drivers/net/mlx5/mlx5_flow.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright 2016 6WIND S.A.
   5  *   Copyright 2016 Mellanox.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of 6WIND S.A. nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  */
  33
  34 #include <sys/queue.h>
  35 #include <string.h>
  36
  37 /* Verbs header. */
  38 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
  39 #ifdef PEDANTIC
  40 #pragma GCC diagnostic ignored "-Wpedantic"
  41 #endif
  42 #include <infiniband/verbs.h>
  43 #ifdef PEDANTIC
  44 #pragma GCC diagnostic error "-Wpedantic"
  45 #endif
  46
  47 #include <rte_ethdev.h>
  48 #include <rte_flow.h>
  49 #include <rte_flow_driver.h>
  50 #include <rte_malloc.h>
  51
  52 #include "mlx5.h"
  53 #include "mlx5_prm.h"
  54
  55 /* Define minimal priority for control plane flows. */
  56 #define MLX5_CTRL_FLOW_PRIORITY 4
  57
  58 static int
  59 mlx5_flow_create_eth(const struct rte_flow_item *item,
  60                      const void *default_mask,
  61                      void *data);
  62
  63 static int
  64 mlx5_flow_create_vlan(const struct rte_flow_item *item,
  65                       const void *default_mask,
  66                       void *data);
  67
  68 static int
  69 mlx5_flow_create_ipv4(const struct rte_flow_item *item,
  70                       const void *default_mask,
  71                       void *data);
  72
  73 static int
  74 mlx5_flow_create_ipv6(const struct rte_flow_item *item,
  75                       const void *default_mask,
  76                       void *data);
  77
  78 static int
  79 mlx5_flow_create_udp(const struct rte_flow_item *item,
  80                      const void *default_mask,
  81                      void *data);
  82
  83 static int
  84 mlx5_flow_create_tcp(const struct rte_flow_item *item,
  85                      const void *default_mask,
  86                      void *data);
  87
  88 static int
  89 mlx5_flow_create_vxlan(const struct rte_flow_item *item,
  90                        const void *default_mask,
  91                        void *data);
  92
  93 /** Structure for Drop queue. */
  94 struct mlx5_hrxq_drop {
  95         struct ibv_rwq_ind_table *ind_table; /**< Indirection table. */
  96         struct ibv_qp *qp; /**< Verbs queue pair. */
  97         struct ibv_wq *wq; /**< Verbs work queue. */
  98         struct ibv_cq *cq; /**< Verbs completion queue. */
  99 };
 100
 101 /* Flows structures. */
 102 struct mlx5_flow {
 103         uint64_t hash_fields; /**< Fields that participate in the hash. */
 104         struct mlx5_hrxq *hrxq; /**< Hash Rx queues. */
 105 };
 106
 107 /* Drop flows structures. */
 108 struct mlx5_flow_drop {
 109         struct mlx5_hrxq_drop hrxq; /**< Drop hash Rx queue. */
 110 };
 111
 112 struct rte_flow {
 113         TAILQ_ENTRY(rte_flow) next; /**< Pointer to the next flow structure. */
 114         uint32_t mark:1; /**< Set if the flow is marked. */
 115         uint32_t drop:1; /**< Drop queue. */
 116         struct ibv_flow_attr *ibv_attr; /**< Pointer to Verbs attributes. */
 117         struct ibv_flow *ibv_flow; /**< Verbs flow. */
 118         uint16_t queues_n; /**< Number of entries in queue[]. */
 119         uint16_t (*queues)[]; /**< Queues indexes to use. */
 120         union {
 121                 struct mlx5_flow frxq; /**< Flow with Rx queue. */
 122                 struct mlx5_flow_drop drxq; /**< Flow with drop Rx queue. */
 123         };
 124 };
 125
 126 /** Static initializer for items. */
 127 #define ITEMS(...) \
 128         (const enum rte_flow_item_type []){ \
 129                 __VA_ARGS__, RTE_FLOW_ITEM_TYPE_END, \
 130         }
 131
 132 /** Structure to generate a simple graph of layers supported by the NIC. */
 133 struct mlx5_flow_items {
 134         /** List of possible actions for these items. */
 135         const enum rte_flow_action_type *const actions;
 136         /** Bit-masks corresponding to the possibilities for the item. */
 137         const void *mask;
 138         /**
 139          * Default bit-masks to use when item->mask is not provided. When
 140          * \default_mask is also NULL, the full supported bit-mask (\mask) is
 141          * used instead.
 142          */
 143         const void *default_mask;
 144         /** Bit-masks size in bytes. */
 145         const unsigned int mask_sz;
 146         /**
 147          * Conversion function from rte_flow to NIC specific flow.
 148          *
 149          * @param item
 150          *   rte_flow item to convert.
 151          * @param default_mask
 152          *   Default bit-masks to use when item->mask is not provided.
 153          * @param data
 154          *   Internal structure to store the conversion.
 155          *
 156          * @return
 157          *   0 on success, negative value otherwise.
 158          */
 159         int (*convert)(const struct rte_flow_item *item,
 160                        const void *default_mask,
 161                        void *data);
 162         /** Size in bytes of the destination structure. */
 163         const unsigned int dst_sz;
 164         /** List of possible following items.  */
 165         const enum rte_flow_item_type *const items;
 166 };
 167
 168 /** Valid action for this PMD. */
 169 static const enum rte_flow_action_type valid_actions[] = {
 170         RTE_FLOW_ACTION_TYPE_DROP,
 171         RTE_FLOW_ACTION_TYPE_QUEUE,
 172         RTE_FLOW_ACTION_TYPE_MARK,
 173         RTE_FLOW_ACTION_TYPE_FLAG,
 174         RTE_FLOW_ACTION_TYPE_END,
 175 };
 176
 177 /** Graph of supported items and associated actions. */
 178 static const struct mlx5_flow_items mlx5_flow_items[] = {
 179         [RTE_FLOW_ITEM_TYPE_END] = {
 180                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_ETH,
 181                                RTE_FLOW_ITEM_TYPE_VXLAN),
 182         },
 183         [RTE_FLOW_ITEM_TYPE_ETH] = {
 184                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_VLAN,
 185                                RTE_FLOW_ITEM_TYPE_IPV4,
 186                                RTE_FLOW_ITEM_TYPE_IPV6),
 187                 .actions = valid_actions,
 188                 .mask = &(const struct rte_flow_item_eth){
 189                         .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 190                         .src.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 191                         .type = -1,
 192                 },
 193                 .default_mask = &rte_flow_item_eth_mask,
 194                 .mask_sz = sizeof(struct rte_flow_item_eth),
 195                 .convert = mlx5_flow_create_eth,
 196                 .dst_sz = sizeof(struct ibv_flow_spec_eth),
 197         },
 198         [RTE_FLOW_ITEM_TYPE_VLAN] = {
 199                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_IPV4,
 200                                RTE_FLOW_ITEM_TYPE_IPV6),
 201                 .actions = valid_actions,
 202                 .mask = &(const struct rte_flow_item_vlan){
 203                         .tci = -1,
 204                 },
 205                 .default_mask = &rte_flow_item_vlan_mask,
 206                 .mask_sz = sizeof(struct rte_flow_item_vlan),
 207                 .convert = mlx5_flow_create_vlan,
 208                 .dst_sz = 0,
 209         },
 210         [RTE_FLOW_ITEM_TYPE_IPV4] = {
 211                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP,
 212                                RTE_FLOW_ITEM_TYPE_TCP),
 213                 .actions = valid_actions,
 214                 .mask = &(const struct rte_flow_item_ipv4){
 215                         .hdr = {
 216                                 .src_addr = -1,
 217                                 .dst_addr = -1,
 218                                 .type_of_service = -1,
 219                                 .next_proto_id = -1,
 220                         },
 221                 },
 222                 .default_mask = &rte_flow_item_ipv4_mask,
 223                 .mask_sz = sizeof(struct rte_flow_item_ipv4),
 224                 .convert = mlx5_flow_create_ipv4,
 225                 .dst_sz = sizeof(struct ibv_flow_spec_ipv4),
 226         },
 227         [RTE_FLOW_ITEM_TYPE_IPV6] = {
 228                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP,
 229                                RTE_FLOW_ITEM_TYPE_TCP),
