net/mlx5: fix shared RSS and mark actions combination
[dpdk.git] / drivers / net / mlx5 / mlx5_flow.c
index b08ee30..2a4073c 100644 (file)
 #include "mlx5_common_os.h"
 #include "rte_pmd_mlx5.h"
 
+struct tunnel_default_miss_ctx {
+       uint16_t *queue;
+       __extension__
+       union {
+               struct rte_flow_action_rss action_rss;
+               struct rte_flow_action_queue miss_queue;
+               struct rte_flow_action_jump miss_jump;
+               uint8_t raw[0];
+       };
+};
+
+static int
+flow_tunnel_add_default_miss(struct rte_eth_dev *dev,
+                            struct rte_flow *flow,
+                            const struct rte_flow_attr *attr,
+                            const struct rte_flow_action *app_actions,
+                            uint32_t flow_idx,
+                            struct tunnel_default_miss_ctx *ctx,
+                            struct rte_flow_error *error);
 static struct mlx5_flow_tunnel *
 mlx5_find_tunnel_id(struct rte_eth_dev *dev, uint32_t id);
 static void
 mlx5_flow_tunnel_free(struct rte_eth_dev *dev, struct mlx5_flow_tunnel *tunnel);
-static const struct mlx5_flow_tbl_data_entry  *
-tunnel_mark_decode(struct rte_eth_dev *dev, uint32_t mark);
-static int
-mlx5_get_flow_tunnel(struct rte_eth_dev *dev,
-                    const struct rte_flow_tunnel *app_tunnel,
-                    struct mlx5_flow_tunnel **tunnel);
+static uint32_t
+tunnel_flow_group_to_flow_table(struct rte_eth_dev *dev,
+                               const struct mlx5_flow_tunnel *tunnel,
+                               uint32_t group, uint32_t *table,
+                               struct rte_flow_error *error);
+
+static struct mlx5_flow_workspace *mlx5_flow_push_thread_workspace(void);
+static void mlx5_flow_pop_thread_workspace(void);
 
 
 /** Device flow drivers. */
@@ -52,7 +73,7 @@ const struct mlx5_flow_driver_ops mlx5_flow_null_drv_ops;
 
 const struct mlx5_flow_driver_ops *flow_drv_ops[] = {
        [MLX5_FLOW_TYPE_MIN] = &mlx5_flow_null_drv_ops,
-#ifdef HAVE_IBV_FLOW_DV_SUPPORT
+#if defined(HAVE_IBV_FLOW_DV_SUPPORT) || !defined(HAVE_INFINIBAND_VERBS_H)
        [MLX5_FLOW_TYPE_DV] = &mlx5_flow_dv_drv_ops,
 #endif
        [MLX5_FLOW_TYPE_VERBS] = &mlx5_flow_verbs_drv_ops,
@@ -191,6 +212,8 @@ mlx5_flow_expand_rss_item_complete(const struct rte_flow_item *item)
        return ret;
 }
 
+#define MLX5_RSS_EXP_ELT_N 8
+
 /**
  * Expand RSS flows into several possible flows according to the RSS hash
  * fields requested and the driver capabilities.
@@ -221,13 +244,12 @@ mlx5_flow_expand_rss(struct mlx5_flow_expand_rss *buf, size_t size,
                     const struct mlx5_flow_expand_node graph[],
                     int graph_root_index)
 {
-       const int elt_n = 8;
        const struct rte_flow_item *item;
        const struct mlx5_flow_expand_node *node = &graph[graph_root_index];
        const int *next_node;
-       const int *stack[elt_n];
+       const int *stack[MLX5_RSS_EXP_ELT_N];
        int stack_pos = 0;
-       struct rte_flow_item flow_items[elt_n];
+       struct rte_flow_item flow_items[MLX5_RSS_EXP_ELT_N];
        unsigned int i;
        size_t lsize;
        size_t user_pattern_size = 0;
@@ -240,10 +262,10 @@ mlx5_flow_expand_rss(struct mlx5_flow_expand_rss *buf, size_t size,
 
        memset(&missed_item, 0, sizeof(missed_item));
        lsize = offsetof(struct mlx5_flow_expand_rss, entry) +
-               elt_n * sizeof(buf->entry[0]);
+               MLX5_RSS_EXP_ELT_N * sizeof(buf->entry[0]);
        if (lsize <= size) {
                buf->entry[0].priority = 0;
-               buf->entry[0].pattern = (void *)&buf->entry[elt_n];
+               buf->entry[0].pattern = (void *)&buf->entry[MLX5_RSS_EXP_ELT_N];
                buf->entries = 0;
                addr = buf->entry[0].pattern;
        }
@@ -346,7 +368,7 @@ mlx5_flow_expand_rss(struct mlx5_flow_expand_rss *buf, size_t size,
                /* Go deeper. */
                if (node->next) {
                        next_node = node->next;
-                       if (stack_pos++ == elt_n) {
+                       if (stack_pos++ == MLX5_RSS_EXP_ELT_N) {
                                rte_errno = E2BIG;
                                return -rte_errno;
                        }
@@ -377,8 +399,6 @@ mlx5_flow_expand_rss(struct mlx5_flow_expand_rss *buf, size_t size,
                                   user_pattern_size);
                        addr = (void *)(((uintptr_t)addr) + user_pattern_size);
                        rte_memcpy(addr, flow_items, elt * sizeof(*item));
-                       addr = (void *)(((uintptr_t)addr) +
-                                       elt * sizeof(*item));
                }
        }
        return lsize;
@@ -580,170 +600,32 @@ static int mlx5_shared_action_query
                                 const struct rte_flow_shared_action *action,
                                 void *data,
                                 struct rte_flow_error *error);
-static inline bool
-mlx5_flow_tunnel_validate(struct rte_eth_dev *dev,
-                         struct rte_flow_tunnel *tunnel,
-                         const char *err_msg)
-{
-       err_msg = NULL;
-       if (!is_tunnel_offload_active(dev)) {
-               err_msg = "tunnel offload was not activated";
-               goto out;
-       } else if (!tunnel) {
-               err_msg = "no application tunnel";
-               goto out;
-       }
-
-       switch (tunnel->type) {
-       default:
-               err_msg = "unsupported tunnel type";
-               goto out;
-       case RTE_FLOW_ITEM_TYPE_VXLAN:
-               break;
-       }
-
-out:
-       return !err_msg;
-}
-
-
 static int
 mlx5_flow_tunnel_decap_set(struct rte_eth_dev *dev,
                    struct rte_flow_tunnel *app_tunnel,
                    struct rte_flow_action **actions,
                    uint32_t *num_of_actions,
-                   struct rte_flow_error *error)
-{
-       int ret;
-       struct mlx5_flow_tunnel *tunnel;
-       const char *err_msg = NULL;
-       bool verdict = mlx5_flow_tunnel_validate(dev, app_tunnel, err_msg);
-
-       if (!verdict)
-               return rte_flow_error_set(error, EINVAL,
-                                         RTE_FLOW_ERROR_TYPE_ACTION_CONF, NULL,
-                                         err_msg);
-       ret = mlx5_get_flow_tunnel(dev, app_tunnel, &tunnel);
-       if (ret < 0) {
-               return rte_flow_error_set(error, ret,
-                                         RTE_FLOW_ERROR_TYPE_ACTION_CONF, NULL,
-                                         "failed to initialize pmd tunnel");
-       }
-       *actions = &tunnel->action;
-       *num_of_actions = 1;
-       return 0;
-}
-
+                   struct rte_flow_error *error);
 static int
 mlx5_flow_tunnel_match(struct rte_eth_dev *dev,
                       struct rte_flow_tunnel *app_tunnel,
                       struct rte_flow_item **items,
                       uint32_t *num_of_items,
-                      struct rte_flow_error *error)
-{
-       int ret;
-       struct mlx5_flow_tunnel *tunnel;
-       const char *err_msg = NULL;
-       bool verdict = mlx5_flow_tunnel_validate(dev, app_tunnel, err_msg);
-
-       if (!verdict)
-               return rte_flow_error_set(error, EINVAL,
-                                         RTE_FLOW_ERROR_TYPE_HANDLE, NULL,
-                                         err_msg);
-       ret = mlx5_get_flow_tunnel(dev, app_tunnel, &tunnel);
-       if (ret < 0) {
-               return rte_flow_error_set(error, ret,
-                                         RTE_FLOW_ERROR_TYPE_HANDLE, NULL,
-                                         "failed to initialize pmd tunnel");
-       }
-       *items = &tunnel->item;
-       *num_of_items = 1;
-       return 0;
-}
-
+                      struct rte_flow_error *error);
 static int
-mlx5_flow_item_release(struct rte_eth_dev *dev,
-                      struct rte_flow_item *pmd_items,
-                      uint32_t num_items, struct rte_flow_error *err)
-{
-       struct mlx5_flow_tunnel_hub *thub = mlx5_tunnel_hub(dev);
-       struct mlx5_flow_tunnel *tun;
-
-       rte_spinlock_lock(&thub->sl);
-       LIST_FOREACH(tun, &thub->tunnels, chain) {
-               if (&tun->item == pmd_items) {
-                       LIST_REMOVE(tun, chain);
-                       break;
-               }
-       }
-       rte_spinlock_unlock(&thub->sl);
-       if (!tun || num_items != 1)
-               return rte_flow_error_set(err, EINVAL,
-                                         RTE_FLOW_ERROR_TYPE_HANDLE, NULL,
-                                         "invalid argument");
-       if (!__atomic_sub_fetch(&tun->refctn, 1, __ATOMIC_RELAXED))
-               mlx5_flow_tunnel_free(dev, tun);
-       return 0;
-}
-
+mlx5_flow_tunnel_item_release(struct rte_eth_dev *dev,
+                             struct rte_flow_item *pmd_items,
+                             uint32_t num_items, struct rte_flow_error *err);
 static int
-mlx5_flow_action_release(struct rte_eth_dev *dev,
-                        struct rte_flow_action *pmd_actions,
-                        uint32_t num_actions, struct rte_flow_error *err)
-{
-       struct mlx5_flow_tunnel_hub *thub = mlx5_tunnel_hub(dev);
-       struct mlx5_flow_tunnel *tun;
-
-       rte_spinlock_lock(&thub->sl);
-       LIST_FOREACH(tun, &thub->tunnels, chain) {
-               if (&tun->action == pmd_actions) {
-                       LIST_REMOVE(tun, chain);
-                       break;
-               }
-       }
-       rte_spinlock_unlock(&thub->sl);
-       if (!tun || num_actions != 1)
-               return rte_flow_error_set(err, EINVAL,
-                                         RTE_FLOW_ERROR_TYPE_HANDLE, NULL,
-                                         "invalid argument");
-       if (!__atomic_sub_fetch(&tun->refctn, 1, __ATOMIC_RELAXED))
-               mlx5_flow_tunnel_free(dev, tun);
-
-       return 0;
-}
-
+mlx5_flow_tunnel_action_release(struct rte_eth_dev *dev,
+                               struct rte_flow_action *pmd_actions,
+                               uint32_t num_actions,
+                               struct rte_flow_error *err);
 static int
 mlx5_flow_tunnel_get_restore_info(struct rte_eth_dev *dev,
                                  struct rte_mbuf *m,
                                  struct rte_flow_restore_info *info,
-                                 struct rte_flow_error *err)
-{
-       uint64_t ol_flags = m->ol_flags;
-       const struct mlx5_flow_tbl_data_entry *tble;
-       const uint64_t mask = PKT_RX_FDIR | PKT_RX_FDIR_ID;
-
-       if ((ol_flags & mask) != mask)
-               goto err;
-       tble = tunnel_mark_decode(dev, m->hash.fdir.hi);
-       if (!tble) {
-               DRV_LOG(DEBUG, "port %u invalid miss tunnel mark %#x",
-                       dev->data->port_id, m->hash.fdir.hi);
-               goto err;
-       }
-       MLX5_ASSERT(tble->tunnel);
-       memcpy(&info->tunnel, &tble->tunnel->app_tunnel, sizeof(info->tunnel));
-       info->group_id = tble->group_id;
-       info->flags = RTE_FLOW_RESTORE_INFO_TUNNEL |
-                     RTE_FLOW_RESTORE_INFO_GROUP_ID |
-                     RTE_FLOW_RESTORE_INFO_ENCAPSULATED;
-
-       return 0;
-
-err:
-       return rte_flow_error_set(err, EINVAL,
-                                 RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
-                                 "failed to get restore info");
-}
+                                 struct rte_flow_error *err);
 
