net/mlx5: make flow table cache thread safe
[dpdk.git] / drivers / net / mlx5 / mlx5_flow_dv.c
index 067ef0f..4c9e45f 100644 (file)
@@ -33,6 +33,7 @@
 #include "mlx5_flow.h"
 #include "mlx5_flow_os.h"
 #include "mlx5_rxtx.h"
+#include "rte_pmd_mlx5.h"
 
 #ifdef HAVE_IBV_FLOW_DV_SUPPORT
 
@@ -73,9 +74,6 @@ static int
 flow_dv_tbl_resource_release(struct rte_eth_dev *dev,
                             struct mlx5_flow_tbl_resource *tbl);
 
-static int
-flow_dv_default_miss_resource_release(struct rte_eth_dev *dev);
-
 static int
 flow_dv_encap_decap_resource_release(struct rte_eth_dev *dev,
                                      uint32_t encap_decap_idx);
@@ -293,7 +291,7 @@ flow_dv_shared_lock(struct rte_eth_dev *dev)
        struct mlx5_priv *priv = dev->data->dev_private;
        struct mlx5_dev_ctx_shared *sh = priv->sh;
 
-       if (sh->dv_refcnt > 1) {
+       if (sh->refcnt > 1) {
                int ret;
 
                ret = pthread_mutex_lock(&sh->dv_mutex);
@@ -308,7 +306,7 @@ flow_dv_shared_unlock(struct rte_eth_dev *dev)
        struct mlx5_priv *priv = dev->data->dev_private;
        struct mlx5_dev_ctx_shared *sh = priv->sh;
 
-       if (sh->dv_refcnt > 1) {
+       if (sh->refcnt > 1) {
                int ret;
 
                ret = pthread_mutex_unlock(&sh->dv_mutex);
@@ -1676,6 +1674,7 @@ flow_dv_validate_item_vlan(const struct rte_flow_item *item,
        const struct rte_flow_item_vlan nic_mask = {
                .tci = RTE_BE16(UINT16_MAX),
                .inner_type = RTE_BE16(UINT16_MAX),
+               .has_more_vlan = 1,
        };
        const int tunnel = !!(item_flags & MLX5_FLOW_LAYER_TUNNEL);
        int ret;
@@ -2872,8 +2871,10 @@ flow_dv_encap_decap_resource_register
                        struct mlx5_flow_dv_encap_decap_resource, entry);
                DRV_LOG(DEBUG, "encap/decap resource %p: refcnt %d++",
                        (void *)cache_resource,
-                       rte_atomic32_read(&cache_resource->refcnt));
-               rte_atomic32_inc(&cache_resource->refcnt);
+                       __atomic_load_n(&cache_resource->refcnt,
+                                       __ATOMIC_RELAXED));
+               __atomic_fetch_add(&cache_resource->refcnt, 1,
+                                  __ATOMIC_RELAXED);
                dev_flow->handle->dvh.rix_encap_decap = cache_resource->idx;
                dev_flow->dv.encap_decap = cache_resource;
                return 0;
@@ -2896,8 +2897,7 @@ flow_dv_encap_decap_resource_register
                                          RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
                                          NULL, "cannot create action");
        }
-       rte_atomic32_init(&cache_resource->refcnt);
-       rte_atomic32_inc(&cache_resource->refcnt);
+       __atomic_store_n(&cache_resource->refcnt, 1, __ATOMIC_RELAXED);
        if (mlx5_hlist_insert_ex(sh->encaps_decaps, &cache_resource->entry,
                                 flow_dv_encap_decap_resource_match,
                                 (void *)cache_resource)) {
@@ -2912,7 +2912,7 @@ flow_dv_encap_decap_resource_register
        dev_flow->dv.encap_decap = cache_resource;
        DRV_LOG(DEBUG, "new encap/decap resource %p: refcnt %d++",
                (void *)cache_resource,
-               rte_atomic32_read(&cache_resource->refcnt));
+               __atomic_load_n(&cache_resource->refcnt, __ATOMIC_RELAXED));
        return 0;
 }
 
@@ -2936,72 +2936,18 @@ flow_dv_jump_tbl_resource_register
                        (struct rte_eth_dev *dev __rte_unused,
                         struct mlx5_flow_tbl_resource *tbl,
                         struct mlx5_flow *dev_flow,
-                        struct rte_flow_error *error)
+                        struct rte_flow_error *error __rte_unused)
 {
        struct mlx5_flow_tbl_data_entry *tbl_data =
                container_of(tbl, struct mlx5_flow_tbl_data_entry, tbl);
-       int cnt, ret;
 
        MLX5_ASSERT(tbl);
-       cnt = rte_atomic32_read(&tbl_data->jump.refcnt);
-       if (!cnt) {
-               ret = mlx5_flow_os_create_flow_action_dest_flow_tbl
-                               (tbl->obj, &tbl_data->jump.action);
-               if (ret)
-                       return rte_flow_error_set(error, ENOMEM,
-                                       RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
-                                       NULL, "cannot create jump action");
-               DRV_LOG(DEBUG, "new jump table resource %p: refcnt %d++",
-                       (void *)&tbl_data->jump, cnt);
-       } else {
-               /* old jump should not make the table ref++. */
-               flow_dv_tbl_resource_release(dev, &tbl_data->tbl);
-               MLX5_ASSERT(tbl_data->jump.action);
-               DRV_LOG(DEBUG, "existed jump table resource %p: refcnt %d++",
-                       (void *)&tbl_data->jump, cnt);
-       }
-       rte_atomic32_inc(&tbl_data->jump.refcnt);
+       MLX5_ASSERT(tbl_data->jump.action);
        dev_flow->handle->rix_jump = tbl_data->idx;
        dev_flow->dv.jump = &tbl_data->jump;
        return 0;
 }
 
-/**
- * Find existing default miss resource or create and register a new one.
- *
- * @param[in, out] dev
- *   Pointer to rte_eth_dev structure.
- * @param[out] error
- *   pointer to error structure.
- *
- * @return
- *   0 on success otherwise -errno and errno is set.
- */
-static int
-flow_dv_default_miss_resource_register(struct rte_eth_dev *dev,
-               struct rte_flow_error *error)
-{
-       struct mlx5_priv *priv = dev->data->dev_private;
-       struct mlx5_dev_ctx_shared *sh = priv->sh;
-       struct mlx5_flow_default_miss_resource *cache_resource =
-                       &sh->default_miss;
-       int cnt = rte_atomic32_read(&cache_resource->refcnt);
-
-       if (!cnt) {
-               MLX5_ASSERT(cache_resource->action);
-               cache_resource->action =
-               mlx5_glue->dr_create_flow_action_default_miss();
-               if (!cache_resource->action)
-                       return rte_flow_error_set(error, ENOMEM,
-                                       RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
-                                       "cannot create default miss action");
-               DRV_LOG(DEBUG, "new default miss resource %p: refcnt %d++",
-                               (void *)cache_resource->action, cnt);
-       }
-       rte_atomic32_inc(&cache_resource->refcnt);
-       return 0;
-}
-
 /**
  * Find existing table port ID resource or create and register a new one.
  *
@@ -3037,8 +2983,10 @@ flow_dv_port_id_action_resource_register
                        DRV_LOG(DEBUG, "port id action resource resource %p: "
                                "refcnt %d++",
                                (void *)cache_resource,
-                               rte_atomic32_read(&cache_resource->refcnt));
-                       rte_atomic32_inc(&cache_resource->refcnt);
+                               __atomic_load_n(&cache_resource->refcnt,
+                                               __ATOMIC_RELAXED));
+                       __atomic_fetch_add(&cache_resource->refcnt, 1,
+                                          __ATOMIC_RELAXED);
                        dev_flow->handle->rix_port_id_action = idx;
                        dev_flow->dv.port_id_action = cache_resource;
                        return 0;
@@ -3061,15 +3009,14 @@ flow_dv_port_id_action_resource_register
                                          RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
                                          NULL, "cannot create action");
        }
-       rte_atomic32_init(&cache_resource->refcnt);
-       rte_atomic32_inc(&cache_resource->refcnt);
+       __atomic_store_n(&cache_resource->refcnt, 1, __ATOMIC_RELAXED);
        ILIST_INSERT(sh->ipool[MLX5_IPOOL_PORT_ID], &sh->port_id_action_list,
                     dev_flow->handle->rix_port_id_action, cache_resource,
                     next);
        dev_flow->dv.port_id_action = cache_resource;
        DRV_LOG(DEBUG, "new port id action resource %p: refcnt %d++",
                (void *)cache_resource,
-               rte_atomic32_read(&cache_resource->refcnt));
+               __atomic_load_n(&cache_resource->refcnt, __ATOMIC_RELAXED));
        return 0;
 }
 
@@ -3110,8 +3057,10 @@ flow_dv_push_vlan_action_resource_register
                        DRV_LOG(DEBUG, "push-VLAN action resource resource %p: "
                                "refcnt %d++",
                                (void *)cache_resource,
-                               rte_atomic32_read(&cache_resource->refcnt));
-                       rte_atomic32_inc(&cache_resource->refcnt);
+                               __atomic_load_n(&cache_resource->refcnt,
+                                               __ATOMIC_RELAXED));
+                       __atomic_fetch_add(&cache_resource->refcnt, 1,
+                                          __ATOMIC_RELAXED);
                        dev_flow->handle->dvh.rix_push_vlan = idx;
                        dev_flow->dv.push_vlan_res = cache_resource;
                        return 0;
@@ -3140,8 +3089,7 @@ flow_dv_push_vlan_action_resource_register
                                          RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
                                          NULL, "cannot create action");
        }
-       rte_atomic32_init(&cache_resource->refcnt);
-       rte_atomic32_inc(&cache_resource->refcnt);
+       __atomic_store_n(&cache_resource->refcnt, 1, __ATOMIC_RELAXED);
        ILIST_INSERT(sh->ipool[MLX5_IPOOL_PUSH_VLAN],
                     &sh->push_vlan_action_list,
                     dev_flow->handle->dvh.rix_push_vlan,
@@ -3149,7 +3097,7 @@ flow_dv_push_vlan_action_resource_register
        dev_flow->dv.push_vlan_res = cache_resource;
        DRV_LOG(DEBUG, "new push vlan action resource %p: refcnt %d++",
                (void *)cache_resource,
-               rte_atomic32_read(&cache_resource->refcnt));
+               __atomic_load_n(&cache_resource->refcnt, __ATOMIC_RELAXED));
        return 0;
 }
 /**
@@ -3947,14 +3895,21 @@ flow_dv_validate_action_modify_ttl(const uint64_t action_flags,
  *   0 on success, a negative errno value otherwise and rte_errno is set.
  */
 static int
-flow_dv_validate_action_jump(const struct rte_flow_action *action,
+flow_dv_validate_action_jump(struct rte_eth_dev *dev,
+                            const struct mlx5_flow_tunnel *tunnel,
+                            const struct rte_flow_action *action,
                             uint64_t action_flags,
                             const struct rte_flow_attr *attributes,
                             bool external, struct rte_flow_error *error)
 {
        uint32_t target_group, table;
        int ret = 0;
-
+       struct flow_grp_info grp_info = {
+               .external = !!external,
+               .transfer = !!attributes->transfer,
+               .fdb_def_rule = 1,
+               .std_tbl_fix = 0
+       };
        if (action_flags & (MLX5_FLOW_FATE_ACTIONS |
                            MLX5_FLOW_FATE_ESWITCH_ACTIONS))
                return rte_flow_error_set(error, EINVAL,
@@ -3977,11 +3932,13 @@ flow_dv_validate_action_jump(const struct rte_flow_action *action,
                                          NULL, "action configuration not set");
        target_group =
                ((const struct rte_flow_action_jump *)action->conf)->group;
-       ret = mlx5_flow_group_to_table(attributes, external, target_group,
-                                      true, &table, error);
+       ret = mlx5_flow_group_to_table(dev, tunnel, target_group, &table,
+                                      grp_info, error);
        if (ret)
                return ret;
-       if (attributes->group == target_group)
+       if (attributes->group == target_group &&
+           !(action_flags & (MLX5_FLOW_ACTION_TUNNEL_SET |
+                             MLX5_FLOW_ACTION_TUNNEL_MATCH)))
                return rte_flow_error_set(error, EINVAL,
                                          RTE_FLOW_ERROR_TYPE_ACTION, NULL,
                                          "target group must be other than"
@@ -4170,7 +4127,7 @@ flow_dv_validate_action_age(uint64_t action_flags,
        struct mlx5_priv *priv = dev->data->dev_private;
        const struct rte_flow_action_age *age = action->conf;
 
-       if (!priv->config.devx || priv->counter_fallback)
+       if (!priv->config.devx || priv->sh->cmng.counter_fallback)
                return rte_flow_error_set(error, ENOTSUP,
                                          RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
                                          NULL,
@@ -4540,8 +4497,10 @@ flow_dv_modify_hdr_resource_register
                                        entry);
                DRV_LOG(DEBUG, "modify-header resource %p: refcnt %d++",
                        (void *)cache_resource,
-                       rte_atomic32_read(&cache_resource->refcnt));
-               rte_atomic32_inc(&cache_resource->refcnt);
+                       __atomic_load_n(&cache_resource->refcnt,
+                                       __ATOMIC_RELAXED));
+               __atomic_fetch_add(&cache_resource->refcnt, 1,
+                                  __ATOMIC_RELAXED);
                dev_flow->handle->dvh.modify_hdr = cache_resource;
                return 0;
 
@@ -4565,8 +4524,7 @@ flow_dv_modify_hdr_resource_register
                                          RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
                                          NULL, "cannot create action");
        }
-       rte_atomic32_init(&cache_resource->refcnt);
-       rte_atomic32_inc(&cache_resource->refcnt);
+       __atomic_store_n(&cache_resource->refcnt, 1, __ATOMIC_RELAXED);
        if (mlx5_hlist_insert_ex(sh->modify_cmds, &cache_resource->entry,
                                 flow_dv_modify_hdr_resource_match,
                                 (void *)cache_resource)) {
@@ -4580,7 +4538,7 @@ flow_dv_modify_hdr_resource_register
        dev_flow->handle->dvh.modify_hdr = cache_resource;
        DRV_LOG(DEBUG, "new modify-header resource %p: refcnt %d++",
                (void *)cache_resource,
-               rte_atomic32_read(&cache_resource->refcnt));
+               __atomic_load_n(&cache_resource->refcnt, __ATOMIC_RELAXED));
        return 0;
 }
 
