LOG(DEBUG, "port %u FDIR created %p", dev->data->port_id,
2887                         (void *)flow);
2888                 return 0;
2889         }
2890         return -rte_errno;
2891 }
2892
2893 /**
2894  * Delete specific filter.
2895  *
2896  * @param dev
2897  *   Pointer to Ethernet device.
2898  * @param fdir_filter
2899  *   Filter to be deleted.
2900  *
2901  * @return
2902  *   0 on success, a negative errno value otherwise and rte_errno is set.
2903  */
2904 static int
2905 mlx5_fdir_filter_delete(struct rte_eth_dev *dev,
2906                         const struct rte_eth_fdir_filter *fdir_filter)
2907 {
2908         struct priv *priv = dev->data->dev_private;
2909         struct mlx5_fdir attributes = {
2910                 .attr.group = 0,
2911         };
2912         struct mlx5_flow_parse parser = {
2913                 .create = 1,
2914                 .layer = HASH_RXQ_ETH,
2915         };
2916         struct rte_flow_error error;
2917         struct rte_flow *flow;
2918         unsigned int i;
2919         int ret;
2920
2921         ret = mlx5_fdir_filter_convert(dev, fdir_filter, &attributes);
2922         if (ret)
2923                 return ret;
2924         ret = mlx5_flow_convert(dev, &attributes.attr, attributes.items,
2925                                 attributes.actions, &error, &parser);
2926         if (ret)
2927                 goto exit;
2928         /*
2929          * Special case for drop action which is only set in the
2930          * specifications when the flow is created.  In this situation the
2931          * drop specification is missing.
2932          */
2933         if (parser.drop) {
2934                 struct ibv_flow_spec_action_drop *drop;
2935
2936                 drop = (void *)((uintptr_t)parser.queue[HASH_RXQ_ETH].ibv_attr +
2937                                 parser.queue[HASH_RXQ_ETH].offset);
2938                 *drop = (struct ibv_flow_spec_action_drop){
2939                         .type = IBV_FLOW_SPEC_ACTION_DROP,
2940                         .size = sizeof(struct ibv_flow_spec_action_drop),
2941                 };
2942                 parser.queue[HASH_RXQ_ETH].ibv_attr->num_of_specs++;
2943         }
2944         TAILQ_FOREACH(flow, &priv->flows, next) {
2945                 struct ibv_flow_attr *attr;
2946                 struct ibv_spec_header *attr_h;
2947                 void *spec;
2948                 struct ibv_flow_attr *flow_attr;
2949                 struct ibv_spec_header *flow_h;
2950                 void *flow_spec;
2951                 unsigned int specs_n;
2952
2953                 attr = parser.queue[HASH_RXQ_ETH].ibv_attr;
2954                 flow_attr = flow->frxq[HASH_RXQ_ETH].ibv_attr;
2955                 /* Compare first the attributes. */
2956                 if (memcmp(attr, flow_attr, sizeof(struct ibv_flow_attr)))
2957                         continue;
2958                 if (attr->num_of_specs == 0)
2959                         continue;
2960                 spec = (void *)((uintptr_t)attr +
2961                                 sizeof(struct ibv_flow_attr));
2962                 flow_spec = (void *)((uintptr_t)flow_attr +
2963                                      sizeof(struct ibv_flow_attr));
2964                 specs_n = RTE_MIN(attr->num_of_specs, flow_attr->num_of_specs);
2965                 for (i = 0; i != specs_n; ++i) {
2966                         attr_h = spec;
2967                         flow_h = flow_spec;
2968                         if (memcmp(spec, flow_spec,
2969                                    RTE_MIN(attr_h->size, flow_h->size)))
2970                                 goto wrong_flow;
2971                         spec = (void *)((uintptr_t)spec + attr_h->size);
2972                         flow_spec = (void *)((uintptr_t)flow_spec +
2973                                              flow_h->size);
2974                 }
2975                 /* At this point, the flow match. */
2976                 break;
2977 wrong_flow:
2978                 /* The flow does not match. */
2979                 continue;
2980         }
2981         ret = rte_errno; /* Save rte_errno before cleanup. */
2982         if (flow)
2983                 mlx5_flow_list_destroy(dev, &priv->flows, flow);
2984 exit:
2985         for (i = 0; i != hash_rxq_init_n; ++i) {
2986                 if (parser.queue[i].ibv_attr)
2987                         rte_free(parser.queue[i].ibv_attr);
2988         }
2989         rte_errno = ret; /* Restore rte_errno. */
2990         return -rte_errno;
2991 }
2992
2993 /**
2994  * Update queue for specific filter.
2995  *
2996  * @param dev
2997  *   Pointer to Ethernet device.
2998  * @param fdir_filter
2999  *   Filter to be updated.
3000  *
3001  * @return
3002  *   0 on success, a negative errno value otherwise and rte_errno is set.
3003  */
3004 static int
3005 mlx5_fdir_filter_update(struct rte_eth_dev *dev,
3006                         const struct rte_eth_fdir_filter *fdir_filter)
3007 {
3008         int ret;
3009
3010         ret = mlx5_fdir_filter_delete(dev, fdir_filter);
3011         if (ret)
3012                 return ret;
3013         return mlx5_fdir_filter_add(dev, fdir_filter);
3014 }
3015
3016 /**
3017  * Flush all filters.
