drivers/net/mlx5/mlx5_flow.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright 2016 6WIND S.A.
   5  *   Copyright 2016 Mellanox.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of 6WIND S.A. nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  */
  33
  34 #include <sys/queue.h>
  35 #include <string.h>
  36
  37 /* Verbs header. */
  38 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
  39 #ifdef PEDANTIC
  40 #pragma GCC diagnostic ignored "-Wpedantic"
  41 #endif
  42 #include <infiniband/verbs.h>
  43 #ifdef PEDANTIC
  44 #pragma GCC diagnostic error "-Wpedantic"
  45 #endif
  46
  47 #include <rte_ethdev.h>
  48 #include <rte_flow.h>
  49 #include <rte_flow_driver.h>
  50 #include <rte_malloc.h>
  51
  52 #include "mlx5.h"
  53 #include "mlx5_prm.h"
  54
  55 /* Define minimal priority for control plane flows. */
  56 #define MLX5_CTRL_FLOW_PRIORITY 4
  57
  58 static int
  59 mlx5_flow_create_eth(const struct rte_flow_item *item,
  60                      const void *default_mask,
  61                      void *data);
  62
  63 static int
  64 mlx5_flow_create_vlan(const struct rte_flow_item *item,
  65                       const void *default_mask,
  66                       void *data);
  67
  68 static int
  69 mlx5_flow_create_ipv4(const struct rte_flow_item *item,
  70                       const void *default_mask,
  71                       void *data);
  72
  73 static int
  74 mlx5_flow_create_ipv6(const struct rte_flow_item *item,
  75                       const void *default_mask,
  76                       void *data);
  77
  78 static int
  79 mlx5_flow_create_udp(const struct rte_flow_item *item,
  80                      const void *default_mask,
  81                      void *data);
  82
  83 static int
  84 mlx5_flow_create_tcp(const struct rte_flow_item *item,
  85                      const void *default_mask,
  86                      void *data);
  87
  88 static int
  89 mlx5_flow_create_vxlan(const struct rte_flow_item *item,
  90                        const void *default_mask,
  91                        void *data);
  92
  93 /** Structure for Drop queue. */
  94 struct mlx5_hrxq_drop {
  95         struct ibv_rwq_ind_table *ind_table; /**< Indirection table. */
  96         struct ibv_qp *qp; /**< Verbs queue pair. */
  97         struct ibv_wq *wq; /**< Verbs work queue. */
  98         struct ibv_cq *cq; /**< Verbs completion queue. */
  99 };
 100
 101 /* Flows structures. */
 102 struct mlx5_flow {
 103         uint64_t hash_fields; /**< Fields that participate in the hash. */
 104         struct mlx5_hrxq *hrxq; /**< Hash Rx queues. */
 105 };
 106
 107 /* Drop flows structures. */
 108 struct mlx5_flow_drop {
 109         struct mlx5_hrxq_drop hrxq; /**< Drop hash Rx queue. */
 110 };
 111
 112 struct rte_flow {
 113         TAILQ_ENTRY(rte_flow) next; /**< Pointer to the next flow structure. */
 114         uint32_t mark:1; /**< Set if the flow is marked. */
 115         uint32_t drop:1; /**< Drop queue. */
 116         struct ibv_flow_attr *ibv_attr; /**< Pointer to Verbs attributes. */
 117         struct ibv_flow *ibv_flow; /**< Verbs flow. */
 118         uint16_t queues_n; /**< Number of entries in queue[]. */
 119         uint16_t (*queues)[]; /**< Queues indexes to use. */
 120         union {
 121                 struct mlx5_flow frxq; /**< Flow with Rx queue. */
 122                 struct mlx5_flow_drop drxq; /**< Flow with drop Rx queue. */
 123         };
 124 };
 125
 126 /** Static initializer for items. */
 127 #define ITEMS(...) \
 128         (const enum rte_flow_item_type []){ \
 129                 __VA_ARGS__, RTE_FLOW_ITEM_TYPE_END, \
 130         }
 131
 132 /** Structure to generate a simple graph of layers supported by the NIC. */
 133 struct mlx5_flow_items {
 134         /** List of possible actions for these items. */
 135         const enum rte_flow_action_type *const actions;
 136         /** Bit-masks corresponding to the possibilities for the item. */
 137         const void *mask;
 138         /**
 139          * Default bit-masks to use when item->mask is not provided. When
 140          * \default_mask is also NULL, the full supported bit-mask (\mask) is
 141          * used instead.
 142          */
 143         const void *default_mask;
 144         /** Bit-masks size in bytes. */
 145         const unsigned int mask_sz;
 146         /**
 147          * Conversion function from rte_flow to NIC specific flow.
 148          *
 149          * @param item
 150          *   rte_flow item to convert.
 151          * @param default_mask
 152          *   Default bit-masks to use when item->mask is not provided.
 153          * @param data
 154          *   Internal structure to store the conversion.
 155          *
 156          * @return
 157          *   0 on success, negative value otherwise.
 158          */
 159         int (*convert)(const struct rte_flow_item *item,
 160                        const void *default_mask,
 161                        void *data);
 162         /** Size in bytes of the destination structure. */
 163         const unsigned int dst_sz;
 164         /** List of possible following items.  */
 165         const enum rte_flow_item_type *const items;
 166 };
 167
 168 /** Valid action for this PMD. */
 169 static const enum rte_flow_action_type valid_actions[] = {
 170         RTE_FLOW_ACTION_TYPE_DROP,
 171         RTE_FLOW_ACTION_TYPE_QUEUE,
 172         RTE_FLOW_ACTION_TYPE_MARK,
 173         RTE_FLOW_ACTION_TYPE_FLAG,
 174         RTE_FLOW_ACTION_TYPE_END,
 175 };
 176
 177 /** Graph of supported items and associated actions. */
 178 static const struct mlx5_flow_items mlx5_flow_items[] = {
 179         [RTE_FLOW_ITEM_TYPE_END] = {
 180                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_ETH,
 181                                RTE_FLOW_ITEM_TYPE_VXLAN),
 182         },
 183         [RTE_FLOW_ITEM_TYPE_ETH] = {
 184                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_VLAN,
 185                                RTE_FLOW_ITEM_TYPE_IPV4,
 186                                RTE_FLOW_ITEM_TYPE_IPV6),
 187                 .actions = valid_actions,
 188                 .mask = &(const struct rte_flow_item_eth){
 189                         .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 190                         .src.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 191                         .type = -1,
 192                 },
 193                 .default_mask = &rte_flow_item_eth_mask,
 194                 .mask_sz = sizeof(struct rte_flow_item_eth),
 195                 .convert = mlx5_flow_create_eth,
 196                 .dst_sz = sizeof(struct ibv_flow_spec_eth),
 197         },
 198         [RTE_FLOW_ITEM_TYPE_VLAN] = {
 199                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_IPV4,
 200                                RTE_FLOW_ITEM_TYPE_IPV6),
 201                 .actions = valid_actions,
 202                 .mask = &(const struct rte_flow_item_vlan){
 203                         .tci = -1,
 204                 },
 205                 .default_mask = &rte_flow_item_vlan_mask,
 206                 .mask_sz = sizeof(struct rte_flow_item_vlan),
 207                 .convert = mlx5_flow_create_vlan,
 208                 .dst_sz = 0,
 209         },
 210         [RTE_FLOW_ITEM_TYPE_IPV4] = {
 211                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP,
 212                                RTE_FLOW_ITEM_TYPE_TCP),
 213                 .actions = valid_actions,
 214                 .mask = &(const struct rte_flow_item_ipv4){
 215                         .hdr = {
 216                                 .src_addr = -1,
 217                                 .dst_addr = -1,
 218                                 .type_of_service = -1,
 219                                 .next_proto_id = -1,
 220                         },
 221                 },
 222                 .default_mask = &rte_flow_item_ipv4_mask,
 223                 .mask_sz = sizeof(struct rte_flow_item_ipv4),
 224                 .convert = mlx5_flow_create_ipv4,
 225                 .dst_sz = sizeof(struct ibv_flow_spec_ipv4),
 226         },
 227         [RTE_FLOW_ITEM_TYPE_IPV6] = {
 228                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP,
 229                                RTE_FLOW_ITEM_TYPE_TCP),