@@ -4739,16 +4697,13 @@ _flow_dv_query_count(struct rte_eth_dev *dev, uint32_t counter, uint64_t *pkts,
        struct mlx5_priv *priv = dev->data->dev_private;
        struct mlx5_flow_counter_pool *pool = NULL;
        struct mlx5_flow_counter *cnt;
-       struct mlx5_flow_counter_ext *cnt_ext = NULL;
        int offset;
 
        cnt = flow_dv_counter_get_by_idx(dev, counter, &pool);
        MLX5_ASSERT(pool);
-       if (priv->counter_fallback) {
-               cnt_ext = MLX5_CNT_TO_CNT_EXT(pool, cnt);
-               return mlx5_devx_cmd_flow_counter_query(cnt_ext->dcs, 0,
+       if (priv->sh->cmng.counter_fallback)
+               return mlx5_devx_cmd_flow_counter_query(cnt->dcs_when_active, 0,
                                        0, pkts, bytes, 0, NULL, NULL, 0);
-       }
        rte_spinlock_lock(&pool->sl);
        if (!pool->raw) {
                *pkts = 0;
@@ -4784,20 +4739,18 @@ flow_dv_pool_create(struct rte_eth_dev *dev, struct mlx5_devx_obj *dcs,
        struct mlx5_priv *priv = dev->data->dev_private;
        struct mlx5_flow_counter_pool *pool;
        struct mlx5_flow_counter_mng *cmng = &priv->sh->cmng;
-       uint32_t fallback = priv->counter_fallback;
+       bool fallback = priv->sh->cmng.counter_fallback;
        uint32_t size = sizeof(*pool);
 
-       size += MLX5_COUNTERS_PER_POOL * CNT_SIZE;
-       size += (!fallback ? 0 : MLX5_COUNTERS_PER_POOL * CNTEXT_SIZE);
-       size += (!age ? 0 : MLX5_COUNTERS_PER_POOL * AGE_SIZE);
+       size += MLX5_COUNTERS_PER_POOL * MLX5_CNT_SIZE;
+       size += (!age ? 0 : MLX5_COUNTERS_PER_POOL * MLX5_AGE_SIZE);
        pool = mlx5_malloc(MLX5_MEM_ZERO, size, 0, SOCKET_ID_ANY);
        if (!pool) {
                rte_errno = ENOMEM;
                return NULL;
        }
        pool->raw = NULL;
-       pool->type = 0;
-       pool->type |= (!age ? 0 :  CNT_POOL_TYPE_AGE);
+       pool->is_aged = !!age;
        pool->query_gen = 0;
        pool->min_dcs = dcs;
        rte_spinlock_init(&pool->sl);
@@ -4822,7 +4775,6 @@ flow_dv_pool_create(struct rte_eth_dev *dev, struct mlx5_devx_obj *dcs,
                if (base > cmng->max_id)
                        cmng->max_id = base + MLX5_COUNTERS_PER_POOL - 1;
                cmng->last_pool_idx = pool->index;
-               pool->type |= CNT_POOL_TYPE_EXT;
        }
        rte_spinlock_unlock(&cmng->pool_update_sl);
        return pool;
@@ -4855,7 +4807,7 @@ flow_dv_counter_pool_prepare(struct rte_eth_dev *dev,
        struct mlx5_flow_counter *cnt;
        enum mlx5_counter_type cnt_type =
                        age ? MLX5_COUNTER_TYPE_AGE : MLX5_COUNTER_TYPE_ORIGIN;
-       uint32_t fallback = priv->counter_fallback;
+       bool fallback = priv->sh->cmng.counter_fallback;
        uint32_t i;
 
        if (fallback) {
@@ -4874,7 +4826,7 @@ flow_dv_counter_pool_prepare(struct rte_eth_dev *dev,
                i = dcs->id % MLX5_COUNTERS_PER_POOL;
                cnt = MLX5_POOL_GET_CNT(pool, i);
                cnt->pool = pool;
-               MLX5_GET_POOL_CNT_EXT(pool, i)->dcs = dcs;
+               cnt->dcs_when_free = dcs;
                *cnt_free = cnt;
                return pool;
        }
@@ -4919,8 +4871,7 @@ flow_dv_counter_alloc(struct rte_eth_dev *dev, uint32_t age)
        struct mlx5_priv *priv = dev->data->dev_private;
        struct mlx5_flow_counter_pool *pool = NULL;
        struct mlx5_flow_counter *cnt_free = NULL;
-       struct mlx5_flow_counter_ext *cnt_ext = NULL;
-       uint32_t fallback = priv->counter_fallback;
+       bool fallback = priv->sh->cmng.counter_fallback;
        struct mlx5_flow_counter_mng *cmng = &priv->sh->cmng;
        enum mlx5_counter_type cnt_type =
                        age ? MLX5_COUNTER_TYPE_AGE : MLX5_COUNTER_TYPE_ORIGIN;
@@ -4940,7 +4891,7 @@ flow_dv_counter_alloc(struct rte_eth_dev *dev, uint32_t age)
                goto err;
        pool = cnt_free->pool;
        if (fallback)
-               cnt_ext = MLX5_CNT_TO_CNT_EXT(pool, cnt_free);
+               cnt_free->dcs_when_active = cnt_free->dcs_when_free;
        /* Create a DV counter action only in the first time usage. */
        if (!cnt_free->action) {
                uint16_t offset;
@@ -4952,7 +4903,7 @@ flow_dv_counter_alloc(struct rte_eth_dev *dev, uint32_t age)
                        dcs = pool->min_dcs;
                } else {
                        offset = 0;
-                       dcs = cnt_ext->dcs;
+                       dcs = cnt_free->dcs_when_free;
                }
                ret = mlx5_flow_os_create_flow_action_count(dcs->obj, offset,
                                                            &cnt_free->action);
@@ -4974,6 +4925,8 @@ flow_dv_counter_alloc(struct rte_eth_dev *dev, uint32_t age)
 err:
        if (cnt_free) {
                cnt_free->pool = pool;
+               if (fallback)
+                       cnt_free->dcs_when_free = cnt_free->dcs_when_active;
                rte_spinlock_lock(&cmng->csl[cnt_type]);
                TAILQ_INSERT_TAIL(&cmng->counters[cnt_type], cnt_free, next);
                rte_spinlock_unlock(&cmng->csl[cnt_type]);
@@ -5117,7 +5070,7 @@ flow_dv_counter_release(struct rte_eth_dev *dev, uint32_t counter)
        if (IS_SHARED_CNT(counter) &&
            mlx5_l3t_clear_entry(priv->sh->cnt_id_tbl, cnt->shared_info.id))
                return;
-       if (IS_AGE_POOL(pool))
+       if (pool->is_aged)
                flow_dv_counter_remove_from_age(dev, counter, cnt);
        cnt->pool = pool;
        /*
@@ -5130,13 +5083,14 @@ flow_dv_counter_release(struct rte_eth_dev *dev, uint32_t counter)
         * function both operate with the different list.
         *
         */
-       if (!priv->counter_fallback) {
+       if (!priv->sh->cmng.counter_fallback) {
                rte_spinlock_lock(&pool->csl);
                TAILQ_INSERT_TAIL(&pool->counters[pool->query_gen], cnt, next);
                rte_spinlock_unlock(&pool->csl);
        } else {
-               cnt_type = IS_AGE_POOL(pool) ? MLX5_COUNTER_TYPE_AGE :
-                                              MLX5_COUNTER_TYPE_ORIGIN;
+               cnt->dcs_when_free = cnt->dcs_when_active;
+               cnt_type = pool->is_aged ? MLX5_COUNTER_TYPE_AGE :
+                                          MLX5_COUNTER_TYPE_ORIGIN;
                rte_spinlock_lock(&priv->sh->cmng.csl[cnt_type]);
                TAILQ_INSERT_TAIL(&priv->sh->cmng.counters[cnt_type],
                                  cnt, next);
@@ -5164,8 +5118,9 @@ flow_dv_counter_release(struct rte_eth_dev *dev, uint32_t counter)
  */
 static int
 flow_dv_validate_attributes(struct rte_eth_dev *dev,
+                           const struct mlx5_flow_tunnel *tunnel,
                            const struct rte_flow_attr *attributes,
-                           bool external __rte_unused,
+                           struct flow_grp_info grp_info,
                            struct rte_flow_error *error)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
@@ -5173,6 +5128,8 @@ flow_dv_validate_attributes(struct rte_eth_dev *dev,
        int ret = 0;
 
 #ifndef HAVE_MLX5DV_DR
+       RTE_SET_USED(tunnel);
+       RTE_SET_USED(grp_info);
        if (attributes->group)
                return rte_flow_error_set(error, ENOTSUP,
                                          RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
@@ -5181,9 +5138,8 @@ flow_dv_validate_attributes(struct rte_eth_dev *dev,
 #else
        uint32_t table = 0;
 
-       ret = mlx5_flow_group_to_table(attributes, external,
-                                      attributes->group, !!priv->fdb_def_rule,
-                                      &table, error);
+       ret = mlx5_flow_group_to_table(dev, tunnel, attributes->group, &table,
+                                      grp_info, error);
        if (ret)
                return ret;
        if (!table)
@@ -5297,10 +5253,29 @@ flow_dv_validate(struct rte_eth_dev *dev, const struct rte_flow_attr *attr,
        const struct rte_flow_item_vlan *vlan_m = NULL;
        int16_t rw_act_num = 0;
        uint64_t is_root;
+       const struct mlx5_flow_tunnel *tunnel;
+       struct flow_grp_info grp_info = {
+               .external = !!external,
+               .transfer = !!attr->transfer,
+               .fdb_def_rule = !!priv->fdb_def_rule,
+       };
+       const struct rte_eth_hairpin_conf *conf;
 
        if (items == NULL)
                return -1;
-       ret = flow_dv_validate_attributes(dev, attr, external, error);
+       if (is_flow_tunnel_match_rule(dev, attr, items, actions)) {
+               tunnel = flow_items_to_tunnel(items);
+               action_flags |= MLX5_FLOW_ACTION_TUNNEL_MATCH |
+                               MLX5_FLOW_ACTION_DECAP;
+       } else if (is_flow_tunnel_steer_rule(dev, attr, items, actions)) {
+               tunnel = flow_actions_to_tunnel(actions);
+               action_flags |= MLX5_FLOW_ACTION_TUNNEL_SET;
+       } else {
+               tunnel = NULL;
+       }
+       grp_info.std_tbl_fix = tunnel_use_standard_attr_group_translate
+                               (dev, tunnel, attr, items, actions);
+       ret = flow_dv_validate_attributes(dev, tunnel, attr, grp_info, error);
        if (ret < 0)
                return ret;
        is_root = (uint64_t)ret;
@@ -5313,6 +5288,15 @@ flow_dv_validate(struct rte_eth_dev *dev, const struct rte_flow_attr *attr,
                                                  RTE_FLOW_ERROR_TYPE_ITEM,
                                                  NULL, "item not supported");
                switch (type) {
+               case MLX5_RTE_FLOW_ITEM_TYPE_TUNNEL:
+                       if (items[0].type != (typeof(items[0].type))
+                                               MLX5_RTE_FLOW_ITEM_TYPE_TUNNEL)
+                               return rte_flow_error_set
+                                               (error, EINVAL,
+                                               RTE_FLOW_ERROR_TYPE_ITEM,
+                                               NULL, "MLX5 private items "
+                                               "must be the first");
+                       break;
                case RTE_FLOW_ITEM_TYPE_VOID:
                        break;
                case RTE_FLOW_ITEM_TYPE_PORT_ID:
@@ -5324,7 +5308,7 @@ flow_dv_validate(struct rte_eth_dev *dev, const struct rte_flow_attr *attr,
                        break;
                case RTE_FLOW_ITEM_TYPE_ETH:
                        ret = mlx5_flow_validate_item_eth(items, item_flags,
-                                                         error);
+                                                         true, error);
                        if (ret < 0)
                                return ret;
                        last_item = tunnel ? MLX5_FLOW_LAYER_INNER_L2 :
@@ -5898,7 +5882,7 @@ flow_dv_validate(struct rte_eth_dev *dev, const struct rte_flow_attr *attr,
                        rw_act_num += MLX5_ACT_NUM_MDF_TTL;
                        break;
                case RTE_FLOW_ACTION_TYPE_JUMP:
-                       ret = flow_dv_validate_action_jump(actions,
+                       ret = flow_dv_validate_action_jump(dev, tunnel, actions,
                                                           action_flags,
                                                           attr, external,
                                                           error);
@@ -6007,6 +5991,17 @@ flow_dv_validate(struct rte_eth_dev *dev, const struct rte_flow_attr *attr,
                        action_flags |= MLX5_FLOW_ACTION_SAMPLE;
                        ++actions_n;
                        break;
+               case MLX5_RTE_FLOW_ACTION_TYPE_TUNNEL_SET:
+                       if (actions[0].type != (typeof(actions[0].type))
+                               MLX5_RTE_FLOW_ACTION_TYPE_TUNNEL_SET)
+                               return rte_flow_error_set
+                                               (error, EINVAL,
+                                               RTE_FLOW_ERROR_TYPE_ACTION,
+                                               NULL, "MLX5 private action "
+                                               "must be the first");
+
+                       action_flags |= MLX5_FLOW_ACTION_TUNNEL_SET;
+                       break;
                default:
                        return rte_flow_error_set(error, ENOTSUP,
                                                  RTE_FLOW_ERROR_TYPE_ACTION,
@@ -6014,6 +6009,54 @@ flow_dv_validate(struct rte_eth_dev *dev, const struct rte_flow_attr *attr,
                                                  "action not supported");
                }
        }
+       /*
+        * Validate actions in flow rules
+        * - Explicit decap action is prohibited by the tunnel offload API.
+        * - Drop action in tunnel steer rule is prohibited by the API.
+        * - Application cannot use MARK action because it's value can mask
+        *   tunnel default miss nitification.
+        * - JUMP in tunnel match rule has no support in current PMD
+        *   implementation.
+        * - TAG & META are reserved for future uses.
+        */
+       if (action_flags & MLX5_FLOW_ACTION_TUNNEL_SET) {
+               uint64_t bad_actions_mask = MLX5_FLOW_ACTION_DECAP    |
+                                           MLX5_FLOW_ACTION_MARK     |
+                                           MLX5_FLOW_ACTION_SET_TAG  |
+                                           MLX5_FLOW_ACTION_SET_META |
+                                           MLX5_FLOW_ACTION_DROP;
+
+               if (action_flags & bad_actions_mask)
+                       return rte_flow_error_set
+                                       (error, EINVAL,
+                                       RTE_FLOW_ERROR_TYPE_ACTION, NULL,
+                                       "Invalid RTE action in tunnel "
+                                       "set decap rule");
+               if (!(action_flags & MLX5_FLOW_ACTION_JUMP))
+                       return rte_flow_error_set
+                                       (error, EINVAL,
+                                       RTE_FLOW_ERROR_TYPE_ACTION, NULL,
+                                       "tunnel set decap rule must terminate "
+                                       "with JUMP");
+               if (!attr->ingress)
+                       return rte_flow_error_set
+                                       (error, EINVAL,
+                                       RTE_FLOW_ERROR_TYPE_ACTION, NULL,
+                                       "tunnel flows for ingress traffic only");
+       }
+       if (action_flags & MLX5_FLOW_ACTION_TUNNEL_MATCH) {
+               uint64_t bad_actions_mask = MLX5_FLOW_ACTION_JUMP    |
+                                           MLX5_FLOW_ACTION_MARK    |
+                                           MLX5_FLOW_ACTION_SET_TAG |
+                                           MLX5_FLOW_ACTION_SET_META;
+
+               if (action_flags & bad_actions_mask)
+                       return rte_flow_error_set
+                                       (error, EINVAL,
+                                       RTE_FLOW_ERROR_TYPE_ACTION, NULL,
+                                       "Invalid RTE action in tunnel "
+                                       "set match rule");
+       }
        /*
         * Validate the drop action mutual exclusion with other actions.
         * Drop action is mutually-exclusive with any other action, except for
@@ -6062,11 +6105,18 @@ flow_dv_validate(struct rte_eth_dev *dev, const struct rte_flow_attr *attr,
                                                  actions,
                                                  "no fate action is found");
        }
-       /* Continue validation for Xcap and VLAN actions.*/
+       /*
+        * Continue validation for Xcap and VLAN actions.
+        * If hairpin is working in explicit TX rule mode, there is no actions
+        * splitting and the validation of hairpin ingress flow should be the
+        * same as other standard flows.
+        */
        if ((action_flags & (MLX5_FLOW_XCAP_ACTIONS |
                             MLX5_FLOW_VLAN_ACTIONS)) &&
            (queue_index == 0xFFFF ||
-            mlx5_rxq_get_type(dev, queue_index) != MLX5_RXQ_TYPE_HAIRPIN)) {
+            mlx5_rxq_get_type(dev, queue_index) != MLX5_RXQ_TYPE_HAIRPIN ||
+            ((conf = mlx5_rxq_get_hairpin_conf(dev, queue_index)) != NULL &&
+            conf->tx_explicit != 0))) {
                if ((action_flags & MLX5_FLOW_XCAP_ACTIONS) ==
                    MLX5_FLOW_XCAP_ACTIONS)
                        return rte_flow_error_set(error, ENOTSUP,
@@ -6095,7 +6145,10 @@ flow_dv_validate(struct rte_eth_dev *dev, const struct rte_flow_attr *attr,
                                                 "multiple VLAN actions");
                }
        }
-       /* Hairpin flow will add one more TAG action. */
+       /*
+        * Hairpin flow will add one more TAG action in TX implicit mode.
+        * In TX explicit mode, there will be no hairpin flow ID.
+        */
        if (hairpin > 0)
                rw_act_num += MLX5_ACT_NUM_SET_TAG;
        /* extra metadata enabled: one more TAG action will be add. */
@@ -6143,9 +6196,11 @@ flow_dv_prepare(struct rte_eth_dev *dev,
        struct mlx5_flow *dev_flow;
        struct mlx5_flow_handle *dev_handle;
        struct mlx5_priv *priv = dev->data->dev_private;
+       struct mlx5_flow_workspace *wks = mlx5_flow_get_thread_workspace();
 
