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
  54 static int
  55 mlx5_flow_create_eth(const struct rte_flow_item *item,
  56                      const void *default_mask,
  57                      void *data);
  58
  59 static int
  60 mlx5_flow_create_vlan(const struct rte_flow_item *item,
  61                       const void *default_mask,
  62                       void *data);
  63
  64 static int
  65 mlx5_flow_create_ipv4(const struct rte_flow_item *item,
  66                       const void *default_mask,
  67                       void *data);
  68
  69 static int
  70 mlx5_flow_create_ipv6(const struct rte_flow_item *item,
  71                       const void *default_mask,
  72                       void *data);
  73
  74 static int
  75 mlx5_flow_create_udp(const struct rte_flow_item *item,
  76                      const void *default_mask,
  77                      void *data);
  78
  79 static int
  80 mlx5_flow_create_tcp(const struct rte_flow_item *item,
  81                      const void *default_mask,
  82                      void *data);
  83
  84 static int
  85 mlx5_flow_create_vxlan(const struct rte_flow_item *item,
  86                        const void *default_mask,
  87                        void *data);
  88
  89 struct rte_flow {
  90         LIST_ENTRY(rte_flow) next; /**< Pointer to the next flow structure. */
  91         struct ibv_exp_flow_attr *ibv_attr; /**< Pointer to Verbs attributes. */
  92         struct ibv_exp_rwq_ind_table *ind_table; /**< Indirection table. */
  93         struct ibv_qp *qp; /**< Verbs queue pair. */
  94         struct ibv_exp_flow *ibv_flow; /**< Verbs flow. */
  95         struct ibv_exp_wq *wq; /**< Verbs work queue. */
  96         struct ibv_cq *cq; /**< Verbs completion queue. */
  97         struct rxq *rxq; /**< Pointer to the queue, NULL if drop queue. */
  98 };
  99
 100 /** Static initializer for items. */
 101 #define ITEMS(...) \
 102         (const enum rte_flow_item_type []){ \
 103                 __VA_ARGS__, RTE_FLOW_ITEM_TYPE_END, \
 104         }
 105
 106 /** Structure to generate a simple graph of layers supported by the NIC. */
 107 struct mlx5_flow_items {
 108         /** List of possible actions for these items. */
 109         const enum rte_flow_action_type *const actions;
 110         /** Bit-masks corresponding to the possibilities for the item. */
 111         const void *mask;
 112         /** Bit-masks size in bytes. */
 113         const unsigned int mask_sz;
 114         /**
 115          * Conversion function from rte_flow to NIC specific flow.
 116          *
 117          * @param item
 118          *   rte_flow item to convert.
 119          * @param default_mask
 120          *   Default bit-masks to use when item->mask is not provided.
 121          * @param data
 122          *   Internal structure to store the conversion.
 123          *
 124          * @return
 125          *   0 on success, negative value otherwise.
 126          */
 127         int (*convert)(const struct rte_flow_item *item,
 128                        const void *default_mask,
 129                        void *data);
 130         /** Size in bytes of the destination structure. */
 131         const unsigned int dst_sz;
 132         /** List of possible following items.  */
 133         const enum rte_flow_item_type *const items;
 134 };
 135
 136 /** Valid action for this PMD. */
 137 static const enum rte_flow_action_type valid_actions[] = {
 138         RTE_FLOW_ACTION_TYPE_DROP,
 139         RTE_FLOW_ACTION_TYPE_QUEUE,
 140         RTE_FLOW_ACTION_TYPE_END,
 141 };
 142
 143 /** Graph of supported items and associated actions. */
 144 static const struct mlx5_flow_items mlx5_flow_items[] = {
 145         [RTE_FLOW_ITEM_TYPE_END] = {
 146                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_ETH,
 147                                RTE_FLOW_ITEM_TYPE_VXLAN),
 148         },
 149         [RTE_FLOW_ITEM_TYPE_ETH] = {
 150                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_VLAN,
 151                                RTE_FLOW_ITEM_TYPE_IPV4,
 152                                RTE_FLOW_ITEM_TYPE_IPV6),
 153                 .actions = valid_actions,
 154                 .mask = &(const struct rte_flow_item_eth){
 155                         .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 156                         .src.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 157                 },
 158                 .mask_sz = sizeof(struct rte_flow_item_eth),
 159                 .convert = mlx5_flow_create_eth,
 160                 .dst_sz = sizeof(struct ibv_exp_flow_spec_eth),
 161         },
 162         [RTE_FLOW_ITEM_TYPE_VLAN] = {
 163                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_IPV4,
 164                                RTE_FLOW_ITEM_TYPE_IPV6),
 165                 .actions = valid_actions,
 166                 .mask = &(const struct rte_flow_item_vlan){
 167                         .tci = -1,
 168                 },
 169                 .mask_sz = sizeof(struct rte_flow_item_vlan),
 170                 .convert = mlx5_flow_create_vlan,
 171                 .dst_sz = 0,
 172         },
 173         [RTE_FLOW_ITEM_TYPE_IPV4] = {
 174                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP,
 175                                RTE_FLOW_ITEM_TYPE_TCP),
 176                 .actions = valid_actions,
 177                 .mask = &(const struct rte_flow_item_ipv4){
 178                         .hdr = {
 179                                 .src_addr = -1,
 180                                 .dst_addr = -1,
 181                         },
 182                 },
 183                 .mask_sz = sizeof(struct rte_flow_item_ipv4),
 184                 .convert = mlx5_flow_create_ipv4,
 185                 .dst_sz = sizeof(struct ibv_exp_flow_spec_ipv4),
