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 static int
  59 mlx5_flow_create_eth(const struct rte_flow_item *item,
  60                      const void *default_mask,
  61                      void *data);
  62
  63 static int
  64 mlx5_flow_create_vlan(const struct rte_flow_item *item,
  65                       const void *default_mask,
  66                       void *data);
  67
  68 static int
  69 mlx5_flow_create_ipv4(const struct rte_flow_item *item,
  70                       const void *default_mask,
  71                       void *data);
  72
  73 static int
  74 mlx5_flow_create_ipv6(const struct rte_flow_item *item,
  75                       const void *default_mask,
  76                       void *data);
  77
  78 static int
  79 mlx5_flow_create_udp(const struct rte_flow_item *item,
  80                      const void *default_mask,
  81                      void *data);
  82
  83 static int
  84 mlx5_flow_create_tcp(const struct rte_flow_item *item,
  85                      const void *default_mask,
  86                      void *data);
  87
  88 static int
  89 mlx5_flow_create_vxlan(const struct rte_flow_item *item,
  90                        const void *default_mask,
  91                        void *data);
  92
  93 /** Structure for Drop queue. */
  94 struct mlx5_hrxq_drop {
  95         struct ibv_rwq_ind_table *ind_table; /**< Indirection table. */
  96         struct ibv_qp *qp; /**< Verbs queue pair. */
  97         struct ibv_wq *wq; /**< Verbs work queue. */
  98         struct ibv_cq *cq; /**< Verbs completion queue. */
  99 };
 100
 101 /* Flows structures. */
 102 struct mlx5_flow {
 103         uint64_t hash_fields; /**< Fields that participate in the hash. */
 104         struct mlx5_hrxq *hrxq; /**< Hash Rx queues. */
 105 };
 106
 107 /* Drop flows structures. */
 108 struct mlx5_flow_drop {
 109         struct mlx5_hrxq_drop hrxq; /**< Drop hash Rx queue. */
 110 };
 111
 112 struct rte_flow {
 113         TAILQ_ENTRY(rte_flow) next; /**< Pointer to the next flow structure. */
 114         uint32_t mark:1; /**< Set if the flow is marked. */
 115         uint32_t drop:1; /**< Drop queue. */
 116         struct ibv_flow_attr *ibv_attr; /**< Pointer to Verbs attributes. */
 117         struct ibv_flow *ibv_flow; /**< Verbs flow. */
 118         uint16_t queues_n; /**< Number of entries in queue[]. */
 119         uint16_t (*queues)[]; /**< Queues indexes to use. */
 120         union {
 121                 struct mlx5_flow frxq; /**< Flow with Rx queue. */
 122                 struct mlx5_flow_drop drxq; /**< Flow with drop Rx queue. */
 123         };
 124 };
 125
 126 /** Static initializer for items. */
 127 #define ITEMS(...) \
 128         (const enum rte_flow_item_type []){ \
 129                 __VA_ARGS__, RTE_FLOW_ITEM_TYPE_END, \
 130         }
 131
 132 /** Structure to generate a simple graph of layers supported by the NIC. */
 133 struct mlx5_flow_items {
 134         /** List of possible actions for these items. */
 135         const enum rte_flow_action_type *const actions;
 136         /** Bit-masks corresponding to the possibilities for the item. */
 137         const void *mask;
 138         /**
 139          * Default bit-masks to use when item->mask is not provided. When
 140          * \default_mask is also NULL, the full supported bit-mask (\mask) is
 141          * used instead.
 142          */
 143         const void *default_mask;
 144         /** Bit-masks size in bytes. */
 145         const unsigned int mask_sz;
 146         /**
 147          * Conversion function from rte_flow to NIC specific flow.
 148          *
 149          * @param item
 150          *   rte_flow item to convert.
 151          * @param default_mask
 152          *   Default bit-masks to use when item->mask is not provided.
 153          * @param data
 154          *   Internal structure to store the conversion.
 155          *
 156          * @return
 157          *   0 on success, negative value otherwise.
 158          */
 159         int (*convert)(const struct rte_flow_item *item,
 160                        const void *default_mask,
 161                        void *data);
 162         /** Size in bytes of the destination structure. */
 163         const unsigned int dst_sz;
 164         /** List of possible following items.  */
 165         const enum rte_flow_item_type *const items;
 166 };
 167
 168 /** Valid action for this PMD. */
 169 static const enum rte_flow_action_type valid_actions[] = {
 170         RTE_FLOW_ACTION_TYPE_DROP,
 171         RTE_FLOW_ACTION_TYPE_QUEUE,
 172         RTE_FLOW_ACTION_TYPE_MARK,
 173         RTE_FLOW_ACTION_TYPE_FLAG,
 174         RTE_FLOW_ACTION_TYPE_END,
 175 };
 176
 177 /** Graph of supported items and associated actions. */
 178 static const struct mlx5_flow_items mlx5_flow_items[] = {
 179         [RTE_FLOW_ITEM_TYPE_END] = {
 180                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_ETH,
 181                                RTE_FLOW_ITEM_TYPE_VXLAN),
 182         },
 183         [RTE_FLOW_ITEM_TYPE_ETH] = {
 184                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_VLAN,
 185                                RTE_FLOW_ITEM_TYPE_IPV4,
 186                                RTE_FLOW_ITEM_TYPE_IPV6),
 187                 .actions = valid_actions,
 188                 .mask = &(const struct rte_flow_item_eth){
 189                         .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 190                         .src.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 191                         .type = -1,
 192                 },
 193                 .default_mask = &rte_flow_item_eth_mask,
 194                 .mask_sz = sizeof(struct rte_flow_item_eth),
 195                 .convert = mlx5_flow_create_eth,
 196                 .dst_sz = sizeof(struct ibv_flow_spec_eth),
 197         },
 198         [RTE_FLOW_ITEM_TYPE_VLAN] = {
 199                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_IPV4,
 200                                RTE_FLOW_ITEM_TYPE_IPV6),
 201                 .actions = valid_actions,
 202                 .mask = &(const struct rte_flow_item_vlan){
 203                         .tci = -1,
 204                 },
 205                 .default_mask = &rte_flow_item_vlan_mask,
 206                 .mask_sz = sizeof(struct rte_flow_item_vlan),
 207                 .convert = mlx5_flow_create_vlan,
 208                 .dst_sz = 0,
 209         },
 210         [RTE_FLOW_ITEM_TYPE_IPV4] = {
 211                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP,
 212                                RTE_FLOW_ITEM_TYPE_TCP),
 213                 .actions = valid_actions,
 214                 .mask = &(const struct rte_flow_item_ipv4){
 215                         .hdr = {
 216                                 .src_addr = -1,
 217                                 .dst_addr = -1,
 218                                 .type_of_service = -1,
 219                                 .next_proto_id = -1,
 220                         },
 221                 },
 222                 .default_mask = &rte_flow_item_ipv4_mask,
 223                 .mask_sz = sizeof(struct rte_flow_item_ipv4),
 224                 .convert = mlx5_flow_create_ipv4,
 225                 .dst_sz = sizeof(struct ibv_flow_spec_ipv4),
 226         },
 227         [RTE_FLOW_ITEM_TYPE_IPV6] = {
 228                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP,
 229                                RTE_FLOW_ITEM_TYPE_TCP),
 230                 .actions = valid_actions,
 231                 .mask = &(const struct rte_flow_item_ipv6){
 232                         .hdr = {
 233                                 .src_addr = {
 234                                         0xff, 0xff, 0xff, 0xff,
 235                                         0xff, 0xff, 0xff, 0xff,
 236                                         0xff, 0xff, 0xff, 0xff,
