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 parse actions. */
 296 struct mlx5_flow_action {
 297         uint32_t queue:1; /**< Target is a receive queue. */
 298         uint32_t drop:1; /**< Target is a drop queue. */
 299         uint32_t mark:1; /**< Mark is present in the flow. */
 300         uint32_t mark_id; /**< Mark identifier. */
 301         uint16_t queues[RTE_MAX_QUEUES_PER_PORT]; /**< Queues indexes to use. */
 302         uint16_t queues_n; /**< Number of entries in queue[]. */
 303 };
 304
 305 /** Structure to pass to the conversion function. */
 306 struct mlx5_flow_parse {
 307         struct ibv_flow_attr *ibv_attr; /**< Verbs attribute. */
 308         unsigned int offset; /**< Offset in bytes in the ibv_attr buffer. */
 309         uint32_t inner; /**< Set once VXLAN is encountered. */
 310         uint32_t create:1; /**< Leave allocated resources on exit. */
 311         uint64_t hash_fields; /**< Fields that participate in the hash. */
 312         struct mlx5_flow_action actions; /**< Parsed action result. */
 313 };
 314
 315 static const struct rte_flow_ops mlx5_flow_ops = {
 316         .validate = mlx5_flow_validate,
 317         .create = mlx5_flow_create,
 318         .destroy = mlx5_flow_destroy,
 319         .flush = mlx5_flow_flush,
 320         .query = NULL,
 321         .isolate = mlx5_flow_isolate,
 322 };
 323
 324 /**
 325  * Manage filter operations.
 326  *
 327  * @param dev
 328  *   Pointer to Ethernet device structure.
 329  * @param filter_type
 330  *   Filter type.
 331  * @param filter_op
 332  *   Operation to perform.
 333  * @param arg
 334  *   Pointer to operation-specific structure.
 335  *
 336  * @return
 337  *   0 on success, negative errno value on failure.
 338  */
 339 int
 340 mlx5_dev_filter_ctrl(struct rte_eth_dev *dev,
 341                      enum rte_filter_type filter_type,
 342                      enum rte_filter_op filter_op,
 343                      void *arg)
 344 {
 345         int ret = EINVAL;
 346
 347         if (filter_type == RTE_ETH_FILTER_GENERIC) {
 348                 if (filter_op != RTE_ETH_FILTER_GET)
 349                         return -EINVAL;
 350                 *(const void **)arg = &mlx5_flow_ops;
 351                 return 0;
 352         }
 353         ERROR("%p: filter type (%d) not supported",
 354               (void *)dev, filter_type);
 355         return -ret;
 356 }
 357
 358 /**
 359  * Check support for a given item.
 360  *
 361  * @param item[in]
 362  *   Item specification.
 363  * @param mask[in]
 364  *   Bit-masks covering supported fields to compare with spec, last and mask in
 365  *   \item.
 366  * @param size
 367  *   Bit-Mask size in bytes.
 368  *
 369  * @return
 370  *   0 on success.
 371  */
 372 static int
 373 mlx5_flow_item_validate(const struct rte_flow_item *item,
 374                         const uint8_t *mask, unsigned int size)
 375 {
 376         int ret = 0;
 377
 378         if (!item->spec && (item->mask || item->last))
 379                 return -1;
 380         if (item->spec && !item->mask) {
 381                 unsigned int i;
 382                 const uint8_t *spec = item->spec;
 383
 384                 for (i = 0; i < size; ++i)
 385                         if ((spec[i] | mask[i]) != mask[i])
 386                                 return -1;
 387         }
 388         if (item->last && !item->mask) {
 389                 unsigned int i;
 390                 const uint8_t *spec = item->last;
 391
 392                 for (i = 0; i < size; ++i)
 393                         if ((spec[i] | mask[i]) != mask[i])
 394                                 return -1;
 395         }
 396         if (item->mask) {
 397                 unsigned int i;
 398                 const uint8_t *spec = item->mask;
 399
 400                 for (i = 0; i < size; ++i)
 401                         if ((spec[i] | mask[i]) != mask[i])
 402                                 return -1;
 403         }
 404         if (item->spec && item->last) {
 405                 uint8_t spec[size];
 406                 uint8_t last[size];
 407                 const uint8_t *apply = mask;
 408                 unsigned int i;
 409
 410                 if (item->mask)
 411                         apply = item->mask;
 412                 for (i = 0; i < size; ++i) {
 413                         spec[i] = ((const uint8_t *)item->spec)[i] & apply[i];
 414                         last[i] = ((const uint8_t *)item->last)[i] & apply[i];
 415                 }
 416                 ret = memcmp(spec, last, size);
 417         }
 418         return ret;
 419 }
 420
 421 /**
 422  * Validate and convert a flow supported by the NIC.
 423  *
 424  * @param priv
 425  *   Pointer to private structure.
 426  * @param[in] attr
 427  *   Flow rule attributes.
 428  * @param[in] pattern
 429  *   Pattern specification (list terminated by the END pattern item).
 430  * @param[in] actions
 431  *   Associated actions (list terminated by the END action).
 432  * @param[out] error
 433  *   Perform verbose error reporting if not NULL.
 434  * @param[in, out] flow
 435  *   Flow structure to update.
 436  *
 437  * @return
 438  *   0 on success, a negative errno value otherwise and rte_errno is set.
 439  */
 440 static int
 441 priv_flow_convert(struct priv *priv,
 442                   const struct rte_flow_attr *attr,
 443                   const struct rte_flow_item items[],
 444                   const struct rte_flow_action actions[],
 445                   struct rte_flow_error *error,
 446                   struct mlx5_flow_parse *flow)
 447 {
 448         const struct mlx5_flow_items *cur_item = mlx5_flow_items;
 449
 450         (void)priv;
 451         *flow = (struct mlx5_flow_parse){
 452                 .ibv_attr = flow->ibv_attr,
 453                 .create = flow->create,
 454                 .offset = sizeof(struct ibv_flow_attr),
 455                 .actions = {
 456                         .mark_id = MLX5_FLOW_MARK_DEFAULT,
 457                 },
 458         };
 459         if (attr->group) {
 460                 rte_flow_error_set(error, ENOTSUP,
 461                                    RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
 462                                    NULL,
 463                                    "groups are not supported");
 464                 return -rte_errno;
 465         }
 466         if (attr->priority && attr->priority != MLX5_CTRL_FLOW_PRIORITY) {
 467                 rte_flow_error_set(error, ENOTSUP,
 468                                    RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
 469                                    NULL,
 470                                    "priorities are not supported");
 471                 return -rte_errno;
 472         }
 473         if (attr->egress) {
 474                 rte_flow_error_set(error, ENOTSUP,
 475                                    RTE_FLOW_ERROR_TYPE_ATTR_EGRESS,
 476                                    NULL,
 477                                    "egress is not supported");
 478                 return -rte_errno;
 479         }
 480         if (!attr->ingress) {
 481                 rte_flow_error_set(error, ENOTSUP,
 482                                    RTE_FLOW_ERROR_TYPE_ATTR_INGRESS,
 483                                    NULL,
 484                                    "only ingress is supported");
 485                 return -rte_errno;
 486         }
 487         for (; items->type != RTE_FLOW_ITEM_TYPE_END; ++items) {
 488                 const struct mlx5_flow_items *token = NULL;
 489                 unsigned int i;
 490                 int err;
 491
 492                 if (items->type == RTE_FLOW_ITEM_TYPE_VOID)
 493                         continue;
 494                 for (i = 0;
 495                      cur_item->items &&
 496                      cur_item->items[i] != RTE_FLOW_ITEM_TYPE_END;
 497                      ++i) {
 498                         if (cur_item->items[i] == items->type) {
 499                                 token = &mlx5_flow_items[items->type];
 500                                 break;
 501                         }
 502                 }
 503                 if (!token)
 504                         goto exit_item_not_supported;
 505                 cur_item = token;
 506                 err = mlx5_flow_item_validate(items,
 507                                               (const uint8_t *)cur_item->mask,
 508                                               cur_item->mask_sz);
 509                 if (err)
 510                         goto exit_item_not_supported;
 511                 if (flow->ibv_attr && cur_item->convert) {
 512                         err = cur_item->convert(items,
 513                                                 (cur_item->default_mask ?
