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 *rxq; /**< Pointer to the queue, NULL if drop queue. */
  99         uint32_t mark:1; /**< Set if the flow is marked. */
 100 };
 101
 102 /** Static initializer for items. */
 103 #define ITEMS(...) \
 104         (const enum rte_flow_item_type []){ \
 105                 __VA_ARGS__, RTE_FLOW_ITEM_TYPE_END, \
 106         }
 107
 108 /** Structure to generate a simple graph of layers supported by the NIC. */
 109 struct mlx5_flow_items {
 110         /** List of possible actions for these items. */
 111         const enum rte_flow_action_type *const actions;
 112         /** Bit-masks corresponding to the possibilities for the item. */
 113         const void *mask;
 114         /**
 115          * Default bit-masks to use when item->mask is not provided. When
 116          * \default_mask is also NULL, the full supported bit-mask (\mask) is
 117          * used instead.
 118          */
 119         const void *default_mask;
 120         /** Bit-masks size in bytes. */
 121         const unsigned int mask_sz;
 122         /**
 123          * Conversion function from rte_flow to NIC specific flow.
 124          *
 125          * @param item
 126          *   rte_flow item to convert.
 127          * @param default_mask
 128          *   Default bit-masks to use when item->mask is not provided.
 129          * @param data
 130          *   Internal structure to store the conversion.
 131          *
 132          * @return
 133          *   0 on success, negative value otherwise.
 134          */
 135         int (*convert)(const struct rte_flow_item *item,
 136                        const void *default_mask,
 137                        void *data);
 138         /** Size in bytes of the destination structure. */
 139         const unsigned int dst_sz;
 140         /** List of possible following items.  */
 141         const enum rte_flow_item_type *const items;
 142 };
 143
 144 /** Valid action for this PMD. */
 145 static const enum rte_flow_action_type valid_actions[] = {
 146         RTE_FLOW_ACTION_TYPE_DROP,
 147         RTE_FLOW_ACTION_TYPE_QUEUE,
 148         RTE_FLOW_ACTION_TYPE_MARK,
 149         RTE_FLOW_ACTION_TYPE_FLAG,
 150         RTE_FLOW_ACTION_TYPE_END,
 151 };
 152
 153 /** Graph of supported items and associated actions. */
 154 static const struct mlx5_flow_items mlx5_flow_items[] = {
 155         [RTE_FLOW_ITEM_TYPE_END] = {
 156                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_ETH,
 157                                RTE_FLOW_ITEM_TYPE_VXLAN),
 158         },
 159         [RTE_FLOW_ITEM_TYPE_ETH] = {
 160                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_VLAN,
 161                                RTE_FLOW_ITEM_TYPE_IPV4,
 162                                RTE_FLOW_ITEM_TYPE_IPV6),
 163                 .actions = valid_actions,
 164                 .mask = &(const struct rte_flow_item_eth){
 165                         .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 166                         .src.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 167                         .type = -1,
 168                 },
 169                 .default_mask = &rte_flow_item_eth_mask,
 170                 .mask_sz = sizeof(struct rte_flow_item_eth),
 171                 .convert = mlx5_flow_create_eth,
 172                 .dst_sz = sizeof(struct ibv_exp_flow_spec_eth),
 173         },
 174         [RTE_FLOW_ITEM_TYPE_VLAN] = {
 175                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_IPV4,
 176                                RTE_FLOW_ITEM_TYPE_IPV6),
 177                 .actions = valid_actions,
 178                 .mask = &(const struct rte_flow_item_vlan){
 179                         .tci = -1,
 180                 },
 181                 .default_mask = &rte_flow_item_vlan_mask,
 182                 .mask_sz = sizeof(struct rte_flow_item_vlan),
 183                 .convert = mlx5_flow_create_vlan,
 184                 .dst_sz = 0,
 185         },
 186         [RTE_FLOW_ITEM_TYPE_IPV4] = {
 187                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP,
 188                                RTE_FLOW_ITEM_TYPE_TCP),
 189                 .actions = valid_actions,
 190                 .mask = &(const struct rte_flow_item_ipv4){
 191                         .hdr = {
 192                                 .src_addr = -1,
 193                                 .dst_addr = -1,
 194                                 .type_of_service = -1,
 195                                 .next_proto_id = -1,
 196                         },
 197                 },
 198                 .default_mask = &rte_flow_item_ipv4_mask,
 199                 .mask_sz = sizeof(struct rte_flow_item_ipv4),
 200                 .convert = mlx5_flow_create_ipv4,
 201                 .dst_sz = sizeof(struct ibv_exp_flow_spec_ipv4_ext),
 202         },
 203         [RTE_FLOW_ITEM_TYPE_IPV6] = {
 204                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP,
 205                                RTE_FLOW_ITEM_TYPE_TCP),
 206                 .actions = valid_actions,
 207                 .mask = &(const struct rte_flow_item_ipv6){
 208                         .hdr = {
 209                                 .src_addr = {
 210                                         0xff, 0xff, 0xff, 0xff,
 211                                         0xff, 0xff, 0xff, 0xff,
 212                                         0xff, 0xff, 0xff, 0xff,
 213                                         0xff, 0xff, 0xff, 0xff,
 214                                 },
 215                                 .dst_addr = {
 216                                         0xff, 0xff, 0xff, 0xff,
 217                                         0xff, 0xff, 0xff, 0xff,
 218                                         0xff, 0xff, 0xff, 0xff,
 219                                         0xff, 0xff, 0xff, 0xff,