 static const struct rte_flow_ops mlx5_flow_ops = {
        .validate = mlx5_flow_validate,
@@ -760,8 +642,8 @@ static const struct rte_flow_ops mlx5_flow_ops = {
        .shared_action_query = mlx5_shared_action_query,
        .tunnel_decap_set = mlx5_flow_tunnel_decap_set,
        .tunnel_match = mlx5_flow_tunnel_match,
-       .tunnel_action_decap_release = mlx5_flow_action_release,
-       .tunnel_item_release = mlx5_flow_item_release,
+       .tunnel_action_decap_release = mlx5_flow_tunnel_action_release,
+       .tunnel_item_release = mlx5_flow_tunnel_item_release,
        .get_restore_info = mlx5_flow_tunnel_get_restore_info,
 };
 
@@ -896,6 +778,7 @@ mlx5_flow_get_reg_id(struct rte_eth_dev *dev,
                        return priv->mtr_color_reg != REG_C_2 ? REG_C_2 :
                               REG_C_3;
        case MLX5_MTR_COLOR:
+       case MLX5_ASO_FLOW_HIT: /* Both features use the same REG_C. */
                MLX5_ASSERT(priv->mtr_color_reg != REG_NON);
                return priv->mtr_color_reg;
        case MLX5_COPY_MARK:
@@ -915,7 +798,7 @@ mlx5_flow_get_reg_id(struct rte_eth_dev *dev,
                start_reg = priv->mtr_color_reg != REG_C_2 ? REG_C_2 :
                            (priv->mtr_reg_share ? REG_C_3 : REG_C_4);
                skip_mtr_reg = !!(priv->mtr_en && start_reg == REG_C_2);
-               if (id > (REG_C_7 - start_reg))
+               if (id > (uint32_t)(REG_C_7 - start_reg))
                        return rte_flow_error_set(error, EINVAL,
                                                  RTE_FLOW_ERROR_TYPE_ITEM,
                                                  NULL, "invalid tag id");
@@ -931,7 +814,7 @@ mlx5_flow_get_reg_id(struct rte_eth_dev *dev,
                 */
                if (skip_mtr_reg && config->flow_mreg_c
                    [id + start_reg - REG_C_0] >= priv->mtr_color_reg) {
-                       if (id >= (REG_C_7 - start_reg))
+                       if (id >= (uint32_t)(REG_C_7 - start_reg))
                                return rte_flow_error_set(error, EINVAL,
                                                       RTE_FLOW_ERROR_TYPE_ITEM,
                                                        NULL, "invalid tag id");
@@ -1119,17 +1002,29 @@ flow_drv_rxq_flags_set(struct rte_eth_dev *dev,
        struct mlx5_priv *priv = dev->data->dev_private;
        const int mark = dev_handle->mark;
        const int tunnel = !!(dev_handle->layers & MLX5_FLOW_LAYER_TUNNEL);
-       struct mlx5_hrxq *hrxq;
+       struct mlx5_ind_table_obj *ind_tbl = NULL;
        unsigned int i;
 
-       if (dev_handle->fate_action != MLX5_FLOW_FATE_QUEUE)
-               return;
-       hrxq = mlx5_ipool_get(priv->sh->ipool[MLX5_IPOOL_HRXQ],
+       if (dev_handle->fate_action == MLX5_FLOW_FATE_QUEUE) {
+               struct mlx5_hrxq *hrxq;
+
+               hrxq = mlx5_ipool_get(priv->sh->ipool[MLX5_IPOOL_HRXQ],
                              dev_handle->rix_hrxq);
-       if (!hrxq)
+               if (hrxq)
+                       ind_tbl = hrxq->ind_table;
+       } else if (dev_handle->fate_action == MLX5_FLOW_FATE_SHARED_RSS) {
+               struct mlx5_shared_action_rss *shared_rss;
+
+               shared_rss = mlx5_ipool_get
+                       (priv->sh->ipool[MLX5_IPOOL_RSS_SHARED_ACTIONS],
+                        dev_handle->rix_srss);
+               if (shared_rss)
+                       ind_tbl = shared_rss->ind_tbl;
+       }
+       if (!ind_tbl)
                return;
-       for (i = 0; i != hrxq->ind_table->queues_n; ++i) {
-               int idx = hrxq->ind_table->queues[i];
+       for (i = 0; i != ind_tbl->queues_n; ++i) {
+               int idx = ind_tbl->queues[i];
                struct mlx5_rxq_ctrl *rxq_ctrl =
                        container_of((*priv->rxqs)[idx],
                                     struct mlx5_rxq_ctrl, rxq);
@@ -1201,18 +1096,30 @@ flow_drv_rxq_flags_trim(struct rte_eth_dev *dev,
        struct mlx5_priv *priv = dev->data->dev_private;
        const int mark = dev_handle->mark;
        const int tunnel = !!(dev_handle->layers & MLX5_FLOW_LAYER_TUNNEL);
-       struct mlx5_hrxq *hrxq;
+       struct mlx5_ind_table_obj *ind_tbl = NULL;
        unsigned int i;
 
