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                 } else if (items->type == RTE_FLOW_ITEM_TYPE_VXLAN) {
 461                         if (flow->inner) {
 462                                 rte_flow_error_set(error, ENOTSUP,
 463                                                    RTE_FLOW_ERROR_TYPE_ITEM,
 464                                                    items,
 465                                                    "cannot recognize multiple"
 466                                                    " VXLAN encapsulations");
 467                                 return -rte_errno;
 468                         }
 469                         flow->inner = 1;
 470                 }
 471                 flow->offset += cur_item->dst_sz;
 472         }
 473         for (; actions->type != RTE_FLOW_ACTION_TYPE_END; ++actions) {
 474                 if (actions->type == RTE_FLOW_ACTION_TYPE_VOID) {
 475                         continue;
 476                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_DROP) {
 477                         action.drop = 1;
 478                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_QUEUE) {
 479                         const struct rte_flow_action_queue *queue =
 480                                 (const struct rte_flow_action_queue *)
 481                                 actions->conf;
 482                         uint16_t n;
 483                         uint16_t found = 0;
 484
 485                         if (!queue || (queue->index > (priv->rxqs_n - 1)))
 486                                 goto exit_action_not_supported;
 487                         for (n = 0; n < action.queues_n; ++n) {
 488                                 if (action.queues[n] == queue->index) {
 489                                         found = 1;
 490                                         break;
 491                                 }
 492                         }
 493                         if (action.queues_n && !found) {
 494                                 rte_flow_error_set(error, ENOTSUP,
 495                                            RTE_FLOW_ERROR_TYPE_ACTION,
 496                                            actions,
 497                                            "queue action not in RSS queues");
 498                                 return -rte_errno;
 499                         }
 500                         action.queue = 1;
 501                         action.queues_n = 1;
 502                         action.queues[0] = queue->index;
 503                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_RSS) {
 504                         const struct rte_flow_action_rss *rss =
 505                                 (const struct rte_flow_action_rss *)
 506                                 actions->conf;
 507                         uint16_t n;
 508
 509                         if (action.queues_n == 1) {
 510                                 uint16_t found = 0;
 511
 512                                 assert(action.queues_n);
 513                                 for (n = 0; n < rss->num; ++n) {
 514                                         if (action.queues[0] == rss->queue[n]) {
 515                                                 found = 1;
 516                                                 break;
 517                                         }
 518                                 }
 519                                 if (!found) {
 520                                         rte_flow_error_set(error, ENOTSUP,
 521                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 522                                                    actions,
 523                                                    "queue action not in RSS"
 524                                                    " queues");
 525                                         return -rte_errno;
 526                                 }
 527                         }
 528                         action.queue = 1;
 529                         for (n = 0; n < rss->num; ++n)
 530                                 action.queues[n] = rss->queue[n];
 531                         action.queues_n = rss->num;
 532                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_MARK) {
 533                         const struct rte_flow_action_mark *mark =
 534                                 (const struct rte_flow_action_mark *)
 535                                 actions->conf;
 536
 537                         if (!mark) {
 538                                 rte_flow_error_set(error, EINVAL,
 539                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 540                                                    actions,
 541                                                    "mark must be defined");
 542                                 return -rte_errno;
 543                         } else if (mark->id >= MLX5_FLOW_MARK_MAX) {
 544                                 rte_flow_error_set(error, ENOTSUP,
 545                                                    RTE_FLOW_ERROR_TYPE_ACTION,
 546                                                    actions,
 547                                                    "mark must be between 0"
 548                                                    " and 16777199");
 549                                 return -rte_errno;
 550                         }
 551                         action.mark = 1;
 552                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_FLAG) {
 553                         action.mark = 1;
 554                 } else {
 555                         goto exit_action_not_supported;
 556                 }
 557         }
 558         if (action.mark && !flow->ibv_attr && !action.drop)
 559                 flow->offset += sizeof(struct ibv_exp_flow_spec_action_tag);
 560         if (!action.queue && !action.drop) {
 561                 rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_HANDLE,
 562                                    NULL, "no valid action");
 563                 return -rte_errno;
 564         }
 565         return 0;
 566 exit_item_not_supported:
 567         rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 568                            items, "item not supported");
 569         return -rte_errno;
 570 exit_action_not_supported:
 571         rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
 572                            actions, "action not supported");
 573         return -rte_errno;
 574 }
 575
 576 /**
 577  * Validate a flow supported by the NIC.
 578  *
 579  * @see rte_flow_validate()
 580  * @see rte_flow_ops
 581  */
 582 int
 583 mlx5_flow_validate(struct rte_eth_dev *dev,
 584                    const struct rte_flow_attr *attr,
 585                    const struct rte_flow_item items[],
 586                    const struct rte_flow_action actions[],
 587                    struct rte_flow_error *error)
 588 {
 589         struct priv *priv = dev->data->dev_private;
 590         int ret;
 591         struct mlx5_flow flow = { .offset = sizeof(struct ibv_exp_flow_attr) };
 592
 593         priv_lock(priv);
 594         ret = priv_flow_validate(priv, attr, items, actions, error, &flow);
 595         priv_unlock(priv);
 596         return ret;
 597 }
 598
 599 /**
 600  * Convert Ethernet item to Verbs specification.
 601  *
 602  * @param item[in]
 603  *   Item specification.
 604  * @param default_mask[in]
 605  *   Default bit-masks to use when item->mask is not provided.
 606  * @param data[in, out]
 607  *   User structure.