 220                                 },
 221                                 .vtc_flow = -1,
 222                                 .proto = -1,
 223                                 .hop_limits = -1,
 224                         },
 225                 },
 226                 .default_mask = &rte_flow_item_ipv6_mask,
 227                 .mask_sz = sizeof(struct rte_flow_item_ipv6),
 228                 .convert = mlx5_flow_create_ipv6,
 229                 .dst_sz = sizeof(struct ibv_exp_flow_spec_ipv6_ext),
 230         },
 231         [RTE_FLOW_ITEM_TYPE_UDP] = {
 232                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_VXLAN),
 233                 .actions = valid_actions,
 234                 .mask = &(const struct rte_flow_item_udp){
 235                         .hdr = {
 236                                 .src_port = -1,
 237                                 .dst_port = -1,
 238                         },
 239                 },
 240                 .default_mask = &rte_flow_item_udp_mask,
 241                 .mask_sz = sizeof(struct rte_flow_item_udp),
 242                 .convert = mlx5_flow_create_udp,
 243                 .dst_sz = sizeof(struct ibv_exp_flow_spec_tcp_udp),
 244         },
 245         [RTE_FLOW_ITEM_TYPE_TCP] = {
 246                 .actions = valid_actions,
 247                 .mask = &(const struct rte_flow_item_tcp){
 248                         .hdr = {
 249                                 .src_port = -1,
 250                                 .dst_port = -1,
 251                         },
 252                 },
 253                 .default_mask = &rte_flow_item_tcp_mask,
 254                 .mask_sz = sizeof(struct rte_flow_item_tcp),
 255                 .convert = mlx5_flow_create_tcp,
 256                 .dst_sz = sizeof(struct ibv_exp_flow_spec_tcp_udp),
 257         },
 258         [RTE_FLOW_ITEM_TYPE_VXLAN] = {
 259                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_ETH),
 260                 .actions = valid_actions,
 261                 .mask = &(const struct rte_flow_item_vxlan){
 262                         .vni = "\xff\xff\xff",
 263                 },
 264                 .default_mask = &rte_flow_item_vxlan_mask,
 265                 .mask_sz = sizeof(struct rte_flow_item_vxlan),
 266                 .convert = mlx5_flow_create_vxlan,
 267                 .dst_sz = sizeof(struct ibv_exp_flow_spec_tunnel),
 268         },
 269 };
 270
 271 /** Structure to pass to the conversion function. */
 272 struct mlx5_flow {
 273         struct ibv_exp_flow_attr *ibv_attr; /**< Verbs attribute. */
 274         unsigned int offset; /**< Offset in bytes in the ibv_attr buffer. */
 275         uint32_t inner; /**< Set once VXLAN is encountered. */
 276 };
 277
 278 struct mlx5_flow_action {
 279         uint32_t queue:1; /**< Target is a receive queue. */
 280         uint32_t drop:1; /**< Target is a drop queue. */
 281         uint32_t mark:1; /**< Mark is present in the flow. */
 282         uint32_t queue_id; /**< Identifier of the queue. */
 283         uint32_t mark_id; /**< Mark identifier. */
 284 };
 285
 286 /**
 287  * Check support for a given item.
 288  *
 289  * @param item[in]
 290  *   Item specification.
 291  * @param mask[in]
 292  *   Bit-masks covering supported fields to compare with spec, last and mask in
 293  *   \item.
 294  * @param size
 295  *   Bit-Mask size in bytes.
 296  *
 297  * @return
 298  *   0 on success.
 299  */
 300 static int
 301 mlx5_flow_item_validate(const struct rte_flow_item *item,
 302                         const uint8_t *mask, unsigned int size)
 303 {
 304         int ret = 0;
 305
 306         if (!item->spec && (item->mask || item->last))
 307                 return -1;
 308         if (item->spec && !item->mask) {
 309                 unsigned int i;
 310                 const uint8_t *spec = item->spec;
 311
 312                 for (i = 0; i < size; ++i)
 313                         if ((spec[i] | mask[i]) != mask[i])
 314                                 return -1;
 315         }
 316         if (item->last && !item->mask) {
 317                 unsigned int i;
 318                 const uint8_t *spec = item->last;
 319
 320                 for (i = 0; i < size; ++i)
 321                         if ((spec[i] | mask[i]) != mask[i])
 322                                 return -1;
 323         }
 324         if (item->mask) {
 325                 unsigned int i;
 326                 const uint8_t *spec = item->mask;
 327
 328                 for (i = 0; i < size; ++i)
 329                         if ((spec[i] | mask[i]) != mask[i])
 330                                 return -1;
 331         }
 332         if (item->spec && item->last) {
 333                 uint8_t spec[size];
 334                 uint8_t last[size];
 335                 const uint8_t *apply = mask;
 336                 unsigned int i;
 337
 338                 if (item->mask)
 339                         apply = item->mask;
 340                 for (i = 0; i < size; ++i) {
 341                         spec[i] = ((const uint8_t *)item->spec)[i] & apply[i];
 342                         last[i] = ((const uint8_t *)item->last)[i] & apply[i];
 343                 }
 344                 ret = memcmp(spec, last, size);
 345         }
 346         return ret;
 347 }
 348
 349 /**
 350  * Validate a flow supported by the NIC.
 351  *
 352  * @param priv
 353  *   Pointer to private structure.
 354  * @param[in] attr
 355  *   Flow rule attributes.
 356  * @param[in] pattern
 357  *   Pattern specification (list terminated by the END pattern item).
 358  * @param[in] actions
 359  *   Associated actions (list terminated by the END action).
 360  * @param[out] error
 361  *   Perform verbose error reporting if not NULL.
 362  * @param[in, out] flow
 363  *   Flow structure to update.
 364  *
 365  * @return
 366  *   0 on success, a negative errno value otherwise and rte_errno is set.
