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 static int
  56 mlx5_flow_create_eth(const struct rte_flow_item *item,
  57                      const void *default_mask,
  58                      void *data);
  59
  60 static int
  61 mlx5_flow_create_vlan(const struct rte_flow_item *item,
  62                       const void *default_mask,
  63                       void *data);
  64
  65 static int
  66 mlx5_flow_create_ipv4(const struct rte_flow_item *item,
  67                       const void *default_mask,
  68                       void *data);
  69
  70 static int
  71 mlx5_flow_create_ipv6(const struct rte_flow_item *item,
  72                       const void *default_mask,
  73                       void *data);
  74
  75 static int
  76 mlx5_flow_create_udp(const struct rte_flow_item *item,
  77                      const void *default_mask,
  78                      void *data);
  79
  80 static int
  81 mlx5_flow_create_tcp(const struct rte_flow_item *item,
  82                      const void *default_mask,
  83                      void *data);
  84
  85 static int
  86 mlx5_flow_create_vxlan(const struct rte_flow_item *item,
  87                        const void *default_mask,
  88                        void *data);
  89
  90 struct rte_flow {
  91         LIST_ENTRY(rte_flow) next; /**< Pointer to the next flow structure. */
  92         struct ibv_exp_flow_attr *ibv_attr; /**< Pointer to Verbs attributes. */
  93         struct ibv_exp_rwq_ind_table *ind_table; /**< Indirection table. */
  94         struct ibv_qp *qp; /**< Verbs queue pair. */
  95         struct ibv_exp_flow *ibv_flow; /**< Verbs flow. */
  96         struct ibv_exp_wq *wq; /**< Verbs work queue. */
  97         struct ibv_cq *cq; /**< Verbs completion queue. */
  98         struct rxq *(*rxqs)[]; /**< Pointer to the queues array. */
  99         uint16_t rxqs_n; /**< Number of queues in this flow, 0 if drop queue. */
 100         uint32_t mark:1; /**< Set if the flow is marked. */
 101         uint32_t drop:1; /**< Drop queue. */
 102         uint64_t hash_fields; /**< Fields that participate in the hash. */
 103 };
 104
 105 /** Static initializer for items. */
 106 #define ITEMS(...) \
 107         (const enum rte_flow_item_type []){ \
 108                 __VA_ARGS__, RTE_FLOW_ITEM_TYPE_END, \
 109         }
 110
 111 /** Structure to generate a simple graph of layers supported by the NIC. */
 112 struct mlx5_flow_items {
 113         /** List of possible actions for these items. */
 114         const enum rte_flow_action_type *const actions;
 115         /** Bit-masks corresponding to the possibilities for the item. */
 116         const void *mask;
 117         /**
 118          * Default bit-masks to use when item->mask is not provided. When
 119          * \default_mask is also NULL, the full supported bit-mask (\mask) is
 120          * used instead.
 121          */
 122         const void *default_mask;
 123         /** Bit-masks size in bytes. */
 124         const unsigned int mask_sz;
 125         /**
 126          * Conversion function from rte_flow to NIC specific flow.
 127          *
 128          * @param item
 129          *   rte_flow item to convert.
 130          * @param default_mask
 131          *   Default bit-masks to use when item->mask is not provided.
 132          * @param data
 133          *   Internal structure to store the conversion.
 134          *
 135          * @return
 136          *   0 on success, negative value otherwise.
 137          */
 138         int (*convert)(const struct rte_flow_item *item,
 139                        const void *default_mask,
 140                        void *data);
 141         /** Size in bytes of the destination structure. */
 142         const unsigned int dst_sz;
 143         /** List of possible following items.  */
 144         const enum rte_flow_item_type *const items;
 145 };
 146
 147 /** Valid action for this PMD. */
 148 static const enum rte_flow_action_type valid_actions[] = {
 149         RTE_FLOW_ACTION_TYPE_DROP,
 150         RTE_FLOW_ACTION_TYPE_QUEUE,
 151         RTE_FLOW_ACTION_TYPE_MARK,
 152         RTE_FLOW_ACTION_TYPE_FLAG,
 153         RTE_FLOW_ACTION_TYPE_END,
 154 };
 155
 156 /** Graph of supported items and associated actions. */
 157 static const struct mlx5_flow_items mlx5_flow_items[] = {
 158         [RTE_FLOW_ITEM_TYPE_END] = {
 159                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_ETH,
 160                                RTE_FLOW_ITEM_TYPE_VXLAN),
 161         },
 162         [RTE_FLOW_ITEM_TYPE_ETH] = {
 163                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_VLAN,
 164                                RTE_FLOW_ITEM_TYPE_IPV4,
 165                                RTE_FLOW_ITEM_TYPE_IPV6),
 166                 .actions = valid_actions,
 167                 .mask = &(const struct rte_flow_item_eth){
 168                         .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 169                         .src.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 170                         .type = -1,
 171                 },
 172                 .default_mask = &rte_flow_item_eth_mask,
 173                 .mask_sz = sizeof(struct rte_flow_item_eth),
 174                 .convert = mlx5_flow_create_eth,
 175                 .dst_sz = sizeof(struct ibv_exp_flow_spec_eth),
 176         },
 177         [RTE_FLOW_ITEM_TYPE_VLAN] = {
 178                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_IPV4,
 179                                RTE_FLOW_ITEM_TYPE_IPV6),
 180                 .actions = valid_actions,
 181                 .mask = &(const struct rte_flow_item_vlan){
 182                         .tci = -1,
 183                 },
 184                 .default_mask = &rte_flow_item_vlan_mask,
 185                 .mask_sz = sizeof(struct rte_flow_item_vlan),
 186                 .convert = mlx5_flow_create_vlan,
 187                 .dst_sz = 0,
 188         },
 189         [RTE_FLOW_ITEM_TYPE_IPV4] = {
 190                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP,
 191                                RTE_FLOW_ITEM_TYPE_TCP),
 192                 .actions = valid_actions,
 193                 .mask = &(const struct rte_flow_item_ipv4){
 194                         .hdr = {
 195                                 .src_addr = -1,
 196                                 .dst_addr = -1,
 197                                 .type_of_service = -1,
 198                                 .next_proto_id = -1,
 199                         },
 200                 },
 201                 .default_mask = &rte_flow_item_ipv4_mask,
 202                 .mask_sz = sizeof(struct rte_flow_item_ipv4),
 203                 .convert = mlx5_flow_create_ipv4,
 204                 .dst_sz = sizeof(struct ibv_exp_flow_spec_ipv4_ext),
 205         },
 206         [RTE_FLOW_ITEM_TYPE_IPV6] = {
 207                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP,