 230                 .actions = valid_actions,
 231                 .mask = &(const struct rte_flow_item_ipv6){
 232                         .hdr = {
 233                                 .src_addr = {
 234                                         0xff, 0xff, 0xff, 0xff,
 235                                         0xff, 0xff, 0xff, 0xff,
 236                                         0xff, 0xff, 0xff, 0xff,
 237                                         0xff, 0xff, 0xff, 0xff,
 238                                 },
 239                                 .dst_addr = {
 240                                         0xff, 0xff, 0xff, 0xff,
 241                                         0xff, 0xff, 0xff, 0xff,
 242                                         0xff, 0xff, 0xff, 0xff,
 243                                         0xff, 0xff, 0xff, 0xff,
 244                                 },
 245                                 .vtc_flow = -1,
 246                                 .proto = -1,
 247                                 .hop_limits = -1,
 248                         },
 249                 },
 250                 .default_mask = &rte_flow_item_ipv6_mask,
 251                 .mask_sz = sizeof(struct rte_flow_item_ipv6),
 252                 .convert = mlx5_flow_create_ipv6,
 253                 .dst_sz = sizeof(struct ibv_flow_spec_ipv6),
 254         },
 255         [RTE_FLOW_ITEM_TYPE_UDP] = {
 256                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_VXLAN),
 257                 .actions = valid_actions,
 258                 .mask = &(const struct rte_flow_item_udp){
 259                         .hdr = {
 260                                 .src_port = -1,
 261                                 .dst_port = -1,
 262                         },
 263                 },
 264                 .default_mask = &rte_flow_item_udp_mask,
 265                 .mask_sz = sizeof(struct rte_flow_item_udp),
 266                 .convert = mlx5_flow_create_udp,
 267                 .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
 268         },
 269         [RTE_FLOW_ITEM_TYPE_TCP] = {
 270                 .actions = valid_actions,
 271                 .mask = &(const struct rte_flow_item_tcp){
 272                         .hdr = {
 273                                 .src_port = -1,
 274                                 .dst_port = -1,
 275                         },
 276                 },
 277                 .default_mask = &rte_flow_item_tcp_mask,
 278                 .mask_sz = sizeof(struct rte_flow_item_tcp),
 279                 .convert = mlx5_flow_create_tcp,
 280                 .dst_sz = sizeof(struct ibv_flow_spec_tcp_udp),
 281         },
 282         [RTE_FLOW_ITEM_TYPE_VXLAN] = {
 283                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_ETH),
 284                 .actions = valid_actions,
 285                 .mask = &(const struct rte_flow_item_vxlan){
 286                         .vni = "\xff\xff\xff",
 287                 },
 288                 .default_mask = &rte_flow_item_vxlan_mask,
 289                 .mask_sz = sizeof(struct rte_flow_item_vxlan),
 290                 .convert = mlx5_flow_create_vxlan,
 291                 .dst_sz = sizeof(struct ibv_flow_spec_tunnel),
 292         },
 293 };
 294
 295 /** Structure to pass to the conversion function. */
 296 struct mlx5_flow_parse {
 297         struct ibv_flow_attr *ibv_attr; /**< Verbs attribute. */
 298         unsigned int offset; /**< Offset in bytes in the ibv_attr buffer. */
 299         uint32_t inner; /**< Set once VXLAN is encountered. */
 300         uint32_t create:1; /**< Leave allocated resources on exit. */
 301         uint32_t queue:1; /**< Target is a receive queue. */
 302         uint32_t drop:1; /**< Target is a drop queue. */
 303         uint32_t mark:1; /**< Mark is present in the flow. */
 304         uint32_t mark_id; /**< Mark identifier. */
 305         uint64_t hash_fields; /**< Fields that participate in the hash. */
 306         uint16_t queues[RTE_MAX_QUEUES_PER_PORT]; /**< Queues indexes to use. */
 307         uint16_t queues_n; /**< Number of entries in queue[]. */
 308 };
 309
 310 static const struct rte_flow_ops mlx5_flow_ops = {
 311         .validate = mlx5_flow_validate,
 312         .create = mlx5_flow_create,
 313         .destroy = mlx5_flow_destroy,
 314         .flush = mlx5_flow_flush,
 315         .query = NULL,
 316         .isolate = mlx5_flow_isolate,
 317 };
 318
 319 /**
 320  * Manage filter operations.
 321  *
 322  * @param dev
 323  *   Pointer to Ethernet device structure.
 324  * @param filter_type
 325  *   Filter type.
 326  * @param filter_op
 327  *   Operation to perform.
 328  * @param arg
 329  *   Pointer to operation-specific structure.
 330  *
 331  * @return
 332  *   0 on success, negative errno value on failure.
 333  */
 334 int
 335 mlx5_dev_filter_ctrl(struct rte_eth_dev *dev,
 336                      enum rte_filter_type filter_type,
 337                      enum rte_filter_op filter_op,
 338                      void *arg)
 339 {
 340         int ret = EINVAL;
 341
 342         if (filter_type == RTE_ETH_FILTER_GENERIC) {
 343                 if (filter_op != RTE_ETH_FILTER_GET)
 344                         return -EINVAL;
 345                 *(const void **)arg = &mlx5_flow_ops;
 346                 return 0;
 347         }
 348         ERROR("%p: filter type (%d) not supported",
 349               (void *)dev, filter_type);
 350         return -ret;
 351 }
 352
 353 /**
 354  * Check support for a given item.
 355  *
 356  * @param item[in]
 357  *   Item specification.
 358  * @param mask[in]
 359  *   Bit-masks covering supported fields to compare with spec, last and mask in
 360  *   \item.
 361  * @param size
 362  *   Bit-Mask size in bytes.
 363  *
 364  * @return
 365  *   0 on success.
 366  */
 367 static int
 368 mlx5_flow_item_validate(const struct rte_flow_item *item,
 369                         const uint8_t *mask, unsigned int size)
 370 {
 371         int ret = 0;
 372
 373         if (!item->spec && (item->mask || item->last))
 374                 return -1;
 375         if (item->spec && !item->mask) {
 376                 unsigned int i;
 377                 const uint8_t *spec = item->spec;
 378
 379                 for (i = 0; i < size; ++i)
 380                         if ((spec[i] | mask[i]) != mask[i])
 381                                 return -1;
 382         }
 383         if (item->last && !item->mask) {
 384                 unsigned int i;
 385                 const uint8_t *spec = item->last;
 386
 387                 for (i = 0; i < size; ++i)
 388                         if ((spec[i] | mask[i]) != mask[i])
 389                                 return -1;
 390         }
 391         if (item->mask) {
 392                 unsigned int i;
 393                 const uint8_t *spec = item->mask;
 394
 395                 for (i = 0; i < size; ++i)
 396                         if ((spec[i] | mask[i]) != mask[i])
 397                                 return -1;
 398         }
 399         if (item->spec && item->last) {
 400                 uint8_t spec[size];
 401                 uint8_t last[size];
 402                 const uint8_t *apply = mask;
 403                 unsigned int i;
 404
 405                 if (item->mask)
 406                         apply = item->mask;
 407                 for (i = 0; i < size; ++i) {
 408                         spec[i] = ((const uint8_t *)item->spec)[i] & apply[i];
 409                         last[i] = ((const uint8_t *)item->last)[i] & apply[i];
 410                 }
 411                 ret = memcmp(spec, last, size);
 412         }
 413         return ret;
 414 }
 415
 416 /**
 417  * Validate and convert a flow supported by the NIC.
 418  *
 419  * @param priv
 420  *   Pointer to private structure.
 421  * @param[in] attr
 422  *   Flow rule attributes.
 423  * @param[in] pattern
 424  *   Pattern specification (list terminated by the END pattern item).
 425  * @param[in] actions
 426  *   Associated actions (list terminated by the END action).
 427  * @param[out] error
 428  *   Perform verbose error reporting if not NULL.
 429  * @param[in, out] flow
 430  *   Flow structure to update.
 431  *
 432  * @return
 433  *   0 on success, a negative errno value otherwise and rte_errno is set.