 230                 .actions = valid_actions,
 231                 .mask = &(const struct rte_flow_item_ipv6){
 232                         .hdr = {
 233                                 .src_addr = {
 234                                         0xff, 0xff, 0xff, 0xff,
 235                                         0xff, 0xff, 0xff, 0xff,
 236                                         0xff, 0xff, 0xff, 0xff,
 237                                         0xff, 0xff, 0xff, 0xff,
 238                                 },
 239                                 .dst_addr = {
 240                                         0xff, 0xff, 0xff, 0xff,
 241                                         0xff, 0xff, 0xff, 0xff,
 242                                         0xff, 0xff, 0xff, 0xff,
 243                                         0xff, 0xff, 0xff, 0xff,
 244                                 },
 245                                 .vtc_flow = -1,
 246                                 .proto = -1,
 247                                 .hop_limits = -1,
 248                         },
 249                 },
 250                 .default_mask = &rte_flow_item_ipv6_mask,
 251                 .mask_sz = sizeof(struct rte_flow_item_ipv6),
 252                 .convert = mlx5_flow_create_ipv6,
 253                 .dst_sz = sizeof(struct ibv_flow_spec_ipv6),
 254         },
 255         [RTE_FLOW_ITEM_TYPE_UDP] = {
 256                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_VXLAN),
 257                 .actions = valid_actions,
 258                 .mask = &(const struct rte_flow_item_udp){
 259                         .hdr = {
 260                                 .src_port = -1,
 261                                 .dst_port = -1,
 262                         },
 263                 },
 264                 .default_mask = &rte_flow_item_udp_mask,
 265                 .mask_sz = sizeof(struct rte_flow_item_udp),
 266                 .convert = mlx5_flow_create_udp,
 267                 .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
 268         },
 269         [RTE_FLOW_ITEM_TYPE_TCP] = {
 270                 .actions = valid_actions,
 271                 .mask = &(const struct rte_flow_item_tcp){
 272                         .hdr = {
 273                                 .src_port = -1,
 274                                 .dst_port = -1,
 275                         },
 276                 },
 277                 .default_mask = &rte_flow_item_tcp_mask,
 278                 .mask_sz = sizeof(struct rte_flow_item_tcp),
 279                 .convert = mlx5_flow_create_tcp,
 280                 .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
 281         },
 282         [RTE_FLOW_ITEM_TYPE_VXLAN] = {
 283                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_ETH),
 284                 .actions = valid_actions,
 285                 .mask = &(const struct rte_flow_item_vxlan){
 286                         .vni = "\xff\xff\xff",
 287                 },
 288                 .default_mask = &rte_flow_item_vxlan_mask,
 289                 .mask_sz = sizeof(struct rte_flow_item_vxlan),
 290                 .convert = mlx5_flow_create_vxlan,
 291                 .dst_sz = sizeof(struct ibv_flow_spec_tunnel),
 292         },
 293 };
 294
 295 /* Structure to parse actions. */
 296 struct mlx5_flow_action {
 297         uint32_t queue:1; /**< Target is a receive queue. */
 298         uint32_t drop:1; /**< Target is a drop queue. */
 299         uint32_t mark:1; /**< Mark is present in the flow. */
 300         uint32_t mark_id; /**< Mark identifier. */
 301         uint16_t queues[RTE_MAX_QUEUES_PER_PORT]; /**< Queues indexes to use. */
 302         uint16_t queues_n; /**< Number of entries in queue[]. */
 303 };
 304
 305 /** Structure to pass to the conversion function. */
 306 struct mlx5_flow_parse {
 307         struct ibv_flow_attr *ibv_attr; /**< Verbs attribute. */
 308         unsigned int offset; /**< Offset in bytes in the ibv_attr buffer. */
 309         uint32_t inner; /**< Set once VXLAN is encountered. */
 310         uint64_t hash_fields; /**< Fields that participate in the hash. */
 311         struct mlx5_flow_action actions; /**< Parsed action result. */
 312 };
 313
 314 static const struct rte_flow_ops mlx5_flow_ops = {
 315         .validate = mlx5_flow_validate,
 316         .create = mlx5_flow_create,
 317         .destroy = mlx5_flow_destroy,
 318         .flush = mlx5_flow_flush,
 319         .query = NULL,
 320         .isolate = mlx5_flow_isolate,
 321 };
 322
 323 /**
 324  * Manage filter operations.
 325  *
 326  * @param dev
 327  *   Pointer to Ethernet device structure.
 328  * @param filter_type
 329  *   Filter type.
 330  * @param filter_op
 331  *   Operation to perform.
 332  * @param arg
 333  *   Pointer to operation-specific structure.
 334  *
 335  * @return
 336  *   0 on success, negative errno value on failure.
 337  */
 338 int
 339 mlx5_dev_filter_ctrl(struct rte_eth_dev *dev,
 340                      enum rte_filter_type filter_type,
 341                      enum rte_filter_op filter_op,
 342                      void *arg)
 343 {
 344         int ret = EINVAL;
 345
 346         if (filter_type == RTE_ETH_FILTER_GENERIC) {
 347                 if (filter_op != RTE_ETH_FILTER_GET)
 348                         return -EINVAL;
 349                 *(const void **)arg = &mlx5_flow_ops;
 350                 return 0;
 351         }
 352         ERROR("%p: filter type (%d) not supported",
 353               (void *)dev, filter_type);
 354         return -ret;
 355 }
 356
 357 /**
 358  * Check support for a given item.
 359  *
 360  * @param item[in]
 361  *   Item specification.
 362  * @param mask[in]
 363  *   Bit-masks covering supported fields to compare with spec, last and mask in
 364  *   \item.
 365  * @param size
 366  *   Bit-Mask size in bytes.
 367  *
 368  * @return
 369  *   0 on success.
 370  */
 371 static int
 372 mlx5_flow_item_validate(const struct rte_flow_item *item,
 373                         const uint8_t *mask, unsigned int size)
 374 {
 375         int ret = 0;
 376
 377         if (!item->spec && (item->mask || item->last))
 378                 return -1;
 379         if (item->spec && !item->mask) {
 380                 unsigned int i;
 381                 const uint8_t *spec = item->spec;
 382
 383                 for (i = 0; i < size; ++i)
 384                         if ((spec[i] | mask[i]) != mask[i])
 385                                 return -1;
 386         }
 387         if (item->last && !item->mask) {
 388                 unsigned int i;
 389                 const uint8_t *spec = item->last;
 390
 391                 for (i = 0; i < size; ++i)
 392                         if ((spec[i] | mask[i]) != mask[i])
 393                                 return -1;
 394         }
 395         if (item->mask) {
 396                 unsigned int i;
 397                 const uint8_t *spec = item->mask;
 398
 399                 for (i = 0; i < size; ++i)
 400                         if ((spec[i] | mask[i]) != mask[i])
 401                                 return -1;
 402         }
 403         if (item->spec && item->last) {
 404                 uint8_t spec[size];
 405                 uint8_t last[size];
 406                 const uint8_t *apply = mask;
 407                 unsigned int i;
 408
 409                 if (item->mask)
 410                         apply = item->mask;
 411                 for (i = 0; i < size; ++i) {
 412                         spec[i] = ((const uint8_t *)item->spec)[i] & apply[i];
 413                         last[i] = ((const uint8_t *)item->last)[i] & apply[i];
 414                 }
 415                 ret = memcmp(spec, last, size);
 416         }
 417         return ret;
 418 }
 419
 420 /**
 421  * Validate a flow supported by the NIC.
 422  *
 423  * @param priv
 424  *   Pointer to private structure.
 425  * @param[in] attr
 426  *   Flow rule attributes.
 427  * @param[in] pattern
 428  *   Pattern specification (list terminated by the END pattern item).
 429  * @param[in] actions
 430  *   Associated actions (list terminated by the END action).
 431  * @param[out] error
 432  *   Perform verbose error reporting if not NULL.
 433  * @param[in, out] flow
 434  *   Flow structure to update.
 435  *
 436  * @return
 437  *   0 on success, a negative errno value otherwise and rte_errno is set.
