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 /* Number of Work Queue necessary for the DROP queue. */
  56 #define MLX5_DROP_WQ_N 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 struct rte_flow {
  94         LIST_ENTRY(rte_flow) next; /**< Pointer to the next flow structure. */
  95         struct ibv_exp_flow_attr *ibv_attr; /**< Pointer to Verbs attributes. */
  96         struct ibv_exp_rwq_ind_table *ind_table; /**< Indirection table. */
  97         struct ibv_qp *qp; /**< Verbs queue pair. */
  98         struct ibv_exp_flow *ibv_flow; /**< Verbs flow. */
  99         struct ibv_exp_wq *wq; /**< Verbs work queue. */
 100         struct ibv_cq *cq; /**< Verbs completion queue. */
 101         struct rxq *(*rxqs)[]; /**< Pointer to the queues array. */
 102         uint16_t rxqs_n; /**< Number of queues in this flow, 0 if drop queue. */
 103         uint32_t mark:1; /**< Set if the flow is marked. */
 104         uint32_t drop:1; /**< Drop queue. */
 105         uint64_t hash_fields; /**< Fields that participate in the hash. */
 106 };
 107
 108 /** Static initializer for items. */
 109 #define ITEMS(...) \
 110         (const enum rte_flow_item_type []){ \
 111                 __VA_ARGS__, RTE_FLOW_ITEM_TYPE_END, \
 112         }
 113
 114 /** Structure to generate a simple graph of layers supported by the NIC. */
 115 struct mlx5_flow_items {
 116         /** List of possible actions for these items. */
 117         const enum rte_flow_action_type *const actions;
 118         /** Bit-masks corresponding to the possibilities for the item. */
 119         const void *mask;
 120         /**
 121          * Default bit-masks to use when item->mask is not provided. When
 122          * \default_mask is also NULL, the full supported bit-mask (\mask) is
 123          * used instead.
 124          */
 125         const void *default_mask;
 126         /** Bit-masks size in bytes. */
 127         const unsigned int mask_sz;
 128         /**
 129          * Conversion function from rte_flow to NIC specific flow.
 130          *
 131          * @param item
 132          *   rte_flow item to convert.
 133          * @param default_mask
 134          *   Default bit-masks to use when item->mask is not provided.
 135          * @param data
 136          *   Internal structure to store the conversion.
 137          *
 138          * @return
 139          *   0 on success, negative value otherwise.
 140          */
 141         int (*convert)(const struct rte_flow_item *item,
 142                        const void *default_mask,
 143                        void *data);
 144         /** Size in bytes of the destination structure. */
 145         const unsigned int dst_sz;
 146         /** List of possible following items.  */
 147         const enum rte_flow_item_type *const items;
 148 };
 149
 150 /** Valid action for this PMD. */
 151 static const enum rte_flow_action_type valid_actions[] = {
 152         RTE_FLOW_ACTION_TYPE_DROP,
 153         RTE_FLOW_ACTION_TYPE_QUEUE,
 154         RTE_FLOW_ACTION_TYPE_MARK,
 155         RTE_FLOW_ACTION_TYPE_FLAG,
 156         RTE_FLOW_ACTION_TYPE_END,
 157 };
 158
 159 /** Graph of supported items and associated actions. */
 160 static const struct mlx5_flow_items mlx5_flow_items[] = {
 161         [RTE_FLOW_ITEM_TYPE_END] = {
 162                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_ETH,
 163                                RTE_FLOW_ITEM_TYPE_VXLAN),
 164         },
 165         [RTE_FLOW_ITEM_TYPE_ETH] = {
 166                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_VLAN,
 167                                RTE_FLOW_ITEM_TYPE_IPV4,
 168                                RTE_FLOW_ITEM_TYPE_IPV6),
 169                 .actions = valid_actions,
 170                 .mask = &(const struct rte_flow_item_eth){
 171                         .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 172                         .src.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 173                         .type = -1,
 174                 },
 175                 .default_mask = &rte_flow_item_eth_mask,
 176                 .mask_sz = sizeof(struct rte_flow_item_eth),
 177                 .convert = mlx5_flow_create_eth,
 178                 .dst_sz = sizeof(struct ibv_exp_flow_spec_eth),
 179         },
 180         [RTE_FLOW_ITEM_TYPE_VLAN] = {
 181                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_IPV4,
 182                                RTE_FLOW_ITEM_TYPE_IPV6),
 183                 .actions = valid_actions,
 184                 .mask = &(const struct rte_flow_item_vlan){
 185                         .tci = -1,
 186                 },
 187                 .default_mask = &rte_flow_item_vlan_mask,
 188                 .mask_sz = sizeof(struct rte_flow_item_vlan),
 189                 .convert = mlx5_flow_create_vlan,
 190                 .dst_sz = 0,
 191         },
 192         [RTE_FLOW_ITEM_TYPE_IPV4] = {
 193                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP,
 194                                RTE_FLOW_ITEM_TYPE_TCP),
 195                 .actions = valid_actions,
 196                 .mask = &(const struct rte_flow_item_ipv4){
 197                         .hdr = {
 198                                 .src_addr = -1,
 199                                 .dst_addr = -1,
 200                                 .type_of_service = -1,
 201                                 .next_proto_id = -1,
 202                         },
 203                 },
 204                 .default_mask = &rte_flow_item_ipv4_mask,
 205                 .mask_sz = sizeof(struct rte_flow_item_ipv4),
 206                 .convert = mlx5_flow_create_ipv4,
 207                 .dst_sz = sizeof(struct ibv_exp_flow_spec_ipv4_ext),
 208         },
 209         [RTE_FLOW_ITEM_TYPE_IPV6] = {
 210                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP,
 211                                RTE_FLOW_ITEM_TYPE_TCP),
 212                 .actions = valid_actions,
 213                 .mask = &(const struct rte_flow_item_ipv6){
 214                         .hdr = {
 215                                 .src_addr = {
 216                                         0xff, 0xff, 0xff, 0xff,
 217                                         0xff, 0xff, 0xff, 0xff,
 218                                         0xff, 0xff, 0xff, 0xff,
 219                                         0xff, 0xff, 0xff, 0xff,
 220                                 },
 221                                 .dst_addr = {
 222                                         0xff, 0xff, 0xff, 0xff,
 223                                         0xff, 0xff, 0xff, 0xff,
 224                                         0xff, 0xff, 0xff, 0xff,
 225                                         0xff, 0xff, 0xff, 0xff,
 226                                 },
 227                                 .vtc_flow = -1,
 228                                 .proto = -1,
 229                                 .hop_limits = -1,
 230                         },
 231                 },
 232                 .default_mask = &rte_flow_item_ipv6_mask,
 233                 .mask_sz = sizeof(struct rte_flow_item_ipv6),
 234                 .convert = mlx5_flow_create_ipv6,
 235                 .dst_sz = sizeof(struct ibv_exp_flow_spec_ipv6_ext),
 236         },
 237         [RTE_FLOW_ITEM_TYPE_UDP] = {
 238                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_VXLAN),
 239                 .actions = valid_actions,
 240                 .mask = &(const struct rte_flow_item_udp){
 241                         .hdr = {
 242                                 .src_port = -1,
 243                                 .dst_port = -1,
 244                         },
 245                 },
 246                 .default_mask = &rte_flow_item_udp_mask,
 247                 .mask_sz = sizeof(struct rte_flow_item_udp),
 248                 .convert = mlx5_flow_create_udp,
 249                 .dst_sz = sizeof(struct ibv_exp_flow_spec_tcp_udp),
 250         },
 251         [RTE_FLOW_ITEM_TYPE_TCP] = {
 252                 .actions = valid_actions,
 253                 .mask = &(const struct rte_flow_item_tcp){
 254                         .hdr = {
 255                                 .src_port = -1,
 256                                 .dst_port = -1,