 186         },
 187         [RTE_FLOW_ITEM_TYPE_IPV6] = {
 188                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP,
 189                                RTE_FLOW_ITEM_TYPE_TCP),
 190                 .actions = valid_actions,
 191                 .mask = &(const struct rte_flow_item_ipv6){
 192                         .hdr = {
 193                                 .src_addr = {
 194                                         0xff, 0xff, 0xff, 0xff,
 195                                         0xff, 0xff, 0xff, 0xff,
 196                                         0xff, 0xff, 0xff, 0xff,
 197                                         0xff, 0xff, 0xff, 0xff,
 198                                 },
 199                                 .dst_addr = {
 200                                         0xff, 0xff, 0xff, 0xff,
 201                                         0xff, 0xff, 0xff, 0xff,
 202                                         0xff, 0xff, 0xff, 0xff,
 203                                         0xff, 0xff, 0xff, 0xff,
 204                                 },
 205                         },
 206                 },
 207                 .mask_sz = sizeof(struct rte_flow_item_ipv6),
 208                 .convert = mlx5_flow_create_ipv6,
 209                 .dst_sz = sizeof(struct ibv_exp_flow_spec_ipv6),
 210         },
 211         [RTE_FLOW_ITEM_TYPE_UDP] = {
 212                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_VXLAN),
 213                 .actions = valid_actions,
 214                 .mask = &(const struct rte_flow_item_udp){
 215                         .hdr = {
 216                                 .src_port = -1,
 217                                 .dst_port = -1,
 218                         },
 219                 },
 220                 .mask_sz = sizeof(struct rte_flow_item_udp),
 221                 .convert = mlx5_flow_create_udp,
 222                 .dst_sz = sizeof(struct ibv_exp_flow_spec_tcp_udp),
 223         },
 224         [RTE_FLOW_ITEM_TYPE_TCP] = {
 225                 .actions = valid_actions,
 226                 .mask = &(const struct rte_flow_item_tcp){
 227                         .hdr = {
 228                                 .src_port = -1,
 229                                 .dst_port = -1,
 230                         },
 231                 },
 232                 .mask_sz = sizeof(struct rte_flow_item_tcp),
 233                 .convert = mlx5_flow_create_tcp,
 234                 .dst_sz = sizeof(struct ibv_exp_flow_spec_tcp_udp),
 235         },
 236         [RTE_FLOW_ITEM_TYPE_VXLAN] = {
 237                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_ETH),
 238                 .actions = valid_actions,
 239                 .mask = &(const struct rte_flow_item_vxlan){
 240                         .vni = "\xff\xff\xff",
 241                 },
 242                 .mask_sz = sizeof(struct rte_flow_item_vxlan),
 243                 .convert = mlx5_flow_create_vxlan,
 244                 .dst_sz = sizeof(struct ibv_exp_flow_spec_tunnel),
 245         },
 246 };
 247
 248 /** Structure to pass to the conversion function. */
 249 struct mlx5_flow {
 250         struct ibv_exp_flow_attr *ibv_attr; /**< Verbs attribute. */
 251         unsigned int offset; /**< Offset in bytes in the ibv_attr buffer. */
 252         uint32_t inner; /**< Set once VXLAN is encountered. */
 253 };
 254
 255 struct mlx5_flow_action {
 256         uint32_t queue:1; /**< Target is a receive queue. */
 257         uint32_t drop:1; /**< Target is a drop queue. */
 258         uint32_t queue_id; /**< Identifier of the queue. */
 259 };
 260
 261 /**
 262  * Check support for a given item.
 263  *
 264  * @param item[in]
 265  *   Item specification.
 266  * @param mask[in]
 267  *   Bit-masks covering supported fields to compare with spec, last and mask in
 268  *   \item.
 269  * @param size
 270  *   Bit-Mask size in bytes.
 271  *
 272  * @return
 273  *   0 on success.
 274  */
 275 static int
 276 mlx5_flow_item_validate(const struct rte_flow_item *item,
 277                         const uint8_t *mask, unsigned int size)
 278 {
 279         int ret = 0;
 280
 281         if (!item->spec && (item->mask || item->last))
 282                 return -1;
 283         if (item->spec && !item->mask) {
 284                 unsigned int i;
 285                 const uint8_t *spec = item->spec;
 286
 287                 for (i = 0; i < size; ++i)
 288                         if ((spec[i] | mask[i]) != mask[i])
 289                                 return -1;
 290         }
 291         if (item->last && !item->mask) {
 292                 unsigned int i;
 293                 const uint8_t *spec = item->last;
 294
 295                 for (i = 0; i < size; ++i)
 296                         if ((spec[i] | mask[i]) != mask[i])
 297                                 return -1;
 298         }
 299         if (item->mask) {
 300                 unsigned int i;
 301                 const uint8_t *spec = item->mask;
 302
 303                 for (i = 0; i < size; ++i)
 304                         if ((spec[i] | mask[i]) != mask[i])
 305                                 return -1;
 306         }
 307         if (item->spec && item->last) {
 308                 uint8_t spec[size];
 309                 uint8_t last[size];
 310                 const uint8_t *apply = mask;
 311                 unsigned int i;
 312
 313                 if (item->mask)
 314                         apply = item->mask;
 315                 for (i = 0; i < size; ++i) {
 316                         spec[i] = ((const uint8_t *)item->spec)[i] & apply[i];
 317                         last[i] = ((const uint8_t *)item->last)[i] & apply[i];
 318                 }
 319                 ret = memcmp(spec, last, size);
 320         }
 321         return ret;
 322 }
 323
 324 /**
 325  * Validate a flow supported by the NIC.