 438  */
 439 static int
 440 priv_flow_validate(struct priv *priv,
 441                    const struct rte_flow_attr *attr,
 442                    const struct rte_flow_item items[],
 443                    const struct rte_flow_action actions[],
 444                    struct rte_flow_error *error,
 445                    struct mlx5_flow_parse *flow)
 446 {
 447         const struct mlx5_flow_items *cur_item = mlx5_flow_items;
 448
 449         (void)priv;
 450         if (attr->group) {
 451                 rte_flow_error_set(error, ENOTSUP,
 452                                    RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
 453                                    NULL,
 454                                    "groups are not supported");
 455                 return -rte_errno;
 456         }
 457         if (attr->priority && attr->priority != MLX5_CTRL_FLOW_PRIORITY) {
 458                 rte_flow_error_set(error, ENOTSUP,
 459                                    RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
 460                                    NULL,
 461                                    "priorities are not supported");
 462                 return -rte_errno;
 463         }
 464         if (attr->egress) {
 465                 rte_flow_error_set(error, ENOTSUP,
 466                                    RTE_FLOW_ERROR_TYPE_ATTR_EGRESS,
 467                                    NULL,
 468                                    "egress is not supported");
 469                 return -rte_errno;
 470         }
 471         if (!attr->ingress) {
 472                 rte_flow_error_set(error, ENOTSUP,
 473                                    RTE_FLOW_ERROR_TYPE_ATTR_INGRESS,
 474                                    NULL,
 475                                    "only ingress is supported");
 476                 return -rte_errno;
 477         }
 478         for (; items->type != RTE_FLOW_ITEM_TYPE_END; ++items) {
 479                 const struct mlx5_flow_items *token = NULL;
 480                 unsigned int i;
 481                 int err;
 482
 483                 if (items->type == RTE_FLOW_ITEM_TYPE_VOID)
 484                         continue;
 485                 for (i = 0;
 486                      cur_item->items &&
 487                      cur_item->items[i] != RTE_FLOW_ITEM_TYPE_END;
 488                      ++i) {
 489                         if (cur_item->items[i] == items->type) {
 490                                 token = &mlx5_flow_items[items->type];
 491                                 break;
 492                         }
 493                 }
 494                 if (!token)
 495                         goto exit_item_not_supported;
 496                 cur_item = token;
 497                 err = mlx5_flow_item_validate(items,
 498                                               (const uint8_t *)cur_item->mask,
 499                                               cur_item->mask_sz);
 500                 if (err)
 501                         goto exit_item_not_supported;
 502                 if (flow->ibv_attr && cur_item->convert) {
 503                         err = cur_item->convert(items,
 504                                                 (cur_item->default_mask ?
 505                                                  cur_item->default_mask :
 506                                                  cur_item->mask),
 507                                                 flow);
 508                         if (err)
 509                                 goto exit_item_not_supported;
 510                 } else if (items->type == RTE_FLOW_ITEM_TYPE_VXLAN) {
 511                         if (flow->inner) {
 512                                 rte_flow_error_set(error, ENOTSUP,
 513                                                    RTE_FLOW_ERROR_TYPE_ITEM,
 514                                                    items,
 515                                                    "cannot recognize multiple"
 516                                                    " VXLAN encapsulations");
 517                                 return -rte_errno;
 518                         }
 519                         flow->inner = 1;
 520                 }
 521                 flow->offset += cur_item->dst_sz;
 522         }
 523         for (; actions->type != RTE_FLOW_ACTION_TYPE_END; ++actions) {
 524                 if (actions->type == RTE_FLOW_ACTION_TYPE_VOID) {
 525                         continue;
 526                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_DROP) {
 527                         flow->actions.drop = 1;
 528                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_QUEUE) {
 529                         const struct rte_flow_action_queue *queue =
 530                                 (const struct rte_flow_action_queue *)
 531                                 actions->conf;
 532                         uint16_t n;
 533                         uint16_t found = 0;
 534
 535                         if (!queue || (queue->index > (priv->rxqs_n - 1)))
 536                                 goto exit_action_not_supported;
 537                         for (n = 0; n < flow->actions.queues_n; ++n) {
 538                                 if (flow->actions.queues[n] == queue->index) {
 539                                         found = 1;
 540                                         break;
 541                                 }
 542                         }
 543                         if (flow->actions.queues_n > 1 && !found) {
 544                                 rte_flow_error_set(error, ENOTSUP,
 545                                            RTE_FLOW_ERROR_TYPE_ACTION,
 546                                            actions,
 547                                            "queue action not in RSS queues");
 548                                 return -rte_errno;
 549                         }
 550                         if (!found) {
 551                                 flow->actions.queue = 1;
 552                                 flow->actions.queues_n = 1;
 553                                 flow->actions.queues[0] = queue->index;
 554                         }
 555                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_RSS) {
 556                         const struct rte_flow_action_rss *rss =
 557                                 (const struct rte_flow_action_rss *)
 558                                 actions->conf;
 559                         uint16_t n;
 560
 561                         if (!rss || !rss->num) {
 562                                 rte_flow_error_set(error, EINVAL,
 563                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 564                                                    actions,
 565                                                    "no valid queues");
 566                                 return -rte_errno;
 567                         }
 568                         if (flow->actions.queues_n == 1) {
 569                                 uint16_t found = 0;
 570
 571                                 assert(flow->actions.queues_n);
 572                                 for (n = 0; n < rss->num; ++n) {
 573                                         if (flow->actions.queues[0] ==
 574                                             rss->queue[n]) {
 575                                                 found = 1;
 576                                                 break;
 577                                         }
 578                                 }
 579                                 if (!found) {
 580                                         rte_flow_error_set(error, ENOTSUP,
 581                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 582                                                    actions,
 583                                                    "queue action not in RSS"
 584                                                    " queues");
 585                                         return -rte_errno;
 586                                 }
 587                         }
 588                         for (n = 0; n < rss->num; ++n) {
 589                                 if (rss->queue[n] >= priv->rxqs_n) {
 590                                         rte_flow_error_set(error, EINVAL,
 591                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 592                                                    actions,
 593                                                    "queue id > number of"
 594                                                    " queues");
 595                                         return -rte_errno;
 596                                 }
 597                         }
 598                         flow->actions.queue = 1;
 599                         for (n = 0; n < rss->num; ++n)
 600                                 flow->actions.queues[n] = rss->queue[n];
 601                         flow->actions.queues_n = rss->num;
 602                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_MARK) {
 603                         const struct rte_flow_action_mark *mark =
 604                                 (const struct rte_flow_action_mark *)
 605                                 actions->conf;
 606
 607                         if (!mark) {
 608                                 rte_flow_error_set(error, EINVAL,
 609                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 610                                                    actions,
 611                                                    "mark must be defined");
 612                                 return -rte_errno;
 613                         } else if (mark->id >= MLX5_FLOW_MARK_MAX) {
 614                                 rte_flow_error_set(error, ENOTSUP,
 615                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 616                                                    actions,
 617                                                    "mark must be between 0"
 618                                                    " and 16777199");
 619                                 return -rte_errno;
 620                         }
 621                         flow->actions.mark = 1;
 622                         flow->actions.mark_id = mark->id;
 623                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_FLAG) {
 624                         flow->actions.mark = 1;
 625                 } else {
 626                         goto exit_action_not_supported;
 627                 }
 628         }
 629         if (flow->actions.mark && !flow->ibv_attr && !flow->actions.drop)
 630                 flow->offset += sizeof(struct ibv_flow_spec_action_tag);
 631         if (!flow->ibv_attr && flow->actions.drop)
 632                 flow->offset += sizeof(struct ibv_flow_spec_action_drop);
 633         if (!flow->actions.queue && !flow->actions.drop) {
 634                 rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_HANDLE,
 635                                    NULL, "no valid action");
 636                 return -rte_errno;
 637         }
 638         return 0;
 639 exit_item_not_supported:
 640         rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 641                            items, "item not supported");
 642         return -rte_errno;
 643 exit_action_not_supported:
 644         rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
 645                            actions, "action not supported");
 646         return -rte_errno;
 647 }
 648
 649 /**
 650  * Validate a flow supported by the NIC.
 651  *
 652  * @see rte_flow_validate()
 653  * @see rte_flow_ops
 654  */
 655 int
 656 mlx5_flow_validate(struct rte_eth_dev *dev,
 657                    const struct rte_flow_attr *attr,
 658                    const struct rte_flow_item items[],
 659                    const struct rte_flow_action actions[],
 660                    struct rte_flow_error *error)
 661 {
 662         struct priv *priv = dev->data->dev_private;
 663         int ret;
 664         struct mlx5_flow_parse flow = {
 665                 .offset = sizeof(struct ibv_flow_attr),
 666                 .actions = {
 667                         .mark_id = MLX5_FLOW_MARK_DEFAULT,
 668                         .queues_n = 0,
 669                 },
 670         };
 671
 672         priv_lock(priv);
 673         ret = priv_flow_validate(priv, attr, items, actions, error, &flow);
 674         priv_unlock(priv);
 675         return ret;
 676 }
 677
 678 /**
 679  * Convert Ethernet item to Verbs specification.