 257                         },
 258                 },
 259                 .default_mask = &rte_flow_item_tcp_mask,
 260                 .mask_sz = sizeof(struct rte_flow_item_tcp),
 261                 .convert = mlx5_flow_create_tcp,
 262                 .dst_sz = sizeof(struct ibv_exp_flow_spec_tcp_udp),
 263         },
 264         [RTE_FLOW_ITEM_TYPE_VXLAN] = {
 265                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_ETH),
 266                 .actions = valid_actions,
 267                 .mask = &(const struct rte_flow_item_vxlan){
 268                         .vni = "\xff\xff\xff",
 269                 },
 270                 .default_mask = &rte_flow_item_vxlan_mask,
 271                 .mask_sz = sizeof(struct rte_flow_item_vxlan),
 272                 .convert = mlx5_flow_create_vxlan,
 273                 .dst_sz = sizeof(struct ibv_exp_flow_spec_tunnel),
 274         },
 275 };
 276
 277 /** Structure to pass to the conversion function. */
 278 struct mlx5_flow {
 279         struct ibv_exp_flow_attr *ibv_attr; /**< Verbs attribute. */
 280         unsigned int offset; /**< Offset in bytes in the ibv_attr buffer. */
 281         uint32_t inner; /**< Set once VXLAN is encountered. */
 282         uint64_t hash_fields; /**< Fields that participate in the hash. */
 283 };
 284
 285 /** Structure for Drop queue. */
 286 struct rte_flow_drop {
 287         struct ibv_exp_rwq_ind_table *ind_table; /**< Indirection table. */
 288         struct ibv_qp *qp; /**< Verbs queue pair. */
 289         struct ibv_exp_wq *wqs[MLX5_DROP_WQ_N]; /**< Verbs work queue. */
 290         struct ibv_cq *cq; /**< Verbs completion queue. */
 291 };
 292
 293 struct mlx5_flow_action {
 294         uint32_t queue:1; /**< Target is a receive queue. */
 295         uint32_t drop:1; /**< Target is a drop queue. */
 296         uint32_t mark:1; /**< Mark is present in the flow. */
 297         uint32_t mark_id; /**< Mark identifier. */
 298         uint16_t queues[RTE_MAX_QUEUES_PER_PORT]; /**< Queues indexes to use. */
 299         uint16_t queues_n; /**< Number of entries in queue[]. */
 300 };
 301
 302 /**
 303  * Check support for a given item.
 304  *
 305  * @param item[in]
 306  *   Item specification.
 307  * @param mask[in]
 308  *   Bit-masks covering supported fields to compare with spec, last and mask in
 309  *   \item.
 310  * @param size
 311  *   Bit-Mask size in bytes.
 312  *
 313  * @return
 314  *   0 on success.
 315  */
 316 static int
 317 mlx5_flow_item_validate(const struct rte_flow_item *item,
 318                         const uint8_t *mask, unsigned int size)
 319 {
 320         int ret = 0;
 321
 322         if (!item->spec && (item->mask || item->last))
 323                 return -1;
 324         if (item->spec && !item->mask) {
 325                 unsigned int i;
 326                 const uint8_t *spec = item->spec;
 327
 328                 for (i = 0; i < size; ++i)
 329                         if ((spec[i] | mask[i]) != mask[i])
 330                                 return -1;
 331         }
 332         if (item->last && !item->mask) {
 333                 unsigned int i;
 334                 const uint8_t *spec = item->last;
 335
 336                 for (i = 0; i < size; ++i)
 337                         if ((spec[i] | mask[i]) != mask[i])
 338                                 return -1;
 339         }
 340         if (item->mask) {
 341                 unsigned int i;
 342                 const uint8_t *spec = item->mask;
 343
 344                 for (i = 0; i < size; ++i)
 345                         if ((spec[i] | mask[i]) != mask[i])
 346                                 return -1;
 347         }
 348         if (item->spec && item->last) {
 349                 uint8_t spec[size];
 350                 uint8_t last[size];
 351                 const uint8_t *apply = mask;
 352                 unsigned int i;
 353
 354                 if (item->mask)
 355                         apply = item->mask;
 356                 for (i = 0; i < size; ++i) {
 357                         spec[i] = ((const uint8_t *)item->spec)[i] & apply[i];
 358                         last[i] = ((const uint8_t *)item->last)[i] & apply[i];
 359                 }
 360                 ret = memcmp(spec, last, size);
 361         }
 362         return ret;
 363 }
 364
 365 /**
 366  * Validate a flow supported by the NIC.
 367  *
 368  * @param priv
 369  *   Pointer to private structure.
 370  * @param[in] attr
 371  *   Flow rule attributes.
 372  * @param[in] pattern
 373  *   Pattern specification (list terminated by the END pattern item).
 374  * @param[in] actions
 375  *   Associated actions (list terminated by the END action).
 376  * @param[out] error
 377  *   Perform verbose error reporting if not NULL.
 378  * @param[in, out] flow
 379  *   Flow structure to update.
 380  *
 381  * @return
 382  *   0 on success, a negative errno value otherwise and rte_errno is set.
 383  */
 384 static int
 385 priv_flow_validate(struct priv *priv,
 386                    const struct rte_flow_attr *attr,
 387                    const struct rte_flow_item items[],
 388                    const struct rte_flow_action actions[],
 389                    struct rte_flow_error *error,
 390                    struct mlx5_flow *flow)
 391 {
 392         const struct mlx5_flow_items *cur_item = mlx5_flow_items;
 393         struct mlx5_flow_action action = {
 394                 .queue = 0,
 395                 .drop = 0,
 396                 .mark = 0,
 397         };
 398
 399         (void)priv;
 400         if (attr->group) {
 401                 rte_flow_error_set(error, ENOTSUP,
 402                                    RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
 403                                    NULL,
 404                                    "groups are not supported");
 405                 return -rte_errno;
 406         }
 407         if (attr->priority) {
 408                 rte_flow_error_set(error, ENOTSUP,
 409                                    RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
 410                                    NULL,
 411                                    "priorities are not supported");
 412                 return -rte_errno;
 413         }
 414         if (attr->egress) {
 415                 rte_flow_error_set(error, ENOTSUP,
 416                                    RTE_FLOW_ERROR_TYPE_ATTR_EGRESS,
 417                                    NULL,
 418                                    "egress is not supported");
 419                 return -rte_errno;
 420         }
 421         if (!attr->ingress) {
 422                 rte_flow_error_set(error, ENOTSUP,
 423                                    RTE_FLOW_ERROR_TYPE_ATTR_INGRESS,
 424                                    NULL,
 425                                    "only ingress is supported");
 426                 return -rte_errno;
 427         }
 428         for (; items->type != RTE_FLOW_ITEM_TYPE_END; ++items) {
 429                 const struct mlx5_flow_items *token = NULL;
 430                 unsigned int i;
 431                 int err;
 432
 433                 if (items->type == RTE_FLOW_ITEM_TYPE_VOID)
 434                         continue;
 435                 for (i = 0;
 436                      cur_item->items &&
 437                      cur_item->items[i] != RTE_FLOW_ITEM_TYPE_END;
 438                      ++i) {
 439                         if (cur_item->items[i] == items->type) {
 440                                 token = &mlx5_flow_items[items->type];
 441                                 break;
 442                         }
 443                 }
 444                 if (!token)
 445                         goto exit_item_not_supported;
 446                 cur_item = token;
 447                 err = mlx5_flow_item_validate(items,
 448                                               (const uint8_t *)cur_item->mask,
 449                                               cur_item->mask_sz);
 450                 if (err)
 451                         goto exit_item_not_supported;
 452                 if (flow->ibv_attr && cur_item->convert) {
 453                         err = cur_item->convert(items,
 454                                                 (cur_item->default_mask ?