 608  */
 609 static int
 610 mlx5_flow_create_eth(const struct rte_flow_item *item,
 611                      const void *default_mask,
 612                      void *data)
 613 {
 614         const struct rte_flow_item_eth *spec = item->spec;
 615         const struct rte_flow_item_eth *mask = item->mask;
 616         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 617         struct ibv_exp_flow_spec_eth *eth;
 618         const unsigned int eth_size = sizeof(struct ibv_exp_flow_spec_eth);
 619         unsigned int i;
 620
 621         ++flow->ibv_attr->num_of_specs;
 622         flow->ibv_attr->priority = 2;
 623         flow->hash_fields = 0;
 624         eth = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 625         *eth = (struct ibv_exp_flow_spec_eth) {
 626                 .type = flow->inner | IBV_EXP_FLOW_SPEC_ETH,
 627                 .size = eth_size,
 628         };
 629         if (!spec)
 630                 return 0;
 631         if (!mask)
 632                 mask = default_mask;
 633         memcpy(eth->val.dst_mac, spec->dst.addr_bytes, ETHER_ADDR_LEN);
 634         memcpy(eth->val.src_mac, spec->src.addr_bytes, ETHER_ADDR_LEN);
 635         eth->val.ether_type = spec->type;
 636         memcpy(eth->mask.dst_mac, mask->dst.addr_bytes, ETHER_ADDR_LEN);
 637         memcpy(eth->mask.src_mac, mask->src.addr_bytes, ETHER_ADDR_LEN);
 638         eth->mask.ether_type = mask->type;
 639         /* Remove unwanted bits from values. */
 640         for (i = 0; i < ETHER_ADDR_LEN; ++i) {
 641                 eth->val.dst_mac[i] &= eth->mask.dst_mac[i];
 642                 eth->val.src_mac[i] &= eth->mask.src_mac[i];
 643         }
 644         eth->val.ether_type &= eth->mask.ether_type;
 645         return 0;
 646 }
 647
 648 /**
 649  * Convert VLAN item to Verbs specification.
 650  *
 651  * @param item[in]
 652  *   Item specification.
 653  * @param default_mask[in]
 654  *   Default bit-masks to use when item->mask is not provided.
 655  * @param data[in, out]
 656  *   User structure.
 657  */
 658 static int
 659 mlx5_flow_create_vlan(const struct rte_flow_item *item,
 660                       const void *default_mask,
 661                       void *data)
 662 {
 663         const struct rte_flow_item_vlan *spec = item->spec;
 664         const struct rte_flow_item_vlan *mask = item->mask;
 665         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 666         struct ibv_exp_flow_spec_eth *eth;
 667         const unsigned int eth_size = sizeof(struct ibv_exp_flow_spec_eth);
 668
 669         eth = (void *)((uintptr_t)flow->ibv_attr + flow->offset - eth_size);
 670         if (!spec)
 671                 return 0;
 672         if (!mask)
 673                 mask = default_mask;
 674         eth->val.vlan_tag = spec->tci;
 675         eth->mask.vlan_tag = mask->tci;
 676         eth->val.vlan_tag &= eth->mask.vlan_tag;
 677         return 0;
 678 }
 679
 680 /**
 681  * Convert IPv4 item to Verbs specification.
 682  *
 683  * @param item[in]
 684  *   Item specification.
 685  * @param default_mask[in]
 686  *   Default bit-masks to use when item->mask is not provided.
 687  * @param data[in, out]
 688  *   User structure.
 689  */
 690 static int
 691 mlx5_flow_create_ipv4(const struct rte_flow_item *item,
 692                       const void *default_mask,
 693                       void *data)
 694 {
 695         const struct rte_flow_item_ipv4 *spec = item->spec;
 696         const struct rte_flow_item_ipv4 *mask = item->mask;
 697         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 698         struct ibv_exp_flow_spec_ipv4_ext *ipv4;
 699         unsigned int ipv4_size = sizeof(struct ibv_exp_flow_spec_ipv4_ext);
 700
 701         ++flow->ibv_attr->num_of_specs;
 702         flow->ibv_attr->priority = 1;
 703         flow->hash_fields = (IBV_EXP_RX_HASH_SRC_IPV4 |
 704                              IBV_EXP_RX_HASH_DST_IPV4);
 705         ipv4 = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 706         *ipv4 = (struct ibv_exp_flow_spec_ipv4_ext) {
 707                 .type = flow->inner | IBV_EXP_FLOW_SPEC_IPV4_EXT,
 708                 .size = ipv4_size,
 709         };
 710         if (!spec)
 711                 return 0;
 712         if (!mask)
 713                 mask = default_mask;
 714         ipv4->val = (struct ibv_exp_flow_ipv4_ext_filter){
 715                 .src_ip = spec->hdr.src_addr,
 716                 .dst_ip = spec->hdr.dst_addr,
 717                 .proto = spec->hdr.next_proto_id,
 718                 .tos = spec->hdr.type_of_service,
 719         };
 720         ipv4->mask = (struct ibv_exp_flow_ipv4_ext_filter){
 721                 .src_ip = mask->hdr.src_addr,
 722                 .dst_ip = mask->hdr.dst_addr,
 723                 .proto = mask->hdr.next_proto_id,
 724                 .tos = mask->hdr.type_of_service,
 725         };
 726         /* Remove unwanted bits from values. */
 727         ipv4->val.src_ip &= ipv4->mask.src_ip;
 728         ipv4->val.dst_ip &= ipv4->mask.dst_ip;
 729         ipv4->val.proto &= ipv4->mask.proto;
 730         ipv4->val.tos &= ipv4->mask.tos;
 731         return 0;
 732 }
 733
 734 /**
 735  * Convert IPv6 item to Verbs specification.