 367  */
 368 static int
 369 priv_flow_validate(struct priv *priv,
 370                    const struct rte_flow_attr *attr,
 371                    const struct rte_flow_item items[],
 372                    const struct rte_flow_action actions[],
 373                    struct rte_flow_error *error,
 374                    struct mlx5_flow *flow)
 375 {
 376         const struct mlx5_flow_items *cur_item = mlx5_flow_items;
 377         struct mlx5_flow_action action = {
 378                 .queue = 0,
 379                 .drop = 0,
 380                 .mark = 0,
 381         };
 382
 383         (void)priv;
 384         if (attr->group) {
 385                 rte_flow_error_set(error, ENOTSUP,
 386                                    RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
 387                                    NULL,
 388                                    "groups are not supported");
 389                 return -rte_errno;
 390         }
 391         if (attr->priority) {
 392                 rte_flow_error_set(error, ENOTSUP,
 393                                    RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
 394                                    NULL,
 395                                    "priorities are not supported");
 396                 return -rte_errno;
 397         }
 398         if (attr->egress) {
 399                 rte_flow_error_set(error, ENOTSUP,
 400                                    RTE_FLOW_ERROR_TYPE_ATTR_EGRESS,
 401                                    NULL,
 402                                    "egress is not supported");
 403                 return -rte_errno;
 404         }
 405         if (!attr->ingress) {
 406                 rte_flow_error_set(error, ENOTSUP,
 407                                    RTE_FLOW_ERROR_TYPE_ATTR_INGRESS,
 408                                    NULL,
 409                                    "only ingress is supported");
 410                 return -rte_errno;
 411         }
 412         for (; items->type != RTE_FLOW_ITEM_TYPE_END; ++items) {
 413                 const struct mlx5_flow_items *token = NULL;
 414                 unsigned int i;
 415                 int err;
 416
 417                 if (items->type == RTE_FLOW_ITEM_TYPE_VOID)
 418                         continue;
 419                 for (i = 0;
 420                      cur_item->items &&
 421                      cur_item->items[i] != RTE_FLOW_ITEM_TYPE_END;
 422                      ++i) {
 423                         if (cur_item->items[i] == items->type) {
 424                                 token = &mlx5_flow_items[items->type];
 425                                 break;
 426                         }
 427                 }
 428                 if (!token)
 429                         goto exit_item_not_supported;
 430                 cur_item = token;
 431                 err = mlx5_flow_item_validate(items,
 432                                               (const uint8_t *)cur_item->mask,
 433                                               cur_item->mask_sz);
 434                 if (err)
 435                         goto exit_item_not_supported;
 436                 if (flow->ibv_attr && cur_item->convert) {
 437                         err = cur_item->convert(items,
 438                                                 (cur_item->default_mask ?
 439                                                  cur_item->default_mask :
 440                                                  cur_item->mask),
 441                                                 flow);
 442                         if (err)
 443                                 goto exit_item_not_supported;
 444                 }
 445                 flow->offset += cur_item->dst_sz;
 446         }
 447         for (; actions->type != RTE_FLOW_ACTION_TYPE_END; ++actions) {
 448                 if (actions->type == RTE_FLOW_ACTION_TYPE_VOID) {
 449                         continue;
 450                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_DROP) {
 451                         action.drop = 1;
 452                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_QUEUE) {
 453                         const struct rte_flow_action_queue *queue =
 454                                 (const struct rte_flow_action_queue *)
 455                                 actions->conf;
 456
 457                         if (!queue || (queue->index > (priv->rxqs_n - 1)))
 458                                 goto exit_action_not_supported;
 459                         action.queue = 1;
 460                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_MARK) {
 461                         const struct rte_flow_action_mark *mark =
 462                                 (const struct rte_flow_action_mark *)
 463                                 actions->conf;
 464
 465                         if (!mark) {
 466                                 rte_flow_error_set(error, EINVAL,
 467                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 468                                                    actions,
 469                                                    "mark must be defined");
 470                                 return -rte_errno;
 471                         } else if (mark->id >= MLX5_FLOW_MARK_MAX) {
 472                                 rte_flow_error_set(error, ENOTSUP,
 473                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 474                                                    actions,
 475                                                    "mark must be between 0"
 476                                                    " and 16777199");
 477                                 return -rte_errno;
 478                         }
 479                         action.mark = 1;
 480                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_FLAG) {
 481                         action.mark = 1;
 482                 } else {
 483                         goto exit_action_not_supported;
 484                 }
 485         }
 486         if (action.mark && !flow->ibv_attr && !action.drop)
 487                 flow->offset += sizeof(struct ibv_exp_flow_spec_action_tag);
 488         if (!action.queue && !action.drop) {
 489                 rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_HANDLE,
 490                                    NULL, "no valid action");
 491                 return -rte_errno;
 492         }
 493         return 0;
 494 exit_item_not_supported:
 495         rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 496                            items, "item not supported");
 497         return -rte_errno;
 498 exit_action_not_supported:
 499         rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
 500                            actions, "action not supported");
 501         return -rte_errno;
 502 }
 503
 504 /**
 505  * Validate a flow supported by the NIC.
 506  *
 507  * @see rte_flow_validate()
 508  * @see rte_flow_ops
 509  */
 510 int
 511 mlx5_flow_validate(struct rte_eth_dev *dev,
 512                    const struct rte_flow_attr *attr,
 513                    const struct rte_flow_item items[],
 514                    const struct rte_flow_action actions[],
 515                    struct rte_flow_error *error)
 516 {
 517         struct priv *priv = dev->data->dev_private;
 518         int ret;
 519         struct mlx5_flow flow = { .offset = sizeof(struct ibv_exp_flow_attr) };
 520
 521         priv_lock(priv);
 522         ret = priv_flow_validate(priv, attr, items, actions, error, &flow);
 523         priv_unlock(priv);
 524         return ret;
 525 }
 526
 527 /**
 528  * Convert Ethernet item to Verbs specification.
 529  *
 530  * @param item[in]
 531  *   Item specification.
 532  * @param default_mask[in]
 533  *   Default bit-masks to use when item->mask is not provided.