 455                                                  cur_item->default_mask :
 456                                                  cur_item->mask),
 457                                                 flow);
 458                         if (err)
 459                                 goto exit_item_not_supported;
 460                 }
 461                 flow->offset += cur_item->dst_sz;
 462         }
 463         for (; actions->type != RTE_FLOW_ACTION_TYPE_END; ++actions) {
 464                 if (actions->type == RTE_FLOW_ACTION_TYPE_VOID) {
 465                         continue;
 466                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_DROP) {
 467                         action.drop = 1;
 468                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_QUEUE) {
 469                         const struct rte_flow_action_queue *queue =
 470                                 (const struct rte_flow_action_queue *)
 471                                 actions->conf;
 472                         uint16_t n;
 473                         uint16_t found = 0;
 474
 475                         if (!queue || (queue->index > (priv->rxqs_n - 1)))
 476                                 goto exit_action_not_supported;
 477                         for (n = 0; n < action.queues_n; ++n) {
 478                                 if (action.queues[n] == queue->index) {
 479                                         found = 1;
 480                                         break;
 481                                 }
 482                         }
 483                         if (action.queues_n && !found) {
 484                                 rte_flow_error_set(error, ENOTSUP,
 485                                            RTE_FLOW_ERROR_TYPE_ACTION,
 486                                            actions,
 487                                            "queue action not in RSS queues");
 488                                 return -rte_errno;
 489                         }
 490                         action.queue = 1;
 491                         action.queues_n = 1;
 492                         action.queues[0] = queue->index;
 493                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_RSS) {
 494                         const struct rte_flow_action_rss *rss =
 495                                 (const struct rte_flow_action_rss *)
 496                                 actions->conf;
 497                         uint16_t n;
 498
 499                         if (action.queues_n == 1) {
 500                                 uint16_t found = 0;
 501
 502                                 assert(action.queues_n);
 503                                 for (n = 0; n < rss->num; ++n) {
 504                                         if (action.queues[0] == rss->queue[n]) {
 505                                                 found = 1;
 506                                                 break;
 507                                         }
 508                                 }
 509                                 if (!found) {
 510                                         rte_flow_error_set(error, ENOTSUP,
 511                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 512                                                    actions,
 513                                                    "queue action not in RSS"
 514                                                    " queues");
 515                                         return -rte_errno;
 516                                 }
 517                         }
 518                         action.queue = 1;
 519                         for (n = 0; n < rss->num; ++n)
 520                                 action.queues[n] = rss->queue[n];
 521                         action.queues_n = rss->num;
 522                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_MARK) {
 523                         const struct rte_flow_action_mark *mark =
 524                                 (const struct rte_flow_action_mark *)
 525                                 actions->conf;
 526
 527                         if (!mark) {
 528                                 rte_flow_error_set(error, EINVAL,
 529                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 530                                                    actions,
 531                                                    "mark must be defined");
 532                                 return -rte_errno;
 533                         } else if (mark->id >= MLX5_FLOW_MARK_MAX) {
 534                                 rte_flow_error_set(error, ENOTSUP,
 535                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 536                                                    actions,
 537                                                    "mark must be between 0"
 538                                                    " and 16777199");
 539                                 return -rte_errno;
 540                         }
 541                         action.mark = 1;
 542                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_FLAG) {
 543                         action.mark = 1;
 544                 } else {
 545                         goto exit_action_not_supported;
 546                 }
 547         }
 548         if (action.mark && !flow->ibv_attr && !action.drop)
 549                 flow->offset += sizeof(struct ibv_exp_flow_spec_action_tag);
 550         if (!action.queue && !action.drop) {
 551                 rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_HANDLE,
 552                                    NULL, "no valid action");
 553                 return -rte_errno;
 554         }
 555         return 0;
 556 exit_item_not_supported:
 557         rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 558                            items, "item not supported");
 559         return -rte_errno;
 560 exit_action_not_supported:
 561         rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
 562                            actions, "action not supported");
 563         return -rte_errno;
 564 }
 565
 566 /**
 567  * Validate a flow supported by the NIC.
 568  *
 569  * @see rte_flow_validate()
 570  * @see rte_flow_ops
 571  */
 572 int
 573 mlx5_flow_validate(struct rte_eth_dev *dev,
 574                    const struct rte_flow_attr *attr,
 575                    const struct rte_flow_item items[],
 576                    const struct rte_flow_action actions[],
 577                    struct rte_flow_error *error)
 578 {
 579         struct priv *priv = dev->data->dev_private;
 580         int ret;
 581         struct mlx5_flow flow = { .offset = sizeof(struct ibv_exp_flow_attr) };
 582
 583         priv_lock(priv);
 584         ret = priv_flow_validate(priv, attr, items, actions, error, &flow);
 585         priv_unlock(priv);
 586         return ret;
 587 }
 588
 589 /**
 590  * Convert Ethernet item to Verbs specification.
 591  *
 592  * @param item[in]
 593  *   Item specification.
 594  * @param default_mask[in]
 595  *   Default bit-masks to use when item->mask is not provided.
 596  * @param data[in, out]
 597  *   User structure.
 598  */
 599 static int
 600 mlx5_flow_create_eth(const struct rte_flow_item *item,
 601                      const void *default_mask,
 602                      void *data)
 603 {
 604         const struct rte_flow_item_eth *spec = item->spec;
 605         const struct rte_flow_item_eth *mask = item->mask;
 606         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 607         struct ibv_exp_flow_spec_eth *eth;
 608         const unsigned int eth_size = sizeof(struct ibv_exp_flow_spec_eth);
 609         unsigned int i;
 610
 611         ++flow->ibv_attr->num_of_specs;
 612         flow->ibv_attr->priority = 2;
 613         flow->hash_fields = 0;
 614         eth = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 615         *eth = (struct ibv_exp_flow_spec_eth) {
 616                 .type = flow->inner | IBV_EXP_FLOW_SPEC_ETH,
 617                 .size = eth_size,
 618         };
 619         if (!spec)
 620                 return 0;
 621         if (!mask)
 622                 mask = default_mask;
 623         memcpy(eth->val.dst_mac, spec->dst.addr_bytes, ETHER_ADDR_LEN);
 624         memcpy(eth->val.src_mac, spec->src.addr_bytes, ETHER_ADDR_LEN);
 625         eth->val.ether_type = spec->type;
 626         memcpy(eth->mask.dst_mac, mask->dst.addr_bytes, ETHER_ADDR_LEN);
 627         memcpy(eth->mask.src_mac, mask->src.addr_bytes, ETHER_ADDR_LEN);
 628         eth->mask.ether_type = mask->type;
 629         /* Remove unwanted bits from values. */
 630         for (i = 0; i < ETHER_ADDR_LEN; ++i) {
 631                 eth->val.dst_mac[i] &= eth->mask.dst_mac[i];
 632                 eth->val.src_mac[i] &= eth->mask.src_mac[i];
 633         }
 634         eth->val.ether_type &= eth->mask.ether_type;
 635         return 0;
 636 }
 637
 638 /**
 639  * Convert VLAN item to Verbs specification.