 237                                         0xff, 0xff, 0xff, 0xff,
 238                                 },
 239                                 .dst_addr = {
 240                                         0xff, 0xff, 0xff, 0xff,
 241                                         0xff, 0xff, 0xff, 0xff,
 242                                         0xff, 0xff, 0xff, 0xff,
 243                                         0xff, 0xff, 0xff, 0xff,
 244                                 },
 245                                 .vtc_flow = -1,
 246                                 .proto = -1,
 247                                 .hop_limits = -1,
 248                         },
 249                 },
 250                 .default_mask = &rte_flow_item_ipv6_mask,
 251                 .mask_sz = sizeof(struct rte_flow_item_ipv6),
 252                 .convert = mlx5_flow_create_ipv6,
 253                 .dst_sz = sizeof(struct ibv_flow_spec_ipv6),
 254         },
 255         [RTE_FLOW_ITEM_TYPE_UDP] = {
 256                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_VXLAN),
 257                 .actions = valid_actions,
 258                 .mask = &(const struct rte_flow_item_udp){
 259                         .hdr = {
 260                                 .src_port = -1,
 261                                 .dst_port = -1,
 262                         },
 263                 },
 264                 .default_mask = &rte_flow_item_udp_mask,
 265                 .mask_sz = sizeof(struct rte_flow_item_udp),
 266                 .convert = mlx5_flow_create_udp,
 267                 .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
 268         },
 269         [RTE_FLOW_ITEM_TYPE_TCP] = {
 270                 .actions = valid_actions,
 271                 .mask = &(const struct rte_flow_item_tcp){
 272                         .hdr = {
 273                                 .src_port = -1,
 274                                 .dst_port = -1,
 275                         },
 276                 },
 277                 .default_mask = &rte_flow_item_tcp_mask,
 278                 .mask_sz = sizeof(struct rte_flow_item_tcp),
 279                 .convert = mlx5_flow_create_tcp,
 280                 .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
 281         },
 282         [RTE_FLOW_ITEM_TYPE_VXLAN] = {
 283                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_ETH),
 284                 .actions = valid_actions,
 285                 .mask = &(const struct rte_flow_item_vxlan){
 286                         .vni = "\xff\xff\xff",
 287                 },
 288                 .default_mask = &rte_flow_item_vxlan_mask,
 289                 .mask_sz = sizeof(struct rte_flow_item_vxlan),
 290                 .convert = mlx5_flow_create_vxlan,
 291                 .dst_sz = sizeof(struct ibv_flow_spec_tunnel),
 292         },
 293 };
 294
 295 /** Structure to pass to the conversion function. */
 296 struct mlx5_flow_parse {
 297         struct ibv_flow_attr *ibv_attr; /**< Verbs attribute. */
 298         unsigned int offset; /**< Offset in bytes in the ibv_attr buffer. */
 299         uint32_t inner; /**< Set once VXLAN is encountered. */
 300         uint32_t create:1; /**< Leave allocated resources on exit. */
 301         uint32_t queue:1; /**< Target is a receive queue. */
 302         uint32_t drop:1; /**< Target is a drop queue. */
 303         uint32_t mark:1; /**< Mark is present in the flow. */
 304         uint32_t mark_id; /**< Mark identifier. */
 305         uint64_t hash_fields; /**< Fields that participate in the hash. */
 306         uint16_t queues[RTE_MAX_QUEUES_PER_PORT]; /**< Queues indexes to use. */
 307         uint16_t queues_n; /**< Number of entries in queue[]. */
 308 };
 309
 310 static const struct rte_flow_ops mlx5_flow_ops = {
 311         .validate = mlx5_flow_validate,
 312         .create = mlx5_flow_create,
 313         .destroy = mlx5_flow_destroy,
 314         .flush = mlx5_flow_flush,
 315         .query = NULL,
 316         .isolate = mlx5_flow_isolate,
 317 };
 318
 319 /**
 320  * Manage filter operations.
 321  *
 322  * @param dev
 323  *   Pointer to Ethernet device structure.
 324  * @param filter_type
 325  *   Filter type.
 326  * @param filter_op
 327  *   Operation to perform.
 328  * @param arg
 329  *   Pointer to operation-specific structure.
 330  *
 331  * @return
 332  *   0 on success, negative errno value on failure.
 333  */
 334 int
 335 mlx5_dev_filter_ctrl(struct rte_eth_dev *dev,
 336                      enum rte_filter_type filter_type,
 337                      enum rte_filter_op filter_op,
 338                      void *arg)
 339 {
 340         int ret = EINVAL;
 341
 342         if (filter_type == RTE_ETH_FILTER_GENERIC) {
 343                 if (filter_op != RTE_ETH_FILTER_GET)
 344                         return -EINVAL;
 345                 *(const void **)arg = &mlx5_flow_ops;
 346                 return 0;
 347         }
 348         ERROR("%p: filter type (%d) not supported",
 349               (void *)dev, filter_type);
 350         return -ret;
 351 }
 352
 353 /**
 354  * Check support for a given item.
 355  *
 356  * @param item[in]
 357  *   Item specification.
 358  * @param mask[in]
 359  *   Bit-masks covering supported fields to compare with spec, last and mask in
 360  *   \item.
 361  * @param size
 362  *   Bit-Mask size in bytes.
 363  *
 364  * @return
 365  *   0 on success.
 366  */
 367 static int
 368 mlx5_flow_item_validate(const struct rte_flow_item *item,
 369                         const uint8_t *mask, unsigned int size)
 370 {
 371         int ret = 0;
 372
 373         if (!item->spec && (item->mask || item->last))
 374                 return -1;
 375         if (item->spec && !item->mask) {
 376                 unsigned int i;
 377                 const uint8_t *spec = item->spec;
 378
 379                 for (i = 0; i < size; ++i)
 380                         if ((spec[i] | mask[i]) != mask[i])
 381                                 return -1;
 382         }
 383         if (item->last && !item->mask) {
 384                 unsigned int i;
 385                 const uint8_t *spec = item->last;
 386
 387                 for (i = 0; i < size; ++i)
 388                         if ((spec[i] | mask[i]) != mask[i])
 389                                 return -1;
 390         }
 391         if (item->mask) {
 392                 unsigned int i;
 393                 const uint8_t *spec = item->mask;
 394
 395                 for (i = 0; i < size; ++i)
 396                         if ((spec[i] | mask[i]) != mask[i])
 397                                 return -1;
 398         }
 399         if (item->spec && item->last) {
 400                 uint8_t spec[size];
 401                 uint8_t last[size];
 402                 const uint8_t *apply = mask;
 403                 unsigned int i;
 404
 405                 if (item->mask)
 406                         apply = item->mask;
 407                 for (i = 0; i < size; ++i) {
 408                         spec[i] = ((const uint8_t *)item->spec)[i] & apply[i];
 409                         last[i] = ((const uint8_t *)item->last)[i] & apply[i];
 410                 }
 411                 ret = memcmp(spec, last, size);
 412         }
 413         return ret;
 414 }
 415
 416 /**
 417  * Validate and convert a flow supported by the NIC.
 418  *
 419  * @param priv
 420  *   Pointer to private structure.
 421  * @param[in] attr
 422  *   Flow rule attributes.
 423  * @param[in] pattern
 424  *   Pattern specification (list terminated by the END pattern item).
 425  * @param[in] actions
 426  *   Associated actions (list terminated by the END action).
 427  * @param[out] error
 428  *   Perform verbose error reporting if not NULL.
 429  * @param[in, out] parser
 430  *   Internal parser structure.
 431  *
 432  * @return
 433  *   0 on success, a negative errno value otherwise and rte_errno is set.