 534  * @param data[in, out]
 535  *   User structure.
 536  */
 537 static int
 538 mlx5_flow_create_eth(const struct rte_flow_item *item,
 539                      const void *default_mask,
 540                      void *data)
 541 {
 542         const struct rte_flow_item_eth *spec = item->spec;
 543         const struct rte_flow_item_eth *mask = item->mask;
 544         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 545         struct ibv_exp_flow_spec_eth *eth;
 546         const unsigned int eth_size = sizeof(struct ibv_exp_flow_spec_eth);
 547         unsigned int i;
 548
 549         ++flow->ibv_attr->num_of_specs;
 550         flow->ibv_attr->priority = 2;
 551         eth = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 552         *eth = (struct ibv_exp_flow_spec_eth) {
 553                 .type = flow->inner | IBV_EXP_FLOW_SPEC_ETH,
 554                 .size = eth_size,
 555         };
 556         if (!spec)
 557                 return 0;
 558         if (!mask)
 559                 mask = default_mask;
 560         memcpy(eth->val.dst_mac, spec->dst.addr_bytes, ETHER_ADDR_LEN);
 561         memcpy(eth->val.src_mac, spec->src.addr_bytes, ETHER_ADDR_LEN);
 562         eth->val.ether_type = spec->type;
 563         memcpy(eth->mask.dst_mac, mask->dst.addr_bytes, ETHER_ADDR_LEN);
 564         memcpy(eth->mask.src_mac, mask->src.addr_bytes, ETHER_ADDR_LEN);
 565         eth->mask.ether_type = mask->type;
 566         /* Remove unwanted bits from values. */
 567         for (i = 0; i < ETHER_ADDR_LEN; ++i) {
 568                 eth->val.dst_mac[i] &= eth->mask.dst_mac[i];
 569                 eth->val.src_mac[i] &= eth->mask.src_mac[i];
 570         }
 571         eth->val.ether_type &= eth->mask.ether_type;
 572         return 0;
 573 }
 574
 575 /**
 576  * Convert VLAN item to Verbs specification.
 577  *
 578  * @param item[in]
 579  *   Item specification.
 580  * @param default_mask[in]
 581  *   Default bit-masks to use when item->mask is not provided.
 582  * @param data[in, out]
 583  *   User structure.
 584  */
 585 static int
 586 mlx5_flow_create_vlan(const struct rte_flow_item *item,
 587                       const void *default_mask,
 588                       void *data)
 589 {
 590         const struct rte_flow_item_vlan *spec = item->spec;
 591         const struct rte_flow_item_vlan *mask = item->mask;
 592         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 593         struct ibv_exp_flow_spec_eth *eth;
 594         const unsigned int eth_size = sizeof(struct ibv_exp_flow_spec_eth);
 595
 596         eth = (void *)((uintptr_t)flow->ibv_attr + flow->offset - eth_size);
 597         if (!spec)
 598                 return 0;
 599         if (!mask)
 600                 mask = default_mask;
 601         eth->val.vlan_tag = spec->tci;
 602         eth->mask.vlan_tag = mask->tci;
 603         eth->val.vlan_tag &= eth->mask.vlan_tag;
 604         return 0;
 605 }
 606
 607 /**
 608  * Convert IPv4 item to Verbs specification.
 609  *
 610  * @param item[in]
 611  *   Item specification.
 612  * @param default_mask[in]
 613  *   Default bit-masks to use when item->mask is not provided.
 614  * @param data[in, out]
 615  *   User structure.
 616  */
 617 static int
 618 mlx5_flow_create_ipv4(const struct rte_flow_item *item,
 619                       const void *default_mask,
 620                       void *data)
 621 {
 622         const struct rte_flow_item_ipv4 *spec = item->spec;
 623         const struct rte_flow_item_ipv4 *mask = item->mask;
 624         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 625         struct ibv_exp_flow_spec_ipv4_ext *ipv4;
 626         unsigned int ipv4_size = sizeof(struct ibv_exp_flow_spec_ipv4_ext);
 627
 628         ++flow->ibv_attr->num_of_specs;
 629         flow->ibv_attr->priority = 1;
 630         ipv4 = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 631         *ipv4 = (struct ibv_exp_flow_spec_ipv4_ext) {
 632                 .type = flow->inner | IBV_EXP_FLOW_SPEC_IPV4_EXT,
 633                 .size = ipv4_size,
 634         };
 635         if (!spec)
 636                 return 0;
 637         if (!mask)
 638                 mask = default_mask;
 639         ipv4->val = (struct ibv_exp_flow_ipv4_ext_filter){
 640                 .src_ip = spec->hdr.src_addr,
 641                 .dst_ip = spec->hdr.dst_addr,
 642                 .proto = spec->hdr.next_proto_id,
 643                 .tos = spec->hdr.type_of_service,
 644         };
 645         ipv4->mask = (struct ibv_exp_flow_ipv4_ext_filter){
 646                 .src_ip = mask->hdr.src_addr,
 647                 .dst_ip = mask->hdr.dst_addr,
 648                 .proto = mask->hdr.next_proto_id,
 649                 .tos = mask->hdr.type_of_service,
 650         };
 651         /* Remove unwanted bits from values. */
 652         ipv4->val.src_ip &= ipv4->mask.src_ip;
 653         ipv4->val.dst_ip &= ipv4->mask.dst_ip;
 654         ipv4->val.proto &= ipv4->mask.proto;
 655         ipv4->val.tos &= ipv4->mask.tos;
 656         return 0;
 657 }
 658
 659 /**
 660  * Convert IPv6 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_ipv6(const struct rte_flow_item *item,
 671                       const void *default_mask,