 680  *
 681  * @param item[in]
 682  *   Item specification.
 683  * @param default_mask[in]
 684  *   Default bit-masks to use when item->mask is not provided.
 685  * @param data[in, out]
 686  *   User structure.
 687  */
 688 static int
 689 mlx5_flow_create_eth(const struct rte_flow_item *item,
 690                      const void *default_mask,
 691                      void *data)
 692 {
 693         const struct rte_flow_item_eth *spec = item->spec;
 694         const struct rte_flow_item_eth *mask = item->mask;
 695         struct mlx5_flow_parse *flow = (struct mlx5_flow_parse *)data;
 696         struct ibv_flow_spec_eth *eth;
 697         const unsigned int eth_size = sizeof(struct ibv_flow_spec_eth);
 698         unsigned int i;
 699
 700         ++flow->ibv_attr->num_of_specs;
 701         flow->ibv_attr->priority = 2;
 702         flow->hash_fields = 0;
 703         eth = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 704         *eth = (struct ibv_flow_spec_eth) {
 705                 .type = flow->inner | IBV_FLOW_SPEC_ETH,
 706                 .size = eth_size,
 707         };
 708         if (!spec)
 709                 return 0;
 710         if (!mask)
 711                 mask = default_mask;
 712         memcpy(eth->val.dst_mac, spec->dst.addr_bytes, ETHER_ADDR_LEN);
 713         memcpy(eth->val.src_mac, spec->src.addr_bytes, ETHER_ADDR_LEN);
 714         eth->val.ether_type = spec->type;
 715         memcpy(eth->mask.dst_mac, mask->dst.addr_bytes, ETHER_ADDR_LEN);
 716         memcpy(eth->mask.src_mac, mask->src.addr_bytes, ETHER_ADDR_LEN);
 717         eth->mask.ether_type = mask->type;
 718         /* Remove unwanted bits from values. */
 719         for (i = 0; i < ETHER_ADDR_LEN; ++i) {
 720                 eth->val.dst_mac[i] &= eth->mask.dst_mac[i];
 721                 eth->val.src_mac[i] &= eth->mask.src_mac[i];
 722         }
 723         eth->val.ether_type &= eth->mask.ether_type;
 724         return 0;
 725 }
 726
 727 /**
 728  * Convert VLAN item to Verbs specification.
 729  *
 730  * @param item[in]
 731  *   Item specification.
 732  * @param default_mask[in]
 733  *   Default bit-masks to use when item->mask is not provided.
 734  * @param data[in, out]
 735  *   User structure.
 736  */
 737 static int
 738 mlx5_flow_create_vlan(const struct rte_flow_item *item,
 739                       const void *default_mask,
 740                       void *data)
 741 {
 742         const struct rte_flow_item_vlan *spec = item->spec;
 743         const struct rte_flow_item_vlan *mask = item->mask;
 744         struct mlx5_flow_parse *flow = (struct mlx5_flow_parse *)data;
 745         struct ibv_flow_spec_eth *eth;
 746         const unsigned int eth_size = sizeof(struct ibv_flow_spec_eth);
 747
 748         eth = (void *)((uintptr_t)flow->ibv_attr + flow->offset - eth_size);
 749         if (!spec)
 750                 return 0;
 751         if (!mask)
 752                 mask = default_mask;
 753         eth->val.vlan_tag = spec->tci;
 754         eth->mask.vlan_tag = mask->tci;
 755         eth->val.vlan_tag &= eth->mask.vlan_tag;
 756         return 0;
 757 }
 758
 759 /**
 760  * Convert IPv4 item to Verbs specification.
 761  *
 762  * @param item[in]
 763  *   Item specification.
 764  * @param default_mask[in]
 765  *   Default bit-masks to use when item->mask is not provided.
 766  * @param data[in, out]
 767  *   User structure.
 768  */
 769 static int
 770 mlx5_flow_create_ipv4(const struct rte_flow_item *item,
 771                       const void *default_mask,
 772                       void *data)
 773 {
 774         const struct rte_flow_item_ipv4 *spec = item->spec;
 775         const struct rte_flow_item_ipv4 *mask = item->mask;
 776         struct mlx5_flow_parse *flow = (struct mlx5_flow_parse *)data;
 777         struct ibv_flow_spec_ipv4_ext *ipv4;
 778         unsigned int ipv4_size = sizeof(struct ibv_flow_spec_ipv4_ext);
 779
 780         ++flow->ibv_attr->num_of_specs;
 781         flow->ibv_attr->priority = 1;
 782         flow->hash_fields = (IBV_RX_HASH_SRC_IPV4 |
 783                              IBV_RX_HASH_DST_IPV4);
 784         ipv4 = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 785         *ipv4 = (struct ibv_flow_spec_ipv4_ext) {
 786                 .type = flow->inner | IBV_FLOW_SPEC_IPV4_EXT,
 787                 .size = ipv4_size,
 788         };
 789         if (!spec)
 790                 return 0;
 791         if (!mask)
 792                 mask = default_mask;
 793         ipv4->val = (struct ibv_flow_ipv4_ext_filter){
 794                 .src_ip = spec->hdr.src_addr,
 795                 .dst_ip = spec->hdr.dst_addr,
 796                 .proto = spec->hdr.next_proto_id,
 797                 .tos = spec->hdr.type_of_service,
 798         };
 799         ipv4->mask = (struct ibv_flow_ipv4_ext_filter){
 800                 .src_ip = mask->hdr.src_addr,
 801                 .dst_ip = mask->hdr.dst_addr,
 802                 .proto = mask->hdr.next_proto_id,
 803                 .tos = mask->hdr.type_of_service,
 804         };
 805         /* Remove unwanted bits from values. */
 806         ipv4->val.src_ip &= ipv4->mask.src_ip;
 807         ipv4->val.dst_ip &= ipv4->mask.dst_ip;
 808         ipv4->val.proto &= ipv4->mask.proto;
 809         ipv4->val.tos &= ipv4->mask.tos;
 810         return 0;
 811 }
 812
 813 /**
 814  * Convert IPv6 item to Verbs specification.
 815  *
 816  * @param item[in]
 817  *   Item specification.
 818  * @param default_mask[in]
 819  *   Default bit-masks to use when item->mask is not provided.
 820  * @param data[in, out]
 821  *   User structure.
 822  */
 823 static int
 824 mlx5_flow_create_ipv6(const struct rte_flow_item *item,
 825                       const void *default_mask,
 826                       void *data)
 827 {
 828         const struct rte_flow_item_ipv6 *spec = item->spec;
 829         const struct rte_flow_item_ipv6 *mask = item->mask;
 830         struct mlx5_flow_parse *flow = (struct mlx5_flow_parse *)data;
 831         struct ibv_flow_spec_ipv6 *ipv6;
 832         unsigned int ipv6_size = sizeof(struct ibv_flow_spec_ipv6);
 833         unsigned int i;
 834
 835         ++flow->ibv_attr->num_of_specs;
 836         flow->ibv_attr->priority = 1;
 837         flow->hash_fields = (IBV_RX_HASH_SRC_IPV6 |
 838                              IBV_RX_HASH_DST_IPV6);
 839         ipv6 = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 840         *ipv6 = (struct ibv_flow_spec_ipv6) {
 841                 .type = flow->inner | IBV_FLOW_SPEC_IPV6,
 842                 .size = ipv6_size,
 843         };
 844         if (!spec)
 845                 return 0;
 846         if (!mask)
 847                 mask = default_mask;
 848         memcpy(ipv6->val.src_ip, spec->hdr.src_addr,
 849                RTE_DIM(ipv6->val.src_ip));
 850         memcpy(ipv6->val.dst_ip, spec->hdr.dst_addr,
 851                RTE_DIM(ipv6->val.dst_ip));
 852         memcpy(ipv6->mask.src_ip, mask->hdr.src_addr,
 853                RTE_DIM(ipv6->mask.src_ip));
 854         memcpy(ipv6->mask.dst_ip, mask->hdr.dst_addr,
 855                RTE_DIM(ipv6->mask.dst_ip));
 856         ipv6->mask.flow_label = mask->hdr.vtc_flow;
 857         ipv6->mask.next_hdr = mask->hdr.proto;
 858         ipv6->mask.hop_limit = mask->hdr.hop_limits;
 859         /* Remove unwanted bits from values. */
 860         for (i = 0; i < RTE_DIM(ipv6->val.src_ip); ++i) {
 861                 ipv6->val.src_ip[i] &= ipv6->mask.src_ip[i];
 862                 ipv6->val.dst_ip[i] &= ipv6->mask.dst_ip[i];
 863         }
 864         ipv6->val.flow_label &= ipv6->mask.flow_label;
 865         ipv6->val.next_hdr &= ipv6->mask.next_hdr;
 866         ipv6->val.hop_limit &= ipv6->mask.hop_limit;
 867         return 0;
 868 }
 869
 870 /**
 871  * Convert UDP item to Verbs specification.