 208                                RTE_FLOW_ITEM_TYPE_TCP),
 209                 .actions = valid_actions,
 210                 .mask = &(const struct rte_flow_item_ipv6){
 211                         .hdr = {
 212                                 .src_addr = {
 213                                         0xff, 0xff, 0xff, 0xff,
 214                                         0xff, 0xff, 0xff, 0xff,
 215                                         0xff, 0xff, 0xff, 0xff,
 216                                         0xff, 0xff, 0xff, 0xff,
 217                                 },
 218                                 .dst_addr = {
 219                                         0xff, 0xff, 0xff, 0xff,
 220                                         0xff, 0xff, 0xff, 0xff,
 221                                         0xff, 0xff, 0xff, 0xff,
 222                                         0xff, 0xff, 0xff, 0xff,
 223                                 },
 224                                 .vtc_flow = -1,
 225                                 .proto = -1,
 226                                 .hop_limits = -1,
 227                         },
 228                 },
 229                 .default_mask = &rte_flow_item_ipv6_mask,
 230                 .mask_sz = sizeof(struct rte_flow_item_ipv6),
 231                 .convert = mlx5_flow_create_ipv6,
 232                 .dst_sz = sizeof(struct ibv_exp_flow_spec_ipv6_ext),
 233         },
 234         [RTE_FLOW_ITEM_TYPE_UDP] = {
 235                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_VXLAN),
 236                 .actions = valid_actions,
 237                 .mask = &(const struct rte_flow_item_udp){
 238                         .hdr = {
 239                                 .src_port = -1,
 240                                 .dst_port = -1,
 241                         },
 242                 },
 243                 .default_mask = &rte_flow_item_udp_mask,
 244                 .mask_sz = sizeof(struct rte_flow_item_udp),
 245                 .convert = mlx5_flow_create_udp,
 246                 .dst_sz = sizeof(struct ibv_exp_flow_spec_tcp_udp),
 247         },
 248         [RTE_FLOW_ITEM_TYPE_TCP] = {
 249                 .actions = valid_actions,
 250                 .mask = &(const struct rte_flow_item_tcp){
 251                         .hdr = {
 252                                 .src_port = -1,
 253                                 .dst_port = -1,
 254                         },
 255                 },
 256                 .default_mask = &rte_flow_item_tcp_mask,
 257                 .mask_sz = sizeof(struct rte_flow_item_tcp),
 258                 .convert = mlx5_flow_create_tcp,
 259                 .dst_sz = sizeof(struct ibv_exp_flow_spec_tcp_udp),
 260         },
 261         [RTE_FLOW_ITEM_TYPE_VXLAN] = {
 262                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_ETH),
 263                 .actions = valid_actions,
 264                 .mask = &(const struct rte_flow_item_vxlan){
 265                         .vni = "\xff\xff\xff",
 266                 },
 267                 .default_mask = &rte_flow_item_vxlan_mask,
 268                 .mask_sz = sizeof(struct rte_flow_item_vxlan),
 269                 .convert = mlx5_flow_create_vxlan,
 270                 .dst_sz = sizeof(struct ibv_exp_flow_spec_tunnel),
 271         },
 272 };
 273
 274 /** Structure to pass to the conversion function. */
 275 struct mlx5_flow {
 276         struct ibv_exp_flow_attr *ibv_attr; /**< Verbs attribute. */
 277         unsigned int offset; /**< Offset in bytes in the ibv_attr buffer. */
 278         uint32_t inner; /**< Set once VXLAN is encountered. */
 279         uint64_t hash_fields; /**< Fields that participate in the hash. */
 280 };
 281
 282 struct mlx5_flow_action {
 283         uint32_t queue:1; /**< Target is a receive queue. */
 284         uint32_t drop:1; /**< Target is a drop queue. */
 285         uint32_t mark:1; /**< Mark is present in the flow. */
 286         uint32_t mark_id; /**< Mark identifier. */
 287         uint16_t queues[RTE_MAX_QUEUES_PER_PORT]; /**< Queues indexes to use. */
 288         uint16_t queues_n; /**< Number of entries in queue[]. */
 289 };
 290
 291 /**
 292  * Check support for a given item.
 293  *
 294  * @param item[in]
 295  *   Item specification.
 296  * @param mask[in]
 297  *   Bit-masks covering supported fields to compare with spec, last and mask in
 298  *   \item.
 299  * @param size
 300  *   Bit-Mask size in bytes.
 301  *
 302  * @return
 303  *   0 on success.
 304  */
 305 static int
 306 mlx5_flow_item_validate(const struct rte_flow_item *item,
 307                         const uint8_t *mask, unsigned int size)
 308 {
 309         int ret = 0;
 310
 311         if (!item->spec && (item->mask || item->last))
 312                 return -1;
 313         if (item->spec && !item->mask) {
 314                 unsigned int i;
 315                 const uint8_t *spec = item->spec;
 316
 317                 for (i = 0; i < size; ++i)
 318                         if ((spec[i] | mask[i]) != mask[i])
 319                                 return -1;
 320         }
 321         if (item->last && !item->mask) {
 322                 unsigned int i;
 323                 const uint8_t *spec = item->last;
 324
 325                 for (i = 0; i < size; ++i)
 326                         if ((spec[i] | mask[i]) != mask[i])
 327                                 return -1;
 328         }
 329         if (item->mask) {
 330                 unsigned int i;
 331                 const uint8_t *spec = item->mask;
 332
 333                 for (i = 0; i < size; ++i)
 334                         if ((spec[i] | mask[i]) != mask[i])
 335                                 return -1;
 336         }
 337         if (item->spec && item->last) {
 338                 uint8_t spec[size];
 339                 uint8_t last[size];
 340                 const uint8_t *apply = mask;
 341                 unsigned int i;
 342
 343                 if (item->mask)
 344                         apply = item->mask;
 345                 for (i = 0; i < size; ++i) {
 346                         spec[i] = ((const uint8_t *)item->spec)[i] & apply[i];
 347                         last[i] = ((const uint8_t *)item->last)[i] & apply[i];
 348                 }
 349                 ret = memcmp(spec, last, size);
 350         }
 351         return ret;
 352 }
 353
 354 /**
 355  * Validate a flow supported by the NIC.
 356  *
 357  * @param priv
 358  *   Pointer to private structure.
 359  * @param[in] attr
 360  *   Flow rule attributes.
 361  * @param[in] pattern
 362  *   Pattern specification (list terminated by the END pattern item).
 363  * @param[in] actions
 364  *   Associated actions (list terminated by the END action).
 365  * @param[out] error
 366  *   Perform verbose error reporting if not NULL.
 367  * @param[in, out] flow
 368  *   Flow structure to update.
 369  *
 370  * @return
 371  *   0 on success, a negative errno value otherwise and rte_errno is set.