 640  *
 641  * @param item[in]
 642  *   Item specification.
 643  * @param default_mask[in]
 644  *   Default bit-masks to use when item->mask is not provided.
 645  * @param data[in, out]
 646  *   User structure.
 647  */
 648 static int
 649 mlx5_flow_create_vlan(const struct rte_flow_item *item,
 650                       const void *default_mask,
 651                       void *data)
 652 {
 653         const struct rte_flow_item_vlan *spec = item->spec;
 654         const struct rte_flow_item_vlan *mask = item->mask;
 655         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 656         struct ibv_exp_flow_spec_eth *eth;
 657         const unsigned int eth_size = sizeof(struct ibv_exp_flow_spec_eth);
 658
 659         eth = (void *)((uintptr_t)flow->ibv_attr + flow->offset - eth_size);
 660         if (!spec)
 661                 return 0;
 662         if (!mask)
 663                 mask = default_mask;
 664         eth->val.vlan_tag = spec->tci;
 665         eth->mask.vlan_tag = mask->tci;
 666         eth->val.vlan_tag &= eth->mask.vlan_tag;
 667         return 0;
 668 }
 669
 670 /**
 671  * Convert IPv4 item to Verbs specification.
 672  *
 673  * @param item[in]
 674  *   Item specification.
 675  * @param default_mask[in]
 676  *   Default bit-masks to use when item->mask is not provided.
 677  * @param data[in, out]
 678  *   User structure.
 679  */
 680 static int
 681 mlx5_flow_create_ipv4(const struct rte_flow_item *item,
 682                       const void *default_mask,
 683                       void *data)
 684 {
 685         const struct rte_flow_item_ipv4 *spec = item->spec;
 686         const struct rte_flow_item_ipv4 *mask = item->mask;
 687         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 688         struct ibv_exp_flow_spec_ipv4_ext *ipv4;
 689         unsigned int ipv4_size = sizeof(struct ibv_exp_flow_spec_ipv4_ext);
 690
 691         ++flow->ibv_attr->num_of_specs;
 692         flow->ibv_attr->priority = 1;
 693         flow->hash_fields = (IBV_EXP_RX_HASH_SRC_IPV4 |
 694                              IBV_EXP_RX_HASH_DST_IPV4);
 695         ipv4 = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 696         *ipv4 = (struct ibv_exp_flow_spec_ipv4_ext) {
 697                 .type = flow->inner | IBV_EXP_FLOW_SPEC_IPV4_EXT,
 698                 .size = ipv4_size,
 699         };
 700         if (!spec)
 701                 return 0;
 702         if (!mask)
 703                 mask = default_mask;
 704         ipv4->val = (struct ibv_exp_flow_ipv4_ext_filter){
 705                 .src_ip = spec->hdr.src_addr,
 706                 .dst_ip = spec->hdr.dst_addr,
 707                 .proto = spec->hdr.next_proto_id,
 708                 .tos = spec->hdr.type_of_service,
 709         };
 710         ipv4->mask = (struct ibv_exp_flow_ipv4_ext_filter){
 711                 .src_ip = mask->hdr.src_addr,
 712                 .dst_ip = mask->hdr.dst_addr,
 713                 .proto = mask->hdr.next_proto_id,
 714                 .tos = mask->hdr.type_of_service,
 715         };
 716         /* Remove unwanted bits from values. */
 717         ipv4->val.src_ip &= ipv4->mask.src_ip;
 718         ipv4->val.dst_ip &= ipv4->mask.dst_ip;
 719         ipv4->val.proto &= ipv4->mask.proto;
 720         ipv4->val.tos &= ipv4->mask.tos;
 721         return 0;
 722 }
 723
 724 /**
 725  * Convert IPv6 item to Verbs specification.
 726  *
 727  * @param item[in]
 728  *   Item specification.
 729  * @param default_mask[in]
 730  *   Default bit-masks to use when item->mask is not provided.
 731  * @param data[in, out]
 732  *   User structure.
 733  */
 734 static int
 735 mlx5_flow_create_ipv6(const struct rte_flow_item *item,
 736                       const void *default_mask,
 737                       void *data)
 738 {
 739         const struct rte_flow_item_ipv6 *spec = item->spec;
 740         const struct rte_flow_item_ipv6 *mask = item->mask;
 741         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 742         struct ibv_exp_flow_spec_ipv6_ext *ipv6;
 743         unsigned int ipv6_size = sizeof(struct ibv_exp_flow_spec_ipv6_ext);
 744
 745         ++flow->ibv_attr->num_of_specs;
 746         flow->ibv_attr->priority = 1;
 747         flow->hash_fields = (IBV_EXP_RX_HASH_SRC_IPV6 |
 748                              IBV_EXP_RX_HASH_DST_IPV6);
 749         ipv6 = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 750         *ipv6 = (struct ibv_exp_flow_spec_ipv6_ext) {
 751                 .type = flow->inner | IBV_EXP_FLOW_SPEC_IPV6_EXT,
 752                 .size = ipv6_size,
 753         };
 754         if (!spec)
 755                 return 0;
 756         if (!mask)
 757                 mask = default_mask;
 758         memcpy(ipv6->val.src_ip, spec->hdr.src_addr,
 759                RTE_DIM(ipv6->val.src_ip));
 760         memcpy(ipv6->val.dst_ip, spec->hdr.dst_addr,
 761                RTE_DIM(ipv6->val.dst_ip));
 762         memcpy(ipv6->mask.src_ip, mask->hdr.src_addr,
 763                RTE_DIM(ipv6->mask.src_ip));
 764         memcpy(ipv6->mask.dst_ip, mask->hdr.dst_addr,
 765                RTE_DIM(ipv6->mask.dst_ip));
 766         ipv6->mask.flow_label = mask->hdr.vtc_flow;
 767         ipv6->mask.next_hdr = mask->hdr.proto;
 768         ipv6->mask.hop_limit = mask->hdr.hop_limits;
 769         ipv6->val.flow_label &= ipv6->mask.flow_label;
 770         ipv6->val.next_hdr &= ipv6->mask.next_hdr;
 771         ipv6->val.hop_limit &= ipv6->mask.hop_limit;
 772         return 0;
 773 }
 774
 775 /**
 776  * Convert UDP item to Verbs specification.
 777  *
 778  * @param item[in]
 779  *   Item specification.
 780  * @param default_mask[in]
 781  *   Default bit-masks to use when item->mask is not provided.
 782  * @param data[in, out]
 783  *   User structure.