-       if (dev_handle->fate_action != MLX5_FLOW_FATE_QUEUE)
-               return;
-       hrxq = mlx5_ipool_get(priv->sh->ipool[MLX5_IPOOL_HRXQ],
+       if (dev_handle->fate_action == MLX5_FLOW_FATE_QUEUE) {
+               struct mlx5_hrxq *hrxq;
+
+               hrxq = mlx5_ipool_get(priv->sh->ipool[MLX5_IPOOL_HRXQ],
                              dev_handle->rix_hrxq);
-       if (!hrxq)
+               if (hrxq)
+                       ind_tbl = hrxq->ind_table;
+       } else if (dev_handle->fate_action == MLX5_FLOW_FATE_SHARED_RSS) {
+               struct mlx5_shared_action_rss *shared_rss;
+
+               shared_rss = mlx5_ipool_get
+                       (priv->sh->ipool[MLX5_IPOOL_RSS_SHARED_ACTIONS],
+                        dev_handle->rix_srss);
+               if (shared_rss)
+                       ind_tbl = shared_rss->ind_tbl;
+       }
+       if (!ind_tbl)
                return;
        MLX5_ASSERT(dev->data->dev_started);
-       for (i = 0; i != hrxq->ind_table->queues_n; ++i) {
-               int idx = hrxq->ind_table->queues[i];
+       for (i = 0; i != ind_tbl->queues_n; ++i) {
+               int idx = ind_tbl->queues[i];
                struct mlx5_rxq_ctrl *rxq_ctrl =
                        container_of((*priv->rxqs)[idx],
                                     struct mlx5_rxq_ctrl, rxq);
@@ -1530,6 +1437,7 @@ mlx5_validate_action_rss(struct rte_eth_dev *dev,
 {
        struct mlx5_priv *priv = dev->data->dev_private;
        const struct rte_flow_action_rss *rss = action->conf;
+       enum mlx5_rxq_type rxq_type = MLX5_RXQ_TYPE_UNDEFINED;
        unsigned int i;
 
        if (rss->func != RTE_ETH_HASH_FUNCTION_DEFAULT &&
@@ -1595,6 +1503,8 @@ mlx5_validate_action_rss(struct rte_eth_dev *dev,
                                          RTE_FLOW_ERROR_TYPE_ACTION_CONF,
                                          NULL, "No queues configured");
        for (i = 0; i != rss->queue_num; ++i) {
+               struct mlx5_rxq_ctrl *rxq_ctrl;
+
                if (rss->queue[i] >= priv->rxqs_n)
                        return rte_flow_error_set
                                (error, EINVAL,
@@ -1604,6 +1514,15 @@ mlx5_validate_action_rss(struct rte_eth_dev *dev,
                        return rte_flow_error_set
                                (error, EINVAL, RTE_FLOW_ERROR_TYPE_ACTION_CONF,
                                 &rss->queue[i], "queue is not configured");
+               rxq_ctrl = container_of((*priv->rxqs)[rss->queue[i]],
+                                       struct mlx5_rxq_ctrl, rxq);
+               if (i == 0)
+                       rxq_type = rxq_ctrl->type;
+               if (rxq_type != rxq_ctrl->type)
+                       return rte_flow_error_set
+                               (error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION_CONF,
+                                &rss->queue[i],
+                                "combining hairpin and regular RSS queues is not supported");
        }
        return 0;
 }
@@ -2746,7 +2665,8 @@ mlx5_flow_validate_item_mpls(struct rte_eth_dev *dev __rte_unused,
        /* MPLS over IP, UDP, GRE is allowed */
        if (!(prev_layer & (MLX5_FLOW_LAYER_OUTER_L3 |
                            MLX5_FLOW_LAYER_OUTER_L4_UDP |
-                           MLX5_FLOW_LAYER_GRE)))
+                           MLX5_FLOW_LAYER_GRE |
+                           MLX5_FLOW_LAYER_GRE_KEY)))
                return rte_flow_error_set(error, EINVAL,
                                          RTE_FLOW_ERROR_TYPE_ITEM, item,
                                          "protocol filtering not compatible"
@@ -2870,17 +2790,20 @@ mlx5_flow_validate_item_ecpri(const struct rte_flow_item *item,
                                        MLX5_FLOW_LAYER_OUTER_VLAN);
        struct rte_flow_item_ecpri mask_lo;
 
+       if (!(last_item & outer_l2_vlan) &&
+           last_item != MLX5_FLOW_LAYER_OUTER_L4_UDP)
+               return rte_flow_error_set(error, EINVAL,
+                                         RTE_FLOW_ERROR_TYPE_ITEM, item,
+                                         "eCPRI can only follow L2/VLAN layer or UDP layer");
        if ((last_item & outer_l2_vlan) && ether_type &&
            ether_type != RTE_ETHER_TYPE_ECPRI)
                return rte_flow_error_set(error, EINVAL,
                                          RTE_FLOW_ERROR_TYPE_ITEM, item,
-                                         "eCPRI cannot follow L2/VLAN layer "
-                                         "which ether type is not 0xAEFE.");
+                                         "eCPRI cannot follow L2/VLAN layer which ether type is not 0xAEFE");
        if (item_flags & MLX5_FLOW_LAYER_TUNNEL)
                return rte_flow_error_set(error, EINVAL,
                                          RTE_FLOW_ERROR_TYPE_ITEM, item,
-                                         "eCPRI with tunnel is not supported "
-                                         "right now.");
+                                         "eCPRI with tunnel is not supported right now");
        if (item_flags & MLX5_FLOW_LAYER_OUTER_L3)
                return rte_flow_error_set(error, ENOTSUP,
                                          RTE_FLOW_ERROR_TYPE_ITEM, item,
@@ -2888,13 +2811,12 @@ mlx5_flow_validate_item_ecpri(const struct rte_flow_item *item,
        else if (item_flags & MLX5_FLOW_LAYER_OUTER_L4_TCP)
                return rte_flow_error_set(error, EINVAL,
                                          RTE_FLOW_ERROR_TYPE_ITEM, item,
-                                         "eCPRI cannot follow a TCP layer.");
+                                         "eCPRI cannot coexist with a TCP layer");
        /* In specification, eCPRI could be over UDP layer. */
        else if (item_flags & MLX5_FLOW_LAYER_OUTER_L4_UDP)
                return rte_flow_error_set(error, EINVAL,
                                          RTE_FLOW_ERROR_TYPE_ITEM, item,
-                                         "eCPRI over UDP layer is not yet "
-                                         "supported right now.");
+                                         "eCPRI over UDP layer is not yet supported right now");
        /* Mask for type field in common header could be zero. */
        if (!mask)
                mask = &rte_flow_item_ecpri_mask;
@@ -2903,13 +2825,11 @@ mlx5_flow_validate_item_ecpri(const struct rte_flow_item *item,
        if (mask_lo.hdr.common.type != 0 && mask_lo.hdr.common.type != 0xff)
                return rte_flow_error_set(error, EINVAL,
                                          RTE_FLOW_ERROR_TYPE_ITEM_MASK, mask,
-                                         "partial mask is not supported "
-                                         "for protocol");
+                                         "partial mask is not supported for protocol");
        else if (mask_lo.hdr.common.type == 0 && mask->hdr.dummy[0] != 0)
                return rte_flow_error_set(error, EINVAL,
                                          RTE_FLOW_ERROR_TYPE_ITEM_MASK, mask,
-                                         "message header mask must be after "
-                                         "a type mask");
+                                         "message header mask must be after a type mask");
        return mlx5_flow_item_acceptable(item, (const uint8_t *)mask,
                                         acc_mask ? (const uint8_t *)acc_mask
                                                  : (const uint8_t *)&nic_mask,
@@ -3266,6 +3186,29 @@ flow_get_rss_action(const struct rte_flow_action actions[])
        return NULL;
 }
 
+/**
+ * Get ASO age action by index.
+ *
+ * @param[in] dev
+ *   Pointer to the Ethernet device structure.
+ * @param[in] age_idx
+ *   Index to the ASO age action.
+ *
+ * @return
+ *   The specified ASO age action.
+ */
+struct mlx5_aso_age_action*
+flow_aso_age_get_by_idx(struct rte_eth_dev *dev, uint32_t age_idx)
+{
+       uint16_t pool_idx = age_idx & UINT16_MAX;
+       uint16_t offset = (age_idx >> 16) & UINT16_MAX;
+       struct mlx5_priv *priv = dev->data->dev_private;
+       struct mlx5_aso_age_mng *mng = priv->sh->aso_age_mng;
+       struct mlx5_aso_age_pool *pool = mng->pools[pool_idx];
+
+       return &pool->actions[offset - 1];
+}
+
 /* maps shared action to translated non shared in some actions array */
 struct mlx5_translated_shared_action {
        struct rte_flow_shared_action *action; /**< Shared action */
@@ -3353,6 +3296,16 @@ flow_shared_actions_translate(struct rte_eth_dev *dev,
                        translated[shared->index].conf =
                                &shared_rss->origin;
                        break;
+               case MLX5_SHARED_ACTION_TYPE_AGE:
+                       if (priv->sh->flow_hit_aso_en) {
+                               translated[shared->index].type =
+                                       (enum rte_flow_action_type)
+                                       MLX5_RTE_FLOW_ACTION_TYPE_AGE;
+                               translated[shared->index].conf =
+                                                        (void *)(uintptr_t)idx;
+                               break;
+                       }
+                       /* Fall-through */
                default:
                        mlx5_free(translated);
                        return rte_flow_error_set
@@ -3620,9 +3573,7 @@ flow_check_hairpin_split(struct rte_eth_dev *dev,
                        break;
                case RTE_FLOW_ACTION_TYPE_RAW_ENCAP:
                        raw_encap = actions->conf;
-                       if (raw_encap->size >
-                           (sizeof(struct rte_flow_item_eth) +
-                            sizeof(struct rte_flow_item_ipv4)))
+                       if (raw_encap->size > MLX5_ENCAPSULATION_DECISION_SIZE)
                                split++;
                        action_n++;
                        break;
@@ -3648,6 +3599,17 @@ static void
 flow_list_destroy(struct rte_eth_dev *dev, uint32_t *list,
                  uint32_t flow_idx);
 