 326  *
 327  * @param priv
 328  *   Pointer to private structure.
 329  * @param[in] attr
 330  *   Flow rule attributes.
 331  * @param[in] pattern
 332  *   Pattern specification (list terminated by the END pattern item).
 333  * @param[in] actions
 334  *   Associated actions (list terminated by the END action).
 335  * @param[out] error
 336  *   Perform verbose error reporting if not NULL.
 337  * @param[in, out] flow
 338  *   Flow structure to update.
 339  *
 340  * @return
 341  *   0 on success, a negative errno value otherwise and rte_errno is set.
 342  */
 343 static int
 344 priv_flow_validate(struct priv *priv,
 345                    const struct rte_flow_attr *attr,
 346                    const struct rte_flow_item items[],
 347                    const struct rte_flow_action actions[],
 348                    struct rte_flow_error *error,
 349                    struct mlx5_flow *flow)
 350 {
 351         const struct mlx5_flow_items *cur_item = mlx5_flow_items;
 352         struct mlx5_flow_action action = {
 353                 .queue = 0,
 354                 .drop = 0,
 355         };
 356
 357         (void)priv;
 358         if (attr->group) {
 359                 rte_flow_error_set(error, ENOTSUP,
 360                                    RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
 361                                    NULL,
 362                                    "groups are not supported");
 363                 return -rte_errno;
 364         }
 365         if (attr->priority) {
 366                 rte_flow_error_set(error, ENOTSUP,
 367                                    RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
 368                                    NULL,
 369                                    "priorities are not supported");
 370                 return -rte_errno;
 371         }
 372         if (attr->egress) {
 373                 rte_flow_error_set(error, ENOTSUP,
 374                                    RTE_FLOW_ERROR_TYPE_ATTR_EGRESS,
 375                                    NULL,
 376                                    "egress is not supported");
 377                 return -rte_errno;
 378         }
 379         if (!attr->ingress) {
 380                 rte_flow_error_set(error, ENOTSUP,
 381                                    RTE_FLOW_ERROR_TYPE_ATTR_INGRESS,
 382                                    NULL,
 383                                    "only ingress is supported");
 384                 return -rte_errno;
 385         }
 386         for (; items->type != RTE_FLOW_ITEM_TYPE_END; ++items) {
 387                 const struct mlx5_flow_items *token = NULL;
 388                 unsigned int i;
 389                 int err;
 390
 391                 if (items->type == RTE_FLOW_ITEM_TYPE_VOID)
 392                         continue;
 393                 /* Handle special situation for VLAN. */
 394                 if (items->type == RTE_FLOW_ITEM_TYPE_VLAN) {
 395                         if (((const struct rte_flow_item_vlan *)items)->tci >
 396                             ETHER_MAX_VLAN_ID) {
 397                                 rte_flow_error_set(error, ENOTSUP,
 398                                                    RTE_FLOW_ERROR_TYPE_ITEM,
 399                                                    items,
 400                                                    "wrong VLAN id value");
 401                                 return -rte_errno;
 402                         }
 403                 }
 404                 for (i = 0;
 405                      cur_item->items &&
 406                      cur_item->items[i] != RTE_FLOW_ITEM_TYPE_END;
 407                      ++i) {
 408                         if (cur_item->items[i] == items->type) {
 409                                 token = &mlx5_flow_items[items->type];
 410                                 break;
 411                         }
 412                 }
 413                 if (!token)
 414                         goto exit_item_not_supported;
 415                 cur_item = token;
 416                 err = mlx5_flow_item_validate(items,
 417                                               (const uint8_t *)cur_item->mask,
 418                                               sizeof(cur_item->mask_sz));
 419                 if (err)
 420                         goto exit_item_not_supported;
 421                 if (flow->ibv_attr && cur_item->convert) {
 422                         err = cur_item->convert(items, cur_item->mask, flow);
 423                         if (err)
 424                                 goto exit_item_not_supported;
 425                 }
 426                 flow->offset += cur_item->dst_sz;
 427         }
 428         for (; actions->type != RTE_FLOW_ACTION_TYPE_END; ++actions) {
 429                 if (actions->type == RTE_FLOW_ACTION_TYPE_VOID) {
 430                         continue;
 431                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_DROP) {
 432                         action.drop = 1;
 433                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_QUEUE) {
 434                         const struct rte_flow_action_queue *queue =
 435                                 (const struct rte_flow_action_queue *)
 436                                 actions->conf;
 437
 438                         if (!queue || (queue->index > (priv->rxqs_n - 1)))
 439                                 goto exit_action_not_supported;
 440                         action.queue = 1;
 441                 } else {
 442                         goto exit_action_not_supported;
 443                 }
 444         }
 445         if (!action.queue && !action.drop) {
 446                 rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_HANDLE,
 447                                    NULL, "no valid action");
 448                 return -rte_errno;
 449         }
 450         return 0;
 451 exit_item_not_supported:
 452         rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 453                            items, "item not supported");
 454         return -rte_errno;
 455 exit_action_not_supported:
 456         rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
 457                            actions, "action not supported");
 458         return -rte_errno;
 459 }
 460
 461 /**
 462  * Validate a flow supported by the NIC.