 434  */
 435 static int
 436 priv_flow_convert(struct priv *priv,
 437                   const struct rte_flow_attr *attr,
 438                   const struct rte_flow_item items[],
 439                   const struct rte_flow_action actions[],
 440                   struct rte_flow_error *error,
 441                   struct mlx5_flow_parse *parser)
 442 {
 443         const struct mlx5_flow_items *cur_item = mlx5_flow_items;
 444
 445         (void)priv;
 446         *parser = (struct mlx5_flow_parse){
 447                 .ibv_attr = parser->ibv_attr,
 448                 .create = parser->create,
 449                 .offset = sizeof(struct ibv_flow_attr),
 450                 .mark_id = MLX5_FLOW_MARK_DEFAULT,
 451         };
 452         if (attr->group) {
 453                 rte_flow_error_set(error, ENOTSUP,
 454                                    RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
 455                                    NULL,
 456                                    "groups are not supported");
 457                 return -rte_errno;
 458         }
 459         if (attr->priority && attr->priority != MLX5_CTRL_FLOW_PRIORITY) {
 460                 rte_flow_error_set(error, ENOTSUP,
 461                                    RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
 462                                    NULL,
 463                                    "priorities are not supported");
 464                 return -rte_errno;
 465         }
 466         if (attr->egress) {
 467                 rte_flow_error_set(error, ENOTSUP,
 468                                    RTE_FLOW_ERROR_TYPE_ATTR_EGRESS,
 469                                    NULL,
 470                                    "egress is not supported");
 471                 return -rte_errno;
 472         }
 473         if (!attr->ingress) {
 474                 rte_flow_error_set(error, ENOTSUP,
 475                                    RTE_FLOW_ERROR_TYPE_ATTR_INGRESS,
 476                                    NULL,
 477                                    "only ingress is supported");
 478                 return -rte_errno;
 479         }
 480         for (; actions->type != RTE_FLOW_ACTION_TYPE_END; ++actions) {
 481                 if (actions->type == RTE_FLOW_ACTION_TYPE_VOID) {
 482                         continue;
 483                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_DROP) {
 484                         parser->drop = 1;
 485                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_QUEUE) {
 486                         const struct rte_flow_action_queue *queue =
 487                                 (const struct rte_flow_action_queue *)
 488                                 actions->conf;
 489                         uint16_t n;
 490                         uint16_t found = 0;
 491
 492                         if (!queue || (queue->index > (priv->rxqs_n - 1)))
 493                                 goto exit_action_not_supported;
 494                         for (n = 0; n < parser->queues_n; ++n) {
 495                                 if (parser->queues[n] == queue->index) {
 496                                         found = 1;
 497                                         break;
 498                                 }
 499                         }
 500                         if (parser->queues_n > 1 && !found) {
 501                                 rte_flow_error_set(error, ENOTSUP,
 502                                            RTE_FLOW_ERROR_TYPE_ACTION,
 503                                            actions,
 504                                            "queue action not in RSS queues");
 505                                 return -rte_errno;
 506                         }
 507                         if (!found) {
 508                                 parser->queue = 1;
 509                                 parser->queues_n = 1;
 510                                 parser->queues[0] = queue->index;
 511                         }
 512                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_RSS) {
 513                         const struct rte_flow_action_rss *rss =
 514                                 (const struct rte_flow_action_rss *)
 515                                 actions->conf;
 516                         uint16_t n;
 517
 518                         if (!rss || !rss->num) {
 519                                 rte_flow_error_set(error, EINVAL,
 520                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 521                                                    actions,
 522                                                    "no valid queues");
 523                                 return -rte_errno;
 524                         }
 525                         if (parser->queues_n == 1) {
 526                                 uint16_t found = 0;
 527
 528                                 assert(parser->queues_n);
 529                                 for (n = 0; n < rss->num; ++n) {
 530                                         if (parser->queues[0] ==
 531                                             rss->queue[n]) {
 532                                                 found = 1;
 533                                                 break;
 534                                         }
 535                                 }
 536                                 if (!found) {
 537                                         rte_flow_error_set(error, ENOTSUP,
 538                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 539                                                    actions,
 540                                                    "queue action not in RSS"
 541                                                    " queues");
 542                                         return -rte_errno;
 543                                 }
 544                         }
 545                         for (n = 0; n < rss->num; ++n) {
 546                                 if (rss->queue[n] >= priv->rxqs_n) {
 547                                         rte_flow_error_set(error, EINVAL,
 548                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 549                                                    actions,
 550                                                    "queue id > number of"
 551                                                    " queues");
 552                                         return -rte_errno;
 553                                 }
 554                         }
 555                         parser->queue = 1;
 556                         for (n = 0; n < rss->num; ++n)
 557                                 parser->queues[n] = rss->queue[n];
 558                         parser->queues_n = rss->num;
 559                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_MARK) {
 560                         const struct rte_flow_action_mark *mark =
 561                                 (const struct rte_flow_action_mark *)
 562                                 actions->conf;
 563
 564                         if (!mark) {
 565                                 rte_flow_error_set(error, EINVAL,
 566                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 567                                                    actions,
 568                                                    "mark must be defined");
 569                                 return -rte_errno;
 570                         } else if (mark->id >= MLX5_FLOW_MARK_MAX) {
 571                                 rte_flow_error_set(error, ENOTSUP,
 572                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 573                                                    actions,
 574                                                    "mark must be between 0"
 575                                                    " and 16777199");
 576                                 return -rte_errno;
 577                         }
 578                         parser->mark = 1;
 579                         parser->mark_id = mark->id;
 580                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_FLAG) {
 581                         parser->mark = 1;
 582                 } else {
 583                         goto exit_action_not_supported;
 584                 }
 585         }
 586         if (parser->mark && !parser->ibv_attr && !parser->drop)
 587                 parser->offset += sizeof(struct ibv_flow_spec_action_tag);
 588         if (!parser->ibv_attr && parser->drop)
 589                 parser->offset += sizeof(struct ibv_flow_spec_action_drop);
 590         if (!parser->queue && !parser->drop) {
 591                 rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_HANDLE,
 592                                    NULL, "no valid action");
 593                 return -rte_errno;
 594         }
 595         for (; items->type != RTE_FLOW_ITEM_TYPE_END; ++items) {
 596                 const struct mlx5_flow_items *token = NULL;
 597                 unsigned int i;
 598                 int err;
 599
 600                 if (items->type == RTE_FLOW_ITEM_TYPE_VOID)
 601                         continue;
 602                 for (i = 0;
 603                      cur_item->items &&
 604                      cur_item->items[i] != RTE_FLOW_ITEM_TYPE_END;
 605                      ++i) {
 606                         if (cur_item->items[i] == items->type) {
 607                                 token = &mlx5_flow_items[items->type];
 608                                 break;
 609                         }
 610                 }
 611                 if (!token)
 612                         goto exit_item_not_supported;
 613                 cur_item = token;
 614                 err = mlx5_flow_item_validate(items,
 615                                               (const uint8_t *)cur_item->mask,
 616                                               cur_item->mask_sz);
 617                 if (err)
 618                         goto exit_item_not_supported;
 619                 if (parser->ibv_attr && cur_item->convert) {
 620                         err = cur_item->convert(items,
 621                                                 (cur_item->default_mask ?
 622                                                  cur_item->default_mask :
 623                                                  cur_item->mask),
 624                                                 parser);
 625                         if (err)
 626                                 goto exit_item_not_supported;
 627                 } else if (items->type == RTE_FLOW_ITEM_TYPE_VXLAN) {
 628                         if (parser->inner) {
 629                                 rte_flow_error_set(error, ENOTSUP,
 630                                                    RTE_FLOW_ERROR_TYPE_ITEM,
 631                                                    items,
 632                                                    "cannot recognize multiple"
 633                                                    " VXLAN encapsulations");
 634                                 return -rte_errno;
 635                         }
 636                         parser->inner = 1;
 637                 }
 638                 parser->offset += cur_item->dst_sz;
 639         }
 640         return 0;
 641 exit_item_not_supported:
 642         rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 643                            items, "item not supported");
 644         return -rte_errno;
 645 exit_action_not_supported:
 646         rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
 647                            actions, "action not supported");
 648         return -rte_errno;
 649 }
 650
 651 /**
 652  * Convert Ethernet item to Verbs specification.
 653  *
 654  * @param item[in]
 655  *   Item specification.
 656  * @param default_mask[in]
 657  *   Default bit-masks to use when item->mask is not provided.
 658  * @param data[in, out]
 659  *   User structure.
 660  */
 661 static int
 662 mlx5_flow_create_eth(const struct rte_flow_item *item,
 663                      const void *default_mask,
 664                      void *data)
 665 {
 666         const struct rte_flow_item_eth *spec = item->spec;
 667         const struct rte_flow_item_eth *mask = item->mask;
 668         struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data;
 669         struct ibv_flow_spec_eth *eth;
 670         const unsigned int eth_size = sizeof(struct ibv_flow_spec_eth);
 671         unsigned int i;
 672
 673         ++parser->ibv_attr->num_of_specs;
 674         parser->hash_fields = 0;
 675         eth = (void *)((uintptr_t)parser->ibv_attr + parser->offset);
 676         *eth = (struct ibv_flow_spec_eth) {
 677                 .type = parser->inner | IBV_FLOW_SPEC_ETH,
 678                 .size = eth_size,
 679         };
 680         if (!spec)
 681                 return 0;
 682         if (!mask)
 683                 mask = default_mask;
 684         memcpy(eth->val.dst_mac, spec->dst.addr_bytes, ETHER_ADDR_LEN);
 685         memcpy(eth->val.src_mac, spec->src.addr_bytes, ETHER_ADDR_LEN);
 686         eth->val.ether_type = spec->type;
 687         memcpy(eth->mask.dst_mac, mask->dst.addr_bytes, ETHER_ADDR_LEN);
 688         memcpy(eth->mask.src_mac, mask->src.addr_bytes, ETHER_ADDR_LEN);
 689         eth->mask.ether_type = mask->type;
 690         /* Remove unwanted bits from values. */
 691         for (i = 0; i < ETHER_ADDR_LEN; ++i) {
 692                 eth->val.dst_mac[i] &= eth->mask.dst_mac[i];
 693                 eth->val.src_mac[i] &= eth->mask.src_mac[i];
 694         }
 695         eth->val.ether_type &= eth->mask.ether_type;
 696         return 0;
 697 }
 698
 699 /**
 700  * Convert VLAN item to Verbs specification.