+int
+flow_dv_mreg_match_cb(struct mlx5_hlist *list __rte_unused,
+                     struct mlx5_hlist_entry *entry,
+                     uint64_t key, void *cb_ctx __rte_unused)
+{
+       struct mlx5_flow_mreg_copy_resource *mcp_res =
+               container_of(entry, typeof(*mcp_res), hlist_ent);
+
+       return mcp_res->mark_id != key;
+}
+
 struct mlx5_hlist_entry *
 flow_dv_mreg_create_cb(struct mlx5_hlist *list, uint64_t key,
                       void *cb_ctx)
@@ -3749,6 +3711,7 @@ flow_dv_mreg_create_cb(struct mlx5_hlist *list, uint64_t key,
                return NULL;
        }
        mcp_res->idx = idx;
+       mcp_res->mark_id = mark_id;
        /*
         * The copy Flows are not included in any list. There
         * ones are referenced from other Flows and can not
@@ -4057,9 +4020,7 @@ flow_hairpin_split(struct rte_eth_dev *dev,
                        break;
                case RTE_FLOW_ACTION_TYPE_RAW_ENCAP:
                        raw_encap = actions->conf;
-                       if (raw_encap->size >
-                           (sizeof(struct rte_flow_item_eth) +
-                            sizeof(struct rte_flow_item_ipv4))) {
+                       if (raw_encap->size > MLX5_ENCAPSULATION_DECISION_SIZE) {
                                memcpy(actions_tx, actions,
                                       sizeof(struct rte_flow_action));
                                actions_tx++;
@@ -4072,9 +4033,7 @@ flow_hairpin_split(struct rte_eth_dev *dev,
                        break;
                case RTE_FLOW_ACTION_TYPE_RAW_DECAP:
                        raw_decap = actions->conf;
-                       if (raw_decap->size <
-                           (sizeof(struct rte_flow_item_eth) +
-                            sizeof(struct rte_flow_item_ipv4))) {
+                       if (raw_decap->size < MLX5_ENCAPSULATION_DECISION_SIZE) {
                                memcpy(actions_tx, actions,
                                       sizeof(struct rte_flow_action));
                                actions_tx++;
@@ -4125,191 +4084,48 @@ flow_hairpin_split(struct rte_eth_dev *dev,
        return 0;
 }
 
-__extension__
-union tunnel_offload_mark {
-       uint32_t val;
-       struct {
-               uint32_t app_reserve:8;
-               uint32_t table_id:15;
-               uint32_t transfer:1;
-               uint32_t _unused_:8;
-       };
-};
-
-struct tunnel_default_miss_ctx {
-       uint16_t *queue;
-       __extension__
-       union {
-               struct rte_flow_action_rss action_rss;
-               struct rte_flow_action_queue miss_queue;
-               struct rte_flow_action_jump miss_jump;
-               uint8_t raw[0];
-       };
-};
-
+/**
+ * The last stage of splitting chain, just creates the subflow
+ * without any modification.
+ *
+ * @param[in] dev
+ *   Pointer to Ethernet device.
+ * @param[in] flow
+ *   Parent flow structure pointer.
+ * @param[in, out] sub_flow
+ *   Pointer to return the created subflow, may be NULL.
+ * @param[in] attr
+ *   Flow rule attributes.
+ * @param[in] items
+ *   Pattern specification (list terminated by the END pattern item).
+ * @param[in] actions
+ *   Associated actions (list terminated by the END action).
+ * @param[in] flow_split_info
+ *   Pointer to flow split info structure.
+ * @param[out] error
+ *   Perform verbose error reporting if not NULL.
+ * @return
+ *   0 on success, negative value otherwise
+ */
 static int
-flow_tunnel_add_default_miss(struct rte_eth_dev *dev,
-                            struct rte_flow *flow,
-                            const struct rte_flow_attr *attr,
-                            const struct rte_flow_action *app_actions,
-                            uint32_t flow_idx,
-                            struct tunnel_default_miss_ctx *ctx,
-                            struct rte_flow_error *error)
-{
-       struct mlx5_priv *priv = dev->data->dev_private;
-       struct mlx5_flow *dev_flow;
-       struct rte_flow_attr miss_attr = *attr;
-       const struct mlx5_flow_tunnel *tunnel = app_actions[0].conf;
-       const struct rte_flow_item miss_items[2] = {
-               {
-                       .type = RTE_FLOW_ITEM_TYPE_ETH,
-                       .spec = NULL,
-                       .last = NULL,
-                       .mask = NULL
-               },
-               {
-                       .type = RTE_FLOW_ITEM_TYPE_END,
-                       .spec = NULL,
-                       .last = NULL,
-                       .mask = NULL
-               }
-       };
-       union tunnel_offload_mark mark_id;
-       struct rte_flow_action_mark miss_mark;
-       struct rte_flow_action miss_actions[3] = {
-               [0] = { .type = RTE_FLOW_ACTION_TYPE_MARK, .conf = &miss_mark },
-               [2] = { .type = RTE_FLOW_ACTION_TYPE_END,  .conf = NULL }
-       };
-       const struct rte_flow_action_jump *jump_data;
-       uint32_t i, flow_table = 0; /* prevent compilation warning */
-       struct flow_grp_info grp_info = {
-               .external = 1,
-               .transfer = attr->transfer,
-               .fdb_def_rule = !!priv->fdb_def_rule,
-               .std_tbl_fix = 0,
-       };
-       int ret;
-
-       if (!attr->transfer) {
-               uint32_t q_size;
-
-               miss_actions[1].type = RTE_FLOW_ACTION_TYPE_RSS;
-               q_size = priv->reta_idx_n * sizeof(ctx->queue[0]);
-               ctx->queue = mlx5_malloc(MLX5_MEM_SYS | MLX5_MEM_ZERO, q_size,
-                                        0, SOCKET_ID_ANY);
-               if (!ctx->queue)
-                       return rte_flow_error_set
-                               (error, ENOMEM,
-                               RTE_FLOW_ERROR_TYPE_ACTION_CONF,
-                               NULL, "invalid default miss RSS");
-               ctx->action_rss.func = RTE_ETH_HASH_FUNCTION_DEFAULT,
-               ctx->action_rss.level = 0,
-               ctx->action_rss.types = priv->rss_conf.rss_hf,
-               ctx->action_rss.key_len = priv->rss_conf.rss_key_len,
-               ctx->action_rss.queue_num = priv->reta_idx_n,
-               ctx->action_rss.key = priv->rss_conf.rss_key,
-               ctx->action_rss.queue = ctx->queue;
-               if (!priv->reta_idx_n || !priv->rxqs_n)
-                       return rte_flow_error_set
-                               (error, EINVAL,
-                               RTE_FLOW_ERROR_TYPE_ACTION_CONF,
-                               NULL, "invalid port configuration");
-               if (!(dev->data->dev_conf.rxmode.mq_mode & ETH_MQ_RX_RSS_FLAG))
-                       ctx->action_rss.types = 0;
-               for (i = 0; i != priv->reta_idx_n; ++i)
-                       ctx->queue[i] = (*priv->reta_idx)[i];
-       } else {
-               miss_actions[1].type = RTE_FLOW_ACTION_TYPE_JUMP;
-               ctx->miss_jump.group = MLX5_TNL_MISS_FDB_JUMP_GRP;
-       }
-       miss_actions[1].conf = (typeof(miss_actions[1].conf))ctx->raw;
-       for (; app_actions->type != RTE_FLOW_ACTION_TYPE_JUMP; app_actions++);
-       jump_data = app_actions->conf;
-       miss_attr.priority = MLX5_TNL_MISS_RULE_PRIORITY;
-       miss_attr.group = jump_data->group;
-       ret = mlx5_flow_group_to_table(dev, tunnel, jump_data->group,
-                                      &flow_table, grp_info, error);
-       if (ret)
-               return rte_flow_error_set(error, EINVAL,
-                                         RTE_FLOW_ERROR_TYPE_ACTION_CONF,
-                                         NULL, "invalid tunnel id");
-       mark_id.app_reserve = 0;
-       mark_id.table_id = tunnel_flow_tbl_to_id(flow_table);
-       mark_id.transfer = !!attr->transfer;
-       mark_id._unused_ = 0;
-       miss_mark.id = mark_id.val;
-       dev_flow = flow_drv_prepare(dev, flow, &miss_attr,
-                                   miss_items, miss_actions, flow_idx, error);
-       if (!dev_flow)
-               return -rte_errno;
-       dev_flow->flow = flow;
-       dev_flow->external = true;
-       dev_flow->tunnel = tunnel;
-       /* Subflow object was created, we must include one in the list. */
-       SILIST_INSERT(&flow->dev_handles, dev_flow->handle_idx,
-                     dev_flow->handle, next);
-       DRV_LOG(DEBUG,
-               "port %u tunnel type=%d id=%u miss rule priority=%u group=%u",
-               dev->data->port_id, tunnel->app_tunnel.type,
-               tunnel->tunnel_id, miss_attr.priority, miss_attr.group);
-       ret = flow_drv_translate(dev, dev_flow, &miss_attr, miss_items,
-                                 miss_actions, error);
-       if (!ret)
-               ret = flow_mreg_update_copy_table(dev, flow, miss_actions,
-                                                 error);
-
-       return ret;
-}
-
-/**
- * The last stage of splitting chain, just creates the subflow
- * without any modification.
- *
- * @param[in] dev
- *   Pointer to Ethernet device.
- * @param[in] flow
- *   Parent flow structure pointer.
- * @param[in, out] sub_flow
- *   Pointer to return the created subflow, may be NULL.
- * @param[in] prefix_layers
- *   Prefix subflow layers, may be 0.
- * @param[in] prefix_mark
- *   Prefix subflow mark flag, may be 0.
- * @param[in] attr
- *   Flow rule attributes.
- * @param[in] items
- *   Pattern specification (list terminated by the END pattern item).
- * @param[in] actions
- *   Associated actions (list terminated by the END action).
- * @param[in] external
- *   This flow rule is created by request external to PMD.
- * @param[in] flow_idx
- *   This memory pool index to the flow.
- * @param[out] error
- *   Perform verbose error reporting if not NULL.
- * @return
- *   0 on success, negative value otherwise
- */
-static int
-flow_create_split_inner(struct rte_eth_dev *dev,
-                       struct rte_flow *flow,
-                       struct mlx5_flow **sub_flow,
-                       uint64_t prefix_layers,
-                       uint32_t prefix_mark,
-                       const struct rte_flow_attr *attr,
-                       const struct rte_flow_item items[],
-                       const struct rte_flow_action actions[],
-                       bool external, uint32_t flow_idx,
-                       struct rte_flow_error *error)
+flow_create_split_inner(struct rte_eth_dev *dev,
+                       struct rte_flow *flow,
+                       struct mlx5_flow **sub_flow,
+                       const struct rte_flow_attr *attr,
+                       const struct rte_flow_item items[],
+                       const struct rte_flow_action actions[],
+                       struct mlx5_flow_split_info *flow_split_info,
+                       struct rte_flow_error *error)
 {
        struct mlx5_flow *dev_flow;
 