 434  */
 435 static int
 436 priv_flow_convert(struct priv *priv,
 437                   const struct rte_flow_attr *attr,
 438                   const struct rte_flow_item items[],
 439                   const struct rte_flow_action actions[],
 440                   struct rte_flow_error *error,
 441                   struct mlx5_flow_parse *flow)
 442 {
 443         const struct mlx5_flow_items *cur_item = mlx5_flow_items;
 444
 445         (void)priv;
 446         *flow = (struct mlx5_flow_parse){
 447                 .ibv_attr = flow->ibv_attr,
 448                 .create = flow->create,
 449                 .offset = sizeof(struct ibv_flow_attr),
 450                 .mark_id = MLX5_FLOW_MARK_DEFAULT,
 451         };
 452         if (attr->group) {
 453                 rte_flow_error_set(error, ENOTSUP,
 454                                    RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
 455                                    NULL,
 456                                    "groups are not supported");
 457                 return -rte_errno;
 458         }
 459         if (attr->priority && attr->priority != MLX5_CTRL_FLOW_PRIORITY) {
 460                 rte_flow_error_set(error, ENOTSUP,
 461                                    RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
 462                                    NULL,
 463                                    "priorities are not supported");
 464                 return -rte_errno;
 465         }
 466         if (attr->egress) {
 467                 rte_flow_error_set(error, ENOTSUP,
 468                                    RTE_FLOW_ERROR_TYPE_ATTR_EGRESS,
 469                                    NULL,
 470                                    "egress is not supported");
 471                 return -rte_errno;
 472         }
 473         if (!attr->ingress) {
 474                 rte_flow_error_set(error, ENOTSUP,
 475                                    RTE_FLOW_ERROR_TYPE_ATTR_INGRESS,
 476                                    NULL,
 477                                    "only ingress is supported");
 478                 return -rte_errno;
 479         }
 480         for (; actions->type != RTE_FLOW_ACTION_TYPE_END; ++actions) {
 481                 if (actions->type == RTE_FLOW_ACTION_TYPE_VOID) {
 482                         continue;
 483                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_DROP) {
 484                         flow->drop = 1;
 485                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_QUEUE) {
 486                         const struct rte_flow_action_queue *queue =
 487                                 (const struct rte_flow_action_queue *)
 488                                 actions->conf;
 489                         uint16_t n;
 490                         uint16_t found = 0;
 491
 492                         if (!queue || (queue->index > (priv->rxqs_n - 1)))
 493                                 goto exit_action_not_supported;
 494                         for (n = 0; n < flow->queues_n; ++n) {
 495                                 if (flow->queues[n] == queue->index) {
 496                                         found = 1;
 497                                         break;
 498                                 }
 499                         }
 500                         if (flow->queues_n > 1 && !found) {
 501                                 rte_flow_error_set(error, ENOTSUP,
 502                                            RTE_FLOW_ERROR_TYPE_ACTION,
 503                                            actions,
 504                                            "queue action not in RSS queues");
 505                                 return -rte_errno;
 506                         }
 507                         if (!found) {
 508                                 flow->queue = 1;
 509                                 flow->queues_n = 1;
 510                                 flow->queues[0] = queue->index;
 511                         }
 512                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_RSS) {
 513                         const struct rte_flow_action_rss *rss =
 514                                 (const struct rte_flow_action_rss *)
 515                                 actions->conf;
 516                         uint16_t n;
 517
 518                         if (!rss || !rss->num) {
 519                                 rte_flow_error_set(error, EINVAL,
 520                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 521                                                    actions,
 522                                                    "no valid queues");
 523                                 return -rte_errno;
 524                         }
 525                         if (flow->queues_n == 1) {
 526                                 uint16_t found = 0;
 527
 528                                 assert(flow->queues_n);
 529                                 for (n = 0; n < rss->num; ++n) {
 530                                         if (flow->queues[0] ==
 531                                             rss->queue[n]) {
 532                                                 found = 1;
 533                                                 break;
 534                                         }
 535                                 }
 536                                 if (!found) {
 537                                         rte_flow_error_set(error, ENOTSUP,
 538                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 539                                                    actions,
 540                                                    "queue action not in RSS"
 541                                                    " queues");
 542                                         return -rte_errno;
 543                                 }
 544                         }
 545                         for (n = 0; n < rss->num; ++n) {
 546                                 if (rss->queue[n] >= priv->rxqs_n) {
 547                                         rte_flow_error_set(error, EINVAL,
 548                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 549                                                    actions,
 550                                                    "queue id > number of"
 551                                                    " queues");
 552                                         return -rte_errno;
 553                                 }
 554                         }
 555                         flow->queue = 1;
 556                         for (n = 0; n < rss->num; ++n)
 557                                 flow->queues[n] = rss->queue[n];
 558                         flow->queues_n = rss->num;
 559                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_MARK) {
 560                         const struct rte_flow_action_mark *mark =
 561                                 (const struct rte_flow_action_mark *)
 562                                 actions->conf;
 563
 564                         if (!mark) {
 565                                 rte_flow_error_set(error, EINVAL,
 566                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 567                                                    actions,
 568                                                    "mark must be defined");
 569                                 return -rte_errno;
 570                         } else if (mark->id >= MLX5_FLOW_MARK_MAX) {
 571                                 rte_flow_error_set(error, ENOTSUP,
 572                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 573                                                    actions,
 574                                                    "mark must be between 0"
 575                                                    " and 16777199");
 576                                 return -rte_errno;
 577                         }
 578                         flow->mark = 1;
 579                         flow->mark_id = mark->id;
 580                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_FLAG) {
 581                         flow->mark = 1;
 582                 } else {
 583                         goto exit_action_not_supported;
 584                 }
 585         }
 586         if (flow->mark && !flow->ibv_attr && !flow->drop)
 587                 flow->offset += sizeof(struct ibv_flow_spec_action_tag);
 588         if (!flow->ibv_attr && flow->drop)
 589                 flow->offset += sizeof(struct ibv_flow_spec_action_drop);
 590         if (!flow->queue && !flow->drop) {
 591                 rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_HANDLE,
 592                                    NULL, "no valid action");
 593                 return -rte_errno;
 594         }
 595         for (; items->type != RTE_FLOW_ITEM_TYPE_END; ++items) {
 596                 const struct mlx5_flow_items *token = NULL;
 597                 unsigned int i;
 598                 int err;
 599
 600                 if (items->type == RTE_FLOW_ITEM_TYPE_VOID)
 601                         continue;
 602                 for (i = 0;
 603                      cur_item->items &&
 604                      cur_item->items[i] != RTE_FLOW_ITEM_TYPE_END;
 605                      ++i) {
 606                         if (cur_item->items[i] == items->type) {
 607                                 token = &mlx5_flow_items[items->type];
 608                                 break;
 609                         }
 610                 }
 611                 if (!token)
 612                         goto exit_item_not_supported;
 613                 cur_item = token;
 614                 err = mlx5_flow_item_validate(items,
 615                                               (const uint8_t *)cur_item->mask,
 616                                               cur_item->mask_sz);
 617                 if (err)
 618                         goto exit_item_not_supported;
 619                 if (flow->ibv_attr && cur_item->convert) {
 620                         err = cur_item->convert(items,
 621                                                 (cur_item->default_mask ?
 622                                                  cur_item->default_mask :
 623                                                  cur_item->mask),
 624                                                 flow);
 625                         if (err)
 626                                 goto exit_item_not_supported;
 627                 } else if (items->type == RTE_FLOW_ITEM_TYPE_VXLAN) {
 628                         if (flow->inner) {
 629                                 rte_flow_error_set(error, ENOTSUP,
 630                                                    RTE_FLOW_ERROR_TYPE_ITEM,
 631                                                    items,
 632                                                    "cannot recognize multiple"
 633                                                    " VXLAN encapsulations");
 634                                 return -rte_errno;
 635                         }
 636                         flow->inner = 1;
 637                 }
 638                 flow->offset += cur_item->dst_sz;
 639         }
 640         return 0;
 641 exit_item_not_supported:
 642         rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 643                            items, "item not supported");
 644         return -rte_errno;
 645 exit_action_not_supported:
 646         rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
 647                            actions, "action not supported");
 648         return -rte_errno;
 649 }
 650
 651 /**
 652  * Convert Ethernet item to Verbs specification.
 653  *
 654  * @param item[in]
 655  *   Item specification.
 656  * @param default_mask[in]
 657  *   Default bit-masks to use when item->mask is not provided.
 658  * @param data[in, out]
 659  *   User structure.