 463  *
 464  * @see rte_flow_validate()
 465  * @see rte_flow_ops
 466  */
 467 int
 468 mlx5_flow_validate(struct rte_eth_dev *dev,
 469                    const struct rte_flow_attr *attr,
 470                    const struct rte_flow_item items[],
 471                    const struct rte_flow_action actions[],
 472                    struct rte_flow_error *error)
 473 {
 474         struct priv *priv = dev->data->dev_private;
 475         int ret;
 476         struct mlx5_flow flow = { .offset = sizeof(struct ibv_exp_flow_attr) };
 477
 478         priv_lock(priv);
 479         ret = priv_flow_validate(priv, attr, items, actions, error, &flow);
 480         priv_unlock(priv);
 481         return ret;
 482 }
 483
 484 /**
 485  * Convert Ethernet item to Verbs specification.
 486  *
 487  * @param item[in]
 488  *   Item specification.
 489  * @param default_mask[in]
 490  *   Default bit-masks to use when item->mask is not provided.
 491  * @param data[in, out]
 492  *   User structure.
 493  */
 494 static int
 495 mlx5_flow_create_eth(const struct rte_flow_item *item,
 496                      const void *default_mask,
 497                      void *data)
 498 {
 499         const struct rte_flow_item_eth *spec = item->spec;
 500         const struct rte_flow_item_eth *mask = item->mask;
 501         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 502         struct ibv_exp_flow_spec_eth *eth;
 503         const unsigned int eth_size = sizeof(struct ibv_exp_flow_spec_eth);
 504         unsigned int i;
 505
 506         ++flow->ibv_attr->num_of_specs;
 507         flow->ibv_attr->priority = 2;
 508         eth = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 509         *eth = (struct ibv_exp_flow_spec_eth) {
 510                 .type = flow->inner | IBV_EXP_FLOW_SPEC_ETH,
 511                 .size = eth_size,
 512         };
 513         if (!spec)
 514                 return 0;
 515         if (!mask)
 516                 mask = default_mask;
 517         memcpy(eth->val.dst_mac, spec->dst.addr_bytes, ETHER_ADDR_LEN);
 518         memcpy(eth->val.src_mac, spec->src.addr_bytes, ETHER_ADDR_LEN);
 519         memcpy(eth->mask.dst_mac, mask->dst.addr_bytes, ETHER_ADDR_LEN);
 520         memcpy(eth->mask.src_mac, mask->src.addr_bytes, ETHER_ADDR_LEN);
 521         /* Remove unwanted bits from values. */
 522         for (i = 0; i < ETHER_ADDR_LEN; ++i) {
 523                 eth->val.dst_mac[i] &= eth->mask.dst_mac[i];
 524                 eth->val.src_mac[i] &= eth->mask.src_mac[i];
 525         }
 526         return 0;
 527 }
 528
 529 /**
 530  * Convert VLAN item to Verbs specification.
 531  *
 532  * @param item[in]
 533  *   Item specification.
 534  * @param default_mask[in]
 535  *   Default bit-masks to use when item->mask is not provided.
 536  * @param data[in, out]
 537  *   User structure.
 538  */
 539 static int
 540 mlx5_flow_create_vlan(const struct rte_flow_item *item,
 541                       const void *default_mask,
 542                       void *data)
 543 {
 544         const struct rte_flow_item_vlan *spec = item->spec;
 545         const struct rte_flow_item_vlan *mask = item->mask;
 546         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 547         struct ibv_exp_flow_spec_eth *eth;
 548         const unsigned int eth_size = sizeof(struct ibv_exp_flow_spec_eth);
 549
 550         eth = (void *)((uintptr_t)flow->ibv_attr + flow->offset - eth_size);
 551         if (!spec)
 552                 return 0;
 553         if (!mask)
 554                 mask = default_mask;
 555         eth->val.vlan_tag = spec->tci;
 556         eth->mask.vlan_tag = mask->tci;
 557         eth->val.vlan_tag &= eth->mask.vlan_tag;
 558         return 0;
 559 }
 560
 561 /**
 562  * Convert IPv4 item to Verbs specification.
 563  *
 564  * @param item[in]
 565  *   Item specification.
 566  * @param default_mask[in]
 567  *   Default bit-masks to use when item->mask is not provided.
 568  * @param data[in, out]
 569  *   User structure.
 570  */
 571 static int
 572 mlx5_flow_create_ipv4(const struct rte_flow_item *item,
 573                       const void *default_mask,
 574                       void *data)
 575 {
 576         const struct rte_flow_item_ipv4 *spec = item->spec;
 577         const struct rte_flow_item_ipv4 *mask = item->mask;
 578         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 579         struct ibv_exp_flow_spec_ipv4 *ipv4;
 580         unsigned int ipv4_size = sizeof(struct ibv_exp_flow_spec_ipv4);
 581
 582         ++flow->ibv_attr->num_of_specs;
 583         flow->ibv_attr->priority = 1;
 584         ipv4 = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 585         *ipv4 = (struct ibv_exp_flow_spec_ipv4) {
 586                 .type = flow->inner | IBV_EXP_FLOW_SPEC_IPV4,
 587                 .size = ipv4_size,
 588         };
 589         if (!spec)
 590                 return 0;
 591         if (!mask)
 592                 mask = default_mask;
 593         ipv4->val = (struct ibv_exp_flow_ipv4_filter){
 594                 .src_ip = spec->hdr.src_addr,
 595                 .dst_ip = spec->hdr.dst_addr,
 596         };
 597         ipv4->mask = (struct ibv_exp_flow_ipv4_filter){
 598                 .src_ip = mask->hdr.src_addr,
 599                 .dst_ip = mask->hdr.dst_addr,
 600         };
 601         /* Remove unwanted bits from values. */
 602         ipv4->val.src_ip &= ipv4->mask.src_ip;
 603         ipv4->val.dst_ip &= ipv4->mask.dst_ip;
 604         return 0;
 605 }
 606
 607 /**
 608  * Convert IPv6 item to Verbs specification.