 872  *
 873  * @param item[in]
 874  *   Item specification.
 875  * @param default_mask[in]
 876  *   Default bit-masks to use when item->mask is not provided.
 877  * @param data[in, out]
 878  *   User structure.
 879  */
 880 static int
 881 mlx5_flow_create_udp(const struct rte_flow_item *item,
 882                      const void *default_mask,
 883                      void *data)
 884 {
 885         const struct rte_flow_item_udp *spec = item->spec;
 886         const struct rte_flow_item_udp *mask = item->mask;
 887         struct mlx5_flow_parse *flow = (struct mlx5_flow_parse *)data;
 888         struct ibv_flow_spec_tcp_udp *udp;
 889         unsigned int udp_size = sizeof(struct ibv_flow_spec_tcp_udp);
 890
 891         ++flow->ibv_attr->num_of_specs;
 892         flow->ibv_attr->priority = 0;
 893         flow->hash_fields |= (IBV_RX_HASH_SRC_PORT_UDP |
 894                               IBV_RX_HASH_DST_PORT_UDP);
 895         udp = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 896         *udp = (struct ibv_flow_spec_tcp_udp) {
 897                 .type = flow->inner | IBV_FLOW_SPEC_UDP,
 898                 .size = udp_size,
 899         };
 900         if (!spec)
 901                 return 0;
 902         if (!mask)
 903                 mask = default_mask;
 904         udp->val.dst_port = spec->hdr.dst_port;
 905         udp->val.src_port = spec->hdr.src_port;
 906         udp->mask.dst_port = mask->hdr.dst_port;
 907         udp->mask.src_port = mask->hdr.src_port;
 908         /* Remove unwanted bits from values. */
 909         udp->val.src_port &= udp->mask.src_port;
 910         udp->val.dst_port &= udp->mask.dst_port;
 911         return 0;
 912 }
 913
 914 /**
 915  * Convert TCP item to Verbs specification.
 916  *
 917  * @param item[in]
 918  *   Item specification.
 919  * @param default_mask[in]
 920  *   Default bit-masks to use when item->mask is not provided.
 921  * @param data[in, out]
 922  *   User structure.
 923  */
 924 static int
 925 mlx5_flow_create_tcp(const struct rte_flow_item *item,
 926                      const void *default_mask,
 927                      void *data)
 928 {
 929         const struct rte_flow_item_tcp *spec = item->spec;
 930         const struct rte_flow_item_tcp *mask = item->mask;
 931         struct mlx5_flow_parse *flow = (struct mlx5_flow_parse *)data;
 932         struct ibv_flow_spec_tcp_udp *tcp;
 933         unsigned int tcp_size = sizeof(struct ibv_flow_spec_tcp_udp);
 934
 935         ++flow->ibv_attr->num_of_specs;
 936         flow->ibv_attr->priority = 0;
 937         flow->hash_fields |= (IBV_RX_HASH_SRC_PORT_TCP |
 938                               IBV_RX_HASH_DST_PORT_TCP);
 939         tcp = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 940         *tcp = (struct ibv_flow_spec_tcp_udp) {
 941                 .type = flow->inner | IBV_FLOW_SPEC_TCP,
 942                 .size = tcp_size,
 943         };
 944         if (!spec)
 945                 return 0;
 946         if (!mask)
 947                 mask = default_mask;
 948         tcp->val.dst_port = spec->hdr.dst_port;
 949         tcp->val.src_port = spec->hdr.src_port;
 950         tcp->mask.dst_port = mask->hdr.dst_port;
 951         tcp->mask.src_port = mask->hdr.src_port;
 952         /* Remove unwanted bits from values. */
 953         tcp->val.src_port &= tcp->mask.src_port;
 954         tcp->val.dst_port &= tcp->mask.dst_port;
 955         return 0;
 956 }
 957
 958 /**
 959  * Convert VXLAN item to Verbs specification.
 960  *
 961  * @param item[in]
 962  *   Item specification.
 963  * @param default_mask[in]
 964  *   Default bit-masks to use when item->mask is not provided.
 965  * @param data[in, out]
 966  *   User structure.
 967  */
 968 static int
 969 mlx5_flow_create_vxlan(const struct rte_flow_item *item,
 970                        const void *default_mask,
 971                        void *data)
 972 {
 973         const struct rte_flow_item_vxlan *spec = item->spec;
 974         const struct rte_flow_item_vxlan *mask = item->mask;
 975         struct mlx5_flow_parse *flow = (struct mlx5_flow_parse *)data;
 976         struct ibv_flow_spec_tunnel *vxlan;
 977         unsigned int size = sizeof(struct ibv_flow_spec_tunnel);
 978         union vni {
 979                 uint32_t vlan_id;
 980                 uint8_t vni[4];
 981         } id;
 982
 983         ++flow->ibv_attr->num_of_specs;
 984         flow->ibv_attr->priority = 0;
 985         id.vni[0] = 0;
 986         vxlan = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 987         *vxlan = (struct ibv_flow_spec_tunnel) {
 988                 .type = flow->inner | IBV_FLOW_SPEC_VXLAN_TUNNEL,
 989                 .size = size,
 990         };
 991         flow->inner = IBV_FLOW_SPEC_INNER;
 992         if (!spec)
 993                 return 0;
 994         if (!mask)
 995                 mask = default_mask;
 996         memcpy(&id.vni[1], spec->vni, 3);
 997         vxlan->val.tunnel_id = id.vlan_id;
 998         memcpy(&id.vni[1], mask->vni, 3);
 999         vxlan->mask.tunnel_id = id.vlan_id;
1000         /* Remove unwanted bits from values. */
1001         vxlan->val.tunnel_id &= vxlan->mask.tunnel_id;
1002         return 0;
1003 }
1004
1005 /**
1006  * Convert mark/flag action to Verbs specification.
1007  *
1008  * @param flow
1009  *   Pointer to MLX5 flow structure.
1010  * @param mark_id
1011  *   Mark identifier.
1012  */
1013 static int
1014 mlx5_flow_create_flag_mark(struct mlx5_flow_parse *flow, uint32_t mark_id)
1015 {
1016         struct ibv_flow_spec_action_tag *tag;
1017         unsigned int size = sizeof(struct ibv_flow_spec_action_tag);
1018
1019         tag = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
1020         *tag = (struct ibv_flow_spec_action_tag){
1021                 .type = IBV_FLOW_SPEC_ACTION_TAG,
1022                 .size = size,
1023                 .tag_id = mlx5_flow_mark_set(mark_id),
1024         };
1025         ++flow->ibv_attr->num_of_specs;
1026         return 0;
1027 }
1028
1029 /**
1030  * Complete flow rule creation with a drop queue.
1031  *
1032  * @param priv
1033  *   Pointer to private structure.
1034  * @param flow
1035  *   MLX5 flow attributes (filled by mlx5_flow_validate()).
1036  * @param[out] error
1037  *   Perform verbose error reporting if not NULL.
1038  *
1039  * @return
1040  *   A flow if the rule could be created.
1041  */
1042 static struct rte_flow *
1043 priv_flow_create_action_queue_drop(struct priv *priv,
1044                                    struct mlx5_flow_parse *flow,
1045                                    struct rte_flow_error *error)
1046 {
1047         struct rte_flow *rte_flow;
1048         struct ibv_flow_spec_action_drop *drop;
1049         unsigned int size = sizeof(struct ibv_flow_spec_action_drop);
1050
1051         assert(priv->pd);
1052         assert(priv->ctx);
1053         rte_flow = rte_calloc(__func__, 1, sizeof(*rte_flow), 0);
1054         if (!rte_flow) {
1055                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1056                                    NULL, "cannot allocate flow memory");
1057                 return NULL;
1058         }
1059         rte_flow->drop = 1;
1060         drop = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
1061         *drop = (struct ibv_flow_spec_action_drop){
1062                         .type = IBV_FLOW_SPEC_ACTION_DROP,
1063                         .size = size,
1064         };
1065         ++flow->ibv_attr->num_of_specs;
1066         flow->offset += sizeof(struct ibv_flow_spec_action_drop);
1067         rte_flow->ibv_attr = flow->ibv_attr;
1068         if (!priv->dev->data->dev_started)
1069                 return rte_flow;
1070         rte_flow->drxq.hrxq.qp = priv->flow_drop_queue->qp;
1071         rte_flow->ibv_flow = ibv_create_flow(rte_flow->drxq.hrxq.qp,
1072                                              rte_flow->ibv_attr);
1073         if (!rte_flow->ibv_flow) {
1074                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1075                                    NULL, "flow rule creation failure");
1076                 goto error;
1077         }
1078         return rte_flow;
1079 error:
1080         assert(rte_flow);
1081         rte_free(rte_flow);
1082         return NULL;
1083 }
1084
1085 /**
1086  * Complete flow rule creation.
1087  *
1088  * @param priv
1089  *   Pointer to private structure.
1090  * @param flow
1091  *   MLX5 flow attributes (filled by mlx5_flow_validate()).
1092  * @param[out] error
1093  *   Perform verbose error reporting if not NULL.
1094  *
1095  * @return
1096  *   A flow if the rule could be created.
1097  */
1098 static struct rte_flow *
1099 priv_flow_create_action_queue(struct priv *priv,
1100                               struct mlx5_flow_parse *flow,
1101                               struct rte_flow_error *error)
1102 {
1103         struct rte_flow *rte_flow;
1104         unsigned int i;
1105
1106         assert(priv->pd);
1107         assert(priv->ctx);
1108         assert(!flow->actions.drop);
1109         rte_flow =
1110                 rte_calloc(__func__, 1,
1111                            sizeof(*flow) +
1112                            flow->actions.queues_n * sizeof(uint16_t),
1113                            0);
1114         if (!rte_flow) {
1115                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1116                                    NULL, "cannot allocate flow memory");
1117                 return NULL;
1118         }
1119         rte_flow->mark = flow->actions.mark;
1120         rte_flow->ibv_attr = flow->ibv_attr;
1121         rte_flow->queues = (uint16_t (*)[])(rte_flow + 1);
1122         memcpy(rte_flow->queues, flow->actions.queues,
1123                flow->actions.queues_n * sizeof(uint16_t));
1124         rte_flow->queues_n = flow->actions.queues_n;
1125         rte_flow->frxq.hash_fields = flow->hash_fields;
1126         rte_flow->frxq.hrxq = mlx5_priv_hrxq_get(priv, rss_hash_default_key,
1127                                                  rss_hash_default_key_len,
1128                                                  flow->hash_fields,
1129                                                  (*rte_flow->queues),
1130                                                  rte_flow->queues_n);
1131         if (rte_flow->frxq.hrxq) {
1132                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1133                                    NULL, "duplicated flow");
1134                 goto error;
1135         }
1136         rte_flow->frxq.hrxq = mlx5_priv_hrxq_new(priv, rss_hash_default_key,
1137                                                  rss_hash_default_key_len,
1138                                                  flow->hash_fields,
1139                                                  (*rte_flow->queues),
1140                                                  rte_flow->queues_n);
1141         if (!rte_flow->frxq.hrxq) {
1142                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1143                                    NULL, "cannot create hash rxq");
1144                 goto error;
1145         }
1146         for (i = 0; i != flow->actions.queues_n; ++i) {
1147                 struct mlx5_rxq_data *q =
1148                         (*priv->rxqs)[flow->actions.queues[i]];
1149
1150                 q->mark |= flow->actions.mark;
1151         }
1152         if (!priv->dev->data->dev_started)
1153                 return rte_flow;
1154         rte_flow->ibv_flow = ibv_create_flow(rte_flow->frxq.hrxq->qp,
1155                                              rte_flow->ibv_attr);
1156         if (!rte_flow->ibv_flow) {
1157                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1158                                    NULL, "flow rule creation failure");
1159                 goto error;
1160         }
1161         return rte_flow;
1162 error:
1163         assert(rte_flow);
1164         if (rte_flow->frxq.hrxq)
1165                 mlx5_priv_hrxq_release(priv, rte_flow->frxq.hrxq);
1166         rte_free(rte_flow);
1167         return NULL;
1168 }
1169
1170 /**
1171  * Convert a flow.
1172  *
1173  * @param priv
1174  *   Pointer to private structure.
1175  * @param list
1176  *   Pointer to a TAILQ flow list.
1177  * @param[in] attr
1178  *   Flow rule attributes.
1179  * @param[in] pattern
1180  *   Pattern specification (list terminated by the END pattern item).
1181  * @param[in] actions
1182  *   Associated actions (list terminated by the END action).
1183  * @param[out] error
1184  *   Perform verbose error reporting if not NULL.
1185  *
1186  * @return
1187  *   A flow on success, NULL otherwise.
1188  */
1189 static struct rte_flow *
1190 priv_flow_create(struct priv *priv,
1191                  struct mlx5_flows *list,
1192                  const struct rte_flow_attr *attr,
1193                  const struct rte_flow_item items[],
1194                  const struct rte_flow_action actions[],
1195                  struct rte_flow_error *error)
1196 {
1197         struct rte_flow *rte_flow;
1198         struct mlx5_flow_parse flow = {
1199                 .offset = sizeof(struct ibv_flow_attr),
1200                 .actions = {
1201                         .mark_id = MLX5_FLOW_MARK_DEFAULT,
1202                         .queues = { 0 },
1203                         .queues_n = 0,
1204                 },
1205         };
1206         int err;
1207
1208         err = priv_flow_validate(priv, attr, items, actions, error, &flow);
1209         if (err)
1210                 goto exit;
1211         flow.ibv_attr = rte_malloc(__func__, flow.offset, 0);
1212         flow.offset = sizeof(struct ibv_flow_attr);
1213         if (!flow.ibv_attr) {
1214                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1215                                    NULL, "cannot allocate ibv_attr memory");
1216                 goto exit;
1217         }
1218         *flow.ibv_attr = (struct ibv_flow_attr){
1219                 .type = IBV_FLOW_ATTR_NORMAL,
1220                 .size = sizeof(struct ibv_flow_attr),
1221                 .priority = attr->priority,
1222                 .num_of_specs = 0,
1223                 .port = 0,
1224                 .flags = 0,
1225         };
1226         flow.inner = 0;
1227         flow.hash_fields = 0;
1228         claim_zero(priv_flow_validate(priv, attr, items, actions,
1229                                       error, &flow));
1230         if (flow.actions.mark && !flow.actions.drop) {
1231                 mlx5_flow_create_flag_mark(&flow, flow.actions.mark_id);
1232                 flow.offset += sizeof(struct ibv_flow_spec_action_tag);
1233         }
1234         if (flow.actions.drop)
1235                 rte_flow =
1236                         priv_flow_create_action_queue_drop(priv, &flow, error);
1237         else
1238                 rte_flow = priv_flow_create_action_queue(priv, &flow, error);
1239         if (!rte_flow)
1240                 goto exit;
1241         if (rte_flow) {
1242                 TAILQ_INSERT_TAIL(list, rte_flow, next);
1243                 DEBUG("Flow created %p", (void *)rte_flow);
1244         }
1245         return rte_flow;
1246 exit:
1247         rte_free(flow.ibv_attr);
1248         return NULL;
1249 }
1250
1251 /**
1252  * Create a flow.
1253  *
1254  * @see rte_flow_create()
1255  * @see rte_flow_ops
1256  */
1257 struct rte_flow *
1258 mlx5_flow_create(struct rte_eth_dev *dev,
1259                  const struct rte_flow_attr *attr,
1260                  const struct rte_flow_item items[],
1261                  const struct rte_flow_action actions[],
1262                  struct rte_flow_error *error)
1263 {
1264         struct priv *priv = dev->data->dev_private;
1265         struct rte_flow *flow;
1266
1267         priv_lock(priv);
1268         flow = priv_flow_create(priv, &priv->flows, attr, items, actions,
1269                                 error);
1270         priv_unlock(priv);
1271         return flow;
1272 }
1273
1274 /**
1275  * Destroy a flow.