 701  *
 702  * @param item[in]
 703  *   Item specification.
 704  * @param default_mask[in]
 705  *   Default bit-masks to use when item->mask is not provided.
 706  * @param data[in, out]
 707  *   User structure.
 708  */
 709 static int
 710 mlx5_flow_create_vlan(const struct rte_flow_item *item,
 711                       const void *default_mask,
 712                       void *data)
 713 {
 714         const struct rte_flow_item_vlan *spec = item->spec;
 715         const struct rte_flow_item_vlan *mask = item->mask;
 716         struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data;
 717         struct ibv_flow_spec_eth *eth;
 718         const unsigned int eth_size = sizeof(struct ibv_flow_spec_eth);
 719
 720         eth = (void *)((uintptr_t)parser->ibv_attr + parser->offset - eth_size);
 721         if (!spec)
 722                 return 0;
 723         if (!mask)
 724                 mask = default_mask;
 725         eth->val.vlan_tag = spec->tci;
 726         eth->mask.vlan_tag = mask->tci;
 727         eth->val.vlan_tag &= eth->mask.vlan_tag;
 728         return 0;
 729 }
 730
 731 /**
 732  * Convert IPv4 item to Verbs specification.
 733  *
 734  * @param item[in]
 735  *   Item specification.
 736  * @param default_mask[in]
 737  *   Default bit-masks to use when item->mask is not provided.
 738  * @param data[in, out]
 739  *   User structure.
 740  */
 741 static int
 742 mlx5_flow_create_ipv4(const struct rte_flow_item *item,
 743                       const void *default_mask,
 744                       void *data)
 745 {
 746         const struct rte_flow_item_ipv4 *spec = item->spec;
 747         const struct rte_flow_item_ipv4 *mask = item->mask;
 748         struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data;
 749         struct ibv_flow_spec_ipv4_ext *ipv4;
 750         unsigned int ipv4_size = sizeof(struct ibv_flow_spec_ipv4_ext);
 751
 752         ++parser->ibv_attr->num_of_specs;
 753         parser->hash_fields = (IBV_RX_HASH_SRC_IPV4 | IBV_RX_HASH_DST_IPV4);
 754         ipv4 = (void *)((uintptr_t)parser->ibv_attr + parser->offset);
 755         *ipv4 = (struct ibv_flow_spec_ipv4_ext) {
 756                 .type = parser->inner | IBV_FLOW_SPEC_IPV4_EXT,
 757                 .size = ipv4_size,
 758         };
 759         if (!spec)
 760                 return 0;
 761         if (!mask)
 762                 mask = default_mask;
 763         ipv4->val = (struct ibv_flow_ipv4_ext_filter){
 764                 .src_ip = spec->hdr.src_addr,
 765                 .dst_ip = spec->hdr.dst_addr,
 766                 .proto = spec->hdr.next_proto_id,
 767                 .tos = spec->hdr.type_of_service,
 768         };
 769         ipv4->mask = (struct ibv_flow_ipv4_ext_filter){
 770                 .src_ip = mask->hdr.src_addr,
 771                 .dst_ip = mask->hdr.dst_addr,
 772                 .proto = mask->hdr.next_proto_id,
 773                 .tos = mask->hdr.type_of_service,
 774         };
 775         /* Remove unwanted bits from values. */
 776         ipv4->val.src_ip &= ipv4->mask.src_ip;
 777         ipv4->val.dst_ip &= ipv4->mask.dst_ip;
 778         ipv4->val.proto &= ipv4->mask.proto;
 779         ipv4->val.tos &= ipv4->mask.tos;
 780         return 0;
 781 }
 782
 783 /**
 784  * Convert IPv6 item to Verbs specification.
 785  *
 786  * @param item[in]
 787  *   Item specification.
 788  * @param default_mask[in]
 789  *   Default bit-masks to use when item->mask is not provided.
 790  * @param data[in, out]
 791  *   User structure.
 792  */
 793 static int
 794 mlx5_flow_create_ipv6(const struct rte_flow_item *item,
 795                       const void *default_mask,
 796                       void *data)
 797 {
 798         const struct rte_flow_item_ipv6 *spec = item->spec;
 799         const struct rte_flow_item_ipv6 *mask = item->mask;
 800         struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data;
 801         struct ibv_flow_spec_ipv6 *ipv6;
 802         unsigned int ipv6_size = sizeof(struct ibv_flow_spec_ipv6);
 803         unsigned int i;
 804
 805         ++parser->ibv_attr->num_of_specs;
 806         parser->hash_fields = (IBV_RX_HASH_SRC_IPV6 | IBV_RX_HASH_DST_IPV6);
 807         ipv6 = (void *)((uintptr_t)parser->ibv_attr + parser->offset);
 808         *ipv6 = (struct ibv_flow_spec_ipv6) {
 809                 .type = parser->inner | IBV_FLOW_SPEC_IPV6,
 810                 .size = ipv6_size,
 811         };
 812         if (!spec)
 813                 return 0;
 814         if (!mask)
 815                 mask = default_mask;
 816         memcpy(ipv6->val.src_ip, spec->hdr.src_addr,
 817                RTE_DIM(ipv6->val.src_ip));
 818         memcpy(ipv6->val.dst_ip, spec->hdr.dst_addr,
 819                RTE_DIM(ipv6->val.dst_ip));
 820         memcpy(ipv6->mask.src_ip, mask->hdr.src_addr,
 821                RTE_DIM(ipv6->mask.src_ip));
 822         memcpy(ipv6->mask.dst_ip, mask->hdr.dst_addr,
 823                RTE_DIM(ipv6->mask.dst_ip));
 824         ipv6->mask.flow_label = mask->hdr.vtc_flow;
 825         ipv6->mask.next_hdr = mask->hdr.proto;
 826         ipv6->mask.hop_limit = mask->hdr.hop_limits;
 827         /* Remove unwanted bits from values. */
 828         for (i = 0; i < RTE_DIM(ipv6->val.src_ip); ++i) {
 829                 ipv6->val.src_ip[i] &= ipv6->mask.src_ip[i];
 830                 ipv6->val.dst_ip[i] &= ipv6->mask.dst_ip[i];
 831         }
 832         ipv6->val.flow_label &= ipv6->mask.flow_label;
 833         ipv6->val.next_hdr &= ipv6->mask.next_hdr;
 834         ipv6->val.hop_limit &= ipv6->mask.hop_limit;
 835         return 0;
 836 }
 837
 838 /**
 839  * Convert UDP item to Verbs specification.
 840  *
 841  * @param item[in]
 842  *   Item specification.
 843  * @param default_mask[in]
 844  *   Default bit-masks to use when item->mask is not provided.
 845  * @param data[in, out]
 846  *   User structure.