 736  *
 737  * @param item[in]
 738  *   Item specification.
 739  * @param default_mask[in]
 740  *   Default bit-masks to use when item->mask is not provided.
 741  * @param data[in, out]
 742  *   User structure.
 743  */
 744 static int
 745 mlx5_flow_create_ipv6(const struct rte_flow_item *item,
 746                       const void *default_mask,
 747                       void *data)
 748 {
 749         const struct rte_flow_item_ipv6 *spec = item->spec;
 750         const struct rte_flow_item_ipv6 *mask = item->mask;
 751         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 752         struct ibv_exp_flow_spec_ipv6_ext *ipv6;
 753         unsigned int ipv6_size = sizeof(struct ibv_exp_flow_spec_ipv6_ext);
 754
 755         ++flow->ibv_attr->num_of_specs;
 756         flow->ibv_attr->priority = 1;
 757         flow->hash_fields = (IBV_EXP_RX_HASH_SRC_IPV6 |
 758                              IBV_EXP_RX_HASH_DST_IPV6);
 759         ipv6 = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 760         *ipv6 = (struct ibv_exp_flow_spec_ipv6_ext) {
 761                 .type = flow->inner | IBV_EXP_FLOW_SPEC_IPV6_EXT,
 762                 .size = ipv6_size,
 763         };
 764         if (!spec)
 765                 return 0;
 766         if (!mask)
 767                 mask = default_mask;
 768         memcpy(ipv6->val.src_ip, spec->hdr.src_addr,
 769                RTE_DIM(ipv6->val.src_ip));
 770         memcpy(ipv6->val.dst_ip, spec->hdr.dst_addr,
 771                RTE_DIM(ipv6->val.dst_ip));
 772         memcpy(ipv6->mask.src_ip, mask->hdr.src_addr,
 773                RTE_DIM(ipv6->mask.src_ip));
 774         memcpy(ipv6->mask.dst_ip, mask->hdr.dst_addr,
 775                RTE_DIM(ipv6->mask.dst_ip));
 776         ipv6->mask.flow_label = mask->hdr.vtc_flow;
 777         ipv6->mask.next_hdr = mask->hdr.proto;
 778         ipv6->mask.hop_limit = mask->hdr.hop_limits;
 779         ipv6->val.flow_label &= ipv6->mask.flow_label;
 780         ipv6->val.next_hdr &= ipv6->mask.next_hdr;
 781         ipv6->val.hop_limit &= ipv6->mask.hop_limit;
 782         return 0;
 783 }
 784
 785 /**
 786  * Convert UDP item to Verbs specification.
 787  *
 788  * @param item[in]
 789  *   Item specification.
 790  * @param default_mask[in]
 791  *   Default bit-masks to use when item->mask is not provided.
 792  * @param data[in, out]
 793  *   User structure.
 794  */
 795 static int
 796 mlx5_flow_create_udp(const struct rte_flow_item *item,
 797                      const void *default_mask,
 798                      void *data)
 799 {
 800         const struct rte_flow_item_udp *spec = item->spec;
 801         const struct rte_flow_item_udp *mask = item->mask;
 802         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 803         struct ibv_exp_flow_spec_tcp_udp *udp;
 804         unsigned int udp_size = sizeof(struct ibv_exp_flow_spec_tcp_udp);
 805
 806         ++flow->ibv_attr->num_of_specs;
 807         flow->ibv_attr->priority = 0;
 808         flow->hash_fields |= (IBV_EXP_RX_HASH_SRC_PORT_UDP |
 809                               IBV_EXP_RX_HASH_DST_PORT_UDP);
 810         udp = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 811         *udp = (struct ibv_exp_flow_spec_tcp_udp) {
 812                 .type = flow->inner | IBV_EXP_FLOW_SPEC_UDP,
 813                 .size = udp_size,
 814         };
 815         if (!spec)
 816                 return 0;
 817         if (!mask)
 818                 mask = default_mask;
 819         udp->val.dst_port = spec->hdr.dst_port;
 820         udp->val.src_port = spec->hdr.src_port;
 821         udp->mask.dst_port = mask->hdr.dst_port;
 822         udp->mask.src_port = mask->hdr.src_port;
 823         /* Remove unwanted bits from values. */
 824         udp->val.src_port &= udp->mask.src_port;
 825         udp->val.dst_port &= udp->mask.dst_port;
 826         return 0;
 827 }
 828
 829 /**
 830  * Convert TCP item to Verbs specification.
 831  *
 832  * @param item[in]
 833  *   Item specification.
 834  * @param default_mask[in]
 835  *   Default bit-masks to use when item->mask is not provided.
 836  * @param data[in, out]
 837  *   User structure.