 784  */
 785 static int
 786 mlx5_flow_create_udp(const struct rte_flow_item *item,
 787                      const void *default_mask,
 788                      void *data)
 789 {
 790         const struct rte_flow_item_udp *spec = item->spec;
 791         const struct rte_flow_item_udp *mask = item->mask;
 792         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 793         struct ibv_exp_flow_spec_tcp_udp *udp;
 794         unsigned int udp_size = sizeof(struct ibv_exp_flow_spec_tcp_udp);
 795
 796         ++flow->ibv_attr->num_of_specs;
 797         flow->ibv_attr->priority = 0;
 798         flow->hash_fields |= (IBV_EXP_RX_HASH_SRC_PORT_UDP |
 799                               IBV_EXP_RX_HASH_DST_PORT_UDP);
 800         udp = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 801         *udp = (struct ibv_exp_flow_spec_tcp_udp) {
 802                 .type = flow->inner | IBV_EXP_FLOW_SPEC_UDP,
 803                 .size = udp_size,
 804         };
 805         if (!spec)
 806                 return 0;
 807         if (!mask)
 808                 mask = default_mask;
 809         udp->val.dst_port = spec->hdr.dst_port;
 810         udp->val.src_port = spec->hdr.src_port;
 811         udp->mask.dst_port = mask->hdr.dst_port;
 812         udp->mask.src_port = mask->hdr.src_port;
 813         /* Remove unwanted bits from values. */
 814         udp->val.src_port &= udp->mask.src_port;
 815         udp->val.dst_port &= udp->mask.dst_port;
 816         return 0;
 817 }
 818
 819 /**
 820  * Convert TCP item to Verbs specification.
 821  *
 822  * @param item[in]
 823  *   Item specification.
 824  * @param default_mask[in]
 825  *   Default bit-masks to use when item->mask is not provided.
 826  * @param data[in, out]
 827  *   User structure.
 828  */
 829 static int
 830 mlx5_flow_create_tcp(const struct rte_flow_item *item,
 831                      const void *default_mask,
 832                      void *data)
 833 {
 834         const struct rte_flow_item_tcp *spec = item->spec;
 835         const struct rte_flow_item_tcp *mask = item->mask;
 836         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 837         struct ibv_exp_flow_spec_tcp_udp *tcp;
 838         unsigned int tcp_size = sizeof(struct ibv_exp_flow_spec_tcp_udp);
 839
 840         ++flow->ibv_attr->num_of_specs;
 841         flow->ibv_attr->priority = 0;
 842         flow->hash_fields |= (IBV_EXP_RX_HASH_SRC_PORT_TCP |
 843                               IBV_EXP_RX_HASH_DST_PORT_TCP);
 844         tcp = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 845         *tcp = (struct ibv_exp_flow_spec_tcp_udp) {
 846                 .type = flow->inner | IBV_EXP_FLOW_SPEC_TCP,
 847                 .size = tcp_size,
 848         };
 849         if (!spec)
 850                 return 0;
 851         if (!mask)
 852                 mask = default_mask;
 853         tcp->val.dst_port = spec->hdr.dst_port;
 854         tcp->val.src_port = spec->hdr.src_port;
 855         tcp->mask.dst_port = mask->hdr.dst_port;
 856         tcp->mask.src_port = mask->hdr.src_port;
 857         /* Remove unwanted bits from values. */
 858         tcp->val.src_port &= tcp->mask.src_port;
 859         tcp->val.dst_port &= tcp->mask.dst_port;
 860         return 0;
 861 }
 862
 863 /**
 864  * Convert VXLAN item to Verbs specification.
 865  *
 866  * @param item[in]
 867  *   Item specification.
 868  * @param default_mask[in]
 869  *   Default bit-masks to use when item->mask is not provided.
 870  * @param data[in, out]
 871  *   User structure.
 872  */
 873 static int
 874 mlx5_flow_create_vxlan(const struct rte_flow_item *item,
 875                        const void *default_mask,
 876                        void *data)
 877 {
 878         const struct rte_flow_item_vxlan *spec = item->spec;
 879         const struct rte_flow_item_vxlan *mask = item->mask;
 880         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 881         struct ibv_exp_flow_spec_tunnel *vxlan;
 882         unsigned int size = sizeof(struct ibv_exp_flow_spec_tunnel);
 883         union vni {
 884                 uint32_t vlan_id;
 885                 uint8_t vni[4];
 886         } id;
 887
 888         ++flow->ibv_attr->num_of_specs;
 889         flow->ibv_attr->priority = 0;
 890         id.vni[0] = 0;
 891         vxlan = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 892         *vxlan = (struct ibv_exp_flow_spec_tunnel) {
 893                 .type = flow->inner | IBV_EXP_FLOW_SPEC_VXLAN_TUNNEL,
 894                 .size = size,
 895         };
 896         flow->inner = IBV_EXP_FLOW_SPEC_INNER;
 897         if (!spec)
 898                 return 0;
 899         if (!mask)
 900                 mask = default_mask;
 901         memcpy(&id.vni[1], spec->vni, 3);
 902         vxlan->val.tunnel_id = id.vlan_id;
 903         memcpy(&id.vni[1], mask->vni, 3);
 904         vxlan->mask.tunnel_id = id.vlan_id;
 905         /* Remove unwanted bits from values. */
 906         vxlan->val.tunnel_id &= vxlan->mask.tunnel_id;
 907         return 0;
 908 }
 909
 910 /**
 911  * Convert mark/flag action to Verbs specification.
 912  *
 913  * @param flow
 914  *   Pointer to MLX5 flow structure.
 915  * @param mark_id
 916  *   Mark identifier.
 917  */
 918 static int
 919 mlx5_flow_create_flag_mark(struct mlx5_flow *flow, uint32_t mark_id)
 920 {
 921         struct ibv_exp_flow_spec_action_tag *tag;
 922         unsigned int size = sizeof(struct ibv_exp_flow_spec_action_tag);
 923
 924         tag = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 925         *tag = (struct ibv_exp_flow_spec_action_tag){
 926                 .type = IBV_EXP_FLOW_SPEC_ACTION_TAG,
 927                 .size = size,
 928                 .tag_id = mlx5_flow_mark_set(mark_id),
 929         };
 930         ++flow->ibv_attr->num_of_specs;
 931         return 0;
 932 }
 933
 934 /**
 935  * Complete flow rule creation with a drop queue.
 936  *
 937  * @param priv
 938  *   Pointer to private structure.
 939  * @param flow
 940  *   MLX5 flow attributes (filled by mlx5_flow_validate()).
 941  * @param[out] error
 942  *   Perform verbose error reporting if not NULL.
 943  *
 944  * @return
 945  *   A flow if the rule could be created.
 946  */
 947 static struct rte_flow *
 948 priv_flow_create_action_queue_drop(struct priv *priv,
 949                                    struct mlx5_flow *flow,
 950                                    struct rte_flow_error *error)
 951 {
 952         struct rte_flow *rte_flow;
 953
 954         assert(priv->pd);
 955         assert(priv->ctx);
 956         rte_flow = rte_calloc(__func__, 1, sizeof(*rte_flow), 0);
 957         if (!rte_flow) {
 958                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
 959                                    NULL, "cannot allocate flow memory");
 960                 return NULL;
 961         }
 962         rte_flow->drop = 1;
 963         rte_flow->ibv_attr = flow->ibv_attr;
 964         rte_flow->qp = priv->flow_drop_queue->qp;
 965         if (!priv->started)
 966                 return rte_flow;
 967         rte_flow->ibv_flow = ibv_exp_create_flow(rte_flow->qp,
 968                                                  rte_flow->ibv_attr);
 969         if (!rte_flow->ibv_flow) {
 970                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
 971                                    NULL, "flow rule creation failure");
 972                 goto error;
 973         }
 974         return rte_flow;
 975 error:
 976         assert(rte_flow);
 977         rte_free(rte_flow);
 978         return NULL;
 979 }
 980
 981 /**
 982  * Complete flow rule creation.