        dev_flow = flow_drv_prepare(dev, flow, attr, items, actions,
-               flow_idx, error);
+                                   flow_split_info->flow_idx, error);
        if (!dev_flow)
                return -rte_errno;
        dev_flow->flow = flow;
-       dev_flow->external = external;
+       dev_flow->external = flow_split_info->external;
+       dev_flow->skip_scale = flow_split_info->skip_scale;
        /* Subflow object was created, we must include one in the list. */
        SILIST_INSERT(&flow->dev_handles, dev_flow->handle_idx,
                      dev_flow->handle, next);
@@ -4318,9 +4134,9 @@ flow_create_split_inner(struct rte_eth_dev *dev,
         * flow may need some user defined item layer flags, and pass the
         * Metadate rxq mark flag to suffix flow as well.
         */
-       if (prefix_layers)
-               dev_flow->handle->layers = prefix_layers;
-       if (prefix_mark)
+       if (flow_split_info->prefix_layers)
+               dev_flow->handle->layers = flow_split_info->prefix_layers;
+       if (flow_split_info->prefix_mark)
                dev_flow->handle->mark = 1;
        if (sub_flow)
                *sub_flow = dev_flow;
@@ -4874,20 +4690,14 @@ flow_sample_split_prep(struct rte_eth_dev *dev,
  *   Pointer to Ethernet device.
  * @param[in] flow
  *   Parent flow structure pointer.
- * @param[in] prefix_layers
- *   Prefix flow layer flags.
- * @param[in] prefix_mark
- *   Prefix subflow mark flag, may be 0.
  * @param[in] attr
  *   Flow rule attributes.
  * @param[in] items
  *   Pattern specification (list terminated by the END pattern item).
  * @param[in] actions
  *   Associated actions (list terminated by the END action).
- * @param[in] external
- *   This flow rule is created by request external to PMD.
- * @param[in] flow_idx
- *   This memory pool index to the flow.
+ * @param[in] flow_split_info
+ *   Pointer to flow split info structure.
  * @param[out] error
  *   Perform verbose error reporting if not NULL.
  * @return
@@ -4896,12 +4706,10 @@ flow_sample_split_prep(struct rte_eth_dev *dev,
 static int
 flow_create_split_metadata(struct rte_eth_dev *dev,
                           struct rte_flow *flow,
-                          uint64_t prefix_layers,
-                          uint32_t prefix_mark,
                           const struct rte_flow_attr *attr,
                           const struct rte_flow_item items[],
                           const struct rte_flow_action actions[],
-                          bool external, uint32_t flow_idx,
+                          struct mlx5_flow_split_info *flow_split_info,
                           struct rte_flow_error *error)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
@@ -4920,10 +4728,8 @@ flow_create_split_metadata(struct rte_eth_dev *dev,
        if (!config->dv_flow_en ||
            config->dv_xmeta_en == MLX5_XMETA_MODE_LEGACY ||
            !mlx5_flow_ext_mreg_supported(dev))
-               return flow_create_split_inner(dev, flow, NULL, prefix_layers,
-                                              prefix_mark, attr, items,
-                                              actions, external, flow_idx,
-                                              error);
+               return flow_create_split_inner(dev, flow, NULL, attr, items,
+                                              actions, flow_split_info, error);
        actions_n = flow_parse_metadata_split_actions_info(actions, &qrss,
                                                           &encap_idx);
        if (qrss) {
@@ -5008,10 +4814,9 @@ flow_create_split_metadata(struct rte_eth_dev *dev,
                        goto exit;
        }
        /* Add the unmodified original or prefix subflow. */
-       ret = flow_create_split_inner(dev, flow, &dev_flow, prefix_layers,
-                                     prefix_mark, attr,
+       ret = flow_create_split_inner(dev, flow, &dev_flow, attr,
                                      items, ext_actions ? ext_actions :
-                                     actions, external, flow_idx, error);
+                                     actions, flow_split_info, error);
        if (ret < 0)
                goto exit;
        MLX5_ASSERT(dev_flow);
@@ -5072,10 +4877,12 @@ flow_create_split_metadata(struct rte_eth_dev *dev,
                }
                dev_flow = NULL;
                /* Add suffix subflow to execute Q/RSS. */
-               ret = flow_create_split_inner(dev, flow, &dev_flow, layers, 0,
+               flow_split_info->prefix_layers = layers;
+               flow_split_info->prefix_mark = 0;
+               ret = flow_create_split_inner(dev, flow, &dev_flow,
                                              &q_attr, mtr_sfx ? items :
                                              q_items, q_actions,
-                                             external, flow_idx, error);
+                                             flow_split_info, error);
                if (ret < 0)
                        goto exit;
                /* qrss ID should be freed if failed. */
@@ -5109,20 +4916,14 @@ exit:
  *   Pointer to Ethernet device.
  * @param[in] flow
  *   Parent flow structure pointer.
- * @param[in] prefix_layers
- *   Prefix subflow layers, may be 0.
- * @param[in] prefix_mark
- *   Prefix subflow mark flag, may be 0.
  * @param[in] attr
  *   Flow rule attributes.
  * @param[in] items
  *   Pattern specification (list terminated by the END pattern item).
  * @param[in] actions
  *   Associated actions (list terminated by the END action).
- * @param[in] external
- *   This flow rule is created by request external to PMD.
- * @param[in] flow_idx
- *   This memory pool index to the flow.
+ * @param[in] flow_split_info
+ *   Pointer to flow split info structure.
  * @param[out] error
  *   Perform verbose error reporting if not NULL.
  * @return
@@ -5131,12 +4932,10 @@ exit:
 static int
 flow_create_split_meter(struct rte_eth_dev *dev,
                        struct rte_flow *flow,
-                       uint64_t prefix_layers,
-                       uint32_t prefix_mark,
                        const struct rte_flow_attr *attr,
                        const struct rte_flow_item items[],
                        const struct rte_flow_action actions[],
-                       bool external, uint32_t flow_idx,
+                       struct mlx5_flow_split_info *flow_split_info,
                        struct rte_flow_error *error)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
@@ -5180,11 +4979,10 @@ flow_create_split_meter(struct rte_eth_dev *dev,
                        goto exit;
                }
                /* Add the prefix subflow. */
+               flow_split_info->prefix_mark = 0;
                ret = flow_create_split_inner(dev, flow, &dev_flow,
-                                             prefix_layers, 0,
-                                             attr, items,
-                                             pre_actions, external,
-                                             flow_idx, error);
+                                             attr, items, pre_actions,
+                                             flow_split_info, error);
                if (ret) {
                        ret = -rte_errno;
                        goto exit;
@@ -5194,16 +4992,16 @@ flow_create_split_meter(struct rte_eth_dev *dev,
                sfx_attr.group = sfx_attr.transfer ?
                                (MLX5_FLOW_TABLE_LEVEL_SUFFIX - 1) :
                                 MLX5_FLOW_TABLE_LEVEL_SUFFIX;
+               flow_split_info->prefix_layers =
+                               flow_get_prefix_layer_flags(dev_flow);
+               flow_split_info->prefix_mark = dev_flow->handle->mark;
        }
        /* Add the prefix subflow. */
-       ret = flow_create_split_metadata(dev, flow, dev_flow ?
-                                        flow_get_prefix_layer_flags(dev_flow) :
-                                        prefix_layers, dev_flow ?
-                                        dev_flow->handle->mark : prefix_mark,
+       ret = flow_create_split_metadata(dev, flow,
                                         &sfx_attr, sfx_items ?
                                         sfx_items : items,
                                         sfx_actions ? sfx_actions : actions,
-                                        external, flow_idx, error);
+                                        flow_split_info, error);
 exit:
        if (sfx_actions)
                mlx5_free(sfx_actions);
@@ -5235,10 +5033,8 @@ exit:
  *   Pattern specification (list terminated by the END pattern item).
  * @param[in] actions
  *   Associated actions (list terminated by the END action).
- * @param[in] external
- *   This flow rule is created by request external to PMD.
- * @param[in] flow_idx
- *   This memory pool index to the flow.
+ * @param[in] flow_split_info
+ *   Pointer to flow split info structure.
  * @param[out] error
  *   Perform verbose error reporting if not NULL.
  * @return
@@ -5250,7 +5046,7 @@ flow_create_split_sample(struct rte_eth_dev *dev,
                         const struct rte_flow_attr *attr,
                         const struct rte_flow_item items[],
                         const struct rte_flow_action actions[],
-                        bool external, uint32_t flow_idx,
+                        struct mlx5_flow_split_info *flow_split_info,
                         struct rte_flow_error *error)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
@@ -5263,7 +5059,6 @@ flow_create_split_sample(struct rte_eth_dev *dev,
        struct mlx5_flow_dv_sample_resource *sample_res;
        struct mlx5_flow_tbl_data_entry *sfx_tbl_data;
        struct mlx5_flow_tbl_resource *sfx_tbl;
-       union mlx5_flow_tbl_key sfx_table_key;
 #endif
        size_t act_size;
        size_t item_size;
@@ -5307,9 +5102,9 @@ flow_create_split_sample(struct rte_eth_dev *dev,
                        goto exit;
                }
                /* Add the prefix subflow. */
-               ret = flow_create_split_inner(dev, flow, &dev_flow, 0, 0, attr,
-                                             items, pre_actions, external,
-                                             flow_idx, error);
+               ret = flow_create_split_inner(dev, flow, &dev_flow, attr,
+                                             items, pre_actions,
+                                             flow_split_info, error);
                if (ret) {
                        ret = -rte_errno;
                        goto exit;
@@ -5323,19 +5118,23 @@ flow_create_split_sample(struct rte_eth_dev *dev,
                                        sample_res->normal_path_tbl;
                sfx_tbl_data = container_of(sfx_tbl,
                                        struct mlx5_flow_tbl_data_entry, tbl);
-               sfx_table_key.v64 = sfx_tbl_data->entry.key;
                sfx_attr.group = sfx_attr.transfer ?
-                                       (sfx_table_key.table_id - 1) :
-                                        sfx_table_key.table_id;
+                                       (sfx_tbl_data->table_id - 1) :
+                                        sfx_tbl_data->table_id;
+               flow_split_info->prefix_layers =
+                               flow_get_prefix_layer_flags(dev_flow);
+               flow_split_info->prefix_mark = dev_flow->handle->mark;
+               /* Suffix group level already be scaled with factor, set
+                * skip_scale to 1 to avoid scale again in translation.
+                */
+               flow_split_info->skip_scale = 1;
 #endif
        }
        /* Add the suffix subflow. */
-       ret = flow_create_split_meter(dev, flow, dev_flow ?
-                                flow_get_prefix_layer_flags(dev_flow) : 0,
-                                dev_flow ? dev_flow->handle->mark : 0,
-                                &sfx_attr, sfx_items ? sfx_items : items,
-                                sfx_actions ? sfx_actions : actions,
-                                external, flow_idx, error);
+       ret = flow_create_split_meter(dev, flow, &sfx_attr,
+                                     sfx_items ? sfx_items : items,
+                                     sfx_actions ? sfx_actions : actions,
+                                     flow_split_info, error);
 exit:
        if (sfx_actions)
                mlx5_free(sfx_actions);
@@ -5370,10 +5169,8 @@ exit:
  *   Pattern specification (list terminated by the END pattern item).
  * @param[in] actions
  *   Associated actions (list terminated by the END action).
- * @param[in] external
- *   This flow rule is created by request external to PMD.
- * @param[in] flow_idx
- *   This memory pool index to the flow.
+ * @param[in] flow_split_info
+ *   Pointer to flow split info structure.
  * @param[out] error
  *   Perform verbose error reporting if not NULL.
  * @return
@@ -5385,13 +5182,13 @@ flow_create_split_outer(struct rte_eth_dev *dev,
                        const struct rte_flow_attr *attr,
                        const struct rte_flow_item items[],
                        const struct rte_flow_action actions[],
-                       bool external, uint32_t flow_idx,
+                       struct mlx5_flow_split_info *flow_split_info,
                        struct rte_flow_error *error)
 {
        int ret;
 