 672                       void *data)
 673 {
 674         const struct rte_flow_item_ipv6 *spec = item->spec;
 675         const struct rte_flow_item_ipv6 *mask = item->mask;
 676         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 677         struct ibv_exp_flow_spec_ipv6_ext *ipv6;
 678         unsigned int ipv6_size = sizeof(struct ibv_exp_flow_spec_ipv6_ext);
 679
 680         ++flow->ibv_attr->num_of_specs;
 681         flow->ibv_attr->priority = 1;
 682         ipv6 = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 683         *ipv6 = (struct ibv_exp_flow_spec_ipv6_ext) {
 684                 .type = flow->inner | IBV_EXP_FLOW_SPEC_IPV6_EXT,
 685                 .size = ipv6_size,
 686         };
 687         if (!spec)
 688                 return 0;
 689         if (!mask)
 690                 mask = default_mask;
 691         memcpy(ipv6->val.src_ip, spec->hdr.src_addr,
 692                RTE_DIM(ipv6->val.src_ip));
 693         memcpy(ipv6->val.dst_ip, spec->hdr.dst_addr,
 694                RTE_DIM(ipv6->val.dst_ip));
 695         memcpy(ipv6->mask.src_ip, mask->hdr.src_addr,
 696                RTE_DIM(ipv6->mask.src_ip));
 697         memcpy(ipv6->mask.dst_ip, mask->hdr.dst_addr,
 698                RTE_DIM(ipv6->mask.dst_ip));
 699         ipv6->mask.flow_label = mask->hdr.vtc_flow;
 700         ipv6->mask.next_hdr = mask->hdr.proto;
 701         ipv6->mask.hop_limit = mask->hdr.hop_limits;
 702         ipv6->val.flow_label &= ipv6->mask.flow_label;
 703         ipv6->val.next_hdr &= ipv6->mask.next_hdr;
 704         ipv6->val.hop_limit &= ipv6->mask.hop_limit;
 705         return 0;
 706 }
 707
 708 /**
 709  * Convert UDP item to Verbs specification.
 710  *
 711  * @param item[in]
 712  *   Item specification.
 713  * @param default_mask[in]
 714  *   Default bit-masks to use when item->mask is not provided.
 715  * @param data[in, out]
 716  *   User structure.
 717  */
 718 static int
 719 mlx5_flow_create_udp(const struct rte_flow_item *item,
 720                      const void *default_mask,
 721                      void *data)
 722 {
 723         const struct rte_flow_item_udp *spec = item->spec;
 724         const struct rte_flow_item_udp *mask = item->mask;
 725         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 726         struct ibv_exp_flow_spec_tcp_udp *udp;
 727         unsigned int udp_size = sizeof(struct ibv_exp_flow_spec_tcp_udp);
 728
 729         ++flow->ibv_attr->num_of_specs;
 730         flow->ibv_attr->priority = 0;
 731         udp = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 732         *udp = (struct ibv_exp_flow_spec_tcp_udp) {
 733                 .type = flow->inner | IBV_EXP_FLOW_SPEC_UDP,
 734                 .size = udp_size,
 735         };
 736         if (!spec)
 737                 return 0;
 738         if (!mask)
 739                 mask = default_mask;
 740         udp->val.dst_port = spec->hdr.dst_port;
 741         udp->val.src_port = spec->hdr.src_port;
 742         udp->mask.dst_port = mask->hdr.dst_port;
 743         udp->mask.src_port = mask->hdr.src_port;
 744         /* Remove unwanted bits from values. */
 745         udp->val.src_port &= udp->mask.src_port;
 746         udp->val.dst_port &= udp->mask.dst_port;
 747         return 0;
 748 }
 749
 750 /**
 751  * Convert TCP item to Verbs specification.
 752  *
 753  * @param item[in]
 754  *   Item specification.
 755  * @param default_mask[in]
 756  *   Default bit-masks to use when item->mask is not provided.
 757  * @param data[in, out]
 758  *   User structure.
 759  */
 760 static int
 761 mlx5_flow_create_tcp(const struct rte_flow_item *item,
 762                      const void *default_mask,
 763                      void *data)
 764 {
 765         const struct rte_flow_item_tcp *spec = item->spec;
 766         const struct rte_flow_item_tcp *mask = item->mask;
 767         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 768         struct ibv_exp_flow_spec_tcp_udp *tcp;
 769         unsigned int tcp_size = sizeof(struct ibv_exp_flow_spec_tcp_udp);
 770
 771         ++flow->ibv_attr->num_of_specs;
 772         flow->ibv_attr->priority = 0;
 773         tcp = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 774         *tcp = (struct ibv_exp_flow_spec_tcp_udp) {
 775                 .type = flow->inner | IBV_EXP_FLOW_SPEC_TCP,
 776                 .size = tcp_size,
 777         };
 778         if (!spec)
 779                 return 0;
 780         if (!mask)
 781                 mask = default_mask;
 782         tcp->val.dst_port = spec->hdr.dst_port;
 783         tcp->val.src_port = spec->hdr.src_port;
 784         tcp->mask.dst_port = mask->hdr.dst_port;
 785         tcp->mask.src_port = mask->hdr.src_port;
 786         /* Remove unwanted bits from values. */
 787         tcp->val.src_port &= tcp->mask.src_port;
 788         tcp->val.dst_port &= tcp->mask.dst_port;
 789         return 0;
 790 }
 791
 792 /**
 793  * Convert VXLAN item to Verbs specification.
 794  *
 795  * @param item[in]
 796  *   Item specification.
 797  * @param default_mask[in]
 798  *   Default bit-masks to use when item->mask is not provided.
 799  * @param data[in, out]
 800  *   User structure.
