drivers/net/mlx5/mlx5_flow.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright 2016 6WIND S.A.
   5  *   Copyright 2016 Mellanox.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of 6WIND S.A. nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  */
  33
  34 #include <sys/queue.h>
  35 #include <string.h>
  36
  37 /* Verbs header. */
  38 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
  39 #ifdef PEDANTIC
  40 #pragma GCC diagnostic ignored "-Wpedantic"
  41 #endif
  42 #include <infiniband/verbs.h>
  43 #ifdef PEDANTIC
  44 #pragma GCC diagnostic error "-Wpedantic"
  45 #endif
  46
  47 #include <rte_ethdev.h>
  48 #include <rte_flow.h>
  49 #include <rte_flow_driver.h>
  50 #include <rte_malloc.h>
  51
  52 #include "mlx5.h"
  53 #include "mlx5_prm.h"
  54
  55 /* Define minimal priority for control plane flows. */
  56 #define MLX5_CTRL_FLOW_PRIORITY 4
  57
  58 /* Internet Protocol versions. */
  59 #define MLX5_IPV4 4
  60 #define MLX5_IPV6 6
  61
  62 #ifndef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
  63 struct ibv_counter_set_init_attr {
  64         int dummy;
  65 };
  66 struct ibv_flow_spec_counter_action {
  67         int dummy;
  68 };
  69 struct ibv_counter_set {
  70         int dummy;
  71 };
  72
  73 static inline int
  74 ibv_destroy_counter_set(struct ibv_counter_set *cs)
  75 {
  76         (void)cs;
  77         return -ENOTSUP;
  78 }
  79 #endif
  80
  81 /* Dev ops structure defined in mlx5.c */
  82 extern const struct eth_dev_ops mlx5_dev_ops;
  83 extern const struct eth_dev_ops mlx5_dev_ops_isolate;
  84
  85 static int
  86 mlx5_flow_create_eth(const struct rte_flow_item *item,
  87                      const void *default_mask,
  88                      void *data);
  89
  90 static int
  91 mlx5_flow_create_vlan(const struct rte_flow_item *item,
  92                       const void *default_mask,
  93                       void *data);
  94
  95 static int
  96 mlx5_flow_create_ipv4(const struct rte_flow_item *item,
  97                       const void *default_mask,
  98                       void *data);
  99
 100 static int
 101 mlx5_flow_create_ipv6(const struct rte_flow_item *item,
 102                       const void *default_mask,
 103                       void *data);
 104
 105 static int
 106 mlx5_flow_create_udp(const struct rte_flow_item *item,
 107                      const void *default_mask,
 108                      void *data);
 109
 110 static int
 111 mlx5_flow_create_tcp(const struct rte_flow_item *item,
 112                      const void *default_mask,
 113                      void *data);
 114
 115 static int
 116 mlx5_flow_create_vxlan(const struct rte_flow_item *item,
 117                        const void *default_mask,
 118                        void *data);
 119
 120 struct mlx5_flow_parse;
 121
 122 static void
 123 mlx5_flow_create_copy(struct mlx5_flow_parse *parser, void *src,
 124                       unsigned int size);
 125
 126 static int
 127 mlx5_flow_create_flag_mark(struct mlx5_flow_parse *parser, uint32_t mark_id);
 128
 129 static int
 130 mlx5_flow_create_count(struct priv *priv, struct mlx5_flow_parse *parser);
 131
 132 /* Hash RX queue types. */
 133 enum hash_rxq_type {
 134         HASH_RXQ_TCPV4,
 135         HASH_RXQ_UDPV4,
 136         HASH_RXQ_IPV4,
 137         HASH_RXQ_TCPV6,
 138         HASH_RXQ_UDPV6,
 139         HASH_RXQ_IPV6,
 140         HASH_RXQ_ETH,
 141 };
 142
 143 /* Initialization data for hash RX queue. */
 144 struct hash_rxq_init {
 145         uint64_t hash_fields; /* Fields that participate in the hash. */
 146         uint64_t dpdk_rss_hf; /* Matching DPDK RSS hash fields. */
 147         unsigned int flow_priority; /* Flow priority to use. */
 148         unsigned int ip_version; /* Internet protocol. */
 149 };
 150
 151 /* Initialization data for hash RX queues. */
 152 const struct hash_rxq_init hash_rxq_init[] = {
 153         [HASH_RXQ_TCPV4] = {
 154                 .hash_fields = (IBV_RX_HASH_SRC_IPV4 |
 155                                 IBV_RX_HASH_DST_IPV4 |
 156                                 IBV_RX_HASH_SRC_PORT_TCP |
 157                                 IBV_RX_HASH_DST_PORT_TCP),
 158                 .dpdk_rss_hf = ETH_RSS_NONFRAG_IPV4_TCP,
 159                 .flow_priority = 0,
 160                 .ip_version = MLX5_IPV4,
 161         },
 162         [HASH_RXQ_UDPV4] = {
 163                 .hash_fields = (IBV_RX_HASH_SRC_IPV4 |
 164                                 IBV_RX_HASH_DST_IPV4 |
 165                                 IBV_RX_HASH_SRC_PORT_UDP |
 166                                 IBV_RX_HASH_DST_PORT_UDP),
 167                 .dpdk_rss_hf = ETH_RSS_NONFRAG_IPV4_UDP,
 168                 .flow_priority = 0,
 169                 .ip_version = MLX5_IPV4,
 170         },
 171         [HASH_RXQ_IPV4] = {
 172                 .hash_fields = (IBV_RX_HASH_SRC_IPV4 |
 173                                 IBV_RX_HASH_DST_IPV4),
 174                 .dpdk_rss_hf = (ETH_RSS_IPV4 |
 175                                 ETH_RSS_FRAG_IPV4),
 176                 .flow_priority = 1,
 177                 .ip_version = MLX5_IPV4,
 178         },
 179         [HASH_RXQ_TCPV6] = {
 180                 .hash_fields = (IBV_RX_HASH_SRC_IPV6 |
 181                                 IBV_RX_HASH_DST_IPV6 |
 182                                 IBV_RX_HASH_SRC_PORT_TCP |
 183                                 IBV_RX_HASH_DST_PORT_TCP),
 184                 .dpdk_rss_hf = ETH_RSS_NONFRAG_IPV6_TCP,
 185                 .flow_priority = 0,
 186                 .ip_version = MLX5_IPV6,
 187         },
 188         [HASH_RXQ_UDPV6] = {
 189                 .hash_fields = (IBV_RX_HASH_SRC_IPV6 |
 190                                 IBV_RX_HASH_DST_IPV6 |
 191                                 IBV_RX_HASH_SRC_PORT_UDP |
 192                                 IBV_RX_HASH_DST_PORT_UDP),
 193                 .dpdk_rss_hf = ETH_RSS_NONFRAG_IPV6_UDP,
 194                 .flow_priority = 0,
 195                 .ip_version = MLX5_IPV6,
 196         },
 197         [HASH_RXQ_IPV6] = {
 198                 .hash_fields = (IBV_RX_HASH_SRC_IPV6 |
 199                                 IBV_RX_HASH_DST_IPV6),
 200                 .dpdk_rss_hf = (ETH_RSS_IPV6 |
 201                                 ETH_RSS_FRAG_IPV6),
 202                 .flow_priority = 1,
 203                 .ip_version = MLX5_IPV6,
 204         },
 205         [HASH_RXQ_ETH] = {
 206                 .hash_fields = 0,
 207                 .dpdk_rss_hf = 0,
 208                 .flow_priority = 2,
 209         },
 210 };
 211
 212 /* Number of entries in hash_rxq_init[]. */
 213 const unsigned int hash_rxq_init_n = RTE_DIM(hash_rxq_init);
 214
 215 /** Structure for holding counter stats. */
 216 struct mlx5_flow_counter_stats {
 217         uint64_t hits; /**< Number of packets matched by the rule. */
 218         uint64_t bytes; /**< Number of bytes matched by the rule. */
 219 };
 220
 221 /** Structure for Drop queue. */
 222 struct mlx5_hrxq_drop {
 223         struct ibv_rwq_ind_table *ind_table; /**< Indirection table. */
 224         struct ibv_qp *qp; /**< Verbs queue pair. */
 225         struct ibv_wq *wq; /**< Verbs work queue. */
 226         struct ibv_cq *cq; /**< Verbs completion queue. */
 227 };
 228
 229 /* Flows structures. */
 230 struct mlx5_flow {
 231         uint64_t hash_fields; /**< Fields that participate in the hash. */
 232         struct ibv_flow_attr *ibv_attr; /**< Pointer to Verbs attributes. */
 233         struct ibv_flow *ibv_flow; /**< Verbs flow. */
 234         struct mlx5_hrxq *hrxq; /**< Hash Rx queues. */
 235 };
 236
 237 /* Drop flows structures. */
 238 struct mlx5_flow_drop {
 239         struct ibv_flow_attr *ibv_attr; /**< Pointer to Verbs attributes. */
 240         struct ibv_flow *ibv_flow; /**< Verbs flow. */
 241 };
 242
 243 struct rte_flow {
 244         TAILQ_ENTRY(rte_flow) next; /**< Pointer to the next flow structure. */
 245         uint32_t mark:1; /**< Set if the flow is marked. */
 246         uint32_t drop:1; /**< Drop queue. */
 247         uint16_t queues_n; /**< Number of entries in queue[]. */
 248         uint16_t (*queues)[]; /**< Queues indexes to use. */
 249         struct rte_eth_rss_conf rss_conf; /**< RSS configuration */
 250         uint8_t rss_key[40]; /**< copy of the RSS key. */
 251         struct ibv_counter_set *cs; /**< Holds the counters for the rule. */
 252         struct mlx5_flow_counter_stats counter_stats;/**<The counter stats. */
 253         union {
 254                 struct mlx5_flow frxq[RTE_DIM(hash_rxq_init)];
 255                 /**< Flow with Rx queue. */
 256                 struct mlx5_flow_drop drxq; /**< Flow with drop Rx queue. */
 257         };
 258 };
 259
 260 /** Static initializer for items. */
 261 #define ITEMS(...) \
 262         (const enum rte_flow_item_type []){ \
 263                 __VA_ARGS__, RTE_FLOW_ITEM_TYPE_END, \
 264         }
 265
 266 /** Structure to generate a simple graph of layers supported by the NIC. */
 267 struct mlx5_flow_items {
 268         /** List of possible actions for these items. */
 269         const enum rte_flow_action_type *const actions;
 270         /** Bit-masks corresponding to the possibilities for the item. */
 271         const void *mask;
 272         /**
 273          * Default bit-masks to use when item->mask is not provided. When
 274          * \default_mask is also NULL, the full supported bit-mask (\mask) is
 275          * used instead.
 276          */
 277         const void *default_mask;
 278         /** Bit-masks size in bytes. */
 279         const unsigned int mask_sz;
 280         /**
 281          * Conversion function from rte_flow to NIC specific flow.
 282          *
 283          * @param item
 284          *   rte_flow item to convert.
 285          * @param default_mask
 286          *   Default bit-masks to use when item->mask is not provided.
 287          * @param data
 288          *   Internal structure to store the conversion.
 289          *
 290          * @return
 291          *   0 on success, negative value otherwise.