        ret = flow_create_split_sample(dev, flow, attr, items,
-                                      actions, external, flow_idx, error);
+                                      actions, flow_split_info, error);
        MLX5_ASSERT(ret <= 0);
        return ret;
 }
@@ -5435,17 +5232,15 @@ flow_rss_workspace_adjust(struct mlx5_flow_workspace *wks,
                          struct mlx5_flow_rss_desc *rss_desc,
                          uint32_t nrssq_num)
 {
-       bool fidx = !!wks->flow_idx;
-
-       if (likely(nrssq_num <= wks->rssq_num[fidx]))
+       if (likely(nrssq_num <= wks->rssq_num))
                return 0;
        rss_desc->queue = realloc(rss_desc->queue,
-                         sizeof(rss_desc->queue[0]) * RTE_ALIGN(nrssq_num, 2));
+                         sizeof(*rss_desc->queue) * RTE_ALIGN(nrssq_num, 2));
        if (!rss_desc->queue) {
                rte_errno = ENOMEM;
                return -1;
        }
-       wks->rssq_num[fidx] = RTE_ALIGN(nrssq_num, 2);
+       wks->rssq_num = RTE_ALIGN(nrssq_num, 2);
        return 0;
 }
 
@@ -5510,17 +5305,22 @@ flow_list_create(struct rte_eth_dev *dev, uint32_t *list,
        uint32_t idx = 0;
        int hairpin_flow;
        struct rte_flow_attr attr_tx = { .priority = 0 };
-       struct rte_flow_attr attr_factor = {0};
        const struct rte_flow_action *actions;
        struct rte_flow_action *translated_actions = NULL;
        struct mlx5_flow_tunnel *tunnel;
        struct tunnel_default_miss_ctx default_miss_ctx = { 0, };
-       struct mlx5_flow_workspace *wks = mlx5_flow_get_thread_workspace();
-       bool fidx = !!wks->flow_idx;
+       struct mlx5_flow_workspace *wks = mlx5_flow_push_thread_workspace();
+       struct mlx5_flow_split_info flow_split_info = {
+               .external = !!external,
+               .skip_scale = 0,
+               .flow_idx = 0,
+               .prefix_mark = 0,
+               .prefix_layers = 0
+       };
        int ret;
 