+       MLX5_ASSERT(wks);
        /* In case of corrupting the memory. */
-       if (priv->flow_idx >= MLX5_NUM_MAX_DEV_FLOWS) {
+       if (wks->flow_idx >= MLX5_NUM_MAX_DEV_FLOWS) {
                rte_flow_error_set(error, ENOSPC,
                                   RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
                                   "not free temporary device flow");
@@ -6159,8 +6214,8 @@ flow_dv_prepare(struct rte_eth_dev *dev,
                                   "not enough memory to create flow handle");
                return NULL;
        }
-       /* No multi-thread supporting. */
-       dev_flow = &((struct mlx5_flow *)priv->inter_flows)[priv->flow_idx++];
+       MLX5_ASSERT(wks->flow_idx + 1 < RTE_DIM(wks->flows));
+       dev_flow = &wks->flows[wks->flow_idx++];
        dev_flow->handle = dev_handle;
        dev_flow->handle_idx = handle_idx;
        /*
@@ -6267,9 +6322,10 @@ flow_dv_translate_item_eth(void *matcher, void *key,
                .dst.addr_bytes = "\xff\xff\xff\xff\xff\xff",
                .src.addr_bytes = "\xff\xff\xff\xff\xff\xff",
                .type = RTE_BE16(0xffff),
+               .has_vlan = 0,
        };
-       void *headers_m;
-       void *headers_v;
+       void *hdrs_m;
+       void *hdrs_v;
        char *l24_v;
        unsigned int i;
 
@@ -6278,38 +6334,26 @@ flow_dv_translate_item_eth(void *matcher, void *key,
        if (!eth_m)
                eth_m = &nic_mask;
        if (inner) {
-               headers_m = MLX5_ADDR_OF(fte_match_param, matcher,
+               hdrs_m = MLX5_ADDR_OF(fte_match_param, matcher,
                                         inner_headers);
-               headers_v = MLX5_ADDR_OF(fte_match_param, key, inner_headers);
+               hdrs_v = MLX5_ADDR_OF(fte_match_param, key, inner_headers);
        } else {
-               headers_m = MLX5_ADDR_OF(fte_match_param, matcher,
+               hdrs_m = MLX5_ADDR_OF(fte_match_param, matcher,
                                         outer_headers);
-               headers_v = MLX5_ADDR_OF(fte_match_param, key, outer_headers);
+               hdrs_v = MLX5_ADDR_OF(fte_match_param, key, outer_headers);
        }
-       memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_m, dmac_47_16),
+       memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, hdrs_m, dmac_47_16),
               &eth_m->dst, sizeof(eth_m->dst));
        /* The value must be in the range of the mask. */
-       l24_v = MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v, dmac_47_16);
+       l24_v = MLX5_ADDR_OF(fte_match_set_lyr_2_4, hdrs_v, dmac_47_16);
        for (i = 0; i < sizeof(eth_m->dst); ++i)
                l24_v[i] = eth_m->dst.addr_bytes[i] & eth_v->dst.addr_bytes[i];
-       memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_m, smac_47_16),
+       memcpy(MLX5_ADDR_OF(fte_match_set_lyr_2_4, hdrs_m, smac_47_16),
               &eth_m->src, sizeof(eth_m->src));
-       l24_v = MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v, smac_47_16);
+       l24_v = MLX5_ADDR_OF(fte_match_set_lyr_2_4, hdrs_v, smac_47_16);
        /* The value must be in the range of the mask. */
        for (i = 0; i < sizeof(eth_m->dst); ++i)
                l24_v[i] = eth_m->src.addr_bytes[i] & eth_v->src.addr_bytes[i];
-       if (eth_v->type) {
-               /* When ethertype is present set mask for tagged VLAN. */
-               MLX5_SET(fte_match_set_lyr_2_4, headers_m, cvlan_tag, 1);
-               /* Set value for tagged VLAN if ethertype is 802.1Q. */
-               if (eth_v->type == RTE_BE16(RTE_ETHER_TYPE_VLAN) ||
-                   eth_v->type == RTE_BE16(RTE_ETHER_TYPE_QINQ)) {
-                       MLX5_SET(fte_match_set_lyr_2_4, headers_v, cvlan_tag,
-                                1);
-                       /* Return here to avoid setting match on ethertype. */
-                       return;
-               }
-       }
        /*
         * HW supports match on one Ethertype, the Ethertype following the last
         * VLAN tag of the packet (see PRM).
@@ -6318,19 +6362,42 @@ flow_dv_translate_item_eth(void *matcher, void *key,
         * ethertype, and use ip_version field instead.
         * eCPRI over Ether layer will use type value 0xAEFE.
         */
-       if (eth_v->type == RTE_BE16(RTE_ETHER_TYPE_IPV4) &&
-           eth_m->type == 0xFFFF) {
-               flow_dv_set_match_ip_version(group, headers_v, headers_m, 4);
-       } else if (eth_v->type == RTE_BE16(RTE_ETHER_TYPE_IPV6) &&
-                  eth_m->type == 0xFFFF) {
-               flow_dv_set_match_ip_version(group, headers_v, headers_m, 6);
-       } else {
-               MLX5_SET(fte_match_set_lyr_2_4, headers_m, ethertype,
-                        rte_be_to_cpu_16(eth_m->type));
-               l24_v = MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
-                                    ethertype);
-               *(uint16_t *)(l24_v) = eth_m->type & eth_v->type;
+       if (eth_m->type == 0xFFFF) {
+               /* Set cvlan_tag mask for any single\multi\un-tagged case. */
+               MLX5_SET(fte_match_set_lyr_2_4, hdrs_m, cvlan_tag, 1);
+               switch (eth_v->type) {
+               case RTE_BE16(RTE_ETHER_TYPE_VLAN):
+                       MLX5_SET(fte_match_set_lyr_2_4, hdrs_v, cvlan_tag, 1);
+                       return;
+               case RTE_BE16(RTE_ETHER_TYPE_QINQ):
+                       MLX5_SET(fte_match_set_lyr_2_4, hdrs_m, svlan_tag, 1);
+                       MLX5_SET(fte_match_set_lyr_2_4, hdrs_v, svlan_tag, 1);
+                       return;
+               case RTE_BE16(RTE_ETHER_TYPE_IPV4):
+                       flow_dv_set_match_ip_version(group, hdrs_v, hdrs_m, 4);
+                       return;
+               case RTE_BE16(RTE_ETHER_TYPE_IPV6):
+                       flow_dv_set_match_ip_version(group, hdrs_v, hdrs_m, 6);
+                       return;
+               default:
+                       break;
+               }
+       }
+       if (eth_m->has_vlan) {
+               MLX5_SET(fte_match_set_lyr_2_4, hdrs_m, cvlan_tag, 1);
+               if (eth_v->has_vlan) {
+                       /*
+                        * Here, when also has_more_vlan field in VLAN item is
+                        * not set, only single-tagged packets will be matched.
+                        */
+                       MLX5_SET(fte_match_set_lyr_2_4, hdrs_v, cvlan_tag, 1);
+                       return;
+               }
        }
+       MLX5_SET(fte_match_set_lyr_2_4, hdrs_m, ethertype,
+                rte_be_to_cpu_16(eth_m->type));
+       l24_v = MLX5_ADDR_OF(fte_match_set_lyr_2_4, hdrs_v, ethertype);
+       *(uint16_t *)(l24_v) = eth_m->type & eth_v->type;
 }
 
 /**
@@ -6355,19 +6422,19 @@ flow_dv_translate_item_vlan(struct mlx5_flow *dev_flow,
 {
        const struct rte_flow_item_vlan *vlan_m = item->mask;
        const struct rte_flow_item_vlan *vlan_v = item->spec;
-       void *headers_m;
-       void *headers_v;
+       void *hdrs_m;
+       void *hdrs_v;
        uint16_t tci_m;
        uint16_t tci_v;
 
        if (inner) {
-               headers_m = MLX5_ADDR_OF(fte_match_param, matcher,
+               hdrs_m = MLX5_ADDR_OF(fte_match_param, matcher,
                                         inner_headers);
-               headers_v = MLX5_ADDR_OF(fte_match_param, key, inner_headers);
+               hdrs_v = MLX5_ADDR_OF(fte_match_param, key, inner_headers);
        } else {
-               headers_m = MLX5_ADDR_OF(fte_match_param, matcher,
+               hdrs_m = MLX5_ADDR_OF(fte_match_param, matcher,
                                         outer_headers);
-               headers_v = MLX5_ADDR_OF(fte_match_param, key, outer_headers);
+               hdrs_v = MLX5_ADDR_OF(fte_match_param, key, outer_headers);
                /*
                 * This is workaround, masks are not supported,
                 * and pre-validated.
@@ -6380,37 +6447,54 @@ flow_dv_translate_item_vlan(struct mlx5_flow *dev_flow,
         * When VLAN item exists in flow, mark packet as tagged,
         * even if TCI is not specified.
         */
-       MLX5_SET(fte_match_set_lyr_2_4, headers_m, cvlan_tag, 1);
-       MLX5_SET(fte_match_set_lyr_2_4, headers_v, cvlan_tag, 1);
+       if (!MLX5_GET(fte_match_set_lyr_2_4, hdrs_v, svlan_tag)) {
+               MLX5_SET(fte_match_set_lyr_2_4, hdrs_m, cvlan_tag, 1);
+               MLX5_SET(fte_match_set_lyr_2_4, hdrs_v, cvlan_tag, 1);
+       }
        if (!vlan_v)
                return;
        if (!vlan_m)
                vlan_m = &rte_flow_item_vlan_mask;
        tci_m = rte_be_to_cpu_16(vlan_m->tci);
        tci_v = rte_be_to_cpu_16(vlan_m->tci & vlan_v->tci);
-       MLX5_SET(fte_match_set_lyr_2_4, headers_m, first_vid, tci_m);
-       MLX5_SET(fte_match_set_lyr_2_4, headers_v, first_vid, tci_v);
-       MLX5_SET(fte_match_set_lyr_2_4, headers_m, first_cfi, tci_m >> 12);
-       MLX5_SET(fte_match_set_lyr_2_4, headers_v, first_cfi, tci_v >> 12);
-       MLX5_SET(fte_match_set_lyr_2_4, headers_m, first_prio, tci_m >> 13);
-       MLX5_SET(fte_match_set_lyr_2_4, headers_v, first_prio, tci_v >> 13);
+       MLX5_SET(fte_match_set_lyr_2_4, hdrs_m, first_vid, tci_m);
+       MLX5_SET(fte_match_set_lyr_2_4, hdrs_v, first_vid, tci_v);
+       MLX5_SET(fte_match_set_lyr_2_4, hdrs_m, first_cfi, tci_m >> 12);
+       MLX5_SET(fte_match_set_lyr_2_4, hdrs_v, first_cfi, tci_v >> 12);
+       MLX5_SET(fte_match_set_lyr_2_4, hdrs_m, first_prio, tci_m >> 13);
+       MLX5_SET(fte_match_set_lyr_2_4, hdrs_v, first_prio, tci_v >> 13);
        /*
         * HW is optimized for IPv4/IPv6. In such cases, avoid setting
         * ethertype, and use ip_version field instead.
         */
-       if (vlan_v->inner_type == RTE_BE16(RTE_ETHER_TYPE_IPV4) &&
-           vlan_m->inner_type == 0xFFFF) {
-               flow_dv_set_match_ip_version(group, headers_v, headers_m, 4);
-       } else if (vlan_v->inner_type == RTE_BE16(RTE_ETHER_TYPE_IPV6) &&
-                  vlan_m->inner_type == 0xFFFF) {
-               flow_dv_set_match_ip_version(group, headers_v, headers_m, 6);
-       } else {
-               MLX5_SET(fte_match_set_lyr_2_4, headers_m, ethertype,
-                        rte_be_to_cpu_16(vlan_m->inner_type));
-               MLX5_SET(fte_match_set_lyr_2_4, headers_v, ethertype,
-                        rte_be_to_cpu_16(vlan_m->inner_type &
-                                         vlan_v->inner_type));
+       if (vlan_m->inner_type == 0xFFFF) {
+               switch (vlan_v->inner_type) {
+               case RTE_BE16(RTE_ETHER_TYPE_VLAN):
+                       MLX5_SET(fte_match_set_lyr_2_4, hdrs_m, svlan_tag, 1);
+                       MLX5_SET(fte_match_set_lyr_2_4, hdrs_v, svlan_tag, 1);
+                       MLX5_SET(fte_match_set_lyr_2_4, hdrs_v, cvlan_tag, 0);
+                       return;
+               case RTE_BE16(RTE_ETHER_TYPE_IPV4):
+                       flow_dv_set_match_ip_version(group, hdrs_v, hdrs_m, 4);
+                       return;
+               case RTE_BE16(RTE_ETHER_TYPE_IPV6):
+                       flow_dv_set_match_ip_version(group, hdrs_v, hdrs_m, 6);
+                       return;
+               default:
+                       break;
+               }
+       }
+       if (vlan_m->has_more_vlan && vlan_v->has_more_vlan) {
+               MLX5_SET(fte_match_set_lyr_2_4, hdrs_m, svlan_tag, 1);
+               MLX5_SET(fte_match_set_lyr_2_4, hdrs_v, svlan_tag, 1);
+               /* Only one vlan_tag bit can be set. */
+               MLX5_SET(fte_match_set_lyr_2_4, hdrs_v, cvlan_tag, 0);
+               return;
        }
+       MLX5_SET(fte_match_set_lyr_2_4, hdrs_m, ethertype,
+                rte_be_to_cpu_16(vlan_m->inner_type));
+       MLX5_SET(fte_match_set_lyr_2_4, hdrs_v, ethertype,
+                rte_be_to_cpu_16(vlan_m->inner_type & vlan_v->inner_type));
 }
 