 660  */
 661 static int
 662 mlx5_flow_create_eth(const struct rte_flow_item *item,
 663                      const void *default_mask,
 664                      void *data)
 665 {
 666         const struct rte_flow_item_eth *spec = item->spec;
 667         const struct rte_flow_item_eth *mask = item->mask;
 668         struct mlx5_flow_parse *flow = (struct mlx5_flow_parse *)data;
 669         struct ibv_flow_spec_eth *eth;
 670         const unsigned int eth_size = sizeof(struct ibv_flow_spec_eth);
 671         unsigned int i;
 672
 673         ++flow->ibv_attr->num_of_specs;
 674         flow->ibv_attr->priority = 2;
 675         flow->hash_fields = 0;
 676         eth = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 677         *eth = (struct ibv_flow_spec_eth) {
 678                 .type = flow->inner | IBV_FLOW_SPEC_ETH,
 679                 .size = eth_size,
 680         };
 681         if (!spec)
 682                 return 0;
 683         if (!mask)
 684                 mask = default_mask;
 685         memcpy(eth->val.dst_mac, spec->dst.addr_bytes, ETHER_ADDR_LEN);
 686         memcpy(eth->val.src_mac, spec->src.addr_bytes, ETHER_ADDR_LEN);
 687         eth->val.ether_type = spec->type;
 688         memcpy(eth->mask.dst_mac, mask->dst.addr_bytes, ETHER_ADDR_LEN);
 689         memcpy(eth->mask.src_mac, mask->src.addr_bytes, ETHER_ADDR_LEN);
 690         eth->mask.ether_type = mask->type;
 691         /* Remove unwanted bits from values. */
 692         for (i = 0; i < ETHER_ADDR_LEN; ++i) {
 693                 eth->val.dst_mac[i] &= eth->mask.dst_mac[i];
 694                 eth->val.src_mac[i] &= eth->mask.src_mac[i];
 695         }
 696         eth->val.ether_type &= eth->mask.ether_type;
 697         return 0;
 698 }
 699
 700 /**
 701  * Convert VLAN item to Verbs specification.
 702  *
 703  * @param item[in]
 704  *   Item specification.
 705  * @param default_mask[in]
 706  *   Default bit-masks to use when item->mask is not provided.
 707  * @param data[in, out]
 708  *   User structure.
 709  */
 710 static int
 711 mlx5_flow_create_vlan(const struct rte_flow_item *item,
 712                       const void *default_mask,
 713                       void *data)
 714 {
 715         const struct rte_flow_item_vlan *spec = item->spec;
 716         const struct rte_flow_item_vlan *mask = item->mask;
 717         struct mlx5_flow_parse *flow = (struct mlx5_flow_parse *)data;
 718         struct ibv_flow_spec_eth *eth;
 719         const unsigned int eth_size = sizeof(struct ibv_flow_spec_eth);
 720
 721         eth = (void *)((uintptr_t)flow->ibv_attr + flow->offset - eth_size);
 722         if (!spec)
 723                 return 0;
 724         if (!mask)
 725                 mask = default_mask;
 726         eth->val.vlan_tag = spec->tci;
 727         eth->mask.vlan_tag = mask->tci;
 728         eth->val.vlan_tag &= eth->mask.vlan_tag;
 729         return 0;
 730 }
 731
 732 /**
 733  * Convert IPv4 item to Verbs specification.
 734  *
 735  * @param item[in]
 736  *   Item specification.
 737  * @param default_mask[in]
 738  *   Default bit-masks to use when item->mask is not provided.
 739  * @param data[in, out]
 740  *   User structure.
 741  */
 742 static int
 743 mlx5_flow_create_ipv4(const struct rte_flow_item *item,
 744                       const void *default_mask,
 745                       void *data)
 746 {
 747         const struct rte_flow_item_ipv4 *spec = item->spec;
 748         const struct rte_flow_item_ipv4 *mask = item->mask;
 749         struct mlx5_flow_parse *flow = (struct mlx5_flow_parse *)data;
 750         struct ibv_flow_spec_ipv4_ext *ipv4;
 751         unsigned int ipv4_size = sizeof(struct ibv_flow_spec_ipv4_ext);
 752
 753         ++flow->ibv_attr->num_of_specs;
 754         flow->ibv_attr->priority = 1;
 755         flow->hash_fields = (IBV_RX_HASH_SRC_IPV4 |
 756                              IBV_RX_HASH_DST_IPV4);
 757         ipv4 = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 758         *ipv4 = (struct ibv_flow_spec_ipv4_ext) {
 759                 .type = flow->inner | IBV_FLOW_SPEC_IPV4_EXT,
 760                 .size = ipv4_size,
 761         };
 762         if (!spec)
 763                 return 0;
 764         if (!mask)
 765                 mask = default_mask;
 766         ipv4->val = (struct ibv_flow_ipv4_ext_filter){
 767                 .src_ip = spec->hdr.src_addr,
 768                 .dst_ip = spec->hdr.dst_addr,
 769                 .proto = spec->hdr.next_proto_id,
 770                 .tos = spec->hdr.type_of_service,
 771         };
 772         ipv4->mask = (struct ibv_flow_ipv4_ext_filter){
 773                 .src_ip = mask->hdr.src_addr,
 774                 .dst_ip = mask->hdr.dst_addr,
 775                 .proto = mask->hdr.next_proto_id,
 776                 .tos = mask->hdr.type_of_service,
 777         };
 778         /* Remove unwanted bits from values. */
 779         ipv4->val.src_ip &= ipv4->mask.src_ip;
 780         ipv4->val.dst_ip &= ipv4->mask.dst_ip;
 781         ipv4->val.proto &= ipv4->mask.proto;
 782         ipv4->val.tos &= ipv4->mask.tos;
 783         return 0;
 784 }
 785
 786 /**
 787  * Convert IPv6 item to Verbs specification.
 788  *
 789  * @param item[in]
 790  *   Item specification.
 791  * @param default_mask[in]
 792  *   Default bit-masks to use when item->mask is not provided.
 793  * @param data[in, out]
 794  *   User structure.
 795  */
 796 static int
 797 mlx5_flow_create_ipv6(const struct rte_flow_item *item,
 798                       const void *default_mask,
 799                       void *data)
 800 {
 801         const struct rte_flow_item_ipv6 *spec = item->spec;
 802         const struct rte_flow_item_ipv6 *mask = item->mask;
 803         struct mlx5_flow_parse *flow = (struct mlx5_flow_parse *)data;
 804         struct ibv_flow_spec_ipv6 *ipv6;
 805         unsigned int ipv6_size = sizeof(struct ibv_flow_spec_ipv6);
 806         unsigned int i;
 807
 808         ++flow->ibv_attr->num_of_specs;
 809         flow->ibv_attr->priority = 1;
 810         flow->hash_fields = (IBV_RX_HASH_SRC_IPV6 |
 811                              IBV_RX_HASH_DST_IPV6);
 812         ipv6 = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 813         *ipv6 = (struct ibv_flow_spec_ipv6) {
 814                 .type = flow->inner | IBV_FLOW_SPEC_IPV6,
 815                 .size = ipv6_size,
 816         };
 817         if (!spec)
 818                 return 0;
 819         if (!mask)
 820                 mask = default_mask;
 821         memcpy(ipv6->val.src_ip, spec->hdr.src_addr,
 822                RTE_DIM(ipv6->val.src_ip));
 823         memcpy(ipv6->val.dst_ip, spec->hdr.dst_addr,
 824                RTE_DIM(ipv6->val.dst_ip));
 825         memcpy(ipv6->mask.src_ip, mask->hdr.src_addr,
 826                RTE_DIM(ipv6->mask.src_ip));
 827         memcpy(ipv6->mask.dst_ip, mask->hdr.dst_addr,
 828                RTE_DIM(ipv6->mask.dst_ip));
 829         ipv6->mask.flow_label = mask->hdr.vtc_flow;
 830         ipv6->mask.next_hdr = mask->hdr.proto;
 831         ipv6->mask.hop_limit = mask->hdr.hop_limits;
 832         /* Remove unwanted bits from values. */
 833         for (i = 0; i < RTE_DIM(ipv6->val.src_ip); ++i) {
 834                 ipv6->val.src_ip[i] &= ipv6->mask.src_ip[i];
 835                 ipv6->val.dst_ip[i] &= ipv6->mask.dst_ip[i];
 836         }
 837         ipv6->val.flow_label &= ipv6->mask.flow_label;
 838         ipv6->val.next_hdr &= ipv6->mask.next_hdr;
 839         ipv6->val.hop_limit &= ipv6->mask.hop_limit;
 840         return 0;
 841 }
 842
 843 /**
 844  * Convert UDP item to Verbs specification.
 845  *
 846  * @param item[in]
 847  *   Item specification.
 848  * @param default_mask[in]
 849  *   Default bit-masks to use when item->mask is not provided.
 850  * @param data[in, out]
 851  *   User structure.