 838  */
 839 static int
 840 mlx5_flow_create_tcp(const struct rte_flow_item *item,
 841                      const void *default_mask,
 842                      void *data)
 843 {
 844         const struct rte_flow_item_tcp *spec = item->spec;
 845         const struct rte_flow_item_tcp *mask = item->mask;
 846         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 847         struct ibv_exp_flow_spec_tcp_udp *tcp;
 848         unsigned int tcp_size = sizeof(struct ibv_exp_flow_spec_tcp_udp);
 849
 850         ++flow->ibv_attr->num_of_specs;
 851         flow->ibv_attr->priority = 0;
 852         flow->hash_fields |= (IBV_EXP_RX_HASH_SRC_PORT_TCP |
 853                               IBV_EXP_RX_HASH_DST_PORT_TCP);
 854         tcp = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 855         *tcp = (struct ibv_exp_flow_spec_tcp_udp) {
 856                 .type = flow->inner | IBV_EXP_FLOW_SPEC_TCP,
 857                 .size = tcp_size,
 858         };
 859         if (!spec)
 860                 return 0;
 861         if (!mask)
 862                 mask = default_mask;
 863         tcp->val.dst_port = spec->hdr.dst_port;
 864         tcp->val.src_port = spec->hdr.src_port;
 865         tcp->mask.dst_port = mask->hdr.dst_port;
 866         tcp->mask.src_port = mask->hdr.src_port;
 867         /* Remove unwanted bits from values. */
 868         tcp->val.src_port &= tcp->mask.src_port;
 869         tcp->val.dst_port &= tcp->mask.dst_port;
 870         return 0;
 871 }
 872
 873 /**
 874  * Convert VXLAN item to Verbs specification.
 875  *
 876  * @param item[in]
 877  *   Item specification.
 878  * @param default_mask[in]
 879  *   Default bit-masks to use when item->mask is not provided.
 880  * @param data[in, out]
 881  *   User structure.
 882  */
 883 static int
 884 mlx5_flow_create_vxlan(const struct rte_flow_item *item,
 885                        const void *default_mask,
 886                        void *data)
 887 {
 888         const struct rte_flow_item_vxlan *spec = item->spec;
 889         const struct rte_flow_item_vxlan *mask = item->mask;
 890         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 891         struct ibv_exp_flow_spec_tunnel *vxlan;
 892         unsigned int size = sizeof(struct ibv_exp_flow_spec_tunnel);
 893         union vni {
 894                 uint32_t vlan_id;
 895                 uint8_t vni[4];
 896         } id;
 897
 898         ++flow->ibv_attr->num_of_specs;
 899         flow->ibv_attr->priority = 0;
 900         id.vni[0] = 0;
 901         vxlan = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 902         *vxlan = (struct ibv_exp_flow_spec_tunnel) {
 903                 .type = flow->inner | IBV_EXP_FLOW_SPEC_VXLAN_TUNNEL,
 904                 .size = size,
 905         };
 906         flow->inner = IBV_EXP_FLOW_SPEC_INNER;
 907         if (!spec)
 908                 return 0;
 909         if (!mask)
 910                 mask = default_mask;
 911         memcpy(&id.vni[1], spec->vni, 3);
 912         vxlan->val.tunnel_id = id.vlan_id;
 913         memcpy(&id.vni[1], mask->vni, 3);
 914         vxlan->mask.tunnel_id = id.vlan_id;
 915         /* Remove unwanted bits from values. */
 916         vxlan->val.tunnel_id &= vxlan->mask.tunnel_id;
 917         return 0;
 918 }
 919
 920 /**
 921  * Convert mark/flag action to Verbs specification.
 922  *
 923  * @param flow
 924  *   Pointer to MLX5 flow structure.
 925  * @param mark_id
 926  *   Mark identifier.
 927  */
 928 static int
 929 mlx5_flow_create_flag_mark(struct mlx5_flow *flow, uint32_t mark_id)
 930 {
 931         struct ibv_exp_flow_spec_action_tag *tag;
 932         unsigned int size = sizeof(struct ibv_exp_flow_spec_action_tag);
 933
 934         tag = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 935         *tag = (struct ibv_exp_flow_spec_action_tag){
 936                 .type = IBV_EXP_FLOW_SPEC_ACTION_TAG,
 937                 .size = size,
 938                 .tag_id = mlx5_flow_mark_set(mark_id),
 939         };
 940         ++flow->ibv_attr->num_of_specs;
 941         return 0;
 942 }
 943
 944 /**
 945  * Complete flow rule creation with a drop queue.
 946  *
 947  * @param priv
 948  *   Pointer to private structure.
 949  * @param flow
 950  *   MLX5 flow attributes (filled by mlx5_flow_validate()).
 951  * @param[out] error
 952  *   Perform verbose error reporting if not NULL.
 953  *
 954  * @return
 955  *   A flow if the rule could be created.
 956  */
 957 static struct rte_flow *
 958 priv_flow_create_action_queue_drop(struct priv *priv,
 959                                    struct mlx5_flow *flow,
 960                                    struct rte_flow_error *error)
 961 {
 962         struct rte_flow *rte_flow;
 963
 964         assert(priv->pd);
 965         assert(priv->ctx);
 966         rte_flow = rte_calloc(__func__, 1, sizeof(*rte_flow), 0);
 967         if (!rte_flow) {
 968                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
 969                                    NULL, "cannot allocate flow memory");
 970                 return NULL;
 971         }
 972         rte_flow->drop = 1;
 973         rte_flow->ibv_attr = flow->ibv_attr;
 974         rte_flow->qp = priv->flow_drop_queue->qp;
 975         if (!priv->started)
 976                 return rte_flow;
 977         rte_flow->ibv_flow = ibv_exp_create_flow(rte_flow->qp,
 978                                                  rte_flow->ibv_attr);
 979         if (!rte_flow->ibv_flow) {
 980                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
 981                                    NULL, "flow rule creation failure");
 982                 goto error;
 983         }
 984         return rte_flow;
 985 error:
 986         assert(rte_flow);
 987         rte_free(rte_flow);
 988         return NULL;
 989 }
 990
 991 /**
 992  * Complete flow rule creation.