 801  */
 802 static int
 803 mlx5_flow_create_vxlan(const struct rte_flow_item *item,
 804                        const void *default_mask,
 805                        void *data)
 806 {
 807         const struct rte_flow_item_vxlan *spec = item->spec;
 808         const struct rte_flow_item_vxlan *mask = item->mask;
 809         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 810         struct ibv_exp_flow_spec_tunnel *vxlan;
 811         unsigned int size = sizeof(struct ibv_exp_flow_spec_tunnel);
 812         union vni {
 813                 uint32_t vlan_id;
 814                 uint8_t vni[4];
 815         } id;
 816
 817         ++flow->ibv_attr->num_of_specs;
 818         flow->ibv_attr->priority = 0;
 819         id.vni[0] = 0;
 820         vxlan = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 821         *vxlan = (struct ibv_exp_flow_spec_tunnel) {
 822                 .type = flow->inner | IBV_EXP_FLOW_SPEC_VXLAN_TUNNEL,
 823                 .size = size,
 824         };
 825         flow->inner = IBV_EXP_FLOW_SPEC_INNER;
 826         if (!spec)
 827                 return 0;
 828         if (!mask)
 829                 mask = default_mask;
 830         memcpy(&id.vni[1], spec->vni, 3);
 831         vxlan->val.tunnel_id = id.vlan_id;
 832         memcpy(&id.vni[1], mask->vni, 3);
 833         vxlan->mask.tunnel_id = id.vlan_id;
 834         /* Remove unwanted bits from values. */
 835         vxlan->val.tunnel_id &= vxlan->mask.tunnel_id;
 836         return 0;
 837 }
 838
 839 /**
 840  * Convert mark/flag action to Verbs specification.
 841  *
 842  * @param flow
 843  *   Pointer to MLX5 flow structure.
 844  * @param mark_id
 845  *   Mark identifier.
 846  */
 847 static int
 848 mlx5_flow_create_flag_mark(struct mlx5_flow *flow, uint32_t mark_id)
 849 {
 850         struct ibv_exp_flow_spec_action_tag *tag;
 851         unsigned int size = sizeof(struct ibv_exp_flow_spec_action_tag);
 852
 853         tag = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 854         *tag = (struct ibv_exp_flow_spec_action_tag){
 855                 .type = IBV_EXP_FLOW_SPEC_ACTION_TAG,
 856                 .size = size,
 857                 .tag_id = mlx5_flow_mark_set(mark_id),
 858         };
 859         ++flow->ibv_attr->num_of_specs;
 860         return 0;
 861 }
 862
 863 /**
 864  * Complete flow rule creation.
 865  *
 866  * @param priv
 867  *   Pointer to private structure.
 868  * @param ibv_attr
 869  *   Verbs flow attributes.
 870  * @param action
 871  *   Target action structure.
 872  * @param[out] error
 873  *   Perform verbose error reporting if not NULL.
 874  *
 875  * @return
 876  *   A flow if the rule could be created.
 877  */
 878 static struct rte_flow *
 879 priv_flow_create_action_queue(struct priv *priv,
 880                               struct ibv_exp_flow_attr *ibv_attr,
 881                               struct mlx5_flow_action *action,
 882                               struct rte_flow_error *error)
 883 {
 884         struct rxq_ctrl *rxq;
 885         struct rte_flow *rte_flow;
 886
 887         assert(priv->pd);
 888         assert(priv->ctx);
 889         rte_flow = rte_calloc(__func__, 1, sizeof(*rte_flow), 0);
 890         if (!rte_flow) {
 891                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
 892                                    NULL, "cannot allocate flow memory");
 893                 return NULL;
 894         }
 895         if (action->drop) {
 896                 rte_flow->cq =
 897                         ibv_exp_create_cq(priv->ctx, 1, NULL, NULL, 0,
 898                                           &(struct ibv_exp_cq_init_attr){
 899                                                   .comp_mask = 0,
 900                                           });
 901                 if (!rte_flow->cq) {
 902                         rte_flow_error_set(error, ENOMEM,
 903                                            RTE_FLOW_ERROR_TYPE_HANDLE,
 904                                            NULL, "cannot allocate CQ");
 905                         goto error;
 906                 }
 907                 rte_flow->wq = ibv_exp_create_wq(priv->ctx,
 908                                                  &(struct ibv_exp_wq_init_attr){
 909                                                  .wq_type = IBV_EXP_WQT_RQ,
 910                                                  .max_recv_wr = 1,
 911                                                  .max_recv_sge = 1,
 912                                                  .pd = priv->pd,
 913                                                  .cq = rte_flow->cq,
 914                                                  });
 915                 if (!rte_flow->wq) {
 916                         rte_flow_error_set(error, ENOMEM,
 917                                            RTE_FLOW_ERROR_TYPE_HANDLE,
 918                                            NULL, "cannot allocate WQ");
 919                         goto error;
 920                 }
 921         } else {
 922                 rxq = container_of((*priv->rxqs)[action->queue_id],
 923                                    struct rxq_ctrl, rxq);
 924                 rte_flow->rxq = &rxq->rxq;
 925                 rxq->rxq.mark |= action->mark;
 926                 rte_flow->wq = rxq->wq;
 927         }
 928         rte_flow->mark = action->mark;
 929         rte_flow->ibv_attr = ibv_attr;
 930         rte_flow->ind_table = ibv_exp_create_rwq_ind_table(
 931                 priv->ctx,
 932                 &(struct ibv_exp_rwq_ind_table_init_attr){
 933                         .pd = priv->pd,
 934                         .log_ind_tbl_size = 0,
 935                         .ind_tbl = &rte_flow->wq,
 936                         .comp_mask = 0,
 937                 });
 938         if (!rte_flow->ind_table) {
 939                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
 940                                    NULL, "cannot allocate indirection table");
 941                 goto error;
 942         }
 943         rte_flow->qp = ibv_exp_create_qp(
 944                 priv->ctx,
 945                 &(struct ibv_exp_qp_init_attr){
 946                         .qp_type = IBV_QPT_RAW_PACKET,
 947                         .comp_mask =
 948                                 IBV_EXP_QP_INIT_ATTR_PD |
 949                                 IBV_EXP_QP_INIT_ATTR_PORT |
 950                                 IBV_EXP_QP_INIT_ATTR_RX_HASH,
 951                         .pd = priv->pd,
 952                         .rx_hash_conf = &(struct ibv_exp_rx_hash_conf){
 953                                 .rx_hash_function =
 954                                         IBV_EXP_RX_HASH_FUNC_TOEPLITZ,
 955                                 .rx_hash_key_len = rss_hash_default_key_len,
 956                                 .rx_hash_key = rss_hash_default_key,
 957                                 .rx_hash_fields_mask = 0,
 958                                 .rwq_ind_tbl = rte_flow->ind_table,
 959                         },
 960                         .port_num = priv->port,
 961                 });
 962         if (!rte_flow->qp) {
 963                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
 964                                    NULL, "cannot allocate QP");
 965                 goto error;
 966         }
 967         if (!priv->started)
 968                 return rte_flow;
 969         rte_flow->ibv_flow = ibv_exp_create_flow(rte_flow->qp,
 970                                                  rte_flow->ibv_attr);
 971         if (!rte_flow->ibv_flow) {
 972                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
 973                                    NULL, "flow rule creation failure");
 974                 goto error;
 975         }
 976         return rte_flow;
 977 error:
 978         assert(rte_flow);
 979         if (rte_flow->qp)
 980                 ibv_destroy_qp(rte_flow->qp);
 981         if (rte_flow->ind_table)
 982                 ibv_exp_destroy_rwq_ind_table(rte_flow->ind_table);
 983         if (!rte_flow->rxq && rte_flow->wq)
 984                 ibv_exp_destroy_wq(rte_flow->wq);
 985         if (!rte_flow->rxq && rte_flow->cq)
 986                 ibv_destroy_cq(rte_flow->cq);
 987         rte_free(rte_flow);
 988         return NULL;
 989 }
 990
 991 /**
 992  * Convert a flow.
 993  *
 994  * @param priv
 995  *   Pointer to private structure.
 996  * @param[in] attr
 997  *   Flow rule attributes.
 998  * @param[in] pattern
 999  *   Pattern specification (list terminated by the END pattern item).
1000  * @param[in] actions
1001  *   Associated actions (list terminated by the END action).
1002  * @param[out] error
1003  *   Perform verbose error reporting if not NULL.
1004  *
1005  * @return
1006  *   A flow on success, NULL otherwise.
1007  */
1008 static struct rte_flow *
1009 priv_flow_create(struct priv *priv,
1010                  const struct rte_flow_attr *attr,
1011                  const struct rte_flow_item items[],
1012                  const struct rte_flow_action actions[],
1013                  struct rte_flow_error *error)
1014 {
1015         struct rte_flow *rte_flow;
1016         struct mlx5_flow_action action;
1017         struct mlx5_flow flow = { .offset = sizeof(struct ibv_exp_flow_attr), };
1018         int err;
1019
1020         err = priv_flow_validate(priv, attr, items, actions, error, &flow);
1021         if (err)
1022                 goto exit;
1023         flow.ibv_attr = rte_malloc(__func__, flow.offset, 0);
1024         flow.offset = sizeof(struct ibv_exp_flow_attr);
1025         if (!flow.ibv_attr) {
1026                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1027                                    NULL, "cannot allocate ibv_attr memory");
1028                 goto exit;
1029         }
1030         *flow.ibv_attr = (struct ibv_exp_flow_attr){
1031                 .type = IBV_EXP_FLOW_ATTR_NORMAL,
1032                 .size = sizeof(struct ibv_exp_flow_attr),
1033                 .priority = attr->priority,
1034                 .num_of_specs = 0,
1035                 .port = 0,
1036                 .flags = 0,
1037                 .reserved = 0,
1038         };
1039         flow.inner = 0;
1040         claim_zero(priv_flow_validate(priv, attr, items, actions,
1041                                       error, &flow));
1042         action = (struct mlx5_flow_action){
1043                 .queue = 0,
1044                 .drop = 0,
1045                 .mark = 0,
1046                 .mark_id = MLX5_FLOW_MARK_DEFAULT,
1047         };
1048         for (; actions->type != RTE_FLOW_ACTION_TYPE_END; ++actions) {
1049                 if (actions->type == RTE_FLOW_ACTION_TYPE_VOID) {
1050                         continue;
1051                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_QUEUE) {
1052                         action.queue = 1;
1053                         action.queue_id =
1054                                 ((const struct rte_flow_action_queue *)
1055                                  actions->conf)->index;
1056                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_DROP) {
1057                         action.drop = 1;
1058                         action.mark = 0;
1059                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_MARK) {
1060                         const struct rte_flow_action_mark *mark =
1061                                 (const struct rte_flow_action_mark *)
1062                                 actions->conf;
1063
1064                         if (mark)
1065                                 action.mark_id = mark->id;
1066                         action.mark = !action.drop;
1067                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_FLAG) {
1068                         action.mark = 1;
1069                 } else {
1070                         rte_flow_error_set(error, ENOTSUP,
1071                                            RTE_FLOW_ERROR_TYPE_ACTION,
1072                                            actions, "unsupported action");
1073                         goto exit;
1074                 }
1075         }
1076         if (action.mark) {
1077                 mlx5_flow_create_flag_mark(&flow, action.mark_id);
1078                 flow.offset += sizeof(struct ibv_exp_flow_spec_action_tag);
1079         }
1080         rte_flow = priv_flow_create_action_queue(priv, flow.ibv_attr,
1081                                                  &action, error);
1082         if (!rte_flow)
1083                 goto exit;
1084         return rte_flow;
1085 exit:
1086         rte_free(flow.ibv_attr);
1087         return NULL;
1088 }
1089
1090 /**
1091  * Create a flow.