 852  */
 853 static int
 854 mlx5_flow_create_udp(const struct rte_flow_item *item,
 855                      const void *default_mask,
 856                      void *data)
 857 {
 858         const struct rte_flow_item_udp *spec = item->spec;
 859         const struct rte_flow_item_udp *mask = item->mask;
 860         struct mlx5_flow_parse *flow = (struct mlx5_flow_parse *)data;
 861         struct ibv_flow_spec_tcp_udp *udp;
 862         unsigned int udp_size = sizeof(struct ibv_flow_spec_tcp_udp);
 863
 864         ++flow->ibv_attr->num_of_specs;
 865         flow->ibv_attr->priority = 0;
 866         flow->hash_fields |= (IBV_RX_HASH_SRC_PORT_UDP |
 867                               IBV_RX_HASH_DST_PORT_UDP);
 868         udp = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 869         *udp = (struct ibv_flow_spec_tcp_udp) {
 870                 .type = flow->inner | IBV_FLOW_SPEC_UDP,
 871                 .size = udp_size,
 872         };
 873         if (!spec)
 874                 return 0;
 875         if (!mask)
 876                 mask = default_mask;
 877         udp->val.dst_port = spec->hdr.dst_port;
 878         udp->val.src_port = spec->hdr.src_port;
 879         udp->mask.dst_port = mask->hdr.dst_port;
 880         udp->mask.src_port = mask->hdr.src_port;
 881         /* Remove unwanted bits from values. */
 882         udp->val.src_port &= udp->mask.src_port;
 883         udp->val.dst_port &= udp->mask.dst_port;
 884         return 0;
 885 }
 886
 887 /**
 888  * Convert TCP item to Verbs specification.
 889  *
 890  * @param item[in]
 891  *   Item specification.
 892  * @param default_mask[in]
 893  *   Default bit-masks to use when item->mask is not provided.
 894  * @param data[in, out]
 895  *   User structure.
 896  */
 897 static int
 898 mlx5_flow_create_tcp(const struct rte_flow_item *item,
 899                      const void *default_mask,
 900                      void *data)
 901 {
 902         const struct rte_flow_item_tcp *spec = item->spec;
 903         const struct rte_flow_item_tcp *mask = item->mask;
 904         struct mlx5_flow_parse *flow = (struct mlx5_flow_parse *)data;
 905         struct ibv_flow_spec_tcp_udp *tcp;
 906         unsigned int tcp_size = sizeof(struct ibv_flow_spec_tcp_udp);
 907
 908         ++flow->ibv_attr->num_of_specs;
 909         flow->ibv_attr->priority = 0;
 910         flow->hash_fields |= (IBV_RX_HASH_SRC_PORT_TCP |
 911                               IBV_RX_HASH_DST_PORT_TCP);
 912         tcp = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 913         *tcp = (struct ibv_flow_spec_tcp_udp) {
 914                 .type = flow->inner | IBV_FLOW_SPEC_TCP,
 915                 .size = tcp_size,
 916         };
 917         if (!spec)
 918                 return 0;
 919         if (!mask)
 920                 mask = default_mask;
 921         tcp->val.dst_port = spec->hdr.dst_port;
 922         tcp->val.src_port = spec->hdr.src_port;
 923         tcp->mask.dst_port = mask->hdr.dst_port;
 924         tcp->mask.src_port = mask->hdr.src_port;
 925         /* Remove unwanted bits from values. */
 926         tcp->val.src_port &= tcp->mask.src_port;
 927         tcp->val.dst_port &= tcp->mask.dst_port;
 928         return 0;
 929 }
 930
 931 /**
 932  * Convert VXLAN item to Verbs specification.
 933  *
 934  * @param item[in]
 935  *   Item specification.
 936  * @param default_mask[in]
 937  *   Default bit-masks to use when item->mask is not provided.
 938  * @param data[in, out]
 939  *   User structure.
 940  */
 941 static int
 942 mlx5_flow_create_vxlan(const struct rte_flow_item *item,
 943                        const void *default_mask,
 944                        void *data)
 945 {
 946         const struct rte_flow_item_vxlan *spec = item->spec;
 947         const struct rte_flow_item_vxlan *mask = item->mask;
 948         struct mlx5_flow_parse *flow = (struct mlx5_flow_parse *)data;
 949         struct ibv_flow_spec_tunnel *vxlan;
 950         unsigned int size = sizeof(struct ibv_flow_spec_tunnel);
 951         union vni {
 952                 uint32_t vlan_id;
 953                 uint8_t vni[4];
 954         } id;
 955
 956         ++flow->ibv_attr->num_of_specs;
 957         flow->ibv_attr->priority = 0;
 958         id.vni[0] = 0;
 959         vxlan = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 960         *vxlan = (struct ibv_flow_spec_tunnel) {
 961                 .type = flow->inner | IBV_FLOW_SPEC_VXLAN_TUNNEL,
 962                 .size = size,
 963         };
 964         flow->inner = IBV_FLOW_SPEC_INNER;
 965         if (!spec)
 966                 return 0;
 967         if (!mask)
 968                 mask = default_mask;
 969         memcpy(&id.vni[1], spec->vni, 3);
 970         vxlan->val.tunnel_id = id.vlan_id;
 971         memcpy(&id.vni[1], mask->vni, 3);
 972         vxlan->mask.tunnel_id = id.vlan_id;
 973         /* Remove unwanted bits from values. */
 974         vxlan->val.tunnel_id &= vxlan->mask.tunnel_id;
 975         return 0;
 976 }
 977
 978 /**
 979  * Convert mark/flag action to Verbs specification.
 980  *
 981  * @param flow
 982  *   Pointer to MLX5 flow structure.
 983  * @param mark_id
 984  *   Mark identifier.
 985  */
 986 static int
 987 mlx5_flow_create_flag_mark(struct mlx5_flow_parse *flow, uint32_t mark_id)
 988 {
 989         struct ibv_flow_spec_action_tag *tag;
 990         unsigned int size = sizeof(struct ibv_flow_spec_action_tag);
 991
 992         assert(flow->mark);
 993         tag = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 994         *tag = (struct ibv_flow_spec_action_tag){
 995                 .type = IBV_FLOW_SPEC_ACTION_TAG,
 996                 .size = size,
 997                 .tag_id = mlx5_flow_mark_set(mark_id),
 998         };
 999         ++flow->ibv_attr->num_of_specs;
1000         flow->offset += size;
1001         return 0;
1002 }
1003
1004 /**
1005  * Complete flow rule creation with a drop queue.
1006  *
1007  * @param priv
1008  *   Pointer to private structure.
1009  * @param flow
1010  *   MLX5 flow attributes (filled by mlx5_flow_validate()).
1011  * @param[out] error
1012  *   Perform verbose error reporting if not NULL.
1013  *
1014  * @return
1015  *   A flow if the rule could be created.
1016  */
1017 static struct rte_flow *
1018 priv_flow_create_action_queue_drop(struct priv *priv,
1019                                    struct mlx5_flow_parse *flow,
1020                                    struct rte_flow_error *error)
1021 {
1022         struct rte_flow *rte_flow;
1023         struct ibv_flow_spec_action_drop *drop;
1024         unsigned int size = sizeof(struct ibv_flow_spec_action_drop);
1025
1026         assert(priv->pd);
1027         assert(priv->ctx);
1028         rte_flow = rte_calloc(__func__, 1, sizeof(*rte_flow), 0);
1029         if (!rte_flow) {
1030                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1031                                    NULL, "cannot allocate flow memory");
1032                 return NULL;
1033         }
1034         rte_flow->drop = 1;
1035         drop = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
1036         *drop = (struct ibv_flow_spec_action_drop){
1037                         .type = IBV_FLOW_SPEC_ACTION_DROP,
1038                         .size = size,
1039         };
1040         ++flow->ibv_attr->num_of_specs;
1041         flow->offset += sizeof(struct ibv_flow_spec_action_drop);
1042         rte_flow->ibv_attr = flow->ibv_attr;
1043         if (!priv->dev->data->dev_started)
1044                 return rte_flow;
1045         rte_flow->drxq.hrxq.qp = priv->flow_drop_queue->qp;
1046         rte_flow->ibv_flow = ibv_create_flow(rte_flow->drxq.hrxq.qp,
1047                                              rte_flow->ibv_attr);
1048         if (!rte_flow->ibv_flow) {
1049                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1050                                    NULL, "flow rule creation failure");
1051                 goto error;
1052         }
1053         return rte_flow;
1054 error:
1055         assert(rte_flow);
1056         rte_free(rte_flow);
1057         return NULL;
1058 }
1059
1060 /**
1061  * Complete flow rule creation.