 847  */
 848 static int
 849 mlx5_flow_create_udp(const struct rte_flow_item *item,
 850                      const void *default_mask,
 851                      void *data)
 852 {
 853         const struct rte_flow_item_udp *spec = item->spec;
 854         const struct rte_flow_item_udp *mask = item->mask;
 855         struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data;
 856         struct ibv_flow_spec_tcp_udp *udp;
 857         unsigned int udp_size = sizeof(struct ibv_flow_spec_tcp_udp);
 858
 859         ++parser->ibv_attr->num_of_specs;
 860         parser->hash_fields |= (IBV_RX_HASH_SRC_PORT_UDP |
 861                                 IBV_RX_HASH_DST_PORT_UDP);
 862         udp = (void *)((uintptr_t)parser->ibv_attr + parser->offset);
 863         *udp = (struct ibv_flow_spec_tcp_udp) {
 864                 .type = parser->inner | IBV_FLOW_SPEC_UDP,
 865                 .size = udp_size,
 866         };
 867         if (!spec)
 868                 return 0;
 869         if (!mask)
 870                 mask = default_mask;
 871         udp->val.dst_port = spec->hdr.dst_port;
 872         udp->val.src_port = spec->hdr.src_port;
 873         udp->mask.dst_port = mask->hdr.dst_port;
 874         udp->mask.src_port = mask->hdr.src_port;
 875         /* Remove unwanted bits from values. */
 876         udp->val.src_port &= udp->mask.src_port;
 877         udp->val.dst_port &= udp->mask.dst_port;
 878         return 0;
 879 }
 880
 881 /**
 882  * Convert TCP item to Verbs specification.
 883  *
 884  * @param item[in]
 885  *   Item specification.
 886  * @param default_mask[in]
 887  *   Default bit-masks to use when item->mask is not provided.
 888  * @param data[in, out]
 889  *   User structure.
 890  */
 891 static int
 892 mlx5_flow_create_tcp(const struct rte_flow_item *item,
 893                      const void *default_mask,
 894                      void *data)
 895 {
 896         const struct rte_flow_item_tcp *spec = item->spec;
 897         const struct rte_flow_item_tcp *mask = item->mask;
 898         struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data;
 899         struct ibv_flow_spec_tcp_udp *tcp;
 900         unsigned int tcp_size = sizeof(struct ibv_flow_spec_tcp_udp);
 901
 902         ++parser->ibv_attr->num_of_specs;
 903         parser->hash_fields |= (IBV_RX_HASH_SRC_PORT_TCP |
 904                                 IBV_RX_HASH_DST_PORT_TCP);
 905         tcp = (void *)((uintptr_t)parser->ibv_attr + parser->offset);
 906         *tcp = (struct ibv_flow_spec_tcp_udp) {
 907                 .type = parser->inner | IBV_FLOW_SPEC_TCP,
 908                 .size = tcp_size,
 909         };
 910         if (!spec)
 911                 return 0;
 912         if (!mask)
 913                 mask = default_mask;
 914         tcp->val.dst_port = spec->hdr.dst_port;
 915         tcp->val.src_port = spec->hdr.src_port;
 916         tcp->mask.dst_port = mask->hdr.dst_port;
 917         tcp->mask.src_port = mask->hdr.src_port;
 918         /* Remove unwanted bits from values. */
 919         tcp->val.src_port &= tcp->mask.src_port;
 920         tcp->val.dst_port &= tcp->mask.dst_port;
 921         return 0;
 922 }
 923
 924 /**
 925  * Convert VXLAN item to Verbs specification.
 926  *
 927  * @param item[in]
 928  *   Item specification.
 929  * @param default_mask[in]
 930  *   Default bit-masks to use when item->mask is not provided.
 931  * @param data[in, out]
 932  *   User structure.
 933  */
 934 static int
 935 mlx5_flow_create_vxlan(const struct rte_flow_item *item,
 936                        const void *default_mask,
 937                        void *data)
 938 {
 939         const struct rte_flow_item_vxlan *spec = item->spec;
 940         const struct rte_flow_item_vxlan *mask = item->mask;
 941         struct mlx5_flow_parse *parser = (struct mlx5_flow_parse *)data;
 942         struct ibv_flow_spec_tunnel *vxlan;
 943         unsigned int size = sizeof(struct ibv_flow_spec_tunnel);
 944         union vni {
 945                 uint32_t vlan_id;
 946                 uint8_t vni[4];
 947         } id;
 948
 949         ++parser->ibv_attr->num_of_specs;
 950         id.vni[0] = 0;
 951         vxlan = (void *)((uintptr_t)parser->ibv_attr + parser->offset);
 952         *vxlan = (struct ibv_flow_spec_tunnel) {
 953                 .type = parser->inner | IBV_FLOW_SPEC_VXLAN_TUNNEL,
 954                 .size = size,
 955         };
 956         parser->inner = IBV_FLOW_SPEC_INNER;
 957         if (!spec)
 958                 return 0;
 959         if (!mask)
 960                 mask = default_mask;
 961         memcpy(&id.vni[1], spec->vni, 3);
 962         vxlan->val.tunnel_id = id.vlan_id;
 963         memcpy(&id.vni[1], mask->vni, 3);
 964         vxlan->mask.tunnel_id = id.vlan_id;
 965         /* Remove unwanted bits from values. */
 966         vxlan->val.tunnel_id &= vxlan->mask.tunnel_id;
 967         return 0;
 968 }
 969
 970 /**
 971  * Convert mark/flag action to Verbs specification.
 972  *
 973  * @param parser
 974  *   Internal parser structure.
 975  * @param mark_id
 976  *   Mark identifier.
 977  */
 978 static int
 979 mlx5_flow_create_flag_mark(struct mlx5_flow_parse *parser, uint32_t mark_id)
 980 {
 981         struct ibv_flow_spec_action_tag *tag;
 982         unsigned int size = sizeof(struct ibv_flow_spec_action_tag);
 983
 984         assert(parser->mark);
 985         tag = (void *)((uintptr_t)parser->ibv_attr + parser->offset);
 986         *tag = (struct ibv_flow_spec_action_tag){
 987                 .type = IBV_FLOW_SPEC_ACTION_TAG,
 988                 .size = size,
 989                 .tag_id = mlx5_flow_mark_set(mark_id),
 990         };
 991         ++parser->ibv_attr->num_of_specs;
 992         parser->offset += size;
 993         return 0;
 994 }
 995
 996 /**
 997  * Complete flow rule creation with a drop queue.
 998  *
 999  * @param priv
1000  *   Pointer to private structure.
1001  * @param parser
1002  *   Internal parser structure.
1003  * @param[out] error
1004  *   Perform verbose error reporting if not NULL.
1005  *
1006  * @return
1007  *   A flow if the rule could be created.
1008  */
1009 static struct rte_flow *
1010 priv_flow_create_action_queue_drop(struct priv *priv,
1011                                    struct mlx5_flow_parse *parser,
1012                                    struct rte_flow_error *error)
1013 {
1014         struct rte_flow *rte_flow;
1015         struct ibv_flow_spec_action_drop *drop;
1016         unsigned int size = sizeof(struct ibv_flow_spec_action_drop);
1017
1018         assert(priv->pd);
1019         assert(priv->ctx);
1020         rte_flow = rte_calloc(__func__, 1, sizeof(*rte_flow), 0);
1021         if (!rte_flow) {
1022                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1023                                    NULL, "cannot allocate flow memory");
1024                 return NULL;
1025         }
1026         rte_flow->drop = 1;
1027         drop = (void *)((uintptr_t)parser->ibv_attr + parser->offset);
1028         *drop = (struct ibv_flow_spec_action_drop){
1029                         .type = IBV_FLOW_SPEC_ACTION_DROP,
1030                         .size = size,
1031         };
1032         ++parser->ibv_attr->num_of_specs;
1033         parser->offset += sizeof(struct ibv_flow_spec_action_drop);
1034         rte_flow->ibv_attr = parser->ibv_attr;
1035         if (!priv->dev->data->dev_started)
1036                 return rte_flow;
1037         rte_flow->drxq.hrxq.qp = priv->flow_drop_queue->qp;
1038         rte_flow->ibv_flow = ibv_create_flow(rte_flow->drxq.hrxq.qp,
1039                                              rte_flow->ibv_attr);
1040         if (!rte_flow->ibv_flow) {
1041                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1042                                    NULL, "flow rule creation failure");
1043                 goto error;
1044         }
1045         return rte_flow;
1046 error:
1047         assert(rte_flow);
1048         rte_free(rte_flow);
1049         return NULL;
1050 }
1051
1052 /**
1053  * Complete flow rule creation.
1054  *
1055  * @param priv
1056  *   Pointer to private structure.
1057  * @param parser
1058  *   MLX5 flow parser attributes (filled by mlx5_flow_validate()).
1059  * @param[out] error
1060  *   Perform verbose error reporting if not NULL.