 609  *
 610  * @param item[in]
 611  *   Item specification.
 612  * @param default_mask[in]
 613  *   Default bit-masks to use when item->mask is not provided.
 614  * @param data[in, out]
 615  *   User structure.
 616  */
 617 static int
 618 mlx5_flow_create_ipv6(const struct rte_flow_item *item,
 619                       const void *default_mask,
 620                       void *data)
 621 {
 622         const struct rte_flow_item_ipv6 *spec = item->spec;
 623         const struct rte_flow_item_ipv6 *mask = item->mask;
 624         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 625         struct ibv_exp_flow_spec_ipv6 *ipv6;
 626         unsigned int ipv6_size = sizeof(struct ibv_exp_flow_spec_ipv6);
 627         unsigned int i;
 628
 629         ++flow->ibv_attr->num_of_specs;
 630         flow->ibv_attr->priority = 1;
 631         ipv6 = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 632         *ipv6 = (struct ibv_exp_flow_spec_ipv6) {
 633                 .type = flow->inner | IBV_EXP_FLOW_SPEC_IPV6,
 634                 .size = ipv6_size,
 635         };
 636         if (!spec)
 637                 return 0;
 638         if (!mask)
 639                 mask = default_mask;
 640         memcpy(ipv6->val.src_ip, spec->hdr.src_addr,
 641                RTE_DIM(ipv6->val.src_ip));
 642         memcpy(ipv6->val.dst_ip, spec->hdr.dst_addr,
 643                RTE_DIM(ipv6->val.dst_ip));
 644         memcpy(ipv6->mask.src_ip, mask->hdr.src_addr,
 645                RTE_DIM(ipv6->mask.src_ip));
 646         memcpy(ipv6->mask.dst_ip, mask->hdr.dst_addr,
 647                RTE_DIM(ipv6->mask.dst_ip));
 648         /* Remove unwanted bits from values. */
 649         for (i = 0; i < RTE_DIM(ipv6->val.src_ip); ++i) {
 650                 ipv6->val.src_ip[i] &= ipv6->mask.src_ip[i];
 651                 ipv6->val.dst_ip[i] &= ipv6->mask.dst_ip[i];
 652         }
 653         return 0;
 654 }
 655
 656 /**
 657  * Convert UDP item to Verbs specification.
 658  *
 659  * @param item[in]
 660  *   Item specification.
 661  * @param default_mask[in]
 662  *   Default bit-masks to use when item->mask is not provided.
 663  * @param data[in, out]
 664  *   User structure.
 665  */
 666 static int
 667 mlx5_flow_create_udp(const struct rte_flow_item *item,
 668                      const void *default_mask,
 669                      void *data)
 670 {
 671         const struct rte_flow_item_udp *spec = item->spec;
 672         const struct rte_flow_item_udp *mask = item->mask;
 673         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 674         struct ibv_exp_flow_spec_tcp_udp *udp;
 675         unsigned int udp_size = sizeof(struct ibv_exp_flow_spec_tcp_udp);
 676
 677         ++flow->ibv_attr->num_of_specs;
 678         flow->ibv_attr->priority = 0;
 679         udp = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 680         *udp = (struct ibv_exp_flow_spec_tcp_udp) {
 681                 .type = flow->inner | IBV_EXP_FLOW_SPEC_UDP,
 682                 .size = udp_size,
 683         };
 684         if (!spec)
 685                 return 0;
 686         if (!mask)
 687                 mask = default_mask;
 688         udp->val.dst_port = spec->hdr.dst_port;
 689         udp->val.src_port = spec->hdr.src_port;
 690         udp->mask.dst_port = mask->hdr.dst_port;
 691         udp->mask.src_port = mask->hdr.src_port;
 692         /* Remove unwanted bits from values. */
 693         udp->val.src_port &= udp->mask.src_port;
 694         udp->val.dst_port &= udp->mask.dst_port;
 695         return 0;
 696 }
 697
 698 /**
 699  * Convert TCP item to Verbs specification.
 700  *
 701  * @param item[in]
 702  *   Item specification.
 703  * @param default_mask[in]
 704  *   Default bit-masks to use when item->mask is not provided.
 705  * @param data[in, out]
 706  *   User structure.