 993  *
 994  * @param priv
 995  *   Pointer to private structure.
 996  * @param flow
 997  *   MLX5 flow attributes (filled by mlx5_flow_validate()).
 998  * @param action
 999  *   Target action structure.
1000  * @param[out] error
1001  *   Perform verbose error reporting if not NULL.
1002  *
1003  * @return
1004  *   A flow if the rule could be created.
1005  */
1006 static struct rte_flow *
1007 priv_flow_create_action_queue(struct priv *priv,
1008                               struct mlx5_flow *flow,
1009                               struct mlx5_flow_action *action,
1010                               struct rte_flow_error *error)
1011 {
1012         struct rte_flow *rte_flow;
1013         unsigned int i;
1014         unsigned int j;
1015         const unsigned int wqs_n = 1 << log2above(action->queues_n);
1016         struct ibv_exp_wq *wqs[wqs_n];
1017
1018         assert(priv->pd);
1019         assert(priv->ctx);
1020         assert(!action->drop);
1021         rte_flow = rte_calloc(__func__, 1,
1022                               sizeof(*rte_flow) + sizeof(struct rxq *) *
1023                               action->queues_n, 0);
1024         if (!rte_flow) {
1025                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1026                                    NULL, "cannot allocate flow memory");
1027                 return NULL;
1028         }
1029         rte_flow->rxqs = (struct rxq *(*)[])((uintptr_t)rte_flow +
1030                                              sizeof(struct rxq *) *
1031                                              action->queues_n);
1032         for (i = 0; i < action->queues_n; ++i) {
1033                 struct rxq_ctrl *rxq;
1034
1035                 rxq = container_of((*priv->rxqs)[action->queues[i]],
1036                                    struct rxq_ctrl, rxq);
1037                 wqs[i] = rxq->wq;
1038                 (*rte_flow->rxqs)[i] = &rxq->rxq;
1039                 ++rte_flow->rxqs_n;
1040                 rxq->rxq.mark |= action->mark;
1041         }
1042         /* finalise indirection table. */
1043         for (j = 0; i < wqs_n; ++i, ++j) {
1044                 wqs[i] = wqs[j];
1045                 if (j == action->queues_n)
1046                         j = 0;
1047         }
1048         rte_flow->mark = action->mark;
1049         rte_flow->ibv_attr = flow->ibv_attr;
1050         rte_flow->hash_fields = flow->hash_fields;
1051         rte_flow->ind_table = ibv_exp_create_rwq_ind_table(
1052                 priv->ctx,
1053                 &(struct ibv_exp_rwq_ind_table_init_attr){
1054                         .pd = priv->pd,
1055                         .log_ind_tbl_size = log2above(action->queues_n),
1056                         .ind_tbl = wqs,
1057                         .comp_mask = 0,
1058                 });
1059         if (!rte_flow->ind_table) {
1060                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1061                                    NULL, "cannot allocate indirection table");
1062                 goto error;
1063         }
1064         rte_flow->qp = ibv_exp_create_qp(
1065                 priv->ctx,
1066                 &(struct ibv_exp_qp_init_attr){
1067                         .qp_type = IBV_QPT_RAW_PACKET,
1068                         .comp_mask =
1069                                 IBV_EXP_QP_INIT_ATTR_PD |
1070                                 IBV_EXP_QP_INIT_ATTR_PORT |
1071                                 IBV_EXP_QP_INIT_ATTR_RX_HASH,
1072                         .pd = priv->pd,
1073                         .rx_hash_conf = &(struct ibv_exp_rx_hash_conf){
1074                                 .rx_hash_function =
1075                                         IBV_EXP_RX_HASH_FUNC_TOEPLITZ,
1076                                 .rx_hash_key_len = rss_hash_default_key_len,
1077                                 .rx_hash_key = rss_hash_default_key,
1078                                 .rx_hash_fields_mask = rte_flow->hash_fields,
1079                                 .rwq_ind_tbl = rte_flow->ind_table,
1080                         },
1081                         .port_num = priv->port,
1082                 });
1083         if (!rte_flow->qp) {
1084                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1085                                    NULL, "cannot allocate QP");
1086                 goto error;
1087         }
1088         if (!priv->started)
1089                 return rte_flow;
1090         rte_flow->ibv_flow = ibv_exp_create_flow(rte_flow->qp,
1091                                                  rte_flow->ibv_attr);
1092         if (!rte_flow->ibv_flow) {
1093                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1094                                    NULL, "flow rule creation failure");
1095                 goto error;
1096         }
1097         return rte_flow;
1098 error:
1099         assert(rte_flow);
1100         if (rte_flow->qp)
1101                 ibv_destroy_qp(rte_flow->qp);
1102         if (rte_flow->ind_table)
1103                 ibv_exp_destroy_rwq_ind_table(rte_flow->ind_table);
1104         rte_free(rte_flow);
1105         return NULL;
1106 }
1107
1108 /**
1109  * Convert a flow.
1110  *
1111  * @param priv
1112  *   Pointer to private structure.
1113  * @param[in] attr
1114  *   Flow rule attributes.
1115  * @param[in] pattern
1116  *   Pattern specification (list terminated by the END pattern item).
1117  * @param[in] actions
1118  *   Associated actions (list terminated by the END action).
1119  * @param[out] error
1120  *   Perform verbose error reporting if not NULL.