 983  *
 984  * @param priv
 985  *   Pointer to private structure.
 986  * @param flow
 987  *   MLX5 flow attributes (filled by mlx5_flow_validate()).
 988  * @param action
 989  *   Target action structure.
 990  * @param[out] error
 991  *   Perform verbose error reporting if not NULL.
 992  *
 993  * @return
 994  *   A flow if the rule could be created.
 995  */
 996 static struct rte_flow *
 997 priv_flow_create_action_queue(struct priv *priv,
 998                               struct mlx5_flow *flow,
 999                               struct mlx5_flow_action *action,
1000                               struct rte_flow_error *error)
1001 {
1002         struct rte_flow *rte_flow;
1003         unsigned int i;
1004         unsigned int j;
1005         const unsigned int wqs_n = 1 << log2above(action->queues_n);
1006         struct ibv_exp_wq *wqs[wqs_n];
1007
1008         assert(priv->pd);
1009         assert(priv->ctx);
1010         assert(!action->drop);
1011         rte_flow = rte_calloc(__func__, 1,
1012                               sizeof(*rte_flow) + sizeof(struct rxq *) *
1013                               action->queues_n, 0);
1014         if (!rte_flow) {
1015                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1016                                    NULL, "cannot allocate flow memory");
1017                 return NULL;
1018         }
1019         rte_flow->rxqs = (struct rxq *(*)[])((uintptr_t)rte_flow +
1020                                              sizeof(struct rxq *) *
1021                                              action->queues_n);
1022         for (i = 0; i < action->queues_n; ++i) {
1023                 struct rxq_ctrl *rxq;
1024
1025                 rxq = container_of((*priv->rxqs)[action->queues[i]],
1026                                    struct rxq_ctrl, rxq);
1027                 wqs[i] = rxq->wq;
1028                 (*rte_flow->rxqs)[i] = &rxq->rxq;
1029                 ++rte_flow->rxqs_n;
1030                 rxq->rxq.mark |= action->mark;
1031         }
1032         /* finalise indirection table. */
1033         for (j = 0; i < wqs_n; ++i, ++j) {
1034                 wqs[i] = wqs[j];
1035                 if (j == action->queues_n)
1036                         j = 0;
1037         }
1038         rte_flow->mark = action->mark;
1039         rte_flow->ibv_attr = flow->ibv_attr;
1040         rte_flow->hash_fields = flow->hash_fields;
1041         rte_flow->ind_table = ibv_exp_create_rwq_ind_table(
1042                 priv->ctx,
1043                 &(struct ibv_exp_rwq_ind_table_init_attr){
1044                         .pd = priv->pd,
1045                         .log_ind_tbl_size = log2above(action->queues_n),
1046                         .ind_tbl = wqs,
1047                         .comp_mask = 0,
1048                 });
1049         if (!rte_flow->ind_table) {
1050                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1051                                    NULL, "cannot allocate indirection table");
1052                 goto error;
1053         }
1054         rte_flow->qp = ibv_exp_create_qp(
1055                 priv->ctx,
1056                 &(struct ibv_exp_qp_init_attr){
1057                         .qp_type = IBV_QPT_RAW_PACKET,
1058                         .comp_mask =
1059                                 IBV_EXP_QP_INIT_ATTR_PD |
1060                                 IBV_EXP_QP_INIT_ATTR_PORT |
1061                                 IBV_EXP_QP_INIT_ATTR_RX_HASH,
1062                         .pd = priv->pd,
1063                         .rx_hash_conf = &(struct ibv_exp_rx_hash_conf){
1064                                 .rx_hash_function =
1065                                         IBV_EXP_RX_HASH_FUNC_TOEPLITZ,
1066                                 .rx_hash_key_len = rss_hash_default_key_len,
1067                                 .rx_hash_key = rss_hash_default_key,
1068                                 .rx_hash_fields_mask = rte_flow->hash_fields,
1069                                 .rwq_ind_tbl = rte_flow->ind_table,
1070                         },
1071                         .port_num = priv->port,
1072                 });
1073         if (!rte_flow->qp) {
1074                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1075                                    NULL, "cannot allocate QP");
1076                 goto error;
1077         }
1078         if (!priv->started)
1079                 return rte_flow;
1080         rte_flow->ibv_flow = ibv_exp_create_flow(rte_flow->qp,
1081                                                  rte_flow->ibv_attr);
1082         if (!rte_flow->ibv_flow) {
1083                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1084                                    NULL, "flow rule creation failure");
1085                 goto error;
1086         }
1087         return rte_flow;
1088 error:
1089         assert(rte_flow);
1090         if (rte_flow->qp)
1091                 ibv_destroy_qp(rte_flow->qp);
1092         if (rte_flow->ind_table)
1093                 ibv_exp_destroy_rwq_ind_table(rte_flow->ind_table);
1094         rte_free(rte_flow);
1095         return NULL;
1096 }
1097
1098 /**
1099  * Convert a flow.
1100  *
1101  * @param priv
1102  *   Pointer to private structure.
1103  * @param[in] attr
1104  *   Flow rule attributes.
1105  * @param[in] pattern
1106  *   Pattern specification (list terminated by the END pattern item).
1107  * @param[in] actions
1108  *   Associated actions (list terminated by the END action).
1109  * @param[out] error
1110  *   Perform verbose error reporting if not NULL.
1111  *
1112  * @return
1113  *   A flow on success, NULL otherwise.