 707  */
 708 static int
 709 mlx5_flow_create_tcp(const struct rte_flow_item *item,
 710                      const void *default_mask,
 711                      void *data)
 712 {
 713         const struct rte_flow_item_tcp *spec = item->spec;
 714         const struct rte_flow_item_tcp *mask = item->mask;
 715         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 716         struct ibv_exp_flow_spec_tcp_udp *tcp;
 717         unsigned int tcp_size = sizeof(struct ibv_exp_flow_spec_tcp_udp);
 718
 719         ++flow->ibv_attr->num_of_specs;
 720         flow->ibv_attr->priority = 0;
 721         tcp = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 722         *tcp = (struct ibv_exp_flow_spec_tcp_udp) {
 723                 .type = flow->inner | IBV_EXP_FLOW_SPEC_TCP,
 724                 .size = tcp_size,
 725         };
 726         if (!spec)
 727                 return 0;
 728         if (!mask)
 729                 mask = default_mask;
 730         tcp->val.dst_port = spec->hdr.dst_port;
 731         tcp->val.src_port = spec->hdr.src_port;
 732         tcp->mask.dst_port = mask->hdr.dst_port;
 733         tcp->mask.src_port = mask->hdr.src_port;
 734         /* Remove unwanted bits from values. */
 735         tcp->val.src_port &= tcp->mask.src_port;
 736         tcp->val.dst_port &= tcp->mask.dst_port;
 737         return 0;
 738 }
 739
 740 /**
 741  * Convert VXLAN item to Verbs specification.
 742  *
 743  * @param item[in]
 744  *   Item specification.
 745  * @param default_mask[in]
 746  *   Default bit-masks to use when item->mask is not provided.
 747  * @param data[in, out]
 748  *   User structure.
 749  */
 750 static int
 751 mlx5_flow_create_vxlan(const struct rte_flow_item *item,
 752                        const void *default_mask,
 753                        void *data)
 754 {
 755         const struct rte_flow_item_vxlan *spec = item->spec;
 756         const struct rte_flow_item_vxlan *mask = item->mask;
 757         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 758         struct ibv_exp_flow_spec_tunnel *vxlan;
 759         unsigned int size = sizeof(struct ibv_exp_flow_spec_tunnel);
 760         union vni {
 761                 uint32_t vlan_id;
 762                 uint8_t vni[4];
 763         } id;
 764
 765         ++flow->ibv_attr->num_of_specs;
 766         flow->ibv_attr->priority = 0;
 767         id.vni[0] = 0;
 768         vxlan = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 769         *vxlan = (struct ibv_exp_flow_spec_tunnel) {
 770                 .type = flow->inner | IBV_EXP_FLOW_SPEC_VXLAN_TUNNEL,
 771                 .size = size,
 772         };
 773         flow->inner = IBV_EXP_FLOW_SPEC_INNER;
 774         if (!spec)
 775                 return 0;
 776         if (!mask)
 777                 mask = default_mask;
 778         memcpy(&id.vni[1], spec->vni, 3);
 779         vxlan->val.tunnel_id = id.vlan_id;
 780         memcpy(&id.vni[1], mask->vni, 3);
 781         vxlan->mask.tunnel_id = id.vlan_id;
 782         /* Remove unwanted bits from values. */
 783         vxlan->val.tunnel_id &= vxlan->mask.tunnel_id;
 784         return 0;
 785 }
 786
 787 /**
 788  * Complete flow rule creation.
 789  *
 790  * @param priv
 791  *   Pointer to private structure.
 792  * @param ibv_attr
 793  *   Verbs flow attributes.
 794  * @param action
 795  *   Target action structure.
 796  * @param[out] error
 797  *   Perform verbose error reporting if not NULL.
 798  *
 799  * @return
 800  *   A flow if the rule could be created.
 801  */
 802 static struct rte_flow *
 803 priv_flow_create_action_queue(struct priv *priv,
 804                               struct ibv_exp_flow_attr *ibv_attr,
 805                               struct mlx5_flow_action *action,
 806                               struct rte_flow_error *error)
 807 {
 808         struct rxq_ctrl *rxq;
 809         struct rte_flow *rte_flow;
 810
 811         assert(priv->pd);
 812         assert(priv->ctx);
 813         rte_flow = rte_calloc(__func__, 1, sizeof(*rte_flow), 0);
 814         if (!rte_flow) {
 815                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
 816                                    NULL, "cannot allocate flow memory");
 817                 return NULL;
 818         }
 819         if (action->drop) {
 820                 rte_flow->cq =
 821                         ibv_exp_create_cq(priv->ctx, 1, NULL, NULL, 0,
 822                                           &(struct ibv_exp_cq_init_attr){
 823                                                   .comp_mask = 0,
 824                                           });
 825                 if (!rte_flow->cq) {
 826                         rte_flow_error_set(error, ENOMEM,
 827                                            RTE_FLOW_ERROR_TYPE_HANDLE,
 828                                            NULL, "cannot allocate CQ");
 829                         goto error;
 830                 }
 831                 rte_flow->wq = ibv_exp_create_wq(priv->ctx,
 832                                                  &(struct ibv_exp_wq_init_attr){
 833                                                  .wq_type = IBV_EXP_WQT_RQ,
 834                                                  .max_recv_wr = 1,
 835                                                  .max_recv_sge = 1,
 836                                                  .pd = priv->pd,
 837                                                  .cq = rte_flow->cq,
 838                                                  });
 839         } else {
 840                 rxq = container_of((*priv->rxqs)[action->queue_id],
 841                                    struct rxq_ctrl, rxq);
 842                 rte_flow->rxq = &rxq->rxq;
 843                 rte_flow->wq = rxq->wq;
 844         }
 845         rte_flow->ibv_attr = ibv_attr;
 846         rte_flow->ind_table = ibv_exp_create_rwq_ind_table(
 847                 priv->ctx,
 848                 &(struct ibv_exp_rwq_ind_table_init_attr){
 849                         .pd = priv->pd,
 850                         .log_ind_tbl_size = 0,