 372  */
 373 static int
 374 priv_flow_validate(struct priv *priv,
 375                    const struct rte_flow_attr *attr,
 376                    const struct rte_flow_item items[],
 377                    const struct rte_flow_action actions[],
 378                    struct rte_flow_error *error,
 379                    struct mlx5_flow *flow)
 380 {
 381         const struct mlx5_flow_items *cur_item = mlx5_flow_items;
 382         struct mlx5_flow_action action = {
 383                 .queue = 0,
 384                 .drop = 0,
 385                 .mark = 0,
 386         };
 387
 388         (void)priv;
 389         if (attr->group) {
 390                 rte_flow_error_set(error, ENOTSUP,
 391                                    RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
 392                                    NULL,
 393                                    "groups are not supported");
 394                 return -rte_errno;
 395         }
 396         if (attr->priority) {
 397                 rte_flow_error_set(error, ENOTSUP,
 398                                    RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
 399                                    NULL,
 400                                    "priorities are not supported");
 401                 return -rte_errno;
 402         }
 403         if (attr->egress) {
 404                 rte_flow_error_set(error, ENOTSUP,
 405                                    RTE_FLOW_ERROR_TYPE_ATTR_EGRESS,
 406                                    NULL,
 407                                    "egress is not supported");
 408                 return -rte_errno;
 409         }
 410         if (!attr->ingress) {
 411                 rte_flow_error_set(error, ENOTSUP,
 412                                    RTE_FLOW_ERROR_TYPE_ATTR_INGRESS,
 413                                    NULL,
 414                                    "only ingress is supported");
 415                 return -rte_errno;
 416         }
 417         for (; items->type != RTE_FLOW_ITEM_TYPE_END; ++items) {
 418                 const struct mlx5_flow_items *token = NULL;
 419                 unsigned int i;
 420                 int err;
 421
 422                 if (items->type == RTE_FLOW_ITEM_TYPE_VOID)
 423                         continue;
 424                 for (i = 0;
 425                      cur_item->items &&
 426                      cur_item->items[i] != RTE_FLOW_ITEM_TYPE_END;
 427                      ++i) {
 428                         if (cur_item->items[i] == items->type) {
 429                                 token = &mlx5_flow_items[items->type];
 430                                 break;
 431                         }
 432                 }
 433                 if (!token)
 434                         goto exit_item_not_supported;
 435                 cur_item = token;
 436                 err = mlx5_flow_item_validate(items,
 437                                               (const uint8_t *)cur_item->mask,
 438                                               cur_item->mask_sz);
 439                 if (err)
 440                         goto exit_item_not_supported;
 441                 if (flow->ibv_attr && cur_item->convert) {
 442                         err = cur_item->convert(items,
 443                                                 (cur_item->default_mask ?
 444                                                  cur_item->default_mask :
 445                                                  cur_item->mask),
 446                                                 flow);
 447                         if (err)
 448                                 goto exit_item_not_supported;
 449                 }
 450                 flow->offset += cur_item->dst_sz;
 451         }
 452         for (; actions->type != RTE_FLOW_ACTION_TYPE_END; ++actions) {
 453                 if (actions->type == RTE_FLOW_ACTION_TYPE_VOID) {
 454                         continue;
 455                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_DROP) {
 456                         action.drop = 1;
 457                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_QUEUE) {
 458                         const struct rte_flow_action_queue *queue =
 459                                 (const struct rte_flow_action_queue *)
 460                                 actions->conf;
 461                         uint16_t n;
 462                         uint16_t found = 0;
 463
 464                         if (!queue || (queue->index > (priv->rxqs_n - 1)))
 465                                 goto exit_action_not_supported;
 466                         for (n = 0; n < action.queues_n; ++n) {
 467                                 if (action.queues[n] == queue->index) {
 468                                         found = 1;
 469                                         break;
 470                                 }
 471                         }
 472                         if (action.queues_n && !found) {
 473                                 rte_flow_error_set(error, ENOTSUP,
 474                                            RTE_FLOW_ERROR_TYPE_ACTION,
 475                                            actions,
 476                                            "queue action not in RSS queues");
 477                                 return -rte_errno;
 478                         }
 479                         action.queue = 1;
 480                         action.queues_n = 1;
 481                         action.queues[0] = queue->index;
 482                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_RSS) {
 483                         const struct rte_flow_action_rss *rss =
 484                                 (const struct rte_flow_action_rss *)
 485                                 actions->conf;
 486                         uint16_t n;
 487
 488                         if (action.queues_n == 1) {
 489                                 uint16_t found = 0;
 490
 491                                 assert(action.queues_n);
 492                                 for (n = 0; n < rss->num; ++n) {
 493                                         if (action.queues[0] == rss->queue[n]) {
 494                                                 found = 1;
 495                                                 break;
 496                                         }
 497                                 }
 498                                 if (!found) {
 499                                         rte_flow_error_set(error, ENOTSUP,
 500                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 501                                                    actions,
 502                                                    "queue action not in RSS"
 503                                                    " queues");
 504                                         return -rte_errno;
 505                                 }
 506                         }
 507                         action.queue = 1;
 508                         for (n = 0; n < rss->num; ++n)
 509                                 action.queues[n] = rss->queue[n];
 510                         action.queues_n = rss->num;
 511                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_MARK) {
 512                         const struct rte_flow_action_mark *mark =
 513                                 (const struct rte_flow_action_mark *)
 514                                 actions->conf;
 515
 516                         if (!mark) {
 517                                 rte_flow_error_set(error, EINVAL,
 518                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 519                                                    actions,
 520                                                    "mark must be defined");
 521                                 return -rte_errno;
 522                         } else if (mark->id >= MLX5_FLOW_MARK_MAX) {
 523                                 rte_flow_error_set(error, ENOTSUP,
 524                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 525                                                    actions,
 526                                                    "mark must be between 0"
 527                                                    " and 16777199");
 528                                 return -rte_errno;
 529                         }
 530                         action.mark = 1;
 531                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_FLAG) {
 532                         action.mark = 1;
 533                 } else {
 534                         goto exit_action_not_supported;
 535                 }
 536         }
 537         if (action.mark && !flow->ibv_attr && !action.drop)
 538                 flow->offset += sizeof(struct ibv_exp_flow_spec_action_tag);
 539         if (!action.queue && !action.drop) {
 540                 rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_HANDLE,
 541                                    NULL, "no valid action");
 542                 return -rte_errno;
 543         }
 544         return 0;
 545 exit_item_not_supported:
 546         rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 547                            items, "item not supported");
 548         return -rte_errno;
 549 exit_action_not_supported:
 550         rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
 551                            actions, "action not supported");
 552         return -rte_errno;
 553 }
 554
 555 /**
 556  * Validate a flow supported by the NIC.
 557  *
 558  * @see rte_flow_validate()
 559  * @see rte_flow_ops
 560  */
 561 int
 562 mlx5_flow_validate(struct rte_eth_dev *dev,
 563                    const struct rte_flow_attr *attr,
 564                    const struct rte_flow_item items[],
 565                    const struct rte_flow_action actions[],
 566                    struct rte_flow_error *error)
 567 {
 568         struct priv *priv = dev->data->dev_private;
 569         int ret;
 570         struct mlx5_flow flow = { .offset = sizeof(struct ibv_exp_flow_attr) };
 571
 572         priv_lock(priv);
 573         ret = priv_flow_validate(priv, attr, items, actions, error, &flow);
 574         priv_unlock(priv);
 575         return ret;
 576 }
 577
 578 /**
 579  * Convert Ethernet item to Verbs specification.
 580  *
 581  * @param item[in]
 582  *   Item specification.
 583  * @param default_mask[in]
 584  *   Default bit-masks to use when item->mask is not provided.
 585  * @param data[in, out]
 586  *   User structure.