1276  *
1277  * @param priv
1278  *   Pointer to private structure.
1279  * @param list
1280  *   Pointer to a TAILQ flow list.
1281  * @param[in] flow
1282  *   Flow to destroy.
1283  */
1284 static void
1285 priv_flow_destroy(struct priv *priv,
1286                   struct mlx5_flows *list,
1287                   struct rte_flow *flow)
1288 {
1289         unsigned int i;
1290         uint16_t *queues;
1291         uint16_t queues_n;
1292
1293         if (flow->drop || !flow->mark)
1294                 goto free;
1295         queues = flow->frxq.hrxq->ind_table->queues;
1296         queues_n = flow->frxq.hrxq->ind_table->queues_n;
1297         for (i = 0; i != queues_n; ++i) {
1298                 struct rte_flow *tmp;
1299                 struct mlx5_rxq_data *rxq_data = (*priv->rxqs)[queues[i]];
1300                 int mark = 0;
1301
1302                 /*
1303                  * To remove the mark from the queue, the queue must not be
1304                  * present in any other marked flow (RSS or not).
1305                  */
1306                 TAILQ_FOREACH(tmp, list, next) {
1307                         unsigned int j;
1308
1309                         if (!tmp->mark)
1310                                 continue;
1311                         for (j = 0;
1312                              (j != tmp->frxq.hrxq->ind_table->queues_n) &&
1313                              !mark;
1314                              j++)
1315                                 if (tmp->frxq.hrxq->ind_table->queues[j] ==
1316                                     queues[i])
1317                                         mark = 1;
1318                 }
1319                 rxq_data->mark = mark;
1320         }
1321 free:
1322         if (flow->ibv_flow)
1323                 claim_zero(ibv_destroy_flow(flow->ibv_flow));
1324         if (!flow->drop)
1325                 mlx5_priv_hrxq_release(priv, flow->frxq.hrxq);
1326         TAILQ_REMOVE(list, flow, next);
1327         rte_free(flow->ibv_attr);
1328         DEBUG("Flow destroyed %p", (void *)flow);
1329         rte_free(flow);
1330 }
1331
1332 /**
1333  * Destroy a flow.
1334  *
1335  * @see rte_flow_destroy()
1336  * @see rte_flow_ops
1337  */
1338 int
1339 mlx5_flow_destroy(struct rte_eth_dev *dev,
1340                   struct rte_flow *flow,
1341                   struct rte_flow_error *error)
1342 {
1343         struct priv *priv = dev->data->dev_private;
1344
1345         (void)error;
1346         priv_lock(priv);
1347         priv_flow_destroy(priv, &priv->flows, flow);
1348         priv_unlock(priv);
1349         return 0;
1350 }
1351
1352 /**
1353  * Destroy all flows.
1354  *
1355  * @param priv
1356  *   Pointer to private structure.
1357  * @param list
1358  *   Pointer to a TAILQ flow list.
1359  */
1360 void
1361 priv_flow_flush(struct priv *priv, struct mlx5_flows *list)
1362 {
1363         while (!TAILQ_EMPTY(list)) {
1364                 struct rte_flow *flow;
1365
1366                 flow = TAILQ_FIRST(list);
1367                 priv_flow_destroy(priv, list, flow);
1368         }
1369 }
1370
1371 /**
1372  * Destroy all flows.
1373  *
1374  * @see rte_flow_flush()
1375  * @see rte_flow_ops
1376  */
1377 int
1378 mlx5_flow_flush(struct rte_eth_dev *dev,
1379                 struct rte_flow_error *error)
1380 {
1381         struct priv *priv = dev->data->dev_private;
1382
1383         (void)error;
1384         priv_lock(priv);
1385         priv_flow_flush(priv, &priv->flows);
1386         priv_unlock(priv);
1387         return 0;
1388 }
1389
1390 /**
1391  * Create drop queue.
1392  *
1393  * @param priv
1394  *   Pointer to private structure.
1395  *
1396  * @return
1397  *   0 on success.
1398  */
1399 static int
1400 priv_flow_create_drop_queue(struct priv *priv)
1401 {
1402         struct mlx5_hrxq_drop *fdq = NULL;
1403
1404         assert(priv->pd);
1405         assert(priv->ctx);
1406         fdq = rte_calloc(__func__, 1, sizeof(*fdq), 0);
1407         if (!fdq) {
1408                 WARN("cannot allocate memory for drop queue");
1409                 goto error;
1410         }
1411         fdq->cq = ibv_create_cq(priv->ctx, 1, NULL, NULL, 0);
1412         if (!fdq->cq) {
1413                 WARN("cannot allocate CQ for drop queue");
1414                 goto error;
1415         }
1416         fdq->wq = ibv_create_wq(priv->ctx,
1417                         &(struct ibv_wq_init_attr){
1418                         .wq_type = IBV_WQT_RQ,
1419                         .max_wr = 1,
1420                         .max_sge = 1,
1421                         .pd = priv->pd,
1422                         .cq = fdq->cq,
1423                         });
1424         if (!fdq->wq) {
1425                 WARN("cannot allocate WQ for drop queue");
1426                 goto error;
1427         }
1428         fdq->ind_table = ibv_create_rwq_ind_table(priv->ctx,
1429                         &(struct ibv_rwq_ind_table_init_attr){
1430                         .log_ind_tbl_size = 0,
1431                         .ind_tbl = &fdq->wq,
1432                         .comp_mask = 0,
1433                         });
1434         if (!fdq->ind_table) {
1435                 WARN("cannot allocate indirection table for drop queue");
1436                 goto error;
1437         }
1438         fdq->qp = ibv_create_qp_ex(priv->ctx,
1439                 &(struct ibv_qp_init_attr_ex){
1440                         .qp_type = IBV_QPT_RAW_PACKET,
1441                         .comp_mask =
1442                                 IBV_QP_INIT_ATTR_PD |
1443                                 IBV_QP_INIT_ATTR_IND_TABLE |
1444                                 IBV_QP_INIT_ATTR_RX_HASH,
1445                         .rx_hash_conf = (struct ibv_rx_hash_conf){
1446                                 .rx_hash_function =
1447                                         IBV_RX_HASH_FUNC_TOEPLITZ,
1448                                 .rx_hash_key_len = rss_hash_default_key_len,
1449                                 .rx_hash_key = rss_hash_default_key,
1450                                 .rx_hash_fields_mask = 0,
1451                                 },
1452                         .rwq_ind_tbl = fdq->ind_table,
1453                         .pd = priv->pd
1454                 });
1455         if (!fdq->qp) {
1456                 WARN("cannot allocate QP for drop queue");
1457                 goto error;
1458         }
1459         priv->flow_drop_queue = fdq;
1460         return 0;
1461 error:
1462         if (fdq->qp)
1463                 claim_zero(ibv_destroy_qp(fdq->qp));
1464         if (fdq->ind_table)
1465                 claim_zero(ibv_destroy_rwq_ind_table(fdq->ind_table));
1466         if (fdq->wq)
1467                 claim_zero(ibv_destroy_wq(fdq->wq));
1468         if (fdq->cq)
1469                 claim_zero(ibv_destroy_cq(fdq->cq));
1470         if (fdq)
1471                 rte_free(fdq);
1472         priv->flow_drop_queue = NULL;
1473         return -1;
1474 }
1475
1476 /**
1477  * Delete drop queue.
1478  *
1479  * @param priv
1480  *   Pointer to private structure.
1481  */
1482 static void
1483 priv_flow_delete_drop_queue(struct priv *priv)
1484 {
1485         struct mlx5_hrxq_drop *fdq = priv->flow_drop_queue;
1486
1487         if (!fdq)
1488                 return;
1489         if (fdq->qp)
1490                 claim_zero(ibv_destroy_qp(fdq->qp));
1491         if (fdq->ind_table)
1492                 claim_zero(ibv_destroy_rwq_ind_table(fdq->ind_table));
1493         if (fdq->wq)
1494                 claim_zero(ibv_destroy_wq(fdq->wq));
1495         if (fdq->cq)
1496                 claim_zero(ibv_destroy_cq(fdq->cq));
1497         rte_free(fdq);
1498         priv->flow_drop_queue = NULL;
1499 }
1500
1501 /**
1502  * Remove all flows.
1503  *
1504  * Called by dev_stop() to remove all flows.
1505  *
1506  * @param priv
1507  *   Pointer to private structure.
1508  * @param list
1509  *   Pointer to a TAILQ flow list.
1510  */
1511 void
1512 priv_flow_stop(struct priv *priv, struct mlx5_flows *list)
1513 {
1514         struct rte_flow *flow;
1515
1516         TAILQ_FOREACH_REVERSE(flow, list, mlx5_flows, next) {
1517                 claim_zero(ibv_destroy_flow(flow->ibv_flow));
1518                 flow->ibv_flow = NULL;
1519                 mlx5_priv_hrxq_release(priv, flow->frxq.hrxq);
1520                 flow->frxq.hrxq = NULL;
1521                 if (flow->mark) {
1522                         unsigned int n;
1523                         struct mlx5_ind_table_ibv *ind_tbl =
1524                                 flow->frxq.hrxq->ind_table;
1525
1526                         for (n = 0; n < ind_tbl->queues_n; ++n)
1527                                 (*priv->rxqs)[ind_tbl->queues[n]]->mark = 0;
1528                 }
1529                 DEBUG("Flow %p removed", (void *)flow);
1530         }
1531         priv_flow_delete_drop_queue(priv);
1532 }
1533
1534 /**
1535  * Add all flows.
1536  *
1537  * @param priv
1538  *   Pointer to private structure.
1539  * @param list
1540  *   Pointer to a TAILQ flow list.
1541  *
1542  * @return
1543  *   0 on success, a errno value otherwise and rte_errno is set.
1544  */
1545 int
1546 priv_flow_start(struct priv *priv, struct mlx5_flows *list)
1547 {
1548         int ret;
1549         struct rte_flow *flow;
1550
1551         ret = priv_flow_create_drop_queue(priv);
1552         if (ret)
1553                 return -1;
1554         TAILQ_FOREACH(flow, list, next) {
1555                 if (flow->frxq.hrxq)
1556                         goto flow_create;
1557                 flow->frxq.hrxq =
1558                         mlx5_priv_hrxq_get(priv, rss_hash_default_key,
1559                                            rss_hash_default_key_len,
1560                                            flow->frxq.hash_fields,
1561                                            (*flow->queues),
1562                                            flow->queues_n);
1563                 if (flow->frxq.hrxq)
1564                         goto flow_create;
1565                 flow->frxq.hrxq =
1566                         mlx5_priv_hrxq_new(priv, rss_hash_default_key,
1567                                            rss_hash_default_key_len,
1568                                            flow->frxq.hash_fields,
1569                                            (*flow->queues),
1570                                            flow->queues_n);
1571                 if (!flow->frxq.hrxq) {
1572                         DEBUG("Flow %p cannot be applied",
1573                               (void *)flow);
1574                         rte_errno = EINVAL;
1575                         return rte_errno;
1576                 }
1577 flow_create:
1578                 flow->ibv_flow = ibv_create_flow(flow->frxq.hrxq->qp,
1579                                                  flow->ibv_attr);
1580                 if (!flow->ibv_flow) {
1581                         DEBUG("Flow %p cannot be applied", (void *)flow);
1582                         rte_errno = EINVAL;
1583                         return rte_errno;
1584                 }
1585                 DEBUG("Flow %p applied", (void *)flow);
1586                 if (flow->mark) {
1587                         unsigned int n;
1588
1589                         for (n = 0;
1590                              n < flow->frxq.hrxq->ind_table->queues_n;
1591                              ++n) {
1592                                 uint16_t idx =
1593                                         flow->frxq.hrxq->ind_table->queues[n];
1594                                 (*priv->rxqs)[idx]->mark = 1;
1595                         }
1596                 }
1597         }
1598         return 0;
1599 }
1600
1601 /**
1602  * Isolated mode.
1603  *
1604  * @see rte_flow_isolate()
1605  * @see rte_flow_ops
1606  */
1607 int
1608 mlx5_flow_isolate(struct rte_eth_dev *dev,
1609                   int enable,
1610                   struct rte_flow_error *error)
1611 {
1612         struct priv *priv = dev->data->dev_private;
1613
1614         priv_lock(priv);
1615         if (dev->data->dev_started) {
1616                 rte_flow_error_set(error, EBUSY,
1617                                    RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
1618                                    NULL,
1619                                    "port must be stopped first");
1620                 priv_unlock(priv);
1621                 return -rte_errno;
1622         }
1623         priv->isolated = !!enable;
1624         priv_unlock(priv);
1625         return 0;
1626 }
1627
1628 /**
1629  * Verify the flow list is empty
1630  *
1631  * @param priv
1632  *  Pointer to private structure.
1633  *
1634  * @return the number of flows not released.
1635  */
1636 int
1637 priv_flow_verify(struct priv *priv)
1638 {
1639         struct rte_flow *flow;
1640         int ret = 0;
1641
1642         TAILQ_FOREACH(flow, &priv->flows, next) {
1643                 DEBUG("%p: flow %p still referenced", (void *)priv,
1644                       (void *)flow);
1645                 ++ret;
1646         }
1647         return ret;
1648 }
1649
1650 /**
1651  * Enable/disable a control flow configured from the control plane.
1652  *
1653  * @param dev
1654  *   Pointer to Ethernet device.
1655  * @param spec
1656  *   An Ethernet flow spec to apply.
1657  * @param mask
1658  *   An Ethernet flow mask to apply.
1659  * @param enable
1660  *   Enable/disable the flow.
1661  *
1662  * @return
1663  *   0 on success.
1664  */
1665 int
1666 mlx5_ctrl_flow(struct rte_eth_dev *dev,
1667                struct rte_flow_item_eth *spec,
1668                struct rte_flow_item_eth *mask,
1669                unsigned int enable)
1670 {
1671         struct priv *priv = dev->data->dev_private;
1672         const struct rte_flow_attr attr = {
1673                 .ingress = 1,
1674                 .priority = MLX5_CTRL_FLOW_PRIORITY,
1675         };
1676         struct rte_flow_item items[] = {
1677                 {
1678                         .type = RTE_FLOW_ITEM_TYPE_ETH,
1679                         .spec = spec,
1680                         .last = NULL,
1681                         .mask = mask,
1682                 },
1683                 {
1684                         .type = RTE_FLOW_ITEM_TYPE_END,
1685                 },
1686         };
1687         struct rte_flow_action actions[] = {
1688                 {
1689                         .type = RTE_FLOW_ACTION_TYPE_QUEUE,
1690                         .conf = &(struct rte_flow_action_queue){
1691                                 .index = 0,
1692                         },
1693                 },
1694                 {
1695                         .type = RTE_FLOW_ACTION_TYPE_END,
1696                 },
1697         };
1698         struct rte_flow *flow;
1699         struct rte_flow_error error;
1700
1701         if (enable) {
1702                 flow = priv_flow_create(priv, &priv->ctrl_flows, &attr, items,
1703                                         actions, &error);
1704                 if (!flow)
1705                         return 1;
1706         } else {
1707                 struct spec {
1708                         struct ibv_flow_attr ibv_attr;
1709                         struct ibv_flow_spec_eth eth;
1710                 } spec;
1711                 struct mlx5_flow_parse parser = {
1712                         .ibv_attr = &spec.ibv_attr,
1713                         .offset = sizeof(struct ibv_flow_attr),
1714                 };
1715                 struct ibv_flow_spec_eth *eth;
1716                 const unsigned int attr_size = sizeof(struct ibv_flow_attr);
1717
1718                 claim_zero(mlx5_flow_create_eth(&items[0], NULL, &parser));
1719                 TAILQ_FOREACH(flow, &priv->ctrl_flows, next) {
1720                         eth = (void *)((uintptr_t)flow->ibv_attr + attr_size);
1721                         assert(eth->type == IBV_FLOW_SPEC_ETH);
1722                         if (!memcmp(eth, &spec.eth, sizeof(*eth)))
1723                                 break;
1724                 }
1725                 if (flow) {
1726                         claim_zero(ibv_destroy_flow(flow->ibv_flow));
1727                         mlx5_priv_hrxq_release(priv, flow->frxq.hrxq);
1728                         rte_free(flow->ibv_attr);
1729                         DEBUG("Control flow destroyed %p", (void *)flow);
1730                         TAILQ_REMOVE(&priv->ctrl_flows, flow, next);
1731                         rte_free(flow);
1732                 }
1733         }
1734         return 0;
1735 }