 851                         .ind_tbl = &rte_flow->wq,
 852                         .comp_mask = 0,
 853                 });
 854         if (!rte_flow->ind_table) {
 855                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
 856                                    NULL, "cannot allocate indirection table");
 857                 goto error;
 858         }
 859         rte_flow->qp = ibv_exp_create_qp(
 860                 priv->ctx,
 861                 &(struct ibv_exp_qp_init_attr){
 862                         .qp_type = IBV_QPT_RAW_PACKET,
 863                         .comp_mask =
 864                                 IBV_EXP_QP_INIT_ATTR_PD |
 865                                 IBV_EXP_QP_INIT_ATTR_PORT |
 866                                 IBV_EXP_QP_INIT_ATTR_RX_HASH,
 867                         .pd = priv->pd,
 868                         .rx_hash_conf = &(struct ibv_exp_rx_hash_conf){
 869                                 .rx_hash_function =
 870                                         IBV_EXP_RX_HASH_FUNC_TOEPLITZ,
 871                                 .rx_hash_key_len = rss_hash_default_key_len,
 872                                 .rx_hash_key = rss_hash_default_key,
 873                                 .rx_hash_fields_mask = 0,
 874                                 .rwq_ind_tbl = rte_flow->ind_table,
 875                         },
 876                         .port_num = priv->port,
 877                 });
 878         if (!rte_flow->qp) {
 879                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
 880                                    NULL, "cannot allocate QP");
 881                 goto error;
 882         }
 883         rte_flow->ibv_flow = ibv_exp_create_flow(rte_flow->qp,
 884                                                  rte_flow->ibv_attr);
 885         if (!rte_flow->ibv_flow) {
 886                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
 887                                    NULL, "flow rule creation failure");
 888                 goto error;
 889         }
 890         return rte_flow;
 891 error:
 892         assert(rte_flow);
 893         if (rte_flow->qp)
 894                 ibv_destroy_qp(rte_flow->qp);
 895         if (rte_flow->ind_table)
 896                 ibv_exp_destroy_rwq_ind_table(rte_flow->ind_table);
 897         if (!rte_flow->rxq && rte_flow->wq)
 898                 ibv_exp_destroy_wq(rte_flow->wq);
 899         if (!rte_flow->rxq && rte_flow->cq)
 900                 ibv_destroy_cq(rte_flow->cq);
 901         rte_free(rte_flow->ibv_attr);
 902         rte_free(rte_flow);
 903         return NULL;
 904 }
 905
 906 /**
 907  * Convert a flow.
 908  *
 909  * @param priv
 910  *   Pointer to private structure.
 911  * @param[in] attr
 912  *   Flow rule attributes.
 913  * @param[in] pattern
 914  *   Pattern specification (list terminated by the END pattern item).
 915  * @param[in] actions
 916  *   Associated actions (list terminated by the END action).
 917  * @param[out] error
 918  *   Perform verbose error reporting if not NULL.
 919  *
 920  * @return
 921  *   A flow on success, NULL otherwise.
 922  */
 923 static struct rte_flow *
 924 priv_flow_create(struct priv *priv,
 925                  const struct rte_flow_attr *attr,
 926                  const struct rte_flow_item items[],
 927                  const struct rte_flow_action actions[],
 928                  struct rte_flow_error *error)
 929 {
 930         struct rte_flow *rte_flow;
 931         struct mlx5_flow_action action;
 932         struct mlx5_flow flow = { .offset = sizeof(struct ibv_exp_flow_attr), };
 933         int err;
 934
 935         err = priv_flow_validate(priv, attr, items, actions, error, &flow);
 936         if (err)
 937                 goto exit;
 938         flow.ibv_attr = rte_malloc(__func__, flow.offset, 0);
 939         flow.offset = sizeof(struct ibv_exp_flow_attr);
 940         if (!flow.ibv_attr) {
 941                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
 942                                    NULL, "cannot allocate ibv_attr memory");
 943                 goto exit;
 944         }
 945         *flow.ibv_attr = (struct ibv_exp_flow_attr){
 946                 .type = IBV_EXP_FLOW_ATTR_NORMAL,
 947                 .size = sizeof(struct ibv_exp_flow_attr),
 948                 .priority = attr->priority,
 949                 .num_of_specs = 0,
 950                 .port = 0,
 951                 .flags = 0,
 952                 .reserved = 0,
 953         };
 954         flow.inner = 0;
 955         claim_zero(priv_flow_validate(priv, attr, items, actions,
 956                                       error, &flow));
 957         action = (struct mlx5_flow_action){
 958                 .queue = 0,
 959                 .drop = 0,
 960         };
 961         for (; actions->type != RTE_FLOW_ACTION_TYPE_END; ++actions) {
 962                 if (actions->type == RTE_FLOW_ACTION_TYPE_VOID) {
 963                         continue;
 964                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_QUEUE) {
 965                         action.queue = 1;
 966                         action.queue_id =
 967                                 ((const struct rte_flow_action_queue *)
 968                                  actions->conf)->index;
 969                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_DROP) {
 970                         action.drop = 1;
 971                 } else {
 972                         rte_flow_error_set(error, ENOTSUP,
 973                                            RTE_FLOW_ERROR_TYPE_ACTION,
 974                                            actions, "unsupported action");
 975                         goto exit;
 976                 }
 977         }
 978         rte_flow = priv_flow_create_action_queue(priv, flow.ibv_attr,
 979                                                  &action, error);
 980         return rte_flow;
 981 exit:
 982         rte_free(flow.ibv_attr);
 983         return NULL;
 984 }
 985
 986 /**
 987  * Create a flow.
 988  *
 989  * @see rte_flow_create()
 990  * @see rte_flow_ops
 991  */
 992 struct rte_flow *
 993 mlx5_flow_create(struct rte_eth_dev *dev,
 994                  const struct rte_flow_attr *attr,
 995                  const struct rte_flow_item items[],
 996                  const struct rte_flow_action actions[],
 997                  struct rte_flow_error *error)
 998 {
 999         struct priv *priv = dev->data->dev_private;
1000         struct rte_flow *flow;
1001
1002         priv_lock(priv);
1003         flow = priv_flow_create(priv, attr, items, actions, error);
1004         if (flow) {
1005                 LIST_INSERT_HEAD(&priv->flows, flow, next);
1006                 DEBUG("Flow created %p", (void *)flow);
1007         }
1008         priv_unlock(priv);
1009         return flow;
1010 }
1011
1012 /**
1013  * Destroy a flow.
1014  *
1015  * @param priv
1016  *   Pointer to private structure.
1017  * @param[in] flow
1018  *   Flow to destroy.
1019  */
1020 static void
1021 priv_flow_destroy(struct priv *priv,
1022                   struct rte_flow *flow)
1023 {
1024         (void)priv;
1025         LIST_REMOVE(flow, next);
1026         if (flow->ibv_flow)
1027                 claim_zero(ibv_exp_destroy_flow(flow->ibv_flow));
1028         if (flow->qp)
1029                 claim_zero(ibv_destroy_qp(flow->qp));
1030         if (flow->ind_table)
1031                 claim_zero(ibv_exp_destroy_rwq_ind_table(flow->ind_table));
1032         if (!flow->rxq && flow->wq)
1033                 claim_zero(ibv_exp_destroy_wq(flow->wq));
1034         if (!flow->rxq && flow->cq)
1035                 claim_zero(ibv_destroy_cq(flow->cq));
1036         rte_free(flow->ibv_attr);
1037         DEBUG("Flow destroyed %p", (void *)flow);
1038         rte_free(flow);
1039 }
1040
1041 /**
1042  * Destroy a flow.
1043  *
1044  * @see rte_flow_destroy()
1045  * @see rte_flow_ops
1046  */
1047 int
1048 mlx5_flow_destroy(struct rte_eth_dev *dev,
1049                   struct rte_flow *flow,
1050                   struct rte_flow_error *error)
1051 {
1052         struct priv *priv = dev->data->dev_private;
1053
1054         (void)error;
1055         priv_lock(priv);
1056         priv_flow_destroy(priv, flow);
1057         priv_unlock(priv);
1058         return 0;
1059 }
1060
1061 /**
1062  * Destroy all flows.
1063  *
1064  * @param priv
1065  *   Pointer to private structure.
1066  */
1067 static void
1068 priv_flow_flush(struct priv *priv)
1069 {
1070         while (!LIST_EMPTY(&priv->flows)) {
1071                 struct rte_flow *flow;
1072
1073                 flow = LIST_FIRST(&priv->flows);
1074                 priv_flow_destroy(priv, flow);
1075         }
1076 }
1077
1078 /**
1079  * Destroy all flows.
1080  *
1081  * @see rte_flow_flush()
1082  * @see rte_flow_ops
1083  */
1084 int
1085 mlx5_flow_flush(struct rte_eth_dev *dev,
1086                 struct rte_flow_error *error)
1087 {
1088         struct priv *priv = dev->data->dev_private;
1089
1090         (void)error;
1091         priv_lock(priv);
1092         priv_flow_flush(priv);
1093         priv_unlock(priv);
1094         return 0;
1095 }
1096
1097 /**
1098  * Remove all flows.
1099  *
1100  * Called by dev_stop() to remove all flows.
1101  *
1102  * @param priv
1103  *   Pointer to private structure.
1104  */
1105 void
1106 priv_flow_stop(struct priv *priv)
1107 {
1108         struct rte_flow *flow;
1109
1110         for (flow = LIST_FIRST(&priv->flows);
1111              flow;
1112              flow = LIST_NEXT(flow, next)) {
1113                 claim_zero(ibv_exp_destroy_flow(flow->ibv_flow));
1114                 flow->ibv_flow = NULL;
1115                 DEBUG("Flow %p removed", (void *)flow);
1116         }
1117 }
1118
1119 /**
1120  * Add all flows.
1121  *
1122  * @param priv
1123  *   Pointer to private structure.
1124  *
1125  * @return
1126  *   0 on success, a errno value otherwise and rte_errno is set.
1127  */
1128 int
1129 priv_flow_start(struct priv *priv)
1130 {
1131         struct rte_flow *flow;
1132
1133         for (flow = LIST_FIRST(&priv->flows);
1134              flow;
1135              flow = LIST_NEXT(flow, next)) {
1136                 flow->ibv_flow = ibv_exp_create_flow(flow->qp,
1137                                                      flow->ibv_attr);
1138                 if (!flow->ibv_flow) {
1139                         DEBUG("Flow %p cannot be applied", (void *)flow);
1140                         rte_errno = EINVAL;
1141                         return rte_errno;
1142                 }
1143                 DEBUG("Flow %p applied", (void *)flow);
1144         }
1145         return 0;
1146 }