 /**
@@ -6422,8 +6506,6 @@ flow_dv_translate_item_vlan(struct mlx5_flow *dev_flow,
  *   Flow matcher value.
  * @param[in] item
  *   Flow pattern to translate.
- * @param[in] item_flags
- *   Bit-fields that holds the items detected until now.
  * @param[in] inner
  *   Item is inner pattern.
  * @param[in] group
@@ -6432,7 +6514,6 @@ flow_dv_translate_item_vlan(struct mlx5_flow *dev_flow,
 static void
 flow_dv_translate_item_ipv4(void *matcher, void *key,
                            const struct rte_flow_item *item,
-                           const uint64_t item_flags,
                            int inner, uint32_t group)
 {
        const struct rte_flow_item_ipv4 *ipv4_m = item->mask;
@@ -6462,13 +6543,6 @@ flow_dv_translate_item_ipv4(void *matcher, void *key,
                headers_v = MLX5_ADDR_OF(fte_match_param, key, outer_headers);
        }
        flow_dv_set_match_ip_version(group, headers_v, headers_m, 4);
-       /*
-        * On outer header (which must contains L2), or inner header with L2,
-        * set cvlan_tag mask bit to mark this packet as untagged.
-        * This should be done even if item->spec is empty.
-        */
-       if (!inner || item_flags & MLX5_FLOW_LAYER_INNER_L2)
-               MLX5_SET(fte_match_set_lyr_2_4, headers_m, cvlan_tag, 1);
        if (!ipv4_v)
                return;
        if (!ipv4_m)
@@ -6515,8 +6589,6 @@ flow_dv_translate_item_ipv4(void *matcher, void *key,
  *   Flow matcher value.
  * @param[in] item
  *   Flow pattern to translate.
- * @param[in] item_flags
- *   Bit-fields that holds the items detected until now.
  * @param[in] inner
  *   Item is inner pattern.
  * @param[in] group
@@ -6525,7 +6597,6 @@ flow_dv_translate_item_ipv4(void *matcher, void *key,
 static void
 flow_dv_translate_item_ipv6(void *matcher, void *key,
                            const struct rte_flow_item *item,
-                           const uint64_t item_flags,
                            int inner, uint32_t group)
 {
        const struct rte_flow_item_ipv6 *ipv6_m = item->mask;
@@ -6564,13 +6635,6 @@ flow_dv_translate_item_ipv6(void *matcher, void *key,
                headers_v = MLX5_ADDR_OF(fte_match_param, key, outer_headers);
        }
        flow_dv_set_match_ip_version(group, headers_v, headers_m, 6);
-       /*
-        * On outer header (which must contains L2), or inner header with L2,
-        * set cvlan_tag mask bit to mark this packet as untagged.
-        * This should be done even if item->spec is empty.
-        */
-       if (!inner || item_flags & MLX5_FLOW_LAYER_INNER_L2)
-               MLX5_SET(fte_match_set_lyr_2_4, headers_m, cvlan_tag, 1);
        if (!ipv6_v)
                return;
        if (!ipv6_m)
@@ -7858,6 +7922,66 @@ flow_dv_matcher_enable(uint32_t *match_criteria)
        return match_criteria_enable;
 }
 
+struct mlx5_hlist_entry *
+flow_dv_tbl_create_cb(struct mlx5_hlist *list, uint64_t key64, void *cb_ctx)
+{
+       struct mlx5_dev_ctx_shared *sh = list->ctx;
+       struct mlx5_flow_cb_ctx *ctx = cb_ctx;
+       struct rte_eth_dev *dev = ctx->dev;
+       struct mlx5_flow_tbl_data_entry *tbl_data;
+       struct mlx5_flow_tbl_tunnel_prm *tt_prm = ctx->data;
+       struct rte_flow_error *error = ctx->error;
+       union mlx5_flow_tbl_key key = { .v64 = key64 };
+       struct mlx5_flow_tbl_resource *tbl;
+       void *domain;
+       uint32_t idx = 0;
+       int ret;
+
+       tbl_data = mlx5_ipool_zmalloc(sh->ipool[MLX5_IPOOL_JUMP], &idx);
+       if (!tbl_data) {
+               rte_flow_error_set(error, ENOMEM,
+                                  RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
+                                  NULL,
+                                  "cannot allocate flow table data entry");
+               return NULL;
+       }
+       tbl_data->idx = idx;
+       tbl_data->tunnel = tt_prm->tunnel;
+       tbl_data->group_id = tt_prm->group_id;
+       tbl_data->external = tt_prm->external;
+       tbl_data->tunnel_offload = is_tunnel_offload_active(dev);
+       tbl = &tbl_data->tbl;
+       if (key.dummy)
+               return &tbl_data->entry;
+       if (key.domain)
+               domain = sh->fdb_domain;
+       else if (key.direction)
+               domain = sh->tx_domain;
+       else
+               domain = sh->rx_domain;
+       ret = mlx5_flow_os_create_flow_tbl(domain, key.table_id, &tbl->obj);
+       if (ret) {
+               rte_flow_error_set(error, ENOMEM,
+                                  RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
+                                  NULL, "cannot create flow table object");
+               mlx5_ipool_free(sh->ipool[MLX5_IPOOL_JUMP], idx);
+               return NULL;
+       }
+       if (key.table_id) {
+               ret = mlx5_flow_os_create_flow_action_dest_flow_tbl
+                                       (tbl->obj, &tbl_data->jump.action);
+               if (ret) {
+                       rte_flow_error_set(error, ENOMEM,
+                                          RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
+                                          NULL,
+                                          "cannot create flow jump action");
+                       mlx5_flow_os_destroy_flow_tbl(tbl->obj);
+                       mlx5_ipool_free(sh->ipool[MLX5_IPOOL_JUMP], idx);
+                       return NULL;
+               }
+       }
+       return &tbl_data->entry;
+}
 
 /**
  * Get a flow table.
@@ -7870,86 +7994,104 @@ flow_dv_matcher_enable(uint32_t *match_criteria)
  *   Direction of the table.
  * @param[in] transfer
  *   E-Switch or NIC flow.
+ * @param[in] dummy
+ *   Dummy entry for dv API.
  * @param[out] error
  *   pointer to error structure.
  *
  * @return
  *   Returns tables resource based on the index, NULL in case of failed.
  */
-static struct mlx5_flow_tbl_resource *
+struct mlx5_flow_tbl_resource *
 flow_dv_tbl_resource_get(struct rte_eth_dev *dev,
                         uint32_t table_id, uint8_t egress,
                         uint8_t transfer,
+                        bool external,
+                        const struct mlx5_flow_tunnel *tunnel,
+                        uint32_t group_id, uint8_t dummy,
                         struct rte_flow_error *error)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
-       struct mlx5_dev_ctx_shared *sh = priv->sh;
-       struct mlx5_flow_tbl_resource *tbl;
        union mlx5_flow_tbl_key table_key = {
                {
                        .table_id = table_id,
-                       .reserved = 0,
+                       .dummy = dummy,
                        .domain = !!transfer,
                        .direction = !!egress,
                }
        };
-       struct mlx5_hlist_entry *pos = mlx5_hlist_lookup(sh->flow_tbls,
-                                                        table_key.v64);
+       struct mlx5_flow_tbl_tunnel_prm tt_prm = {
+               .tunnel = tunnel,
+               .group_id = group_id,
+               .external = external,
+       };
+       struct mlx5_flow_cb_ctx ctx = {
+               .dev = dev,
+               .error = error,
+               .data = &tt_prm,
+       };
+       struct mlx5_hlist_entry *entry;
        struct mlx5_flow_tbl_data_entry *tbl_data;
-       uint32_t idx = 0;
-       int ret;
-       void *domain;
 
-       if (pos) {
-               tbl_data = container_of(pos, struct mlx5_flow_tbl_data_entry,
-                                       entry);
-               tbl = &tbl_data->tbl;
-               rte_atomic32_inc(&tbl->refcnt);
-               return tbl;
-       }
-       tbl_data = mlx5_ipool_zmalloc(sh->ipool[MLX5_IPOOL_JUMP], &idx);
-       if (!tbl_data) {
-               rte_flow_error_set(error, ENOMEM,
-                                  RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
-                                  NULL,
-                                  "cannot allocate flow table data entry");
-               return NULL;
-       }
-       tbl_data->idx = idx;
-       tbl = &tbl_data->tbl;
-       pos = &tbl_data->entry;
-       if (transfer)
-               domain = sh->fdb_domain;
-       else if (egress)
-               domain = sh->tx_domain;
-       else
-               domain = sh->rx_domain;
-       ret = mlx5_flow_os_create_flow_tbl(domain, table_id, &tbl->obj);
-       if (ret) {
+       entry = mlx5_hlist_register(priv->sh->flow_tbls, table_key.v64, &ctx);
+       if (!entry) {
                rte_flow_error_set(error, ENOMEM,
-                                  RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
-                                  NULL, "cannot create flow table object");
-               mlx5_ipool_free(sh->ipool[MLX5_IPOOL_JUMP], idx);
+                                  RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
+                                  "cannot get table");
                return NULL;
        }
-       /*
-        * No multi-threads now, but still better to initialize the reference
-        * count before insert it into the hash list.
-        */
-       rte_atomic32_init(&tbl->refcnt);
-       /* Jump action reference count is initialized here. */
-       rte_atomic32_init(&tbl_data->jump.refcnt);
-       pos->key = table_key.v64;
-       ret = mlx5_hlist_insert(sh->flow_tbls, pos);
-       if (ret < 0) {
-               rte_flow_error_set(error, -ret,
-                                  RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
-                                  "cannot insert flow table data entry");
-               mlx5_flow_os_destroy_flow_tbl(tbl->obj);
-               mlx5_ipool_free(sh->ipool[MLX5_IPOOL_JUMP], idx);
+       tbl_data = container_of(entry, struct mlx5_flow_tbl_data_entry, entry);
+       return &tbl_data->tbl;
+}
+
+void
+flow_dv_tbl_remove_cb(struct mlx5_hlist *list,
+                     struct mlx5_hlist_entry *entry)
+{
+       struct mlx5_dev_ctx_shared *sh = list->ctx;
+       struct mlx5_flow_tbl_data_entry *tbl_data =
+               container_of(entry, struct mlx5_flow_tbl_data_entry, entry);
+
+       MLX5_ASSERT(entry && sh);
+       if (tbl_data->jump.action)
+               mlx5_flow_os_destroy_flow_action(tbl_data->jump.action);
+       if (tbl_data->tbl.obj)
+               mlx5_flow_os_destroy_flow_tbl(tbl_data->tbl.obj);
+       if (tbl_data->tunnel_offload && tbl_data->external) {
+               struct mlx5_hlist_entry *he;
+               struct mlx5_hlist *tunnel_grp_hash;
+               struct mlx5_flow_tunnel_hub *thub = sh->tunnel_hub;
+               union tunnel_tbl_key tunnel_key = {
+                       .tunnel_id = tbl_data->tunnel ?
+                                       tbl_data->tunnel->tunnel_id : 0,
+                       .group = tbl_data->group_id
+               };
+               union mlx5_flow_tbl_key table_key = {
+                       .v64 = entry->key
+               };
+               uint32_t table_id = table_key.table_id;
+
+               tunnel_grp_hash = tbl_data->tunnel ?
+                                       tbl_data->tunnel->groups :
+                                       thub->groups;
+               he = mlx5_hlist_lookup(tunnel_grp_hash, tunnel_key.val, NULL);
+               if (he) {
+                       struct tunnel_tbl_entry *tte;
+                       tte = container_of(he, typeof(*tte), hash);
+                       MLX5_ASSERT(tte->flow_table == table_id);
+                       mlx5_hlist_remove(tunnel_grp_hash, he);
+                       mlx5_free(tte);
+               }
+               mlx5_ipool_free(sh->ipool[MLX5_IPOOL_TNL_TBL_ID],
+                               tunnel_flow_tbl_to_id(table_id));
+               DRV_LOG(DEBUG,
+                       "Table_id %#x tunnel %u group %u released.",
+                       table_id,
+                       tbl_data->tunnel ?
+                       tbl_data->tunnel->tunnel_id : 0,
+                       tbl_data->group_id);
        }
-       rte_atomic32_inc(&tbl->refcnt);
-       return tbl;
+       mlx5_ipool_free(sh->ipool[MLX5_IPOOL_JUMP], tbl_data->idx);
 }
 
 /**
@@ -7974,18 +8116,7 @@ flow_dv_tbl_resource_release(struct rte_eth_dev *dev,
 
        if (!tbl)
                return 0;
-       if (rte_atomic32_dec_and_test(&tbl->refcnt)) {
-               struct mlx5_hlist_entry *pos = &tbl_data->entry;
-
-               mlx5_flow_os_destroy_flow_tbl(tbl->obj);
-               tbl->obj = NULL;
-               /* remove the entry from the hash list and free memory. */
-               mlx5_hlist_remove(sh->flow_tbls, pos);
-               mlx5_ipool_free(priv->sh->ipool[MLX5_IPOOL_JUMP],
-                               tbl_data->idx);
-               return 0;
-       }
-       return 1;
+       return mlx5_hlist_unregister(sh->flow_tbls, &tbl_data->entry);
 }
 