1061  *
1062  * @return
1063  *   A flow if the rule could be created.
1064  */
1065 static struct rte_flow *
1066 priv_flow_create_action_queue(struct priv *priv,
1067                               struct mlx5_flow_parse *parser,
1068                               struct rte_flow_error *error)
1069 {
1070         struct rte_flow *rte_flow;
1071         unsigned int i;
1072
1073         assert(priv->pd);
1074         assert(priv->ctx);
1075         assert(!parser->drop);
1076         rte_flow = rte_calloc(__func__, 1,
1077                               sizeof(*rte_flow) +
1078                               parser->queues_n * sizeof(uint16_t),
1079                               0);
1080         if (!rte_flow) {
1081                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1082                                    NULL, "cannot allocate flow memory");
1083                 return NULL;
1084         }
1085         rte_flow->mark = parser->mark;
1086         rte_flow->ibv_attr = parser->ibv_attr;
1087         rte_flow->queues = (uint16_t (*)[])(rte_flow + 1);
1088         memcpy(rte_flow->queues, parser->queues,
1089                parser->queues_n * sizeof(uint16_t));
1090         rte_flow->queues_n = parser->queues_n;
1091         rte_flow->frxq.hash_fields = parser->hash_fields;
1092         rte_flow->frxq.hrxq = mlx5_priv_hrxq_get(priv, rss_hash_default_key,
1093                                                  rss_hash_default_key_len,
1094                                                  parser->hash_fields,
1095                                                  (*rte_flow->queues),
1096                                                  rte_flow->queues_n);
1097         if (!rte_flow->frxq.hrxq) {
1098                 rte_flow->frxq.hrxq =
1099                         mlx5_priv_hrxq_new(priv, rss_hash_default_key,
1100                                            rss_hash_default_key_len,
1101                                            parser->hash_fields,
1102                                            (*rte_flow->queues),
1103                                            rte_flow->queues_n);
1104                 if (!rte_flow->frxq.hrxq) {
1105                         rte_flow_error_set(error, ENOMEM,
1106                                            RTE_FLOW_ERROR_TYPE_HANDLE,
1107                                            NULL, "cannot create hash rxq");
1108                         goto error;
1109                 }
1110         }
1111         for (i = 0; i != parser->queues_n; ++i) {
1112                 struct mlx5_rxq_data *q =
1113                         (*priv->rxqs)[parser->queues[i]];
1114
1115                 q->mark |= parser->mark;
1116         }
1117         if (!priv->dev->data->dev_started)
1118                 return rte_flow;
1119         rte_flow->ibv_flow = ibv_create_flow(rte_flow->frxq.hrxq->qp,
1120                                              rte_flow->ibv_attr);
1121         if (!rte_flow->ibv_flow) {
1122                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1123                                    NULL, "flow rule creation failure");
1124                 goto error;
1125         }
1126         return rte_flow;
1127 error:
1128         assert(rte_flow);
1129         if (rte_flow->frxq.hrxq)
1130                 mlx5_priv_hrxq_release(priv, rte_flow->frxq.hrxq);
1131         rte_free(rte_flow);
1132         return NULL;
1133 }
1134
1135 /**
1136  * Validate a flow.
1137  *
1138  * @param priv
1139  *   Pointer to private structure.
1140  * @param[in] attr
1141  *   Flow rule attributes.
1142  * @param[in] pattern
1143  *   Pattern specification (list terminated by the END pattern item).
1144  * @param[in] actions
1145  *   Associated actions (list terminated by the END action).
1146  * @param[out] error
1147  *   Perform verbose error reporting if not NULL.
1148  * @param[in,out] parser
1149  *   MLX5 parser structure.
1150  *
1151  * @return
1152  *   0 on success, negative errno value on failure.
1153  */
1154 static int
1155 priv_flow_validate(struct priv *priv,
1156                    const struct rte_flow_attr *attr,
1157                    const struct rte_flow_item items[],
1158                    const struct rte_flow_action actions[],
1159                    struct rte_flow_error *error,
1160                    struct mlx5_flow_parse *parser)
1161 {
1162         int err;
1163
1164         err = priv_flow_convert(priv, attr, items, actions, error, parser);
1165         if (err)
1166                 goto exit;
1167         if (parser->mark)
1168                 parser->offset += sizeof(struct ibv_flow_spec_action_tag);
1169         parser->ibv_attr = rte_malloc(__func__, parser->offset, 0);
1170         if (!parser->ibv_attr) {
1171                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1172                                    NULL, "cannot allocate ibv_attr memory");
1173                 err = rte_errno;
1174                 goto exit;
1175         }
1176         *parser->ibv_attr = (struct ibv_flow_attr){
1177                 .type = IBV_FLOW_ATTR_NORMAL,
1178                 .size = sizeof(struct ibv_flow_attr),
1179                 .priority = attr->priority,
1180                 .num_of_specs = 0,
1181                 .port = 0,
1182                 .flags = 0,
1183         };
1184         err = priv_flow_convert(priv, attr, items, actions, error, parser);
1185         if (err || parser->create)
1186                 goto exit;
1187         if (parser->mark)
1188                 mlx5_flow_create_flag_mark(parser, parser->mark_id);
1189         return 0;
1190 exit:
1191         if (parser->ibv_attr)
1192                 rte_free(parser->ibv_attr);
1193         return err;
1194 }
1195
1196 /**
1197  * Convert a flow.
1198  *
1199  * @param priv
1200  *   Pointer to private structure.
1201  * @param list
1202  *   Pointer to a TAILQ flow list.
1203  * @param[in] attr
1204  *   Flow rule attributes.
1205  * @param[in] pattern
1206  *   Pattern specification (list terminated by the END pattern item).
1207  * @param[in] actions
1208  *   Associated actions (list terminated by the END action).
1209  * @param[out] error
1210  *   Perform verbose error reporting if not NULL.
1211  *
1212  * @return
1213  *   A flow on success, NULL otherwise.
1214  */
1215 static struct rte_flow *
1216 priv_flow_create(struct priv *priv,
1217                  struct mlx5_flows *list,
1218                  const struct rte_flow_attr *attr,
1219                  const struct rte_flow_item items[],
1220                  const struct rte_flow_action actions[],
1221                  struct rte_flow_error *error)
1222 {
1223         struct mlx5_flow_parse parser = { .create = 1, };
1224         struct rte_flow *flow;
1225         int err;
1226
1227         err = priv_flow_validate(priv, attr, items, actions, error, &parser);
1228         if (err)
1229                 goto exit;
1230         if (parser.drop)
1231                 flow = priv_flow_create_action_queue_drop(priv, &parser, error);
1232         else
1233                 flow = priv_flow_create_action_queue(priv, &parser, error);
1234         if (!flow)
1235                 goto exit;
1236         TAILQ_INSERT_TAIL(list, flow, next);
1237         DEBUG("Flow created %p", (void *)flow);
1238         return flow;
1239 exit:
1240         if (parser.ibv_attr)
1241                 rte_free(parser.ibv_attr);
1242         return NULL;
1243 }
1244
1245 /**
1246  * Validate a flow supported by the NIC.
1247  *
1248  * @see rte_flow_validate()
1249  * @see rte_flow_ops
1250  */
1251 int
1252 mlx5_flow_validate(struct rte_eth_dev *dev,
1253                    const struct rte_flow_attr *attr,
1254                    const struct rte_flow_item items[],
1255                    const struct rte_flow_action actions[],
1256                    struct rte_flow_error *error)
1257 {
1258         struct priv *priv = dev->data->dev_private;
1259         int ret;
1260         struct mlx5_flow_parse parser = { .create = 0, };
1261
1262         priv_lock(priv);
1263         ret = priv_flow_validate(priv, attr, items, actions, error, &parser);
1264         priv_unlock(priv);
1265         return ret;
1266 }
1267
1268 /**
1269  * Create a flow.
1270  *
1271  * @see rte_flow_create()
1272  * @see rte_flow_ops
1273  */
1274 struct rte_flow *
1275 mlx5_flow_create(struct rte_eth_dev *dev,
1276                  const struct rte_flow_attr *attr,
1277                  const struct rte_flow_item items[],
1278                  const struct rte_flow_action actions[],
1279                  struct rte_flow_error *error)
1280 {
1281         struct priv *priv = dev->data->dev_private;
1282         struct rte_flow *flow;
1283
1284         priv_lock(priv);
1285         flow = priv_flow_create(priv, &priv->flows, attr, items, actions,
1286                                 error);
1287         priv_unlock(priv);
1288         return flow;
1289 }
1290
1291 /**
1292  * Destroy a flow.