 292          */
 293         int (*convert)(const struct rte_flow_item *item,
 294                        const void *default_mask,
 295                        void *data);
 296         /** Size in bytes of the destination structure. */
 297         const unsigned int dst_sz;
 298         /** List of possible following items.  */
 299         const enum rte_flow_item_type *const items;
 300 };
 301
 302 /** Valid action for this PMD. */
 303 static const enum rte_flow_action_type valid_actions[] = {
 304         RTE_FLOW_ACTION_TYPE_DROP,
 305         RTE_FLOW_ACTION_TYPE_QUEUE,
 306         RTE_FLOW_ACTION_TYPE_MARK,
 307         RTE_FLOW_ACTION_TYPE_FLAG,
 308 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
 309         RTE_FLOW_ACTION_TYPE_COUNT,
 310 #endif
 311         RTE_FLOW_ACTION_TYPE_END,
 312 };
 313
 314 /** Graph of supported items and associated actions. */
 315 static const struct mlx5_flow_items mlx5_flow_items[] = {
 316         [RTE_FLOW_ITEM_TYPE_END] = {
 317                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_ETH,
 318                                RTE_FLOW_ITEM_TYPE_VXLAN),
 319         },
 320         [RTE_FLOW_ITEM_TYPE_ETH] = {
 321                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_VLAN,
 322                                RTE_FLOW_ITEM_TYPE_IPV4,
 323                                RTE_FLOW_ITEM_TYPE_IPV6),
 324                 .actions = valid_actions,
 325                 .mask = &(const struct rte_flow_item_eth){
 326                         .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 327                         .src.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 328                         .type = -1,
 329                 },
 330                 .default_mask = &rte_flow_item_eth_mask,
 331                 .mask_sz = sizeof(struct rte_flow_item_eth),
 332                 .convert = mlx5_flow_create_eth,
 333                 .dst_sz = sizeof(struct ibv_flow_spec_eth),
 334         },
 335         [RTE_FLOW_ITEM_TYPE_VLAN] = {
 336                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_IPV4,
 337                                RTE_FLOW_ITEM_TYPE_IPV6),
 338                 .actions = valid_actions,
 339                 .mask = &(const struct rte_flow_item_vlan){
 340                         .tci = -1,
 341                 },
 342                 .default_mask = &rte_flow_item_vlan_mask,
 343                 .mask_sz = sizeof(struct rte_flow_item_vlan),
 344                 .convert = mlx5_flow_create_vlan,
 345                 .dst_sz = 0,
 346         },
 347         [RTE_FLOW_ITEM_TYPE_IPV4] = {
 348                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP,
 349                                RTE_FLOW_ITEM_TYPE_TCP),
 350                 .actions = valid_actions,
 351                 .mask = &(const struct rte_flow_item_ipv4){
 352                         .hdr = {
 353                                 .src_addr = -1,
 354                                 .dst_addr = -1,
 355                                 .type_of_service = -1,
 356                                 .next_proto_id = -1,
 357                         },
 358                 },
 359                 .default_mask = &rte_flow_item_ipv4_mask,
 360                 .mask_sz = sizeof(struct rte_flow_item_ipv4),
 361                 .convert = mlx5_flow_create_ipv4,
 362                 .dst_sz = sizeof(struct ibv_flow_spec_ipv4_ext),
 363         },
 364         [RTE_FLOW_ITEM_TYPE_IPV6] = {
 365                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP,
 366                                RTE_FLOW_ITEM_TYPE_TCP),
 367                 .actions = valid_actions,
 368                 .mask = &(const struct rte_flow_item_ipv6){
 369                         .hdr = {
 370                                 .src_addr = {
 371                                         0xff, 0xff, 0xff, 0xff,
 372                                         0xff, 0xff, 0xff, 0xff,
 373                                         0xff, 0xff, 0xff, 0xff,
 374                                         0xff, 0xff, 0xff, 0xff,
 375                                 },
 376                                 .dst_addr = {
 377                                         0xff, 0xff, 0xff, 0xff,
 378                                         0xff, 0xff, 0xff, 0xff,
 379                                         0xff, 0xff, 0xff, 0xff,
 380                                         0xff, 0xff, 0xff, 0xff,
 381                                 },
 382                                 .vtc_flow = -1,
 383                                 .proto = -1,
 384                                 .hop_limits = -1,
 385                         },
 386                 },
 387                 .default_mask = &rte_flow_item_ipv6_mask,
 388                 .mask_sz = sizeof(struct rte_flow_item_ipv6),
 389                 .convert = mlx5_flow_create_ipv6,
 390                 .dst_sz = sizeof(struct ibv_flow_spec_ipv6),
 391         },
 392         [RTE_FLOW_ITEM_TYPE_UDP] = {
 393                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_VXLAN),
 394                 .actions = valid_actions,
 395                 .mask = &(const struct rte_flow_item_udp){
 396                         .hdr = {
 397                                 .src_port = -1,
 398                                 .dst_port = -1,
 399                         },
 400                 },
 401                 .default_mask = &rte_flow_item_udp_mask,
 402                 .mask_sz = sizeof(struct rte_flow_item_udp),
 403                 .convert = mlx5_flow_create_udp,
 404                 .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
 405         },
 406         [RTE_FLOW_ITEM_TYPE_TCP] = {
 407                 .actions = valid_actions,
 408                 .mask = &(const struct rte_flow_item_tcp){
 409                         .hdr = {
 410                                 .src_port = -1,
 411                                 .dst_port = -1,
 412                         },
 413                 },
 414                 .default_mask = &rte_flow_item_tcp_mask,
 415                 .mask_sz = sizeof(struct rte_flow_item_tcp),
 416                 .convert = mlx5_flow_create_tcp,
 417                 .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
 418         },
 419         [RTE_FLOW_ITEM_TYPE_VXLAN] = {
 420                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_ETH),
 421                 .actions = valid_actions,
 422                 .mask = &(const struct rte_flow_item_vxlan){
 423                         .vni = "\xff\xff\xff",
 424                 },
 425                 .default_mask = &rte_flow_item_vxlan_mask,
 426                 .mask_sz = sizeof(struct rte_flow_item_vxlan),
 427                 .convert = mlx5_flow_create_vxlan,
 428                 .dst_sz = sizeof(struct ibv_flow_spec_tunnel),
 429         },
 430 };
 431
 432 /** Structure to pass to the conversion function. */
 433 struct mlx5_flow_parse {
 434         uint32_t inner; /**< Set once VXLAN is encountered. */
 435         uint32_t create:1;
 436         /**< Whether resources should remain after a validate. */
 437         uint32_t drop:1; /**< Target is a drop queue. */
 438         uint32_t mark:1; /**< Mark is present in the flow. */
 439         uint32_t count:1; /**< Count is present in the flow. */
 440         uint32_t mark_id; /**< Mark identifier. */
 441         uint16_t queues[RTE_MAX_QUEUES_PER_PORT]; /**< Queues indexes to use. */
 442         uint16_t queues_n; /**< Number of entries in queue[]. */
 443         struct rte_eth_rss_conf rss_conf; /**< RSS configuration */
 444         uint8_t rss_key[40]; /**< copy of the RSS key. */
 445         enum hash_rxq_type layer; /**< Last pattern layer detected. */
 446         struct ibv_counter_set *cs; /**< Holds the counter set for the rule */
 447         union {
 448                 struct {
 449                         struct ibv_flow_attr *ibv_attr;
 450                         /**< Pointer to Verbs attributes. */
 451                         unsigned int offset;
 452                         /**< Current position or total size of the attribute. */
 453                 } queue[RTE_DIM(hash_rxq_init)];
 454                 struct {
 455                         struct ibv_flow_attr *ibv_attr;
 456                         /**< Pointer to Verbs attributes. */
 457                         unsigned int offset;
 458                         /**< Current position or total size of the attribute. */
 459                 } drop_q;
 460         };
 461 };
 462
 463 static const struct rte_flow_ops mlx5_flow_ops = {
 464         .validate = mlx5_flow_validate,
 465         .create = mlx5_flow_create,
 466         .destroy = mlx5_flow_destroy,
 467         .flush = mlx5_flow_flush,
 468 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
 469         .query = mlx5_flow_query,
 470 #else
 471         .query = NULL,
 472 #endif
 473         .isolate = mlx5_flow_isolate,
 474 };
 475
 476 /* Convert FDIR request to Generic flow. */
 477 struct mlx5_fdir {
 478         struct rte_flow_attr attr;
 479         struct rte_flow_action actions[2];
 480         struct rte_flow_item items[4];
 481         struct rte_flow_item_eth l2;
 482         union {
 483                 struct rte_flow_item_ipv4 ipv4;
 484                 struct rte_flow_item_ipv6 ipv6;
 485         } l3;
 486         union {
 487                 struct rte_flow_item_udp udp;
 488                 struct rte_flow_item_tcp tcp;
 489         } l4;
 490         struct rte_flow_action_queue queue;
 491 };
 492
 493 /* Verbs specification header. */
 494 struct ibv_spec_header {
 495         enum ibv_flow_spec_type type;
 496         uint16_t size;
 497 };
 498
 499 /**
 500  * Check support for a given item.
 501  *
 502  * @param item[in]
 503  *   Item specification.
 504  * @param mask[in]
 505  *   Bit-masks covering supported fields to compare with spec, last and mask in
 506  *   \item.
 507  * @param size
 508  *   Bit-Mask size in bytes.
 509  *
 510  * @return
 511  *   0 on success.
 512  */
 513 static int
 514 mlx5_flow_item_validate(const struct rte_flow_item *item,
 515                         const uint8_t *mask, unsigned int size)
 516 {
 517         int ret = 0;
 518
 519         if (!item->spec && (item->mask || item->last))
 520                 return -1;
 521         if (item->spec && !item->mask) {
 522                 unsigned int i;
 523                 const uint8_t *spec = item->spec;
 524
 525                 for (i = 0; i < size; ++i)
 526                         if ((spec[i] | mask[i]) != mask[i])
 527                                 return -1;
 528         }
 529         if (item->last && !item->mask) {
 530                 unsigned int i;
 531                 const uint8_t *spec = item->last;
 532
 533                 for (i = 0; i < size; ++i)
 534                         if ((spec[i] | mask[i]) != mask[i])
 535                                 return -1;
 536         }
 537         if (item->mask) {
 538                 unsigned int i;
 539                 const uint8_t *spec = item->mask;
 540
 541                 for (i = 0; i < size; ++i)
 542                         if ((spec[i] | mask[i]) != mask[i])
 543                                 return -1;
 544         }
 545         if (item->spec && item->last) {
 546                 uint8_t spec[size];
 547                 uint8_t last[size];
 548                 const uint8_t *apply = mask;
 549                 unsigned int i;
 550
 551                 if (item->mask)
 552                         apply = item->mask;
 553                 for (i = 0; i < size; ++i) {
 554                         spec[i] = ((const uint8_t *)item->spec)[i] & apply[i];
 555                         last[i] = ((const uint8_t *)item->last)[i] & apply[i];
 556                 }
 557                 ret = memcmp(spec, last, size);
 558         }
 559         return ret;
 560 }
 561
 562 /**
 563  * Copy the RSS configuration from the user ones.
 564  *
 565  * @param priv
 566  *   Pointer to private structure.
 567  * @param parser
 568  *   Internal parser structure.
 569  * @param rss_conf
 570  *   User RSS configuration to save.
 571  *
 572  * @return
 573  *   0 on success, errno value on failure.
 574  */
 575 static int
 576 priv_flow_convert_rss_conf(struct priv *priv,
 577                            struct mlx5_flow_parse *parser,
 578                            const struct rte_eth_rss_conf *rss_conf)
 579 {
 580         const struct rte_eth_rss_conf *rss =
 581                 rss_conf ? rss_conf : &priv->rss_conf;
 582
 583         if (rss->rss_key_len > 40)
 584                 return EINVAL;
 585         parser->rss_conf.rss_key_len = rss->rss_key_len;
 586         parser->rss_conf.rss_hf = rss->rss_hf;
 587         memcpy(parser->rss_key, rss->rss_key, rss->rss_key_len);
 588         parser->rss_conf.rss_key = parser->rss_key;
 589         return 0;
 590 }
 591
 592 /**
 593  * Extract attribute to the parser.
 594  *
 595  * @param priv
 596  *   Pointer to private structure.
 597  * @param[in] attr
 598  *   Flow rule attributes.
 599  * @param[out] error
 600  *   Perform verbose error reporting if not NULL.
 601  * @param[in, out] parser
 602  *   Internal parser structure.
 603  *
 604  * @return
 605  *   0 on success, a negative errno value otherwise and rte_errno is set.
 606  */
 607 static int
 608 priv_flow_convert_attributes(struct priv *priv,
 609                              const struct rte_flow_attr *attr,
 610                              struct rte_flow_error *error,
 611                              struct mlx5_flow_parse *parser)
 612 {
 613         (void)priv;
 614         (void)parser;
 615         if (attr->group) {
 616                 rte_flow_error_set(error, ENOTSUP,
 617                                    RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
 618                                    NULL,
 619                                    "groups are not supported");
 620                 return -rte_errno;
 621         }
 622         if (attr->priority && attr->priority != MLX5_CTRL_FLOW_PRIORITY) {
 623                 rte_flow_error_set(error, ENOTSUP,
 624                                    RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
 625                                    NULL,
 626                                    "priorities are not supported");
 627                 return -rte_errno;
 628         }
 629         if (attr->egress) {
 630                 rte_flow_error_set(error, ENOTSUP,
 631                                    RTE_FLOW_ERROR_TYPE_ATTR_EGRESS,
 632                                    NULL,
 633                                    "egress is not supported");
 634                 return -rte_errno;
 635         }
 636         if (!attr->ingress) {
 637                 rte_flow_error_set(error, ENOTSUP,
 638                                    RTE_FLOW_ERROR_TYPE_ATTR_INGRESS,
 639                                    NULL,
 640                                    "only ingress is supported");
 641                 return -rte_errno;
 642         }
 643         return 0;
 644 }
 645
 646 /**
 647  * Extract actions request to the parser.
 648  *
 649  * @param priv
 650  *   Pointer to private structure.
 651  * @param[in] actions
 652  *   Associated actions (list terminated by the END action).
 653  * @param[out] error
 654  *   Perform verbose error reporting if not NULL.
 655  * @param[in, out] parser
 656  *   Internal parser structure.
 657  *
 658  * @return
 659  *   0 on success, a negative errno value otherwise and rte_errno is set.
 660  */
 661 static int
 662 priv_flow_convert_actions(struct priv *priv,
 663                           const struct rte_flow_action actions[],
 664                           struct rte_flow_error *error,
 665                           struct mlx5_flow_parse *parser)
 666 {
 667         /*
 668          * Add default RSS configuration necessary for Verbs to create QP even
 669          * if no RSS is necessary.
 670          */
 671         priv_flow_convert_rss_conf(priv, parser,
 672                                    (const struct rte_eth_rss_conf *)
 673                                    &priv->rss_conf);
 674         for (; actions->type != RTE_FLOW_ACTION_TYPE_END; ++actions) {
 675                 if (actions->type == RTE_FLOW_ACTION_TYPE_VOID) {
 676                         continue;
 677                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_DROP) {
 678                         parser->drop = 1;
 679                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_QUEUE) {
 680                         const struct rte_flow_action_queue *queue =
 681                                 (const struct rte_flow_action_queue *)
 682                                 actions->conf;
 683                         uint16_t n;
 684                         uint16_t found = 0;
 685
 686                         if (!queue || (queue->index > (priv->rxqs_n - 1)))
 687                                 goto exit_action_not_supported;
 688                         for (n = 0; n < parser->queues_n; ++n) {
 689                                 if (parser->queues[n] == queue->index) {
 690                                         found = 1;
 691                                         break;
 692                                 }
 693                         }
 694                         if (parser->queues_n > 1 && !found) {
 695                                 rte_flow_error_set(error, ENOTSUP,
 696                                            RTE_FLOW_ERROR_TYPE_ACTION,
 697                                            actions,
 698                                            "queue action not in RSS queues");
 699                                 return -rte_errno;
 700                         }
 701                         if (!found) {
 702                                 parser->queues_n = 1;
 703                                 parser->queues[0] = queue->index;
 704                         }
 705                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_RSS) {
 706                         const struct rte_flow_action_rss *rss =
 707                                 (const struct rte_flow_action_rss *)
 708                                 actions->conf;
 709                         uint16_t n;
 710
 711                         if (!rss || !rss->num) {
 712                                 rte_flow_error_set(error, EINVAL,
 713                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 714                                                    actions,
 715                                                    "no valid queues");
 716                                 return -rte_errno;
 717                         }
 718                         if (parser->queues_n == 1) {
 719                                 uint16_t found = 0;
 720
 721                                 assert(parser->queues_n);
 722                                 for (n = 0; n < rss->num; ++n) {
 723                                         if (parser->queues[0] ==
 724                                             rss->queue[n]) {
 725                                                 found = 1;
 726                                                 break;
 727                                         }
 728                                 }
 729                                 if (!found) {
 730                                         rte_flow_error_set(error, ENOTSUP,
 731                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 732                                                    actions,
 733                                                    "queue action not in RSS"
 734                                                    " queues");
 735                                         return -rte_errno;
 736                                 }
 737                         }
 738                         for (n = 0; n < rss->num; ++n) {
 739                                 if (rss->queue[n] >= priv->rxqs_n) {
 740                                         rte_flow_error_set(error, EINVAL,
 741                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 742                                                    actions,
 743                                                    "queue id > number of"
 744                                                    " queues");
 745                                         return -rte_errno;
 746                                 }
 747                         }
 748                         for (n = 0; n < rss->num; ++n)
 749                                 parser->queues[n] = rss->queue[n];
 750                         parser->queues_n = rss->num;
 751                         if (priv_flow_convert_rss_conf(priv, parser,
 752                                                        rss->rss_conf)) {
 753                                 rte_flow_error_set(error, EINVAL,
 754                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 755                                                    actions,
 756                                                    "wrong RSS configuration");
 757                                 return -rte_errno;
 758                         }
 759                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_MARK) {
 760                         const struct rte_flow_action_mark *mark =
 761                                 (const struct rte_flow_action_mark *)
 762                                 actions->conf;
 763
 764                         if (!mark) {
 765                                 rte_flow_error_set(error, EINVAL,
 766                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 767                                                    actions,
 768                                                    "mark must be defined");
 769                                 return -rte_errno;
 770                         } else if (mark->id >= MLX5_FLOW_MARK_MAX) {
 771                                 rte_flow_error_set(error, ENOTSUP,
 772                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 773                                                    actions,
 774                                                    "mark must be between 0"
 775                                                    " and 16777199");
 776                                 return -rte_errno;
 777                         }
 778                         parser->mark = 1;
 779                         parser->mark_id = mark->id;
 780                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_FLAG) {
 781                         parser->mark = 1;
 782                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_COUNT &&
 783                            priv->counter_set_supported) {
 784                         parser->count = 1;
 785                 } else {
 786                         goto exit_action_not_supported;
 787                 }
 788         }
 789         if (parser->drop && parser->mark)
 790                 parser->mark = 0;
 791         if (!parser->queues_n && !parser->drop) {
 792                 rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_HANDLE,
 793                                    NULL, "no valid action");
 794                 return -rte_errno;
 795         }
 796         return 0;
 797 exit_action_not_supported:
 798         rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
 799                            actions, "action not supported");
 800         return -rte_errno;
 801 }
 802
 803 /**
 804  * Validate items.