1062  *
1063  * @param priv
1064  *   Pointer to private structure.
1065  * @param flow
1066  *   MLX5 flow attributes (filled by mlx5_flow_validate()).
1067  * @param[out] error
1068  *   Perform verbose error reporting if not NULL.
1069  *
1070  * @return
1071  *   A flow if the rule could be created.
1072  */
1073 static struct rte_flow *
1074 priv_flow_create_action_queue(struct priv *priv,
1075                               struct mlx5_flow_parse *flow,
1076                               struct rte_flow_error *error)
1077 {
1078         struct rte_flow *rte_flow;
1079         unsigned int i;
1080
1081         assert(priv->pd);
1082         assert(priv->ctx);
1083         assert(!flow->drop);
1084         rte_flow = rte_calloc(__func__, 1,
1085                               sizeof(*rte_flow) +
1086                               flow->queues_n * sizeof(uint16_t),
1087                               0);
1088         if (!rte_flow) {
1089                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1090                                    NULL, "cannot allocate flow memory");
1091                 return NULL;
1092         }
1093         rte_flow->mark = flow->mark;
1094         rte_flow->ibv_attr = flow->ibv_attr;
1095         rte_flow->queues = (uint16_t (*)[])(rte_flow + 1);
1096         memcpy(rte_flow->queues, flow->queues,
1097                flow->queues_n * sizeof(uint16_t));
1098         rte_flow->queues_n = flow->queues_n;
1099         rte_flow->frxq.hash_fields = flow->hash_fields;
1100         rte_flow->frxq.hrxq = mlx5_priv_hrxq_get(priv, rss_hash_default_key,
1101                                                  rss_hash_default_key_len,
1102                                                  flow->hash_fields,
1103                                                  (*rte_flow->queues),
1104                                                  rte_flow->queues_n);
1105         if (!rte_flow->frxq.hrxq) {
1106                 rte_flow->frxq.hrxq =
1107                         mlx5_priv_hrxq_new(priv, rss_hash_default_key,
1108                                            rss_hash_default_key_len,
1109                                            flow->hash_fields,
1110                                            (*rte_flow->queues),
1111                                            rte_flow->queues_n);
1112                 if (!rte_flow->frxq.hrxq) {
1113                         rte_flow_error_set(error, ENOMEM,
1114                                            RTE_FLOW_ERROR_TYPE_HANDLE,
1115                                            NULL, "cannot create hash rxq");
1116                         goto error;
1117                 }
1118         }
1119         for (i = 0; i != flow->queues_n; ++i) {
1120                 struct mlx5_rxq_data *q =
1121                         (*priv->rxqs)[flow->queues[i]];
1122
1123                 q->mark |= flow->mark;
1124         }
1125         if (!priv->dev->data->dev_started)
1126                 return rte_flow;
1127         rte_flow->ibv_flow = ibv_create_flow(rte_flow->frxq.hrxq->qp,
1128                                              rte_flow->ibv_attr);
1129         if (!rte_flow->ibv_flow) {
1130                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1131                                    NULL, "flow rule creation failure");
1132                 goto error;
1133         }
1134         return rte_flow;
1135 error:
1136         assert(rte_flow);
1137         if (rte_flow->frxq.hrxq)
1138                 mlx5_priv_hrxq_release(priv, rte_flow->frxq.hrxq);
1139         rte_free(rte_flow);
1140         return NULL;
1141 }
1142
1143 /**
1144  * Validate a flow.
1145  *
1146  * @param priv
1147  *   Pointer to private structure.
1148  * @param[in] attr
1149  *   Flow rule attributes.
1150  * @param[in] pattern
1151  *   Pattern specification (list terminated by the END pattern item).
1152  * @param[in] actions
1153  *   Associated actions (list terminated by the END action).
1154  * @param[out] error
1155  *   Perform verbose error reporting if not NULL.
1156  * @param[in,out] parser
1157  *   MLX5 parser structure.
1158  *
1159  * @return
1160  *   0 on success, negative errno value on failure.
1161  */
1162 static int
1163 priv_flow_validate(struct priv *priv,
1164                    const struct rte_flow_attr *attr,
1165                    const struct rte_flow_item items[],
1166                    const struct rte_flow_action actions[],
1167                    struct rte_flow_error *error,
1168                    struct mlx5_flow_parse *parser)
1169 {
1170         int err;
1171
1172         err = priv_flow_convert(priv, attr, items, actions, error, parser);
1173         if (err)
1174                 goto exit;
1175         if (parser->mark)
1176                 parser->offset += sizeof(struct ibv_flow_spec_action_tag);
1177         parser->ibv_attr = rte_malloc(__func__, parser->offset, 0);
1178         if (!parser->ibv_attr) {
1179                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1180                                    NULL, "cannot allocate ibv_attr memory");
1181                 err = rte_errno;
1182                 goto exit;
1183         }
1184         *parser->ibv_attr = (struct ibv_flow_attr){
1185                 .type = IBV_FLOW_ATTR_NORMAL,
1186                 .size = sizeof(struct ibv_flow_attr),
1187                 .priority = attr->priority,
1188                 .num_of_specs = 0,
1189                 .port = 0,
1190                 .flags = 0,
1191         };
1192         err = priv_flow_convert(priv, attr, items, actions, error, parser);
1193         if (err || parser->create)
1194                 goto exit;
1195         if (parser->mark)
1196                 mlx5_flow_create_flag_mark(parser, parser->mark_id);
1197         return 0;
1198 exit:
1199         if (parser->ibv_attr)
1200                 rte_free(parser->ibv_attr);
1201         return err;
1202 }
1203
1204 /**
1205  * Convert a flow.
1206  *
1207  * @param priv
1208  *   Pointer to private structure.
1209  * @param list
1210  *   Pointer to a TAILQ flow list.
1211  * @param[in] attr
1212  *   Flow rule attributes.
1213  * @param[in] pattern
1214  *   Pattern specification (list terminated by the END pattern item).
1215  * @param[in] actions
1216  *   Associated actions (list terminated by the END action).
1217  * @param[out] error
1218  *   Perform verbose error reporting if not NULL.
1219  *
1220  * @return
1221  *   A flow on success, NULL otherwise.
1222  */
1223 static struct rte_flow *
1224 priv_flow_create(struct priv *priv,
1225                  struct mlx5_flows *list,
1226                  const struct rte_flow_attr *attr,
1227                  const struct rte_flow_item items[],
1228                  const struct rte_flow_action actions[],
1229                  struct rte_flow_error *error)
1230 {
1231         struct mlx5_flow_parse parser = { .create = 1, };
1232         struct rte_flow *flow;
1233         int err;
1234
1235         err = priv_flow_validate(priv, attr, items, actions, error, &parser);
1236         if (err)
1237                 goto exit;
1238         if (parser.drop)
1239                 flow = priv_flow_create_action_queue_drop(priv, &parser, error);
1240         else
1241                 flow = priv_flow_create_action_queue(priv, &parser, error);
1242         if (!flow)
1243                 goto exit;
1244         TAILQ_INSERT_TAIL(list, flow, next);
1245         DEBUG("Flow created %p", (void *)flow);
1246         return flow;
1247 exit:
1248         if (parser.ibv_attr)
1249                 rte_free(parser.ibv_attr);
1250         return NULL;
1251 }
1252
1253 /**
1254  * Validate a flow supported by the NIC.
1255  *
1256  * @see rte_flow_validate()
1257  * @see rte_flow_ops
1258  */
1259 int
1260 mlx5_flow_validate(struct rte_eth_dev *dev,
1261                    const struct rte_flow_attr *attr,
1262                    const struct rte_flow_item items[],
1263                    const struct rte_flow_action actions[],
1264                    struct rte_flow_error *error)
1265 {
1266         struct priv *priv = dev->data->dev_private;
1267         int ret;
1268         struct mlx5_flow_parse parser = { .create = 0, };
1269
1270         priv_lock(priv);
1271         ret = priv_flow_validate(priv, attr, items, actions, error, &parser);
1272         priv_unlock(priv);
1273         return ret;
1274 }
1275
1276 /**
1277  * Create a flow.