 /**
@@ -8024,7 +8155,7 @@ flow_dv_matcher_register(struct rte_eth_dev *dev,
        int ret;
 
        tbl = flow_dv_tbl_resource_get(dev, key->table_id, key->direction,
-                                      key->domain, error);
+                                      key->domain, false, NULL, 0, 0, error);
        if (!tbl)
                return -rte_errno;      /* No need to refill the error info */
        tbl_data = container_of(tbl, struct mlx5_flow_tbl_data_entry, tbl);
@@ -8042,8 +8173,10 @@ flow_dv_matcher_register(struct rte_eth_dev *dev,
                                cache_matcher->priority,
                                key->direction ? "tx" : "rx",
                                (void *)cache_matcher,
-                               rte_atomic32_read(&cache_matcher->refcnt));
-                       rte_atomic32_inc(&cache_matcher->refcnt);
+                               __atomic_load_n(&cache_matcher->refcnt,
+                                               __ATOMIC_RELAXED));
+                       __atomic_fetch_add(&cache_matcher->refcnt, 1,
+                                          __ATOMIC_RELAXED);
                        dev_flow->handle->dvh.matcher = cache_matcher;
                        /* old matcher should not make the table ref++. */
                        flow_dv_tbl_resource_release(dev, tbl);
@@ -8078,16 +8211,15 @@ flow_dv_matcher_register(struct rte_eth_dev *dev,
        }
        /* Save the table information */
        cache_matcher->tbl = tbl;
-       rte_atomic32_init(&cache_matcher->refcnt);
        /* only matcher ref++, table ref++ already done above in get API. */
-       rte_atomic32_inc(&cache_matcher->refcnt);
+       __atomic_store_n(&cache_matcher->refcnt, 1, __ATOMIC_RELAXED);
        LIST_INSERT_HEAD(&tbl_data->matchers, cache_matcher, next);
        dev_flow->handle->dvh.matcher = cache_matcher;
        DRV_LOG(DEBUG, "%s group %u priority %hd new %s matcher %p: refcnt %d",
                key->domain ? "FDB" : "NIC", key->table_id,
                cache_matcher->priority,
                key->direction ? "tx" : "rx", (void *)cache_matcher,
-               rte_atomic32_read(&cache_matcher->refcnt));
+               __atomic_load_n(&cache_matcher->refcnt, __ATOMIC_RELAXED));
        return 0;
 }
 
@@ -8120,16 +8252,18 @@ flow_dv_tag_resource_register
        int ret;
 
        /* Lookup a matching resource from cache. */
-       entry = mlx5_hlist_lookup(sh->tag_table, (uint64_t)tag_be24);
+       entry = mlx5_hlist_lookup(sh->tag_table, (uint64_t)tag_be24, NULL);
        if (entry) {
                cache_resource = container_of
                        (entry, struct mlx5_flow_dv_tag_resource, entry);
-               rte_atomic32_inc(&cache_resource->refcnt);
+               __atomic_fetch_add(&cache_resource->refcnt, 1,
+                                  __ATOMIC_RELAXED);
                dev_flow->handle->dvh.rix_tag = cache_resource->idx;
                dev_flow->dv.tag_resource = cache_resource;
                DRV_LOG(DEBUG, "cached tag resource %p: refcnt now %d++",
                        (void *)cache_resource,
-                       rte_atomic32_read(&cache_resource->refcnt));
+                       __atomic_load_n(&cache_resource->refcnt,
+                                       __ATOMIC_RELAXED));
                return 0;
        }
        /* Register new resource. */
@@ -8148,8 +8282,7 @@ flow_dv_tag_resource_register
                                          RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
                                          NULL, "cannot create action");
        }
-       rte_atomic32_init(&cache_resource->refcnt);
-       rte_atomic32_inc(&cache_resource->refcnt);
+       __atomic_store_n(&cache_resource->refcnt, 1, __ATOMIC_RELAXED);
        if (mlx5_hlist_insert(sh->tag_table, &cache_resource->entry)) {
                mlx5_flow_os_destroy_flow_action(cache_resource->action);
                mlx5_free(cache_resource);
@@ -8160,7 +8293,7 @@ flow_dv_tag_resource_register
        dev_flow->dv.tag_resource = cache_resource;
        DRV_LOG(DEBUG, "new tag resource %p: refcnt now %d++",
                (void *)cache_resource,
-               rte_atomic32_read(&cache_resource->refcnt));
+               __atomic_load_n(&cache_resource->refcnt, __ATOMIC_RELAXED));
        return 0;
 }
 
@@ -8188,8 +8321,8 @@ flow_dv_tag_release(struct rte_eth_dev *dev,
                return 0;
        DRV_LOG(DEBUG, "port %u tag %p: refcnt %d--",
                dev->data->port_id, (void *)tag,
-               rte_atomic32_read(&tag->refcnt));
-       if (rte_atomic32_dec_and_test(&tag->refcnt)) {
+               __atomic_load_n(&tag->refcnt, __ATOMIC_RELAXED));
+       if (__atomic_sub_fetch(&tag->refcnt, 1, __ATOMIC_RELAXED) == 0) {
                claim_zero(mlx5_flow_os_destroy_flow_action(tag->action));
                mlx5_hlist_remove(sh->tag_table, &tag->entry);
                DRV_LOG(DEBUG, "port %u tag %p: removed",
@@ -8429,20 +8562,16 @@ flow_dv_handle_rx_queue(struct rte_eth_dev *dev,
        struct mlx5_hrxq *hrxq;
 
        MLX5_ASSERT(rss_desc->queue_num);
-       *hrxq_idx = mlx5_hrxq_get(dev, rss_desc->key,
-                                 MLX5_RSS_HASH_KEY_LEN,
+       *hrxq_idx = mlx5_hrxq_get(dev, rss_desc->key, MLX5_RSS_HASH_KEY_LEN,
                                  dev_flow->hash_fields,
-                                 rss_desc->queue,
-                                 rss_desc->queue_num);
+                                 rss_desc->queue, rss_desc->queue_num);
        if (!*hrxq_idx) {
                *hrxq_idx = mlx5_hrxq_new
-                               (dev, rss_desc->key,
-                                MLX5_RSS_HASH_KEY_LEN,
+                               (dev, rss_desc->key, MLX5_RSS_HASH_KEY_LEN,
                                 dev_flow->hash_fields,
-                                rss_desc->queue,
-                                rss_desc->queue_num,
-                                !!(dh->layers &
-                                MLX5_FLOW_LAYER_TUNNEL));
+                                rss_desc->queue, rss_desc->queue_num,
+                                !!(dh->layers & MLX5_FLOW_LAYER_TUNNEL),
+                                false);
                if (!*hrxq_idx)
                        return NULL;
        }
@@ -8519,7 +8648,8 @@ flow_dv_sample_resource_register(struct rte_eth_dev *dev,
        *cache_resource = *resource;
        /* Create normal path table level */
        tbl = flow_dv_tbl_resource_get(dev, next_ft_id,
-                                       attr->egress, attr->transfer, error);
+                                       attr->egress, attr->transfer,
+                                       dev_flow->external, NULL, 0, 0, error);
        if (!tbl) {
                rte_flow_error_set(error, ENOMEM,
                                          RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
@@ -8783,11 +8913,12 @@ flow_dv_translate_action_sample(struct rte_eth_dev *dev,
        const struct rte_flow_action_queue *queue;
        struct mlx5_flow_sub_actions_list *sample_act;
        struct mlx5_flow_sub_actions_idx *sample_idx;
-       struct mlx5_flow_rss_desc *rss_desc = &((struct mlx5_flow_rss_desc *)
-                                             priv->rss_desc)
-                                             [!!priv->flow_nested_idx];
+       struct mlx5_flow_workspace *wks = mlx5_flow_get_thread_workspace();
+       struct mlx5_flow_rss_desc *rss_desc;
        uint64_t action_flags = 0;
 
+       MLX5_ASSERT(wks);
+       rss_desc = &wks->rss_desc[!!wks->flow_nested_idx];
        sample_act = &res->sample_act;
        sample_idx = &res->sample_idx;
        sample_action = (const struct rte_flow_action_sample *)action->conf;
@@ -8989,18 +9120,18 @@ flow_dv_create_action_sample(struct rte_eth_dev *dev,
                             uint64_t action_flags,
                             struct rte_flow_error *error)
 {
-       struct mlx5_priv *priv = dev->data->dev_private;
        /* update normal path action resource into last index of array */
        uint32_t dest_index = MLX5_MAX_DEST_NUM - 1;
        struct mlx5_flow_sub_actions_list *sample_act =
                                        &mdest_res->sample_act[dest_index];
-       struct mlx5_flow_rss_desc *rss_desc = &((struct mlx5_flow_rss_desc *)
-                                             priv->rss_desc)
-                                             [!!priv->flow_nested_idx];
+       struct mlx5_flow_workspace *wks = mlx5_flow_get_thread_workspace();
+       struct mlx5_flow_rss_desc *rss_desc;
        uint32_t normal_idx = 0;
        struct mlx5_hrxq *hrxq;
        uint32_t hrxq_idx;
 
+       MLX5_ASSERT(wks);
+       rss_desc = &wks->rss_desc[!!wks->flow_nested_idx];
        if (num_of_dest > 1) {
                if (sample_act->action_flags & MLX5_FLOW_ACTION_QUEUE) {
                        /* Handle QP action for mirroring */
@@ -9090,9 +9221,8 @@ __flow_dv_translate(struct rte_eth_dev *dev,
        struct mlx5_dev_config *dev_conf = &priv->config;
        struct rte_flow *flow = dev_flow->flow;
        struct mlx5_flow_handle *handle = dev_flow->handle;
-       struct mlx5_flow_rss_desc *rss_desc = &((struct mlx5_flow_rss_desc *)
-                                             priv->rss_desc)
-                                             [!!priv->flow_nested_idx];
+       struct mlx5_flow_workspace *wks = mlx5_flow_get_thread_workspace();
+       struct mlx5_flow_rss_desc *rss_desc;
        uint64_t item_flags = 0;
        uint64_t last_item = 0;
        uint64_t action_flags = 0;
@@ -9131,15 +9261,32 @@ __flow_dv_translate(struct rte_eth_dev *dev,
        int tmp_actions_n = 0;
        uint32_t table;
        int ret = 0;
+       const struct mlx5_flow_tunnel *tunnel;
+       struct flow_grp_info grp_info = {
+               .external = !!dev_flow->external,
+               .transfer = !!attr->transfer,
+               .fdb_def_rule = !!priv->fdb_def_rule,
+       };
 
+       MLX5_ASSERT(wks);
+       rss_desc = &wks->rss_desc[!!wks->flow_nested_idx];
        memset(&mdest_res, 0, sizeof(struct mlx5_flow_dv_dest_array_resource));
        memset(&sample_res, 0, sizeof(struct mlx5_flow_dv_sample_resource));
        mhdr_res->ft_type = attr->egress ? MLX5DV_FLOW_TABLE_TYPE_NIC_TX :
                                           MLX5DV_FLOW_TABLE_TYPE_NIC_RX;
        /* update normal path action resource into last index of array */
        sample_act = &mdest_res.sample_act[MLX5_MAX_DEST_NUM - 1];
-       ret = mlx5_flow_group_to_table(attr, dev_flow->external, attr->group,
-                                      !!priv->fdb_def_rule, &table, error);
+       tunnel = is_flow_tunnel_match_rule(dev, attr, items, actions) ?
+                flow_items_to_tunnel(items) :
+                is_flow_tunnel_steer_rule(dev, attr, items, actions) ?
+                flow_actions_to_tunnel(actions) :
+                dev_flow->tunnel ? dev_flow->tunnel : NULL;
+       mhdr_res->ft_type = attr->egress ? MLX5DV_FLOW_TABLE_TYPE_NIC_TX :
+                                          MLX5DV_FLOW_TABLE_TYPE_NIC_RX;
+       grp_info.std_tbl_fix = tunnel_use_standard_attr_group_translate
+                               (dev, tunnel, attr, items, actions);
+       ret = mlx5_flow_group_to_table(dev, tunnel, attr->group, &table,
+                                      grp_info, error);
        if (ret)
                return ret;
        dev_flow->dv.group = table;
@@ -9149,6 +9296,45 @@ __flow_dv_translate(struct rte_eth_dev *dev,
                priority = dev_conf->flow_prio - 1;
        /* number of actions must be set to 0 in case of dirty stack. */
        mhdr_res->actions_num = 0;
+       if (is_flow_tunnel_match_rule(dev, attr, items, actions)) {
+               /*
+                * do not add decap action if match rule drops packet
+                * HW rejects rules with decap & drop
+                */
+               bool add_decap = true;
+               const struct rte_flow_action *ptr = actions;
+               struct mlx5_flow_tbl_resource *tbl;
+
+               for (; ptr->type != RTE_FLOW_ACTION_TYPE_END; ptr++) {
+                       if (ptr->type == RTE_FLOW_ACTION_TYPE_DROP) {
+                               add_decap = false;
+                               break;
+                       }
+               }
+               if (add_decap) {
+                       if (flow_dv_create_action_l2_decap(dev, dev_flow,
+                                                          attr->transfer,
+                                                          error))
+                               return -rte_errno;
+                       dev_flow->dv.actions[actions_n++] =
+                                       dev_flow->dv.encap_decap->action;
+                       action_flags |= MLX5_FLOW_ACTION_DECAP;
+               }
+               /*
+                * bind table_id with <group, table> for tunnel match rule.
+                * Tunnel set rule establishes that bind in JUMP action handler.
+                * Required for scenario when application creates tunnel match
+                * rule before tunnel set rule.
+                */
+               tbl = flow_dv_tbl_resource_get(dev, table, attr->egress,
+                                              attr->transfer,
+                                              !!dev_flow->external, tunnel,
+                                              attr->group, 0, error);
+               if (!tbl)
+                       return rte_flow_error_set
+                              (error, EINVAL, RTE_FLOW_ERROR_TYPE_ACTION,
+                              actions, "cannot register tunnel group");
+       }
        for (; !actions_end ; actions++) {
                const struct rte_flow_action_queue *queue;
                const struct rte_flow_action_rss *rss;
@@ -9169,6 +9355,9 @@ __flow_dv_translate(struct rte_eth_dev *dev,
                                                  actions,
                                                  "action not supported");
                switch (action_type) {
+               case MLX5_RTE_FLOW_ACTION_TYPE_TUNNEL_SET:
+                       action_flags |= MLX5_FLOW_ACTION_TUNNEL_SET;
+                       break;
                case RTE_FLOW_ACTION_TYPE_VOID:
                        break;
                case RTE_FLOW_ACTION_TYPE_PORT_ID:
@@ -9412,18 +9601,18 @@ __flow_dv_translate(struct rte_eth_dev *dev,
                case RTE_FLOW_ACTION_TYPE_JUMP:
                        jump_group = ((const struct rte_flow_action_jump *)
                                                        action->conf)->group;
-                       if (dev_flow->external && jump_group <
-                                       MLX5_MAX_TABLES_EXTERNAL)
-                               jump_group *= MLX5_FLOW_TABLE_FACTOR;
-                       ret = mlx5_flow_group_to_table(attr, dev_flow->external,
+                       grp_info.std_tbl_fix = 0;
+                       ret = mlx5_flow_group_to_table(dev, tunnel,
                                                       jump_group,
-                                                      !!priv->fdb_def_rule,
-                                                      &table, error);
+                                                      &table,
+                                                      grp_info, error);
                        if (ret)
                                return ret;
-                       tbl = flow_dv_tbl_resource_get(dev, table,
-                                                      attr->egress,
-                                                      attr->transfer, error);
+                       tbl = flow_dv_tbl_resource_get(dev, table, attr->egress,
+                                                      attr->transfer,
+                                                      !!dev_flow->external,
+                                                      tunnel, jump_group, 0,
+                                                      error);
                        if (!tbl)
                                return rte_flow_error_set
                                                (error, errno,
@@ -9727,7 +9916,7 @@ __flow_dv_translate(struct rte_eth_dev *dev,
                        mlx5_flow_tunnel_ip_check(items, next_protocol,
                                                  &item_flags, &tunnel);
                        flow_dv_translate_item_ipv4(match_mask, match_value,
-                                                   items, item_flags, tunnel,
+                                                   items, tunnel,
                                                    dev_flow->dv.group);
                        matcher.priority = MLX5_PRIORITY_MAP_L3;
                        last_item = tunnel ? MLX5_FLOW_LAYER_INNER_L3_IPV4 :
@@ -9750,7 +9939,7 @@ __flow_dv_translate(struct rte_eth_dev *dev,
                        mlx5_flow_tunnel_ip_check(items, next_protocol,
                                                  &item_flags, &tunnel);
                        flow_dv_translate_item_ipv6(match_mask, match_value,
-                                                   items, item_flags, tunnel,
+                                                   items, tunnel,
                                                    dev_flow->dv.group);
                        matcher.priority = MLX5_PRIORITY_MAP_L3;
                        last_item = tunnel ? MLX5_FLOW_LAYER_INNER_L3_IPV6 :
@@ -9952,88 +10141,208 @@ __flow_dv_translate(struct rte_eth_dev *dev,
 }
 