1293  *
1294  * @param priv
1295  *   Pointer to private structure.
1296  * @param list
1297  *   Pointer to a TAILQ flow list.
1298  * @param[in] flow
1299  *   Flow to destroy.
1300  */
1301 static void
1302 priv_flow_destroy(struct priv *priv,
1303                   struct mlx5_flows *list,
1304                   struct rte_flow *flow)
1305 {
1306         unsigned int i;
1307         uint16_t *queues;
1308         uint16_t queues_n;
1309
1310         if (flow->drop || !flow->mark)
1311                 goto free;
1312         queues = flow->frxq.hrxq->ind_table->queues;
1313         queues_n = flow->frxq.hrxq->ind_table->queues_n;
1314         for (i = 0; i != queues_n; ++i) {
1315                 struct rte_flow *tmp;
1316                 struct mlx5_rxq_data *rxq_data = (*priv->rxqs)[queues[i]];
1317                 int mark = 0;
1318
1319                 /*
1320                  * To remove the mark from the queue, the queue must not be
1321                  * present in any other marked flow (RSS or not).
1322                  */
1323                 TAILQ_FOREACH(tmp, list, next) {
1324                         unsigned int j;
1325
1326                         if (!tmp->mark)
1327                                 continue;
1328                         for (j = 0;
1329                              (j != tmp->frxq.hrxq->ind_table->queues_n) &&
1330                              !mark;
1331                              j++)
1332                                 if (tmp->frxq.hrxq->ind_table->queues[j] ==
1333                                     queues[i])
1334                                         mark = 1;
1335                 }
1336                 rxq_data->mark = mark;
1337         }
1338 free:
1339         if (flow->ibv_flow)
1340                 claim_zero(ibv_destroy_flow(flow->ibv_flow));
1341         if (!flow->drop)
1342                 mlx5_priv_hrxq_release(priv, flow->frxq.hrxq);
1343         TAILQ_REMOVE(list, flow, next);
1344         rte_free(flow->ibv_attr);
1345         DEBUG("Flow destroyed %p", (void *)flow);
1346         rte_free(flow);
1347 }
1348
1349 /**
1350  * Destroy a flow.
1351  *
1352  * @see rte_flow_destroy()
1353  * @see rte_flow_ops
1354  */
1355 int
1356 mlx5_flow_destroy(struct rte_eth_dev *dev,
1357                   struct rte_flow *flow,
1358                   struct rte_flow_error *error)
1359 {
1360         struct priv *priv = dev->data->dev_private;
1361
1362         (void)error;
1363         priv_lock(priv);
1364         priv_flow_destroy(priv, &priv->flows, flow);
1365         priv_unlock(priv);
1366         return 0;
1367 }
1368
1369 /**
1370  * Destroy all flows.
1371  *
1372  * @param priv
1373  *   Pointer to private structure.
1374  * @param list
1375  *   Pointer to a TAILQ flow list.
1376  */
1377 void
1378 priv_flow_flush(struct priv *priv, struct mlx5_flows *list)
1379 {
1380         while (!TAILQ_EMPTY(list)) {
1381                 struct rte_flow *flow;
1382
1383                 flow = TAILQ_FIRST(list);
1384                 priv_flow_destroy(priv, list, flow);
1385         }
1386 }
1387
1388 /**
1389  * Destroy all flows.
1390  *
1391  * @see rte_flow_flush()
1392  * @see rte_flow_ops
1393  */
1394 int
1395 mlx5_flow_flush(struct rte_eth_dev *dev,
1396                 struct rte_flow_error *error)
1397 {
1398         struct priv *priv = dev->data->dev_private;
1399
1400         (void)error;
1401         priv_lock(priv);
1402         priv_flow_flush(priv, &priv->flows);
1403         priv_unlock(priv);
1404         return 0;
1405 }
1406
1407 /**
1408  * Create drop queue.
1409  *
1410  * @param priv
1411  *   Pointer to private structure.
1412  *
1413  * @return
1414  *   0 on success.
1415  */
1416 int
1417 priv_flow_create_drop_queue(struct priv *priv)
1418 {
1419         struct mlx5_hrxq_drop *fdq = NULL;
1420
1421         assert(priv->pd);
1422         assert(priv->ctx);
1423         fdq = rte_calloc(__func__, 1, sizeof(*fdq), 0);
1424         if (!fdq) {
1425                 WARN("cannot allocate memory for drop queue");
1426                 goto error;
1427         }
1428         fdq->cq = ibv_create_cq(priv->ctx, 1, NULL, NULL, 0);
1429         if (!fdq->cq) {
1430                 WARN("cannot allocate CQ for drop queue");
1431                 goto error;
1432         }
1433         fdq->wq = ibv_create_wq(priv->ctx,
1434                         &(struct ibv_wq_init_attr){
1435                         .wq_type = IBV_WQT_RQ,
1436                         .max_wr = 1,
1437                         .max_sge = 1,
1438                         .pd = priv->pd,
1439                         .cq = fdq->cq,
1440                         });
1441         if (!fdq->wq) {
1442                 WARN("cannot allocate WQ for drop queue");
1443                 goto error;
1444         }
1445         fdq->ind_table = ibv_create_rwq_ind_table(priv->ctx,
1446                         &(struct ibv_rwq_ind_table_init_attr){
1447                         .log_ind_tbl_size = 0,
1448                         .ind_tbl = &fdq->wq,
1449                         .comp_mask = 0,
1450                         });
1451         if (!fdq->ind_table) {
1452                 WARN("cannot allocate indirection table for drop queue");
1453                 goto error;
1454         }
1455         fdq->qp = ibv_create_qp_ex(priv->ctx,
1456                 &(struct ibv_qp_init_attr_ex){
1457                         .qp_type = IBV_QPT_RAW_PACKET,
1458                         .comp_mask =
1459                                 IBV_QP_INIT_ATTR_PD |
1460                                 IBV_QP_INIT_ATTR_IND_TABLE |
1461                                 IBV_QP_INIT_ATTR_RX_HASH,
1462                         .rx_hash_conf = (struct ibv_rx_hash_conf){
1463                                 .rx_hash_function =
1464                                         IBV_RX_HASH_FUNC_TOEPLITZ,
1465                                 .rx_hash_key_len = rss_hash_default_key_len,
1466                                 .rx_hash_key = rss_hash_default_key,
1467                                 .rx_hash_fields_mask = 0,
1468                                 },
1469                         .rwq_ind_tbl = fdq->ind_table,
1470                         .pd = priv->pd
1471                 });
1472         if (!fdq->qp) {
1473                 WARN("cannot allocate QP for drop queue");
1474                 goto error;
1475         }
1476         priv->flow_drop_queue = fdq;
1477         return 0;
1478 error:
1479         if (fdq->qp)
1480                 claim_zero(ibv_destroy_qp(fdq->qp));
1481         if (fdq->ind_table)
1482                 claim_zero(ibv_destroy_rwq_ind_table(fdq->ind_table));
1483         if (fdq->wq)
1484                 claim_zero(ibv_destroy_wq(fdq->wq));
1485         if (fdq->cq)
1486                 claim_zero(ibv_destroy_cq(fdq->cq));
1487         if (fdq)
1488                 rte_free(fdq);
1489         priv->flow_drop_queue = NULL;
1490         return -1;
1491 }
1492
1493 /**
1494  * Delete drop queue.
1495  *
1496  * @param priv
1497  *   Pointer to private structure.
1498  */
1499 void
1500 priv_flow_delete_drop_queue(struct priv *priv)
1501 {
1502         struct mlx5_hrxq_drop *fdq = priv->flow_drop_queue;
1503
1504         if (!fdq)
1505                 return;
1506         if (fdq->qp)
1507                 claim_zero(ibv_destroy_qp(fdq->qp));
1508         if (fdq->ind_table)
1509                 claim_zero(ibv_destroy_rwq_ind_table(fdq->ind_table));
1510         if (fdq->wq)
1511                 claim_zero(ibv_destroy_wq(fdq->wq));
1512         if (fdq->cq)
1513                 claim_zero(ibv_destroy_cq(fdq->cq));
1514         rte_free(fdq);
1515         priv->flow_drop_queue = NULL;
1516 }
1517
1518 /**
1519  * Remove all flows.