1121  *
1122  * @return
1123  *   A flow on success, NULL otherwise.
1124  */
1125 static struct rte_flow *
1126 priv_flow_create(struct priv *priv,
1127                  const struct rte_flow_attr *attr,
1128                  const struct rte_flow_item items[],
1129                  const struct rte_flow_action actions[],
1130                  struct rte_flow_error *error)
1131 {
1132         struct rte_flow *rte_flow;
1133         struct mlx5_flow_action action;
1134         struct mlx5_flow flow = { .offset = sizeof(struct ibv_exp_flow_attr), };
1135         int err;
1136
1137         err = priv_flow_validate(priv, attr, items, actions, error, &flow);
1138         if (err)
1139                 goto exit;
1140         flow.ibv_attr = rte_malloc(__func__, flow.offset, 0);
1141         flow.offset = sizeof(struct ibv_exp_flow_attr);
1142         if (!flow.ibv_attr) {
1143                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
1144                                    NULL, "cannot allocate ibv_attr memory");
1145                 goto exit;
1146         }
1147         *flow.ibv_attr = (struct ibv_exp_flow_attr){
1148                 .type = IBV_EXP_FLOW_ATTR_NORMAL,
1149                 .size = sizeof(struct ibv_exp_flow_attr),
1150                 .priority = attr->priority,
1151                 .num_of_specs = 0,
1152                 .port = 0,
1153                 .flags = 0,
1154                 .reserved = 0,
1155         };
1156         flow.inner = 0;
1157         flow.hash_fields = 0;
1158         claim_zero(priv_flow_validate(priv, attr, items, actions,
1159                                       error, &flow));
1160         action = (struct mlx5_flow_action){
1161                 .queue = 0,
1162                 .drop = 0,
1163                 .mark = 0,
1164                 .mark_id = MLX5_FLOW_MARK_DEFAULT,
1165         };
1166         for (; actions->type != RTE_FLOW_ACTION_TYPE_END; ++actions) {
1167                 if (actions->type == RTE_FLOW_ACTION_TYPE_VOID) {
1168                         continue;
1169                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_QUEUE) {
1170                         action.queue = 1;
1171                         action.queues[action.queues_n++] =
1172                                 ((const struct rte_flow_action_queue *)
1173                                  actions->conf)->index;
1174                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_RSS) {
1175                         const struct rte_flow_action_rss *rss =
1176                                 (const struct rte_flow_action_rss *)
1177                                  actions->conf;
1178                         uint16_t n;
1179
1180                         action.queue = 1;
1181                         action.queues_n = rss->num;
1182                         for (n = 0; n < rss->num; ++n)
1183                                 action.queues[n] = rss->queue[n];
1184                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_DROP) {
1185                         action.drop = 1;
1186                         action.mark = 0;
1187                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_MARK) {
1188                         const struct rte_flow_action_mark *mark =
1189                                 (const struct rte_flow_action_mark *)
1190                                 actions->conf;
1191
1192                         if (mark)
1193                                 action.mark_id = mark->id;
1194                         action.mark = !action.drop;
1195                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_FLAG) {
1196                         action.mark = 1;
1197                 } else {
1198                         rte_flow_error_set(error, ENOTSUP,
1199                                            RTE_FLOW_ERROR_TYPE_ACTION,
1200                                            actions, "unsupported action");
1201                         goto exit;
1202                 }
1203         }
1204         if (action.mark) {
1205                 mlx5_flow_create_flag_mark(&flow, action.mark_id);
1206                 flow.offset += sizeof(struct ibv_exp_flow_spec_action_tag);
1207         }
1208         if (action.drop)
1209                 rte_flow =
1210                         priv_flow_create_action_queue_drop(priv, &flow, error);
1211         else
1212                 rte_flow = priv_flow_create_action_queue(priv, &flow, &action,
1213                                                          error);
1214         if (!rte_flow)
1215                 goto exit;
1216         return rte_flow;
1217 exit:
1218         rte_free(flow.ibv_attr);
1219         return NULL;
1220 }
1221
1222 /**
1223  * Create a flow.
1224  *
1225  * @see rte_flow_create()
1226  * @see rte_flow_ops
1227  */
1228 struct rte_flow *
1229 mlx5_flow_create(struct rte_eth_dev *dev,
1230                  const struct rte_flow_attr *attr,
1231                  const struct rte_flow_item items[],
1232                  const struct rte_flow_action actions[],
1233                  struct rte_flow_error *error)
1234 {
1235         struct priv *priv = dev->data->dev_private;
1236         struct rte_flow *flow;
1237
1238         priv_lock(priv);
1239         flow = priv_flow_create(priv, attr, items, actions, error);
1240         if (flow) {
1241                 LIST_INSERT_HEAD(&priv->flows, flow, next);
1242                 DEBUG("Flow created %p", (void *)flow);
1243         }
1244         priv_unlock(priv);
1245         return flow;
1246 }
1247
1248 /**
1249  * Destroy a flow.
1250  *
1251  * @param priv
1252  *   Pointer to private structure.
1253  * @param[in] flow
1254  *   Flow to destroy.
1255  */
1256 static void
1257 priv_flow_destroy(struct priv *priv,
1258                   struct rte_flow *flow)
1259 {
1260         (void)priv;
1261         LIST_REMOVE(flow, next);
1262         if (flow->ibv_flow)
1263                 claim_zero(ibv_exp_destroy_flow(flow->ibv_flow));
1264         if (flow->drop)
1265                 goto free;
1266         if (flow->qp)
1267                 claim_zero(ibv_destroy_qp(flow->qp));
1268         if (flow->ind_table)
1269                 claim_zero(ibv_exp_destroy_rwq_ind_table(flow->ind_table));
1270         if (flow->drop && flow->wq)
1271                 claim_zero(ibv_exp_destroy_wq(flow->wq));
1272         if (flow->drop && flow->cq)
1273                 claim_zero(ibv_destroy_cq(flow->cq));
1274         if (flow->mark) {
1275                 struct rte_flow *tmp;
1276                 struct rxq *rxq;
1277                 uint32_t mark_n = 0;
1278                 uint32_t queue_n;
1279
1280                 /*
1281                  * To remove the mark from the queue, the queue must not be
1282                  * present in any other marked flow (RSS or not).
1283                  */
1284                 for (queue_n = 0; queue_n < flow->rxqs_n; ++queue_n) {
1285                         rxq = (*flow->rxqs)[queue_n];
1286                         for (tmp = LIST_FIRST(&priv->flows);
1287                              tmp;
1288                              tmp = LIST_NEXT(tmp, next)) {
1289                                 uint32_t tqueue_n;
1290
1291                                 if (tmp->drop)
1292                                         continue;
1293                                 for (tqueue_n = 0;
1294                                      tqueue_n < tmp->rxqs_n;
1295                                      ++tqueue_n) {
1296                                         struct rxq *trxq;
1297
1298                                         trxq = (*tmp->rxqs)[tqueue_n];
1299                                         if (rxq == trxq)
1300                                                 ++mark_n;
1301                                 }
1302                         }
1303                         rxq->mark = !!mark_n;
1304                 }
1305         }
1306 free:
1307         rte_free(flow->ibv_attr);
1308         DEBUG("Flow destroyed %p", (void *)flow);
1309         rte_free(flow);
1310 }
1311
1312 /**
1313  * Destroy a flow.
1314  *
1315  * @see rte_flow_destroy()
1316  * @see rte_flow_ops
1317  */
1318 int
1319 mlx5_flow_destroy(struct rte_eth_dev *dev,
1320                   struct rte_flow *flow,
1321                   struct rte_flow_error *error)
1322 {
1323         struct priv *priv = dev->data->dev_private;
1324
1325         (void)error;
1326         priv_lock(priv);
1327         priv_flow_destroy(priv, flow);
1328         priv_unlock(priv);
1329         return 0;
1330 }
1331
1332 /**
1333  * Destroy all flows.
1334  *
1335  * @param priv
1336  *   Pointer to private structure.
1337  */
1338 static void
1339 priv_flow_flush(struct priv *priv)
1340 {
1341         while (!LIST_EMPTY(&priv->flows)) {
1342                 struct rte_flow *flow;
1343
1344                 flow = LIST_FIRST(&priv->flows);
1345                 priv_flow_destroy(priv, flow);
1346         }
1347 }
1348
1349 /**
1350  * Destroy all flows.
1351  *
1352  * @see rte_flow_flush()
1353  * @see rte_flow_ops
1354  */
1355 int
1356 mlx5_flow_flush(struct rte_eth_dev *dev,
1357                 struct rte_flow_error *error)
1358 {
1359         struct priv *priv = dev->data->dev_private;
1360
1361         (void)error;
1362         priv_lock(priv);
1363         priv_flow_flush(priv);
1364         priv_unlock(priv);
1365         return 0;
1366 }
1367
1368 /**
1369  * Create drop queue.
1370  *
1371  * @param priv
1372  *   Pointer to private structure.
1373  *
1374  * @return
1375  *   0 on success.
1376  */
1377 static int
1378 priv_flow_create_drop_queue(struct priv *priv)
1379 {
1380         struct rte_flow_drop *fdq = NULL;
1381         unsigned int i;
1382
1383         assert(priv->pd);
1384         assert(priv->ctx);
1385         fdq = rte_calloc(__func__, 1, sizeof(*fdq), 0);
1386         if (!fdq) {
1387                 WARN("cannot allocate memory for drop queue");
1388                 goto error;
1389         }
1390         fdq->cq = ibv_exp_create_cq(priv->ctx, 1, NULL, NULL, 0,
1391                         &(struct ibv_exp_cq_init_attr){
1392                         .comp_mask = 0,
1393                         });
1394         if (!fdq->cq) {
1395                 WARN("cannot allocate CQ for drop queue");
1396                 goto error;
1397         }
1398         for (i = 0; i != MLX5_DROP_WQ_N; ++i) {
1399                 fdq->wqs[i] = ibv_exp_create_wq(priv->ctx,
1400                                 &(struct ibv_exp_wq_init_attr){
1401                                 .wq_type = IBV_EXP_WQT_RQ,
1402                                 .max_recv_wr = 1,
1403                                 .max_recv_sge = 1,
1404                                 .pd = priv->pd,
1405                                 .cq = fdq->cq,
1406                                 });
1407                 if (!fdq->wqs[i]) {
1408                         WARN("cannot allocate WQ for drop queue");
1409                         goto error;
1410                 }
1411         }
1412         fdq->ind_table = ibv_exp_create_rwq_ind_table(priv->ctx,