3018  *
3019  * @param dev
3020  *   Pointer to Ethernet device.
3021  */
3022 static void
3023 mlx5_fdir_filter_flush(struct rte_eth_dev *dev)
3024 {
3025         struct priv *priv = dev->data->dev_private;
3026
3027         mlx5_flow_list_flush(dev, &priv->flows);
3028 }
3029
3030 /**
3031  * Get flow director information.
3032  *
3033  * @param dev
3034  *   Pointer to Ethernet device.
3035  * @param[out] fdir_info
3036  *   Resulting flow director information.
3037  */
3038 static void
3039 mlx5_fdir_info_get(struct rte_eth_dev *dev, struct rte_eth_fdir_info *fdir_info)
3040 {
3041         struct priv *priv = dev->data->dev_private;
3042         struct rte_eth_fdir_masks *mask =
3043                 &priv->dev->data->dev_conf.fdir_conf.mask;
3044
3045         fdir_info->mode = priv->dev->data->dev_conf.fdir_conf.mode;
3046         fdir_info->guarant_spc = 0;
3047         rte_memcpy(&fdir_info->mask, mask, sizeof(fdir_info->mask));
3048         fdir_info->max_flexpayload = 0;
3049         fdir_info->flow_types_mask[0] = 0;
3050         fdir_info->flex_payload_unit = 0;
3051         fdir_info->max_flex_payload_segment_num = 0;
3052         fdir_info->flex_payload_limit = 0;
3053         memset(&fdir_info->flex_conf, 0, sizeof(fdir_info->flex_conf));
3054 }
3055
3056 /**
3057  * Deal with flow director operations.
3058  *
3059  * @param dev
3060  *   Pointer to Ethernet device.
3061  * @param filter_op
3062  *   Operation to perform.
3063  * @param arg
3064  *   Pointer to operation-specific structure.
3065  *
3066  * @return
3067  *   0 on success, a negative errno value otherwise and rte_errno is set.
3068  */
3069 static int
3070 mlx5_fdir_ctrl_func(struct rte_eth_dev *dev, enum rte_filter_op filter_op,
3071                     void *arg)
3072 {
3073         struct priv *priv = dev->data->dev_private;
3074         enum rte_fdir_mode fdir_mode =
3075                 priv->dev->data->dev_conf.fdir_conf.mode;
3076
3077         if (filter_op == RTE_ETH_FILTER_NOP)
3078                 return 0;
3079         if (fdir_mode != RTE_FDIR_MODE_PERFECT &&
3080             fdir_mode != RTE_FDIR_MODE_PERFECT_MAC_VLAN) {
3081                 DRV_LOG(ERR, "port %u flow director mode %d not supported",
3082                         dev->data->port_id, fdir_mode);
3083                 rte_errno = EINVAL;
3084                 return -rte_errno;
3085         }
3086         switch (filter_op) {
3087         case RTE_ETH_FILTER_ADD:
3088                 return mlx5_fdir_filter_add(dev, arg);
3089         case RTE_ETH_FILTER_UPDATE:
3090                 return mlx5_fdir_filter_update(dev, arg);
3091         case RTE_ETH_FILTER_DELETE:
3092                 return mlx5_fdir_filter_delete(dev, arg);
3093         case RTE_ETH_FILTER_FLUSH:
3094                 mlx5_fdir_filter_flush(dev);
3095                 break;
3096         case RTE_ETH_FILTER_INFO:
3097                 mlx5_fdir_info_get(dev, arg);
3098                 break;
3099         default:
3100                 DRV_LOG(DEBUG, "port %u unknown operation %u",
3101                         dev->data->port_id, filter_op);
3102                 rte_errno = EINVAL;
3103                 return -rte_errno;
3104         }
3105         return 0;
3106 }
3107
3108 /**
3109  * Manage filter operations.
3110  *
3111  * @param dev
3112  *   Pointer to Ethernet device structure.
3113  * @param filter_type
3114  *   Filter type.
3115  * @param filter_op
3116  *   Operation to perform.
3117  * @param arg
3118  *   Pointer to operation-specific structure.
3119  *
3120  * @return
3121  *   0 on success, a negative errno value otherwise and rte_errno is set.
3122  */
3123 int
3124 mlx5_dev_filter_ctrl(struct rte_eth_dev *dev,
3125                      enum rte_filter_type filter_type,
3126                      enum rte_filter_op filter_op,
3127                      void *arg)
3128 {
3129         switch (filter_type) {
3130         case RTE_ETH_FILTER_GENERIC:
3131                 if (filter_op != RTE_ETH_FILTER_GET) {
3132                         rte_errno = EINVAL;
3133                         return -rte_errno;
3134                 }
3135                 *(const void **)arg = &mlx5_flow_ops;
3136                 return 0;
3137         case RTE_ETH_FILTER_FDIR:
3138                 return mlx5_fdir_ctrl_func(dev, filter_op, arg);
3139         default:
3140                 DRV_LOG(ERR, "port %u filter type (%d) not supported",
3141                         dev->data->port_id, filter_type);
3142                 rte_errno = ENOTSUP;
3143                 return -rte_errno;
3144         }
3145         return 0;
3146 }