 514                                                  cur_item->default_mask :
 515                                                  cur_item->mask),
 516                                                 flow);
 517                         if (err)
 518                                 goto exit_item_not_supported;
 519                 } else if (items->type == RTE_FLOW_ITEM_TYPE_VXLAN) {
 520                         if (flow->inner) {
 521                                 rte_flow_error_set(error, ENOTSUP,
 522                                                    RTE_FLOW_ERROR_TYPE_ITEM,
 523                                                    items,
 524                                                    "cannot recognize multiple"
 525                                                    " VXLAN encapsulations");
 526                                 return -rte_errno;
 527                         }
 528                         flow->inner = 1;
 529                 }
 530                 flow->offset += cur_item->dst_sz;
 531         }
 532         for (; actions->type != RTE_FLOW_ACTION_TYPE_END; ++actions) {
 533                 if (actions->type == RTE_FLOW_ACTION_TYPE_VOID) {
 534                         continue;
 535                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_DROP) {
 536                         flow->actions.drop = 1;
 537                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_QUEUE) {
 538                         const struct rte_flow_action_queue *queue =
 539                                 (const struct rte_flow_action_queue *)
 540                                 actions->conf;
 541                         uint16_t n;
 542                         uint16_t found = 0;
 543
 544                         if (!queue || (queue->index > (priv->rxqs_n - 1)))
 545                                 goto exit_action_not_supported;
 546                         for (n = 0; n < flow->actions.queues_n; ++n) {
 547                                 if (flow->actions.queues[n] == queue->index) {
 548                                         found = 1;
 549                                         break;
 550                                 }
 551                         }
 552                         if (flow->actions.queues_n > 1 && !found) {
 553                                 rte_flow_error_set(error, ENOTSUP,
 554                                            RTE_FLOW_ERROR_TYPE_ACTION,
 555                                            actions,
 556                                            "queue action not in RSS queues");
 557                                 return -rte_errno;
 558                         }
 559                         if (!found) {
 560                                 flow->actions.queue = 1;
 561                                 flow->actions.queues_n = 1;
 562                                 flow->actions.queues[0] = queue->index;
 563                         }
 564                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_RSS) {
 565                         const struct rte_flow_action_rss *rss =
 566                                 (const struct rte_flow_action_rss *)
 567                                 actions->conf;
 568                         uint16_t n;
 569
 570                         if (!rss || !rss->num) {
 571                                 rte_flow_error_set(error, EINVAL,
 572                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 573                                                    actions,
 574                                                    "no valid queues");
 575                                 return -rte_errno;
 576                         }
 577                         if (flow->actions.queues_n == 1) {
 578                                 uint16_t found = 0;
 579
 580                                 assert(flow->actions.queues_n);
 581                                 for (n = 0; n < rss->num; ++n) {
 582                                         if (flow->actions.queues[0] ==
 583                                             rss->queue[n]) {
 584                                                 found = 1;
 585                                                 break;
 586                                         }
 587                                 }
 588                                 if (!found) {
 589                                         rte_flow_error_set(error, ENOTSUP,
 590                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 591                                                    actions,
 592                                                    "queue action not in RSS"
 593                                                    " queues");
 594                                         return -rte_errno;
 595                                 }
 596                         }
 597                         for (n = 0; n < rss->num; ++n) {
 598                                 if (rss->queue[n] >= priv->rxqs_n) {
 599                                         rte_flow_error_set(error, EINVAL,
 600                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 601                                                    actions,
 602                                                    "queue id > number of"
 603                                                    " queues");
 604                                         return -rte_errno;
 605                                 }
 606                         }
 607                         flow->actions.queue = 1;
 608                         for (n = 0; n < rss->num; ++n)
 609                                 flow->actions.queues[n] = rss->queue[n];
 610                         flow->actions.queues_n = rss->num;
 611                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_MARK) {
 612                         const struct rte_flow_action_mark *mark =
 613                                 (const struct rte_flow_action_mark *)
 614                                 actions->conf;
 615
 616                         if (!mark) {
 617                                 rte_flow_error_set(error, EINVAL,
 618                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 619                                                    actions,
 620                                                    "mark must be defined");
 621                                 return -rte_errno;
 622                         } else if (mark->id >= MLX5_FLOW_MARK_MAX) {
 623                                 rte_flow_error_set(error, ENOTSUP,
 624                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 625                                                    actions,
 626                                                    "mark must be between 0"
 627                                                    " and 16777199");
 628                                 return -rte_errno;
 629                         }
 630                         flow->actions.mark = 1;
 631                         flow->actions.mark_id = mark->id;
 632                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_FLAG) {
 633                         flow->actions.mark = 1;
 634                 } else {
 635                         goto exit_action_not_supported;
 636                 }
 637         }
 638         if (flow->actions.mark && !flow->ibv_attr && !flow->actions.drop)
 639                 flow->offset += sizeof(struct ibv_flow_spec_action_tag);
 640         if (!flow->ibv_attr && flow->actions.drop)
 641                 flow->offset += sizeof(struct ibv_flow_spec_action_drop);
 642         if (!flow->actions.queue && !flow->actions.drop) {
 643                 rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_HANDLE,
 644                                    NULL, "no valid action");
 645                 return -rte_errno;
 646         }
 647         return 0;
 648 exit_item_not_supported:
 649         rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 650                            items, "item not supported");
 651         return -rte_errno;
 652 exit_action_not_supported:
 653         rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
 654                            actions, "action not supported");
 655         return -rte_errno;
 656 }
 657
 658 /**
 659  * Convert Ethernet item to Verbs specification.
 660  *
 661  * @param item[in]
 662  *   Item specification.
 663  * @param default_mask[in]
 664  *   Default bit-masks to use when item->mask is not provided.
 665  * @param data[in, out]
 666  *   User structure.