 587  */
 588 static int
 589 mlx5_flow_create_eth(const struct rte_flow_item *item,
 590                      const void *default_mask,
 591                      void *data)
 592 {
 593         const struct rte_flow_item_eth *spec = item->spec;
 594         const struct rte_flow_item_eth *mask = item->mask;
 595         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 596         struct ibv_exp_flow_spec_eth *eth;
 597         const unsigned int eth_size = sizeof(struct ibv_exp_flow_spec_eth);
 598         unsigned int i;
 599
 600         ++flow->ibv_attr->num_of_specs;
 601         flow->ibv_attr->priority = 2;
 602         flow->hash_fields = 0;
 603         eth = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 604         *eth = (struct ibv_exp_flow_spec_eth) {
 605                 .type = flow->inner | IBV_EXP_FLOW_SPEC_ETH,
 606                 .size = eth_size,
 607         };
 608         if (!spec)
 609                 return 0;
 610         if (!mask)
 611                 mask = default_mask;
 612         memcpy(eth->val.dst_mac, spec->dst.addr_bytes, ETHER_ADDR_LEN);
 613         memcpy(eth->val.src_mac, spec->src.addr_bytes, ETHER_ADDR_LEN);
 614         eth->val.ether_type = spec->type;
 615         memcpy(eth->mask.dst_mac, mask->dst.addr_bytes, ETHER_ADDR_LEN);
 616         memcpy(eth->mask.src_mac, mask->src.addr_bytes, ETHER_ADDR_LEN);
 617         eth->mask.ether_type = mask->type;
 618         /* Remove unwanted bits from values. */
 619         for (i = 0; i < ETHER_ADDR_LEN; ++i) {
 620                 eth->val.dst_mac[i] &= eth->mask.dst_mac[i];
 621                 eth->val.src_mac[i] &= eth->mask.src_mac[i];
 622         }
 623         eth->val.ether_type &= eth->mask.ether_type;
 624         return 0;
 625 }
 626
 627 /**
 628  * Convert VLAN item to Verbs specification.
 629  *
 630  * @param item[in]
 631  *   Item specification.
 632  * @param default_mask[in]
 633  *   Default bit-masks to use when item->mask is not provided.
 634  * @param data[in, out]
 635  *   User structure.
 636  */
 637 static int
 638 mlx5_flow_create_vlan(const struct rte_flow_item *item,
 639                       const void *default_mask,
 640                       void *data)
 641 {
 642         const struct rte_flow_item_vlan *spec = item->spec;
 643         const struct rte_flow_item_vlan *mask = item->mask;
 644         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 645         struct ibv_exp_flow_spec_eth *eth;
 646         const unsigned int eth_size = sizeof(struct ibv_exp_flow_spec_eth);
 647
 648         eth = (void *)((uintptr_t)flow->ibv_attr + flow->offset - eth_size);
 649         if (!spec)
 650                 return 0;
 651         if (!mask)
 652                 mask = default_mask;
 653         eth->val.vlan_tag = spec->tci;
 654         eth->mask.vlan_tag = mask->tci;
 655         eth->val.vlan_tag &= eth->mask.vlan_tag;
 656         return 0;
 657 }
 658
 659 /**
 660  * Convert IPv4 item to Verbs specification.
 661  *
 662  * @param item[in]
 663  *   Item specification.
 664  * @param default_mask[in]
 665  *   Default bit-masks to use when item->mask is not provided.
 666  * @param data[in, out]
 667  *   User structure.
 668  */
 669 static int
 670 mlx5_flow_create_ipv4(const struct rte_flow_item *item,
 671                       const void *default_mask,
 672                       void *data)
 673 {
 674         const struct rte_flow_item_ipv4 *spec = item->spec;
 675         const struct rte_flow_item_ipv4 *mask = item->mask;
 676         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 677         struct ibv_exp_flow_spec_ipv4_ext *ipv4;
 678         unsigned int ipv4_size = sizeof(struct ibv_exp_flow_spec_ipv4_ext);
 679
 680         ++flow->ibv_attr->num_of_specs;
 681         flow->ibv_attr->priority = 1;
 682         flow->hash_fields = (IBV_EXP_RX_HASH_SRC_IPV4 |
 683                              IBV_EXP_RX_HASH_DST_IPV4);
 684         ipv4 = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 685         *ipv4 = (struct ibv_exp_flow_spec_ipv4_ext) {
 686                 .type = flow->inner | IBV_EXP_FLOW_SPEC_IPV4_EXT,
 687                 .size = ipv4_size,
 688         };
 689         if (!spec)
 690                 return 0;
 691         if (!mask)
 692                 mask = default_mask;
 693         ipv4->val = (struct ibv_exp_flow_ipv4_ext_filter){
 694                 .src_ip = spec->hdr.src_addr,
 695                 .dst_ip = spec->hdr.dst_addr,
 696                 .proto = spec->hdr.next_proto_id,
 697                 .tos = spec->hdr.type_of_service,
 698         };
 699         ipv4->mask = (struct ibv_exp_flow_ipv4_ext_filter){
 700                 .src_ip = mask->hdr.src_addr,
 701                 .dst_ip = mask->hdr.dst_addr,
 702                 .proto = mask->hdr.next_proto_id,
 703                 .tos = mask->hdr.type_of_service,
 704         };
 705         /* Remove unwanted bits from values. */
 706         ipv4->val.src_ip &= ipv4->mask.src_ip;
 707         ipv4->val.dst_ip &= ipv4->mask.dst_ip;
 708         ipv4->val.proto &= ipv4->mask.proto;
 709         ipv4->val.tos &= ipv4->mask.tos;
 710         return 0;
 711 }
 712
 713 /**
 714  * Convert IPv6 item to Verbs specification.
 715  *
 716  * @param item[in]
 717  *   Item specification.
 718  * @param default_mask[in]
 719  *   Default bit-masks to use when item->mask is not provided.
 720  * @param data[in, out]
 721  *   User structure.
 722  */
 723 static int
 724 mlx5_flow_create_ipv6(const struct rte_flow_item *item,
 725                       const void *default_mask,
 726                       void *data)
 727 {
 728         const struct rte_flow_item_ipv6 *spec = item->spec;
 729         const struct rte_flow_item_ipv6 *mask = item->mask;
 730         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 731         struct ibv_exp_flow_spec_ipv6_ext *ipv6;
 732         unsigned int ipv6_size = sizeof(struct ibv_exp_flow_spec_ipv6_ext);
 733
 734         ++flow->ibv_attr->num_of_specs;
 735         flow->ibv_attr->priority = 1;
 736         flow->hash_fields = (IBV_EXP_RX_HASH_SRC_IPV6 |
 737                              IBV_EXP_RX_HASH_DST_IPV6);
 738         ipv6 = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 739         *ipv6 = (struct ibv_exp_flow_spec_ipv6_ext) {
 740                 .type = flow->inner | IBV_EXP_FLOW_SPEC_IPV6_EXT,
 741                 .size = ipv6_size,
 742         };
 743         if (!spec)
 744                 return 0;
 745         if (!mask)
 746                 mask = default_mask;
 747         memcpy(ipv6->val.src_ip, spec->hdr.src_addr,
 748                RTE_DIM(ipv6->val.src_ip));
 749         memcpy(ipv6->val.dst_ip, spec->hdr.dst_addr,
 750                RTE_DIM(ipv6->val.dst_ip));
 751         memcpy(ipv6->mask.src_ip, mask->hdr.src_addr,
 752                RTE_DIM(ipv6->mask.src_ip));
 753         memcpy(ipv6->mask.dst_ip, mask->hdr.dst_addr,
 754                RTE_DIM(ipv6->mask.dst_ip));
 755         ipv6->mask.flow_label = mask->hdr.vtc_flow;
 756         ipv6->mask.next_hdr = mask->hdr.proto;
 757         ipv6->mask.hop_limit = mask->hdr.hop_limits;
 758         ipv6->val.flow_label &= ipv6->mask.flow_label;
 759         ipv6->val.next_hdr &= ipv6->mask.next_hdr;
 760         ipv6->val.hop_limit &= ipv6->mask.hop_limit;
 761         return 0;
 762 }
 763
 764 /**
 765  * Convert UDP item to Verbs specification.