1413                         &(struct ibv_exp_rwq_ind_table_init_attr){
1414                         .pd = priv->pd,
1415                         .log_ind_tbl_size = 0,
1416                         .ind_tbl = fdq->wqs,
1417                         .comp_mask = 0,
1418                         });
1419         if (!fdq->ind_table) {
1420                 WARN("cannot allocate indirection table for drop queue");
1421                 goto error;
1422         }
1423         fdq->qp = ibv_exp_create_qp(priv->ctx,
1424                 &(struct ibv_exp_qp_init_attr){
1425                         .qp_type = IBV_QPT_RAW_PACKET,
1426                         .comp_mask =
1427                                 IBV_EXP_QP_INIT_ATTR_PD |
1428                                 IBV_EXP_QP_INIT_ATTR_PORT |
1429                                 IBV_EXP_QP_INIT_ATTR_RX_HASH,
1430                         .pd = priv->pd,
1431                         .rx_hash_conf = &(struct ibv_exp_rx_hash_conf){
1432                                 .rx_hash_function =
1433                                         IBV_EXP_RX_HASH_FUNC_TOEPLITZ,
1434                                 .rx_hash_key_len = rss_hash_default_key_len,
1435                                 .rx_hash_key = rss_hash_default_key,
1436                                 .rx_hash_fields_mask = 0,
1437                                 .rwq_ind_tbl = fdq->ind_table,
1438                                 },
1439                         .port_num = priv->port,
1440                         });
1441         if (!fdq->qp) {
1442                 WARN("cannot allocate QP for drop queue");
1443                 goto error;
1444         }
1445         priv->flow_drop_queue = fdq;
1446         return 0;
1447 error:
1448         if (fdq->qp)
1449                 claim_zero(ibv_destroy_qp(fdq->qp));
1450         if (fdq->ind_table)
1451                 claim_zero(ibv_exp_destroy_rwq_ind_table(fdq->ind_table));
1452         for (i = 0; i != MLX5_DROP_WQ_N; ++i) {
1453                 if (fdq->wqs[i])
1454                         claim_zero(ibv_exp_destroy_wq(fdq->wqs[i]));
1455         }
1456         if (fdq->cq)
1457                 claim_zero(ibv_destroy_cq(fdq->cq));
1458         if (fdq)
1459                 rte_free(fdq);
1460         priv->flow_drop_queue = NULL;
1461         return -1;
1462 }
1463
1464 /**
1465  * Delete drop queue.
1466  *
1467  * @param priv
1468  *   Pointer to private structure.
1469  */
1470 static void
1471 priv_flow_delete_drop_queue(struct priv *priv)
1472 {
1473         struct rte_flow_drop *fdq = priv->flow_drop_queue;
1474         unsigned int i;
1475
1476         claim_zero(ibv_destroy_qp(fdq->qp));
1477         claim_zero(ibv_exp_destroy_rwq_ind_table(fdq->ind_table));
1478         for (i = 0; i != MLX5_DROP_WQ_N; ++i) {
1479                 assert(fdq->wqs[i]);
1480                 claim_zero(ibv_exp_destroy_wq(fdq->wqs[i]));
1481         }
1482         claim_zero(ibv_destroy_cq(fdq->cq));
1483         rte_free(fdq);
1484         priv->flow_drop_queue = NULL;
1485 }
1486
1487 /**
1488  * Remove all flows.
1489  *
1490  * Called by dev_stop() to remove all flows.
1491  *
1492  * @param priv
1493  *   Pointer to private structure.
1494  */
1495 void
1496 priv_flow_stop(struct priv *priv)
1497 {
1498         struct rte_flow *flow;
1499
1500         for (flow = LIST_FIRST(&priv->flows);
1501              flow;
1502              flow = LIST_NEXT(flow, next)) {
1503                 claim_zero(ibv_exp_destroy_flow(flow->ibv_flow));
1504                 flow->ibv_flow = NULL;
1505                 if (flow->mark) {
1506                         unsigned int n;
1507
1508                         for (n = 0; n < flow->rxqs_n; ++n)
1509                                 (*flow->rxqs)[n]->mark = 0;
1510                 }
1511                 DEBUG("Flow %p removed", (void *)flow);
1512         }
1513         priv_flow_delete_drop_queue(priv);
1514 }
1515
1516 /**
1517  * Add all flows.
1518  *
1519  * @param priv
1520  *   Pointer to private structure.
1521  *
1522  * @return
1523  *   0 on success, a errno value otherwise and rte_errno is set.
1524  */
1525 int
1526 priv_flow_start(struct priv *priv)
1527 {
1528         int ret;
1529         struct rte_flow *flow;
1530
1531         ret = priv_flow_create_drop_queue(priv);
1532         if (ret)
1533                 return -1;
1534         for (flow = LIST_FIRST(&priv->flows);
1535              flow;
1536              flow = LIST_NEXT(flow, next)) {
1537                 struct ibv_qp *qp;
1538
1539                 if (flow->drop)
1540                         qp = priv->flow_drop_queue->qp;
1541                 else
1542                         qp = flow->qp;
1543                 flow->ibv_flow = ibv_exp_create_flow(qp, flow->ibv_attr);
1544                 if (!flow->ibv_flow) {
1545                         DEBUG("Flow %p cannot be applied", (void *)flow);
1546                         rte_errno = EINVAL;
1547                         return rte_errno;
1548                 }
1549                 DEBUG("Flow %p applied", (void *)flow);
1550                 if (flow->mark) {
1551                         unsigned int n;
1552
1553                         for (n = 0; n < flow->rxqs_n; ++n)
1554                                 (*flow->rxqs)[n]->mark = 1;
1555                 }
1556         }
1557         return 0;
1558 }