 667  */
 668 static int
 669 mlx5_flow_create_eth(const struct rte_flow_item *item,
 670                      const void *default_mask,
 671                      void *data)
 672 {
 673         const struct rte_flow_item_eth *spec = item->spec;
 674         const struct rte_flow_item_eth *mask = item->mask;
 675         struct mlx5_flow_parse *flow = (struct mlx5_flow_parse *)data;
 676         struct ibv_flow_spec_eth *eth;
 677         const unsigned int eth_size = sizeof(struct ibv_flow_spec_eth);
 678         unsigned int i;
 679
 680         ++flow->ibv_attr->num_of_specs;
 681         flow->ibv_attr->priority = 2;
 682         flow->hash_fields = 0;
 683         eth = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 684         *eth = (struct ibv_flow_spec_eth) {
 685                 .type = flow->inner | IBV_FLOW_SPEC_ETH,
 686                 .size = eth_size,
 687         };
 688         if (!spec)
 689                 return 0;
 690         if (!mask)
 691                 mask = default_mask;
 692         memcpy(eth->val.dst_mac, spec->dst.addr_bytes, ETHER_ADDR_LEN);
 693         memcpy(eth->val.src_mac, spec->src.addr_bytes, ETHER_ADDR_LEN);
 694         eth->val.ether_type = spec->type;
 695         memcpy(eth->mask.dst_mac, mask->dst.addr_bytes, ETHER_ADDR_LEN);
 696         memcpy(eth->mask.src_mac, mask->src.addr_bytes, ETHER_ADDR_LEN);
 697         eth->mask.ether_type = mask->type;
 698         /* Remove unwanted bits from values. */
 699         for (i = 0; i < ETHER_ADDR_LEN; ++i) {
 700                 eth->val.dst_mac[i] &= eth->mask.dst_mac[i];
 701                 eth->val.src_mac[i] &= eth->mask.src_mac[i];
 702         }
 703         eth->val.ether_type &= eth->mask.ether_type;
 704         return 0;
 705 }
 706
 707 /**
 708  * Convert VLAN item to Verbs specification.
 709  *
 710  * @param item[in]
 711  *   Item specification.
 712  * @param default_mask[in]
 713  *   Default bit-masks to use when item->mask is not provided.
 714  * @param data[in, out]
 715  *   User structure.
 716  */
 717 static int
 718 mlx5_flow_create_vlan(const struct rte_flow_item *item,
 719                       const void *default_mask,
 720                       void *data)
 721 {
 722         const struct rte_flow_item_vlan *spec = item->spec;
 723         const struct rte_flow_item_vlan *mask = item->mask;
 724         struct mlx5_flow_parse *flow = (struct mlx5_flow_parse *)data;
 725         struct ibv_flow_spec_eth *eth;
 726         const unsigned int eth_size = sizeof(struct ibv_flow_spec_eth);
 727
 728         eth = (void *)((uintptr_t)flow->ibv_attr + flow->offset - eth_size);
 729         if (!spec)
 730                 return 0;
 731         if (!mask)
 732                 mask = default_mask;
 733         eth->val.vlan_tag = spec->tci;
 734         eth->mask.vlan_tag = mask->tci;
 735         eth->val.vlan_tag &= eth->mask.vlan_tag;
 736         return 0;
 737 }
 738
 739 /**
 740  * Convert IPv4 item to Verbs specification.
 741  *
 742  * @param item[in]
 743  *   Item specification.
 744  * @param default_mask[in]
 745  *   Default bit-masks to use when item->mask is not provided.
 746  * @param data[in, out]
 747  *   User structure.
 748  */
 749 static int
 750 mlx5_flow_create_ipv4(const struct rte_flow_item *item,
 751                       const void *default_mask,
 752                       void *data)
 753 {
 754         const struct rte_flow_item_ipv4 *spec = item->spec;
 755         const struct rte_flow_item_ipv4 *mask = item->mask;
 756         struct mlx5_flow_parse *flow = (struct mlx5_flow_parse *)data;
 757         struct ibv_flow_spec_ipv4_ext *ipv4;
 758         unsigned int ipv4_size = sizeof(struct ibv_flow_spec_ipv4_ext);
 759
 760         ++flow->ibv_attr->num_of_specs;
 761         flow->ibv_attr->priority = 1;
 762         flow->hash_fields = (IBV_RX_HASH_SRC_IPV4 |
 763                              IBV_RX_HASH_DST_IPV4);
 764         ipv4 = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 765         *ipv4 = (struct ibv_flow_spec_ipv4_ext) {
 766                 .type = flow->inner | IBV_FLOW_SPEC_IPV4_EXT,
 767                 .size = ipv4_size,
 768         };
 769         if (!spec)
 770                 return 0;
 771         if (!mask)
 772                 mask = default_mask;
 773         ipv4->val = (struct ibv_flow_ipv4_ext_filter){
 774                 .src_ip = spec->hdr.src_addr,
 775                 .dst_ip = spec->hdr.dst_addr,
 776                 .proto = spec->hdr.next_proto_id,
 777                 .tos = spec->hdr.type_of_service,
 778         };
 779         ipv4->mask = (struct ibv_flow_ipv4_ext_filter){
 780                 .src_ip = mask->hdr.src_addr,
 781                 .dst_ip = mask->hdr.dst_addr,
 782                 .proto = mask->hdr.next_proto_id,
 783                 .tos = mask->hdr.type_of_service,
 784         };
 785         /* Remove unwanted bits from values. */
 786         ipv4->val.src_ip &= ipv4->mask.src_ip;
 787         ipv4->val.dst_ip &= ipv4->mask.dst_ip;
 788         ipv4->val.proto &= ipv4->mask.proto;
 789         ipv4->val.tos &= ipv4->mask.tos;
 790         return 0;
 791 }
 792
 793 /**
 794  * Convert IPv6 item to Verbs specification.
 795  *
 796  * @param item[in]
 797  *   Item specification.
 798  * @param default_mask[in]
 799  *   Default bit-masks to use when item->mask is not provided.
 800  * @param data[in, out]
 801  *   User structure.