 805  *
 806  * @param priv
 807  *   Pointer to private structure.
 808  * @param[in] items
 809  *   Pattern specification (list terminated by the END pattern item).
 810  * @param[out] error
 811  *   Perform verbose error reporting if not NULL.
 812  * @param[in, out] parser
 813  *   Internal parser structure.
 814  *
 815  * @return
 816  *   0 on success, a negative errno value otherwise and rte_errno is set.
 817  */
 818 static int
 819 priv_flow_convert_items_validate(struct priv *priv,
 820                                  const struct rte_flow_item items[],
 821                                  struct rte_flow_error *error,
 822                                  struct mlx5_flow_parse *parser)
 823 {
 824         const struct mlx5_flow_items *cur_item = mlx5_flow_items;
 825         unsigned int i;
 826
 827         (void)priv;
 828         /* Initialise the offsets to start after verbs attribute. */
 829         if (parser->drop) {
 830                 parser->drop_q.offset = sizeof(struct ibv_flow_attr);
 831         } else {
 832                 for (i = 0; i != hash_rxq_init_n; ++i)
 833                         parser->queue[i].offset = sizeof(struct ibv_flow_attr);
 834         }
 835         for (; items->type != RTE_FLOW_ITEM_TYPE_END; ++items) {
 836                 const struct mlx5_flow_items *token = NULL;
 837                 unsigned int n;
 838                 int err;
 839
 840                 if (items->type == RTE_FLOW_ITEM_TYPE_VOID)
 841                         continue;
 842                 for (i = 0;
 843                      cur_item->items &&
 844                      cur_item->items[i] != RTE_FLOW_ITEM_TYPE_END;
 845                      ++i) {
 846                         if (cur_item->items[i] == items->type) {
 847                                 token = &mlx5_flow_items[items->type];
 848                                 break;
 849                         }
 850                 }
 851                 if (!token)
 852                         goto exit_item_not_supported;
 853                 cur_item = token;
 854                 err = mlx5_flow_item_validate(items,
 855                                               (const uint8_t *)cur_item->mask,
 856                                               cur_item->mask_sz);
 857                 if (err)
 858                         goto exit_item_not_supported;
 859                 if (items->type == RTE_FLOW_ITEM_TYPE_VXLAN) {
 860                         if (parser->inner) {
 861                                 rte_flow_error_set(error, ENOTSUP,
 862                                                    RTE_FLOW_ERROR_TYPE_ITEM,
 863                                                    items,
 864                                                    "cannot recognize multiple"
 865                                                    " VXLAN encapsulations");
 866                                 return -rte_errno;
 867                         }
 868                         parser->inner = 1;
 869                 }
 870                 if (parser->drop) {
 871                         parser->drop_q.offset += cur_item->dst_sz;
 872                 } else if (parser->queues_n == 1) {
 873                         parser->queue[HASH_RXQ_ETH].offset += cur_item->dst_sz;
 874                 } else {
 875                         for (n = 0; n != hash_rxq_init_n; ++n)
 876                                 parser->queue[n].offset += cur_item->dst_sz;
 877                 }
 878         }
 879         if (parser->mark) {
 880                 for (i = 0; i != hash_rxq_init_n; ++i)
 881                         parser->queue[i].offset +=
 882                                 sizeof(struct ibv_flow_spec_action_tag);
 883         }
 884         if (parser->count) {
 885                 unsigned int size = sizeof(struct ibv_flow_spec_counter_action);
 886
 887                 if (parser->drop) {
 888                         parser->drop_q.offset += size;
 889                 } else {
 890                         for (i = 0; i != hash_rxq_init_n; ++i)
 891                                 parser->queue[i].offset += size;
 892                 }
 893         }
 894         return 0;
 895 exit_item_not_supported:
 896         rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 897                            items, "item not supported");
 898         return -rte_errno;
 899 }
 900
 901 /**
 902  * Allocate memory space to store verbs flow attributes.
 903  *
 904  * @param priv
 905  *   Pointer to private structure.
 906  * @param[in] priority
 907  *   Flow priority.
 908  * @param[in] size
 909  *   Amount of byte to allocate.
 910  * @param[out] error
 911  *   Perform verbose error reporting if not NULL.
 912  *
 913  * @return
 914  *   A verbs flow attribute on success, NULL otherwise.
 915  */
 916 static struct ibv_flow_attr*
 917 priv_flow_convert_allocate(struct priv *priv,
 918                            unsigned int priority,
 919                            unsigned int size,
 920                            struct rte_flow_error *error)
 921 {
 922         struct ibv_flow_attr *ibv_attr;
 923
 924         (void)priv;
 925         ibv_attr = rte_calloc(__func__, 1, size, 0);
 926         if (!ibv_attr) {
 927                 rte_flow_error_set(error, ENOMEM,
 928                                    RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
 929                                    NULL,
 930                                    "cannot allocate verbs spec attributes.");
 931                 return NULL;
 932         }
 933         ibv_attr->priority = priority;
 934         return ibv_attr;
 935 }
 936
 937 /**
 938  * Finalise verbs flow attributes.
 939  *
 940  * @param priv
 941  *   Pointer to private structure.
 942  * @param[in, out] parser
 943  *   Internal parser structure.
 944  */
 945 static void
 946 priv_flow_convert_finalise(struct priv *priv, struct mlx5_flow_parse *parser)
 947 {
 948         const unsigned int ipv4 =
 949                 hash_rxq_init[parser->layer].ip_version == MLX5_IPV4;
 950         const enum hash_rxq_type hmin = ipv4 ? HASH_RXQ_TCPV4 : HASH_RXQ_TCPV6;
 951         const enum hash_rxq_type hmax = ipv4 ? HASH_RXQ_IPV4 : HASH_RXQ_IPV6;
 952         const enum hash_rxq_type ohmin = ipv4 ? HASH_RXQ_TCPV6 : HASH_RXQ_TCPV4;
 953         const enum hash_rxq_type ohmax = ipv4 ? HASH_RXQ_IPV6 : HASH_RXQ_IPV4;
 954         const enum hash_rxq_type ip = ipv4 ? HASH_RXQ_IPV4 : HASH_RXQ_IPV6;
 955         unsigned int i;
 956
 957         (void)priv;
 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.rss_hf) {
 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->queues_n = 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                                                 parser->inner,
1014                                         .size = size,
1015                                 };
1016                         } else {
1017                                 size = sizeof(struct ibv_flow_spec_ipv6);
1018                                 specs.ipv6 = (struct ibv_flow_spec_ipv6){
1019                                         .type = IBV_FLOW_SPEC_IPV6 |
1020                                                 parser->inner,
1021                                         .size = size,
1022                                 };
1023                         }
1024                         if (parser->queue[i].ibv_attr) {
1025                                 dst = (void *)((uintptr_t)
1026                                                parser->queue[i].ibv_attr +
1027                                                parser->queue[i].offset);
1028                                 memcpy(dst, &specs, size);
1029                                 ++parser->queue[i].ibv_attr->num_of_specs;
1030                         }
1031                         parser->queue[i].offset += size;
1032                 }
1033                 if ((i == HASH_RXQ_UDPV4) || (i == HASH_RXQ_TCPV4) ||
1034                     (i == HASH_RXQ_UDPV6) || (i == HASH_RXQ_TCPV6)) {
1035                         size = sizeof(struct ibv_flow_spec_tcp_udp);
1036                         specs.udp_tcp = (struct ibv_flow_spec_tcp_udp) {
1037                                 .type = ((i == HASH_RXQ_UDPV4 ||
1038                                           i == HASH_RXQ_UDPV6) ?
1039                                          IBV_FLOW_SPEC_UDP :
1040                                          IBV_FLOW_SPEC_TCP) |
1041                                         parser->inner,
1042                                 .size = size,
1043                         };
1044                         if (parser->queue[i].ibv_attr) {
1045                                 dst = (void *)((uintptr_t)
1046                                                parser->queue[i].ibv_attr +
1047                                                parser->queue[i].offset);
1048                                 memcpy(dst, &specs, size);
1049                                 ++parser->queue[i].ibv_attr->num_of_specs;
1050                         }
1051                         parser->queue[i].offset += size;
1052                 }
1053         }
1054 }
1055
1056 /**
1057  * Validate and convert a flow supported by the NIC.
1058  *
1059  * @param priv
1060  *   Pointer to private structure.
1061  * @param[in] attr
1062  *   Flow rule attributes.
1063  * @param[in] pattern
1064  *   Pattern specification (list terminated by the END pattern item).
1065  * @param[in] actions
1066  *   Associated actions (list terminated by the END action).
1067  * @param[out] error
1068  *   Perform verbose error reporting if not NULL.
1069  * @param[in, out] parser
1070  *   Internal parser structure.
1071  *
1072  * @return
1073  *   0 on success, a negative errno value otherwise and rte_errno is set.
1074  */
1075 static int
1076 priv_flow_convert(struct priv *priv,
1077                   const struct rte_flow_attr *attr,
1078                   const struct rte_flow_item items[],
1079                   const struct rte_flow_action actions[],
1080                   struct rte_flow_error *error,
1081                   struct mlx5_flow_parse *parser)
1082 {
1083         const struct mlx5_flow_items *cur_item = mlx5_flow_items;
1084         unsigned int i;
1085         int ret;
1086
1087         /* First step. Validate the attributes, items and actions. */
1088         *parser = (struct mlx5_flow_parse){
1089                 .create = parser->create,
1090                 .layer = HASH_RXQ_ETH,
1091                 .mark_id = MLX5_FLOW_MARK_DEFAULT,
1092         };
1093         ret = priv_flow_convert_attributes(priv, attr, error, parser);
1094         if (ret)
1095                 return ret;
1096         ret = priv_flow_convert_actions(priv, actions, error, parser);
1097         if (ret)
1098                 return ret;
1099         ret = priv_flow_convert_items_validate(priv, items, error, parser);
1100         if (ret)
1101                 return ret;
1102         priv_flow_convert_finalise(priv, parser);
1103         /*
1104          * Second step.
1105          * Allocate the memory space to store verbs specifications.
1106          */
1107         if (parser->drop) {
1108                 parser->drop_q.ibv_attr =
1109                         priv_flow_convert_allocate(priv, attr->priority,
1110                                                    parser->drop_q.offset,
1111                                                    error);
1112                 if (!parser->drop_q.ibv_attr)
1113                         return ENOMEM;
1114                 parser->drop_q.offset = sizeof(struct ibv_flow_attr);
1115         } else if (parser->queues_n == 1) {
1116                 unsigned int priority =
1117                         attr->priority +
1118                         hash_rxq_init[HASH_RXQ_ETH].flow_priority;
1119                 unsigned int offset = parser->queue[HASH_RXQ_ETH].offset;
1120
1121                 parser->queue[HASH_RXQ_ETH].ibv_attr =
1122                         priv_flow_convert_allocate(priv, priority,
1123                                                    offset, error);
1124                 if (!parser->queue[HASH_RXQ_ETH].ibv_attr)
1125                         return ENOMEM;
1126                 parser->queue[HASH_RXQ_ETH].offset =
1127                         sizeof(struct ibv_flow_attr);
1128         } else {
1129                 for (i = 0; i != hash_rxq_init_n; ++i) {
1130                         unsigned int priority =
1131                                 attr->priority +
1132                                 hash_rxq_init[i].flow_priority;
1133                         unsigned int offset;
1134
1135                         if (!(parser->rss_conf.rss_hf &
1136                               hash_rxq_init[i].dpdk_rss_hf) &&
1137                             (i != HASH_RXQ_ETH))
1138                                 continue;
1139                         offset = parser->queue[i].offset;
1140                         parser->queue[i].ibv_attr =
1141                                 priv_flow_convert_allocate(priv, priority,
1142                                                            offset, error);
1143                         if (!parser->queue[i].ibv_attr)
1144                                 goto exit_enomem;
1145                         parser->queue[i].offset = sizeof(struct ibv_flow_attr);
1146                 }
1147         }
1148         /* Third step. Conversion parse, fill the specifications. */
1149         parser->inner = 0;
1150         for (; items->type != RTE_FLOW_ITEM_TYPE_END; ++items) {
1151                 if (items->type == RTE_FLOW_ITEM_TYPE_VOID)
1152                         continue;
1153                 cur_item = &mlx5_flow_items[items->type];
1154                 ret = cur_item->convert(items,
1155                                         (cur_item->default_mask ?
