drivers/net/mlx5/mlx5_flow_verbs.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright 2018 Mellanox Technologies, Ltd
   3  */
   4
   5 #include <netinet/in.h>
   6 #include <sys/queue.h>
   7 #include <stdalign.h>
   8 #include <stdint.h>
   9 #include <string.h>
  10
  11 /* Verbs header. */
  12 /* ISO C doesn't support unnamed structs/unions, disabling -pedantic. */
  13 #ifdef PEDANTIC
  14 #pragma GCC diagnostic ignored "-Wpedantic"
  15 #endif
  16 #include <infiniband/verbs.h>
  17 #ifdef PEDANTIC
  18 #pragma GCC diagnostic error "-Wpedantic"
  19 #endif
  20
  21 #include <rte_common.h>
  22 #include <rte_ether.h>
  23 #include <rte_ethdev_driver.h>
  24 #include <rte_flow.h>
  25 #include <rte_flow_driver.h>
  26 #include <rte_malloc.h>
  27 #include <rte_ip.h>
  28
  29 #include <mlx5_glue.h>
  30 #include <mlx5_prm.h>
  31
  32 #include "mlx5_defs.h"
  33 #include "mlx5.h"
  34 #include "mlx5_flow.h"
  35 #include "mlx5_rxtx.h"
  36
  37 #define VERBS_SPEC_INNER(item_flags) \
  38         (!!((item_flags) & MLX5_FLOW_LAYER_TUNNEL) ? IBV_FLOW_SPEC_INNER : 0)
  39
  40 /**
  41  * Get Verbs flow counter by index.
  42  *
  43  * @param[in] dev
  44  *   Pointer to the Ethernet device structure.
  45  * @param[in] idx
  46  *   mlx5 flow counter index in the container.
  47  * @param[out] ppool
  48  *   mlx5 flow counter pool in the container,
  49  *
  50  * @return
  51  *   A pointer to the counter, NULL otherwise.
  52  */
  53 static struct mlx5_flow_counter *
  54 flow_verbs_counter_get_by_idx(struct rte_eth_dev *dev,
  55                               uint32_t idx,
  56                               struct mlx5_flow_counter_pool **ppool)
  57 {
  58         struct mlx5_priv *priv = dev->data->dev_private;
  59         struct mlx5_pools_container *cont = MLX5_CNT_CONTAINER(priv->sh, 0, 0);
  60         struct mlx5_flow_counter_pool *pool;
  61
  62         idx--;
  63         pool = cont->pools[idx / MLX5_COUNTERS_PER_POOL];
  64         MLX5_ASSERT(pool);
  65         if (ppool)
  66                 *ppool = pool;
  67         return &pool->counters_raw[idx % MLX5_COUNTERS_PER_POOL];
  68 }
  69
  70 /**
  71  * Create Verbs flow counter with Verbs library.
  72  *
  73  * @param[in] dev
  74  *   Pointer to the Ethernet device structure.
  75  * @param[in, out] counter
  76  *   mlx5 flow counter object, contains the counter id,
  77  *   handle of created Verbs flow counter is returned
  78  *   in cs field (if counters are supported).
  79  *
  80  * @return
  81  *   0 On success else a negative errno value is returned
  82  *   and rte_errno is set.
  83  */
  84 static int
  85 flow_verbs_counter_create(struct rte_eth_dev *dev,
  86                           struct mlx5_flow_counter_ext *counter)
  87 {
  88 #if defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42)
  89         struct mlx5_priv *priv = dev->data->dev_private;
  90         struct ibv_context *ctx = priv->sh->ctx;
  91         struct ibv_counter_set_init_attr init = {
  92                          .counter_set_id = counter->id};
  93
  94         counter->cs = mlx5_glue->create_counter_set(ctx, &init);
  95         if (!counter->cs) {
  96                 rte_errno = ENOTSUP;
  97                 return -ENOTSUP;
  98         }
  99         return 0;
 100 #elif defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
 101         struct mlx5_priv *priv = dev->data->dev_private;
 102         struct ibv_context *ctx = priv->sh->ctx;
 103         struct ibv_counters_init_attr init = {0};
 104         struct ibv_counter_attach_attr attach;
 105         int ret;
 106
 107         memset(&attach, 0, sizeof(attach));
 108         counter->cs = mlx5_glue->create_counters(ctx, &init);
 109         if (!counter->cs) {
 110                 rte_errno = ENOTSUP;
 111                 return -ENOTSUP;
 112         }
 113         attach.counter_desc = IBV_COUNTER_PACKETS;
 114         attach.index = 0;
 115         ret = mlx5_glue->attach_counters(counter->cs, &attach, NULL);
 116         if (!ret) {
 117                 attach.counter_desc = IBV_COUNTER_BYTES;
 118                 attach.index = 1;
 119                 ret = mlx5_glue->attach_counters
 120                                         (counter->cs, &attach, NULL);
 121         }
 122         if (ret) {
 123                 claim_zero(mlx5_glue->destroy_counters(counter->cs));
 124                 counter->cs = NULL;
 125                 rte_errno = ret;
 126                 return -ret;
 127         }
 128         return 0;
 129 #else
 130         (void)dev;
 131         (void)counter;
 132         rte_errno = ENOTSUP;
 133         return -ENOTSUP;
 134 #endif
 135 }
 136
 137 /**
 138  * Get a flow counter.
 139  *
 140  * @param[in] dev
 141  *   Pointer to the Ethernet device structure.
 142  * @param[in] shared
 143  *   Indicate if this counter is shared with other flows.
 144  * @param[in] id
 145  *   Counter identifier.
 146  *
 147  * @return
 148  *   Index to the counter, 0 otherwise and rte_errno is set.
 149  */
 150 static uint32_t
 151 flow_verbs_counter_new(struct rte_eth_dev *dev, uint32_t shared, uint32_t id)
 152 {
 153         struct mlx5_priv *priv = dev->data->dev_private;
 154         struct mlx5_pools_container *cont = MLX5_CNT_CONTAINER(priv->sh, 0, 0);
 155         struct mlx5_flow_counter_pool *pool = NULL;
 156         struct mlx5_flow_counter_ext *cnt_ext = NULL;
 157         struct mlx5_flow_counter *cnt = NULL;
 158         uint32_t n_valid = rte_atomic16_read(&cont->n_valid);
 159         uint32_t pool_idx;
 160         uint32_t i;
 161         int ret;
 162
 163         if (shared) {
 164                 for (pool_idx = 0; pool_idx < n_valid; ++pool_idx) {
 165                         pool = cont->pools[pool_idx];
 166                         for (i = 0; i < MLX5_COUNTERS_PER_POOL; ++i) {
 167                                 cnt_ext = MLX5_GET_POOL_CNT_EXT(pool, i);
 168                                 if (cnt_ext->shared && cnt_ext->id == id) {
 169                                         cnt_ext->ref_cnt++;
 170                                         return MLX5_MAKE_CNT_IDX(pool_idx, i);
 171                                 }
 172                         }
 173                 }
 174         }
 175         for (pool_idx = 0; pool_idx < n_valid; ++pool_idx) {
 176                 pool = cont->pools[pool_idx];
 177                 if (!pool)
 178                         continue;
 179                 cnt = TAILQ_FIRST(&pool->counters);
 180                 if (cnt)
 181                         break;
 182         }
 183         if (!cnt) {
 184                 struct mlx5_flow_counter_pool **pools;
 185                 uint32_t size;
 186
 187                 if (n_valid == cont->n) {
 188                         /* Resize the container pool array. */
 189                         size = sizeof(struct mlx5_flow_counter_pool *) *
 190                                      (n_valid + MLX5_CNT_CONTAINER_RESIZE);
 191                         pools = rte_zmalloc(__func__, size, 0);
 192                         if (!pools)
 193                                 return 0;
 194                         if (n_valid) {
 195                                 memcpy(pools, cont->pools,
 196                                        sizeof(struct mlx5_flow_counter_pool *) *
 197                                        n_valid);
 198                                 rte_free(cont->pools);
 199                         }
 200                         cont->pools = pools;
 201                         cont->n += MLX5_CNT_CONTAINER_RESIZE;
 202                 }
 203                 /* Allocate memory for new pool*/
 204                 size = sizeof(*pool) + sizeof(*cnt_ext) *
 205                        MLX5_COUNTERS_PER_POOL;
 206                 pool = rte_calloc(__func__, 1, size, 0);
 207                 if (!pool)
 208                         return 0;
 209                 for (i = 0; i < MLX5_COUNTERS_PER_POOL; ++i) {
 210                         cnt = &pool->counters_raw[i];
 211                         TAILQ_INSERT_HEAD(&pool->counters, cnt, next);
 212                 }
 213                 cnt = &pool->counters_raw[0];
 214                 cont->pools[n_valid] = pool;
 215                 pool_idx = n_valid;
 216                 rte_atomic16_add(&cont->n_valid, 1);
 217                 TAILQ_INSERT_HEAD(&cont->pool_list, pool, next);
 218         }
 219         i = cnt - pool->counters_raw;
 220         cnt_ext = MLX5_GET_POOL_CNT_EXT(pool, i);
 221         cnt_ext->id = id;
 222         cnt_ext->shared = shared;
 223         cnt_ext->ref_cnt = 1;
 224         cnt->hits = 0;
 225         cnt->bytes = 0;
 226         /* Create counter with Verbs. */
 227         ret = flow_verbs_counter_create(dev, cnt_ext);
 228         if (!ret) {
 229                 TAILQ_REMOVE(&pool->counters, cnt, next);
 230                 return MLX5_MAKE_CNT_IDX(pool_idx, i);
 231         }
 232         /* Some error occurred in Verbs library. */
 233         rte_errno = -ret;
 234         return 0;
 235 }
 236
 237 /**
 238  * Release a flow counter.
 239  *
 240  * @param[in] dev
 241  *   Pointer to the Ethernet device structure.
 242  * @param[in] counter
 243  *   Index to the counter handler.
 244  */
 245 static void
 246 flow_verbs_counter_release(struct rte_eth_dev *dev, uint32_t counter)
 247 {
 248         struct mlx5_flow_counter_pool *pool;
 249         struct mlx5_flow_counter *cnt;
 250         struct mlx5_flow_counter_ext *cnt_ext;
 251
 252         cnt = flow_verbs_counter_get_by_idx(dev, counter,
 253                                             &pool);
 254         cnt_ext = MLX5_CNT_TO_CNT_EXT(pool, cnt);
 255         if (--cnt_ext->ref_cnt == 0) {
 256 #if defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42)
 257                 claim_zero(mlx5_glue->destroy_counter_set(cnt_ext->cs));
 258                 cnt_ext->cs = NULL;
 259 #elif defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
 260                 claim_zero(mlx5_glue->destroy_counters(cnt_ext->cs));
 261                 cnt_ext->cs = NULL;
 262 #endif
 263                 TAILQ_INSERT_HEAD(&pool->counters, cnt, next);
 264         }
 265 }
 266
 267 /**
 268  * Query a flow counter via Verbs library call.
 269  *
 270  * @see rte_flow_query()
 271  * @see rte_flow_ops
 272  */
 273 static int
 274 flow_verbs_counter_query(struct rte_eth_dev *dev __rte_unused,
 275                          struct rte_flow *flow, void *data,
 276                          struct rte_flow_error *error)
 277 {
 278 #if defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) || \
 279         defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
 280         if (flow->counter) {
 281                 struct mlx5_flow_counter_pool *pool;
 282                 struct mlx5_flow_counter *cnt = flow_verbs_counter_get_by_idx
 283                                                 (dev, flow->counter, &pool);
 284                 struct mlx5_flow_counter_ext *cnt_ext = MLX5_CNT_TO_CNT_EXT
 285                                                         (pool, cnt);
 286                 struct rte_flow_query_count *qc = data;
 287                 uint64_t counters[2] = {0, 0};
 288 #if defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42)
 289                 struct ibv_query_counter_set_attr query_cs_attr = {
 290                         .cs = cnt_ext->cs,
 291                         .query_flags = IBV_COUNTER_SET_FORCE_UPDATE,
 292                 };
 293                 struct ibv_counter_set_data query_out = {
 294                         .out = counters,
 295                         .outlen = 2 * sizeof(uint64_t),
 296                 };
 297                 int err = mlx5_glue->query_counter_set(&query_cs_attr,
 298                                                        &query_out);
 299 #elif defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
 300                 int err = mlx5_glue->query_counters
 301                                (cnt_ext->cs, counters,
 302                                 RTE_DIM(counters),
 303                                 IBV_READ_COUNTERS_ATTR_PREFER_CACHED);
 304 #endif
 305                 if (err)
 306                         return rte_flow_error_set
 307                                 (error, err,
 308                                  RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
 309                                  NULL,
 310                                  "cannot read counter");
 311                 qc->hits_set = 1;
 312                 qc->bytes_set = 1;
 313                 qc->hits = counters[0] - cnt->hits;
 314                 qc->bytes = counters[1] - cnt->bytes;
 315                 if (qc->reset) {
 316                         cnt->hits = counters[0];
 317                         cnt->bytes = counters[1];
 318                 }
 319                 return 0;
 320         }
 321         return rte_flow_error_set(error, EINVAL,
 322                                   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
 323                                   NULL,
 324                                   "flow does not have counter");
 325 #else
 326         (void)flow;
 327         (void)data;
 328         return rte_flow_error_set(error, ENOTSUP,
 329                                   RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
 330                                   NULL,
 331                                   "counters are not available");
 332 #endif
 333 }
 334
 335 /**
 336  * Add a verbs item specification into @p verbs.
 337  *
 338  * @param[out] verbs
 339  *   Pointer to verbs structure.
 340  * @param[in] src
 341  *   Create specification.
 342  * @param[in] size
 343  *   Size in bytes of the specification to copy.
 344  */
 345 static void
 346 flow_verbs_spec_add(struct mlx5_flow_verbs_workspace *verbs,
 347                     void *src, unsigned int size)
 348 {
 349         void *dst;
 350
 351         if (!verbs)
 352                 return;
 353         MLX5_ASSERT(verbs->specs);
 354         dst = (void *)(verbs->specs + verbs->size);
 355         memcpy(dst, src, size);
 356         ++verbs->attr.num_of_specs;
 357         verbs->size += size;
 358 }
 359
 360 /**
 361  * Convert the @p item into a Verbs specification. This function assumes that
 362  * the input is valid and that there is space to insert the requested item
 363  * into the flow.
 364  *
 365  * @param[in, out] dev_flow
 366  *   Pointer to dev_flow structure.
 367  * @param[in] item
 368  *   Item specification.
 369  * @param[in] item_flags
 370  *   Parsed item flags.
 371  */
 372 static void
 373 flow_verbs_translate_item_eth(struct mlx5_flow *dev_flow,
 374                               const struct rte_flow_item *item,
 375                               uint64_t item_flags)
 376 {
 377         const struct rte_flow_item_eth *spec = item->spec;
 378         const struct rte_flow_item_eth *mask = item->mask;
 379         const unsigned int size = sizeof(struct ibv_flow_spec_eth);
 380         struct ibv_flow_spec_eth eth = {
 381                 .type = IBV_FLOW_SPEC_ETH | VERBS_SPEC_INNER(item_flags),
 382                 .size = size,
 383         };
 384
 385         if (!mask)
 386                 mask = &rte_flow_item_eth_mask;
 387         if (spec) {
 388                 unsigned int i;
 389
 390                 memcpy(&eth.val.dst_mac, spec->dst.addr_bytes,
 391                         RTE_ETHER_ADDR_LEN);
 392                 memcpy(&eth.val.src_mac, spec->src.addr_bytes,
 393                         RTE_ETHER_ADDR_LEN);
 394                 eth.val.ether_type = spec->type;
 395                 memcpy(&eth.mask.dst_mac, mask->dst.addr_bytes,
 396                         RTE_ETHER_ADDR_LEN);
 397                 memcpy(&eth.mask.src_mac, mask->src.addr_bytes,
 398                         RTE_ETHER_ADDR_LEN);
 399                 eth.mask.ether_type = mask->type;
 400                 /* Remove unwanted bits from values. */
 401                 for (i = 0; i < RTE_ETHER_ADDR_LEN; ++i) {
 402                         eth.val.dst_mac[i] &= eth.mask.dst_mac[i];
 403                         eth.val.src_mac[i] &= eth.mask.src_mac[i];
 404                 }
 405                 eth.val.ether_type &= eth.mask.ether_type;
 406         }
 407         flow_verbs_spec_add(&dev_flow->verbs, &eth, size);
 408 }
 409
 410 /**
 411  * Update the VLAN tag in the Verbs Ethernet specification.
 412  * This function assumes that the input is valid and there is space to add
 413  * the requested item.
 414  *
 415  * @param[in, out] attr
 416  *   Pointer to Verbs attributes structure.
 417  * @param[in] eth
 418  *   Verbs structure containing the VLAN information to copy.
 419  */
 420 static void
 421 flow_verbs_item_vlan_update(struct ibv_flow_attr *attr,
 422                             struct ibv_flow_spec_eth *eth)
 423 {
 424         unsigned int i;
 425         const enum ibv_flow_spec_type search = eth->type;
 426         struct ibv_spec_header *hdr = (struct ibv_spec_header *)
 427                 ((uint8_t *)attr + sizeof(struct ibv_flow_attr));
 428
 429         for (i = 0; i != attr->num_of_specs; ++i) {
 430                 if (hdr->type == search) {
 431                         struct ibv_flow_spec_eth *e =
 432                                 (struct ibv_flow_spec_eth *)hdr;
 433
 434                         e->val.vlan_tag = eth->val.vlan_tag;
 435                         e->mask.vlan_tag = eth->mask.vlan_tag;
 436                         e->val.ether_type = eth->val.ether_type;
 437                         e->mask.ether_type = eth->mask.ether_type;
 438                         break;
 439                 }
 440                 hdr = (struct ibv_spec_header *)((uint8_t *)hdr + hdr->size);
 441         }
 442 }
 443
 444 /**
 445  * Convert the @p item into a Verbs specification. This function assumes that
 446  * the input is valid and that there is space to insert the requested item
 447  * into the flow.
 448  *
 449  * @param[in, out] dev_flow
 450  *   Pointer to dev_flow structure.
 451  * @param[in] item
 452  *   Item specification.
 453  * @param[in] item_flags
 454  *   Parsed item flags.
 455  */
 456 static void
 457 flow_verbs_translate_item_vlan(struct mlx5_flow *dev_flow,
 458                                const struct rte_flow_item *item,
 459                                uint64_t item_flags)
 460 {
 461         const struct rte_flow_item_vlan *spec = item->spec;
 462         const struct rte_flow_item_vlan *mask = item->mask;
 463         unsigned int size = sizeof(struct ibv_flow_spec_eth);
 464         const int tunnel = !!(item_flags & MLX5_FLOW_LAYER_TUNNEL);
 465         struct ibv_flow_spec_eth eth = {
 466                 .type = IBV_FLOW_SPEC_ETH | VERBS_SPEC_INNER(item_flags),
 467                 .size = size,
 468         };
 469         const uint32_t l2m = tunnel ? MLX5_FLOW_LAYER_INNER_L2 :
 470                                       MLX5_FLOW_LAYER_OUTER_L2;
 471
 472         if (!mask)
 473                 mask = &rte_flow_item_vlan_mask;
 474         if (spec) {
 475                 eth.val.vlan_tag = spec->tci;
 476                 eth.mask.vlan_tag = mask->tci;
 477                 eth.val.vlan_tag &= eth.mask.vlan_tag;
 478                 eth.val.ether_type = spec->inner_type;
 479                 eth.mask.ether_type = mask->inner_type;
 480                 eth.val.ether_type &= eth.mask.ether_type;
 481         }
 482         if (!(item_flags & l2m))
 483                 flow_verbs_spec_add(&dev_flow->verbs, &eth, size);
 484         else
 485                 flow_verbs_item_vlan_update(&dev_flow->verbs.attr, &eth);
 486         if (!tunnel)
 487                 dev_flow->handle->vf_vlan.tag =
 488                         rte_be_to_cpu_16(spec->tci) & 0x0fff;
 489 }
 490
 491 /**
 492  * Convert the @p item into a Verbs specification. This function assumes that
 493  * the input is valid and that there is space to insert the requested item
 494  * into the flow.
 495  *
 496  * @param[in, out] dev_flow
 497  *   Pointer to dev_flow structure.
 498  * @param[in] item
 499  *   Item specification.
 500  * @param[in] item_flags
 501  *   Parsed item flags.
 502  */
 503 static void
 504 flow_verbs_translate_item_ipv4(struct mlx5_flow *dev_flow,
 505                                const struct rte_flow_item *item,
 506                                uint64_t item_flags)
 507 {
 508         const struct rte_flow_item_ipv4 *spec = item->spec;
 509         const struct rte_flow_item_ipv4 *mask = item->mask;
 510         unsigned int size = sizeof(struct ibv_flow_spec_ipv4_ext);
 511         struct ibv_flow_spec_ipv4_ext ipv4 = {
 512                 .type = IBV_FLOW_SPEC_IPV4_EXT | VERBS_SPEC_INNER(item_flags),
 513                 .size = size,
 514         };
 515
 516         if (!mask)
 517                 mask = &rte_flow_item_ipv4_mask;
 518         if (spec) {
 519                 ipv4.val = (struct ibv_flow_ipv4_ext_filter){
 520                         .src_ip = spec->hdr.src_addr,
 521                         .dst_ip = spec->hdr.dst_addr,
 522                         .proto = spec->hdr.next_proto_id,
 523                         .tos = spec->hdr.type_of_service,
 524                 };
 525                 ipv4.mask = (struct ibv_flow_ipv4_ext_filter){
 526                         .src_ip = mask->hdr.src_addr,
 527                         .dst_ip = mask->hdr.dst_addr,
 528                         .proto = mask->hdr.next_proto_id,
 529                         .tos = mask->hdr.type_of_service,
 530                 };
 531                 /* Remove unwanted bits from values. */
 532                 ipv4.val.src_ip &= ipv4.mask.src_ip;
 533                 ipv4.val.dst_ip &= ipv4.mask.dst_ip;
 534                 ipv4.val.proto &= ipv4.mask.proto;
 535                 ipv4.val.tos &= ipv4.mask.tos;
 536         }
 537         flow_verbs_spec_add(&dev_flow->verbs, &ipv4, size);
 538 }
 539
 540 /**
 541  * Convert the @p item into a Verbs specification. This function assumes that
 542  * the input is valid and that there is space to insert the requested item
 543  * into the flow.
 544  *
 545  * @param[in, out] dev_flow
 546  *   Pointer to dev_flow structure.
 547  * @param[in] item
 548  *   Item specification.
 549  * @param[in] item_flags
 550  *   Parsed item flags.
 551  */
 552 static void
 553 flow_verbs_translate_item_ipv6(struct mlx5_flow *dev_flow,
 554                                const struct rte_flow_item *item,
 555                                uint64_t item_flags)
 556 {
 557         const struct rte_flow_item_ipv6 *spec = item->spec;
 558         const struct rte_flow_item_ipv6 *mask = item->mask;
 559         unsigned int size = sizeof(struct ibv_flow_spec_ipv6);
 560         struct ibv_flow_spec_ipv6 ipv6 = {
 561                 .type = IBV_FLOW_SPEC_IPV6 | VERBS_SPEC_INNER(item_flags),
 562                 .size = size,
 563         };
 564
 565         if (!mask)
 566                 mask = &rte_flow_item_ipv6_mask;
 567         if (spec) {
 568                 unsigned int i;
 569                 uint32_t vtc_flow_val;
 570                 uint32_t vtc_flow_mask;
 571
 572                 memcpy(&ipv6.val.src_ip, spec->hdr.src_addr,
 573                        RTE_DIM(ipv6.val.src_ip));
 574                 memcpy(&ipv6.val.dst_ip, spec->hdr.dst_addr,
 575                        RTE_DIM(ipv6.val.dst_ip));
 576                 memcpy(&ipv6.mask.src_ip, mask->hdr.src_addr,
 577                        RTE_DIM(ipv6.mask.src_ip));
 578                 memcpy(&ipv6.mask.dst_ip, mask->hdr.dst_addr,
 579                        RTE_DIM(ipv6.mask.dst_ip));
 580                 vtc_flow_val = rte_be_to_cpu_32(spec->hdr.vtc_flow);
 581                 vtc_flow_mask = rte_be_to_cpu_32(mask->hdr.vtc_flow);
 582                 ipv6.val.flow_label =
 583                         rte_cpu_to_be_32((vtc_flow_val & RTE_IPV6_HDR_FL_MASK) >>
 584                                          RTE_IPV6_HDR_FL_SHIFT);
 585                 ipv6.val.traffic_class = (vtc_flow_val & RTE_IPV6_HDR_TC_MASK) >>
 586                                          RTE_IPV6_HDR_TC_SHIFT;
 587                 ipv6.val.next_hdr = spec->hdr.proto;
 588                 ipv6.mask.flow_label =
 589                         rte_cpu_to_be_32((vtc_flow_mask & RTE_IPV6_HDR_FL_MASK) >>
 590                                          RTE_IPV6_HDR_FL_SHIFT);
 591                 ipv6.mask.traffic_class = (vtc_flow_mask & RTE_IPV6_HDR_TC_MASK) >>
 592                                           RTE_IPV6_HDR_TC_SHIFT;
 593                 ipv6.mask.next_hdr = mask->hdr.proto;
 594                 /* Remove unwanted bits from values. */
 595                 for (i = 0; i < RTE_DIM(ipv6.val.src_ip); ++i) {
 596                         ipv6.val.src_ip[i] &= ipv6.mask.src_ip[i];
 597                         ipv6.val.dst_ip[i] &= ipv6.mask.dst_ip[i];
 598                 }
 599                 ipv6.val.flow_label &= ipv6.mask.flow_label;
 600                 ipv6.val.traffic_class &= ipv6.mask.traffic_class;
 601                 ipv6.val.next_hdr &= ipv6.mask.next_hdr;
 602         }
 603         flow_verbs_spec_add(&dev_flow->verbs, &ipv6, size);
 604 }
 605
 606 /**
 607  * Convert the @p item into a Verbs specification. This function assumes that
 608  * the input is valid and that there is space to insert the requested item
 609  * into the flow.
 610  *
 611  * @param[in, out] dev_flow
 612  *   Pointer to dev_flow structure.
 613  * @param[in] item
 614  *   Item specification.
 615  * @param[in] item_flags
 616  *   Parsed item flags.
 617  */
 618 static void
 619 flow_verbs_translate_item_tcp(struct mlx5_flow *dev_flow,
 620                               const struct rte_flow_item *item,
 621                               uint64_t item_flags __rte_unused)
 622 {
 623         const struct rte_flow_item_tcp *spec = item->spec;
 624         const struct rte_flow_item_tcp *mask = item->mask;
 625         unsigned int size = sizeof(struct ibv_flow_spec_tcp_udp);
 626         struct ibv_flow_spec_tcp_udp tcp = {
 627                 .type = IBV_FLOW_SPEC_TCP | VERBS_SPEC_INNER(item_flags),
 628                 .size = size,
 629         };
 630
 631         if (!mask)
 632                 mask = &rte_flow_item_tcp_mask;
 633         if (spec) {
 634                 tcp.val.dst_port = spec->hdr.dst_port;
 635                 tcp.val.src_port = spec->hdr.src_port;
 636                 tcp.mask.dst_port = mask->hdr.dst_port;
 637                 tcp.mask.src_port = mask->hdr.src_port;
 638                 /* Remove unwanted bits from values. */
 639                 tcp.val.src_port &= tcp.mask.src_port;
 640                 tcp.val.dst_port &= tcp.mask.dst_port;
 641         }
 642         flow_verbs_spec_add(&dev_flow->verbs, &tcp, size);
 643 }
 644
 645 /**
 646  * Convert the @p item into a Verbs specification. This function assumes that
 647  * the input is valid and that there is space to insert the requested item
 648  * into the flow.
 649  *
 650  * @param[in, out] dev_flow
 651  *   Pointer to dev_flow structure.
 652  * @param[in] item
 653  *   Item specification.
 654  * @param[in] item_flags
 655  *   Parsed item flags.
 656  */
 657 static void
 658 flow_verbs_translate_item_udp(struct mlx5_flow *dev_flow,
 659                               const struct rte_flow_item *item,
 660                               uint64_t item_flags __rte_unused)
 661 {
 662         const struct rte_flow_item_udp *spec = item->spec;
 663         const struct rte_flow_item_udp *mask = item->mask;
 664         unsigned int size = sizeof(struct ibv_flow_spec_tcp_udp);
 665         struct ibv_flow_spec_tcp_udp udp = {
 666                 .type = IBV_FLOW_SPEC_UDP | VERBS_SPEC_INNER(item_flags),
 667                 .size = size,
 668         };
 669
 670         if (!mask)
 671                 mask = &rte_flow_item_udp_mask;
 672         if (spec) {
 673                 udp.val.dst_port = spec->hdr.dst_port;
 674                 udp.val.src_port = spec->hdr.src_port;
 675                 udp.mask.dst_port = mask->hdr.dst_port;
 676                 udp.mask.src_port = mask->hdr.src_port;
 677                 /* Remove unwanted bits from values. */
 678                 udp.val.src_port &= udp.mask.src_port;
 679                 udp.val.dst_port &= udp.mask.dst_port;
 680         }
 681         flow_verbs_spec_add(&dev_flow->verbs, &udp, size);
 682 }
 683
 684 /**
 685  * Convert the @p item into a Verbs specification. This function assumes that
 686  * the input is valid and that there is space to insert the requested item
 687  * into the flow.
 688  *
 689  * @param[in, out] dev_flow
 690  *   Pointer to dev_flow structure.
 691  * @param[in] item
 692  *   Item specification.
 693  * @param[in] item_flags
 694  *   Parsed item flags.
 695  */
 696 static void
 697 flow_verbs_translate_item_vxlan(struct mlx5_flow *dev_flow,
 698                                 const struct rte_flow_item *item,
 699                                 uint64_t item_flags __rte_unused)
 700 {
 701         const struct rte_flow_item_vxlan *spec = item->spec;
 702         const struct rte_flow_item_vxlan *mask = item->mask;
 703         unsigned int size = sizeof(struct ibv_flow_spec_tunnel);
 704         struct ibv_flow_spec_tunnel vxlan = {
 705                 .type = IBV_FLOW_SPEC_VXLAN_TUNNEL,
 706                 .size = size,
 707         };
 708         union vni {
 709                 uint32_t vlan_id;
 710                 uint8_t vni[4];
 711         } id = { .vlan_id = 0, };
 712
 713         if (!mask)
 714                 mask = &rte_flow_item_vxlan_mask;
 715         if (spec) {
 716                 memcpy(&id.vni[1], spec->vni, 3);
 717                 vxlan.val.tunnel_id = id.vlan_id;
 718                 memcpy(&id.vni[1], mask->vni, 3);
 719                 vxlan.mask.tunnel_id = id.vlan_id;
 720                 /* Remove unwanted bits from values. */
 721                 vxlan.val.tunnel_id &= vxlan.mask.tunnel_id;
 722         }
 723         flow_verbs_spec_add(&dev_flow->verbs, &vxlan, size);
 724 }
 725
 726 /**
 727  * Convert the @p item into a Verbs specification. This function assumes that
 728  * the input is valid and that there is space to insert the requested item
 729  * into the flow.
 730  *
 731  * @param[in, out] dev_flow
 732  *   Pointer to dev_flow structure.
 733  * @param[in] item
 734  *   Item specification.
 735  * @param[in] item_flags
 736  *   Parsed item flags.
 737  */
 738 static void
 739 flow_verbs_translate_item_vxlan_gpe(struct mlx5_flow *dev_flow,
 740                                     const struct rte_flow_item *item,
 741                                     uint64_t item_flags __rte_unused)
 742 {
 743         const struct rte_flow_item_vxlan_gpe *spec = item->spec;
 744         const struct rte_flow_item_vxlan_gpe *mask = item->mask;
 745         unsigned int size = sizeof(struct ibv_flow_spec_tunnel);
 746         struct ibv_flow_spec_tunnel vxlan_gpe = {
 747                 .type = IBV_FLOW_SPEC_VXLAN_TUNNEL,
 748                 .size = size,
 749         };
 750         union vni {
 751                 uint32_t vlan_id;
 752                 uint8_t vni[4];
 753         } id = { .vlan_id = 0, };
 754
 755         if (!mask)
 756                 mask = &rte_flow_item_vxlan_gpe_mask;
 757         if (spec) {
 758                 memcpy(&id.vni[1], spec->vni, 3);
 759                 vxlan_gpe.val.tunnel_id = id.vlan_id;
 760                 memcpy(&id.vni[1], mask->vni, 3);
 761                 vxlan_gpe.mask.tunnel_id = id.vlan_id;
 762                 /* Remove unwanted bits from values. */
 763                 vxlan_gpe.val.tunnel_id &= vxlan_gpe.mask.tunnel_id;
 764         }
 765         flow_verbs_spec_add(&dev_flow->verbs, &vxlan_gpe, size);
 766 }
 767
 768 /**
 769  * Update the protocol in Verbs IPv4/IPv6 spec.
 770  *
 771  * @param[in, out] attr
 772  *   Pointer to Verbs attributes structure.
 773  * @param[in] search
 774  *   Specification type to search in order to update the IP protocol.
 775  * @param[in] protocol
 776  *   Protocol value to set if none is present in the specification.
 777  */
 778 static void
 779 flow_verbs_item_gre_ip_protocol_update(struct ibv_flow_attr *attr,
 780                                        enum ibv_flow_spec_type search,
 781                                        uint8_t protocol)
 782 {
 783         unsigned int i;
 784         struct ibv_spec_header *hdr = (struct ibv_spec_header *)
 785                 ((uint8_t *)attr + sizeof(struct ibv_flow_attr));
 786
 787         if (!attr)
 788                 return;
 789         for (i = 0; i != attr->num_of_specs; ++i) {
 790                 if (hdr->type == search) {
 791                         union {
 792                                 struct ibv_flow_spec_ipv4_ext *ipv4;
 793                                 struct ibv_flow_spec_ipv6 *ipv6;
 794                         } ip;
 795
 796                         switch (search) {
 797                         case IBV_FLOW_SPEC_IPV4_EXT:
 798                                 ip.ipv4 = (struct ibv_flow_spec_ipv4_ext *)hdr;
 799                                 if (!ip.ipv4->val.proto) {
 800                                         ip.ipv4->val.proto = protocol;
 801                                         ip.ipv4->mask.proto = 0xff;
 802                                 }
 803                                 break;
 804                         case IBV_FLOW_SPEC_IPV6:
 805                                 ip.ipv6 = (struct ibv_flow_spec_ipv6 *)hdr;
 806                                 if (!ip.ipv6->val.next_hdr) {
 807                                         ip.ipv6->val.next_hdr = protocol;
 808                                         ip.ipv6->mask.next_hdr = 0xff;
 809                                 }
 810                                 break;
 811                         default:
 812                                 break;
 813                         }
 814                         break;
 815                 }
 816                 hdr = (struct ibv_spec_header *)((uint8_t *)hdr + hdr->size);
 817         }
 818 }
 819
 820 /**
 821  * Convert the @p item into a Verbs specification. This function assumes that
 822  * the input is valid and that there is space to insert the requested item
 823  * into the flow.
 824  *
 825  * @param[in, out] dev_flow
 826  *   Pointer to dev_flow structure.
 827  * @param[in] item
 828  *   Item specification.
 829  * @param[in] item_flags
 830  *   Parsed item flags.
 831  */
 832 static void
 833 flow_verbs_translate_item_gre(struct mlx5_flow *dev_flow,
 834                               const struct rte_flow_item *item __rte_unused,
 835                               uint64_t item_flags)
 836 {
 837         struct mlx5_flow_verbs_workspace *verbs = &dev_flow->verbs;
 838 #ifndef HAVE_IBV_DEVICE_MPLS_SUPPORT
 839         unsigned int size = sizeof(struct ibv_flow_spec_tunnel);
 840         struct ibv_flow_spec_tunnel tunnel = {
 841                 .type = IBV_FLOW_SPEC_VXLAN_TUNNEL,
 842                 .size = size,
 843         };
 844 #else
 845         const struct rte_flow_item_gre *spec = item->spec;
 846         const struct rte_flow_item_gre *mask = item->mask;
 847         unsigned int size = sizeof(struct ibv_flow_spec_gre);
 848         struct ibv_flow_spec_gre tunnel = {
 849                 .type = IBV_FLOW_SPEC_GRE,
 850                 .size = size,
 851         };
 852
 853         if (!mask)
 854                 mask = &rte_flow_item_gre_mask;
 855         if (spec) {
 856                 tunnel.val.c_ks_res0_ver = spec->c_rsvd0_ver;
 857                 tunnel.val.protocol = spec->protocol;
 858                 tunnel.mask.c_ks_res0_ver = mask->c_rsvd0_ver;
 859                 tunnel.mask.protocol = mask->protocol;
 860                 /* Remove unwanted bits from values. */
 861                 tunnel.val.c_ks_res0_ver &= tunnel.mask.c_ks_res0_ver;
 862                 tunnel.val.protocol &= tunnel.mask.protocol;
 863                 tunnel.val.key &= tunnel.mask.key;
 864         }
 865 #endif
 866         if (item_flags & MLX5_FLOW_LAYER_OUTER_L3_IPV4)
 867                 flow_verbs_item_gre_ip_protocol_update(&verbs->attr,
 868                                                        IBV_FLOW_SPEC_IPV4_EXT,
 869                                                        IPPROTO_GRE);
 870         else
 871                 flow_verbs_item_gre_ip_protocol_update(&verbs->attr,
 872                                                        IBV_FLOW_SPEC_IPV6,
 873                                                        IPPROTO_GRE);
 874         flow_verbs_spec_add(verbs, &tunnel, size);
 875 }
 876
 877 /**
 878  * Convert the @p action into a Verbs specification. This function assumes that
 879  * the input is valid and that there is space to insert the requested action
 880  * into the flow. This function also return the action that was added.
 881  *
 882  * @param[in, out] dev_flow
 883  *   Pointer to dev_flow structure.
 884  * @param[in] item
 885  *   Item specification.
 886  * @param[in] item_flags
 887  *   Parsed item flags.
 888  */
 889 static void
 890 flow_verbs_translate_item_mpls(struct mlx5_flow *dev_flow __rte_unused,
 891                                const struct rte_flow_item *item __rte_unused,
 892                                uint64_t item_flags __rte_unused)
 893 {
 894 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
 895         const struct rte_flow_item_mpls *spec = item->spec;
 896         const struct rte_flow_item_mpls *mask = item->mask;
 897         unsigned int size = sizeof(struct ibv_flow_spec_mpls);
 898         struct ibv_flow_spec_mpls mpls = {
 899                 .type = IBV_FLOW_SPEC_MPLS,
 900                 .size = size,
 901         };
 902
 903         if (!mask)
 904                 mask = &rte_flow_item_mpls_mask;
 905         if (spec) {
 906                 memcpy(&mpls.val.label, spec, sizeof(mpls.val.label));
 907                 memcpy(&mpls.mask.label, mask, sizeof(mpls.mask.label));
 908                 /* Remove unwanted bits from values.  */
 909                 mpls.val.label &= mpls.mask.label;
 910         }
 911         flow_verbs_spec_add(&dev_flow->verbs, &mpls, size);
 912 #endif
 913 }
 914
 915 /**
 916  * Convert the @p action into a Verbs specification. This function assumes that
 917  * the input is valid and that there is space to insert the requested action
 918  * into the flow.
 919  *
 920  * @param[in] dev_flow
 921  *   Pointer to mlx5_flow.
 922  * @param[in] action
 923  *   Action configuration.
 924  */
 925 static void
 926 flow_verbs_translate_action_drop
 927         (struct mlx5_flow *dev_flow,
 928          const struct rte_flow_action *action __rte_unused)
 929 {
 930         unsigned int size = sizeof(struct ibv_flow_spec_action_drop);
 931         struct ibv_flow_spec_action_drop drop = {
 932                         .type = IBV_FLOW_SPEC_ACTION_DROP,
 933                         .size = size,
 934         };
 935
 936         flow_verbs_spec_add(&dev_flow->verbs, &drop, size);
 937 }
 938
 939 /**
 940  * Convert the @p action into a Verbs specification. This function assumes that
 941  * the input is valid and that there is space to insert the requested action
 942  * into the flow.
 943  *
 944  * @param[in] rss_desc
 945  *   Pointer to mlx5_flow_rss_desc.
 946  * @param[in] action
 947  *   Action configuration.
 948  */
 949 static void
 950 flow_verbs_translate_action_queue(struct mlx5_flow_rss_desc *rss_desc,
 951                                   const struct rte_flow_action *action)
 952 {
 953         const struct rte_flow_action_queue *queue = action->conf;
 954
 955         rss_desc->queue[0] = queue->index;
 956         rss_desc->queue_num = 1;
 957 }
 958
 959 /**
 960  * Convert the @p action into a Verbs specification. This function assumes that
 961  * the input is valid and that there is space to insert the requested action
 962  * into the flow.
 963  *
 964  * @param[in] rss_desc
 965  *   Pointer to mlx5_flow_rss_desc.
 966  * @param[in] action
 967  *   Action configuration.
 968  */
 969 static void
 970 flow_verbs_translate_action_rss(struct mlx5_flow_rss_desc *rss_desc,
 971                                 const struct rte_flow_action *action)
 972 {
 973         const struct rte_flow_action_rss *rss = action->conf;
 974         const uint8_t *rss_key;
 975
 976         memcpy(rss_desc->queue, rss->queue, rss->queue_num * sizeof(uint16_t));
 977         rss_desc->queue_num = rss->queue_num;
 978         /* NULL RSS key indicates default RSS key. */
 979         rss_key = !rss->key ? rss_hash_default_key : rss->key;
 980         memcpy(rss_desc->key, rss_key, MLX5_RSS_HASH_KEY_LEN);
 981         /*
 982          * rss->level and rss.types should be set in advance when expanding
 983          * items for RSS.
 984          */
 985 }
 986
 987 /**
 988  * Convert the @p action into a Verbs specification. This function assumes that
 989  * the input is valid and that there is space to insert the requested action
 990  * into the flow.
 991  *
 992  * @param[in] dev_flow
 993  *   Pointer to mlx5_flow.
 994  * @param[in] action
 995  *   Action configuration.
 996  */
 997 static void
 998 flow_verbs_translate_action_flag
 999         (struct mlx5_flow *dev_flow,
1000          const struct rte_flow_action *action __rte_unused)
1001 {
1002         unsigned int size = sizeof(struct ibv_flow_spec_action_tag);
1003         struct ibv_flow_spec_action_tag tag = {
1004                 .type = IBV_FLOW_SPEC_ACTION_TAG,
1005                 .size = size,
1006                 .tag_id = mlx5_flow_mark_set(MLX5_FLOW_MARK_DEFAULT),
1007         };
1008
1009         flow_verbs_spec_add(&dev_flow->verbs, &tag, size);
1010 }
1011
1012 /**
1013  * Convert the @p action into a Verbs specification. This function assumes that
1014  * the input is valid and that there is space to insert the requested action
1015  * into the flow.
1016  *
1017  * @param[in] dev_flow
1018  *   Pointer to mlx5_flow.
1019  * @param[in] action
1020  *   Action configuration.
1021  */
1022 static void
1023 flow_verbs_translate_action_mark(struct mlx5_flow *dev_flow,
1024                                  const struct rte_flow_action *action)
1025 {
1026         const struct rte_flow_action_mark *mark = action->conf;
1027         unsigned int size = sizeof(struct ibv_flow_spec_action_tag);
1028         struct ibv_flow_spec_action_tag tag = {
1029                 .type = IBV_FLOW_SPEC_ACTION_TAG,
1030                 .size = size,
1031                 .tag_id = mlx5_flow_mark_set(mark->id),
1032         };
1033
1034         flow_verbs_spec_add(&dev_flow->verbs, &tag, size);
1035 }
1036
1037 /**
1038  * Convert the @p action into a Verbs specification. This function assumes that
1039  * the input is valid and that there is space to insert the requested action
1040  * into the flow.
1041  *
1042  * @param[in] dev
1043  *   Pointer to the Ethernet device structure.
1044  * @param[in] action
1045  *   Action configuration.
1046  * @param[in] dev_flow
1047  *   Pointer to mlx5_flow.
1048  * @param[out] error
1049  *   Pointer to error structure.
1050  *
1051  * @return
1052  *   0 On success else a negative errno value is returned and rte_errno is set.
1053  */
1054 static int
1055 flow_verbs_translate_action_count(struct mlx5_flow *dev_flow,
1056                                   const struct rte_flow_action *action,
1057                                   struct rte_eth_dev *dev,
1058                                   struct rte_flow_error *error)
1059 {
1060         const struct rte_flow_action_count *count = action->conf;
1061         struct rte_flow *flow = dev_flow->flow;
1062 #if defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) || \
1063         defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1064         struct mlx5_flow_counter_pool *pool;
1065         struct mlx5_flow_counter *cnt = NULL;
1066         struct mlx5_flow_counter_ext *cnt_ext;
1067         unsigned int size = sizeof(struct ibv_flow_spec_counter_action);
1068         struct ibv_flow_spec_counter_action counter = {
1069                 .type = IBV_FLOW_SPEC_ACTION_COUNT,
1070                 .size = size,
1071         };
1072 #endif
1073
1074         if (!flow->counter) {
1075                 flow->counter = flow_verbs_counter_new(dev, count->shared,
1076                                                        count->id);
1077                 if (!flow->counter)
1078                         return rte_flow_error_set(error, rte_errno,
1079                                                   RTE_FLOW_ERROR_TYPE_ACTION,
1080                                                   action,
1081                                                   "cannot get counter"
1082                                                   " context.");
1083         }
1084 #if defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42)
1085         cnt = flow_verbs_counter_get_by_idx(dev, flow->counter, &pool);
1086         cnt_ext = MLX5_CNT_TO_CNT_EXT(pool, cnt);
1087         counter.counter_set_handle = cnt_ext->cs->handle;
1088         flow_verbs_spec_add(&dev_flow->verbs, &counter, size);
1089 #elif defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1090         cnt = flow_verbs_counter_get_by_idx(dev, flow->counter, &pool);
1091         cnt_ext = MLX5_CNT_TO_CNT_EXT(pool, cnt);
1092         counter.counters = cnt_ext->cs;
1093         flow_verbs_spec_add(&dev_flow->verbs, &counter, size);
1094 #endif
1095         return 0;
1096 }
1097
1098 /**
1099  * Internal validation function. For validating both actions and items.
1100  *
1101  * @param[in] dev
1102  *   Pointer to the Ethernet device structure.
1103  * @param[in] attr
1104  *   Pointer to the flow attributes.
1105  * @param[in] items
1106  *   Pointer to the list of items.
1107  * @param[in] actions
1108  *   Pointer to the list of actions.
1109  * @param[in] external
1110  *   This flow rule is created by request external to PMD.
1111  * @param[in] hairpin
1112  *   Number of hairpin TX actions, 0 means classic flow.
1113  * @param[out] error
1114  *   Pointer to the error structure.
1115  *
1116  * @return
1117  *   0 on success, a negative errno value otherwise and rte_errno is set.
1118  */
1119 static int
1120 flow_verbs_validate(struct rte_eth_dev *dev,
1121                     const struct rte_flow_attr *attr,
1122                     const struct rte_flow_item items[],
1123                     const struct rte_flow_action actions[],
1124                     bool external __rte_unused,
1125                     int hairpin __rte_unused,
1126                     struct rte_flow_error *error)
1127 {
1128         int ret;
1129         uint64_t action_flags = 0;
1130         uint64_t item_flags = 0;
1131         uint64_t last_item = 0;
1132         uint8_t next_protocol = 0xff;
1133         uint16_t ether_type = 0;
1134
1135         if (items == NULL)
1136                 return -1;
1137         ret = mlx5_flow_validate_attributes(dev, attr, error);
1138         if (ret < 0)
1139                 return ret;
1140         for (; items->type != RTE_FLOW_ITEM_TYPE_END; items++) {
1141                 int tunnel = !!(item_flags & MLX5_FLOW_LAYER_TUNNEL);
1142                 int ret = 0;
1143
1144                 switch (items->type) {
1145                 case RTE_FLOW_ITEM_TYPE_VOID:
1146                         break;
1147                 case RTE_FLOW_ITEM_TYPE_ETH:
1148                         ret = mlx5_flow_validate_item_eth(items, item_flags,
1149                                                           error);
1150                         if (ret < 0)
1151                                 return ret;
1152                         last_item = tunnel ? MLX5_FLOW_LAYER_INNER_L2 :
1153                                              MLX5_FLOW_LAYER_OUTER_L2;
1154                         if (items->mask != NULL && items->spec != NULL) {
1155                                 ether_type =
1156                                         ((const struct rte_flow_item_eth *)
1157                                          items->spec)->type;
1158                                 ether_type &=
1159                                         ((const struct rte_flow_item_eth *)
1160                                          items->mask)->type;
1161                                 ether_type = rte_be_to_cpu_16(ether_type);
1162                         } else {
1163                                 ether_type = 0;
1164                         }
1165                         break;
1166                 case RTE_FLOW_ITEM_TYPE_VLAN:
1167                         ret = mlx5_flow_validate_item_vlan(items, item_flags,
1168                                                            dev, error);
1169                         if (ret < 0)
1170                                 return ret;
1171                         last_item = tunnel ? (MLX5_FLOW_LAYER_INNER_L2 |
1172                                               MLX5_FLOW_LAYER_INNER_VLAN) :
1173                                              (MLX5_FLOW_LAYER_OUTER_L2 |
1174                                               MLX5_FLOW_LAYER_OUTER_VLAN);
1175                         if (items->mask != NULL && items->spec != NULL) {
1176                                 ether_type =
1177                                         ((const struct rte_flow_item_vlan *)
1178                                          items->spec)->inner_type;
1179                                 ether_type &=
1180                                         ((const struct rte_flow_item_vlan *)
1181                                          items->mask)->inner_type;
1182                                 ether_type = rte_be_to_cpu_16(ether_type);
1183                         } else {
1184                                 ether_type = 0;
1185                         }
1186                         break;
1187                 case RTE_FLOW_ITEM_TYPE_IPV4:
1188                         ret = mlx5_flow_validate_item_ipv4(items, item_flags,
1189                                                            last_item,
1190                                                            ether_type, NULL,
1191                                                            error);
1192                         if (ret < 0)
1193                                 return ret;
1194                         last_item = tunnel ? MLX5_FLOW_LAYER_INNER_L3_IPV4 :
1195                                              MLX5_FLOW_LAYER_OUTER_L3_IPV4;
1196                         if (items->mask != NULL &&
1197                             ((const struct rte_flow_item_ipv4 *)
1198                              items->mask)->hdr.next_proto_id) {
1199                                 next_protocol =
1200                                         ((const struct rte_flow_item_ipv4 *)
1201                                          (items->spec))->hdr.next_proto_id;
1202                                 next_protocol &=
1203                                         ((const struct rte_flow_item_ipv4 *)
1204                                          (items->mask))->hdr.next_proto_id;
1205                         } else {
1206                                 /* Reset for inner layer. */
1207                                 next_protocol = 0xff;
1208                         }
1209                         break;
1210                 case RTE_FLOW_ITEM_TYPE_IPV6:
1211                         ret = mlx5_flow_validate_item_ipv6(items, item_flags,
1212                                                            last_item,
1213                                                            ether_type, NULL,
1214                                                            error);
1215                         if (ret < 0)
1216                                 return ret;
1217                         last_item = tunnel ? MLX5_FLOW_LAYER_INNER_L3_IPV6 :
1218                                              MLX5_FLOW_LAYER_OUTER_L3_IPV6;
1219                         if (items->mask != NULL &&
1220                             ((const struct rte_flow_item_ipv6 *)
1221                              items->mask)->hdr.proto) {
1222                                 next_protocol =
1223                                         ((const struct rte_flow_item_ipv6 *)
1224                                          items->spec)->hdr.proto;
1225                                 next_protocol &=
1226                                         ((const struct rte_flow_item_ipv6 *)
1227                                          items->mask)->hdr.proto;
1228                         } else {
1229                                 /* Reset for inner layer. */
1230                                 next_protocol = 0xff;
1231                         }
1232                         break;
1233                 case RTE_FLOW_ITEM_TYPE_UDP:
1234                         ret = mlx5_flow_validate_item_udp(items, item_flags,
1235                                                           next_protocol,
1236                                                           error);
1237                         if (ret < 0)
1238                                 return ret;
1239                         last_item = tunnel ? MLX5_FLOW_LAYER_INNER_L4_UDP :
1240                                              MLX5_FLOW_LAYER_OUTER_L4_UDP;
1241                         break;
1242                 case RTE_FLOW_ITEM_TYPE_TCP:
1243                         ret = mlx5_flow_validate_item_tcp
1244                                                 (items, item_flags,
1245                                                  next_protocol,
1246                                                  &rte_flow_item_tcp_mask,
1247                                                  error);
1248                         if (ret < 0)
1249                                 return ret;
1250                         last_item = tunnel ? MLX5_FLOW_LAYER_INNER_L4_TCP :
1251                                              MLX5_FLOW_LAYER_OUTER_L4_TCP;
1252                         break;
1253                 case RTE_FLOW_ITEM_TYPE_VXLAN:
1254                         ret = mlx5_flow_validate_item_vxlan(items, item_flags,
1255                                                             error);
1256                         if (ret < 0)
1257                                 return ret;
1258                         last_item = MLX5_FLOW_LAYER_VXLAN;
1259                         break;
1260                 case RTE_FLOW_ITEM_TYPE_VXLAN_GPE:
1261                         ret = mlx5_flow_validate_item_vxlan_gpe(items,
1262                                                                 item_flags,
1263                                                                 dev, error);
1264                         if (ret < 0)
1265                                 return ret;
1266                         last_item = MLX5_FLOW_LAYER_VXLAN_GPE;
1267                         break;
1268                 case RTE_FLOW_ITEM_TYPE_GRE:
1269                         ret = mlx5_flow_validate_item_gre(items, item_flags,
1270                                                           next_protocol, error);
1271                         if (ret < 0)
1272                                 return ret;
1273                         last_item = MLX5_FLOW_LAYER_GRE;
1274                         break;
1275                 case RTE_FLOW_ITEM_TYPE_MPLS:
1276                         ret = mlx5_flow_validate_item_mpls(dev, items,
1277                                                            item_flags,
1278                                                            last_item, error);
1279                         if (ret < 0)
1280                                 return ret;
1281                         last_item = MLX5_FLOW_LAYER_MPLS;
1282                         break;
1283                 default:
1284                         return rte_flow_error_set(error, ENOTSUP,
1285                                                   RTE_FLOW_ERROR_TYPE_ITEM,
1286                                                   NULL, "item not supported");
1287                 }
1288                 item_flags |= last_item;
1289         }
1290         for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
1291                 switch (actions->type) {
1292                 case RTE_FLOW_ACTION_TYPE_VOID:
1293                         break;
1294                 case RTE_FLOW_ACTION_TYPE_FLAG:
1295                         ret = mlx5_flow_validate_action_flag(action_flags,
1296                                                              attr,
1297                                                              error);
1298                         if (ret < 0)
1299                                 return ret;
1300                         action_flags |= MLX5_FLOW_ACTION_FLAG;
1301                         break;
1302                 case RTE_FLOW_ACTION_TYPE_MARK:
1303                         ret = mlx5_flow_validate_action_mark(actions,
1304                                                              action_flags,
1305                                                              attr,
1306                                                              error);
1307                         if (ret < 0)
1308                                 return ret;
1309                         action_flags |= MLX5_FLOW_ACTION_MARK;
1310                         break;
1311                 case RTE_FLOW_ACTION_TYPE_DROP:
1312                         ret = mlx5_flow_validate_action_drop(action_flags,
1313                                                              attr,
1314                                                              error);
1315                         if (ret < 0)
1316                                 return ret;
1317                         action_flags |= MLX5_FLOW_ACTION_DROP;
1318                         break;
1319                 case RTE_FLOW_ACTION_TYPE_QUEUE:
1320                         ret = mlx5_flow_validate_action_queue(actions,
1321                                                               action_flags, dev,
1322                                                               attr,
1323                                                               error);
1324                         if (ret < 0)
1325                                 return ret;
1326                         action_flags |= MLX5_FLOW_ACTION_QUEUE;
1327                         break;
1328                 case RTE_FLOW_ACTION_TYPE_RSS:
1329                         ret = mlx5_flow_validate_action_rss(actions,
1330                                                             action_flags, dev,
1331                                                             attr, item_flags,
1332                                                             error);
1333                         if (ret < 0)
1334                                 return ret;
1335                         action_flags |= MLX5_FLOW_ACTION_RSS;
1336                         break;
1337                 case RTE_FLOW_ACTION_TYPE_COUNT:
1338                         ret = mlx5_flow_validate_action_count(dev, attr, error);
1339                         if (ret < 0)
1340                                 return ret;
1341                         action_flags |= MLX5_FLOW_ACTION_COUNT;
1342                         break;
1343                 default:
1344                         return rte_flow_error_set(error, ENOTSUP,
1345                                                   RTE_FLOW_ERROR_TYPE_ACTION,
1346                                                   actions,
1347                                                   "action not supported");
1348                 }
1349         }
1350         /*
1351          * Validate the drop action mutual exclusion with other actions.
1352          * Drop action is mutually-exclusive with any other action, except for
1353          * Count action.
1354          */
1355         if ((action_flags & MLX5_FLOW_ACTION_DROP) &&
1356             (action_flags & ~(MLX5_FLOW_ACTION_DROP | MLX5_FLOW_ACTION_COUNT)))
1357                 return rte_flow_error_set(error, EINVAL,
1358                                           RTE_FLOW_ERROR_TYPE_ACTION, NULL,
1359                                           "Drop action is mutually-exclusive "
1360                                           "with any other action, except for "
1361                                           "Count action");
1362         if (!(action_flags & MLX5_FLOW_FATE_ACTIONS))
1363                 return rte_flow_error_set(error, EINVAL,
1364                                           RTE_FLOW_ERROR_TYPE_ACTION, actions,
1365                                           "no fate action is found");
1366         return 0;
1367 }
1368
1369 /**
1370  * Calculate the required bytes that are needed for the action part of the verbs
1371  * flow.
1372  *
1373  * @param[in] actions
1374  *   Pointer to the list of actions.
1375  *
1376  * @return
1377  *   The size of the memory needed for all actions.
1378  */
1379 static int
1380 flow_verbs_get_actions_size(const struct rte_flow_action actions[])
1381 {
1382         int size = 0;
1383
1384         for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
1385                 switch (actions->type) {
1386                 case RTE_FLOW_ACTION_TYPE_VOID:
1387                         break;
1388                 case RTE_FLOW_ACTION_TYPE_FLAG:
1389                         size += sizeof(struct ibv_flow_spec_action_tag);
1390                         break;
1391                 case RTE_FLOW_ACTION_TYPE_MARK:
1392                         size += sizeof(struct ibv_flow_spec_action_tag);
1393                         break;
1394                 case RTE_FLOW_ACTION_TYPE_DROP:
1395                         size += sizeof(struct ibv_flow_spec_action_drop);
1396                         break;
1397                 case RTE_FLOW_ACTION_TYPE_QUEUE:
1398                         break;
1399                 case RTE_FLOW_ACTION_TYPE_RSS:
1400                         break;
1401                 case RTE_FLOW_ACTION_TYPE_COUNT:
1402 #if defined(HAVE_IBV_DEVICE_COUNTERS_SET_V42) || \
1403         defined(HAVE_IBV_DEVICE_COUNTERS_SET_V45)
1404                         size += sizeof(struct ibv_flow_spec_counter_action);
1405 #endif
1406                         break;
1407                 default:
1408                         break;
1409                 }
1410         }
1411         return size;
1412 }
1413
1414 /**
1415  * Calculate the required bytes that are needed for the item part of the verbs
1416  * flow.
1417  *
1418  * @param[in] items
1419  *   Pointer to the list of items.
1420  *
1421  * @return
1422  *   The size of the memory needed for all items.
1423  */
1424 static int
1425 flow_verbs_get_items_size(const struct rte_flow_item items[])
1426 {
1427         int size = 0;
1428
1429         for (; items->type != RTE_FLOW_ITEM_TYPE_END; items++) {
1430                 switch (items->type) {
1431                 case RTE_FLOW_ITEM_TYPE_VOID:
1432                         break;
1433                 case RTE_FLOW_ITEM_TYPE_ETH:
1434                         size += sizeof(struct ibv_flow_spec_eth);
1435                         break;
1436                 case RTE_FLOW_ITEM_TYPE_VLAN:
1437                         size += sizeof(struct ibv_flow_spec_eth);
1438                         break;
1439                 case RTE_FLOW_ITEM_TYPE_IPV4:
1440                         size += sizeof(struct ibv_flow_spec_ipv4_ext);
1441                         break;
1442                 case RTE_FLOW_ITEM_TYPE_IPV6:
1443                         size += sizeof(struct ibv_flow_spec_ipv6);
1444                         break;
1445                 case RTE_FLOW_ITEM_TYPE_UDP:
1446                         size += sizeof(struct ibv_flow_spec_tcp_udp);
1447                         break;
1448                 case RTE_FLOW_ITEM_TYPE_TCP:
1449                         size += sizeof(struct ibv_flow_spec_tcp_udp);
1450                         break;
1451                 case RTE_FLOW_ITEM_TYPE_VXLAN:
1452                         size += sizeof(struct ibv_flow_spec_tunnel);
1453                         break;
1454                 case RTE_FLOW_ITEM_TYPE_VXLAN_GPE:
1455                         size += sizeof(struct ibv_flow_spec_tunnel);
1456                         break;
1457 #ifdef HAVE_IBV_DEVICE_MPLS_SUPPORT
1458                 case RTE_FLOW_ITEM_TYPE_GRE:
1459                         size += sizeof(struct ibv_flow_spec_gre);
1460                         break;
1461                 case RTE_FLOW_ITEM_TYPE_MPLS:
1462                         size += sizeof(struct ibv_flow_spec_mpls);
1463                         break;
1464 #else
1465                 case RTE_FLOW_ITEM_TYPE_GRE:
1466                         size += sizeof(struct ibv_flow_spec_tunnel);
1467                         break;
1468 #endif
1469                 default:
1470                         break;
1471                 }
1472         }
1473         return size;
1474 }
1475
1476 /**
1477  * Internal preparation function. Allocate mlx5_flow with the required size.
1478  * The required size is calculate based on the actions and items. This function
1479  * also returns the detected actions and items for later use.
1480  *
1481  * @param[in] dev
1482  *   Pointer to Ethernet device.
1483  * @param[in] attr
1484  *   Pointer to the flow attributes.
1485  * @param[in] items
1486  *   Pointer to the list of items.
1487  * @param[in] actions
1488  *   Pointer to the list of actions.
1489  * @param[out] error
1490  *   Pointer to the error structure.
1491  *
1492  * @return
1493  *   Pointer to mlx5_flow object on success, otherwise NULL and rte_errno
1494  *   is set.
1495  */
1496 static struct mlx5_flow *
1497 flow_verbs_prepare(struct rte_eth_dev *dev,
1498                    const struct rte_flow_attr *attr __rte_unused,
1499                    const struct rte_flow_item items[],
1500                    const struct rte_flow_action actions[],
1501                    struct rte_flow_error *error)
1502 {
1503         size_t size = 0;
1504         uint32_t handle_idx = 0;
1505         struct mlx5_flow *dev_flow;
1506         struct mlx5_flow_handle *dev_handle;
1507         struct mlx5_priv *priv = dev->data->dev_private;
1508
1509         size += flow_verbs_get_actions_size(actions);
1510         size += flow_verbs_get_items_size(items);
1511         if (size > MLX5_VERBS_MAX_SPEC_ACT_SIZE) {
1512                 rte_flow_error_set(error, E2BIG,
1513                                    RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
1514                                    "Verbs spec/action size too large");
1515                 return NULL;
1516         }
1517         /* In case of corrupting the memory. */
1518         if (priv->flow_idx >= MLX5_NUM_MAX_DEV_FLOWS) {
1519                 rte_flow_error_set(error, ENOSPC,
1520                                    RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
1521                                    "not free temporary device flow");
1522                 return NULL;
1523         }
1524         dev_handle = mlx5_ipool_zmalloc(priv->sh->ipool[MLX5_IPOOL_MLX5_FLOW],
1525                                    &handle_idx);
1526         if (!dev_handle) {
1527                 rte_flow_error_set(error, ENOMEM,
1528                                    RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
1529                                    "not enough memory to create flow handle");
1530                 return NULL;
1531         }
1532         /* No multi-thread supporting. */
1533         dev_flow = &((struct mlx5_flow *)priv->inter_flows)[priv->flow_idx++];
1534         dev_flow->handle = dev_handle;
1535         dev_flow->handle_idx = handle_idx;
1536         /* Memcpy is used, only size needs to be cleared to 0. */
1537         dev_flow->verbs.size = 0;
1538         dev_flow->verbs.attr.num_of_specs = 0;
1539         dev_flow->ingress = attr->ingress;
1540         /* Need to set transfer attribute: not supported in Verbs mode. */
1541         return dev_flow;
1542 }
1543
1544 /**
1545  * Fill the flow with verb spec.
1546  *
1547  * @param[in] dev
1548  *   Pointer to Ethernet device.
1549  * @param[in, out] dev_flow
1550  *   Pointer to the mlx5 flow.
1551  * @param[in] attr
1552  *   Pointer to the flow attributes.
1553  * @param[in] items
1554  *   Pointer to the list of items.
1555  * @param[in] actions
1556  *   Pointer to the list of actions.
1557  * @param[out] error
1558  *   Pointer to the error structure.
1559  *
1560  * @return
1561  *   0 on success, else a negative errno value otherwise and rte_errno is set.
1562  */
1563 static int
1564 flow_verbs_translate(struct rte_eth_dev *dev,
1565                      struct mlx5_flow *dev_flow,
1566                      const struct rte_flow_attr *attr,
1567                      const struct rte_flow_item items[],
1568                      const struct rte_flow_action actions[],
1569                      struct rte_flow_error *error)
1570 {
1571         uint64_t item_flags = 0;
1572         uint64_t action_flags = 0;
1573         uint64_t priority = attr->priority;
1574         uint32_t subpriority = 0;
1575         struct mlx5_priv *priv = dev->data->dev_private;
1576         struct mlx5_flow_rss_desc *rss_desc = &((struct mlx5_flow_rss_desc *)
1577                                               priv->rss_desc)
1578                                               [!!priv->flow_nested_idx];
1579
1580         if (priority == MLX5_FLOW_PRIO_RSVD)
1581                 priority = priv->config.flow_prio - 1;
1582         for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
1583                 int ret;
1584
1585                 switch (actions->type) {
1586                 case RTE_FLOW_ACTION_TYPE_VOID:
1587                         break;
1588                 case RTE_FLOW_ACTION_TYPE_FLAG:
1589                         flow_verbs_translate_action_flag(dev_flow, actions);
1590                         action_flags |= MLX5_FLOW_ACTION_FLAG;
1591                         dev_flow->handle->mark = 1;
1592                         break;
1593                 case RTE_FLOW_ACTION_TYPE_MARK:
1594                         flow_verbs_translate_action_mark(dev_flow, actions);
1595                         action_flags |= MLX5_FLOW_ACTION_MARK;
1596                         dev_flow->handle->mark = 1;
1597                         break;
1598                 case RTE_FLOW_ACTION_TYPE_DROP:
1599                         flow_verbs_translate_action_drop(dev_flow, actions);
1600                         action_flags |= MLX5_FLOW_ACTION_DROP;
1601                         dev_flow->handle->fate_action = MLX5_FLOW_FATE_DROP;
1602                         break;
1603                 case RTE_FLOW_ACTION_TYPE_QUEUE:
1604                         flow_verbs_translate_action_queue(rss_desc, actions);
1605                         action_flags |= MLX5_FLOW_ACTION_QUEUE;
1606                         dev_flow->handle->fate_action = MLX5_FLOW_FATE_QUEUE;
1607                         break;
1608                 case RTE_FLOW_ACTION_TYPE_RSS:
1609                         flow_verbs_translate_action_rss(rss_desc, actions);
1610                         action_flags |= MLX5_FLOW_ACTION_RSS;
1611                         dev_flow->handle->fate_action = MLX5_FLOW_FATE_QUEUE;
1612                         break;
1613                 case RTE_FLOW_ACTION_TYPE_COUNT:
1614                         ret = flow_verbs_translate_action_count(dev_flow,
1615                                                                 actions,
1616                                                                 dev, error);
1617                         if (ret < 0)
1618                                 return ret;
1619                         action_flags |= MLX5_FLOW_ACTION_COUNT;
1620                         break;
1621                 default:
1622                         return rte_flow_error_set(error, ENOTSUP,
1623                                                   RTE_FLOW_ERROR_TYPE_ACTION,
1624                                                   actions,
1625                                                   "action not supported");
1626                 }
1627         }
1628         dev_flow->act_flags = action_flags;
1629         for (; items->type != RTE_FLOW_ITEM_TYPE_END; items++) {
1630                 int tunnel = !!(item_flags & MLX5_FLOW_LAYER_TUNNEL);
1631
1632                 switch (items->type) {
1633                 case RTE_FLOW_ITEM_TYPE_VOID:
1634                         break;
1635                 case RTE_FLOW_ITEM_TYPE_ETH:
1636                         flow_verbs_translate_item_eth(dev_flow, items,
1637                                                       item_flags);
1638                         subpriority = MLX5_PRIORITY_MAP_L2;
1639                         item_flags |= tunnel ? MLX5_FLOW_LAYER_INNER_L2 :
1640                                                MLX5_FLOW_LAYER_OUTER_L2;
1641                         break;
1642                 case RTE_FLOW_ITEM_TYPE_VLAN:
1643                         flow_verbs_translate_item_vlan(dev_flow, items,
1644                                                        item_flags);
1645                         subpriority = MLX5_PRIORITY_MAP_L2;
1646                         item_flags |= tunnel ? (MLX5_FLOW_LAYER_INNER_L2 |
1647                                                 MLX5_FLOW_LAYER_INNER_VLAN) :
1648                                                (MLX5_FLOW_LAYER_OUTER_L2 |
1649                                                 MLX5_FLOW_LAYER_OUTER_VLAN);
1650                         break;
1651                 case RTE_FLOW_ITEM_TYPE_IPV4:
1652                         flow_verbs_translate_item_ipv4(dev_flow, items,
1653                                                        item_flags);
1654                         subpriority = MLX5_PRIORITY_MAP_L3;
1655                         dev_flow->hash_fields |=
1656                                 mlx5_flow_hashfields_adjust
1657                                         (rss_desc, tunnel,
1658                                          MLX5_IPV4_LAYER_TYPES,
1659                                          MLX5_IPV4_IBV_RX_HASH);
1660                         item_flags |= tunnel ? MLX5_FLOW_LAYER_INNER_L3_IPV4 :
1661                                                MLX5_FLOW_LAYER_OUTER_L3_IPV4;
1662                         break;
1663                 case RTE_FLOW_ITEM_TYPE_IPV6:
1664                         flow_verbs_translate_item_ipv6(dev_flow, items,
1665                                                        item_flags);
1666                         subpriority = MLX5_PRIORITY_MAP_L3;
1667                         dev_flow->hash_fields |=
1668                                 mlx5_flow_hashfields_adjust
1669                                         (rss_desc, tunnel,
1670                                          MLX5_IPV6_LAYER_TYPES,
1671                                          MLX5_IPV6_IBV_RX_HASH);
1672                         item_flags |= tunnel ? MLX5_FLOW_LAYER_INNER_L3_IPV6 :
1673                                                MLX5_FLOW_LAYER_OUTER_L3_IPV6;
1674                         break;
1675                 case RTE_FLOW_ITEM_TYPE_TCP:
1676                         flow_verbs_translate_item_tcp(dev_flow, items,
1677                                                       item_flags);
1678                         subpriority = MLX5_PRIORITY_MAP_L4;
1679                         dev_flow->hash_fields |=
1680                                 mlx5_flow_hashfields_adjust
1681                                         (rss_desc, tunnel, ETH_RSS_TCP,
1682                                          (IBV_RX_HASH_SRC_PORT_TCP |
1683                                           IBV_RX_HASH_DST_PORT_TCP));
1684                         item_flags |= tunnel ? MLX5_FLOW_LAYER_INNER_L4_TCP :
1685                                                MLX5_FLOW_LAYER_OUTER_L4_TCP;
1686                         break;
1687                 case RTE_FLOW_ITEM_TYPE_UDP:
1688                         flow_verbs_translate_item_udp(dev_flow, items,
1689                                                       item_flags);
1690                         subpriority = MLX5_PRIORITY_MAP_L4;
1691                         dev_flow->hash_fields |=
1692                                 mlx5_flow_hashfields_adjust
1693                                         (rss_desc, tunnel, ETH_RSS_UDP,
1694                                          (IBV_RX_HASH_SRC_PORT_UDP |
1695                                           IBV_RX_HASH_DST_PORT_UDP));
1696                         item_flags |= tunnel ? MLX5_FLOW_LAYER_INNER_L4_UDP :
1697                                                MLX5_FLOW_LAYER_OUTER_L4_UDP;
1698                         break;
1699                 case RTE_FLOW_ITEM_TYPE_VXLAN:
1700                         flow_verbs_translate_item_vxlan(dev_flow, items,
1701                                                         item_flags);
1702                         subpriority = MLX5_PRIORITY_MAP_L2;
1703                         item_flags |= MLX5_FLOW_LAYER_VXLAN;
1704                         break;
1705                 case RTE_FLOW_ITEM_TYPE_VXLAN_GPE:
1706                         flow_verbs_translate_item_vxlan_gpe(dev_flow, items,
1707                                                             item_flags);
1708                         subpriority = MLX5_PRIORITY_MAP_L2;
1709                         item_flags |= MLX5_FLOW_LAYER_VXLAN_GPE;
1710                         break;
1711                 case RTE_FLOW_ITEM_TYPE_GRE:
1712                         flow_verbs_translate_item_gre(dev_flow, items,
1713                                                       item_flags);
1714                         subpriority = MLX5_PRIORITY_MAP_L2;
1715                         item_flags |= MLX5_FLOW_LAYER_GRE;
1716                         break;
1717                 case RTE_FLOW_ITEM_TYPE_MPLS:
1718                         flow_verbs_translate_item_mpls(dev_flow, items,
1719                                                        item_flags);
1720                         subpriority = MLX5_PRIORITY_MAP_L2;
1721                         item_flags |= MLX5_FLOW_LAYER_MPLS;
1722                         break;
1723                 default:
1724                         return rte_flow_error_set(error, ENOTSUP,
1725                                                   RTE_FLOW_ERROR_TYPE_ITEM,
1726                                                   NULL,
1727                                                   "item not supported");
1728                 }
1729         }
1730         dev_flow->handle->layers = item_flags;
1731         /* Other members of attr will be ignored. */
1732         dev_flow->verbs.attr.priority =
1733                 mlx5_flow_adjust_priority(dev, priority, subpriority);
1734         dev_flow->verbs.attr.port = (uint8_t)priv->ibv_port;
1735         return 0;
1736 }
1737
1738 /**
1739  * Remove the flow from the NIC but keeps it in memory.
1740  *
1741  * @param[in] dev
1742  *   Pointer to the Ethernet device structure.
1743  * @param[in, out] flow
1744  *   Pointer to flow structure.
1745  */
1746 static void
1747 flow_verbs_remove(struct rte_eth_dev *dev, struct rte_flow *flow)
1748 {
1749         struct mlx5_priv *priv = dev->data->dev_private;
1750         struct mlx5_flow_handle *handle;
1751         uint32_t handle_idx;
1752
1753         if (!flow)
1754                 return;
1755         SILIST_FOREACH(priv->sh->ipool[MLX5_IPOOL_MLX5_FLOW], flow->dev_handles,
1756                        handle_idx, handle, next) {
1757                 if (handle->ib_flow) {
1758                         claim_zero(mlx5_glue->destroy_flow(handle->ib_flow));
1759                         handle->ib_flow = NULL;
1760                 }
1761                 /* hrxq is union, don't touch it only the flag is set. */
1762                 if (handle->rix_hrxq) {
1763                         if (handle->fate_action == MLX5_FLOW_FATE_DROP) {
1764                                 mlx5_hrxq_drop_release(dev);
1765                                 handle->rix_hrxq = 0;
1766                         } else if (handle->fate_action ==
1767                                    MLX5_FLOW_FATE_QUEUE) {
1768                                 mlx5_hrxq_release(dev, handle->rix_hrxq);
1769                                 handle->rix_hrxq = 0;
1770                         }
1771                 }
1772                 if (handle->vf_vlan.tag && handle->vf_vlan.created)
1773                         mlx5_vlan_vmwa_release(dev, &handle->vf_vlan);
1774         }
1775 }
1776
1777 /**
1778  * Remove the flow from the NIC and the memory.
1779  *
1780  * @param[in] dev
1781  *   Pointer to the Ethernet device structure.
1782  * @param[in, out] flow
1783  *   Pointer to flow structure.
1784  */
1785 static void
1786 flow_verbs_destroy(struct rte_eth_dev *dev, struct rte_flow *flow)
1787 {
1788         struct mlx5_priv *priv = dev->data->dev_private;
1789         struct mlx5_flow_handle *handle;
1790
1791         if (!flow)
1792                 return;
1793         flow_verbs_remove(dev, flow);
1794         while (flow->dev_handles) {
1795                 uint32_t tmp_idx = flow->dev_handles;
1796
1797                 handle = mlx5_ipool_get(priv->sh->ipool[MLX5_IPOOL_MLX5_FLOW],
1798                                    tmp_idx);
1799                 if (!handle)
1800                         return;
1801                 flow->dev_handles = handle->next.next;
1802                 mlx5_ipool_free(priv->sh->ipool[MLX5_IPOOL_MLX5_FLOW],
1803                            tmp_idx);
1804         }
1805         if (flow->counter) {
1806                 flow_verbs_counter_release(dev, flow->counter);
1807                 flow->counter = 0;
1808         }
1809 }
1810
1811 /**
1812  * Apply the flow to the NIC.
1813  *
1814  * @param[in] dev
1815  *   Pointer to the Ethernet device structure.
1816  * @param[in, out] flow
1817  *   Pointer to flow structure.
1818  * @param[out] error
1819  *   Pointer to error structure.
1820  *
1821  * @return
1822  *   0 on success, a negative errno value otherwise and rte_errno is set.
1823  */
1824 static int
1825 flow_verbs_apply(struct rte_eth_dev *dev, struct rte_flow *flow,
1826                  struct rte_flow_error *error)
1827 {
1828         struct mlx5_priv *priv = dev->data->dev_private;
1829         struct mlx5_flow_handle *handle;
1830         struct mlx5_flow *dev_flow;
1831         struct mlx5_hrxq *hrxq;
1832         uint32_t dev_handles;
1833         int err;
1834         int idx;
1835
1836         for (idx = priv->flow_idx - 1; idx >= priv->flow_nested_idx; idx--) {
1837                 dev_flow = &((struct mlx5_flow *)priv->inter_flows)[idx];
1838                 handle = dev_flow->handle;
1839                 if (handle->fate_action == MLX5_FLOW_FATE_DROP) {
1840                         hrxq = mlx5_hrxq_drop_new(dev);
1841                         if (!hrxq) {
1842                                 rte_flow_error_set
1843                                         (error, errno,
1844                                          RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
1845                                          "cannot get drop hash queue");
1846                                 goto error;
1847                         }
1848                 } else {
1849                         uint32_t hrxq_idx;
1850                         struct mlx5_flow_rss_desc *rss_desc =
1851                                 &((struct mlx5_flow_rss_desc *)priv->rss_desc)
1852                                 [!!priv->flow_nested_idx];
1853
1854                         MLX5_ASSERT(rss_desc->queue_num);
1855                         hrxq_idx = mlx5_hrxq_get(dev, rss_desc->key,
1856                                              MLX5_RSS_HASH_KEY_LEN,
1857                                              dev_flow->hash_fields,
1858                                              rss_desc->queue,
1859                                              rss_desc->queue_num);
1860                         if (!hrxq_idx)
1861                                 hrxq_idx = mlx5_hrxq_new(dev, rss_desc->key,
1862                                                 MLX5_RSS_HASH_KEY_LEN,
1863                                                 dev_flow->hash_fields,
1864                                                 rss_desc->queue,
1865                                                 rss_desc->queue_num,
1866                                                 !!(handle->layers &
1867                                                 MLX5_FLOW_LAYER_TUNNEL));
1868                         hrxq = mlx5_ipool_get(priv->sh->ipool[MLX5_IPOOL_HRXQ],
1869                                          hrxq_idx);
1870                         if (!hrxq) {
1871                                 rte_flow_error_set
1872                                         (error, rte_errno,
1873                                          RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
1874                                          "cannot get hash queue");
1875                                 goto error;
1876                         }
1877                         handle->rix_hrxq = hrxq_idx;
1878                 }
1879                 MLX5_ASSERT(hrxq);
1880                 handle->ib_flow = mlx5_glue->create_flow(hrxq->qp,
1881                                                      &dev_flow->verbs.attr);
1882                 if (!handle->ib_flow) {
1883                         rte_flow_error_set(error, errno,
1884                                            RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
1885                                            NULL,
1886                                            "hardware refuses to create flow");
1887                         goto error;
1888                 }
1889                 if (priv->vmwa_context &&
1890                     handle->vf_vlan.tag && !handle->vf_vlan.created) {
1891                         /*
1892                          * The rule contains the VLAN pattern.
1893                          * For VF we are going to create VLAN
1894                          * interface to make hypervisor set correct
1895                          * e-Switch vport context.
1896                          */
1897                         mlx5_vlan_vmwa_acquire(dev, &handle->vf_vlan);
1898                 }
1899         }
1900         return 0;
1901 error:
1902         err = rte_errno; /* Save rte_errno before cleanup. */
1903         SILIST_FOREACH(priv->sh->ipool[MLX5_IPOOL_MLX5_FLOW], flow->dev_handles,
1904                        dev_handles, handle, next) {
1905                 /* hrxq is union, don't touch it only the flag is set. */
1906                 if (handle->rix_hrxq) {
1907                         if (handle->fate_action == MLX5_FLOW_FATE_DROP) {
1908                                 mlx5_hrxq_drop_release(dev);
1909                                 handle->rix_hrxq = 0;
1910                         } else if (handle->fate_action ==
1911                                    MLX5_FLOW_FATE_QUEUE) {
1912                                 mlx5_hrxq_release(dev, handle->rix_hrxq);
1913                                 handle->rix_hrxq = 0;
1914                         }
1915                 }
1916                 if (handle->vf_vlan.tag && handle->vf_vlan.created)
1917                         mlx5_vlan_vmwa_release(dev, &handle->vf_vlan);
1918         }
1919         rte_errno = err; /* Restore rte_errno. */
1920         return -rte_errno;
1921 }
1922
1923 /**
1924  * Query a flow.
1925  *
1926  * @see rte_flow_query()
1927  * @see rte_flow_ops
1928  */
1929 static int
1930 flow_verbs_query(struct rte_eth_dev *dev,
1931                  struct rte_flow *flow,
1932                  const struct rte_flow_action *actions,
1933                  void *data,
1934                  struct rte_flow_error *error)
1935 {
1936         int ret = -EINVAL;
1937
1938         for (; actions->type != RTE_FLOW_ACTION_TYPE_END; actions++) {
1939                 switch (actions->type) {
1940                 case RTE_FLOW_ACTION_TYPE_VOID:
1941                         break;
1942                 case RTE_FLOW_ACTION_TYPE_COUNT:
1943                         ret = flow_verbs_counter_query(dev, flow, data, error);
1944                         break;
1945                 default:
1946                         return rte_flow_error_set(error, ENOTSUP,
1947                                                   RTE_FLOW_ERROR_TYPE_ACTION,
1948                                                   actions,
1949                                                   "action not supported");
1950                 }
1951         }
1952         return ret;
1953 }
1954
1955 const struct mlx5_flow_driver_ops mlx5_flow_verbs_drv_ops = {
1956         .validate = flow_verbs_validate,
1957         .prepare = flow_verbs_prepare,
1958         .translate = flow_verbs_translate,
1959         .apply = flow_verbs_apply,
1960         .remove = flow_verbs_remove,
1961         .destroy = flow_verbs_destroy,
1962         .query = flow_verbs_query,
1963 };