1114  */
1115 static struct rte_flow *
1116 priv_flow_create(struct priv *priv,
1117                  const struct rte_flow_attr *attr,
1118                  const struct rte_flow_item items[],
1119                  const struct rte_flow_action actions[],
1120                  struct rte_flow_error *error)
1121 {
1122         struct rte_flow *rte_flow;
1123         struct mlx5_flow_action action;
1124         struct mlx5_flow flow = { .offset = sizeof(struct ibv_exp_flow_attr), };
1125         int err;
1126
1127         err = priv_flow_validate(priv, attr, items, actions, error, &flow);
1128         if (err)
1129                 goto exit;
1130         flow.ibv_attr = rte_malloc(__func__, flow.offset, 0);
1131         flow.offset = sizeof(struct ibv_exp_flow_attr);
1132         if (!flow.ibv_attr) {
1133                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1134                                    NULL, "cannot allocate ibv_attr memory");
1135                 goto exit;
1136         }
1137         *flow.ibv_attr = (struct ibv_exp_flow_attr){
1138                 .type = IBV_EXP_FLOW_ATTR_NORMAL,
1139                 .size = sizeof(struct ibv_exp_flow_attr),
1140                 .priority = attr->priority,
1141                 .num_of_specs = 0,
1142                 .port = 0,
1143                 .flags = 0,
1144                 .reserved = 0,
1145         };
1146         flow.inner = 0;
1147         flow.hash_fields = 0;
1148         claim_zero(priv_flow_validate(priv, attr, items, actions,
1149                                       error, &flow));
1150         action = (struct mlx5_flow_action){
1151                 .queue = 0,
1152                 .drop = 0,
1153                 .mark = 0,
1154                 .mark_id = MLX5_FLOW_MARK_DEFAULT,
1155         };
1156         for (; actions->type != RTE_FLOW_ACTION_TYPE_END; ++actions) {
1157                 if (actions->type == RTE_FLOW_ACTION_TYPE_VOID) {
1158                         continue;
1159                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_QUEUE) {
1160                         action.queue = 1;
1161                         action.queues[action.queues_n++] =
1162                                 ((const struct rte_flow_action_queue *)
1163                                  actions->conf)->index;
1164                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_RSS) {
1165                         const struct rte_flow_action_rss *rss =
1166                                 (const struct rte_flow_action_rss *)
1167                                  actions->conf;
1168                         uint16_t n;
1169
1170                         action.queue = 1;
1171                         action.queues_n = rss->num;
1172                         for (n = 0; n < rss->num; ++n)
1173                                 action.queues[n] = rss->queue[n];
1174                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_DROP) {
1175                         action.drop = 1;
1176                         action.mark = 0;
1177                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_MARK) {
1178                         const struct rte_flow_action_mark *mark =
1179                                 (const struct rte_flow_action_mark *)
1180                                 actions->conf;
1181
1182                         if (mark)
1183                                 action.mark_id = mark->id;
1184                         action.mark = !action.drop;
1185                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_FLAG) {
1186                         action.mark = 1;
1187                 } else {
1188                         rte_flow_error_set(error, ENOTSUP,
1189                                            RTE_FLOW_ERROR_TYPE_ACTION,
1190                                            actions, "unsupported action");
1191                         goto exit;
1192                 }
1193         }
1194         if (action.mark) {
1195                 mlx5_flow_create_flag_mark(&flow, action.mark_id);
1196                 flow.offset += sizeof(struct ibv_exp_flow_spec_action_tag);
1197         }
1198         if (action.drop)
1199                 rte_flow =
1200                         priv_flow_create_action_queue_drop(priv, &flow, error);
1201         else
1202                 rte_flow = priv_flow_create_action_queue(priv, &flow, &action,
1203                                                          error);
1204         if (!rte_flow)
1205                 goto exit;
1206         return rte_flow;
1207 exit:
1208         rte_free(flow.ibv_attr);
1209         return NULL;
1210 }
1211
1212 /**
1213  * Create a flow.
1214  *
1215  * @see rte_flow_create()
1216  * @see rte_flow_ops
1217  */
1218 struct rte_flow *
1219 mlx5_flow_create(struct rte_eth_dev *dev,
1220                  const struct rte_flow_attr *attr,
1221                  const struct rte_flow_item items[],
1222                  const struct rte_flow_action actions[],
1223                  struct rte_flow_error *error)
1224 {
1225         struct priv *priv = dev->data->dev_private;
1226         struct rte_flow *flow;
1227
1228         priv_lock(priv);
1229         flow = priv_flow_create(priv, attr, items, actions, error);
1230         if (flow) {
1231                 LIST_INSERT_HEAD(&priv->flows, flow, next);
1232                 DEBUG("Flow created %p", (void *)flow);
1233         }
1234         priv_unlock(priv);
1235         return flow;
1236 }
1237
1238 /**
1239  * Destroy a flow.
1240  *
1241  * @param priv
1242  *   Pointer to private structure.
1243  * @param[in] flow
1244  *   Flow to destroy.
1245  */
1246 static void
1247 priv_flow_destroy(struct priv *priv,
1248                   struct rte_flow *flow)
1249 {
1250         (void)priv;
1251         LIST_REMOVE(flow, next);
1252         if (flow->ibv_flow)
1253                 claim_zero(ibv_exp_destroy_flow(flow->ibv_flow));
1254         if (flow->drop)
1255                 goto free;
1256         if (flow->qp)
1257                 claim_zero(ibv_destroy_qp(flow->qp));
1258         if (flow->ind_table)
1259                 claim_zero(ibv_exp_destroy_rwq_ind_table(flow->ind_table));
1260         if (flow->drop && flow->wq)
1261                 claim_zero(ibv_exp_destroy_wq(flow->wq));
1262         if (flow->drop && flow->cq)
1263                 claim_zero(ibv_destroy_cq(flow->cq));
1264         if (flow->mark) {
1265                 struct rte_flow *tmp;
1266                 struct rxq *rxq;
1267                 uint32_t mark_n = 0;
1268                 uint32_t queue_n;
1269
1270                 /*
1271                  * To remove the mark from the queue, the queue must not be
1272                  * present in any other marked flow (RSS or not).
1273                  */
1274                 for (queue_n = 0; queue_n < flow->rxqs_n; ++queue_n) {
1275                         rxq = (*flow->rxqs)[queue_n];
1276                         for (tmp = LIST_FIRST(&priv->flows);
1277                              tmp;
1278                              tmp = LIST_NEXT(tmp, next)) {
1279                                 uint32_t tqueue_n;
1280
1281                                 if (tmp->drop)
1282                                         continue;
1283                                 for (tqueue_n = 0;
1284                                      tqueue_n < tmp->rxqs_n;
1285                                      ++tqueue_n) {
1286                                         struct rxq *trxq;
1287
1288                                         trxq = (*tmp->rxqs)[tqueue_n];
1289                                         if (rxq == trxq)
1290                                                 ++mark_n;
1291                                 }
1292                         }
1293                         rxq->mark = !!mark_n;
1294                 }
1295         }
1296 free:
1297         rte_free(flow->ibv_attr);
1298         DEBUG("Flow destroyed %p", (void *)flow);
1299         rte_free(flow);
1300 }
1301
1302 /**
1303  * Destroy a flow.
1304  *
1305  * @see rte_flow_destroy()
1306  * @see rte_flow_ops
1307  */
1308 int
1309 mlx5_flow_destroy(struct rte_eth_dev *dev,
1310                   struct rte_flow *flow,
1311                   struct rte_flow_error *error)
1312 {
1313         struct priv *priv = dev->data->dev_private;
1314
1315         (void)error;
1316         priv_lock(priv);
1317         priv_flow_destroy(priv, flow);
1318         priv_unlock(priv);
1319         return 0;
1320 }
1321
1322 /**
1323  * Destroy all flows.
1324  *
1325  * @param priv
1326  *   Pointer to private structure.
1327  */
1328 static void
1329 priv_flow_flush(struct priv *priv)
1330 {
1331         while (!LIST_EMPTY(&priv->flows)) {
1332                 struct rte_flow *flow;
1333
1334                 flow = LIST_FIRST(&priv->flows);
1335                 priv_flow_destroy(priv, flow);
1336         }
1337 }
1338
1339 /**
1340  * Destroy all flows.
1341  *
1342  * @see rte_flow_flush()
1343  * @see rte_flow_ops
1344  */
1345 int
1346 mlx5_flow_flush(struct rte_eth_dev *dev,
1347                 struct rte_flow_error *error)
1348 {
1349         struct priv *priv = dev->data->dev_private;
1350
1351         (void)error;
1352         priv_lock(priv);
1353         priv_flow_flush(priv);
1354         priv_unlock(priv);
1355         return 0;
1356 }
1357
1358 /**
1359  * Create drop queue.