 802  */
 803 static int
 804 mlx5_flow_create_ipv6(const struct rte_flow_item *item,
 805                       const void *default_mask,
 806                       void *data)
 807 {
 808         const struct rte_flow_item_ipv6 *spec = item->spec;
 809         const struct rte_flow_item_ipv6 *mask = item->mask;
 810         struct mlx5_flow_parse *flow = (struct mlx5_flow_parse *)data;
 811         struct ibv_flow_spec_ipv6 *ipv6;
 812         unsigned int ipv6_size = sizeof(struct ibv_flow_spec_ipv6);
 813         unsigned int i;
 814
 815         ++flow->ibv_attr->num_of_specs;
 816         flow->ibv_attr->priority = 1;
 817         flow->hash_fields = (IBV_RX_HASH_SRC_IPV6 |
 818                              IBV_RX_HASH_DST_IPV6);
 819         ipv6 = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 820         *ipv6 = (struct ibv_flow_spec_ipv6) {
 821                 .type = flow->inner | IBV_FLOW_SPEC_IPV6,
 822                 .size = ipv6_size,
 823         };
 824         if (!spec)
 825                 return 0;
 826         if (!mask)
 827                 mask = default_mask;
 828         memcpy(ipv6->val.src_ip, spec->hdr.src_addr,
 829                RTE_DIM(ipv6->val.src_ip));
 830         memcpy(ipv6->val.dst_ip, spec->hdr.dst_addr,
 831                RTE_DIM(ipv6->val.dst_ip));
 832         memcpy(ipv6->mask.src_ip, mask->hdr.src_addr,
 833                RTE_DIM(ipv6->mask.src_ip));
 834         memcpy(ipv6->mask.dst_ip, mask->hdr.dst_addr,
 835                RTE_DIM(ipv6->mask.dst_ip));
 836         ipv6->mask.flow_label = mask->hdr.vtc_flow;
 837         ipv6->mask.next_hdr = mask->hdr.proto;
 838         ipv6->mask.hop_limit = mask->hdr.hop_limits;
 839         /* Remove unwanted bits from values. */
 840         for (i = 0; i < RTE_DIM(ipv6->val.src_ip); ++i) {
 841                 ipv6->val.src_ip[i] &= ipv6->mask.src_ip[i];
 842                 ipv6->val.dst_ip[i] &= ipv6->mask.dst_ip[i];
 843         }
 844         ipv6->val.flow_label &= ipv6->mask.flow_label;
 845         ipv6->val.next_hdr &= ipv6->mask.next_hdr;
 846         ipv6->val.hop_limit &= ipv6->mask.hop_limit;
 847         return 0;
 848 }
 849
 850 /**
 851  * Convert UDP item to Verbs specification.
 852  *
 853  * @param item[in]
 854  *   Item specification.
 855  * @param default_mask[in]
 856  *   Default bit-masks to use when item->mask is not provided.
 857  * @param data[in, out]
 858  *   User structure.
 859  */
 860 static int
 861 mlx5_flow_create_udp(const struct rte_flow_item *item,
 862                      const void *default_mask,
 863                      void *data)
 864 {
 865         const struct rte_flow_item_udp *spec = item->spec;
 866         const struct rte_flow_item_udp *mask = item->mask;
 867         struct mlx5_flow_parse *flow = (struct mlx5_flow_parse *)data;
 868         struct ibv_flow_spec_tcp_udp *udp;
 869         unsigned int udp_size = sizeof(struct ibv_flow_spec_tcp_udp);
 870
 871         ++flow->ibv_attr->num_of_specs;
 872         flow->ibv_attr->priority = 0;
 873         flow->hash_fields |= (IBV_RX_HASH_SRC_PORT_UDP |
 874                               IBV_RX_HASH_DST_PORT_UDP);
 875         udp = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 876         *udp = (struct ibv_flow_spec_tcp_udp) {
 877                 .type = flow->inner | IBV_FLOW_SPEC_UDP,
 878                 .size = udp_size,
 879         };
 880         if (!spec)
 881                 return 0;
 882         if (!mask)
 883                 mask = default_mask;
 884         udp->val.dst_port = spec->hdr.dst_port;
 885         udp->val.src_port = spec->hdr.src_port;
 886         udp->mask.dst_port = mask->hdr.dst_port;
 887         udp->mask.src_port = mask->hdr.src_port;
 888         /* Remove unwanted bits from values. */
 889         udp->val.src_port &= udp->mask.src_port;
 890         udp->val.dst_port &= udp->mask.dst_port;
 891         return 0;
 892 }
 893
 894 /**
 895  * Convert TCP item to Verbs specification.
 896  *
 897  * @param item[in]
 898  *   Item specification.
 899  * @param default_mask[in]
 900  *   Default bit-masks to use when item->mask is not provided.
 901  * @param data[in, out]
 902  *   User structure.
 903  */
 904 static int
 905 mlx5_flow_create_tcp(const struct rte_flow_item *item,
 906                      const void *default_mask,
 907                      void *data)
 908 {
 909         const struct rte_flow_item_tcp *spec = item->spec;
 910         const struct rte_flow_item_tcp *mask = item->mask;
 911         struct mlx5_flow_parse *flow = (struct mlx5_flow_parse *)data;
 912         struct ibv_flow_spec_tcp_udp *tcp;
 913         unsigned int tcp_size = sizeof(struct ibv_flow_spec_tcp_udp);
 914
 915         ++flow->ibv_attr->num_of_specs;
 916         flow->ibv_attr->priority = 0;
 917         flow->hash_fields |= (IBV_RX_HASH_SRC_PORT_TCP |
 918                               IBV_RX_HASH_DST_PORT_TCP);
 919         tcp = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 920         *tcp = (struct ibv_flow_spec_tcp_udp) {
 921                 .type = flow->inner | IBV_FLOW_SPEC_TCP,
 922                 .size = tcp_size,
 923         };
 924         if (!spec)
 925                 return 0;
 926         if (!mask)
 927                 mask = default_mask;
 928         tcp->val.dst_port = spec->hdr.dst_port;
 929         tcp->val.src_port = spec->hdr.src_port;
 930         tcp->mask.dst_port = mask->hdr.dst_port;
 931         tcp->mask.src_port = mask->hdr.src_port;
 932         /* Remove unwanted bits from values. */
 933         tcp->val.src_port &= tcp->mask.src_port;
 934         tcp->val.dst_port &= tcp->mask.dst_port;
 935         return 0;
 936 }
 937
 938 /**
 939  * Convert VXLAN item to Verbs specification.
 940  *
 941  * @param item[in]
 942  *   Item specification.
 943  * @param default_mask[in]
 944  *   Default bit-masks to use when item->mask is not provided.
 945  * @param data[in, out]
 946  *   User structure.
 947  */
 948 static int
 949 mlx5_flow_create_vxlan(const struct rte_flow_item *item,
 950                        const void *default_mask,
 951                        void *data)
 952 {
 953         const struct rte_flow_item_vxlan *spec = item->spec;
 954         const struct rte_flow_item_vxlan *mask = item->mask;
 955         struct mlx5_flow_parse *flow = (struct mlx5_flow_parse *)data;
 956         struct ibv_flow_spec_tunnel *vxlan;
 957         unsigned int size = sizeof(struct ibv_flow_spec_tunnel);
 958         union vni {
 959                 uint32_t vlan_id;
 960                 uint8_t vni[4];
 961         } id;
 962
 963         ++flow->ibv_attr->num_of_specs;
 964         flow->ibv_attr->priority = 0;
 965         id.vni[0] = 0;
 966         vxlan = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 967         *vxlan = (struct ibv_flow_spec_tunnel) {
 968                 .type = flow->inner | IBV_FLOW_SPEC_VXLAN_TUNNEL,
 969                 .size = size,
 970         };
 971         flow->inner = IBV_FLOW_SPEC_INNER;
 972         if (!spec)
 973                 return 0;
 974         if (!mask)
 975                 mask = default_mask;
 976         memcpy(&id.vni[1], spec->vni, 3);
 977         vxlan->val.tunnel_id = id.vlan_id;
 978         memcpy(&id.vni[1], mask->vni, 3);
 979         vxlan->mask.tunnel_id = id.vlan_id;
 980         /* Remove unwanted bits from values. */
 981         vxlan->val.tunnel_id &= vxlan->mask.tunnel_id;
 982         return 0;
 983 }
 984
 985 /**
 986  * Convert mark/flag action to Verbs specification.
 987  *
 988  * @param flow
 989  *   Pointer to MLX5 flow structure.
 990  * @param mark_id
 991  *   Mark identifier.
 992  */
 993 static int
 994 mlx5_flow_create_flag_mark(struct mlx5_flow_parse *flow, uint32_t mark_id)
 995 {
 996         struct ibv_flow_spec_action_tag *tag;
 997         unsigned int size = sizeof(struct ibv_flow_spec_action_tag);
 998
 999         assert(flow->actions.mark);
1000         tag = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
1001         *tag = (struct ibv_flow_spec_action_tag){
1002                 .type = IBV_FLOW_SPEC_ACTION_TAG,
1003                 .size = size,
1004                 .tag_id = mlx5_flow_mark_set(mark_id),
1005         };
1006         ++flow->ibv_attr->num_of_specs;
1007         flow->offset += size;
1008         return 0;
1009 }
1010
1011 /**
1012  * Complete flow rule creation with a drop queue.
1013  *
1014  * @param priv
1015  *   Pointer to private structure.
1016  * @param flow
1017  *   MLX5 flow attributes (filled by mlx5_flow_validate()).
1018  * @param[out] error
1019  *   Perform verbose error reporting if not NULL.
1020  *
1021  * @return
1022  *   A flow if the rule could be created.
1023  */
1024 static struct rte_flow *
1025 priv_flow_create_action_queue_drop(struct priv *priv,
1026                                    struct mlx5_flow_parse *flow,
1027                                    struct rte_flow_error *error)
1028 {
1029         struct rte_flow *rte_flow;
1030         struct ibv_flow_spec_action_drop *drop;
1031         unsigned int size = sizeof(struct ibv_flow_spec_action_drop);
1032
1033         assert(priv->pd);
1034         assert(priv->ctx);
1035         rte_flow = rte_calloc(__func__, 1, sizeof(*rte_flow), 0);
1036         if (!rte_flow) {
1037                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1038                                    NULL, "cannot allocate flow memory");
1039                 return NULL;
1040         }
1041         rte_flow->drop = 1;
1042         drop = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
1043         *drop = (struct ibv_flow_spec_action_drop){
1044                         .type = IBV_FLOW_SPEC_ACTION_DROP,
1045                         .size = size,
1046         };
1047         ++flow->ibv_attr->num_of_specs;
1048         flow->offset += sizeof(struct ibv_flow_spec_action_drop);
1049         rte_flow->ibv_attr = flow->ibv_attr;
1050         if (!priv->dev->data->dev_started)
1051                 return rte_flow;
1052         rte_flow->drxq.hrxq.qp = priv->flow_drop_queue->qp;
1053         rte_flow->ibv_flow = ibv_create_flow(rte_flow->drxq.hrxq.qp,
1054                                              rte_flow->ibv_attr);
1055         if (!rte_flow->ibv_flow) {
1056                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1057                                    NULL, "flow rule creation failure");
1058                 goto error;
1059         }
1060         return rte_flow;
1061 error:
1062         assert(rte_flow);
1063         rte_free(rte_flow);
1064         return NULL;
1065 }
1066
1067 /**
1068  * Complete flow rule creation.
1069  *
1070  * @param priv
1071  *   Pointer to private structure.
1072  * @param flow
1073  *   MLX5 flow attributes (filled by mlx5_flow_validate()).
1074  * @param[out] error
1075  *   Perform verbose error reporting if not NULL.
1076  *
1077  * @return
1078  *   A flow if the rule could be created.
1079  */
1080 static struct rte_flow *
1081 priv_flow_create_action_queue(struct priv *priv,
1082                               struct mlx5_flow_parse *flow,
1083                               struct rte_flow_error *error)
1084 {
1085         struct rte_flow *rte_flow;
1086         unsigned int i;
1087
1088         assert(priv->pd);
1089         assert(priv->ctx);
1090         assert(!flow->actions.drop);
1091         rte_flow =
1092                 rte_calloc(__func__, 1,
1093                            sizeof(*flow) +
1094                            flow->actions.queues_n * sizeof(uint16_t),
1095                            0);
1096         if (!rte_flow) {
1097                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1098                                    NULL, "cannot allocate flow memory");
1099                 return NULL;
1100         }
1101         rte_flow->mark = flow->actions.mark;
1102         rte_flow->ibv_attr = flow->ibv_attr;
1103         rte_flow->queues = (uint16_t (*)[])(rte_flow + 1);
1104         memcpy(rte_flow->queues, flow->actions.queues,
1105                flow->actions.queues_n * sizeof(uint16_t));
1106         rte_flow->queues_n = flow->actions.queues_n;
1107         rte_flow->frxq.hash_fields = flow->hash_fields;
1108         rte_flow->frxq.hrxq = mlx5_priv_hrxq_get(priv, rss_hash_default_key,
1109                                                  rss_hash_default_key_len,
1110                                                  flow->hash_fields,
1111                                                  (*rte_flow->queues),
1112                                                  rte_flow->queues_n);
1113         if (!rte_flow->frxq.hrxq) {
1114                 rte_flow->frxq.hrxq =
1115                         mlx5_priv_hrxq_new(priv, rss_hash_default_key,
1116                                            rss_hash_default_key_len,
1117                                            flow->hash_fields,
1118                                            (*rte_flow->queues),
1119                                            rte_flow->queues_n);
1120                 if (!rte_flow->frxq.hrxq) {
1121                         rte_flow_error_set(error, ENOMEM,
1122                                            RTE_FLOW_ERROR_TYPE_HANDLE,
1123                                            NULL, "cannot create hash rxq");
1124                         goto error;
1125                 }
1126         }
1127         for (i = 0; i != flow->actions.queues_n; ++i) {
1128                 struct mlx5_rxq_data *q =
1129                         (*priv->rxqs)[flow->actions.queues[i]];
1130
1131                 q->mark |= flow->actions.mark;
1132         }
1133         if (!priv->dev->data->dev_started)
1134                 return rte_flow;
1135         rte_flow->ibv_flow = ibv_create_flow(rte_flow->frxq.hrxq->qp,
1136                                              rte_flow->ibv_attr);
1137         if (!rte_flow->ibv_flow) {
1138                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1139                                    NULL, "flow rule creation failure");
1140                 goto error;
1141         }
1142         return rte_flow;
1143 error:
1144         assert(rte_flow);
1145         if (rte_flow->frxq.hrxq)
1146                 mlx5_priv_hrxq_release(priv, rte_flow->frxq.hrxq);
1147         rte_free(rte_flow);
1148         return NULL;
1149 }
1150
1151 /**
1152  * Validate a flow.
1153  *
1154  * @param priv
1155  *   Pointer to private structure.
1156  * @param[in] attr
1157  *   Flow rule attributes.
1158  * @param[in] pattern
1159  *   Pattern specification (list terminated by the END pattern item).
1160  * @param[in] actions
1161  *   Associated actions (list terminated by the END action).
1162  * @param[out] error
1163  *   Perform verbose error reporting if not NULL.
1164  * @param[in,out] parser
1165  *   MLX5 parser structure.
1166  *
1167  * @return
1168  *   0 on success, negative errno value on failure.
1169  */
1170 static int
1171 priv_flow_validate(struct priv *priv,
1172                    const struct rte_flow_attr *attr,
1173                    const struct rte_flow_item items[],
1174                    const struct rte_flow_action actions[],
1175                    struct rte_flow_error *error,
1176                    struct mlx5_flow_parse *parser)
1177 {
1178         int err;
1179
1180         err = priv_flow_convert(priv, attr, items, actions, error, parser);
1181         if (err)
1182                 goto exit;
1183         if (parser->actions.mark)
1184                 parser->offset += sizeof(struct ibv_flow_spec_action_tag);
1185         parser->ibv_attr = rte_malloc(__func__, parser->offset, 0);
1186         if (!parser->ibv_attr) {
1187                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1188                                    NULL, "cannot allocate ibv_attr memory");
1189                 err = rte_errno;
1190                 goto exit;
1191         }
1192         *parser->ibv_attr = (struct ibv_flow_attr){
1193                 .type = IBV_FLOW_ATTR_NORMAL,
1194                 .size = sizeof(struct ibv_flow_attr),
1195                 .priority = attr->priority,
1196                 .num_of_specs = 0,
1197                 .port = 0,
1198                 .flags = 0,
1199         };
1200         err = priv_flow_convert(priv, attr, items, actions, error, parser);
1201         if (err || parser->create)
1202                 goto exit;
1203         if (parser->actions.mark)
1204                 mlx5_flow_create_flag_mark(parser, parser->actions.mark_id);
1205         return 0;
1206 exit:
1207         if (parser->ibv_attr)
1208                 rte_free(parser->ibv_attr);
1209         return err;
1210 }
1211
1212 /**
1213  * Convert a flow.
1214  *
1215  * @param priv
1216  *   Pointer to private structure.
1217  * @param list
1218  *   Pointer to a TAILQ flow list.
1219  * @param[in] attr
1220  *   Flow rule attributes.
1221  * @param[in] pattern
1222  *   Pattern specification (list terminated by the END pattern item).
1223  * @param[in] actions
1224  *   Associated actions (list terminated by the END action).
1225  * @param[out] error
1226  *   Perform verbose error reporting if not NULL.
1227  *
1228  * @return
1229  *   A flow on success, NULL otherwise.
1230  */
1231 static struct rte_flow *
1232 priv_flow_create(struct priv *priv,
1233                  struct mlx5_flows *list,
1234                  const struct rte_flow_attr *attr,
1235                  const struct rte_flow_item items[],
1236                  const struct rte_flow_action actions[],
1237                  struct rte_flow_error *error)
1238 {
1239         struct mlx5_flow_parse parser = { .create = 1, };
1240         struct rte_flow *flow;
1241         int err;
1242
1243         err = priv_flow_validate(priv, attr, items, actions, error, &parser);
1244         if (err)
1245                 goto exit;
1246         if (parser.actions.drop)
1247                 flow = priv_flow_create_action_queue_drop(priv, &parser, error);
1248         else
1249                 flow = priv_flow_create_action_queue(priv, &parser, error);
1250         if (!flow)
1251                 goto exit;
1252         TAILQ_INSERT_TAIL(list, flow, next);
1253         DEBUG("Flow created %p", (void *)flow);
1254         return flow;
1255 exit:
1256         if (parser.ibv_attr)
1257                 rte_free(parser.ibv_attr);
1258         return NULL;
1259 }
1260
1261 /**
1262  * Validate a flow supported by the NIC.
1263  *
1264  * @see rte_flow_validate()
1265  * @see rte_flow_ops
1266  */
1267 int
1268 mlx5_flow_validate(struct rte_eth_dev *dev,
1269                    const struct rte_flow_attr *attr,
1270                    const struct rte_flow_item items[],
1271                    const struct rte_flow_action actions[],
1272                    struct rte_flow_error *error)
1273 {
1274         struct priv *priv = dev->data->dev_private;
1275         int ret;
1276         struct mlx5_flow_parse parser = { .create = 0, };
1277
1278         priv_lock(priv);
1279         ret = priv_flow_validate(priv, attr, items, actions, error, &parser);
1280         priv_unlock(priv);
1281         return ret;
1282 }
1283
1284 /**
1285  * Create a flow.
1286  *
1287  * @see rte_flow_create()
1288  * @see rte_flow_ops
1289  */
1290 struct rte_flow *
1291 mlx5_flow_create(struct rte_eth_dev *dev,
1292                  const struct rte_flow_attr *attr,
1293                  const struct rte_flow_item items[],
1294                  const struct rte_flow_action actions[],
1295                  struct rte_flow_error *error)
1296 {
1297         struct priv *priv = dev->data->dev_private;
1298         struct rte_flow *flow;
1299
1300         priv_lock(priv);
1301         flow = priv_flow_create(priv, &priv->flows, attr, items, actions,
1302                                 error);
1303         priv_unlock(priv);
1304         return flow;
1305 }
1306
1307 /**
1308  * Destroy a flow.
1309  *
1310  * @param priv
1311  *   Pointer to private structure.
1312  * @param list
1313  *   Pointer to a TAILQ flow list.
1314  * @param[in] flow
1315  *   Flow to destroy.
1316  */
1317 static void
1318 priv_flow_destroy(struct priv *priv,
1319                   struct mlx5_flows *list,
1320                   struct rte_flow *flow)
1321 {
1322         unsigned int i;
1323         uint16_t *queues;
1324         uint16_t queues_n;
1325
1326         if (flow->drop || !flow->mark)
1327                 goto free;
1328         queues = flow->frxq.hrxq->ind_table->queues;
1329         queues_n = flow->frxq.hrxq->ind_table->queues_n;
1330         for (i = 0; i != queues_n; ++i) {
1331                 struct rte_flow *tmp;
1332                 struct mlx5_rxq_data *rxq_data = (*priv->rxqs)[queues[i]];
1333                 int mark = 0;
1334
1335                 /*
1336                  * To remove the mark from the queue, the queue must not be
1337                  * present in any other marked flow (RSS or not).
1338                  */
1339                 TAILQ_FOREACH(tmp, list, next) {
1340                         unsigned int j;
1341
1342                         if (!tmp->mark)
1343                                 continue;
1344                         for (j = 0;
1345                              (j != tmp->frxq.hrxq->ind_table->queues_n) &&
1346                              !mark;
1347                              j++)
1348                                 if (tmp->frxq.hrxq->ind_table->queues[j] ==
1349                                     queues[i])
1350                                         mark = 1;
1351                 }
1352                 rxq_data->mark = mark;
1353         }
1354 free:
1355         if (flow->ibv_flow)
1356                 claim_zero(ibv_destroy_flow(flow->ibv_flow));
1357         if (!flow->drop)
1358                 mlx5_priv_hrxq_release(priv, flow->frxq.hrxq);
1359         TAILQ_REMOVE(list, flow, next);
1360         rte_free(flow->ibv_attr);
1361         DEBUG("Flow destroyed %p", (void *)flow);
1362         rte_free(flow);
1363 }
1364
1365 /**
1366  * Destroy a flow.
1367  *
1368  * @see rte_flow_destroy()
1369  * @see rte_flow_ops
1370  */
1371 int
1372 mlx5_flow_destroy(struct rte_eth_dev *dev,
1373                   struct rte_flow *flow,
1374                   struct rte_flow_error *error)
1375 {
1376         struct priv *priv = dev->data->dev_private;
1377
1378         (void)error;
1379         priv_lock(priv);
1380         priv_flow_destroy(priv, &priv->flows, flow);
1381         priv_unlock(priv);
1382         return 0;
1383 }
1384
1385 /**
1386  * Destroy all flows.
1387  *
1388  * @param priv
1389  *   Pointer to private structure.
1390  * @param list
1391  *   Pointer to a TAILQ flow list.
1392  */
1393 void
1394 priv_flow_flush(struct priv *priv, struct mlx5_flows *list)
1395 {
1396         while (!TAILQ_EMPTY(list)) {
1397                 struct rte_flow *flow;
1398
1399                 flow = TAILQ_FIRST(list);
1400                 priv_flow_destroy(priv, list, flow);
1401         }
1402 }
1403
1404 /**
1405  * Destroy all flows.
1406  *
1407  * @see rte_flow_flush()
1408  * @see rte_flow_ops
1409  */
1410 int
1411 mlx5_flow_flush(struct rte_eth_dev *dev,
1412                 struct rte_flow_error *error)
1413 {
1414         struct priv *priv = dev->data->dev_private;
1415
1416         (void)error;
1417         priv_lock(priv);
1418         priv_flow_flush(priv, &priv->flows);
1419         priv_unlock(priv);
1420         return 0;
1421 }
1422
1423 /**
1424  * Create drop queue.
1425  *
1426  * @param priv
1427  *   Pointer to private structure.
1428  *
1429  * @return
1430  *   0 on success.
1431  */
1432 int
1433 priv_flow_create_drop_queue(struct priv *priv)
1434 {
1435         struct mlx5_hrxq_drop *fdq = NULL;
1436
1437         assert(priv->pd);
1438         assert(priv->ctx);
1439         fdq = rte_calloc(__func__, 1, sizeof(*fdq), 0);
1440         if (!fdq) {
1441                 WARN("cannot allocate memory for drop queue");
1442                 goto error;
1443         }
1444         fdq->cq = ibv_create_cq(priv->ctx, 1, NULL, NULL, 0);
1445         if (!fdq->cq) {
1446                 WARN("cannot allocate CQ for drop queue");
1447                 goto error;
1448         }
1449         fdq->wq = ibv_create_wq(priv->ctx,
1450                         &(struct ibv_wq_init_attr){
1451                         .wq_type = IBV_WQT_RQ,
1452                         .max_wr = 1,
1453                         .max_sge = 1,
1454                         .pd = priv->pd,
1455                         .cq = fdq->cq,
1456                         });
1457         if (!fdq->wq) {
1458                 WARN("cannot allocate WQ for drop queue");
1459                 goto error;
1460         }
1461         fdq->ind_table = ibv_create_rwq_ind_table(priv->ctx,
1462                         &(struct ibv_rwq_ind_table_init_attr){
1463                         .log_ind_tbl_size = 0,
1464                         .ind_tbl = &fdq->wq,
1465                         .comp_mask = 0,
1466                         });
1467         if (!fdq->ind_table) {
1468                 WARN("cannot allocate indirection table for drop queue");
1469                 goto error;
1470         }
1471         fdq->qp = ibv_create_qp_ex(priv->ctx,
1472                 &(struct ibv_qp_init_attr_ex){
1473                         .qp_type = IBV_QPT_RAW_PACKET,
1474                         .comp_mask =
1475                                 IBV_QP_INIT_ATTR_PD |
1476                                 IBV_QP_INIT_ATTR_IND_TABLE |
1477                                 IBV_QP_INIT_ATTR_RX_HASH,
1478                         .rx_hash_conf = (struct ibv_rx_hash_conf){
1479                                 .rx_hash_function =
1480                                         IBV_RX_HASH_FUNC_TOEPLITZ,
1481                                 .rx_hash_key_len = rss_hash_default_key_len,
1482                                 .rx_hash_key = rss_hash_default_key,
1483                                 .rx_hash_fields_mask = 0,
1484                                 },
1485                         .rwq_ind_tbl = fdq->ind_table,
1486                         .pd = priv->pd
1487                 });
1488         if (!fdq->qp) {
1489                 WARN("cannot allocate QP for drop queue");
1490                 goto error;
1491         }
1492         priv->flow_drop_queue = fdq;
1493         return 0;
1494 error:
1495         if (fdq->qp)
1496                 claim_zero(ibv_destroy_qp(fdq->qp));
1497         if (fdq->ind_table)
1498                 claim_zero(ibv_destroy_rwq_ind_table(fdq->ind_table));
1499         if (fdq->wq)
1500                 claim_zero(ibv_destroy_wq(fdq->wq));
1501         if (fdq->cq)
1502                 claim_zero(ibv_destroy_cq(fdq->cq));
1503         if (fdq)
1504                 rte_free(fdq);
1505         priv->flow_drop_queue = NULL;
1506         return -1;
1507 }
1508
1509 /**
1510  * Delete drop queue.
1511  *
1512  * @param priv
1513  *   Pointer to private structure.
1514  */
1515 void
1516 priv_flow_delete_drop_queue(struct priv *priv)
1517 {
1518         struct mlx5_hrxq_drop *fdq = priv->flow_drop_queue;
1519
1520         if (!fdq)
1521                 return;
1522         if (fdq->qp)
1523                 claim_zero(ibv_destroy_qp(fdq->qp));
1524         if (fdq->ind_table)
1525                 claim_zero(ibv_destroy_rwq_ind_table(fdq->ind_table));
1526         if (fdq->wq)
1527                 claim_zero(ibv_destroy_wq(fdq->wq));
1528         if (fdq->cq)
1529                 claim_zero(ibv_destroy_cq(fdq->cq));
1530         rte_free(fdq);
1531         priv->flow_drop_queue = NULL;
1532 }
1533
1534 /**
1535  * Remove all flows.
1536  *
1537  * @param priv
1538  *   Pointer to private structure.
1539  * @param list
1540  *   Pointer to a TAILQ flow list.
1541  */
1542 void
1543 priv_flow_stop(struct priv *priv, struct mlx5_flows *list)
1544 {
1545         struct rte_flow *flow;
1546
1547         TAILQ_FOREACH_REVERSE(flow, list, mlx5_flows, next) {
1548                 claim_zero(ibv_destroy_flow(flow->ibv_flow));
1549                 flow->ibv_flow = NULL;
1550                 mlx5_priv_hrxq_release(priv, flow->frxq.hrxq);
1551                 flow->frxq.hrxq = NULL;
1552                 if (flow->mark) {
1553                         unsigned int n;
1554                         struct mlx5_ind_table_ibv *ind_tbl =
1555                                 flow->frxq.hrxq->ind_table;
1556
1557                         for (n = 0; n < ind_tbl->queues_n; ++n)
1558                                 (*priv->rxqs)[ind_tbl->queues[n]]->mark = 0;
1559                 }
1560                 DEBUG("Flow %p removed", (void *)flow);
1561         }
1562 }
1563
1564 /**
1565  * Add all flows.
1566  *
1567  * @param priv
1568  *   Pointer to private structure.
1569  * @param list
1570  *   Pointer to a TAILQ flow list.
1571  *
1572  * @return
1573  *   0 on success, a errno value otherwise and rte_errno is set.
1574  */
1575 int
1576 priv_flow_start(struct priv *priv, struct mlx5_flows *list)
1577 {
1578         struct rte_flow *flow;
1579
1580         TAILQ_FOREACH(flow, list, next) {
1581                 if (flow->frxq.hrxq)
1582                         goto flow_create;
1583                 flow->frxq.hrxq =
1584                         mlx5_priv_hrxq_get(priv, rss_hash_default_key,
1585                                            rss_hash_default_key_len,
1586                                            flow->frxq.hash_fields,
1587                                            (*flow->queues),
1588                                            flow->queues_n);
1589                 if (flow->frxq.hrxq)
1590                         goto flow_create;
1591                 flow->frxq.hrxq =
1592                         mlx5_priv_hrxq_new(priv, rss_hash_default_key,
1593                                            rss_hash_default_key_len,
1594                                            flow->frxq.hash_fields,
1595                                            (*flow->queues),
1596                                            flow->queues_n);
1597                 if (!flow->frxq.hrxq) {
1598                         DEBUG("Flow %p cannot be applied",
1599                               (void *)flow);
1600                         rte_errno = EINVAL;
1601                         return rte_errno;
1602                 }
1603 flow_create:
1604                 flow->ibv_flow = ibv_create_flow(flow->frxq.hrxq->qp,
1605                                                  flow->ibv_attr);
1606                 if (!flow->ibv_flow) {
1607                         DEBUG("Flow %p cannot be applied", (void *)flow);
1608                         rte_errno = EINVAL;
1609                         return rte_errno;
1610                 }
1611                 DEBUG("Flow %p applied", (void *)flow);
1612                 if (flow->mark) {
1613                         unsigned int n;
1614
1615                         for (n = 0;
1616                              n < flow->frxq.hrxq->ind_table->queues_n;
1617                              ++n) {
1618                                 uint16_t idx =
1619                                         flow->frxq.hrxq->ind_table->queues[n];
1620                                 (*priv->rxqs)[idx]->mark = 1;
1621                         }
1622                 }
1623         }
1624         return 0;
1625 }
1626
1627 /**
1628  * Isolated mode.
1629  *
1630  * @see rte_flow_isolate()
1631  * @see rte_flow_ops
1632  */
1633 int
1634 mlx5_flow_isolate(struct rte_eth_dev *dev,
1635                   int enable,
1636                   struct rte_flow_error *error)
1637 {
1638         struct priv *priv = dev->data->dev_private;
1639
1640         priv_lock(priv);
1641         if (dev->data->dev_started) {
1642                 rte_flow_error_set(error, EBUSY,
1643                                    RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
1644                                    NULL,
1645                                    "port must be stopped first");
1646                 priv_unlock(priv);
1647                 return -rte_errno;
1648         }
1649         priv->isolated = !!enable;
1650         priv_unlock(priv);
1651         return 0;
1652 }
1653
1654 /**
1655  * Verify the flow list is empty
1656  *
1657  * @param priv
1658  *  Pointer to private structure.
1659  *
1660  * @return the number of flows not released.
1661  */
1662 int
1663 priv_flow_verify(struct priv *priv)
1664 {
1665         struct rte_flow *flow;
1666         int ret = 0;
1667
1668         TAILQ_FOREACH(flow, &priv->flows, next) {
1669                 DEBUG("%p: flow %p still referenced", (void *)priv,
1670                       (void *)flow);
1671                 ++ret;
1672         }
1673         return ret;
1674 }
1675
1676 /**
1677  * Enable a control flow configured from the control plane.
1678  *
1679  * @param dev
1680  *   Pointer to Ethernet device.
1681  * @param eth_spec
1682  *   An Ethernet flow spec to apply.
1683  * @param eth_mask
1684  *   An Ethernet flow mask to apply.
1685  * @param vlan_spec
1686  *   A VLAN flow spec to apply.
1687  * @param vlan_mask
1688  *   A VLAN flow mask to apply.
1689  *
1690  * @return
1691  *   0 on success.
1692  */
1693 int
1694 mlx5_ctrl_flow_vlan(struct rte_eth_dev *dev,
1695                     struct rte_flow_item_eth *eth_spec,
1696                     struct rte_flow_item_eth *eth_mask,
1697                     struct rte_flow_item_vlan *vlan_spec,
1698                     struct rte_flow_item_vlan *vlan_mask)
1699 {
1700         struct priv *priv = dev->data->dev_private;
1701         const struct rte_flow_attr attr = {
1702                 .ingress = 1,
1703                 .priority = MLX5_CTRL_FLOW_PRIORITY,
1704         };
1705         struct rte_flow_item items[] = {
1706                 {
1707                         .type = RTE_FLOW_ITEM_TYPE_ETH,
1708                         .spec = eth_spec,
1709                         .last = NULL,
1710                         .mask = eth_mask,
1711                 },
1712                 {
1713                         .type = (vlan_spec) ? RTE_FLOW_ITEM_TYPE_VLAN :
1714                                 RTE_FLOW_ITEM_TYPE_END,
1715                         .spec = vlan_spec,
1716                         .last = NULL,
1717                         .mask = vlan_mask,
1718                 },
1719                 {
1720                         .type = RTE_FLOW_ITEM_TYPE_END,
1721                 },
1722         };
1723         struct rte_flow_action actions[] = {
1724                 {
1725                         .type = RTE_FLOW_ACTION_TYPE_QUEUE,
1726                         .conf = &(struct rte_flow_action_queue){
1727                                 .index = 0,
1728                         },
1729                 },
1730                 {
1731                         .type = RTE_FLOW_ACTION_TYPE_END,
1732                 },
1733         };
1734         struct rte_flow *flow;
1735         struct rte_flow_error error;
1736
1737         flow = priv_flow_create(priv, &priv->ctrl_flows, &attr, items, actions,
1738                                 &error);
1739         if (!flow)
1740                 return rte_errno;
1741         return 0;
1742 }
1743
1744 /**
1745  * Enable a flow control configured from the control plane.
1746  *
1747  * @param dev
1748  *   Pointer to Ethernet device.
1749  * @param eth_spec
1750  *   An Ethernet flow spec to apply.
1751  * @param eth_mask
1752  *   An Ethernet flow mask to apply.
1753  *
1754  * @return
1755  *   0 on success.
1756  */
1757 int
1758 mlx5_ctrl_flow(struct rte_eth_dev *dev,
1759                struct rte_flow_item_eth *eth_spec,
1760                struct rte_flow_item_eth *eth_mask)
1761 {
1762         return mlx5_ctrl_flow_vlan(dev, eth_spec, eth_mask, NULL, NULL);
1763 }