 766  *
 767  * @param item[in]
 768  *   Item specification.
 769  * @param default_mask[in]
 770  *   Default bit-masks to use when item->mask is not provided.
 771  * @param data[in, out]
 772  *   User structure.
 773  */
 774 static int
 775 mlx5_flow_create_udp(const struct rte_flow_item *item,
 776                      const void *default_mask,
 777                      void *data)
 778 {
 779         const struct rte_flow_item_udp *spec = item->spec;
 780         const struct rte_flow_item_udp *mask = item->mask;
 781         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 782         struct ibv_exp_flow_spec_tcp_udp *udp;
 783         unsigned int udp_size = sizeof(struct ibv_exp_flow_spec_tcp_udp);
 784
 785         ++flow->ibv_attr->num_of_specs;
 786         flow->ibv_attr->priority = 0;
 787         flow->hash_fields |= (IBV_EXP_RX_HASH_SRC_PORT_UDP |
 788                               IBV_EXP_RX_HASH_DST_PORT_UDP);
 789         udp = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 790         *udp = (struct ibv_exp_flow_spec_tcp_udp) {
 791                 .type = flow->inner | IBV_EXP_FLOW_SPEC_UDP,
 792                 .size = udp_size,
 793         };
 794         if (!spec)
 795                 return 0;
 796         if (!mask)
 797                 mask = default_mask;
 798         udp->val.dst_port = spec->hdr.dst_port;
 799         udp->val.src_port = spec->hdr.src_port;
 800         udp->mask.dst_port = mask->hdr.dst_port;
 801         udp->mask.src_port = mask->hdr.src_port;
 802         /* Remove unwanted bits from values. */
 803         udp->val.src_port &= udp->mask.src_port;
 804         udp->val.dst_port &= udp->mask.dst_port;
 805         return 0;
 806 }
 807
 808 /**
 809  * Convert TCP item to Verbs specification.
 810  *
 811  * @param item[in]
 812  *   Item specification.
 813  * @param default_mask[in]
 814  *   Default bit-masks to use when item->mask is not provided.
 815  * @param data[in, out]
 816  *   User structure.
 817  */
 818 static int
 819 mlx5_flow_create_tcp(const struct rte_flow_item *item,
 820                      const void *default_mask,
 821                      void *data)
 822 {
 823         const struct rte_flow_item_tcp *spec = item->spec;
 824         const struct rte_flow_item_tcp *mask = item->mask;
 825         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 826         struct ibv_exp_flow_spec_tcp_udp *tcp;
 827         unsigned int tcp_size = sizeof(struct ibv_exp_flow_spec_tcp_udp);
 828
 829         ++flow->ibv_attr->num_of_specs;
 830         flow->ibv_attr->priority = 0;
 831         flow->hash_fields |= (IBV_EXP_RX_HASH_SRC_PORT_TCP |
 832                               IBV_EXP_RX_HASH_DST_PORT_TCP);
 833         tcp = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 834         *tcp = (struct ibv_exp_flow_spec_tcp_udp) {
 835                 .type = flow->inner | IBV_EXP_FLOW_SPEC_TCP,
 836                 .size = tcp_size,
 837         };
 838         if (!spec)
 839                 return 0;
 840         if (!mask)
 841                 mask = default_mask;
 842         tcp->val.dst_port = spec->hdr.dst_port;
 843         tcp->val.src_port = spec->hdr.src_port;
 844         tcp->mask.dst_port = mask->hdr.dst_port;
 845         tcp->mask.src_port = mask->hdr.src_port;
 846         /* Remove unwanted bits from values. */
 847         tcp->val.src_port &= tcp->mask.src_port;
 848         tcp->val.dst_port &= tcp->mask.dst_port;
 849         return 0;
 850 }
 851
 852 /**
 853  * Convert VXLAN item to Verbs specification.
 854  *
 855  * @param item[in]
 856  *   Item specification.
 857  * @param default_mask[in]
 858  *   Default bit-masks to use when item->mask is not provided.
 859  * @param data[in, out]
 860  *   User structure.
 861  */
 862 static int
 863 mlx5_flow_create_vxlan(const struct rte_flow_item *item,
 864                        const void *default_mask,
 865                        void *data)
 866 {
 867         const struct rte_flow_item_vxlan *spec = item->spec;
 868         const struct rte_flow_item_vxlan *mask = item->mask;
 869         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 870         struct ibv_exp_flow_spec_tunnel *vxlan;
 871         unsigned int size = sizeof(struct ibv_exp_flow_spec_tunnel);
 872         union vni {
 873                 uint32_t vlan_id;
 874                 uint8_t vni[4];
 875         } id;
 876
 877         ++flow->ibv_attr->num_of_specs;
 878         flow->ibv_attr->priority = 0;
 879         id.vni[0] = 0;
 880         vxlan = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 881         *vxlan = (struct ibv_exp_flow_spec_tunnel) {
 882                 .type = flow->inner | IBV_EXP_FLOW_SPEC_VXLAN_TUNNEL,
 883                 .size = size,
 884         };
 885         flow->inner = IBV_EXP_FLOW_SPEC_INNER;
 886         if (!spec)
 887                 return 0;
 888         if (!mask)
 889                 mask = default_mask;
 890         memcpy(&id.vni[1], spec->vni, 3);
 891         vxlan->val.tunnel_id = id.vlan_id;
 892         memcpy(&id.vni[1], mask->vni, 3);
 893         vxlan->mask.tunnel_id = id.vlan_id;
 894         /* Remove unwanted bits from values. */
 895         vxlan->val.tunnel_id &= vxlan->mask.tunnel_id;
 896         return 0;
 897 }
 898
 899 /**
 900  * Convert mark/flag action to Verbs specification.
 901  *
 902  * @param flow
 903  *   Pointer to MLX5 flow structure.
 904  * @param mark_id
 905  *   Mark identifier.
 906  */
 907 static int
 908 mlx5_flow_create_flag_mark(struct mlx5_flow *flow, uint32_t mark_id)
 909 {
 910         struct ibv_exp_flow_spec_action_tag *tag;
 911         unsigned int size = sizeof(struct ibv_exp_flow_spec_action_tag);
 912
 913         tag = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 914         *tag = (struct ibv_exp_flow_spec_action_tag){
 915                 .type = IBV_EXP_FLOW_SPEC_ACTION_TAG,
 916                 .size = size,
 917                 .tag_id = mlx5_flow_mark_set(mark_id),
 918         };
 919         ++flow->ibv_attr->num_of_specs;
 920         return 0;
 921 }
 922
 923 /**
 924  * Complete flow rule creation with a drop queue.