1156                                          cur_item->default_mask :
1157                                          cur_item->mask),
1158                                         parser);
1159                 if (ret) {
1160                         rte_flow_error_set(error, ENOTSUP,
1161                                            RTE_FLOW_ERROR_TYPE_ITEM,
1162                                            items, "item not supported");
1163                         goto exit_free;
1164                 }
1165         }
1166         if (parser->mark)
1167                 mlx5_flow_create_flag_mark(parser, parser->mark_id);
1168         if (parser->count && parser->create) {
1169                 mlx5_flow_create_count(priv, parser);
1170                 if (!parser->cs)
1171                         goto exit_count_error;
1172         }
1173         /*
1174          * Last step. Complete missing specification to reach the RSS
1175          * configuration.
1176          */
1177         if (parser->drop) {
1178                 /*
1179                  * Drop queue priority needs to be adjusted to
1180                  * their most specific layer priority.
1181                  */
1182                 parser->drop_q.ibv_attr->priority =
1183                         attr->priority +
1184                         hash_rxq_init[parser->layer].flow_priority;
1185         } else if (parser->queues_n > 1) {
1186                 priv_flow_convert_finalise(priv, parser);
1187         } else {
1188                 /*
1189                  * Action queue have their priority overridden with
1190                  * Ethernet priority, this priority needs to be adjusted to
1191                  * their most specific layer priority.
1192                  */
1193                 parser->queue[HASH_RXQ_ETH].ibv_attr->priority =
1194                         attr->priority +
1195                         hash_rxq_init[parser->layer].flow_priority;
1196         }
1197 exit_free:
1198         /* Only verification is expected, all resources should be released. */
1199         if (!parser->create) {
1200                 if (parser->drop) {
1201                         rte_free(parser->drop_q.ibv_attr);
1202                         parser->drop_q.ibv_attr = NULL;
1203                 }
1204                 for (i = 0; i != hash_rxq_init_n; ++i) {
1205                         if (parser->queue[i].ibv_attr) {
1206                                 rte_free(parser->queue[i].ibv_attr);
1207                                 parser->queue[i].ibv_attr = NULL;
1208                         }
1209                 }
1210         }
1211         return ret;
1212 exit_enomem:
1213         for (i = 0; i != hash_rxq_init_n; ++i) {
1214                 if (parser->queue[i].ibv_attr) {
1215                         rte_free(parser->queue[i].ibv_attr);
1216                         parser->queue[i].ibv_attr = NULL;
1217                 }
1218         }
1219         rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
1220                            NULL, "cannot allocate verbs spec attributes.");
1221         return ret;
1222 exit_count_error:
1223         rte_flow_error_set(error, EINVAL, RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
1224                            NULL, "cannot create counter.");
1225         return rte_errno;
1226 }
1227
1228 /**
1229  * Copy the specification created into the flow.
1230  *
1231  * @param parser
1232  *   Internal parser structure.
1233  * @param src
1234  *   Create specification.
1235  * @param size
1236  *   Size in bytes of the specification to copy.
1237  */
1238 static void
1239 mlx5_flow_create_copy(struct mlx5_flow_parse *parser, void *src,
1240                       unsigned int size)
1241 {
1242         unsigned int i;
1243         void *dst;
1244
1245         if (parser->drop) {
1246                 dst = (void *)((uintptr_t)parser->drop_q.ibv_attr +
1247                                 parser->drop_q.offset);
1248                 memcpy(dst, src, size);
1249                 ++parser->drop_q.ibv_attr->num_of_specs;
1250                 parser->drop_q.offset += size;
1251                 return;
1252         }
1253         for (i = 0; i != hash_rxq_init_n; ++i) {
1254                 if (!parser->queue[i].ibv_attr)
1255                         continue;
1256                 /* Specification must be the same l3 type or none. */
1257                 if (parser->layer == HASH_RXQ_ETH ||
1258                     (hash_rxq_init[parser->layer].ip_version ==
1259                      hash_rxq_init[i].ip_version) ||
1260                     (hash_rxq_init[i].ip_version == 0)) {
1261                         dst = (void *)((uintptr_t)parser->queue[i].ibv_attr +
1262                                         parser->queue[i].offset);
1263                         memcpy(dst, src, size);
1264                         ++parser->queue[i].ibv_attr->num_of_specs;
1265                         parser->queue[i].offset += size;
1266                 }
1267         }
1268 }
1269
1270 /**
1271  * Convert Ethernet item to Verbs specification.
1272  *
1273  * @param item[in]
1274  *   Item specification.
1275  * @param default_mask[in]
1276  *   Default bit-masks to use when item->mask is not provided.
1277  * @param data[in, out]
1278  *   User structure.
1279  */
1280 static int
1281 mlx5_flow_create_eth(const struct rte_flow_item *item,
1282                      const void *default_mask,
1283                      void *data)
1284 {
1285         const struct rte_flow_item_eth *spec = item->spec;
1286         const struct rte_flow_item_eth *mask = item->mask;
1287         struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data;
1288         const unsigned int eth_size = sizeof(struct ibv_flow_spec_eth);
1289         struct ibv_flow_spec_eth eth = {
1290                 .type = parser->inner | IBV_FLOW_SPEC_ETH,
1291                 .size = eth_size,
1292         };
1293
1294         parser->layer = HASH_RXQ_ETH;
1295         if (spec) {
1296                 unsigned int i;
1297
1298                 if (!mask)
1299                         mask = default_mask;
1300                 memcpy(&eth.val.dst_mac, spec->dst.addr_bytes, ETHER_ADDR_LEN);
1301                 memcpy(&eth.val.src_mac, spec->src.addr_bytes, ETHER_ADDR_LEN);
1302                 eth.val.ether_type = spec->type;
1303                 memcpy(&eth.mask.dst_mac, mask->dst.addr_bytes, ETHER_ADDR_LEN);
1304                 memcpy(&eth.mask.src_mac, mask->src.addr_bytes, ETHER_ADDR_LEN);
1305                 eth.mask.ether_type = mask->type;
1306                 /* Remove unwanted bits from values. */
1307                 for (i = 0; i < ETHER_ADDR_LEN; ++i) {
1308                         eth.val.dst_mac[i] &= eth.mask.dst_mac[i];
1309                         eth.val.src_mac[i] &= eth.mask.src_mac[i];
1310                 }
1311                 eth.val.ether_type &= eth.mask.ether_type;
1312         }
1313         mlx5_flow_create_copy(parser, &eth, eth_size);
1314         return 0;
1315 }
1316
1317 /**
1318  * Convert VLAN item to Verbs specification.
1319  *
1320  * @param item[in]
1321  *   Item specification.
1322  * @param default_mask[in]
1323  *   Default bit-masks to use when item->mask is not provided.
1324  * @param data[in, out]
1325  *   User structure.
1326  */
1327 static int
1328 mlx5_flow_create_vlan(const struct rte_flow_item *item,
1329                       const void *default_mask,
1330                       void *data)
1331 {
1332         const struct rte_flow_item_vlan *spec = item->spec;
1333         const struct rte_flow_item_vlan *mask = item->mask;
1334         struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data;
1335         struct ibv_flow_spec_eth *eth;
1336         const unsigned int eth_size = sizeof(struct ibv_flow_spec_eth);
1337
1338         if (spec) {
1339                 unsigned int i;
1340                 if (!mask)
1341                         mask = default_mask;
1342
1343                 if (parser->drop) {
1344                         eth = (void *)((uintptr_t)parser->drop_q.ibv_attr +
1345                                        parser->drop_q.offset - eth_size);
1346                         eth->val.vlan_tag = spec->tci;
1347                         eth->mask.vlan_tag = mask->tci;
1348                         eth->val.vlan_tag &= eth->mask.vlan_tag;
1349                         return 0;
1350                 }
1351                 for (i = 0; i != hash_rxq_init_n; ++i) {
1352                         if (!parser->queue[i].ibv_attr)
1353                                 continue;
1354
1355                         eth = (void *)((uintptr_t)parser->queue[i].ibv_attr +
1356                                        parser->queue[i].offset - eth_size);
1357                         eth->val.vlan_tag = spec->tci;
1358                         eth->mask.vlan_tag = mask->tci;
1359                         eth->val.vlan_tag &= eth->mask.vlan_tag;
1360                 }
1361         }
1362         return 0;
1363 }
1364
1365 /**
1366  * Convert IPv4 item to Verbs specification.
1367  *
1368  * @param item[in]
1369  *   Item specification.
1370  * @param default_mask[in]
1371  *   Default bit-masks to use when item->mask is not provided.
1372  * @param data[in, out]
1373  *   User structure.
1374  */
1375 static int
1376 mlx5_flow_create_ipv4(const struct rte_flow_item *item,
1377                       const void *default_mask,
1378                       void *data)
1379 {
1380         const struct rte_flow_item_ipv4 *spec = item->spec;
1381         const struct rte_flow_item_ipv4 *mask = item->mask;
1382         struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data;
1383         unsigned int ipv4_size = sizeof(struct ibv_flow_spec_ipv4_ext);
1384         struct ibv_flow_spec_ipv4_ext ipv4 = {
1385                 .type = parser->inner | IBV_FLOW_SPEC_IPV4_EXT,
1386                 .size = ipv4_size,
1387         };
1388
1389         parser->layer = HASH_RXQ_IPV4;
1390         if (spec) {
1391                 if (!mask)
1392                         mask = default_mask;
1393                 ipv4.val = (struct ibv_flow_ipv4_ext_filter){
1394                         .src_ip = spec->hdr.src_addr,
1395                         .dst_ip = spec->hdr.dst_addr,
1396                         .proto = spec->hdr.next_proto_id,
1397                         .tos = spec->hdr.type_of_service,
1398                 };
1399                 ipv4.mask = (struct ibv_flow_ipv4_ext_filter){
1400                         .src_ip = mask->hdr.src_addr,
1401                         .dst_ip = mask->hdr.dst_addr,
1402                         .proto = mask->hdr.next_proto_id,
1403                         .tos = mask->hdr.type_of_service,
1404                 };
1405                 /* Remove unwanted bits from values. */
1406                 ipv4.val.src_ip &= ipv4.mask.src_ip;
1407                 ipv4.val.dst_ip &= ipv4.mask.dst_ip;
1408                 ipv4.val.proto &= ipv4.mask.proto;
1409                 ipv4.val.tos &= ipv4.mask.tos;
1410         }
1411         mlx5_flow_create_copy(parser, &ipv4, ipv4_size);
1412         return 0;
1413 }
1414
1415 /**
1416  * Convert IPv6 item to Verbs specification.
1417  *
1418  * @param item[in]
1419  *   Item specification.
1420  * @param default_mask[in]
1421  *   Default bit-masks to use when item->mask is not provided.
1422  * @param data[in, out]
1423  *   User structure.
1424  */
1425 static int
1426 mlx5_flow_create_ipv6(const struct rte_flow_item *item,
1427                       const void *default_mask,
1428                       void *data)
1429 {
1430         const struct rte_flow_item_ipv6 *spec = item->spec;
1431         const struct rte_flow_item_ipv6 *mask = item->mask;
1432         struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data;
1433         unsigned int ipv6_size = sizeof(struct ibv_flow_spec_ipv6);
1434         struct ibv_flow_spec_ipv6 ipv6 = {
1435                 .type = parser->inner | IBV_FLOW_SPEC_IPV6,
1436                 .size = ipv6_size,
1437         };
1438
1439         parser->layer = HASH_RXQ_IPV6;
1440         if (spec) {
1441                 unsigned int i;
1442
1443                 if (!mask)
1444                         mask = default_mask;
1445                 memcpy(&ipv6.val.src_ip, spec->hdr.src_addr,
1446                        RTE_DIM(ipv6.val.src_ip));
1447                 memcpy(&ipv6.val.dst_ip, spec->hdr.dst_addr,
1448                        RTE_DIM(ipv6.val.dst_ip));
1449                 memcpy(&ipv6.mask.src_ip, mask->hdr.src_addr,
1450                        RTE_DIM(ipv6.mask.src_ip));
1451                 memcpy(&ipv6.mask.dst_ip, mask->hdr.dst_addr,
1452                        RTE_DIM(ipv6.mask.dst_ip));
1453                 ipv6.mask.flow_label = mask->hdr.vtc_flow;
1454                 ipv6.mask.next_hdr = mask->hdr.proto;
1455                 ipv6.mask.hop_limit = mask->hdr.hop_limits;
1456                 /* Remove unwanted bits from values. */
1457                 for (i = 0; i < RTE_DIM(ipv6.val.src_ip); ++i) {
1458                         ipv6.val.src_ip[i] &= ipv6.mask.src_ip[i];
1459                         ipv6.val.dst_ip[i] &= ipv6.mask.dst_ip[i];
1460                 }
1461                 ipv6.val.flow_label &= ipv6.mask.flow_label;
1462                 ipv6.val.next_hdr &= ipv6.mask.next_hdr;
1463                 ipv6.val.hop_limit &= ipv6.mask.hop_limit;
1464         }
1465         mlx5_flow_create_copy(parser, &ipv6, ipv6_size);
1466         return 0;
1467 }
1468
1469 /**
1470  * Convert UDP item to Verbs specification.
1471  *
1472  * @param item[in]
1473  *   Item specification.
1474  * @param default_mask[in]
1475  *   Default bit-masks to use when item->mask is not provided.
1476  * @param data[in, out]
1477  *   User structure.
1478  */
1479 static int
1480 mlx5_flow_create_udp(const struct rte_flow_item *item,
1481                      const void *default_mask,
1482                      void *data)
1483 {
1484         const struct rte_flow_item_udp *spec = item->spec;
1485         const struct rte_flow_item_udp *mask = item->mask;
1486         struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data;
1487         unsigned int udp_size = sizeof(struct ibv_flow_spec_tcp_udp);
1488         struct ibv_flow_spec_tcp_udp udp = {
1489                 .type = parser->inner | IBV_FLOW_SPEC_UDP,
1490                 .size = udp_size,
1491         };
1492
1493         if (parser->layer == HASH_RXQ_IPV4)
1494                 parser->layer = HASH_RXQ_UDPV4;
1495         else
1496                 parser->layer = HASH_RXQ_UDPV6;
1497         if (spec) {
1498                 if (!mask)
1499                         mask = default_mask;
1500                 udp.val.dst_port = spec->hdr.dst_port;
1501                 udp.val.src_port = spec->hdr.src_port;
1502                 udp.mask.dst_port = mask->hdr.dst_port;
1503                 udp.mask.src_port = mask->hdr.src_port;
1504                 /* Remove unwanted bits from values. */
1505                 udp.val.src_port &= udp.mask.src_port;
1506                 udp.val.dst_port &= udp.mask.dst_port;
1507         }
1508         mlx5_flow_create_copy(parser, &udp, udp_size);
1509         return 0;
1510 }
1511
1512 /**
1513  * Convert TCP item to Verbs specification.