1092  *
1093  * @see rte_flow_create()
1094  * @see rte_flow_ops
1095  */
1096 struct rte_flow *
1097 mlx5_flow_create(struct rte_eth_dev *dev,
1098                  const struct rte_flow_attr *attr,
1099                  const struct rte_flow_item items[],
1100                  const struct rte_flow_action actions[],
1101                  struct rte_flow_error *error)
1102 {
1103         struct priv *priv = dev->data->dev_private;
1104         struct rte_flow *flow;
1105
1106         priv_lock(priv);
1107         flow = priv_flow_create(priv, attr, items, actions, error);
1108         if (flow) {
1109                 LIST_INSERT_HEAD(&priv->flows, flow, next);
1110                 DEBUG("Flow created %p", (void *)flow);
1111         }
1112         priv_unlock(priv);
1113         return flow;
1114 }
1115
1116 /**
1117  * Destroy a flow.
1118  *
1119  * @param priv
1120  *   Pointer to private structure.
1121  * @param[in] flow
1122  *   Flow to destroy.
1123  */
1124 static void
1125 priv_flow_destroy(struct priv *priv,
1126                   struct rte_flow *flow)
1127 {
1128         (void)priv;
1129         LIST_REMOVE(flow, next);
1130         if (flow->ibv_flow)
1131                 claim_zero(ibv_exp_destroy_flow(flow->ibv_flow));
1132         if (flow->qp)
1133                 claim_zero(ibv_destroy_qp(flow->qp));
1134         if (flow->ind_table)
1135                 claim_zero(ibv_exp_destroy_rwq_ind_table(flow->ind_table));
1136         if (!flow->rxq && flow->wq)
1137                 claim_zero(ibv_exp_destroy_wq(flow->wq));
1138         if (!flow->rxq && flow->cq)
1139                 claim_zero(ibv_destroy_cq(flow->cq));
1140         if (flow->mark) {
1141                 struct rte_flow *tmp;
1142                 uint32_t mark_n = 0;
1143
1144                 for (tmp = LIST_FIRST(&priv->flows);
1145                      tmp;
1146                      tmp = LIST_NEXT(tmp, next)) {
1147                         if ((flow->rxq == tmp->rxq) && tmp->mark)
1148                                 ++mark_n;
1149                 }
1150                 flow->rxq->mark = !!mark_n;
1151         }
1152         rte_free(flow->ibv_attr);
1153         DEBUG("Flow destroyed %p", (void *)flow);
1154         rte_free(flow);
1155 }
1156
1157 /**
1158  * Destroy a flow.
1159  *
1160  * @see rte_flow_destroy()
1161  * @see rte_flow_ops
1162  */
1163 int
1164 mlx5_flow_destroy(struct rte_eth_dev *dev,
1165                   struct rte_flow *flow,
1166                   struct rte_flow_error *error)
1167 {
1168         struct priv *priv = dev->data->dev_private;
1169
1170         (void)error;
1171         priv_lock(priv);
1172         priv_flow_destroy(priv, flow);
1173         priv_unlock(priv);
1174         return 0;
1175 }
1176
1177 /**
1178  * Destroy all flows.
1179  *
1180  * @param priv
1181  *   Pointer to private structure.
1182  */
1183 static void
1184 priv_flow_flush(struct priv *priv)
1185 {
1186         while (!LIST_EMPTY(&priv->flows)) {
1187                 struct rte_flow *flow;
1188
1189                 flow = LIST_FIRST(&priv->flows);
1190                 priv_flow_destroy(priv, flow);
1191         }
1192 }
1193
1194 /**
1195  * Destroy all flows.
1196  *
1197  * @see rte_flow_flush()
1198  * @see rte_flow_ops
1199  */
1200 int
1201 mlx5_flow_flush(struct rte_eth_dev *dev,
1202                 struct rte_flow_error *error)
1203 {
1204         struct priv *priv = dev->data->dev_private;
1205
1206         (void)error;
1207         priv_lock(priv);
1208         priv_flow_flush(priv);
1209         priv_unlock(priv);
1210         return 0;
1211 }
1212
1213 /**
1214  * Remove all flows.
1215  *
1216  * Called by dev_stop() to remove all flows.
1217  *
1218  * @param priv
1219  *   Pointer to private structure.
1220  */
1221 void
1222 priv_flow_stop(struct priv *priv)
1223 {
1224         struct rte_flow *flow;
1225
1226         for (flow = LIST_FIRST(&priv->flows);
1227              flow;
1228              flow = LIST_NEXT(flow, next)) {
1229                 claim_zero(ibv_exp_destroy_flow(flow->ibv_flow));
1230                 flow->ibv_flow = NULL;
1231                 if (flow->mark)
1232                         flow->rxq->mark = 0;
1233                 DEBUG("Flow %p removed", (void *)flow);
1234         }
1235 }
1236
1237 /**
1238  * Add all flows.
1239  *
1240  * @param priv
1241  *   Pointer to private structure.
1242  *
1243  * @return
1244  *   0 on success, a errno value otherwise and rte_errno is set.
1245  */
1246 int
1247 priv_flow_start(struct priv *priv)
1248 {
1249         struct rte_flow *flow;
1250
1251         for (flow = LIST_FIRST(&priv->flows);
1252              flow;
1253              flow = LIST_NEXT(flow, next)) {
1254                 flow->ibv_flow = ibv_exp_create_flow(flow->qp,
1255                                                      flow->ibv_attr);
1256                 if (!flow->ibv_flow) {
1257                         DEBUG("Flow %p cannot be applied", (void *)flow);
1258                         rte_errno = EINVAL;
1259                         return rte_errno;
1260                 }
1261                 DEBUG("Flow %p applied", (void *)flow);
1262                 if (flow->rxq)
1263                         flow->rxq->mark |= flow->mark;
1264         }
1265         return 0;
1266 }