 925  *
 926  * @param priv
 927  *   Pointer to private structure.
 928  * @param flow
 929  *   MLX5 flow attributes (filled by mlx5_flow_validate()).
 930  * @param[out] error
 931  *   Perform verbose error reporting if not NULL.
 932  *
 933  * @return
 934  *   A flow if the rule could be created.
 935  */
 936 static struct rte_flow *
 937 priv_flow_create_action_queue_drop(struct priv *priv,
 938                                    struct mlx5_flow *flow,
 939                                    struct rte_flow_error *error)
 940 {
 941         struct rte_flow *rte_flow;
 942
 943         assert(priv->pd);
 944         assert(priv->ctx);
 945         rte_flow = rte_calloc(__func__, 1, sizeof(*rte_flow), 0);
 946         if (!rte_flow) {
 947                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
 948                                    NULL, "cannot allocate flow memory");
 949                 return NULL;
 950         }
 951         rte_flow->cq =
 952                 ibv_exp_create_cq(priv->ctx, 1, NULL, NULL, 0,
 953                                   &(struct ibv_exp_cq_init_attr){
 954                                           .comp_mask = 0,
 955                                   });
 956         if (!rte_flow->cq) {
 957                 rte_flow_error_set(error, ENOMEM,
 958                                    RTE_FLOW_ERROR_TYPE_HANDLE,
 959                                    NULL, "cannot allocate CQ");
 960                 goto error;
 961         }
 962         rte_flow->wq = ibv_exp_create_wq(priv->ctx,
 963                                          &(struct ibv_exp_wq_init_attr){
 964                                          .wq_type = IBV_EXP_WQT_RQ,
 965                                          .max_recv_wr = 1,
 966                                          .max_recv_sge = 1,
 967                                          .pd = priv->pd,
 968                                          .cq = rte_flow->cq,
 969                                          });
 970         if (!rte_flow->wq) {
 971                 rte_flow_error_set(error, ENOMEM,
 972                                    RTE_FLOW_ERROR_TYPE_HANDLE,
 973                                    NULL, "cannot allocate WQ");
 974                 goto error;
 975         }
 976         rte_flow->drop = 1;
 977         rte_flow->ibv_attr = flow->ibv_attr;
 978         rte_flow->ind_table = ibv_exp_create_rwq_ind_table(
 979                 priv->ctx,
 980                 &(struct ibv_exp_rwq_ind_table_init_attr){
 981                         .pd = priv->pd,
 982                         .log_ind_tbl_size = 0,
 983                         .ind_tbl = &rte_flow->wq,
 984                         .comp_mask = 0,
 985                 });
 986         if (!rte_flow->ind_table) {
 987                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
 988                                    NULL, "cannot allocate indirection table");
 989                 goto error;
 990         }
 991         rte_flow->qp = ibv_exp_create_qp(
 992                 priv->ctx,
 993                 &(struct ibv_exp_qp_init_attr){
 994                         .qp_type = IBV_QPT_RAW_PACKET,
 995                         .comp_mask =
 996                                 IBV_EXP_QP_INIT_ATTR_PD |
 997                                 IBV_EXP_QP_INIT_ATTR_PORT |
 998                                 IBV_EXP_QP_INIT_ATTR_RX_HASH,
 999                         .pd = priv->pd,
1000                         .rx_hash_conf = &(struct ibv_exp_rx_hash_conf){
1001                                 .rx_hash_function =
1002                                         IBV_EXP_RX_HASH_FUNC_TOEPLITZ,
1003                                 .rx_hash_key_len = rss_hash_default_key_len,
1004                                 .rx_hash_key = rss_hash_default_key,
1005                                 .rx_hash_fields_mask = 0,
1006                                 .rwq_ind_tbl = rte_flow->ind_table,
1007                         },
1008                         .port_num = priv->port,
1009                 });
1010         if (!rte_flow->qp) {
1011                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1012                                    NULL, "cannot allocate QP");
1013                 goto error;
1014         }
1015         if (!priv->started)
1016                 return rte_flow;
1017         rte_flow->ibv_flow = ibv_exp_create_flow(rte_flow->qp,
1018                                                  rte_flow->ibv_attr);
1019         if (!rte_flow->ibv_flow) {
1020                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1021                                    NULL, "flow rule creation failure");
1022                 goto error;
1023         }
1024         return rte_flow;
1025 error:
1026         assert(rte_flow);
1027         if (rte_flow->qp)
1028                 ibv_destroy_qp(rte_flow->qp);
1029         if (rte_flow->ind_table)
1030                 ibv_exp_destroy_rwq_ind_table(rte_flow->ind_table);
1031         if (rte_flow->wq)
1032                 ibv_exp_destroy_wq(rte_flow->wq);
1033         if (rte_flow->cq)
1034                 ibv_destroy_cq(rte_flow->cq);
1035         rte_free(rte_flow);
1036         return NULL;
1037 }
1038
1039 /**
1040  * Complete flow rule creation.
1041  *
1042  * @param priv
1043  *   Pointer to private structure.
1044  * @param flow
1045  *   MLX5 flow attributes (filled by mlx5_flow_validate()).
1046  * @param action
1047  *   Target action structure.
1048  * @param[out] error
1049  *   Perform verbose error reporting if not NULL.
1050  *
1051  * @return
1052  *   A flow if the rule could be created.