1520  *
1521  * @param priv
1522  *   Pointer to private structure.
1523  * @param list
1524  *   Pointer to a TAILQ flow list.
1525  */
1526 void
1527 priv_flow_stop(struct priv *priv, struct mlx5_flows *list)
1528 {
1529         struct rte_flow *flow;
1530
1531         TAILQ_FOREACH_REVERSE(flow, list, mlx5_flows, next) {
1532                 claim_zero(ibv_destroy_flow(flow->ibv_flow));
1533                 flow->ibv_flow = NULL;
1534                 mlx5_priv_hrxq_release(priv, flow->frxq.hrxq);
1535                 flow->frxq.hrxq = NULL;
1536                 if (flow->mark) {
1537                         unsigned int n;
1538                         struct mlx5_ind_table_ibv *ind_tbl =
1539                                 flow->frxq.hrxq->ind_table;
1540
1541                         for (n = 0; n < ind_tbl->queues_n; ++n)
1542                                 (*priv->rxqs)[ind_tbl->queues[n]]->mark = 0;
1543                 }
1544                 DEBUG("Flow %p removed", (void *)flow);
1545         }
1546 }
1547
1548 /**
1549  * Add all flows.
1550  *
1551  * @param priv
1552  *   Pointer to private structure.
1553  * @param list
1554  *   Pointer to a TAILQ flow list.
1555  *
1556  * @return
1557  *   0 on success, a errno value otherwise and rte_errno is set.
1558  */
1559 int
1560 priv_flow_start(struct priv *priv, struct mlx5_flows *list)
1561 {
1562         struct rte_flow *flow;
1563
1564         TAILQ_FOREACH(flow, list, next) {
1565                 if (flow->frxq.hrxq)
1566                         goto flow_create;
1567                 flow->frxq.hrxq =
1568                         mlx5_priv_hrxq_get(priv, rss_hash_default_key,
1569                                            rss_hash_default_key_len,
1570                                            flow->frxq.hash_fields,
1571                                            (*flow->queues),
1572                                            flow->queues_n);
1573                 if (flow->frxq.hrxq)
1574                         goto flow_create;
1575                 flow->frxq.hrxq =
1576                         mlx5_priv_hrxq_new(priv, rss_hash_default_key,
1577                                            rss_hash_default_key_len,
1578                                            flow->frxq.hash_fields,
1579                                            (*flow->queues),
1580                                            flow->queues_n);
1581                 if (!flow->frxq.hrxq) {
1582                         DEBUG("Flow %p cannot be applied",
1583                               (void *)flow);
1584                         rte_errno = EINVAL;
1585                         return rte_errno;
1586                 }
1587 flow_create:
1588                 flow->ibv_flow = ibv_create_flow(flow->frxq.hrxq->qp,
1589                                                  flow->ibv_attr);
1590                 if (!flow->ibv_flow) {
1591                         DEBUG("Flow %p cannot be applied", (void *)flow);
1592                         rte_errno = EINVAL;
1593                         return rte_errno;
1594                 }
1595                 DEBUG("Flow %p applied", (void *)flow);
1596                 if (flow->mark) {
1597                         unsigned int n;
1598
1599                         for (n = 0;
1600                              n < flow->frxq.hrxq->ind_table->queues_n;
1601                              ++n) {
1602                                 uint16_t idx =
1603                                         flow->frxq.hrxq->ind_table->queues[n];
1604                                 (*priv->rxqs)[idx]->mark = 1;
1605                         }
1606                 }
1607         }
1608         return 0;
1609 }
1610
1611 /**
1612  * Isolated mode.
1613  *
1614  * @see rte_flow_isolate()
1615  * @see rte_flow_ops
1616  */
1617 int
1618 mlx5_flow_isolate(struct rte_eth_dev *dev,
1619                   int enable,
1620                   struct rte_flow_error *error)
1621 {
1622         struct priv *priv = dev->data->dev_private;
1623
1624         priv_lock(priv);
1625         if (dev->data->dev_started) {
1626                 rte_flow_error_set(error, EBUSY,
1627                                    RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
1628                                    NULL,
1629                                    "port must be stopped first");
1630                 priv_unlock(priv);
1631                 return -rte_errno;
1632         }
1633         priv->isolated = !!enable;
1634         priv_unlock(priv);
1635         return 0;
1636 }
1637
1638 /**
1639  * Verify the flow list is empty
1640  *
1641  * @param priv
1642  *  Pointer to private structure.
1643  *
1644  * @return the number of flows not released.
1645  */
1646 int
1647 priv_flow_verify(struct priv *priv)
1648 {
1649         struct rte_flow *flow;
1650         int ret = 0;
1651
1652         TAILQ_FOREACH(flow, &priv->flows, next) {
1653                 DEBUG("%p: flow %p still referenced", (void *)priv,
1654                       (void *)flow);
1655                 ++ret;
1656         }
1657         return ret;
1658 }
1659
1660 /**
1661  * Enable a control flow configured from the control plane.
1662  *
1663  * @param dev
1664  *   Pointer to Ethernet device.
1665  * @param eth_spec
1666  *   An Ethernet flow spec to apply.
1667  * @param eth_mask
1668  *   An Ethernet flow mask to apply.
1669  * @param vlan_spec
1670  *   A VLAN flow spec to apply.
1671  * @param vlan_mask
1672  *   A VLAN flow mask to apply.
1673  *
1674  * @return
1675  *   0 on success.
1676  */
1677 int
1678 mlx5_ctrl_flow_vlan(struct rte_eth_dev *dev,
1679                     struct rte_flow_item_eth *eth_spec,
1680                     struct rte_flow_item_eth *eth_mask,
1681                     struct rte_flow_item_vlan *vlan_spec,
1682                     struct rte_flow_item_vlan *vlan_mask)
1683 {
1684         struct priv *priv = dev->data->dev_private;
1685         const struct rte_flow_attr attr = {
1686                 .ingress = 1,
1687                 .priority = MLX5_CTRL_FLOW_PRIORITY,
1688         };
1689         struct rte_flow_item items[] = {
1690                 {
1691                         .type = RTE_FLOW_ITEM_TYPE_ETH,
1692                         .spec = eth_spec,
1693                         .last = NULL,
1694                         .mask = eth_mask,
1695                 },
1696                 {
1697                         .type = (vlan_spec) ? RTE_FLOW_ITEM_TYPE_VLAN :
1698                                 RTE_FLOW_ITEM_TYPE_END,
1699                         .spec = vlan_spec,
1700                         .last = NULL,
1701                         .mask = vlan_mask,
1702                 },
1703                 {
1704                         .type = RTE_FLOW_ITEM_TYPE_END,
1705                 },
1706         };
1707         struct rte_flow_action actions[] = {
1708                 {
1709                         .type = RTE_FLOW_ACTION_TYPE_QUEUE,
1710                         .conf = &(struct rte_flow_action_queue){
1711                                 .index = 0,
1712                         },
1713                 },
1714                 {
1715                         .type = RTE_FLOW_ACTION_TYPE_END,
1716                 },
1717         };
1718         struct rte_flow *flow;
1719         struct rte_flow_error error;
1720
1721         flow = priv_flow_create(priv, &priv->ctrl_flows, &attr, items, actions,
1722                                 &error);
1723         if (!flow)
1724                 return rte_errno;
1725         return 0;
1726 }
1727
1728 /**
1729  * Enable a flow control configured from the control plane.
1730  *
1731  * @param dev
1732  *   Pointer to Ethernet device.
1733  * @param eth_spec
1734  *   An Ethernet flow spec to apply.
1735  * @param eth_mask
1736  *   An Ethernet flow mask to apply.
1737  *
1738  * @return
1739  *   0 on success.
1740  */
1741 int
1742 mlx5_ctrl_flow(struct rte_eth_dev *dev,
1743                struct rte_flow_item_eth *eth_spec,
1744                struct rte_flow_item_eth *eth_mask)
1745 {
1746         return mlx5_ctrl_flow_vlan(dev, eth_spec, eth_mask, NULL, NULL);
1747 }