 /**
- * Apply the flow to the NIC, lock free,
- * (mutex should be acquired by caller).
+ * Set hash RX queue by hash fields (see enum ibv_rx_hash_fields)
+ * and tunnel.
  *
- * @param[in] dev
- *   Pointer to the Ethernet device structure.
- * @param[in, out] flow
- *   Pointer to flow structure.
- * @param[out] error
- *   Pointer to error structure.
+ * @param[in, out] action
+ *   Shred RSS action holding hash RX queue objects.
+ * @param[in] hash_fields
+ *   Defines combination of packet fields to participate in RX hash.
+ * @param[in] tunnel
+ *   Tunnel type
+ * @param[in] hrxq_idx
+ *   Hash RX queue index to set.
  *
  * @return
- *   0 on success, a negative errno value otherwise and rte_errno is set.
+ *   0 on success, otherwise negative errno value.
  */
 static int
-__flow_dv_apply(struct rte_eth_dev *dev, struct rte_flow *flow,
-               struct rte_flow_error *error)
+__flow_dv_action_rss_hrxq_set(struct mlx5_shared_action_rss *action,
+                             const uint64_t hash_fields,
+                             const int tunnel,
+                             uint32_t hrxq_idx)
 {
-       struct mlx5_flow_dv_workspace *dv;
-       struct mlx5_flow_handle *dh;
-       struct mlx5_flow_handle_dv *dv_h;
-       struct mlx5_flow *dev_flow;
-       struct mlx5_priv *priv = dev->data->dev_private;
-       uint32_t handle_idx;
-       int n;
-       int err;
-       int idx;
+       uint32_t *hrxqs = tunnel ? action->hrxq : action->hrxq_tunnel;
 
-       for (idx = priv->flow_idx - 1; idx >= priv->flow_nested_idx; idx--) {
-               dev_flow = &((struct mlx5_flow *)priv->inter_flows)[idx];
-               dv = &dev_flow->dv;
-               dh = dev_flow->handle;
-               dv_h = &dh->dvh;
+       switch (hash_fields & ~IBV_RX_HASH_INNER) {
+       case MLX5_RSS_HASH_IPV4:
+               hrxqs[0] = hrxq_idx;
+               return 0;
+       case MLX5_RSS_HASH_IPV4_TCP:
+               hrxqs[1] = hrxq_idx;
+               return 0;
+       case MLX5_RSS_HASH_IPV4_UDP:
+               hrxqs[2] = hrxq_idx;
+               return 0;
+       case MLX5_RSS_HASH_IPV6:
+               hrxqs[3] = hrxq_idx;
+               return 0;
+       case MLX5_RSS_HASH_IPV6_TCP:
+               hrxqs[4] = hrxq_idx;
+               return 0;
+       case MLX5_RSS_HASH_IPV6_UDP:
+               hrxqs[5] = hrxq_idx;
+               return 0;
+       case MLX5_RSS_HASH_NONE:
+               hrxqs[6] = hrxq_idx;
+               return 0;
+       default:
+               return -1;
+       }
+}
+
+/**
+ * Look up for hash RX queue by hash fields (see enum ibv_rx_hash_fields)
+ * and tunnel.
+ *
+ * @param[in] action
+ *   Shred RSS action holding hash RX queue objects.
+ * @param[in] hash_fields
+ *   Defines combination of packet fields to participate in RX hash.
+ * @param[in] tunnel
+ *   Tunnel type
+ *
+ * @return
+ *   Valid hash RX queue index, otherwise 0.
+ */
+static uint32_t
+__flow_dv_action_rss_hrxq_lookup(const struct mlx5_shared_action_rss *action,
+                                const uint64_t hash_fields,
+                                const int tunnel)
+{
+       const uint32_t *hrxqs = tunnel ? action->hrxq : action->hrxq_tunnel;
+
+       switch (hash_fields & ~IBV_RX_HASH_INNER) {
+       case MLX5_RSS_HASH_IPV4:
+               return hrxqs[0];
+       case MLX5_RSS_HASH_IPV4_TCP:
+               return hrxqs[1];
+       case MLX5_RSS_HASH_IPV4_UDP:
+               return hrxqs[2];
+       case MLX5_RSS_HASH_IPV6:
+               return hrxqs[3];
+       case MLX5_RSS_HASH_IPV6_TCP:
+               return hrxqs[4];
+       case MLX5_RSS_HASH_IPV6_UDP:
+               return hrxqs[5];
+       case MLX5_RSS_HASH_NONE:
+               return hrxqs[6];
+       default:
+               return 0;
+       }
+}
+
+/**
+ * Retrieves hash RX queue suitable for the *flow*.
+ * If shared action configured for *flow* suitable hash RX queue will be
+ * retrieved from attached shared action.
+ *
+ * @param[in] flow
+ *   Shred RSS action holding hash RX queue objects.
+ * @param[in] dev_flow
+ *   Pointer to the sub flow.
+ * @param[out] hrxq
+ *   Pointer to retrieved hash RX queue object.
+ *
+ * @return
+ *   Valid hash RX queue index, otherwise 0 and rte_errno is set.
+ */
+static uint32_t
+__flow_dv_rss_get_hrxq(struct rte_eth_dev *dev, struct rte_flow *flow,
+                          struct mlx5_flow *dev_flow,
+                          struct mlx5_hrxq **hrxq)
+{
+       struct mlx5_priv *priv = dev->data->dev_private;
+       struct mlx5_flow_workspace *wks = mlx5_flow_get_thread_workspace();
+       uint32_t hrxq_idx;
+
+       if (flow->shared_rss) {
+               hrxq_idx = __flow_dv_action_rss_hrxq_lookup
+                               (flow->shared_rss, dev_flow->hash_fields,
+                                !!(dev_flow->handle->layers &
+                                   MLX5_FLOW_LAYER_TUNNEL));
+               if (hrxq_idx) {
+                       *hrxq = mlx5_ipool_get(priv->sh->ipool[MLX5_IPOOL_HRXQ],
+                                              hrxq_idx);
+                       __atomic_fetch_add(&(*hrxq)->refcnt, 1,
+                                          __ATOMIC_RELAXED);
+               }
+       } else {
+               struct mlx5_flow_rss_desc *rss_desc =
+                               &wks->rss_desc[!!wks->flow_nested_idx];
+
+               MLX5_ASSERT(rss_desc->queue_num);
+               hrxq_idx = mlx5_hrxq_get(dev, rss_desc->key,
+                                        MLX5_RSS_HASH_KEY_LEN,
+                                        dev_flow->hash_fields,
+                                        rss_desc->queue, rss_desc->queue_num);
+               if (!hrxq_idx) {
+                       hrxq_idx = mlx5_hrxq_new(dev,
+                                                rss_desc->key,
+                                                MLX5_RSS_HASH_KEY_LEN,
+                                                dev_flow->hash_fields,
+                                                rss_desc->queue,
+                                                rss_desc->queue_num,
+                                                !!(dev_flow->handle->layers &
+                                                MLX5_FLOW_LAYER_TUNNEL),
+                                                false);
+               }
+               *hrxq = mlx5_ipool_get(priv->sh->ipool[MLX5_IPOOL_HRXQ],
+                                      hrxq_idx);
+       }
+       return hrxq_idx;
+}
+
+/**
+ * Apply the flow to the NIC, lock free,
+ * (mutex should be acquired by caller).
+ *
+ * @param[in] dev
+ *   Pointer to the Ethernet device structure.
+ * @param[in, out] flow
+ *   Pointer to flow structure.
+ * @param[out] error
+ *   Pointer to error structure.
+ *
+ * @return
+ *   0 on success, a negative errno value otherwise and rte_errno is set.
+ */
+static int
+__flow_dv_apply(struct rte_eth_dev *dev, struct rte_flow *flow,
+               struct rte_flow_error *error)
+{
+       struct mlx5_flow_dv_workspace *dv;
+       struct mlx5_flow_handle *dh;
+       struct mlx5_flow_handle_dv *dv_h;
+       struct mlx5_flow *dev_flow;
+       struct mlx5_priv *priv = dev->data->dev_private;
+       uint32_t handle_idx;
+       int n;
+       int err;
+       int idx;
+       struct mlx5_flow_workspace *wks = mlx5_flow_get_thread_workspace();
+
+       MLX5_ASSERT(wks);
+       for (idx = wks->flow_idx - 1; idx >= wks->flow_nested_idx; idx--) {
+               dev_flow = &wks->flows[idx];
+               dv = &dev_flow->dv;
+               dh = dev_flow->handle;
+               dv_h = &dh->dvh;
                n = dv->actions_n;
                if (dh->fate_action == MLX5_FLOW_FATE_DROP) {
                        if (dv->transfer) {
                                dv->actions[n++] = priv->sh->esw_drop_action;
                        } else {
-                               struct mlx5_hrxq *drop_hrxq;
-                               drop_hrxq = mlx5_drop_action_create(dev);
-                               if (!drop_hrxq) {
-                                       rte_flow_error_set
-                                               (error, errno,
-                                                RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
-                                                NULL,
-                                                "cannot get drop hash queue");
-                                       goto error;
-                               }
-                               /*
-                                * Drop queues will be released by the specify
-                                * mlx5_drop_action_destroy() function. Assign
-                                * the special index to hrxq to mark the queue
-                                * has been allocated.
-                                */
-                               dh->rix_hrxq = UINT32_MAX;
-                               dv->actions[n++] = drop_hrxq->action;
+                               MLX5_ASSERT(priv->drop_queue.hrxq);
+                               dv->actions[n++] =
+                                               priv->drop_queue.hrxq->action;
                        }
                } else if (dh->fate_action == MLX5_FLOW_FATE_QUEUE &&
                           !dv_h->rix_sample && !dv_h->rix_dest_array) {
-                       struct mlx5_hrxq *hrxq;
-                       uint32_t hrxq_idx;
-                       struct mlx5_flow_rss_desc *rss_desc =
-                               &((struct mlx5_flow_rss_desc *)priv->rss_desc)
-                               [!!priv->flow_nested_idx];
+                       struct mlx5_hrxq *hrxq = NULL;
+                       uint32_t hrxq_idx = __flow_dv_rss_get_hrxq
+                                               (dev, flow, dev_flow, &hrxq);
 
-                       MLX5_ASSERT(rss_desc->queue_num);
-                       hrxq_idx = mlx5_hrxq_get(dev, rss_desc->key,
-                                                MLX5_RSS_HASH_KEY_LEN,
-                                                dev_flow->hash_fields,
-                                                rss_desc->queue,
-                                                rss_desc->queue_num);
-                       if (!hrxq_idx) {
-                               hrxq_idx = mlx5_hrxq_new
-                                               (dev, rss_desc->key,
-                                                MLX5_RSS_HASH_KEY_LEN,
-                                                dev_flow->hash_fields,
-                                                rss_desc->queue,
-                                                rss_desc->queue_num,
-                                                !!(dh->layers &
-                                                MLX5_FLOW_LAYER_TUNNEL));
-                       }
-                       hrxq = mlx5_ipool_get(priv->sh->ipool[MLX5_IPOOL_HRXQ],
-                                             hrxq_idx);
                        if (!hrxq) {
                                rte_flow_error_set
                                        (error, rte_errno,
@@ -10044,16 +10353,14 @@ __flow_dv_apply(struct rte_eth_dev *dev, struct rte_flow *flow,
                        dh->rix_hrxq = hrxq_idx;
                        dv->actions[n++] = hrxq->action;
                } else if (dh->fate_action == MLX5_FLOW_FATE_DEFAULT_MISS) {
-                       if (flow_dv_default_miss_resource_register
-                                       (dev, error)) {
+                       if (!priv->sh->default_miss_action) {
                                rte_flow_error_set
                                        (error, rte_errno,
                                         RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
-                                        "cannot create default miss resource");
-                               goto error_default_miss;
+                                        "default miss action not be created.");
+                               goto error;
                        }
-                       dh->rix_default_fate =  MLX5_FLOW_FATE_DEFAULT_MISS;
-                       dv->actions[n++] = priv->sh->default_miss.action;
+                       dv->actions[n++] = priv->sh->default_miss_action;
                }
                err = mlx5_flow_os_create_flow(dv_h->matcher->matcher_object,
                                               (void *)&dv->value, n,
@@ -10078,21 +10385,13 @@ __flow_dv_apply(struct rte_eth_dev *dev, struct rte_flow *flow,
        }
        return 0;
 error:
-       if (dh->fate_action == MLX5_FLOW_FATE_DEFAULT_MISS)
-               flow_dv_default_miss_resource_release(dev);
-error_default_miss:
        err = rte_errno; /* Save rte_errno before cleanup. */
        SILIST_FOREACH(priv->sh->ipool[MLX5_IPOOL_MLX5_FLOW], flow->dev_handles,
                       handle_idx, dh, next) {
                /* hrxq is union, don't clear it if the flag is not set. */
-               if (dh->rix_hrxq) {
-                       if (dh->fate_action == MLX5_FLOW_FATE_DROP) {
-                               mlx5_drop_action_destroy(dev);
-                               dh->rix_hrxq = 0;
-                       } else if (dh->fate_action == MLX5_FLOW_FATE_QUEUE) {
-                               mlx5_hrxq_release(dev, dh->rix_hrxq);
-                               dh->rix_hrxq = 0;
-                       }
+               if (dh->fate_action == MLX5_FLOW_FATE_QUEUE && dh->rix_hrxq) {
+                       mlx5_hrxq_release(dev, dh->rix_hrxq);
+                       dh->rix_hrxq = 0;
                }
                if (dh->vf_vlan.tag && dh->vf_vlan.created)
                        mlx5_vlan_vmwa_release(dev, &dh->vf_vlan);
@@ -10121,8 +10420,8 @@ flow_dv_matcher_release(struct rte_eth_dev *dev,
        MLX5_ASSERT(matcher->matcher_object);
        DRV_LOG(DEBUG, "port %u matcher %p: refcnt %d--",
                dev->data->port_id, (void *)matcher,
-               rte_atomic32_read(&matcher->refcnt));
-       if (rte_atomic32_dec_and_test(&matcher->refcnt)) {
+               __atomic_load_n(&matcher->refcnt, __ATOMIC_RELAXED));
+       if (__atomic_sub_fetch(&matcher->refcnt, 1, __ATOMIC_RELAXED) == 0) {
                claim_zero(mlx5_flow_os_destroy_flow_matcher
                           (matcher->matcher_object));
                LIST_REMOVE(matcher, next);
@@ -10162,8 +10461,9 @@ flow_dv_encap_decap_resource_release(struct rte_eth_dev *dev,
        MLX5_ASSERT(cache_resource->action);
        DRV_LOG(DEBUG, "encap/decap resource %p: refcnt %d--",
                (void *)cache_resource,
-               rte_atomic32_read(&cache_resource->refcnt));
-       if (rte_atomic32_dec_and_test(&cache_resource->refcnt)) {
+               __atomic_load_n(&cache_resource->refcnt, __ATOMIC_RELAXED));
+       if (__atomic_sub_fetch(&cache_resource->refcnt, 1,
+                              __ATOMIC_RELAXED) == 0) {
                claim_zero(mlx5_flow_os_destroy_flow_action
                                                (cache_resource->action));
                mlx5_hlist_remove(priv->sh->encaps_decaps,
@@ -10192,58 +10492,13 @@ flow_dv_jump_tbl_resource_release(struct rte_eth_dev *dev,
                                  struct mlx5_flow_handle *handle)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
-       struct mlx5_flow_dv_jump_tbl_resource *cache_resource;
        struct mlx5_flow_tbl_data_entry *tbl_data;
 