1053  */
1054 static struct rte_flow *
1055 priv_flow_create_action_queue(struct priv *priv,
1056                               struct mlx5_flow *flow,
1057                               struct mlx5_flow_action *action,
1058                               struct rte_flow_error *error)
1059 {
1060         struct rte_flow *rte_flow;
1061         unsigned int i;
1062         unsigned int j;
1063         const unsigned int wqs_n = 1 << log2above(action->queues_n);
1064         struct ibv_exp_wq *wqs[wqs_n];
1065
1066         assert(priv->pd);
1067         assert(priv->ctx);
1068         assert(!action->drop);
1069         rte_flow = rte_calloc(__func__, 1,
1070                               sizeof(*rte_flow) + sizeof(struct rxq *) *
1071                               action->queues_n, 0);
1072         if (!rte_flow) {
1073                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1074                                    NULL, "cannot allocate flow memory");
1075                 return NULL;
1076         }
1077         rte_flow->rxqs = (struct rxq *(*)[])((uintptr_t)rte_flow +
1078                                              sizeof(struct rxq *) *
1079                                              action->queues_n);
1080         for (i = 0; i < action->queues_n; ++i) {
1081                 struct rxq_ctrl *rxq;
1082
1083                 rxq = container_of((*priv->rxqs)[action->queues[i]],
1084                                    struct rxq_ctrl, rxq);
1085                 wqs[i] = rxq->wq;
1086                 (*rte_flow->rxqs)[i] = &rxq->rxq;
1087                 ++rte_flow->rxqs_n;
1088                 rxq->rxq.mark |= action->mark;
1089         }
1090         /* finalise indirection table. */
1091         for (j = 0; i < wqs_n; ++i, ++j) {
1092                 wqs[i] = wqs[j];
1093                 if (j == action->queues_n)
1094                         j = 0;
1095         }
1096         rte_flow->mark = action->mark;
1097         rte_flow->ibv_attr = flow->ibv_attr;
1098         rte_flow->hash_fields = flow->hash_fields;
1099         rte_flow->ind_table = ibv_exp_create_rwq_ind_table(
1100                 priv->ctx,
1101                 &(struct ibv_exp_rwq_ind_table_init_attr){
1102                         .pd = priv->pd,
1103                         .log_ind_tbl_size = log2above(action->queues_n),
1104                         .ind_tbl = wqs,
1105                         .comp_mask = 0,
1106                 });
1107         if (!rte_flow->ind_table) {
1108                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1109                                    NULL, "cannot allocate indirection table");
1110                 goto error;
1111         }
1112         rte_flow->qp = ibv_exp_create_qp(
1113                 priv->ctx,
1114                 &(struct ibv_exp_qp_init_attr){
1115                         .qp_type = IBV_QPT_RAW_PACKET,
1116                         .comp_mask =
1117                                 IBV_EXP_QP_INIT_ATTR_PD |
1118                                 IBV_EXP_QP_INIT_ATTR_PORT |
1119                                 IBV_EXP_QP_INIT_ATTR_RX_HASH,
1120                         .pd = priv->pd,
1121                         .rx_hash_conf = &(struct ibv_exp_rx_hash_conf){
1122                                 .rx_hash_function =
1123                                         IBV_EXP_RX_HASH_FUNC_TOEPLITZ,
1124                                 .rx_hash_key_len = rss_hash_default_key_len,
1125                                 .rx_hash_key = rss_hash_default_key,
1126                                 .rx_hash_fields_mask = rte_flow->hash_fields,
1127                                 .rwq_ind_tbl = rte_flow->ind_table,
1128                         },
1129                         .port_num = priv->port,
1130                 });
1131         if (!rte_flow->qp) {
1132                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1133                                    NULL, "cannot allocate QP");
1134                 goto error;
1135         }
1136         if (!priv->started)
1137                 return rte_flow;
1138         rte_flow->ibv_flow = ibv_exp_create_flow(rte_flow->qp,
1139                                                  rte_flow->ibv_attr);
1140         if (!rte_flow->ibv_flow) {
1141                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1142                                    NULL, "flow rule creation failure");
1143                 goto error;
1144         }
1145         return rte_flow;
1146 error:
1147         assert(rte_flow);
1148         if (rte_flow->qp)
1149                 ibv_destroy_qp(rte_flow->qp);
1150         if (rte_flow->ind_table)
1151                 ibv_exp_destroy_rwq_ind_table(rte_flow->ind_table);
1152         rte_free(rte_flow);
1153         return NULL;
1154 }
1155
1156 /**
1157  * Convert a flow.
1158  *
1159  * @param priv
1160  *   Pointer to private structure.
1161  * @param[in] attr
1162  *   Flow rule attributes.
1163  * @param[in] pattern
1164  *   Pattern specification (list terminated by the END pattern item).
1165  * @param[in] actions
1166  *   Associated actions (list terminated by the END action).
1167  * @param[out] error
1168  *   Perform verbose error reporting if not NULL.
1169  *
1170  * @return
1171  *   A flow on success, NULL otherwise.
1172  */
1173 static struct rte_flow *
1174 priv_flow_create(struct priv *priv,
1175                  const struct rte_flow_attr *attr,
1176                  const struct rte_flow_item items[],
1177                  const struct rte_flow_action actions[],
1178                  struct rte_flow_error *error)
1179 {
1180         struct rte_flow *rte_flow;
1181         struct mlx5_flow_action action;
1182         struct mlx5_flow flow = { .offset = sizeof(struct ibv_exp_flow_attr), };
1183         int err;
1184
1185         err = priv_flow_validate(priv, attr, items, actions, error, &flow);
1186         if (err)
1187                 goto exit;
1188         flow.ibv_attr = rte_malloc(__func__, flow.offset, 0);
1189         flow.offset = sizeof(struct ibv_exp_flow_attr);
1190         if (!flow.ibv_attr) {
1191                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1192                                    NULL, "cannot allocate ibv_attr memory");
1193                 goto exit;
1194         }
1195         *flow.ibv_attr = (struct ibv_exp_flow_attr){
1196                 .type = IBV_EXP_FLOW_ATTR_NORMAL,
1197                 .size = sizeof(struct ibv_exp_flow_attr),
1198                 .priority = attr->priority,
1199                 .num_of_specs = 0,
1200                 .port = 0,
1201                 .flags = 0,
1202                 .reserved = 0,
1203         };
1204         flow.inner = 0;
1205         flow.hash_fields = 0;
1206         claim_zero(priv_flow_validate(priv, attr, items, actions,
1207                                       error, &flow));
1208         action = (struct mlx5_flow_action){
1209                 .queue = 0,
1210                 .drop = 0,
1211                 .mark = 0,
1212                 .mark_id = MLX5_FLOW_MARK_DEFAULT,
1213         };
1214         for (; actions->type != RTE_FLOW_ACTION_TYPE_END; ++actions) {
1215                 if (actions->type == RTE_FLOW_ACTION_TYPE_VOID) {
1216                         continue;
1217                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_QUEUE) {
1218                         action.queue = 1;
1219                         action.queues[action.queues_n++] =
1220                                 ((const struct rte_flow_action_queue *)
1221                                  actions->conf)->index;
1222                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_RSS) {
1223                         const struct rte_flow_action_rss *rss =
1224                                 (const struct rte_flow_action_rss *)
1225                                  actions->conf;
1226                         uint16_t n;
1227
1228                         action.queue = 1;
1229                         action.queues_n = rss->num;
1230                         for (n = 0; n < rss->num; ++n)
1231                                 action.queues[n] = rss->queue[n];
1232                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_DROP) {
1233                         action.drop = 1;
1234                         action.mark = 0;
1235                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_MARK) {
1236                         const struct rte_flow_action_mark *mark =
1237                                 (const struct rte_flow_action_mark *)
1238                                 actions->conf;
1239
1240                         if (mark)
1241                                 action.mark_id = mark->id;
1242                         action.mark = !action.drop;
1243                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_FLAG) {
1244                         action.mark = 1;
1245                 } else {
1246                         rte_flow_error_set(error, ENOTSUP,
1247                                            RTE_FLOW_ERROR_TYPE_ACTION,
1248                                            actions, "unsupported action");
1249                         goto exit;
1250                 }
1251         }
1252         if (action.mark) {
1253                 mlx5_flow_create_flag_mark(&flow, action.mark_id);
1254                 flow.offset += sizeof(struct ibv_exp_flow_spec_action_tag);
1255         }
1256         if (action.drop)
1257                 rte_flow =
1258                         priv_flow_create_action_queue_drop(priv, &flow, error);
1259         else
1260                 rte_flow = priv_flow_create_action_queue(priv, &flow, &action,
1261                                                          error);
1262         if (!rte_flow)
1263                 goto exit;
1264         return rte_flow;
1265 exit:
1266         rte_free(flow.ibv_attr);
1267         return NULL;
1268 }
1269
1270 /**
1271  * Create a flow.