1360  *
1361  * @param priv
1362  *   Pointer to private structure.
1363  *
1364  * @return
1365  *   0 on success.
1366  */
1367 static int
1368 priv_flow_create_drop_queue(struct priv *priv)
1369 {
1370         struct rte_flow_drop *fdq = NULL;
1371         unsigned int i;
1372
1373         assert(priv->pd);
1374         assert(priv->ctx);
1375         fdq = rte_calloc(__func__, 1, sizeof(*fdq), 0);
1376         if (!fdq) {
1377                 WARN("cannot allocate memory for drop queue");
1378                 goto error;
1379         }
1380         fdq->cq = ibv_exp_create_cq(priv->ctx, 1, NULL, NULL, 0,
1381                         &(struct ibv_exp_cq_init_attr){
1382                         .comp_mask = 0,
1383                         });
1384         if (!fdq->cq) {
1385                 WARN("cannot allocate CQ for drop queue");
1386                 goto error;
1387         }
1388         for (i = 0; i != MLX5_DROP_WQ_N; ++i) {
1389                 fdq->wqs[i] = ibv_exp_create_wq(priv->ctx,
1390                                 &(struct ibv_exp_wq_init_attr){
1391                                 .wq_type = IBV_EXP_WQT_RQ,
1392                                 .max_recv_wr = 1,
1393                                 .max_recv_sge = 1,
1394                                 .pd = priv->pd,
1395                                 .cq = fdq->cq,
1396                                 });
1397                 if (!fdq->wqs[i]) {
1398                         WARN("cannot allocate WQ for drop queue");
1399                         goto error;
1400                 }
1401         }
1402         fdq->ind_table = ibv_exp_create_rwq_ind_table(priv->ctx,
1403                         &(struct ibv_exp_rwq_ind_table_init_attr){
1404                         .pd = priv->pd,
1405                         .log_ind_tbl_size = 0,
1406                         .ind_tbl = fdq->wqs,
1407                         .comp_mask = 0,
1408                         });
1409         if (!fdq->ind_table) {
1410                 WARN("cannot allocate indirection table for drop queue");
1411                 goto error;
1412         }
1413         fdq->qp = ibv_exp_create_qp(priv->ctx,
1414                 &(struct ibv_exp_qp_init_attr){
1415                         .qp_type = IBV_QPT_RAW_PACKET,
1416                         .comp_mask =
1417                                 IBV_EXP_QP_INIT_ATTR_PD |
1418                                 IBV_EXP_QP_INIT_ATTR_PORT |
1419                                 IBV_EXP_QP_INIT_ATTR_RX_HASH,
1420                         .pd = priv->pd,
1421                         .rx_hash_conf = &(struct ibv_exp_rx_hash_conf){
1422                                 .rx_hash_function =
1423                                         IBV_EXP_RX_HASH_FUNC_TOEPLITZ,
1424                                 .rx_hash_key_len = rss_hash_default_key_len,
1425                                 .rx_hash_key = rss_hash_default_key,
1426                                 .rx_hash_fields_mask = 0,
1427                                 .rwq_ind_tbl = fdq->ind_table,
1428                                 },
1429                         .port_num = priv->port,
1430                         });
1431         if (!fdq->qp) {
1432                 WARN("cannot allocate QP for drop queue");
1433                 goto error;
1434         }
1435         priv->flow_drop_queue = fdq;
1436         return 0;
1437 error:
1438         if (fdq->qp)
1439                 claim_zero(ibv_destroy_qp(fdq->qp));
1440         if (fdq->ind_table)
1441                 claim_zero(ibv_exp_destroy_rwq_ind_table(fdq->ind_table));
1442         for (i = 0; i != MLX5_DROP_WQ_N; ++i) {
1443                 if (fdq->wqs[i])
1444                         claim_zero(ibv_exp_destroy_wq(fdq->wqs[i]));
1445         }
1446         if (fdq->cq)
1447                 claim_zero(ibv_destroy_cq(fdq->cq));
1448         if (fdq)
1449                 rte_free(fdq);
1450         priv->flow_drop_queue = NULL;
1451         return -1;
1452 }
1453
1454 /**
1455  * Delete drop queue.
1456  *
1457  * @param priv
1458  *   Pointer to private structure.
1459  */
1460 static void
1461 priv_flow_delete_drop_queue(struct priv *priv)
1462 {
1463         struct rte_flow_drop *fdq = priv->flow_drop_queue;
1464         unsigned int i;
1465
1466         claim_zero(ibv_destroy_qp(fdq->qp));
1467         claim_zero(ibv_exp_destroy_rwq_ind_table(fdq->ind_table));
1468         for (i = 0; i != MLX5_DROP_WQ_N; ++i) {
1469                 assert(fdq->wqs[i]);
1470                 claim_zero(ibv_exp_destroy_wq(fdq->wqs[i]));
1471         }
1472         claim_zero(ibv_destroy_cq(fdq->cq));
1473         rte_free(fdq);
1474         priv->flow_drop_queue = NULL;
1475 }
1476
1477 /**
1478  * Remove all flows.
1479  *
1480  * Called by dev_stop() to remove all flows.
1481  *
1482  * @param priv
1483  *   Pointer to private structure.
1484  */
1485 void
1486 priv_flow_stop(struct priv *priv)
1487 {
1488         struct rte_flow *flow;
1489
1490         for (flow = LIST_FIRST(&priv->flows);
1491              flow;
1492              flow = LIST_NEXT(flow, next)) {
1493                 claim_zero(ibv_exp_destroy_flow(flow->ibv_flow));
1494                 flow->ibv_flow = NULL;
1495                 if (flow->mark) {
1496                         unsigned int n;
1497
1498                         for (n = 0; n < flow->rxqs_n; ++n)
1499                                 (*flow->rxqs)[n]->mark = 0;
1500                 }
1501                 DEBUG("Flow %p removed", (void *)flow);
1502         }
1503         priv_flow_delete_drop_queue(priv);
1504 }
1505
1506 /**
1507  * Add all flows.
1508  *
1509  * @param priv
1510  *   Pointer to private structure.
1511  *
1512  * @return
1513  *   0 on success, a errno value otherwise and rte_errno is set.
1514  */
1515 int
1516 priv_flow_start(struct priv *priv)
1517 {
1518         int ret;
1519         struct rte_flow *flow;
1520
1521         ret = priv_flow_create_drop_queue(priv);
1522         if (ret)
1523                 return -1;
1524         for (flow = LIST_FIRST(&priv->flows);
1525              flow;
1526              flow = LIST_NEXT(flow, next)) {
1527                 struct ibv_qp *qp;
1528
1529                 if (flow->drop)
1530                         qp = priv->flow_drop_queue->qp;
1531                 else
1532                         qp = flow->qp;
1533                 flow->ibv_flow = ibv_exp_create_flow(qp, flow->ibv_attr);
1534                 if (!flow->ibv_flow) {
1535                         DEBUG("Flow %p cannot be applied", (void *)flow);
1536                         rte_errno = EINVAL;
1537                         return rte_errno;
1538                 }
1539                 DEBUG("Flow %p applied", (void *)flow);
1540                 if (flow->mark) {
1541                         unsigned int n;
1542
1543                         for (n = 0; n < flow->rxqs_n; ++n)
1544                                 (*flow->rxqs)[n]->mark = 1;
1545                 }
1546         }
1547         return 0;
1548 }