        tbl_data = mlx5_ipool_get(priv->sh->ipool[MLX5_IPOOL_JUMP],
                             handle->rix_jump);
        if (!tbl_data)
                return 0;
-       cache_resource = &tbl_data->jump;
-       MLX5_ASSERT(cache_resource->action);
-       DRV_LOG(DEBUG, "jump table resource %p: refcnt %d--",
-               (void *)cache_resource,
-               rte_atomic32_read(&cache_resource->refcnt));
-       if (rte_atomic32_dec_and_test(&cache_resource->refcnt)) {
-               claim_zero(mlx5_flow_os_destroy_flow_action
-                                               (cache_resource->action));
-               /* jump action memory free is inside the table release. */
-               flow_dv_tbl_resource_release(dev, &tbl_data->tbl);
-               DRV_LOG(DEBUG, "jump table resource %p: removed",
-                       (void *)cache_resource);
-               return 0;
-       }
-       return 1;
-}
-
-/**
- * Release a default miss resource.
- *
- * @param dev
- *   Pointer to Ethernet device.
- * @return
- *   1 while a reference on it exists, 0 when freed.
- */
-static int
-flow_dv_default_miss_resource_release(struct rte_eth_dev *dev)
-{
-       struct mlx5_priv *priv = dev->data->dev_private;
-       struct mlx5_dev_ctx_shared *sh = priv->sh;
-       struct mlx5_flow_default_miss_resource *cache_resource =
-                       &sh->default_miss;
-
-       MLX5_ASSERT(cache_resource->action);
-       DRV_LOG(DEBUG, "default miss resource %p: refcnt %d--",
-                       (void *)cache_resource->action,
-                       rte_atomic32_read(&cache_resource->refcnt));
-       if (rte_atomic32_dec_and_test(&cache_resource->refcnt)) {
-               claim_zero(mlx5_glue->destroy_flow_action
-                               (cache_resource->action));
-               DRV_LOG(DEBUG, "default miss resource %p: removed",
-                               (void *)cache_resource->action);
-               return 0;
-       }
-       return 1;
+       return flow_dv_tbl_resource_release(dev, &tbl_data->tbl);
 }
 
 /**
@@ -10268,8 +10523,9 @@ flow_dv_modify_hdr_resource_release(struct rte_eth_dev *dev,
        MLX5_ASSERT(cache_resource->action);
        DRV_LOG(DEBUG, "modify-header resource %p: refcnt %d--",
                (void *)cache_resource,
-               rte_atomic32_read(&cache_resource->refcnt));
-       if (rte_atomic32_dec_and_test(&cache_resource->refcnt)) {
+               __atomic_load_n(&cache_resource->refcnt, __ATOMIC_RELAXED));
+       if (__atomic_sub_fetch(&cache_resource->refcnt, 1,
+                               __ATOMIC_RELAXED) == 0) {
                claim_zero(mlx5_flow_os_destroy_flow_action
                                                (cache_resource->action));
                mlx5_hlist_remove(priv->sh->modify_cmds,
@@ -10308,8 +10564,9 @@ flow_dv_port_id_action_resource_release(struct rte_eth_dev *dev,
        MLX5_ASSERT(cache_resource->action);
        DRV_LOG(DEBUG, "port ID action resource %p: refcnt %d--",
                (void *)cache_resource,
-               rte_atomic32_read(&cache_resource->refcnt));
-       if (rte_atomic32_dec_and_test(&cache_resource->refcnt)) {
+               __atomic_load_n(&cache_resource->refcnt, __ATOMIC_RELAXED));
+       if (__atomic_sub_fetch(&cache_resource->refcnt, 1,
+                              __ATOMIC_RELAXED) == 0) {
                claim_zero(mlx5_flow_os_destroy_flow_action
                                                (cache_resource->action));
                ILIST_REMOVE(priv->sh->ipool[MLX5_IPOOL_PORT_ID],
@@ -10349,8 +10606,9 @@ flow_dv_push_vlan_action_resource_release(struct rte_eth_dev *dev,
        MLX5_ASSERT(cache_resource->action);
        DRV_LOG(DEBUG, "push VLAN action resource %p: refcnt %d--",
                (void *)cache_resource,
-               rte_atomic32_read(&cache_resource->refcnt));
-       if (rte_atomic32_dec_and_test(&cache_resource->refcnt)) {
+               __atomic_load_n(&cache_resource->refcnt, __ATOMIC_RELAXED));
+       if (__atomic_sub_fetch(&cache_resource->refcnt, 1,
+                              __ATOMIC_RELAXED) == 0) {
                claim_zero(mlx5_flow_os_destroy_flow_action
                                                (cache_resource->action));
                ILIST_REMOVE(priv->sh->ipool[MLX5_IPOOL_PUSH_VLAN],
@@ -10379,9 +10637,6 @@ flow_dv_fate_resource_release(struct rte_eth_dev *dev,
        if (!handle->rix_fate)
                return;
        switch (handle->fate_action) {
-       case MLX5_FLOW_FATE_DROP:
-               mlx5_drop_action_destroy(dev);
-               break;
        case MLX5_FLOW_FATE_QUEUE:
                mlx5_hrxq_release(dev, handle->rix_hrxq);
                break;
@@ -10392,9 +10647,6 @@ flow_dv_fate_resource_release(struct rte_eth_dev *dev,
                flow_dv_port_id_action_resource_release(dev,
                                handle->rix_port_id_action);
                break;
-       case MLX5_FLOW_FATE_DEFAULT_MISS:
-               flow_dv_default_miss_resource_release(dev);
-               break;
        default:
                DRV_LOG(DEBUG, "Incorrect fate action:%d", handle->fate_action);
                break;
@@ -10564,9 +10816,7 @@ __flow_dv_remove(struct rte_eth_dev *dev, struct rte_flow *flow)
                        claim_zero(mlx5_flow_os_destroy_flow(dh->drv_flow));
                        dh->drv_flow = NULL;
                }
-               if (dh->fate_action == MLX5_FLOW_FATE_DROP ||
-                   dh->fate_action == MLX5_FLOW_FATE_QUEUE ||
-                   dh->fate_action == MLX5_FLOW_FATE_DEFAULT_MISS)
+               if (dh->fate_action == MLX5_FLOW_FATE_QUEUE)
                        flow_dv_fate_resource_release(dev, dh);
                if (dh->vf_vlan.tag && dh->vf_vlan.created)
                        mlx5_vlan_vmwa_release(dev, &dh->vf_vlan);
@@ -10586,12 +10836,16 @@ __flow_dv_remove(struct rte_eth_dev *dev, struct rte_flow *flow)
 static void
 __flow_dv_destroy(struct rte_eth_dev *dev, struct rte_flow *flow)
 {
+       struct rte_flow_shared_action *shared;
        struct mlx5_flow_handle *dev_handle;
        struct mlx5_priv *priv = dev->data->dev_private;
 
        if (!flow)
                return;
        __flow_dv_remove(dev, flow);
+       shared = mlx5_flow_get_shared_rss(flow);
+       if (shared)
+               __atomic_sub_fetch(&shared->refcnt, 1, __ATOMIC_RELAXED);
        if (flow->counter) {
                flow_dv_counter_release(dev, flow->counter);
                flow->counter = 0;
@@ -10636,6 +10890,423 @@ __flow_dv_destroy(struct rte_eth_dev *dev, struct rte_flow *flow)
        }
 }
 