1514  *
1515  * @param item[in]
1516  *   Item specification.
1517  * @param default_mask[in]
1518  *   Default bit-masks to use when item->mask is not provided.
1519  * @param data[in, out]
1520  *   User structure.
1521  */
1522 static int
1523 mlx5_flow_create_tcp(const struct rte_flow_item *item,
1524                      const void *default_mask,
1525                      void *data)
1526 {
1527         const struct rte_flow_item_tcp *spec = item->spec;
1528         const struct rte_flow_item_tcp *mask = item->mask;
1529         struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data;
1530         unsigned int tcp_size = sizeof(struct ibv_flow_spec_tcp_udp);
1531         struct ibv_flow_spec_tcp_udp tcp = {
1532                 .type = parser->inner | IBV_FLOW_SPEC_TCP,
1533                 .size = tcp_size,
1534         };
1535
1536         if (parser->layer == HASH_RXQ_IPV4)
1537                 parser->layer = HASH_RXQ_TCPV4;
1538         else
1539                 parser->layer = HASH_RXQ_TCPV6;
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 static int
1566 mlx5_flow_create_vxlan(const struct rte_flow_item *item,
1567                        const void *default_mask,
1568                        void *data)
1569 {
1570         const struct rte_flow_item_vxlan *spec = item->spec;
1571         const struct rte_flow_item_vxlan *mask = item->mask;
1572         struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data;
1573         unsigned int size = sizeof(struct ibv_flow_spec_tunnel);
1574         struct ibv_flow_spec_tunnel vxlan = {
1575                 .type = parser->inner | IBV_FLOW_SPEC_VXLAN_TUNNEL,
1576                 .size = size,
1577         };
1578         union vni {
1579                 uint32_t vlan_id;
1580                 uint8_t vni[4];
1581         } id;
1582
1583         id.vni[0] = 0;
1584         parser->inner = IBV_FLOW_SPEC_INNER;
1585         if (spec) {
1586                 if (!mask)
1587                         mask = default_mask;
1588                 memcpy(&id.vni[1], spec->vni, 3);
1589                 vxlan.val.tunnel_id = id.vlan_id;
1590                 memcpy(&id.vni[1], mask->vni, 3);
1591                 vxlan.mask.tunnel_id = id.vlan_id;
1592                 /* Remove unwanted bits from values. */
1593                 vxlan.val.tunnel_id &= vxlan.mask.tunnel_id;
1594         }
1595         mlx5_flow_create_copy(parser, &vxlan, size);
1596         return 0;
1597 }
1598
1599 /**
1600  * Convert mark/flag action to Verbs specification.
1601  *
1602  * @param parser
1603  *   Internal parser structure.
1604  * @param mark_id
1605  *   Mark identifier.
1606  */
1607 static int
1608 mlx5_flow_create_flag_mark(struct mlx5_flow_parse *parser, uint32_t mark_id)
1609 {
1610         unsigned int size = sizeof(struct ibv_flow_spec_action_tag);
1611         struct ibv_flow_spec_action_tag tag = {
1612                 .type = IBV_FLOW_SPEC_ACTION_TAG,
1613                 .size = size,
1614                 .tag_id = mlx5_flow_mark_set(mark_id),
1615         };
1616
1617         assert(parser->mark);
1618         mlx5_flow_create_copy(parser, &tag, size);
1619         return 0;
1620 }
1621
1622 /**
1623  * Convert count action to Verbs specification.
1624  *
1625  * @param priv
1626  *   Pointer to private structure.
1627  * @param parser
1628  *   Pointer to MLX5 flow parser structure.
1629  *
1630  * @return
1631  *   0 on success, errno value on failure.
1632  */
1633 static int
1634 mlx5_flow_create_count(struct priv *priv __rte_unused,
1635                        struct mlx5_flow_parse *parser __rte_unused)
1636 {
1637 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
1638         unsigned int size = sizeof(struct ibv_flow_spec_counter_action);
1639         struct ibv_counter_set_init_attr init_attr = {0};
1640         struct ibv_flow_spec_counter_action counter = {
1641                 .type = IBV_FLOW_SPEC_ACTION_COUNT,
1642                 .size = size,
1643                 .counter_set_handle = 0,
1644         };
1645
1646         init_attr.counter_set_id = 0;
1647         parser->cs = ibv_create_counter_set(priv->ctx, &init_attr);
1648         if (!parser->cs)
1649                 return EINVAL;
1650         counter.counter_set_handle = parser->cs->handle;
1651         mlx5_flow_create_copy(parser, &counter, size);
1652 #endif
1653         return 0;
1654 }
1655
1656 /**
1657  * Complete flow rule creation with a drop queue.
1658  *
1659  * @param priv
1660  *   Pointer to private structure.
1661  * @param parser
1662  *   Internal parser structure.
1663  * @param flow
1664  *   Pointer to the rte_flow.
1665  * @param[out] error
1666  *   Perform verbose error reporting if not NULL.
1667  *
1668  * @return
1669  *   0 on success, errno value on failure.
1670  */
1671 static int
1672 priv_flow_create_action_queue_drop(struct priv *priv,
1673                                    struct mlx5_flow_parse *parser,
1674                                    struct rte_flow *flow,
1675                                    struct rte_flow_error *error)
1676 {
1677         struct ibv_flow_spec_action_drop *drop;
1678         unsigned int size = sizeof(struct ibv_flow_spec_action_drop);
1679         int err = 0;
1680
1681         assert(priv->pd);
1682         assert(priv->ctx);
1683         flow->drop = 1;
1684         drop = (void *)((uintptr_t)parser->drop_q.ibv_attr +
1685                         parser->drop_q.offset);
1686         *drop = (struct ibv_flow_spec_action_drop){
1687                         .type = IBV_FLOW_SPEC_ACTION_DROP,
1688                         .size = size,
1689         };
1690         ++parser->drop_q.ibv_attr->num_of_specs;
1691         parser->drop_q.offset += size;
1692         flow->drxq.ibv_attr = parser->drop_q.ibv_attr;
1693         if (!priv->dev->data->dev_started)
1694                 return 0;
1695         parser->drop_q.ibv_attr = NULL;
1696         flow->drxq.ibv_flow = ibv_create_flow(priv->flow_drop_queue->qp,
1697                                               flow->drxq.ibv_attr);
1698         if (parser->count)
1699                 flow->cs = parser->cs;
1700         if (!flow->drxq.ibv_flow) {
1701                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1702                                    NULL, "flow rule creation failure");
1703                 err = ENOMEM;
1704                 goto error;
1705         }
1706         return 0;
1707 error:
1708         assert(flow);
1709         if (flow->drxq.ibv_flow) {
1710                 claim_zero(ibv_destroy_flow(flow->drxq.ibv_flow));
1711                 flow->drxq.ibv_flow = NULL;
1712         }
1713         if (flow->drxq.ibv_attr) {
1714                 rte_free(flow->drxq.ibv_attr);
1715                 flow->drxq.ibv_attr = NULL;
1716         }
1717         if (flow->cs) {
1718                 claim_zero(ibv_destroy_counter_set(flow->cs));
1719                 flow->cs = NULL;
1720                 parser->cs = NULL;
1721         }
1722         return err;
1723 }
1724
1725 /**
1726  * Create hash Rx queues when RSS is enabled.
1727  *
1728  * @param priv
1729  *   Pointer to private structure.
1730  * @param parser
1731  *   Internal parser structure.
1732  * @param flow
1733  *   Pointer to the rte_flow.
1734  * @param[out] error
1735  *   Perform verbose error reporting if not NULL.
1736  *
1737  * @return
1738  *   0 on success, a errno value otherwise and rte_errno is set.
1739  */
1740 static int
1741 priv_flow_create_action_queue_rss(struct priv *priv,
1742                                   struct mlx5_flow_parse *parser,
1743                                   struct rte_flow *flow,
1744                                   struct rte_flow_error *error)
1745 {
1746         unsigned int i;
1747
1748         for (i = 0; i != hash_rxq_init_n; ++i) {
1749                 uint64_t hash_fields;
1750
1751                 if (!parser->queue[i].ibv_attr)
1752                         continue;
1753                 flow->frxq[i].ibv_attr = parser->queue[i].ibv_attr;
1754                 parser->queue[i].ibv_attr = NULL;
1755                 hash_fields = hash_rxq_init[i].hash_fields;
1756                 if (!priv->dev->data->dev_started)
1757                         continue;
1758                 flow->frxq[i].hrxq =
1759                         mlx5_priv_hrxq_get(priv,
1760                                            parser->rss_conf.rss_key,
1761                                            parser->rss_conf.rss_key_len,
1762                                            hash_fields,
1763                                            parser->queues,
1764                                            hash_fields ? parser->queues_n : 1);
1765                 if (flow->frxq[i].hrxq)
1766                         continue;
1767                 flow->frxq[i].hrxq =
1768                         mlx5_priv_hrxq_new(priv,
1769                                            parser->rss_conf.rss_key,
1770                                            parser->rss_conf.rss_key_len,
1771                                            hash_fields,
1772                                            parser->queues,
1773                                            hash_fields ? parser->queues_n : 1);
1774                 if (!flow->frxq[i].hrxq) {
1775                         rte_flow_error_set(error, ENOMEM,
1776                                            RTE_FLOW_ERROR_TYPE_HANDLE,
1777                                            NULL, "cannot create hash rxq");
1778                         return ENOMEM;
1779                 }
1780         }
1781         return 0;
1782 }
1783
1784 /**
1785  * Complete flow rule creation.
1786  *
1787  * @param priv
1788  *   Pointer to private structure.
1789  * @param parser
1790  *   Internal parser structure.
1791  * @param flow
1792  *   Pointer to the rte_flow.
1793  * @param[out] error
1794  *   Perform verbose error reporting if not NULL.
1795  *
1796  * @return
1797  *   0 on success, a errno value otherwise and rte_errno is set.
1798  */
1799 static int
1800 priv_flow_create_action_queue(struct priv *priv,
1801                               struct mlx5_flow_parse *parser,
1802                               struct rte_flow *flow,
1803                               struct rte_flow_error *error)
1804 {
1805         int err = 0;
1806         unsigned int i;
1807
1808         assert(priv->pd);
1809         assert(priv->ctx);
1810         assert(!parser->drop);
1811         err = priv_flow_create_action_queue_rss(priv, parser, flow, error);
1812         if (err)
1813                 goto error;
1814         if (parser->count)
1815                 flow->cs = parser->cs;
1816         if (!priv->dev->data->dev_started)
1817                 return 0;
1818         for (i = 0; i != hash_rxq_init_n; ++i) {
1819                 if (!flow->frxq[i].hrxq)
1820                         continue;
1821                 flow->frxq[i].ibv_flow =
1822                         ibv_create_flow(flow->frxq[i].hrxq->qp,
1823                                         flow->frxq[i].ibv_attr);
1824                 if (!flow->frxq[i].ibv_flow) {
1825                         rte_flow_error_set(error, ENOMEM,
1826                                            RTE_FLOW_ERROR_TYPE_HANDLE,
1827                                            NULL, "flow rule creation failure");
1828                         err = ENOMEM;
1829                         goto error;
1830                 }
1831                 DEBUG("%p type %d QP %p ibv_flow %p",
1832                       (void *)flow, i,
1833                       (void *)flow->frxq[i].hrxq,
1834                       (void *)flow->frxq[i].ibv_flow);
1835         }
1836         for (i = 0; i != parser->queues_n; ++i) {
1837                 struct mlx5_rxq_data *q =
1838                         (*priv->rxqs)[parser->queues[i]];
1839
1840                 q->mark |= parser->mark;
1841         }
1842         return 0;
1843 error:
1844         assert(flow);
1845         for (i = 0; i != hash_rxq_init_n; ++i) {
1846                 if (flow->frxq[i].ibv_flow) {
1847                         struct ibv_flow *ibv_flow = flow->frxq[i].ibv_flow;
1848
1849                         claim_zero(ibv_destroy_flow(ibv_flow));
1850                 }
1851                 if (flow->frxq[i].hrxq)
1852                         mlx5_priv_hrxq_release(priv, flow->frxq[i].hrxq);
1853                 if (flow->frxq[i].ibv_attr)
1854                         rte_free(flow->frxq[i].ibv_attr);
1855         }
1856         if (flow->cs) {
1857                 claim_zero(ibv_destroy_counter_set(flow->cs));
1858                 flow->cs = NULL;
1859                 parser->cs = NULL;
1860         }
1861         return err;
1862 }
1863
1864 /**
1865  * Convert a flow.
1866  *
1867  * @param priv
1868  *   Pointer to private structure.
1869  * @param list
1870  *   Pointer to a TAILQ flow list.
1871  * @param[in] attr
1872  *   Flow rule attributes.
1873  * @param[in] pattern
1874  *   Pattern specification (list terminated by the END pattern item).
1875  * @param[in] actions
1876  *   Associated actions (list terminated by the END action).
1877  * @param[out] error
1878  *   Perform verbose error reporting if not NULL.
1879  *
1880  * @return
1881  *   A flow on success, NULL otherwise.
1882  */
1883 static struct rte_flow *
1884 priv_flow_create(struct priv *priv,
1885                  struct mlx5_flows *list,
1886                  const struct rte_flow_attr *attr,
1887                  const struct rte_flow_item items[],
1888                  const struct rte_flow_action actions[],
1889                  struct rte_flow_error *error)
1890 {
1891         struct mlx5_flow_parse parser = { .create = 1, };
1892         struct rte_flow *flow = NULL;
1893         unsigned int i;
1894         int err;
1895
1896         err = priv_flow_convert(priv, attr, items, actions, error, &parser);
1897         if (err)
1898                 goto exit;
1899         flow = rte_calloc(__func__, 1,
1900                           sizeof(*flow) + parser.queues_n * sizeof(uint16_t),
1901                           0);
1902         if (!flow) {
1903                 rte_flow_error_set(error, ENOMEM,
1904                                    RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
1905                                    NULL,
1906                                    "cannot allocate flow memory");
1907                 return NULL;
1908         }
1909         /* Copy queues configuration. */
1910         flow->queues = (uint16_t (*)[])(flow + 1);
1911         memcpy(flow->queues, parser.queues, parser.queues_n * sizeof(uint16_t));
1912         flow->queues_n = parser.queues_n;
1913         /* Copy RSS configuration. */
1914         flow->rss_conf = parser.rss_conf;
1915         flow->rss_conf.rss_key = flow->rss_key;
1916         memcpy(flow->rss_key, parser.rss_key, parser.rss_conf.rss_key_len);
1917         /* finalise the flow. */
1918         if (parser.drop)
1919                 err = priv_flow_create_action_queue_drop(priv, &parser, flow,
1920                                                          error);
1921         else
1922                 err = priv_flow_create_action_queue(priv, &parser, flow, error);
1923         if (err)
1924                 goto exit;
1925         TAILQ_INSERT_TAIL(list, flow, next);
1926         DEBUG("Flow created %p", (void *)flow);
1927         return flow;
1928 exit:
1929         if (parser.drop) {
1930                 rte_free(parser.drop_q.ibv_attr);
1931         } else {
1932                 for (i = 0; i != hash_rxq_init_n; ++i) {
1933                         if (parser.queue[i].ibv_attr)
1934                                 rte_free(parser.queue[i].ibv_attr);
1935                 }
1936         }
1937         rte_free(flow);
1938         return NULL;
1939 }
1940
1941 /**
1942  * Validate a flow supported by the NIC.
1943  *
1944  * @see rte_flow_validate()
1945  * @see rte_flow_ops
1946  */
1947 int
1948 mlx5_flow_validate(struct rte_eth_dev *dev,
1949                    const struct rte_flow_attr *attr,
1950                    const struct rte_flow_item items[],
1951                    const struct rte_flow_action actions[],
1952                    struct rte_flow_error *error)
1953 {
1954         struct priv *priv = dev->data->dev_private;
1955         int ret;
1956         struct mlx5_flow_parse parser = { .create = 0, };
1957
1958         priv_lock(priv);
1959         ret = priv_flow_convert(priv, attr, items, actions, error, &parser);
1960         priv_unlock(priv);
1961         return ret;
1962 }
1963
1964 /**
1965  * Create a flow.
1966  *
1967  * @see rte_flow_create()
1968  * @see rte_flow_ops
1969  */
1970 struct rte_flow *
1971 mlx5_flow_create(struct rte_eth_dev *dev,
1972                  const struct rte_flow_attr *attr,
1973                  const struct rte_flow_item items[],
1974                  const struct rte_flow_action actions[],
1975                  struct rte_flow_error *error)
1976 {
1977         struct priv *priv = dev->data->dev_private;
1978         struct rte_flow *flow;
1979
1980         priv_lock(priv);
1981         flow = priv_flow_create(priv, &priv->flows, attr, items, actions,
1982                                 error);
1983         priv_unlock(priv);
1984         return flow;
1985 }
1986
1987 /**
1988  * Destroy a flow.
1989  *
1990  * @param priv
1991  *   Pointer to private structure.
1992  * @param list
1993  *   Pointer to a TAILQ flow list.
1994  * @param[in] flow
1995  *   Flow to destroy.
1996  */
1997 static void
1998 priv_flow_destroy(struct priv *priv,
1999                   struct mlx5_flows *list,
2000                   struct rte_flow *flow)
2001 {
2002         unsigned int i;
2003
2004         if (flow->cs) {
2005                 claim_zero(ibv_destroy_counter_set(flow->cs));
2006                 flow->cs = NULL;
2007         }
2008         if (flow->drop || !flow->mark)
2009                 goto free;
2010         for (i = 0; i != flow->queues_n; ++i) {
2011                 struct rte_flow *tmp;
2012                 int mark = 0;
2013
2014                 /*
2015                  * To remove the mark from the queue, the queue must not be
2016                  * present in any other marked flow (RSS or not).
2017                  */
2018                 TAILQ_FOREACH(tmp, list, next) {
2019                         unsigned int j;
2020                         uint16_t *tqs = NULL;
2021                         uint16_t tq_n = 0;
2022
2023                         if (!tmp->mark)
2024                                 continue;
2025                         for (j = 0; j != hash_rxq_init_n; ++j) {
2026                                 if (!tmp->frxq[j].hrxq)
2027                                         continue;
2028                                 tqs = tmp->frxq[j].hrxq->ind_table->queues;
2029                                 tq_n = tmp->frxq[j].hrxq->ind_table->queues_n;
2030                         }
2031                         if (!tq_n)
2032                                 continue;
2033                         for (j = 0; (j != tq_n) && !mark; j++)
2034                                 if (tqs[j] == (*flow->queues)[i])
2035                                         mark = 1;
2036                 }
2037                 (*priv->rxqs)[(*flow->queues)[i]]->mark = mark;
2038         }
2039 free:
2040         if (flow->drop) {
2041                 if (flow->drxq.ibv_flow)
2042                         claim_zero(ibv_destroy_flow(flow->drxq.ibv_flow));
2043                 rte_free(flow->drxq.ibv_attr);
2044         } else {
2045                 for (i = 0; i != hash_rxq_init_n; ++i) {
2046                         struct mlx5_flow *frxq = &flow->frxq[i];
2047
2048                         if (frxq->ibv_flow)
2049                                 claim_zero(ibv_destroy_flow(frxq->ibv_flow));
2050                         if (frxq->hrxq)
2051                                 mlx5_priv_hrxq_release(priv, frxq->hrxq);
2052                         if (frxq->ibv_attr)
2053                                 rte_free(frxq->ibv_attr);
2054                 }
2055         }
2056         TAILQ_REMOVE(list, flow, next);
2057         DEBUG("Flow destroyed %p", (void *)flow);
2058         rte_free(flow);
2059 }
2060
2061 /**
2062  * Destroy all flows.
2063  *
2064  * @param priv
2065  *   Pointer to private structure.
2066  * @param list
2067  *   Pointer to a TAILQ flow list.
2068  */
2069 void
2070 priv_flow_flush(struct priv *priv, struct mlx5_flows *list)
2071 {
2072         while (!TAILQ_EMPTY(list)) {
2073                 struct rte_flow *flow;
2074
2075                 flow = TAILQ_FIRST(list);
2076                 priv_flow_destroy(priv, list, flow);
2077         }
2078 }
2079
2080 /**
2081  * Create drop queue.
2082  *
2083  * @param priv
2084  *   Pointer to private structure.
2085  *
2086  * @return
2087  *   0 on success.
2088  */
2089 int
2090 priv_flow_create_drop_queue(struct priv *priv)
2091 {
2092         struct mlx5_hrxq_drop *fdq = NULL;
2093
2094         assert(priv->pd);
2095         assert(priv->ctx);
2096         fdq = rte_calloc(__func__, 1, sizeof(*fdq), 0);
2097         if (!fdq) {
2098                 WARN("cannot allocate memory for drop queue");
2099                 goto error;
2100         }
2101         fdq->cq = ibv_create_cq(priv->ctx, 1, NULL, NULL, 0);
2102         if (!fdq->cq) {
2103                 WARN("cannot allocate CQ for drop queue");
2104                 goto error;
2105         }
2106         fdq->wq = ibv_create_wq(priv->ctx,
2107                         &(struct ibv_wq_init_attr){
2108                         .wq_type = IBV_WQT_RQ,
2109                         .max_wr = 1,
2110                         .max_sge = 1,
2111                         .pd = priv->pd,
2112                         .cq = fdq->cq,
2113                         });
2114         if (!fdq->wq) {
2115                 WARN("cannot allocate WQ for drop queue");
2116                 goto error;
2117         }
2118         fdq->ind_table = ibv_create_rwq_ind_table(priv->ctx,
2119                         &(struct ibv_rwq_ind_table_init_attr){
2120                         .log_ind_tbl_size = 0,
2121                         .ind_tbl = &fdq->wq,
2122                         .comp_mask = 0,
2123                         });
2124         if (!fdq->ind_table) {
2125                 WARN("cannot allocate indirection table for drop queue");
2126                 goto error;
2127         }
2128         fdq->qp = ibv_create_qp_ex(priv->ctx,
2129                 &(struct ibv_qp_init_attr_ex){
2130                         .qp_type = IBV_QPT_RAW_PACKET,
2131                         .comp_mask =
2132                                 IBV_QP_INIT_ATTR_PD |
2133                                 IBV_QP_INIT_ATTR_IND_TABLE |
2134                                 IBV_QP_INIT_ATTR_RX_HASH,
2135                         .rx_hash_conf = (struct ibv_rx_hash_conf){
2136                                 .rx_hash_function =
2137                                         IBV_RX_HASH_FUNC_TOEPLITZ,
2138                                 .rx_hash_key_len = rss_hash_default_key_len,
2139                                 .rx_hash_key = rss_hash_default_key,
2140                                 .rx_hash_fields_mask = 0,
2141                                 },
2142                         .rwq_ind_tbl = fdq->ind_table,
2143                         .pd = priv->pd
2144                 });
2145         if (!fdq->qp) {
2146                 WARN("cannot allocate QP for drop queue");
2147                 goto error;
2148         }
2149         priv->flow_drop_queue = fdq;
2150         return 0;
2151 error:
2152         if (fdq->qp)
2153                 claim_zero(ibv_destroy_qp(fdq->qp));
2154         if (fdq->ind_table)
2155                 claim_zero(ibv_destroy_rwq_ind_table(fdq->ind_table));
2156         if (fdq->wq)
2157                 claim_zero(ibv_destroy_wq(fdq->wq));
2158         if (fdq->cq)
2159                 claim_zero(ibv_destroy_cq(fdq->cq));
2160         if (fdq)
2161                 rte_free(fdq);
2162         priv->flow_drop_queue = NULL;
2163         return -1;
2164 }
2165
2166 /**
2167  * Delete drop queue.
2168  *
2169  * @param priv
2170  *   Pointer to private structure.
2171  */
2172 void
2173 priv_flow_delete_drop_queue(struct priv *priv)
2174 {
2175         struct mlx5_hrxq_drop *fdq = priv->flow_drop_queue;
2176
2177         if (!fdq)
2178                 return;
2179         if (fdq->qp)
2180                 claim_zero(ibv_destroy_qp(fdq->qp));
2181         if (fdq->ind_table)
2182                 claim_zero(ibv_destroy_rwq_ind_table(fdq->ind_table));
2183         if (fdq->wq)
2184                 claim_zero(ibv_destroy_wq(fdq->wq));
2185         if (fdq->cq)
2186                 claim_zero(ibv_destroy_cq(fdq->cq));
2187         rte_free(fdq);
2188         priv->flow_drop_queue = NULL;
2189 }
2190
2191 /**
2192  * Remove all flows.
2193  *
2194  * @param priv
2195  *   Pointer to private structure.
2196  * @param list
2197  *   Pointer to a TAILQ flow list.
2198  */
2199 void
2200 priv_flow_stop(struct priv *priv, struct mlx5_flows *list)
2201 {
2202         struct rte_flow *flow;
2203
2204         TAILQ_FOREACH_REVERSE(flow, list, mlx5_flows, next) {
2205                 unsigned int i;
2206
2207                 if (flow->drop) {
2208                         if (!flow->drxq.ibv_flow)
2209                                 continue;
2210                         claim_zero(ibv_destroy_flow(flow->drxq.ibv_flow));
2211                         flow->drxq.ibv_flow = NULL;
2212                         /* Next flow. */
2213                         continue;
2214                 }
2215                 if (flow->mark) {
2216                         struct mlx5_ind_table_ibv *ind_tbl = NULL;
2217
2218                         for (i = 0; i != hash_rxq_init_n; ++i) {
2219                                 if (!flow->frxq[i].hrxq)
2220                                         continue;
2221                                 ind_tbl = flow->frxq[i].hrxq->ind_table;
2222                         }
2223                         assert(ind_tbl);
2224                         for (i = 0; i != ind_tbl->queues_n; ++i)
2225                                 (*priv->rxqs)[ind_tbl->queues[i]]->mark = 0;
2226                 }
2227                 for (i = 0; i != hash_rxq_init_n; ++i) {
2228                         if (!flow->frxq[i].ibv_flow)
2229                                 continue;
2230                         claim_zero(ibv_destroy_flow(flow->frxq[i].ibv_flow));
2231                         flow->frxq[i].ibv_flow = NULL;
2232                         mlx5_priv_hrxq_release(priv, flow->frxq[i].hrxq);
2233                         flow->frxq[i].hrxq = NULL;
2234                 }
2235                 DEBUG("Flow %p removed", (void *)flow);
2236         }
2237 }
2238
2239 /**
2240  * Add all flows.
2241  *
2242  * @param priv
2243  *   Pointer to private structure.
2244  * @param list
2245  *   Pointer to a TAILQ flow list.
2246  *
2247  * @return
2248  *   0 on success, a errno value otherwise and rte_errno is set.
2249  */
2250 int
2251 priv_flow_start(struct priv *priv, struct mlx5_flows *list)
2252 {
2253         struct rte_flow *flow;
2254
2255         TAILQ_FOREACH(flow, list, next) {
2256                 unsigned int i;
2257
2258                 if (flow->drop) {
2259                         flow->drxq.ibv_flow =
2260                                 ibv_create_flow(priv->flow_drop_queue->qp,
2261                                                 flow->drxq.ibv_attr);
2262                         if (!flow->drxq.ibv_flow) {
2263                                 DEBUG("Flow %p cannot be applied",
2264                                       (void *)flow);
2265                                 rte_errno = EINVAL;
2266                                 return rte_errno;
2267                         }
2268                         DEBUG("Flow %p applied", (void *)flow);
2269                         /* Next flow. */
2270                         continue;
2271                 }
2272                 for (i = 0; i != hash_rxq_init_n; ++i) {
2273                         if (!flow->frxq[i].ibv_attr)
2274                                 continue;
2275                         flow->frxq[i].hrxq =
2276                                 mlx5_priv_hrxq_get(priv, flow->rss_conf.rss_key,
2277                                                    flow->rss_conf.rss_key_len,
2278                                                    hash_rxq_init[i].hash_fields,
2279                                                    (*flow->queues),
2280                                                    flow->queues_n);
2281                         if (flow->frxq[i].hrxq)
2282                                 goto flow_create;
2283                         flow->frxq[i].hrxq =
2284                                 mlx5_priv_hrxq_new(priv, flow->rss_conf.rss_key,
2285                                                    flow->rss_conf.rss_key_len,
2286                                                    hash_rxq_init[i].hash_fields,
2287                                                    (*flow->queues),
2288                                                    flow->queues_n);
2289                         if (!flow->frxq[i].hrxq) {
2290                                 DEBUG("Flow %p cannot be applied",
2291                                       (void *)flow);
2292                                 rte_errno = EINVAL;
2293                                 return rte_errno;
2294                         }
2295 flow_create:
2296                         flow->frxq[i].ibv_flow =
2297                                 ibv_create_flow(flow->frxq[i].hrxq->qp,
2298                                                 flow->frxq[i].ibv_attr);
2299                         if (!flow->frxq[i].ibv_flow) {
2300                                 DEBUG("Flow %p cannot be applied",
2301                                       (void *)flow);
2302                                 rte_errno = EINVAL;
2303                                 return rte_errno;
2304                         }
2305                         DEBUG("Flow %p applied", (void *)flow);
2306                 }
2307                 if (!flow->mark)
2308                         continue;
2309                 for (i = 0; i != flow->queues_n; ++i)
2310                         (*priv->rxqs)[(*flow->queues)[i]]->mark = 1;
2311         }
2312         return 0;
2313 }
2314
2315 /**
2316  * Verify the flow list is empty
2317  *
2318  * @param priv
2319  *  Pointer to private structure.
2320  *
2321  * @return the number of flows not released.
2322  */
2323 int
2324 priv_flow_verify(struct priv *priv)
2325 {
2326         struct rte_flow *flow;
2327         int ret = 0;
2328
2329         TAILQ_FOREACH(flow, &priv->flows, next) {
2330                 DEBUG("%p: flow %p still referenced", (void *)priv,
2331                       (void *)flow);
2332                 ++ret;
2333         }
2334         return ret;
2335 }
2336
2337 /**
2338  * Enable a control flow configured from the control plane.
2339  *
2340  * @param dev
2341  *   Pointer to Ethernet device.
2342  * @param eth_spec
2343  *   An Ethernet flow spec to apply.
2344  * @param eth_mask
2345  *   An Ethernet flow mask to apply.
2346  * @param vlan_spec
2347  *   A VLAN flow spec to apply.
2348  * @param vlan_mask
2349  *   A VLAN flow mask to apply.
2350  *
2351  * @return
2352  *   0 on success.
2353  */
2354 int
2355 mlx5_ctrl_flow_vlan(struct rte_eth_dev *dev,
2356                     struct rte_flow_item_eth *eth_spec,
2357                     struct rte_flow_item_eth *eth_mask,
2358                     struct rte_flow_item_vlan *vlan_spec,
2359                     struct rte_flow_item_vlan *vlan_mask)
2360 {
2361         struct priv *priv = dev->data->dev_private;
2362         const struct rte_flow_attr attr = {
2363                 .ingress = 1,
2364                 .priority = MLX5_CTRL_FLOW_PRIORITY,
2365         };
2366         struct rte_flow_item items[] = {
2367                 {
2368                         .type = RTE_FLOW_ITEM_TYPE_ETH,
2369                         .spec = eth_spec,
2370                         .last = NULL,
2371                         .mask = eth_mask,
2372                 },
2373                 {
2374                         .type = (vlan_spec) ? RTE_FLOW_ITEM_TYPE_VLAN :
2375                                 RTE_FLOW_ITEM_TYPE_END,
2376                         .spec = vlan_spec,
2377                         .last = NULL,
2378                         .mask = vlan_mask,
2379                 },
2380                 {
2381                         .type = RTE_FLOW_ITEM_TYPE_END,
2382                 },
2383         };
2384         struct rte_flow_action actions[] = {
2385                 {
2386                         .type = RTE_FLOW_ACTION_TYPE_RSS,
2387                 },
2388                 {
2389                         .type = RTE_FLOW_ACTION_TYPE_END,
2390                 },
2391         };
2392         struct rte_flow *flow;
2393         struct rte_flow_error error;
2394         unsigned int i;
2395         union {
2396                 struct rte_flow_action_rss rss;
2397                 struct {
2398                         const struct rte_eth_rss_conf *rss_conf;
2399                         uint16_t num;
2400                         uint16_t queue[RTE_MAX_QUEUES_PER_PORT];
2401                 } local;
2402         } action_rss;
2403
2404         if (!priv->reta_idx_n)
2405                 return EINVAL;
2406         for (i = 0; i != priv->reta_idx_n; ++i)
2407                 action_rss.local.queue[i] = (*priv->reta_idx)[i];
2408         action_rss.local.rss_conf = &priv->rss_conf;
2409         action_rss.local.num = priv->reta_idx_n;
2410         actions[0].conf = (const void *)&action_rss.rss;
2411         flow = priv_flow_create(priv, &priv->ctrl_flows, &attr, items, actions,
2412                                 &error);
2413         if (!flow)
2414                 return rte_errno;
2415         return 0;
2416 }
2417
2418 /**
2419  * Enable a flow control configured from the control plane.
2420  *
2421  * @param dev
2422  *   Pointer to Ethernet device.
2423  * @param eth_spec
2424  *   An Ethernet flow spec to apply.
2425  * @param eth_mask
2426  *   An Ethernet flow mask to apply.
2427  *
2428  * @return
2429  *   0 on success.
2430  */
2431 int
2432 mlx5_ctrl_flow(struct rte_eth_dev *dev,
2433                struct rte_flow_item_eth *eth_spec,
2434                struct rte_flow_item_eth *eth_mask)
2435 {
2436         return mlx5_ctrl_flow_vlan(dev, eth_spec, eth_mask, NULL, NULL);
2437 }
2438
2439 /**
2440  * Destroy a flow.
2441  *
2442  * @see rte_flow_destroy()
2443  * @see rte_flow_ops
2444  */
2445 int
2446 mlx5_flow_destroy(struct rte_eth_dev *dev,
2447                   struct rte_flow *flow,
2448                   struct rte_flow_error *error)
2449 {
2450         struct priv *priv = dev->data->dev_private;
2451
2452         (void)error;
2453         priv_lock(priv);
2454         priv_flow_destroy(priv, &priv->flows, flow);
2455         priv_unlock(priv);
2456         return 0;
2457 }
2458
2459 /**
2460  * Destroy all flows.
2461  *
2462  * @see rte_flow_flush()
2463  * @see rte_flow_ops
2464  */
2465 int
2466 mlx5_flow_flush(struct rte_eth_dev *dev,
2467                 struct rte_flow_error *error)
2468 {
2469         struct priv *priv = dev->data->dev_private;
2470
2471         (void)error;
2472         priv_lock(priv);
2473         priv_flow_flush(priv, &priv->flows);
2474         priv_unlock(priv);
2475         return 0;
2476 }
2477
2478 #ifdef HAVE_IBV_DEVICE_COUNTERS_SET_SUPPORT
2479 /**
2480  * Query flow counter.
2481  *
2482  * @param cs
2483  *   the counter set.
2484  * @param counter_value
2485  *   returned data from the counter.
2486  *
2487  * @return
2488  *   0 on success, a errno value otherwise and rte_errno is set.
2489  */
2490 static int
2491 priv_flow_query_count(struct ibv_counter_set *cs,
2492                       struct mlx5_flow_counter_stats *counter_stats,
2493                       struct rte_flow_query_count *query_count,
2494                       struct rte_flow_error *error)
2495 {
2496         uint64_t counters[2];
2497         struct ibv_query_counter_set_attr query_cs_attr = {
2498                 .cs = cs,
2499                 .query_flags = IBV_COUNTER_SET_FORCE_UPDATE,
2500         };
2501         struct ibv_counter_set_data query_out = {
2502                 .out = counters,
2503                 .outlen = 2 * sizeof(uint64_t),
2504         };
2505         int res = ibv_query_counter_set(&query_cs_attr, &query_out);
2506
2507         if (res) {
2508                 rte_flow_error_set(error, -res,
2509                                    RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
2510                                    NULL,
2511                                    "cannot read counter");
2512                 return -res;
2513         }
2514         query_count->hits_set = 1;
2515         query_count->bytes_set = 1;
2516         query_count->hits = counters[0] - counter_stats->hits;
2517         query_count->bytes = counters[1] - counter_stats->bytes;
2518         if (query_count->reset) {
2519                 counter_stats->hits = counters[0];
2520                 counter_stats->bytes = counters[1];
2521         }
2522         return 0;
2523 }
2524
2525 /**
2526  * Query a flows.
2527  *
2528  * @see rte_flow_query()
2529  * @see rte_flow_ops
2530  */
2531 int
2532 mlx5_flow_query(struct rte_eth_dev *dev,
2533                 struct rte_flow *flow,
2534                 enum rte_flow_action_type action __rte_unused,
2535                 void *data,
2536                 struct rte_flow_error *error)
2537 {
2538         struct priv *priv = dev->data->dev_private;
2539         int res = EINVAL;
2540
2541         priv_lock(priv);
2542         if (flow->cs) {
2543                 res = priv_flow_query_count(flow->cs,
2544                                         &flow->counter_stats,
2545                                         (struct rte_flow_query_count *)data,
2546                                         error);
2547         } else {
2548                 rte_flow_error_set(error, res,
2549                                    RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
2550                                    NULL,
2551                                    "no counter found for flow");
2552         }
2553         priv_unlock(priv);
2554         return -res;
2555 }
2556 #endif
2557
2558 /**
2559  * Isolated mode.
2560  *
2561  * @see rte_flow_isolate()
2562  * @see rte_flow_ops
2563  */
2564 int
2565 mlx5_flow_isolate(struct rte_eth_dev *dev,
2566                   int enable,
2567                   struct rte_flow_error *error)
2568 {
2569         struct priv *priv = dev->data->dev_private;
2570
2571         priv_lock(priv);
2572         if (dev->data->dev_started) {
2573                 rte_flow_error_set(error, EBUSY,
2574                                    RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
2575                                    NULL,
2576                                    "port must be stopped first");
2577                 priv_unlock(priv);
2578                 return -rte_errno;
2579         }
2580         priv->isolated = !!enable;
2581         if (enable)
2582                 priv->dev->dev_ops = &mlx5_dev_ops_isolate;
2583         else
2584                 priv->dev->dev_ops = &mlx5_dev_ops;
2585         priv_unlock(priv);
2586         return 0;
2587 }
2588
2589 /**
2590  * Convert a flow director filter to a generic flow.
2591  *
2592  * @param priv
2593  *   Private structure.
2594  * @param fdir_filter
2595  *   Flow director filter to add.
2596  * @param attributes
2597  *   Generic flow parameters structure.
2598  *
2599  * @return
2600  *  0 on success, errno value on error.
2601  */
2602 static int
2603 priv_fdir_filter_convert(struct priv *priv,
2604                          const struct rte_eth_fdir_filter *fdir_filter,
2605                          struct mlx5_fdir *attributes)
2606 {
2607         const struct rte_eth_fdir_input *input = &fdir_filter->input;
2608
2609         /* Validate queue number. */
2610         if (fdir_filter->action.rx_queue >= priv->rxqs_n) {
2611                 ERROR("invalid queue number %d", fdir_filter->action.rx_queue);
2612                 return EINVAL;
2613         }
2614         attributes->attr.ingress = 1;
2615         attributes->items[0] = (struct rte_flow_item) {
2616                 .type = RTE_FLOW_ITEM_TYPE_ETH,
2617                 .spec = &attributes->l2,
2618         };
2619         switch (fdir_filter->action.behavior) {
2620         case RTE_ETH_FDIR_ACCEPT:
2621                 attributes->actions[0] = (struct rte_flow_action){
2622                         .type = RTE_FLOW_ACTION_TYPE_QUEUE,
2623                         .conf = &attributes->queue,
2624                 };
2625                 break;
2626         case RTE_ETH_FDIR_REJECT:
2627                 attributes->actions[0] = (struct rte_flow_action){
2628                         .type = RTE_FLOW_ACTION_TYPE_DROP,
2629                 };
2630                 break;
2631         default:
2632                 ERROR("invalid behavior %d", fdir_filter->action.behavior);
2633                 return ENOTSUP;
2634         }
2635         attributes->queue.index = fdir_filter->action.rx_queue;
2636         switch (fdir_filter->input.flow_type) {
2637         case RTE_ETH_FLOW_NONFRAG_IPV4_UDP:
2638                 attributes->l3.ipv4.hdr = (struct ipv4_hdr){
2639                         .src_addr = input->flow.udp4_flow.ip.src_ip,
2640                         .dst_addr = input->flow.udp4_flow.ip.dst_ip,
2641                         .time_to_live = input->flow.udp4_flow.ip.ttl,
2642                         .type_of_service = input->flow.udp4_flow.ip.tos,
2643                         .next_proto_id = input->flow.udp4_flow.ip.proto,
2644                 };
2645                 attributes->l4.udp.hdr = (struct udp_hdr){
2646                         .src_port = input->flow.udp4_flow.src_port,
2647                         .dst_port = input->flow.udp4_flow.dst_port,
2648                 };
2649                 attributes->items[1] = (struct rte_flow_item){
2650                         .type = RTE_FLOW_ITEM_TYPE_IPV4,
2651                         .spec = &attributes->l3,
2652                 };
2653                 attributes->items[2] = (struct rte_flow_item){
2654                         .type = RTE_FLOW_ITEM_TYPE_UDP,
2655                         .spec = &attributes->l4,
2656                 };
2657                 break;
2658         case RTE_ETH_FLOW_NONFRAG_IPV4_TCP:
2659                 attributes->l3.ipv4.hdr = (struct ipv4_hdr){
2660                         .src_addr = input->flow.tcp4_flow.ip.src_ip,
2661                         .dst_addr = input->flow.tcp4_flow.ip.dst_ip,
2662                         .time_to_live = input->flow.tcp4_flow.ip.ttl,
2663                         .type_of_service = input->flow.tcp4_flow.ip.tos,
2664                         .next_proto_id = input->flow.tcp4_flow.ip.proto,
2665                 };
2666                 attributes->l4.tcp.hdr = (struct tcp_hdr){
2667                         .src_port = input->flow.tcp4_flow.src_port,
2668                         .dst_port = input->flow.tcp4_flow.dst_port,
2669                 };
2670                 attributes->items[1] = (struct rte_flow_item){
2671                         .type = RTE_FLOW_ITEM_TYPE_IPV4,
2672                         .spec = &attributes->l3,
2673                 };
2674                 attributes->items[2] = (struct rte_flow_item){
2675                         .type = RTE_FLOW_ITEM_TYPE_TCP,
2676                         .spec = &attributes->l4,
2677                 };
2678                 break;
2679         case RTE_ETH_FLOW_NONFRAG_IPV4_OTHER:
2680                 attributes->l3.ipv4.hdr = (struct ipv4_hdr){
2681                         .src_addr = input->flow.ip4_flow.src_ip,
2682                         .dst_addr = input->flow.ip4_flow.dst_ip,
2683                         .time_to_live = input->flow.ip4_flow.ttl,
2684                         .type_of_service = input->flow.ip4_flow.tos,
2685                         .next_proto_id = input->flow.ip4_flow.proto,
2686                 };
2687                 attributes->items[1] = (struct rte_flow_item){
2688                         .type = RTE_FLOW_ITEM_TYPE_IPV4,
2689                         .spec = &attributes->l3,
2690                 };
2691                 break;
2692         case RTE_ETH_FLOW_NONFRAG_IPV6_UDP:
2693                 attributes->l3.ipv6.hdr = (struct ipv6_hdr){
2694                         .hop_limits = input->flow.udp6_flow.ip.hop_limits,
2695                         .proto = input->flow.udp6_flow.ip.proto,
2696                 };
2697                 memcpy(attributes->l3.ipv6.hdr.src_addr,
2698                        input->flow.udp6_flow.ip.src_ip,
2699                        RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
2700                 memcpy(attributes->l3.ipv6.hdr.dst_addr,
2701                        input->flow.udp6_flow.ip.dst_ip,
2702                        RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
2703                 attributes->l4.udp.hdr = (struct udp_hdr){
2704                         .src_port = input->flow.udp6_flow.src_port,
2705                         .dst_port = input->flow.udp6_flow.dst_port,
2706                 };
2707                 attributes->items[1] = (struct rte_flow_item){
2708                         .type = RTE_FLOW_ITEM_TYPE_IPV6,
2709                         .spec = &attributes->l3,
2710                 };
2711                 attributes->items[2] = (struct rte_flow_item){
2712                         .type = RTE_FLOW_ITEM_TYPE_UDP,
2713                         .spec = &attributes->l4,
2714                 };
2715                 break;
2716         case RTE_ETH_FLOW_NONFRAG_IPV6_TCP:
2717                 attributes->l3.ipv6.hdr = (struct ipv6_hdr){
2718                         .hop_limits = input->flow.tcp6_flow.ip.hop_limits,
2719                         .proto = input->flow.tcp6_flow.ip.proto,
2720                 };
2721                 memcpy(attributes->l3.ipv6.hdr.src_addr,
2722                        input->flow.tcp6_flow.ip.src_ip,
2723                        RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
2724                 memcpy(attributes->l3.ipv6.hdr.dst_addr,
2725                        input->flow.tcp6_flow.ip.dst_ip,
2726                        RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
2727                 attributes->l4.tcp.hdr = (struct tcp_hdr){
2728                         .src_port = input->flow.tcp6_flow.src_port,
2729                         .dst_port = input->flow.tcp6_flow.dst_port,
2730                 };
2731                 attributes->items[1] = (struct rte_flow_item){
2732                         .type = RTE_FLOW_ITEM_TYPE_IPV6,
2733                         .spec = &attributes->l3,
2734                 };
2735                 attributes->items[2] = (struct rte_flow_item){
2736                         .type = RTE_FLOW_ITEM_TYPE_UDP,
2737                         .spec = &attributes->l4,
2738                 };
2739                 break;
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                 memcpy(attributes->l3.ipv6.hdr.src_addr,
2746                        input->flow.ipv6_flow.src_ip,
2747                        RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
2748                 memcpy(attributes->l3.ipv6.hdr.dst_addr,
2749                        input->flow.ipv6_flow.dst_ip,
2750                        RTE_DIM(attributes->l3.ipv6.hdr.src_addr));
2751                 attributes->items[1] = (struct rte_flow_item){
2752                         .type = RTE_FLOW_ITEM_TYPE_IPV6,
2753                         .spec = &attributes->l3,
2754                 };
2755                 break;
2756         default:
2757                 ERROR("invalid flow type%d",
2758                       fdir_filter->input.flow_type);
2759                 return ENOTSUP;
2760         }
2761         return 0;
2762 }
2763
2764 /**
2765  * Add new flow director filter and store it in list.
2766  *
2767  * @param priv
2768  *   Private structure.
2769  * @param fdir_filter
2770  *   Flow director filter to add.
2771  *
2772  * @return
2773  *   0 on success, errno value on failure.
2774  */
2775 static int
2776 priv_fdir_filter_add(struct priv *priv,
2777                      const struct rte_eth_fdir_filter *fdir_filter)
2778 {
2779         struct mlx5_fdir attributes = {
2780                 .attr.group = 0,
2781         };
2782         struct mlx5_flow_parse parser = {
2783                 .layer = HASH_RXQ_ETH,
2784         };
2785         struct rte_flow_error error;
2786         struct rte_flow *flow;
2787         int ret;
2788
2789         ret = priv_fdir_filter_convert(priv, fdir_filter, &attributes);
2790         if (ret)
2791                 return -ret;
2792         ret = priv_flow_convert(priv, &attributes.attr, attributes.items,
2793                                 attributes.actions, &error, &parser);
2794         if (ret)
2795                 return -ret;
2796         flow = priv_flow_create(priv,
2797                                 &priv->flows,
2798                                 &attributes.attr,
2799                                 attributes.items,
2800                                 attributes.actions,
2801                                 &error);
2802         if (flow) {
2803                 DEBUG("FDIR created %p", (void *)flow);
2804                 return 0;
2805         }
2806         return ENOTSUP;
2807 }
2808
2809 /**
2810  * Delete specific filter.
2811  *
2812  * @param priv
2813  *   Private structure.
2814  * @param fdir_filter
2815  *   Filter to be deleted.
2816  *
2817  * @return
2818  *   0 on success, errno value on failure.
2819  */
2820 static int
2821 priv_fdir_filter_delete(struct priv *priv,
2822                         const struct rte_eth_fdir_filter *fdir_filter)
2823 {
2824         struct mlx5_fdir attributes;
2825         struct mlx5_flow_parse parser = {
2826                 .create = 1,
2827                 .layer = HASH_RXQ_ETH,
2828         };
2829         struct rte_flow_error error;
2830         struct rte_flow *flow;
2831         unsigned int i;
2832         int ret;
2833
2834         ret = priv_fdir_filter_convert(priv, fdir_filter, &attributes);
2835         if (ret)
2836                 return -ret;
2837         ret = priv_flow_convert(priv, &attributes.attr, attributes.items,
2838                                 attributes.actions, &error, &parser);
2839         if (ret)
2840                 goto exit;
2841         TAILQ_FOREACH(flow, &priv->flows, next) {
2842                 struct ibv_flow_attr *attr;
2843                 struct ibv_spec_header *attr_h;
2844                 void *spec;
2845                 struct ibv_flow_attr *flow_attr;
2846                 struct ibv_spec_header *flow_h;
2847                 void *flow_spec;
2848                 unsigned int specs_n;
2849
2850                 if (parser.drop)
2851                         attr = parser.drop_q.ibv_attr;
2852                 else
2853                         attr = parser.queue[HASH_RXQ_ETH].ibv_attr;
2854                 if (flow->drop)
2855                         flow_attr = flow->drxq.ibv_attr;
2856                 else
2857                         flow_attr = flow->frxq[HASH_RXQ_ETH].ibv_attr;
2858                 /* Compare first the attributes. */
2859                 if (memcmp(attr, flow_attr, sizeof(struct ibv_flow_attr)))
2860                         continue;
2861                 if (attr->num_of_specs == 0)
2862                         continue;
2863                 spec = (void *)((uintptr_t)attr +
2864                                 sizeof(struct ibv_flow_attr));
2865                 flow_spec = (void *)((uintptr_t)flow_attr +
2866                                      sizeof(struct ibv_flow_attr));
2867                 specs_n = RTE_MIN(attr->num_of_specs, flow_attr->num_of_specs);
2868                 for (i = 0; i != specs_n; ++i) {
2869                         attr_h = spec;
2870                         flow_h = flow_spec;
2871                         if (memcmp(spec, flow_spec,
2872                                    RTE_MIN(attr_h->size, flow_h->size)))
2873                                 continue;
2874                         spec = (void *)((uintptr_t)attr + attr_h->size);
2875                         flow_spec = (void *)((uintptr_t)flow_attr +
2876                                              flow_h->size);
2877                 }
2878                 /* At this point, the flow match. */
2879                 break;
2880         }
2881         if (flow)
2882                 priv_flow_destroy(priv, &priv->flows, flow);
2883 exit:
2884         if (parser.drop) {
2885                 rte_free(parser.drop_q.ibv_attr);
2886         } else {
2887                 for (i = 0; i != hash_rxq_init_n; ++i) {
2888                         if (parser.queue[i].ibv_attr)
2889                                 rte_free(parser.queue[i].ibv_attr);
2890                 }
2891         }
2892         return -ret;
2893 }
2894
2895 /**
2896  * Update queue for specific filter.
2897  *
2898  * @param priv
2899  *   Private structure.
2900  * @param fdir_filter
2901  *   Filter to be updated.
2902  *
2903  * @return
2904  *   0 on success, errno value on failure.
2905  */
2906 static int
2907 priv_fdir_filter_update(struct priv *priv,
2908                         const struct rte_eth_fdir_filter *fdir_filter)
2909 {
2910         int ret;
2911
2912         ret = priv_fdir_filter_delete(priv, fdir_filter);
2913         if (ret)
2914                 return ret;
2915         ret = priv_fdir_filter_add(priv, fdir_filter);
2916         return ret;
2917 }
2918
2919 /**
2920  * Flush all filters.
2921  *
2922  * @param priv
2923  *   Private structure.
2924  */
2925 static void
2926 priv_fdir_filter_flush(struct priv *priv)
2927 {
2928         priv_flow_flush(priv, &priv->flows);
2929 }
2930
2931 /**
2932  * Get flow director information.
2933  *
2934  * @param priv
2935  *   Private structure.
2936  * @param[out] fdir_info
2937  *   Resulting flow director information.
2938  */
2939 static void
2940 priv_fdir_info_get(struct priv *priv, struct rte_eth_fdir_info *fdir_info)
2941 {
2942         struct rte_eth_fdir_masks *mask =
2943                 &priv->dev->data->dev_conf.fdir_conf.mask;
2944
2945         fdir_info->mode = priv->dev->data->dev_conf.fdir_conf.mode;
2946         fdir_info->guarant_spc = 0;
2947         rte_memcpy(&fdir_info->mask, mask, sizeof(fdir_info->mask));
2948         fdir_info->max_flexpayload = 0;
2949         fdir_info->flow_types_mask[0] = 0;
2950         fdir_info->flex_payload_unit = 0;
2951         fdir_info->max_flex_payload_segment_num = 0;
2952         fdir_info->flex_payload_limit = 0;
2953         memset(&fdir_info->flex_conf, 0, sizeof(fdir_info->flex_conf));
2954 }
2955
2956 /**
2957  * Deal with flow director operations.
2958  *
2959  * @param priv
2960  *   Pointer to private structure.
2961  * @param filter_op
2962  *   Operation to perform.
2963  * @param arg
2964  *   Pointer to operation-specific structure.
2965  *
2966  * @return
2967  *   0 on success, errno value on failure.
2968  */
2969 static int
2970 priv_fdir_ctrl_func(struct priv *priv, enum rte_filter_op filter_op, void *arg)
2971 {
2972         enum rte_fdir_mode fdir_mode =
2973                 priv->dev->data->dev_conf.fdir_conf.mode;
2974         int ret = 0;
2975
2976         if (filter_op == RTE_ETH_FILTER_NOP)
2977                 return 0;
2978         if (fdir_mode != RTE_FDIR_MODE_PERFECT &&
2979             fdir_mode != RTE_FDIR_MODE_PERFECT_MAC_VLAN) {
2980                 ERROR("%p: flow director mode %d not supported",
2981                       (void *)priv, fdir_mode);
2982                 return EINVAL;
2983         }
2984         switch (filter_op) {
2985         case RTE_ETH_FILTER_ADD:
2986                 ret = priv_fdir_filter_add(priv, arg);
2987                 break;
2988         case RTE_ETH_FILTER_UPDATE:
2989                 ret = priv_fdir_filter_update(priv, arg);
2990                 break;
2991         case RTE_ETH_FILTER_DELETE:
2992                 ret = priv_fdir_filter_delete(priv, arg);
2993                 break;
2994         case RTE_ETH_FILTER_FLUSH:
2995                 priv_fdir_filter_flush(priv);
2996                 break;
2997         case RTE_ETH_FILTER_INFO:
2998                 priv_fdir_info_get(priv, arg);
2999                 break;
3000         default:
3001                 DEBUG("%p: unknown operation %u", (void *)priv,
3002                       filter_op);
3003                 ret = EINVAL;
3004                 break;
3005         }
3006         return ret;
3007 }
3008
3009 /**
3010  * Manage filter operations.
3011  *
3012  * @param dev
3013  *   Pointer to Ethernet device structure.
3014  * @param filter_type
3015  *   Filter type.
3016  * @param filter_op
3017  *   Operation to perform.
3018  * @param arg
3019  *   Pointer to operation-specific structure.
3020  *
3021  * @return
3022  *   0 on success, negative errno value on failure.
3023  */
3024 int
3025 mlx5_dev_filter_ctrl(struct rte_eth_dev *dev,
3026                      enum rte_filter_type filter_type,
3027                      enum rte_filter_op filter_op,
3028                      void *arg)
3029 {
3030         int ret = EINVAL;
3031         struct priv *priv = dev->data->dev_private;
3032
3033         switch (filter_type) {
3034         case RTE_ETH_FILTER_GENERIC:
3035                 if (filter_op != RTE_ETH_FILTER_GET)
3036                         return -EINVAL;
3037                 *(const void **)arg = &mlx5_flow_ops;
3038                 return 0;
3039         case RTE_ETH_FILTER_FDIR:
3040                 priv_lock(priv);
3041                 ret = priv_fdir_ctrl_func(priv, filter_op, arg);
3042                 priv_unlock(priv);
3043                 break;
3044         default:
3045                 ERROR("%p: filter type (%d) not supported",
3046                       (void *)dev, filter_type);
3047                 break;
3048         }
3049         return -ret;
3050 }