1278  *
1279  * @see rte_flow_create()
1280  * @see rte_flow_ops
1281  */
1282 struct rte_flow *
1283 mlx5_flow_create(struct rte_eth_dev *dev,
1284                  const struct rte_flow_attr *attr,
1285                  const struct rte_flow_item items[],
1286                  const struct rte_flow_action actions[],
1287                  struct rte_flow_error *error)
1288 {
1289         struct priv *priv = dev->data->dev_private;
1290         struct rte_flow *flow;
1291
1292         priv_lock(priv);
1293         flow = priv_flow_create(priv, &priv->flows, attr, items, actions,
1294                                 error);
1295         priv_unlock(priv);
1296         return flow;
1297 }
1298
1299 /**
1300  * Destroy a flow.
1301  *
1302  * @param priv
1303  *   Pointer to private structure.
1304  * @param list
1305  *   Pointer to a TAILQ flow list.
1306  * @param[in] flow
1307  *   Flow to destroy.
1308  */
1309 static void
1310 priv_flow_destroy(struct priv *priv,
1311                   struct mlx5_flows *list,
1312                   struct rte_flow *flow)
1313 {
1314         unsigned int i;
1315         uint16_t *queues;
1316         uint16_t queues_n;
1317
1318         if (flow->drop || !flow->mark)
1319                 goto free;
1320         queues = flow->frxq.hrxq->ind_table->queues;
1321         queues_n = flow->frxq.hrxq->ind_table->queues_n;
1322         for (i = 0; i != queues_n; ++i) {
1323                 struct rte_flow *tmp;
1324                 struct mlx5_rxq_data *rxq_data = (*priv->rxqs)[queues[i]];
1325                 int mark = 0;
1326
1327                 /*
1328                  * To remove the mark from the queue, the queue must not be
1329                  * present in any other marked flow (RSS or not).
1330                  */
1331                 TAILQ_FOREACH(tmp, list, next) {
1332                         unsigned int j;
1333
1334                         if (!tmp->mark)
1335                                 continue;
1336                         for (j = 0;
1337                              (j != tmp->frxq.hrxq->ind_table->queues_n) &&
1338                              !mark;
1339                              j++)
1340                                 if (tmp->frxq.hrxq->ind_table->queues[j] ==
1341                                     queues[i])
1342                                         mark = 1;
1343                 }
1344                 rxq_data->mark = mark;
1345         }
1346 free:
1347         if (flow->ibv_flow)
1348                 claim_zero(ibv_destroy_flow(flow->ibv_flow));
1349         if (!flow->drop)
1350                 mlx5_priv_hrxq_release(priv, flow->frxq.hrxq);
1351         TAILQ_REMOVE(list, flow, next);
1352         rte_free(flow->ibv_attr);
1353         DEBUG("Flow destroyed %p", (void *)flow);
1354         rte_free(flow);
1355 }
1356
1357 /**
1358  * Destroy a flow.
1359  *
1360  * @see rte_flow_destroy()
1361  * @see rte_flow_ops
1362  */
1363 int
1364 mlx5_flow_destroy(struct rte_eth_dev *dev,
1365                   struct rte_flow *flow,
1366                   struct rte_flow_error *error)
1367 {
1368         struct priv *priv = dev->data->dev_private;
1369
1370         (void)error;
1371         priv_lock(priv);
1372         priv_flow_destroy(priv, &priv->flows, flow);
1373         priv_unlock(priv);
1374         return 0;
1375 }
1376
1377 /**
1378  * Destroy all flows.
1379  *
1380  * @param priv
1381  *   Pointer to private structure.
1382  * @param list
1383  *   Pointer to a TAILQ flow list.
1384  */
1385 void
1386 priv_flow_flush(struct priv *priv, struct mlx5_flows *list)
1387 {
1388         while (!TAILQ_EMPTY(list)) {
1389                 struct rte_flow *flow;
1390
1391                 flow = TAILQ_FIRST(list);
1392                 priv_flow_destroy(priv, list, flow);
1393         }
1394 }
1395
1396 /**
1397  * Destroy all flows.
1398  *
1399  * @see rte_flow_flush()
1400  * @see rte_flow_ops
1401  */
1402 int
1403 mlx5_flow_flush(struct rte_eth_dev *dev,
1404                 struct rte_flow_error *error)
1405 {
1406         struct priv *priv = dev->data->dev_private;
1407
1408         (void)error;
1409         priv_lock(priv);
1410         priv_flow_flush(priv, &priv->flows);
1411         priv_unlock(priv);
1412         return 0;
1413 }
1414
1415 /**
1416  * Create drop queue.
1417  *
1418  * @param priv
1419  *   Pointer to private structure.
1420  *
1421  * @return
1422  *   0 on success.
1423  */
1424 int
1425 priv_flow_create_drop_queue(struct priv *priv)
1426 {
1427         struct mlx5_hrxq_drop *fdq = NULL;
1428
1429         assert(priv->pd);
1430         assert(priv->ctx);
1431         fdq = rte_calloc(__func__, 1, sizeof(*fdq), 0);
1432         if (!fdq) {
1433                 WARN("cannot allocate memory for drop queue");
1434                 goto error;
1435         }
1436         fdq->cq = ibv_create_cq(priv->ctx, 1, NULL, NULL, 0);
1437         if (!fdq->cq) {
1438                 WARN("cannot allocate CQ for drop queue");
1439                 goto error;
1440         }
1441         fdq->wq = ibv_create_wq(priv->ctx,
1442                         &(struct ibv_wq_init_attr){
1443                         .wq_type = IBV_WQT_RQ,
1444                         .max_wr = 1,
1445                         .max_sge = 1,
1446                         .pd = priv->pd,
1447                         .cq = fdq->cq,
1448                         });
1449         if (!fdq->wq) {
1450                 WARN("cannot allocate WQ for drop queue");
1451                 goto error;
1452         }
1453         fdq->ind_table = ibv_create_rwq_ind_table(priv->ctx,
1454                         &(struct ibv_rwq_ind_table_init_attr){
1455                         .log_ind_tbl_size = 0,
1456                         .ind_tbl = &fdq->wq,
1457                         .comp_mask = 0,
1458                         });
1459         if (!fdq->ind_table) {
1460                 WARN("cannot allocate indirection table for drop queue");
1461                 goto error;
1462         }
1463         fdq->qp = ibv_create_qp_ex(priv->ctx,
1464                 &(struct ibv_qp_init_attr_ex){
1465                         .qp_type = IBV_QPT_RAW_PACKET,
1466                         .comp_mask =
1467                                 IBV_QP_INIT_ATTR_PD |
1468                                 IBV_QP_INIT_ATTR_IND_TABLE |
1469                                 IBV_QP_INIT_ATTR_RX_HASH,
1470                         .rx_hash_conf = (struct ibv_rx_hash_conf){
1471                                 .rx_hash_function =
1472                                         IBV_RX_HASH_FUNC_TOEPLITZ,
1473                                 .rx_hash_key_len = rss_hash_default_key_len,
1474                                 .rx_hash_key = rss_hash_default_key,
1475                                 .rx_hash_fields_mask = 0,
1476                                 },
1477                         .rwq_ind_tbl = fdq->ind_table,
1478                         .pd = priv->pd
1479                 });
1480         if (!fdq->qp) {
1481                 WARN("cannot allocate QP for drop queue");
1482                 goto error;
1483         }
1484         priv->flow_drop_queue = fdq;
1485         return 0;
1486 error:
1487         if (fdq->qp)
1488                 claim_zero(ibv_destroy_qp(fdq->qp));
1489         if (fdq->ind_table)
1490                 claim_zero(ibv_destroy_rwq_ind_table(fdq->ind_table));
1491         if (fdq->wq)
1492                 claim_zero(ibv_destroy_wq(fdq->wq));
1493         if (fdq->cq)
1494                 claim_zero(ibv_destroy_cq(fdq->cq));
1495         if (fdq)
1496                 rte_free(fdq);
1497         priv->flow_drop_queue = NULL;
1498         return -1;
1499 }
1500
1501 /**
1502  * Delete drop queue.
1503  *
1504  * @param priv
1505  *   Pointer to private structure.
1506  */
1507 void
1508 priv_flow_delete_drop_queue(struct priv *priv)
1509 {
1510         struct mlx5_hrxq_drop *fdq = priv->flow_drop_queue;
1511
1512         if (!fdq)
1513                 return;
1514         if (fdq->qp)
1515                 claim_zero(ibv_destroy_qp(fdq->qp));
1516         if (fdq->ind_table)
1517                 claim_zero(ibv_destroy_rwq_ind_table(fdq->ind_table));
1518         if (fdq->wq)
1519                 claim_zero(ibv_destroy_wq(fdq->wq));
1520         if (fdq->cq)
1521                 claim_zero(ibv_destroy_cq(fdq->cq));
1522         rte_free(fdq);
1523         priv->flow_drop_queue = NULL;
1524 }
1525
1526 /**
1527  * Remove all flows.
1528  *
1529  * @param priv
1530  *   Pointer to private structure.
1531  * @param list
1532  *   Pointer to a TAILQ flow list.
1533  */
1534 void
1535 priv_flow_stop(struct priv *priv, struct mlx5_flows *list)
1536 {
1537         struct rte_flow *flow;
1538
1539         TAILQ_FOREACH_REVERSE(flow, list, mlx5_flows, next) {
1540                 claim_zero(ibv_destroy_flow(flow->ibv_flow));
1541                 flow->ibv_flow = NULL;
1542                 mlx5_priv_hrxq_release(priv, flow->frxq.hrxq);
1543                 flow->frxq.hrxq = NULL;
1544                 if (flow->mark) {
1545                         unsigned int n;
1546                         struct mlx5_ind_table_ibv *ind_tbl =
1547                                 flow->frxq.hrxq->ind_table;
1548
1549                         for (n = 0; n < ind_tbl->queues_n; ++n)
1550                                 (*priv->rxqs)[ind_tbl->queues[n]]->mark = 0;
1551                 }
1552                 DEBUG("Flow %p removed", (void *)flow);
1553         }
1554 }
1555
1556 /**
1557  * Add all flows.
1558  *
1559  * @param priv
1560  *   Pointer to private structure.
1561  * @param list
1562  *   Pointer to a TAILQ flow list.
1563  *
1564  * @return
1565  *   0 on success, a errno value otherwise and rte_errno is set.
1566  */
1567 int
1568 priv_flow_start(struct priv *priv, struct mlx5_flows *list)
1569 {
1570         struct rte_flow *flow;
1571
1572         TAILQ_FOREACH(flow, list, next) {
1573                 if (flow->frxq.hrxq)
1574                         goto flow_create;
1575                 flow->frxq.hrxq =
1576                         mlx5_priv_hrxq_get(priv, rss_hash_default_key,
1577                                            rss_hash_default_key_len,
1578                                            flow->frxq.hash_fields,
1579                                            (*flow->queues),
1580                                            flow->queues_n);
1581                 if (flow->frxq.hrxq)
1582                         goto flow_create;
1583                 flow->frxq.hrxq =
1584                         mlx5_priv_hrxq_new(priv, rss_hash_default_key,
1585                                            rss_hash_default_key_len,
1586                                            flow->frxq.hash_fields,
1587                                            (*flow->queues),
1588                                            flow->queues_n);
1589                 if (!flow->frxq.hrxq) {
1590                         DEBUG("Flow %p cannot be applied",
1591                               (void *)flow);
1592                         rte_errno = EINVAL;
1593                         return rte_errno;
1594                 }
1595 flow_create:
1596                 flow->ibv_flow = ibv_create_flow(flow->frxq.hrxq->qp,
1597                                                  flow->ibv_attr);
1598                 if (!flow->ibv_flow) {
1599                         DEBUG("Flow %p cannot be applied", (void *)flow);
1600                         rte_errno = EINVAL;
1601                         return rte_errno;
1602                 }
1603                 DEBUG("Flow %p applied", (void *)flow);
1604                 if (flow->mark) {
1605                         unsigned int n;
1606
1607                         for (n = 0;
1608                              n < flow->frxq.hrxq->ind_table->queues_n;
1609                              ++n) {
1610                                 uint16_t idx =
1611                                         flow->frxq.hrxq->ind_table->queues[n];
1612                                 (*priv->rxqs)[idx]->mark = 1;
1613                         }
1614                 }
1615         }
1616         return 0;
1617 }
1618
1619 /**
1620  * Isolated mode.
1621  *
1622  * @see rte_flow_isolate()
1623  * @see rte_flow_ops
1624  */
1625 int
1626 mlx5_flow_isolate(struct rte_eth_dev *dev,
1627                   int enable,
1628                   struct rte_flow_error *error)
1629 {
1630         struct priv *priv = dev->data->dev_private;
1631
1632         priv_lock(priv);
1633         if (dev->data->dev_started) {
1634                 rte_flow_error_set(error, EBUSY,
1635                                    RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
1636                                    NULL,
1637                                    "port must be stopped first");
1638                 priv_unlock(priv);
1639                 return -rte_errno;
1640         }
1641         priv->isolated = !!enable;
1642         priv_unlock(priv);
1643         return 0;
1644 }
1645
1646 /**
1647  * Verify the flow list is empty
1648  *
1649  * @param priv
1650  *  Pointer to private structure.
1651  *
1652  * @return the number of flows not released.
1653  */
1654 int
1655 priv_flow_verify(struct priv *priv)
1656 {
1657         struct rte_flow *flow;
1658         int ret = 0;
1659
1660         TAILQ_FOREACH(flow, &priv->flows, next) {
1661                 DEBUG("%p: flow %p still referenced", (void *)priv,
1662                       (void *)flow);
1663                 ++ret;
1664         }
1665         return ret;
1666 }
1667
1668 /**
1669  * Enable a control flow configured from the control plane.
1670  *
1671  * @param dev
1672  *   Pointer to Ethernet device.
1673  * @param eth_spec
1674  *   An Ethernet flow spec to apply.
1675  * @param eth_mask
1676  *   An Ethernet flow mask to apply.
1677  * @param vlan_spec
1678  *   A VLAN flow spec to apply.
1679  * @param vlan_mask
1680  *   A VLAN flow mask to apply.
1681  *
1682  * @return
1683  *   0 on success.
1684  */
1685 int
1686 mlx5_ctrl_flow_vlan(struct rte_eth_dev *dev,
1687                     struct rte_flow_item_eth *eth_spec,
1688                     struct rte_flow_item_eth *eth_mask,
1689                     struct rte_flow_item_vlan *vlan_spec,
1690                     struct rte_flow_item_vlan *vlan_mask)
1691 {
1692         struct priv *priv = dev->data->dev_private;
1693         const struct rte_flow_attr attr = {
1694                 .ingress = 1,
1695                 .priority = MLX5_CTRL_FLOW_PRIORITY,
1696         };
1697         struct rte_flow_item items[] = {
1698                 {
1699                         .type = RTE_FLOW_ITEM_TYPE_ETH,
1700                         .spec = eth_spec,
1701                         .last = NULL,
1702                         .mask = eth_mask,
1703                 },
1704                 {
1705                         .type = (vlan_spec) ? RTE_FLOW_ITEM_TYPE_VLAN :
1706                                 RTE_FLOW_ITEM_TYPE_END,
1707                         .spec = vlan_spec,
1708                         .last = NULL,
1709                         .mask = vlan_mask,
1710                 },
1711                 {
1712                         .type = RTE_FLOW_ITEM_TYPE_END,
1713                 },
1714         };
1715         struct rte_flow_action actions[] = {
1716                 {
1717                         .type = RTE_FLOW_ACTION_TYPE_QUEUE,
1718                         .conf = &(struct rte_flow_action_queue){
1719                                 .index = 0,
1720                         },
1721                 },
1722                 {
1723                         .type = RTE_FLOW_ACTION_TYPE_END,
1724                 },
1725         };
1726         struct rte_flow *flow;
1727         struct rte_flow_error error;
1728
1729         flow = priv_flow_create(priv, &priv->ctrl_flows, &attr, items, actions,
1730                                 &error);
1731         if (!flow)
1732                 return rte_errno;
1733         return 0;
1734 }
1735
1736 /**
1737  * Enable a flow control configured from the control plane.
1738  *
1739  * @param dev
1740  *   Pointer to Ethernet device.
1741  * @param eth_spec
1742  *   An Ethernet flow spec to apply.
1743  * @param eth_mask
1744  *   An Ethernet flow mask to apply.
1745  *
1746  * @return
1747  *   0 on success.
1748  */
1749 int
1750 mlx5_ctrl_flow(struct rte_eth_dev *dev,
1751                struct rte_flow_item_eth *eth_spec,
1752                struct rte_flow_item_eth *eth_mask)
1753 {
1754         return mlx5_ctrl_flow_vlan(dev, eth_spec, eth_mask, NULL, NULL);
1755 }