+/**
+ * Release array of hash RX queue objects.
+ * Helper function.
+ *
+ * @param[in] dev
+ *   Pointer to the Ethernet device structure.
+ * @param[in, out] hrxqs
+ *   Array of hash RX queue objects.
+ *
+ * @return
+ *   Total number of references to hash RX queue objects in *hrxqs* array
+ *   after this operation.
+ */
+static int
+__flow_dv_hrxqs_release(struct rte_eth_dev *dev,
+                       uint32_t (*hrxqs)[MLX5_RSS_HASH_FIELDS_LEN])
+{
+       size_t i;
+       int remaining = 0;
+
+       for (i = 0; i < RTE_DIM(*hrxqs); i++) {
+               int ret = mlx5_hrxq_release(dev, (*hrxqs)[i]);
+
+               if (!ret)
+                       (*hrxqs)[i] = 0;
+               remaining += ret;
+       }
+       return remaining;
+}
+
+/**
+ * Release all hash RX queue objects representing shared RSS action.
+ *
+ * @param[in] dev
+ *   Pointer to the Ethernet device structure.
+ * @param[in, out] action
+ *   Shared RSS action to remove hash RX queue objects from.
+ *
+ * @return
+ *   Total number of references to hash RX queue objects stored in *action*
+ *   after this operation.
+ *   Expected to be 0 if no external references held.
+ */
+static int
+__flow_dv_action_rss_hrxqs_release(struct rte_eth_dev *dev,
+                                struct mlx5_shared_action_rss *action)
+{
+       return __flow_dv_hrxqs_release(dev, &action->hrxq) +
+               __flow_dv_hrxqs_release(dev, &action->hrxq_tunnel);
+}
+
+/**
+ * Setup shared RSS action.
+ * Prepare set of hash RX queue objects sufficient to handle all valid
+ * hash_fields combinations (see enum ibv_rx_hash_fields).
+ *
+ * @param[in] dev
+ *   Pointer to the Ethernet device structure.
+ * @param[in, out] action
+ *   Partially initialized shared RSS action.
+ * @param[out] error
+ *   Perform verbose error reporting if not NULL. Initialized in case of
+ *   error only.
+ *
+ * @return
+ *   0 on success, otherwise negative errno value.
+ */
+static int
+__flow_dv_action_rss_setup(struct rte_eth_dev *dev,
+                       struct mlx5_shared_action_rss *action,
+                       struct rte_flow_error *error)
+{
+       size_t i;
+       int err;
+
+       for (i = 0; i < MLX5_RSS_HASH_FIELDS_LEN; i++) {
+               uint32_t hrxq_idx;
+               uint64_t hash_fields = mlx5_rss_hash_fields[i];
+               int tunnel;
+
+               for (tunnel = 0; tunnel < 2; tunnel++) {
+                       hrxq_idx = mlx5_hrxq_new(dev, action->origin.key,
+                                       MLX5_RSS_HASH_KEY_LEN,
+                                       hash_fields,
+                                       action->origin.queue,
+                                       action->origin.queue_num,
+                                       tunnel, true);
+                       if (!hrxq_idx) {
+                               rte_flow_error_set
+                                       (error, rte_errno,
+                                        RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
+                                        "cannot get hash queue");
+                               goto error_hrxq_new;
+                       }
+                       err = __flow_dv_action_rss_hrxq_set
+                               (action, hash_fields, tunnel, hrxq_idx);
+                       MLX5_ASSERT(!err);
+               }
+       }
+       return 0;
+error_hrxq_new:
+       err = rte_errno;
+       __flow_dv_action_rss_hrxqs_release(dev, action);
+       rte_errno = err;
+       return -rte_errno;
+}
+
+/**
+ * Create shared RSS action.
+ *
+ * @param[in] dev
+ *   Pointer to the Ethernet device structure.
+ * @param[in] conf
+ *   Shared action configuration.
+ * @param[in] rss
+ *   RSS action specification used to create shared action.
+ * @param[out] error
+ *   Perform verbose error reporting if not NULL. Initialized in case of
+ *   error only.
+ *
+ * @return
+ *   A valid shared action handle in case of success, NULL otherwise and
+ *   rte_errno is set.
+ */
+static struct rte_flow_shared_action *
+__flow_dv_action_rss_create(struct rte_eth_dev *dev,
+                           const struct rte_flow_shared_action_conf *conf,
+                           const struct rte_flow_action_rss *rss,
+                           struct rte_flow_error *error)
+{
+       struct rte_flow_shared_action *shared_action = NULL;
+       void *queue = NULL;
+       struct mlx5_shared_action_rss *shared_rss;
+       struct rte_flow_action_rss *origin;
+       const uint8_t *rss_key;
+       uint32_t queue_size = rss->queue_num * sizeof(uint16_t);
+
+       RTE_SET_USED(conf);
+       queue = mlx5_malloc(0, RTE_ALIGN_CEIL(queue_size, sizeof(void *)),
+                           0, SOCKET_ID_ANY);
+       shared_action = mlx5_malloc(MLX5_MEM_ZERO, sizeof(*shared_action), 0,
+                                   SOCKET_ID_ANY);
+       if (!shared_action || !queue) {
+               rte_flow_error_set(error, ENOMEM,
+                                  RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
+                                  "cannot allocate resource memory");
+               goto error_rss_init;
+       }
+       shared_rss = &shared_action->rss;
+       shared_rss->queue = queue;
+       origin = &shared_rss->origin;
+       origin->func = rss->func;
+       origin->level = rss->level;
+       /* RSS type 0 indicates default RSS type (ETH_RSS_IP). */
+       origin->types = !rss->types ? ETH_RSS_IP : rss->types;
+       /* NULL RSS key indicates default RSS key. */
+       rss_key = !rss->key ? rss_hash_default_key : rss->key;
+       memcpy(shared_rss->key, rss_key, MLX5_RSS_HASH_KEY_LEN);
+       origin->key = &shared_rss->key[0];
+       origin->key_len = MLX5_RSS_HASH_KEY_LEN;
+       memcpy(shared_rss->queue, rss->queue, queue_size);
+       origin->queue = shared_rss->queue;
+       origin->queue_num = rss->queue_num;
+       if (__flow_dv_action_rss_setup(dev, shared_rss, error))
+               goto error_rss_init;
+       shared_action->type = MLX5_RTE_FLOW_ACTION_TYPE_SHARED_RSS;
+       return shared_action;
+error_rss_init:
+       mlx5_free(shared_action);
+       mlx5_free(queue);
+       return NULL;
+}
+
+/**
+ * Destroy the shared RSS action.
+ * Release related hash RX queue objects.
+ *
+ * @param[in] dev
+ *   Pointer to the Ethernet device structure.
+ * @param[in] shared_rss
+ *   The shared RSS action object to be removed.
+ * @param[out] error
+ *   Perform verbose error reporting if not NULL. Initialized in case of
+ *   error only.
+ *
+ * @return
+ *   0 on success, otherwise negative errno value.
+ */
+static int
+__flow_dv_action_rss_release(struct rte_eth_dev *dev,
+                        struct mlx5_shared_action_rss *shared_rss,
+                        struct rte_flow_error *error)
+{
+       struct rte_flow_shared_action *shared_action = NULL;
+       uint32_t old_refcnt = 1;
+       int remaining = __flow_dv_action_rss_hrxqs_release(dev, shared_rss);
+
+       if (remaining) {
+               return rte_flow_error_set(error, ETOOMANYREFS,
+                                         RTE_FLOW_ERROR_TYPE_ACTION,
+                                         NULL,
+                                         "shared rss hrxq has references");
+       }
+       shared_action = container_of(shared_rss,
+                                    struct rte_flow_shared_action, rss);
+       if (!__atomic_compare_exchange_n(&shared_action->refcnt, &old_refcnt,
+                                        0, 0,
+                                        __ATOMIC_ACQUIRE, __ATOMIC_RELAXED)) {
+               return rte_flow_error_set(error, ETOOMANYREFS,
+                                         RTE_FLOW_ERROR_TYPE_ACTION,
+                                         NULL,
+                                         "shared rss has references");
+       }
+       rte_free(shared_rss->queue);
+       return 0;
+}
+
+/**
+ * Create shared action, lock free,
+ * (mutex should be acquired by caller).
+ * Dispatcher for action type specific call.
+ *
+ * @param[in] dev
+ *   Pointer to the Ethernet device structure.
+ * @param[in] conf
+ *   Shared action configuration.
+ * @param[in] action
+ *   Action specification used to create shared action.
+ * @param[out] error
+ *   Perform verbose error reporting if not NULL. Initialized in case of
+ *   error only.
+ *
+ * @return
+ *   A valid shared action handle in case of success, NULL otherwise and
+ *   rte_errno is set.
+ */
+static struct rte_flow_shared_action *
+__flow_dv_action_create(struct rte_eth_dev *dev,
+                       const struct rte_flow_shared_action_conf *conf,
+                       const struct rte_flow_action *action,
+                       struct rte_flow_error *error)
+{
+       struct rte_flow_shared_action *shared_action = NULL;
+       struct mlx5_priv *priv = dev->data->dev_private;
+
+       switch (action->type) {
+       case RTE_FLOW_ACTION_TYPE_RSS:
+               shared_action = __flow_dv_action_rss_create(dev, conf,
+                                                           action->conf,
+                                                           error);
+               break;
+       default:
+               rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
+                                  NULL, "action type not supported");
+               break;
+       }
+       if (shared_action) {
+               __atomic_add_fetch(&shared_action->refcnt, 1,
+                                  __ATOMIC_RELAXED);
+               LIST_INSERT_HEAD(&priv->shared_actions, shared_action, next);
+       }
+       return shared_action;
+}
+
+/**
+ * Destroy the shared action.
+ * Release action related resources on the NIC and the memory.
+ * Lock free, (mutex should be acquired by caller).
+ * Dispatcher for action type specific call.
+ *
+ * @param[in] dev
+ *   Pointer to the Ethernet device structure.
+ * @param[in] action
+ *   The shared action object to be removed.
+ * @param[out] error
+ *   Perform verbose error reporting if not NULL. Initialized in case of
+ *   error only.
+ *
+ * @return
+ *   0 on success, otherwise negative errno value.
+ */
+static int
+__flow_dv_action_destroy(struct rte_eth_dev *dev,
+                        struct rte_flow_shared_action *action,
+                        struct rte_flow_error *error)
+{
+       int ret;
+
+       switch (action->type) {
+       case MLX5_RTE_FLOW_ACTION_TYPE_SHARED_RSS:
+               ret = __flow_dv_action_rss_release(dev, &action->rss, error);
+               break;
+       default:
+               return rte_flow_error_set(error, ENOTSUP,
+                                         RTE_FLOW_ERROR_TYPE_ACTION,
+                                         NULL,
+                                         "action type not supported");
+       }
+       if (ret)
+               return ret;
+       LIST_REMOVE(action, next);
+       rte_free(action);
+       return 0;
+}
+
+/**
+ * Updates in place shared RSS action configuration.
+ *
+ * @param[in] dev
+ *   Pointer to the Ethernet device structure.
+ * @param[in] shared_rss
+ *   The shared RSS action object to be updated.
+ * @param[in] action_conf
+ *   RSS action specification used to modify *shared_rss*.
+ * @param[out] error
+ *   Perform verbose error reporting if not NULL. Initialized in case of
+ *   error only.
+ *
+ * @return
+ *   0 on success, otherwise negative errno value.
+ * @note: currently only support update of RSS queues.
+ */
+static int
+__flow_dv_action_rss_update(struct rte_eth_dev *dev,
+                           struct mlx5_shared_action_rss *shared_rss,
+                           const struct rte_flow_action_rss *action_conf,
+                           struct rte_flow_error *error)
+{
+       size_t i;
+       int ret;
+       void *queue = NULL;
+       const uint8_t *rss_key;
+       uint32_t rss_key_len;
+       uint32_t queue_size = action_conf->queue_num * sizeof(uint16_t);
+
+       queue = mlx5_malloc(MLX5_MEM_ZERO,
+                           RTE_ALIGN_CEIL(queue_size, sizeof(void *)),
+                           0, SOCKET_ID_ANY);
+       if (!queue)
+               return rte_flow_error_set(error, ENOMEM,
+                                         RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
+                                         NULL,
+                                         "cannot allocate resource memory");
+       if (action_conf->key) {
+               rss_key = action_conf->key;
+               rss_key_len = action_conf->key_len;
+       } else {
+               rss_key = rss_hash_default_key;
+               rss_key_len = MLX5_RSS_HASH_KEY_LEN;
+       }
+       for (i = 0; i < MLX5_RSS_HASH_FIELDS_LEN; i++) {
+               uint32_t hrxq_idx;
+               uint64_t hash_fields = mlx5_rss_hash_fields[i];
+               int tunnel;
+
+               for (tunnel = 0; tunnel < 2; tunnel++) {
+                       hrxq_idx = __flow_dv_action_rss_hrxq_lookup
+                                       (shared_rss, hash_fields, tunnel);
+                       MLX5_ASSERT(hrxq_idx);
+                       ret = mlx5_hrxq_modify
+                               (dev, hrxq_idx,
+                                rss_key, rss_key_len,
+                                hash_fields,
+                                action_conf->queue, action_conf->queue_num);
+                       if (ret) {
+                               mlx5_free(queue);
+                               return rte_flow_error_set
+                                       (error, rte_errno,
+                                        RTE_FLOW_ERROR_TYPE_ACTION, NULL,
+                                        "cannot update hash queue");
+                       }
+               }
+       }
+       mlx5_free(shared_rss->queue);
+       shared_rss->queue = queue;
+       memcpy(shared_rss->queue, action_conf->queue, queue_size);
+       shared_rss->origin.queue = shared_rss->queue;
+       shared_rss->origin.queue_num = action_conf->queue_num;
+       return 0;
+}
+
+/**
+ * Updates in place shared action configuration, lock free,
+ * (mutex should be acquired by caller).
+ *
+ * @param[in] dev
+ *   Pointer to the Ethernet device structure.
+ * @param[in] action
+ *   The shared action object to be updated.
+ * @param[in] action_conf
+ *   Action specification used to modify *action*.
+ *   *action_conf* should be of type correlating with type of the *action*,
+ *   otherwise considered as invalid.
+ * @param[out] error
+ *   Perform verbose error reporting if not NULL. Initialized in case of
+ *   error only.
+ *
+ * @return
+ *   0 on success, otherwise negative errno value.
+ */
+static int
+__flow_dv_action_update(struct rte_eth_dev *dev,
+                       struct rte_flow_shared_action *action,
+                       const void *action_conf,
+                       struct rte_flow_error *error)
+{
+       switch (action->type) {
+       case MLX5_RTE_FLOW_ACTION_TYPE_SHARED_RSS:
+               return __flow_dv_action_rss_update(dev, &action->rss,
+                                                  action_conf, error);
+       default:
+               return rte_flow_error_set(error, ENOTSUP,
+                                         RTE_FLOW_ERROR_TYPE_ACTION,
+                                         NULL,
+                                         "action type not supported");
+       }
+}
 /**
  * Query a dv flow  rule for its statistics via devx.
  *
@@ -10894,7 +11565,8 @@ flow_dv_prepare_mtr_tables(struct rte_eth_dev *dev,
                dtb = &mtb->ingress;
        /* Create the meter table with METER level. */
        dtb->tbl = flow_dv_tbl_resource_get(dev, MLX5_FLOW_TABLE_LEVEL_METER,
-                                           egress, transfer, &error);
+                                           egress, transfer, false, NULL, 0,
+                                           0, &error);
        if (!dtb->tbl) {
                DRV_LOG(ERR, "Failed to create meter policer table.");
                return -1;
@@ -10902,7 +11574,8 @@ flow_dv_prepare_mtr_tables(struct rte_eth_dev *dev,
        /* Create the meter suffix table with SUFFIX level. */
        dtb->sfx_tbl = flow_dv_tbl_resource_get(dev,
                                            MLX5_FLOW_TABLE_LEVEL_SUFFIX,
-                                           egress, transfer, &error);
+                                           egress, transfer, false, NULL, 0,
+                                           0, &error);
        if (!dtb->sfx_tbl) {
                DRV_LOG(ERR, "Failed to create meter suffix table.");
                return -1;
@@ -11221,10 +11894,11 @@ mlx5_flow_dv_discover_counter_offset_support(struct rte_eth_dev *dev)
        void *flow = NULL;
        int i, ret = -1;
 
-       tbl = flow_dv_tbl_resource_get(dev, 0, 0, 0, NULL);
+       tbl = flow_dv_tbl_resource_get(dev, 0, 0, 0, false, NULL, 0, 0, NULL);
        if (!tbl)
                goto err;
-       dest_tbl = flow_dv_tbl_resource_get(dev, 1, 0, 0, NULL);
+       dest_tbl = flow_dv_tbl_resource_get(dev, 1, 0, 0, false,
+                                           NULL, 0, 0, NULL);
        if (!dest_tbl)
                goto err;
        dcs = mlx5_devx_cmd_flow_counter_alloc(priv->sh->ctx, 0x4);
@@ -11460,6 +12134,117 @@ flow_dv_counter_free(struct rte_eth_dev *dev, uint32_t cnt)
        flow_dv_shared_unlock(dev);
 }
 
+/**
+ * Validate shared action.
+ * Dispatcher for action type specific validation.
+ *
+ * @param[in] dev
+ *   Pointer to the Ethernet device structure.
+ * @param[in] conf
+ *   Shared action configuration.
+ * @param[in] action
+ *   The shared action object to validate.
+ * @param[out] error
+ *   Perform verbose error reporting if not NULL. Initialized in case of
+ *   error only.
+ *
+ * @return
+ *   0 on success, otherwise negative errno value.
+ */
+static int
+flow_dv_action_validate(struct rte_eth_dev *dev,
+                       const struct rte_flow_shared_action_conf *conf,
+                       const struct rte_flow_action *action,
+                       struct rte_flow_error *error)
+{
+       RTE_SET_USED(conf);
+       switch (action->type) {
+       case RTE_FLOW_ACTION_TYPE_RSS:
+               return mlx5_validate_action_rss(dev, action, error);
+       default:
+               return rte_flow_error_set(error, ENOTSUP,
+                                         RTE_FLOW_ERROR_TYPE_ACTION,
+                                         NULL,
+                                         "action type not supported");
+       }
+}
+
+/*
+ * Mutex-protected thunk to lock-free  __flow_dv_action_create().
+ */
+static struct rte_flow_shared_action *
+flow_dv_action_create(struct rte_eth_dev *dev,
+                     const struct rte_flow_shared_action_conf *conf,
+                     const struct rte_flow_action *action,
+                     struct rte_flow_error *error)
+{
+       struct rte_flow_shared_action *shared_action = NULL;
+
+       flow_dv_shared_lock(dev);
+       shared_action = __flow_dv_action_create(dev, conf, action, error);
+       flow_dv_shared_unlock(dev);
+       return shared_action;
+}
+
+/*
+ * Mutex-protected thunk to lock-free  __flow_dv_action_destroy().
+ */
+static int
+flow_dv_action_destroy(struct rte_eth_dev *dev,
+                      struct rte_flow_shared_action *action,
+                      struct rte_flow_error *error)
+{
+       int ret;
+
+       flow_dv_shared_lock(dev);
+       ret = __flow_dv_action_destroy(dev, action, error);
+       flow_dv_shared_unlock(dev);
+       return ret;
+}
+
+/*
+ * Mutex-protected thunk to lock-free  __flow_dv_action_update().
+ */
+static int
+flow_dv_action_update(struct rte_eth_dev *dev,
+                     struct rte_flow_shared_action *action,
+                     const void *action_conf,
+                     struct rte_flow_error *error)
+{
+       int ret;
+
+       flow_dv_shared_lock(dev);
+       ret = __flow_dv_action_update(dev, action, action_conf,
+                                     error);
+       flow_dv_shared_unlock(dev);
+       return ret;
+}
+
+static int
+flow_dv_sync_domain(struct rte_eth_dev *dev, uint32_t domains, uint32_t flags)
+{
+       struct mlx5_priv *priv = dev->data->dev_private;
+       int ret = 0;
+
+       if ((domains & MLX5_DOMAIN_BIT_NIC_RX) && priv->sh->rx_domain != NULL) {
+               ret = mlx5_glue->dr_sync_domain(priv->sh->rx_domain,
+                                               flags);
+               if (ret != 0)
+                       return ret;
+       }
+       if ((domains & MLX5_DOMAIN_BIT_NIC_TX) && priv->sh->tx_domain != NULL) {
+               ret = mlx5_glue->dr_sync_domain(priv->sh->tx_domain, flags);
+               if (ret != 0)
+                       return ret;
+       }
+       if ((domains & MLX5_DOMAIN_BIT_FDB) && priv->sh->fdb_domain != NULL) {
+               ret = mlx5_glue->dr_sync_domain(priv->sh->fdb_domain, flags);
+               if (ret != 0)
+                       return ret;
+       }
+       return 0;
+}
+
 const struct mlx5_flow_driver_ops mlx5_flow_dv_drv_ops = {
        .validate = flow_dv_validate,
        .prepare = flow_dv_prepare,
@@ -11476,6 +12261,11 @@ const struct mlx5_flow_driver_ops mlx5_flow_dv_drv_ops = {
        .counter_free = flow_dv_counter_free,
        .counter_query = flow_dv_counter_query,
        .get_aged_flows = flow_get_aged_flows,
+       .action_validate = flow_dv_action_validate,
+       .action_create = flow_dv_action_create,
+       .action_destroy = flow_dv_action_destroy,
+       .action_update = flow_dv_action_update,
+       .sync_domain = flow_dv_sync_domain,
 };
 
 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */