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_ipv4(const struct rte_flow_item *item,
  61                       const void *default_mask,
  62                       void *data);
  63
  64 static int
  65 mlx5_flow_create_ipv6(const struct rte_flow_item *item,
  66                       const void *default_mask,
  67                       void *data);
  68
  69 static int
  70 mlx5_flow_create_udp(const struct rte_flow_item *item,
  71                      const void *default_mask,
  72                      void *data);
  73
  74 static int
  75 mlx5_flow_create_tcp(const struct rte_flow_item *item,
  76                      const void *default_mask,
  77                      void *data);
  78
  79 struct rte_flow {
  80         LIST_ENTRY(rte_flow) next; /**< Pointer to the next flow structure. */
  81         struct ibv_exp_flow_attr *ibv_attr; /**< Pointer to Verbs attributes. */
  82         struct ibv_exp_rwq_ind_table *ind_table; /**< Indirection table. */
  83         struct ibv_qp *qp; /**< Verbs queue pair. */
  84         struct ibv_exp_flow *ibv_flow; /**< Verbs flow. */
  85         struct ibv_exp_wq *wq; /**< Verbs work queue. */
  86         struct ibv_cq *cq; /**< Verbs completion queue. */
  87         struct rxq *rxq; /**< Pointer to the queue, NULL if drop queue. */
  88 };
  89
  90 /** Static initializer for items. */
  91 #define ITEMS(...) \
  92         (const enum rte_flow_item_type []){ \
  93                 __VA_ARGS__, RTE_FLOW_ITEM_TYPE_END, \
  94         }
  95
  96 /** Structure to generate a simple graph of layers supported by the NIC. */
  97 struct mlx5_flow_items {
  98         /** List of possible actions for these items. */
  99         const enum rte_flow_action_type *const actions;
 100         /** Bit-masks corresponding to the possibilities for the item. */
 101         const void *mask;
 102         /** Bit-masks size in bytes. */
 103         const unsigned int mask_sz;
 104         /**
 105          * Conversion function from rte_flow to NIC specific flow.
 106          *
 107          * @param item
 108          *   rte_flow item to convert.
 109          * @param default_mask
 110          *   Default bit-masks to use when item->mask is not provided.
 111          * @param data
 112          *   Internal structure to store the conversion.
 113          *
 114          * @return
 115          *   0 on success, negative value otherwise.
 116          */
 117         int (*convert)(const struct rte_flow_item *item,
 118                        const void *default_mask,
 119                        void *data);
 120         /** Size in bytes of the destination structure. */
 121         const unsigned int dst_sz;
 122         /** List of possible following items.  */
 123         const enum rte_flow_item_type *const items;
 124 };
 125
 126 /** Valid action for this PMD. */
 127 static const enum rte_flow_action_type valid_actions[] = {
 128         RTE_FLOW_ACTION_TYPE_DROP,
 129         RTE_FLOW_ACTION_TYPE_QUEUE,
 130         RTE_FLOW_ACTION_TYPE_END,
 131 };
 132
 133 /** Graph of supported items and associated actions. */
 134 static const struct mlx5_flow_items mlx5_flow_items[] = {
 135         [RTE_FLOW_ITEM_TYPE_END] = {
 136                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_ETH),
 137         },
 138         [RTE_FLOW_ITEM_TYPE_ETH] = {
 139                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_IPV4,
 140                                RTE_FLOW_ITEM_TYPE_IPV6),
 141                 .actions = valid_actions,
 142                 .mask = &(const struct rte_flow_item_eth){
 143                         .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 144                         .src.addr_bytes = "\xff\xff\xff\xff\xff\xff",
 145                 },
 146                 .mask_sz = sizeof(struct rte_flow_item_eth),
 147                 .convert = mlx5_flow_create_eth,
 148                 .dst_sz = sizeof(struct ibv_exp_flow_spec_eth),
 149         },
 150         [RTE_FLOW_ITEM_TYPE_IPV4] = {
 151                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP,
 152                                RTE_FLOW_ITEM_TYPE_TCP),
 153                 .actions = valid_actions,
 154                 .mask = &(const struct rte_flow_item_ipv4){
 155                         .hdr = {
 156                                 .src_addr = -1,
 157                                 .dst_addr = -1,
 158                         },
 159                 },
 160                 .mask_sz = sizeof(struct rte_flow_item_ipv4),
 161                 .convert = mlx5_flow_create_ipv4,
 162                 .dst_sz = sizeof(struct ibv_exp_flow_spec_ipv4),
 163         },
 164         [RTE_FLOW_ITEM_TYPE_IPV6] = {
 165                 .items = ITEMS(RTE_FLOW_ITEM_TYPE_UDP,
 166                                RTE_FLOW_ITEM_TYPE_TCP),
 167                 .actions = valid_actions,
 168                 .mask = &(const struct rte_flow_item_ipv6){
 169                         .hdr = {
 170                                 .src_addr = {
 171                                         0xff, 0xff, 0xff, 0xff,
 172                                         0xff, 0xff, 0xff, 0xff,
 173                                         0xff, 0xff, 0xff, 0xff,
 174                                         0xff, 0xff, 0xff, 0xff,
 175                                 },
 176                                 .dst_addr = {
 177                                         0xff, 0xff, 0xff, 0xff,
 178                                         0xff, 0xff, 0xff, 0xff,
 179                                         0xff, 0xff, 0xff, 0xff,
 180                                         0xff, 0xff, 0xff, 0xff,
 181                                 },
 182                         },
 183                 },
 184                 .mask_sz = sizeof(struct rte_flow_item_ipv6),
 185                 .convert = mlx5_flow_create_ipv6,
 186                 .dst_sz = sizeof(struct ibv_exp_flow_spec_ipv6),
 187         },
 188         [RTE_FLOW_ITEM_TYPE_UDP] = {
 189                 .actions = valid_actions,
 190                 .mask = &(const struct rte_flow_item_udp){
 191                         .hdr = {
 192                                 .src_port = -1,
 193                                 .dst_port = -1,
 194                         },
 195                 },
 196                 .mask_sz = sizeof(struct rte_flow_item_udp),
 197                 .convert = mlx5_flow_create_udp,
 198                 .dst_sz = sizeof(struct ibv_exp_flow_spec_tcp_udp),
 199         },
 200         [RTE_FLOW_ITEM_TYPE_TCP] = {
 201                 .actions = valid_actions,
 202                 .mask = &(const struct rte_flow_item_tcp){
 203                         .hdr = {
 204                                 .src_port = -1,
 205                                 .dst_port = -1,
 206                         },
 207                 },
 208                 .mask_sz = sizeof(struct rte_flow_item_tcp),
 209                 .convert = mlx5_flow_create_tcp,
 210                 .dst_sz = sizeof(struct ibv_exp_flow_spec_tcp_udp),
 211         },
 212 };
 213
 214 /** Structure to pass to the conversion function. */
 215 struct mlx5_flow {
 216         struct ibv_exp_flow_attr *ibv_attr; /**< Verbs attribute. */
 217         unsigned int offset; /**< Offset in bytes in the ibv_attr buffer. */
 218 };
 219
 220 struct mlx5_flow_action {
 221         uint32_t queue:1; /**< Target is a receive queue. */
 222         uint32_t drop:1; /**< Target is a drop queue. */
 223         uint32_t queue_id; /**< Identifier of the queue. */
 224 };
 225
 226 /**
 227  * Check support for a given item.
 228  *
 229  * @param item[in]
 230  *   Item specification.
 231  * @param mask[in]
 232  *   Bit-masks covering supported fields to compare with spec, last and mask in
 233  *   \item.
 234  * @param size
 235  *   Bit-Mask size in bytes.
 236  *
 237  * @return
 238  *   0 on success.
 239  */
 240 static int
 241 mlx5_flow_item_validate(const struct rte_flow_item *item,
 242                         const uint8_t *mask, unsigned int size)
 243 {
 244         int ret = 0;
 245
 246         if (!item->spec && (item->mask || item->last))
 247                 return -1;
 248         if (item->spec && !item->mask) {
 249                 unsigned int i;
 250                 const uint8_t *spec = item->spec;
 251
 252                 for (i = 0; i < size; ++i)
 253                         if ((spec[i] | mask[i]) != mask[i])
 254                                 return -1;
 255         }
 256         if (item->last && !item->mask) {
 257                 unsigned int i;
 258                 const uint8_t *spec = item->last;
 259
 260                 for (i = 0; i < size; ++i)
 261                         if ((spec[i] | mask[i]) != mask[i])
 262                                 return -1;
 263         }
 264         if (item->mask) {
 265                 unsigned int i;
 266                 const uint8_t *spec = item->mask;
 267
 268                 for (i = 0; i < size; ++i)
 269                         if ((spec[i] | mask[i]) != mask[i])
 270                                 return -1;
 271         }
 272         if (item->spec && item->last) {
 273                 uint8_t spec[size];
 274                 uint8_t last[size];
 275                 const uint8_t *apply = mask;
 276                 unsigned int i;
 277
 278                 if (item->mask)
 279                         apply = item->mask;
 280                 for (i = 0; i < size; ++i) {
 281                         spec[i] = ((const uint8_t *)item->spec)[i] & apply[i];
 282                         last[i] = ((const uint8_t *)item->last)[i] & apply[i];
 283                 }
 284                 ret = memcmp(spec, last, size);
 285         }
 286         return ret;
 287 }
 288
 289 /**
 290  * Validate a flow supported by the NIC.
 291  *
 292  * @param priv
 293  *   Pointer to private structure.
 294  * @param[in] attr
 295  *   Flow rule attributes.
 296  * @param[in] pattern
 297  *   Pattern specification (list terminated by the END pattern item).
 298  * @param[in] actions
 299  *   Associated actions (list terminated by the END action).
 300  * @param[out] error
 301  *   Perform verbose error reporting if not NULL.
 302  * @param[in, out] flow
 303  *   Flow structure to update.
 304  *
 305  * @return
 306  *   0 on success, a negative errno value otherwise and rte_errno is set.
 307  */
 308 static int
 309 priv_flow_validate(struct priv *priv,
 310                    const struct rte_flow_attr *attr,
 311                    const struct rte_flow_item items[],
 312                    const struct rte_flow_action actions[],
 313                    struct rte_flow_error *error,
 314                    struct mlx5_flow *flow)
 315 {
 316         const struct mlx5_flow_items *cur_item = mlx5_flow_items;
 317         struct mlx5_flow_action action = {
 318                 .queue = 0,
 319                 .drop = 0,
 320         };
 321
 322         (void)priv;
 323         if (attr->group) {
 324                 rte_flow_error_set(error, ENOTSUP,
 325                                    RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
 326                                    NULL,
 327                                    "groups are not supported");
 328                 return -rte_errno;
 329         }
 330         if (attr->priority) {
 331                 rte_flow_error_set(error, ENOTSUP,
 332                                    RTE_FLOW_ERROR_TYPE_ATTR_PRIORITY,
 333                                    NULL,
 334                                    "priorities are not supported");
 335                 return -rte_errno;
 336         }
 337         if (attr->egress) {
 338                 rte_flow_error_set(error, ENOTSUP,
 339                                    RTE_FLOW_ERROR_TYPE_ATTR_EGRESS,
 340                                    NULL,
 341                                    "egress is not supported");
 342                 return -rte_errno;
 343         }
 344         if (!attr->ingress) {
 345                 rte_flow_error_set(error, ENOTSUP,
 346                                    RTE_FLOW_ERROR_TYPE_ATTR_INGRESS,
 347                                    NULL,
 348                                    "only ingress is supported");
 349                 return -rte_errno;
 350         }
 351         for (; items->type != RTE_FLOW_ITEM_TYPE_END; ++items) {
 352                 const struct mlx5_flow_items *token = NULL;
 353                 unsigned int i;
 354                 int err;
 355
 356                 if (items->type == RTE_FLOW_ITEM_TYPE_VOID)
 357                         continue;
 358                 for (i = 0;
 359                      cur_item->items &&
 360                      cur_item->items[i] != RTE_FLOW_ITEM_TYPE_END;
 361                      ++i) {
 362                         if (cur_item->items[i] == items->type) {
 363                                 token = &mlx5_flow_items[items->type];
 364                                 break;
 365                         }
 366                 }
 367                 if (!token)
 368                         goto exit_item_not_supported;
 369                 cur_item = token;
 370                 err = mlx5_flow_item_validate(items,
 371                                               (const uint8_t *)cur_item->mask,
 372                                               sizeof(cur_item->mask_sz));
 373                 if (err)
 374                         goto exit_item_not_supported;
 375                 if (flow->ibv_attr && cur_item->convert) {
 376                         err = cur_item->convert(items, cur_item->mask, flow);
 377                         if (err)
 378                                 goto exit_item_not_supported;
 379                 }
 380                 flow->offset += cur_item->dst_sz;
 381         }
 382         for (; actions->type != RTE_FLOW_ACTION_TYPE_END; ++actions) {
 383                 if (actions->type == RTE_FLOW_ACTION_TYPE_VOID) {
 384                         continue;
 385                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_DROP) {
 386                         action.drop = 1;
 387                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_QUEUE) {
 388                         const struct rte_flow_action_queue *queue =
 389                                 (const struct rte_flow_action_queue *)
 390                                 actions->conf;
 391
 392                         if (!queue || (queue->index > (priv->rxqs_n - 1)))
 393                                 goto exit_action_not_supported;
 394                         action.queue = 1;
 395                 } else {
 396                         goto exit_action_not_supported;
 397                 }
 398         }
 399         if (!action.queue && !action.drop) {
 400                 rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_HANDLE,
 401                                    NULL, "no valid action");
 402                 return -rte_errno;
 403         }
 404         return 0;
 405 exit_item_not_supported:
 406         rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ITEM,
 407                            items, "item not supported");
 408         return -rte_errno;
 409 exit_action_not_supported:
 410         rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
 411                            actions, "action not supported");
 412         return -rte_errno;
 413 }
 414
 415 /**
 416  * Validate a flow supported by the NIC.
 417  *
 418  * @see rte_flow_validate()
 419  * @see rte_flow_ops
 420  */
 421 int
 422 mlx5_flow_validate(struct rte_eth_dev *dev,
 423                    const struct rte_flow_attr *attr,
 424                    const struct rte_flow_item items[],
 425                    const struct rte_flow_action actions[],
 426                    struct rte_flow_error *error)
 427 {
 428         struct priv *priv = dev->data->dev_private;
 429         int ret;
 430         struct mlx5_flow flow = { .offset = sizeof(struct ibv_exp_flow_attr) };
 431
 432         priv_lock(priv);
 433         ret = priv_flow_validate(priv, attr, items, actions, error, &flow);
 434         priv_unlock(priv);
 435         return ret;
 436 }
 437
 438 /**
 439  * Convert Ethernet item to Verbs specification.
 440  *
 441  * @param item[in]
 442  *   Item specification.
 443  * @param default_mask[in]
 444  *   Default bit-masks to use when item->mask is not provided.
 445  * @param data[in, out]
 446  *   User structure.
 447  */
 448 static int
 449 mlx5_flow_create_eth(const struct rte_flow_item *item,
 450                      const void *default_mask,
 451                      void *data)
 452 {
 453         const struct rte_flow_item_eth *spec = item->spec;
 454         const struct rte_flow_item_eth *mask = item->mask;
 455         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 456         struct ibv_exp_flow_spec_eth *eth;
 457         const unsigned int eth_size = sizeof(struct ibv_exp_flow_spec_eth);
 458         unsigned int i;
 459
 460         ++flow->ibv_attr->num_of_specs;
 461         flow->ibv_attr->priority = 2;
 462         eth = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 463         *eth = (struct ibv_exp_flow_spec_eth) {
 464                 .type = IBV_EXP_FLOW_SPEC_ETH,
 465                 .size = eth_size,
 466         };
 467         if (!spec)
 468                 return 0;
 469         if (!mask)
 470                 mask = default_mask;
 471         memcpy(eth->val.dst_mac, spec->dst.addr_bytes, ETHER_ADDR_LEN);
 472         memcpy(eth->val.src_mac, spec->src.addr_bytes, ETHER_ADDR_LEN);
 473         memcpy(eth->mask.dst_mac, mask->dst.addr_bytes, ETHER_ADDR_LEN);
 474         memcpy(eth->mask.src_mac, mask->src.addr_bytes, ETHER_ADDR_LEN);
 475         /* Remove unwanted bits from values. */
 476         for (i = 0; i < ETHER_ADDR_LEN; ++i) {
 477                 eth->val.dst_mac[i] &= eth->mask.dst_mac[i];
 478                 eth->val.src_mac[i] &= eth->mask.src_mac[i];
 479         }
 480         return 0;
 481 }
 482
 483 /**
 484  * Convert IPv4 item to Verbs specification.
 485  *
 486  * @param item[in]
 487  *   Item specification.
 488  * @param default_mask[in]
 489  *   Default bit-masks to use when item->mask is not provided.
 490  * @param data[in, out]
 491  *   User structure.
 492  */
 493 static int
 494 mlx5_flow_create_ipv4(const struct rte_flow_item *item,
 495                       const void *default_mask,
 496                       void *data)
 497 {
 498         const struct rte_flow_item_ipv4 *spec = item->spec;
 499         const struct rte_flow_item_ipv4 *mask = item->mask;
 500         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 501         struct ibv_exp_flow_spec_ipv4 *ipv4;
 502         unsigned int ipv4_size = sizeof(struct ibv_exp_flow_spec_ipv4);
 503
 504         ++flow->ibv_attr->num_of_specs;
 505         flow->ibv_attr->priority = 1;
 506         ipv4 = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 507         *ipv4 = (struct ibv_exp_flow_spec_ipv4) {
 508                 .type = IBV_EXP_FLOW_SPEC_IPV4,
 509                 .size = ipv4_size,
 510         };
 511         if (!spec)
 512                 return 0;
 513         if (!mask)
 514                 mask = default_mask;
 515         ipv4->val = (struct ibv_exp_flow_ipv4_filter){
 516                 .src_ip = spec->hdr.src_addr,
 517                 .dst_ip = spec->hdr.dst_addr,
 518         };
 519         ipv4->mask = (struct ibv_exp_flow_ipv4_filter){
 520                 .src_ip = mask->hdr.src_addr,
 521                 .dst_ip = mask->hdr.dst_addr,
 522         };
 523         /* Remove unwanted bits from values. */
 524         ipv4->val.src_ip &= ipv4->mask.src_ip;
 525         ipv4->val.dst_ip &= ipv4->mask.dst_ip;
 526         return 0;
 527 }
 528
 529 /**
 530  * Convert IPv6 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_ipv6(const struct rte_flow_item *item,
 541                       const void *default_mask,
 542                       void *data)
 543 {
 544         const struct rte_flow_item_ipv6 *spec = item->spec;
 545         const struct rte_flow_item_ipv6 *mask = item->mask;
 546         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 547         struct ibv_exp_flow_spec_ipv6 *ipv6;
 548         unsigned int ipv6_size = sizeof(struct ibv_exp_flow_spec_ipv6);
 549         unsigned int i;
 550
 551         ++flow->ibv_attr->num_of_specs;
 552         flow->ibv_attr->priority = 1;
 553         ipv6 = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 554         *ipv6 = (struct ibv_exp_flow_spec_ipv6) {
 555                 .type = IBV_EXP_FLOW_SPEC_IPV6,
 556                 .size = ipv6_size,
 557         };
 558         if (!spec)
 559                 return 0;
 560         if (!mask)
 561                 mask = default_mask;
 562         memcpy(ipv6->val.src_ip, spec->hdr.src_addr,
 563                RTE_DIM(ipv6->val.src_ip));
 564         memcpy(ipv6->val.dst_ip, spec->hdr.dst_addr,
 565                RTE_DIM(ipv6->val.dst_ip));
 566         memcpy(ipv6->mask.src_ip, mask->hdr.src_addr,
 567                RTE_DIM(ipv6->mask.src_ip));
 568         memcpy(ipv6->mask.dst_ip, mask->hdr.dst_addr,
 569                RTE_DIM(ipv6->mask.dst_ip));
 570         /* Remove unwanted bits from values. */
 571         for (i = 0; i < RTE_DIM(ipv6->val.src_ip); ++i) {
 572                 ipv6->val.src_ip[i] &= ipv6->mask.src_ip[i];
 573                 ipv6->val.dst_ip[i] &= ipv6->mask.dst_ip[i];
 574         }
 575         return 0;
 576 }
 577
 578 /**
 579  * Convert UDP item to Verbs specification.
 580  *
 581  * @param item[in]
 582  *   Item specification.
 583  * @param default_mask[in]
 584  *   Default bit-masks to use when item->mask is not provided.
 585  * @param data[in, out]
 586  *   User structure.
 587  */
 588 static int
 589 mlx5_flow_create_udp(const struct rte_flow_item *item,
 590                      const void *default_mask,
 591                      void *data)
 592 {
 593         const struct rte_flow_item_udp *spec = item->spec;
 594         const struct rte_flow_item_udp *mask = item->mask;
 595         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 596         struct ibv_exp_flow_spec_tcp_udp *udp;
 597         unsigned int udp_size = sizeof(struct ibv_exp_flow_spec_tcp_udp);
 598
 599         ++flow->ibv_attr->num_of_specs;
 600         flow->ibv_attr->priority = 0;
 601         udp = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 602         *udp = (struct ibv_exp_flow_spec_tcp_udp) {
 603                 .type = IBV_EXP_FLOW_SPEC_UDP,
 604                 .size = udp_size,
 605         };
 606         if (!spec)
 607                 return 0;
 608         if (!mask)
 609                 mask = default_mask;
 610         udp->val.dst_port = spec->hdr.dst_port;
 611         udp->val.src_port = spec->hdr.src_port;
 612         udp->mask.dst_port = mask->hdr.dst_port;
 613         udp->mask.src_port = mask->hdr.src_port;
 614         /* Remove unwanted bits from values. */
 615         udp->val.src_port &= udp->mask.src_port;
 616         udp->val.dst_port &= udp->mask.dst_port;
 617         return 0;
 618 }
 619
 620 /**
 621  * Convert TCP item to Verbs specification.
 622  *
 623  * @param item[in]
 624  *   Item specification.
 625  * @param default_mask[in]
 626  *   Default bit-masks to use when item->mask is not provided.
 627  * @param data[in, out]
 628  *   User structure.
 629  */
 630 static int
 631 mlx5_flow_create_tcp(const struct rte_flow_item *item,
 632                      const void *default_mask,
 633                      void *data)
 634 {
 635         const struct rte_flow_item_tcp *spec = item->spec;
 636         const struct rte_flow_item_tcp *mask = item->mask;
 637         struct mlx5_flow *flow = (struct mlx5_flow *)data;
 638         struct ibv_exp_flow_spec_tcp_udp *tcp;
 639         unsigned int tcp_size = sizeof(struct ibv_exp_flow_spec_tcp_udp);
 640
 641         ++flow->ibv_attr->num_of_specs;
 642         flow->ibv_attr->priority = 0;
 643         tcp = (void *)((uintptr_t)flow->ibv_attr + flow->offset);
 644         *tcp = (struct ibv_exp_flow_spec_tcp_udp) {
 645                 .type = IBV_EXP_FLOW_SPEC_TCP,
 646                 .size = tcp_size,
 647         };
 648         if (!spec)
 649                 return 0;
 650         if (!mask)
 651                 mask = default_mask;
 652         tcp->val.dst_port = spec->hdr.dst_port;
 653         tcp->val.src_port = spec->hdr.src_port;
 654         tcp->mask.dst_port = mask->hdr.dst_port;
 655         tcp->mask.src_port = mask->hdr.src_port;
 656         /* Remove unwanted bits from values. */
 657         tcp->val.src_port &= tcp->mask.src_port;
 658         tcp->val.dst_port &= tcp->mask.dst_port;
 659         return 0;
 660 }
 661
 662 /**
 663  * Complete flow rule creation.
 664  *
 665  * @param priv
 666  *   Pointer to private structure.
 667  * @param ibv_attr
 668  *   Verbs flow attributes.
 669  * @param action
 670  *   Target action structure.
 671  * @param[out] error
 672  *   Perform verbose error reporting if not NULL.
 673  *
 674  * @return
 675  *   A flow if the rule could be created.
 676  */
 677 static struct rte_flow *
 678 priv_flow_create_action_queue(struct priv *priv,
 679                               struct ibv_exp_flow_attr *ibv_attr,
 680                               struct mlx5_flow_action *action,
 681                               struct rte_flow_error *error)
 682 {
 683         struct rxq_ctrl *rxq;
 684         struct rte_flow *rte_flow;
 685
 686         assert(priv->pd);
 687         assert(priv->ctx);
 688         rte_flow = rte_calloc(__func__, 1, sizeof(*rte_flow), 0);
 689         if (!rte_flow) {
 690                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
 691                                    NULL, "cannot allocate flow memory");
 692                 return NULL;
 693         }
 694         if (action->drop) {
 695                 rte_flow->cq =
 696                         ibv_exp_create_cq(priv->ctx, 1, NULL, NULL, 0,
 697                                           &(struct ibv_exp_cq_init_attr){
 698                                                   .comp_mask = 0,
 699                                           });
 700                 if (!rte_flow->cq) {
 701                         rte_flow_error_set(error, ENOMEM,
 702                                            RTE_FLOW_ERROR_TYPE_HANDLE,
 703                                            NULL, "cannot allocate CQ");
 704                         goto error;
 705                 }
 706                 rte_flow->wq = ibv_exp_create_wq(priv->ctx,
 707                                                  &(struct ibv_exp_wq_init_attr){
 708                                                  .wq_type = IBV_EXP_WQT_RQ,
 709                                                  .max_recv_wr = 1,
 710                                                  .max_recv_sge = 1,
 711                                                  .pd = priv->pd,
 712                                                  .cq = rte_flow->cq,
 713                                                  });
 714         } else {
 715                 rxq = container_of((*priv->rxqs)[action->queue_id],
 716                                    struct rxq_ctrl, rxq);
 717                 rte_flow->rxq = &rxq->rxq;
 718                 rte_flow->wq = rxq->wq;
 719         }
 720         rte_flow->ibv_attr = ibv_attr;
 721         rte_flow->ind_table = ibv_exp_create_rwq_ind_table(
 722                 priv->ctx,
 723                 &(struct ibv_exp_rwq_ind_table_init_attr){
 724                         .pd = priv->pd,
 725                         .log_ind_tbl_size = 0,
 726                         .ind_tbl = &rte_flow->wq,
 727                         .comp_mask = 0,
 728                 });
 729         if (!rte_flow->ind_table) {
 730                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
 731                                    NULL, "cannot allocate indirection table");
 732                 goto error;
 733         }
 734         rte_flow->qp = ibv_exp_create_qp(
 735                 priv->ctx,
 736                 &(struct ibv_exp_qp_init_attr){
 737                         .qp_type = IBV_QPT_RAW_PACKET,
 738                         .comp_mask =
 739                                 IBV_EXP_QP_INIT_ATTR_PD |
 740                                 IBV_EXP_QP_INIT_ATTR_PORT |
 741                                 IBV_EXP_QP_INIT_ATTR_RX_HASH,
 742                         .pd = priv->pd,
 743                         .rx_hash_conf = &(struct ibv_exp_rx_hash_conf){
 744                                 .rx_hash_function =
 745                                         IBV_EXP_RX_HASH_FUNC_TOEPLITZ,
 746                                 .rx_hash_key_len = rss_hash_default_key_len,
 747                                 .rx_hash_key = rss_hash_default_key,
 748                                 .rx_hash_fields_mask = 0,
 749                                 .rwq_ind_tbl = rte_flow->ind_table,
 750                         },
 751                         .port_num = priv->port,
 752                 });
 753         if (!rte_flow->qp) {
 754                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
 755                                    NULL, "cannot allocate QP");
 756                 goto error;
 757         }
 758         rte_flow->ibv_flow = ibv_exp_create_flow(rte_flow->qp,
 759                                                  rte_flow->ibv_attr);
 760         if (!rte_flow->ibv_flow) {
 761                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
 762                                    NULL, "flow rule creation failure");
 763                 goto error;
 764         }
 765         return rte_flow;
 766 error:
 767         assert(rte_flow);
 768         if (rte_flow->qp)
 769                 ibv_destroy_qp(rte_flow->qp);
 770         if (rte_flow->ind_table)
 771                 ibv_exp_destroy_rwq_ind_table(rte_flow->ind_table);
 772         if (!rte_flow->rxq && rte_flow->wq)
 773                 ibv_exp_destroy_wq(rte_flow->wq);
 774         if (!rte_flow->rxq && rte_flow->cq)
 775                 ibv_destroy_cq(rte_flow->cq);
 776         rte_free(rte_flow->ibv_attr);
 777         rte_free(rte_flow);
 778         return NULL;
 779 }
 780
 781 /**
 782  * Convert a flow.
 783  *
 784  * @param priv
 785  *   Pointer to private structure.
 786  * @param[in] attr
 787  *   Flow rule attributes.
 788  * @param[in] pattern
 789  *   Pattern specification (list terminated by the END pattern item).
 790  * @param[in] actions
 791  *   Associated actions (list terminated by the END action).
 792  * @param[out] error
 793  *   Perform verbose error reporting if not NULL.
 794  *
 795  * @return
 796  *   A flow on success, NULL otherwise.
 797  */
 798 static struct rte_flow *
 799 priv_flow_create(struct priv *priv,
 800                  const struct rte_flow_attr *attr,
 801                  const struct rte_flow_item items[],
 802                  const struct rte_flow_action actions[],
 803                  struct rte_flow_error *error)
 804 {
 805         struct rte_flow *rte_flow;
 806         struct mlx5_flow_action action;
 807         struct mlx5_flow flow = { .offset = sizeof(struct ibv_exp_flow_attr), };
 808         int err;
 809
 810         err = priv_flow_validate(priv, attr, items, actions, error, &flow);
 811         if (err)
 812                 goto exit;
 813         flow.ibv_attr = rte_malloc(__func__, flow.offset, 0);
 814         flow.offset = sizeof(struct ibv_exp_flow_attr);
 815         if (!flow.ibv_attr) {
 816                 rte_flow_error_set(error, ENOMEM, RTE_FLOW_ERROR_TYPE_HANDLE,
 817                                    NULL, "cannot allocate ibv_attr memory");
 818                 goto exit;
 819         }
 820         *flow.ibv_attr = (struct ibv_exp_flow_attr){
 821                 .type = IBV_EXP_FLOW_ATTR_NORMAL,
 822                 .size = sizeof(struct ibv_exp_flow_attr),
 823                 .priority = attr->priority,
 824                 .num_of_specs = 0,
 825                 .port = 0,
 826                 .flags = 0,
 827                 .reserved = 0,
 828         };
 829         claim_zero(priv_flow_validate(priv, attr, items, actions,
 830                                       error, &flow));
 831         action = (struct mlx5_flow_action){
 832                 .queue = 0,
 833                 .drop = 0,
 834         };
 835         for (; actions->type != RTE_FLOW_ACTION_TYPE_END; ++actions) {
 836                 if (actions->type == RTE_FLOW_ACTION_TYPE_VOID) {
 837                         continue;
 838                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_QUEUE) {
 839                         action.queue = 1;
 840                         action.queue_id =
 841                                 ((const struct rte_flow_action_queue *)
 842                                  actions->conf)->index;
 843                 } else if (actions->type == RTE_FLOW_ACTION_TYPE_DROP) {
 844                         action.drop = 1;
 845                 } else {
 846                         rte_flow_error_set(error, ENOTSUP,
 847                                            RTE_FLOW_ERROR_TYPE_ACTION,
 848                                            actions, "unsupported action");
 849                         goto exit;
 850                 }
 851         }
 852         rte_flow = priv_flow_create_action_queue(priv, flow.ibv_attr,
 853                                                  &action, error);
 854         return rte_flow;
 855 exit:
 856         rte_free(flow.ibv_attr);
 857         return NULL;
 858 }
 859
 860 /**
 861  * Create a flow.
 862  *
 863  * @see rte_flow_create()
 864  * @see rte_flow_ops
 865  */
 866 struct rte_flow *
 867 mlx5_flow_create(struct rte_eth_dev *dev,
 868                  const struct rte_flow_attr *attr,
 869                  const struct rte_flow_item items[],
 870                  const struct rte_flow_action actions[],
 871                  struct rte_flow_error *error)
 872 {
 873         struct priv *priv = dev->data->dev_private;
 874         struct rte_flow *flow;
 875
 876         priv_lock(priv);
 877         flow = priv_flow_create(priv, attr, items, actions, error);
 878         if (flow) {
 879                 LIST_INSERT_HEAD(&priv->flows, flow, next);
 880                 DEBUG("Flow created %p", (void *)flow);
 881         }
 882         priv_unlock(priv);
 883         return flow;
 884 }
 885
 886 /**
 887  * Destroy a flow.
 888  *
 889  * @param priv
 890  *   Pointer to private structure.
 891  * @param[in] flow
 892  *   Flow to destroy.
 893  */
 894 static void
 895 priv_flow_destroy(struct priv *priv,
 896                   struct rte_flow *flow)
 897 {
 898         (void)priv;
 899         LIST_REMOVE(flow, next);
 900         if (flow->ibv_flow)
 901                 claim_zero(ibv_exp_destroy_flow(flow->ibv_flow));
 902         if (flow->qp)
 903                 claim_zero(ibv_destroy_qp(flow->qp));
 904         if (flow->ind_table)
 905                 claim_zero(ibv_exp_destroy_rwq_ind_table(flow->ind_table));
 906         if (!flow->rxq && flow->wq)
 907                 claim_zero(ibv_exp_destroy_wq(flow->wq));
 908         if (!flow->rxq && flow->cq)
 909                 claim_zero(ibv_destroy_cq(flow->cq));
 910         rte_free(flow->ibv_attr);
 911         DEBUG("Flow destroyed %p", (void *)flow);
 912         rte_free(flow);
 913 }
 914
 915 /**
 916  * Destroy a flow.
 917  *
 918  * @see rte_flow_destroy()
 919  * @see rte_flow_ops
 920  */
 921 int
 922 mlx5_flow_destroy(struct rte_eth_dev *dev,
 923                   struct rte_flow *flow,
 924                   struct rte_flow_error *error)
 925 {
 926         struct priv *priv = dev->data->dev_private;
 927
 928         (void)error;
 929         priv_lock(priv);
 930         priv_flow_destroy(priv, flow);
 931         priv_unlock(priv);
 932         return 0;
 933 }
 934
 935 /**
 936  * Destroy all flows.
 937  *
 938  * @param priv
 939  *   Pointer to private structure.
 940  */
 941 static void
 942 priv_flow_flush(struct priv *priv)
 943 {
 944         while (!LIST_EMPTY(&priv->flows)) {
 945                 struct rte_flow *flow;
 946
 947                 flow = LIST_FIRST(&priv->flows);
 948                 priv_flow_destroy(priv, flow);
 949         }
 950 }
 951
 952 /**
 953  * Destroy all flows.
 954  *
 955  * @see rte_flow_flush()
 956  * @see rte_flow_ops
 957  */
 958 int
 959 mlx5_flow_flush(struct rte_eth_dev *dev,
 960                 struct rte_flow_error *error)
 961 {
 962         struct priv *priv = dev->data->dev_private;
 963
 964         (void)error;
 965         priv_lock(priv);
 966         priv_flow_flush(priv);
 967         priv_unlock(priv);
 968         return 0;
 969 }
 970
 971 /**
 972  * Remove all flows.
 973  *
 974  * Called by dev_stop() to remove all flows.
 975  *
 976  * @param priv
 977  *   Pointer to private structure.
 978  */
 979 void
 980 priv_flow_stop(struct priv *priv)
 981 {
 982         struct rte_flow *flow;
 983
 984         for (flow = LIST_FIRST(&priv->flows);
 985              flow;
 986              flow = LIST_NEXT(flow, next)) {
 987                 claim_zero(ibv_exp_destroy_flow(flow->ibv_flow));
 988                 flow->ibv_flow = NULL;
 989                 DEBUG("Flow %p removed", (void *)flow);
 990         }
 991 }
 992
 993 /**
 994  * Add all flows.
 995  *
 996  * @param priv
 997  *   Pointer to private structure.
 998  *
 999  * @return
1000  *   0 on success, a errno value otherwise and rte_errno is set.
1001  */
1002 int
1003 priv_flow_start(struct priv *priv)
1004 {
1005         struct rte_flow *flow;
1006
1007         for (flow = LIST_FIRST(&priv->flows);
1008              flow;
1009              flow = LIST_NEXT(flow, next)) {
1010                 flow->ibv_flow = ibv_exp_create_flow(flow->qp,
1011                                                      flow->ibv_attr);
1012                 if (!flow->ibv_flow) {
1013                         DEBUG("Flow %p cannot be applied", (void *)flow);
1014                         rte_errno = EINVAL;
1015                         return rte_errno;
1016                 }
1017                 DEBUG("Flow %p applied", (void *)flow);
1018         }
1019         return 0;
1020 }