1272  *
1273  * @see rte_flow_create()
1274  * @see rte_flow_ops
1275  */
1276 struct rte_flow *
1277 mlx5_flow_create(struct rte_eth_dev *dev,
1278                  const struct rte_flow_attr *attr,
1279                  const struct rte_flow_item items[],
1280                  const struct rte_flow_action actions[],
1281                  struct rte_flow_error *error)
1282 {
1283         struct priv *priv = dev->data->dev_private;
1284         struct rte_flow *flow;
1285
1286         priv_lock(priv);
1287         flow = priv_flow_create(priv, attr, items, actions, error);
1288         if (flow) {
1289                 LIST_INSERT_HEAD(&priv->flows, flow, next);
1290                 DEBUG("Flow created %p", (void *)flow);
1291         }
1292         priv_unlock(priv);
1293         return flow;
1294 }
1295
1296 /**
1297  * Destroy a flow.
1298  *
1299  * @param priv
1300  *   Pointer to private structure.
1301  * @param[in] flow
1302  *   Flow to destroy.
1303  */
1304 static void
1305 priv_flow_destroy(struct priv *priv,
1306                   struct rte_flow *flow)
1307 {
1308         (void)priv;
1309         LIST_REMOVE(flow, next);
1310         if (flow->ibv_flow)
1311                 claim_zero(ibv_exp_destroy_flow(flow->ibv_flow));
1312         if (flow->qp)
1313                 claim_zero(ibv_destroy_qp(flow->qp));
1314         if (flow->ind_table)
1315                 claim_zero(ibv_exp_destroy_rwq_ind_table(flow->ind_table));
1316         if (flow->drop && flow->wq)
1317                 claim_zero(ibv_exp_destroy_wq(flow->wq));
1318         if (flow->drop && flow->cq)
1319                 claim_zero(ibv_destroy_cq(flow->cq));
1320         if (flow->mark) {
1321                 struct rte_flow *tmp;
1322                 struct rxq *rxq;
1323                 uint32_t mark_n = 0;
1324                 uint32_t queue_n;
1325
1326                 /*
1327                  * To remove the mark from the queue, the queue must not be
1328                  * present in any other marked flow (RSS or not).
1329                  */
1330                 for (queue_n = 0; queue_n < flow->rxqs_n; ++queue_n) {
1331                         rxq = (*flow->rxqs)[queue_n];
1332                         for (tmp = LIST_FIRST(&priv->flows);
1333                              tmp;
1334                              tmp = LIST_NEXT(tmp, next)) {
1335                                 uint32_t tqueue_n;
1336
1337                                 if (tmp->drop)
1338                                         continue;
1339                                 for (tqueue_n = 0;
1340                                      tqueue_n < tmp->rxqs_n;
1341                                      ++tqueue_n) {
1342                                         struct rxq *trxq;
1343
1344                                         trxq = (*tmp->rxqs)[tqueue_n];
1345                                         if (rxq == trxq)
1346                                                 ++mark_n;
1347                                 }
1348                         }
1349                         rxq->mark = !!mark_n;
1350                 }
1351         }
1352         rte_free(flow->ibv_attr);
1353         DEBUG("Flow destroyed %p", (void *)flow);
1354         rte_free(flow);
1355 }
1356
1357 /**
1358  * Destroy a flow.
1359  *
1360  * @see rte_flow_destroy()
1361  * @see rte_flow_ops
1362  */
1363 int
1364 mlx5_flow_destroy(struct rte_eth_dev *dev,
1365                   struct rte_flow *flow,
1366                   struct rte_flow_error *error)
1367 {
1368         struct priv *priv = dev->data->dev_private;
1369
1370         (void)error;
1371         priv_lock(priv);
1372         priv_flow_destroy(priv, flow);
1373         priv_unlock(priv);
1374         return 0;
1375 }
1376
1377 /**
1378  * Destroy all flows.
1379  *
1380  * @param priv
1381  *   Pointer to private structure.
1382  */
1383 static void
1384 priv_flow_flush(struct priv *priv)
1385 {
1386         while (!LIST_EMPTY(&priv->flows)) {
1387                 struct rte_flow *flow;
1388
1389                 flow = LIST_FIRST(&priv->flows);
1390                 priv_flow_destroy(priv, flow);
1391         }
1392 }
1393
1394 /**
1395  * Destroy all flows.
1396  *
1397  * @see rte_flow_flush()
1398  * @see rte_flow_ops
1399  */
1400 int
1401 mlx5_flow_flush(struct rte_eth_dev *dev,
1402                 struct rte_flow_error *error)
1403 {
1404         struct priv *priv = dev->data->dev_private;
1405
1406         (void)error;
1407         priv_lock(priv);
1408         priv_flow_flush(priv);
1409         priv_unlock(priv);
1410         return 0;
1411 }
1412
1413 /**
1414  * Remove all flows.
1415  *
1416  * Called by dev_stop() to remove all flows.
1417  *
1418  * @param priv
1419  *   Pointer to private structure.
1420  */
1421 void
1422 priv_flow_stop(struct priv *priv)
1423 {
1424         struct rte_flow *flow;
1425
1426         for (flow = LIST_FIRST(&priv->flows);
1427              flow;
1428              flow = LIST_NEXT(flow, next)) {
1429                 claim_zero(ibv_exp_destroy_flow(flow->ibv_flow));
1430                 flow->ibv_flow = NULL;
1431                 if (flow->mark) {
1432                         unsigned int n;
1433
1434                         for (n = 0; n < flow->rxqs_n; ++n)
1435                                 (*flow->rxqs)[n]->mark = 0;
1436                 }
1437                 DEBUG("Flow %p removed", (void *)flow);
1438         }
1439 }
1440
1441 /**
1442  * Add all flows.
1443  *
1444  * @param priv
1445  *   Pointer to private structure.
1446  *
1447  * @return
1448  *   0 on success, a errno value otherwise and rte_errno is set.
1449  */
1450 int
1451 priv_flow_start(struct priv *priv)
1452 {
1453         struct rte_flow *flow;
1454
1455         for (flow = LIST_FIRST(&priv->flows);
1456              flow;
1457              flow = LIST_NEXT(flow, next)) {
1458                 flow->ibv_flow = ibv_exp_create_flow(flow->qp,
1459                                                      flow->ibv_attr);
1460                 if (!flow->ibv_flow) {
1461                         DEBUG("Flow %p cannot be applied", (void *)flow);
1462                         rte_errno = EINVAL;
1463                         return rte_errno;
1464                 }
1465                 DEBUG("Flow %p applied", (void *)flow);
1466                 if (flow->mark) {
1467                         unsigned int n;
1468
1469                         for (n = 0; n < flow->rxqs_n; ++n)
1470                                 (*flow->rxqs)[n]->mark = 1;
1471                 }
1472         }
1473         return 0;
1474 }