        MLX5_ASSERT(wks);
-       rss_desc = &wks->rss_desc[fidx];
+       rss_desc = &wks->rss_desc;
        ret = flow_shared_actions_translate(dev, original_actions,
                                            shared_actions,
                                            &shared_actions_n,
@@ -5530,10 +5330,9 @@ flow_list_create(struct rte_eth_dev *dev, uint32_t *list,
                return 0;
        }
        actions = translated_actions ? translated_actions : original_actions;
-       memcpy((void *)&attr_factor, (const void *)attr, sizeof(*attr));
        p_actions_rx = actions;
-       hairpin_flow = flow_check_hairpin_split(dev, &attr_factor, actions);
-       ret = flow_drv_validate(dev, &attr_factor, items, p_actions_rx,
+       hairpin_flow = flow_check_hairpin_split(dev, attr, actions);
+       ret = flow_drv_validate(dev, attr, items, p_actions_rx,
                                external, hairpin_flow, error);
        if (ret < 0)
                goto error_before_hairpin_split;
@@ -5552,7 +5351,8 @@ flow_list_create(struct rte_eth_dev *dev, uint32_t *list,
                                   idx);
                p_actions_rx = actions_rx.actions;
        }
-       flow->drv_type = flow_get_drv_type(dev, &attr_factor);
+       flow_split_info.flow_idx = idx;
+       flow->drv_type = flow_get_drv_type(dev, attr);
        MLX5_ASSERT(flow->drv_type > MLX5_FLOW_TYPE_MIN &&
                    flow->drv_type < MLX5_FLOW_TYPE_MAX);
        memset(rss_desc, 0, offsetof(struct mlx5_flow_rss_desc, queue));
@@ -5582,26 +5382,21 @@ flow_list_create(struct rte_eth_dev *dev, uint32_t *list,
                buf->entries = 1;
                buf->entry[0].pattern = (void *)(uintptr_t)items;
        }
-       flow->shared_rss = flow_get_shared_rss_action(dev, shared_actions,
+       rss_desc->shared_rss = flow_get_shared_rss_action(dev, shared_actions,
                                                      shared_actions_n);
-       /*
-        * Record the start index when there is a nested call. All sub-flows
-        * need to be translated before another calling.
-        * No need to use ping-pong buffer to save memory here.
-        */
-       if (fidx) {
-               MLX5_ASSERT(!wks->flow_nested_idx);
-               wks->flow_nested_idx = fidx;
-       }
        for (i = 0; i < buf->entries; ++i) {
+               /* Initialize flow split data. */
+               flow_split_info.prefix_layers = 0;
+               flow_split_info.prefix_mark = 0;
+               flow_split_info.skip_scale = 0;
                /*
                 * The splitter may create multiple dev_flows,
                 * depending on configuration. In the simplest
                 * case it just creates unmodified original flow.
                 */
-               ret = flow_create_split_outer(dev, flow, &attr_factor,
+               ret = flow_create_split_outer(dev, flow, attr,
                                              buf->entry[i].pattern,
-                                             p_actions_rx, external, idx,
+                                             p_actions_rx, &flow_split_info,
                                              error);
                if (ret < 0)
                        goto error;
@@ -5649,17 +5444,19 @@ flow_list_create(struct rte_eth_dev *dev, uint32_t *list,
         * the egress Flows belong to the different device and
         * copy table should be updated in peer NIC Rx domain.
         */
-       if (attr_factor.ingress &&
-           (external || attr_factor.group != MLX5_FLOW_MREG_CP_TABLE_GROUP)) {
+       if (attr->ingress &&
+           (external || attr->group != MLX5_FLOW_MREG_CP_TABLE_GROUP)) {
                ret = flow_mreg_update_copy_table(dev, flow, actions, error);
                if (ret)
                        goto error;
        }
        /*
-        * If the flow is external (from application) OR device is started, then
-        * the flow will be applied immediately.
+        * If the flow is external (from application) OR device is started,
+        * OR mreg discover, then apply immediately.
         */
-       if (external || dev->data->dev_started) {
+       if (external || dev->data->dev_started ||
+           (attr->group == MLX5_FLOW_MREG_CP_TABLE_GROUP &&
+            attr->priority == MLX5_FLOW_PRIO_RSVD)) {
                ret = flow_drv_apply(dev, flow, error);
                if (ret < 0)
                        goto error;
@@ -5672,10 +5469,6 @@ flow_list_create(struct rte_eth_dev *dev, uint32_t *list,
        }
        flow_rxq_flags_set(dev, flow);
        rte_free(translated_actions);
-       /* Nested flow creation index recovery. */
-       wks->flow_idx = wks->flow_nested_idx;
-       if (wks->flow_nested_idx)
-               wks->flow_nested_idx = 0;
        tunnel = flow_tunnel_from_rule(dev, attr, items, actions);
        if (tunnel) {
                flow->tunnel = 1;
@@ -5683,19 +5476,23 @@ flow_list_create(struct rte_eth_dev *dev, uint32_t *list,
                __atomic_add_fetch(&tunnel->refctn, 1, __ATOMIC_RELAXED);
                mlx5_free(default_miss_ctx.queue);
        }
+       mlx5_flow_pop_thread_workspace();
        return idx;
 error:
        MLX5_ASSERT(flow);
        ret = rte_errno; /* Save rte_errno before cleanup. */
        flow_mreg_del_copy_action(dev, flow);
        flow_drv_destroy(dev, flow);
+       if (rss_desc->shared_rss)
+               __atomic_sub_fetch(&((struct mlx5_shared_action_rss *)
+                       mlx5_ipool_get
+                       (priv->sh->ipool[MLX5_IPOOL_RSS_SHARED_ACTIONS],
+                       rss_desc->shared_rss))->refcnt, 1, __ATOMIC_RELAXED);
        mlx5_ipool_free(priv->sh->ipool[MLX5_IPOOL_RTE_FLOW], idx);
        rte_errno = ret; /* Restore rte_errno. */
        ret = rte_errno;
        rte_errno = ret;
-       wks->flow_idx = wks->flow_nested_idx;
-       if (wks->flow_nested_idx)
-               wks->flow_nested_idx = 0;
+       mlx5_flow_pop_thread_workspace();
 error_before_hairpin_split:
        rte_free(translated_actions);
        return 0;
@@ -5855,19 +5652,16 @@ flow_list_destroy(struct rte_eth_dev *dev, uint32_t *list,
                             flow_idx, flow, next);
                rte_spinlock_unlock(&priv->flow_list_lock);
        }
-       flow_mreg_del_copy_action(dev, flow);
-       mlx5_ipool_free(priv->sh->ipool[MLX5_IPOOL_RTE_FLOW], flow_idx);
        if (flow->tunnel) {
                struct mlx5_flow_tunnel *tunnel;
 
-               rte_spinlock_lock(&mlx5_tunnel_hub(dev)->sl);
                tunnel = mlx5_find_tunnel_id(dev, flow->tunnel_id);
                RTE_VERIFY(tunnel);
-               LIST_REMOVE(tunnel, chain);
-               rte_spinlock_unlock(&mlx5_tunnel_hub(dev)->sl);
                if (!__atomic_sub_fetch(&tunnel->refctn, 1, __ATOMIC_RELAXED))
                        mlx5_flow_tunnel_free(dev, tunnel);
        }
+       flow_mreg_del_copy_action(dev, flow);
+       mlx5_ipool_free(priv->sh->ipool[MLX5_IPOOL_RTE_FLOW], flow_idx);
 }
 
 /**
@@ -5932,12 +5726,14 @@ static void
 flow_release_workspace(void *data)
 {
        struct mlx5_flow_workspace *wks = data;
+       struct mlx5_flow_workspace *next;
 
-       if (!wks)
-               return;
-       free(wks->rss_desc[0].queue);
-       free(wks->rss_desc[1].queue);
-       free(wks);
+       while (wks) {
+               next = wks->next;
+               free(wks->rss_desc.queue);
+               free(wks);
+               wks = next;
+       }
 }
 
 /**
@@ -5951,52 +5747,116 @@ flow_alloc_workspace(void)
 }
 
 /**
- * Get thread specific flow workspace.
+ * Get thread specific current flow workspace.
  *
- * @return pointer to thread specific flowworkspace data, NULL on error.
+ * @return pointer to thread specific flow workspace data, NULL on error.
  */
 struct mlx5_flow_workspace*
 mlx5_flow_get_thread_workspace(void)
 {
        struct mlx5_flow_workspace *data;
 
-       if (pthread_once(&key_workspace_init, flow_alloc_workspace)) {
-               DRV_LOG(ERR, "Failed to init flow workspace data thread key.");
-               return NULL;
-       }
        data = pthread_getspecific(key_workspace);
+       MLX5_ASSERT(data && data->inuse);
+       if (!data || !data->inuse)
+               DRV_LOG(ERR, "flow workspace not initialized.");
+       return data;
+}
+
+/**
+ * Allocate and init new flow workspace.
+ *
+ * @return pointer to flow workspace data, NULL on error.
+ */
+static struct mlx5_flow_workspace*
+flow_alloc_thread_workspace(void)
+{
+       struct mlx5_flow_workspace *data = calloc(1, sizeof(*data));
+
        if (!data) {
-               data = calloc(1, sizeof(*data));
-               if (!data) {
-                       DRV_LOG(ERR, "Failed to allocate flow workspace "
-                               "memory.");
-                       return NULL;
-               }
-               data->rss_desc[0].queue = calloc(1,
-                               sizeof(uint16_t) * MLX5_RSSQ_DEFAULT_NUM);
-               if (!data->rss_desc[0].queue)
-                       goto err;
-               data->rss_desc[1].queue = calloc(1,
-                               sizeof(uint16_t) * MLX5_RSSQ_DEFAULT_NUM);
-               if (!data->rss_desc[1].queue)
-                       goto err;
-               data->rssq_num[0] = MLX5_RSSQ_DEFAULT_NUM;
-               data->rssq_num[1] = MLX5_RSSQ_DEFAULT_NUM;
-               if (pthread_setspecific(key_workspace, data)) {
-                       DRV_LOG(ERR, "Failed to set flow workspace to thread.");
-                       goto err;
-               }
+               DRV_LOG(ERR, "Failed to allocate flow workspace "
+                       "memory.");
+               return NULL;
        }
+       data->rss_desc.queue = calloc(1,
+                       sizeof(uint16_t) * MLX5_RSSQ_DEFAULT_NUM);
+       if (!data->rss_desc.queue)
+               goto err;
+       data->rssq_num = MLX5_RSSQ_DEFAULT_NUM;
        return data;
 err:
-       if (data->rss_desc[0].queue)
-               free(data->rss_desc[0].queue);
-       if (data->rss_desc[1].queue)
-               free(data->rss_desc[1].queue);
+       if (data->rss_desc.queue)
+               free(data->rss_desc.queue);
        free(data);
        return NULL;
 }
 
+/**
+ * Get new thread specific flow workspace.
+ *
+ * If current workspace inuse, create new one and set as current.
+ *
+ * @return pointer to thread specific flow workspace data, NULL on error.
+ */
+static struct mlx5_flow_workspace*
+mlx5_flow_push_thread_workspace(void)
+{
+       struct mlx5_flow_workspace *curr;
+       struct mlx5_flow_workspace *data;
+
+       if (pthread_once(&key_workspace_init, flow_alloc_workspace)) {
+               DRV_LOG(ERR, "Failed to init flow workspace data thread key.");
+               return NULL;
+       }
+       curr = pthread_getspecific(key_workspace);
+       if (!curr) {
+               data = flow_alloc_thread_workspace();
+               if (!data)
+                       return NULL;
+       } else if (!curr->inuse) {
+               data = curr;
+       } else if (curr->next) {
+               data = curr->next;
+       } else {
+               data = flow_alloc_thread_workspace();
+               if (!data)
+                       return NULL;
+               curr->next = data;
+               data->prev = curr;
+       }
+       data->inuse = 1;
+       data->flow_idx = 0;
+       /* Set as current workspace */
+       if (pthread_setspecific(key_workspace, data))
+               DRV_LOG(ERR, "Failed to set flow workspace to thread.");
+       return data;
+}
+
+/**
+ * Close current thread specific flow workspace.
+ *
+ * If previous workspace available, set it as current.
+ *
+ * @return pointer to thread specific flow workspace data, NULL on error.
+ */
+static void
+mlx5_flow_pop_thread_workspace(void)
+{
+       struct mlx5_flow_workspace *data = mlx5_flow_get_thread_workspace();
+
+       if (!data)
+               return;
+       if (!data->inuse) {
+               DRV_LOG(ERR, "Failed to close unused flow workspace.");
+               return;
+       }
+       data->inuse = 0;
+       if (!data->prev)
+               return;
+       if (pthread_setspecific(key_workspace, data->prev))
+               DRV_LOG(ERR, "Failed to set flow workspace to thread.");
+}
+
 /**
  * Verify the flow list is empty
  *
@@ -6299,9 +6159,9 @@ mlx5_flow_isolate(struct rte_eth_dev *dev,
        }
        priv->isolated = !!enable;
        if (enable)
-               dev->dev_ops = &mlx5_os_dev_ops_isolate;
+               dev->dev_ops = &mlx5_dev_ops_isolate;
        else
-               dev->dev_ops = &mlx5_os_dev_ops;
+               dev->dev_ops = &mlx5_dev_ops;
 
        dev->rx_descriptor_status = mlx5_rx_descriptor_status;
        dev->tx_descriptor_status = mlx5_tx_descriptor_status;
@@ -6610,7 +6470,7 @@ mlx5_flow_create_counter_stat_mem_mng(struct mlx5_dev_ctx_shared *sh)
        }
        mem_mng = (struct mlx5_counter_stats_mem_mng *)(mem + size) - 1;
        size = sizeof(*raw_data) * MLX5_COUNTERS_PER_POOL * raws_n;
-       mem_mng->umem = mlx5_glue->devx_umem_reg(sh->ctx, mem, size,
+       mem_mng->umem = mlx5_os_umem_reg(sh->ctx, mem, size,
                                                 IBV_ACCESS_LOCAL_WRITE);
        if (!mem_mng->umem) {
                rte_errno = errno;
@@ -6625,10 +6485,11 @@ mlx5_flow_create_counter_stat_mem_mng(struct mlx5_dev_ctx_shared *sh)
        mkey_attr.pg_access = 0;
        mkey_attr.klm_array = NULL;
        mkey_attr.klm_num = 0;
-       mkey_attr.relaxed_ordering = sh->cmng.relaxed_ordering;
+       mkey_attr.relaxed_ordering_write = sh->cmng.relaxed_ordering_write;
+       mkey_attr.relaxed_ordering_read = sh->cmng.relaxed_ordering_read;
        mem_mng->dm = mlx5_devx_cmd_mkey_create(sh->ctx, &mkey_attr);
        if (!mem_mng->dm) {
-               mlx5_glue->devx_umem_dereg(mem_mng->umem);
+               mlx5_os_umem_dereg(mem_mng->umem);
                rte_errno = errno;
                mlx5_free(mem);
                return -rte_errno;
@@ -6875,116 +6736,25 @@ mlx5_flow_async_pool_query_handle(struct mlx5_dev_ctx_shared *sh,
        sh->cmng.pending_queries--;
 }
 
-static const struct mlx5_flow_tbl_data_entry  *
-tunnel_mark_decode(struct rte_eth_dev *dev, uint32_t mark)
+static int
+flow_group_to_table(uint32_t port_id, uint32_t group, uint32_t *table,
+                   const struct flow_grp_info *grp_info,
+                   struct rte_flow_error *error)
 {
-       struct mlx5_priv *priv = dev->data->dev_private;
-       struct mlx5_dev_ctx_shared *sh = priv->sh;
-       struct mlx5_hlist_entry *he;
-       union tunnel_offload_mark mbits = { .val = mark };
-       union mlx5_flow_tbl_key table_key = {
-               {
-                       .table_id = tunnel_id_to_flow_tbl(mbits.table_id),
-                       .dummy = 0,
-                       .domain = !!mbits.transfer,
-                       .direction = 0,
-               }
-       };
-       he = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64, NULL);
-       return he ?
-              container_of(he, struct mlx5_flow_tbl_data_entry, entry) : NULL;
-}
-
-static void
-mlx5_flow_tunnel_grp2tbl_remove_cb(struct mlx5_hlist *list,
-                                  struct mlx5_hlist_entry *entry)
-{
-       struct mlx5_dev_ctx_shared *sh = list->ctx;
-       struct tunnel_tbl_entry *tte = container_of(entry, typeof(*tte), hash);
-
-       mlx5_ipool_free(sh->ipool[MLX5_IPOOL_TNL_TBL_ID],
-                       tunnel_flow_tbl_to_id(tte->flow_table));
-       mlx5_free(tte);
-}
-
-static struct mlx5_hlist_entry *
-mlx5_flow_tunnel_grp2tbl_create_cb(struct mlx5_hlist *list,
-                                  uint64_t key __rte_unused,
-                                  void *ctx __rte_unused)
-{
-       struct mlx5_dev_ctx_shared *sh = list->ctx;
-       struct tunnel_tbl_entry *tte;
-
-       tte = mlx5_malloc(MLX5_MEM_SYS | MLX5_MEM_ZERO,
-                         sizeof(*tte), 0,
-                         SOCKET_ID_ANY);
-       if (!tte)
-               goto err;
-       mlx5_ipool_malloc(sh->ipool[MLX5_IPOOL_TNL_TBL_ID],
-                         &tte->flow_table);
-       if (tte->flow_table >= MLX5_MAX_TABLES) {
-               DRV_LOG(ERR, "Tunnel TBL ID %d exceed max limit.",
-                       tte->flow_table);
-               mlx5_ipool_free(sh->ipool[MLX5_IPOOL_TNL_TBL_ID],
-                               tte->flow_table);
-               goto err;
-       } else if (!tte->flow_table) {
-               goto err;
-       }
-       tte->flow_table = tunnel_id_to_flow_tbl(tte->flow_table);
-       return &tte->hash;
-err:
-       if (tte)
-               mlx5_free(tte);
-       return NULL;
-}
-
-static uint32_t
-tunnel_flow_group_to_flow_table(struct rte_eth_dev *dev,
-                               const struct mlx5_flow_tunnel *tunnel,
-                               uint32_t group, uint32_t *table,
-                               struct rte_flow_error *error)
-{
-       struct mlx5_hlist_entry *he;
-       struct tunnel_tbl_entry *tte;
-       union tunnel_tbl_key key = {
-               .tunnel_id = tunnel ? tunnel->tunnel_id : 0,
-               .group = group
-       };
-       struct mlx5_flow_tunnel_hub *thub = mlx5_tunnel_hub(dev);
-       struct mlx5_hlist *group_hash;
-
-       group_hash = tunnel ? tunnel->groups : thub->groups;
-       he = mlx5_hlist_register(group_hash, key.val, NULL);
-       if (!he)
-               return rte_flow_error_set(error, EINVAL,
-                                         RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
-                                         NULL,
-                                         "tunnel group index not supported");
-       tte = container_of(he, typeof(*tte), hash);
-       *table = tte->flow_table;
-       DRV_LOG(DEBUG, "port %u tunnel %u group=%#x table=%#x",
-               dev->data->port_id, key.tunnel_id, group, *table);
-       return 0;
-}
-
-static int
-flow_group_to_table(uint32_t port_id, uint32_t group, uint32_t *table,
-                   struct flow_grp_info grp_info, struct rte_flow_error *error)
-{
-       if (grp_info.transfer && grp_info.external && grp_info.fdb_def_rule) {
-               if (group == UINT32_MAX)
-                       return rte_flow_error_set
-                                               (error, EINVAL,
-                                                RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
-                                                NULL,
-                                                "group index not supported");
-               *table = group + 1;
-       } else {
-               *table = group;
-       }
-       DRV_LOG(DEBUG, "port %u group=%#x table=%#x", port_id, group, *table);
-       return 0;
+       if (grp_info->transfer && grp_info->external &&
+           grp_info->fdb_def_rule) {
+               if (group == UINT32_MAX)
+                       return rte_flow_error_set
+                                               (error, EINVAL,
+                                                RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
+                                                NULL,
+                                                "group index not supported");
+               *table = group + 1;
+       } else {
+               *table = group;
+       }
+       DRV_LOG(DEBUG, "port %u group=%#x table=%#x", port_id, group, *table);
+       return 0;
 }
 
 /**
@@ -7029,24 +6799,25 @@ int
 mlx5_flow_group_to_table(struct rte_eth_dev *dev,
                         const struct mlx5_flow_tunnel *tunnel,
                         uint32_t group, uint32_t *table,
-                        struct flow_grp_info grp_info,
+                        const struct flow_grp_info *grp_info,
                         struct rte_flow_error *error)
 {
        int ret;
        bool standard_translation;
 
-       if (grp_info.external && group < MLX5_MAX_TABLES_EXTERNAL)
+       if (!grp_info->skip_scale && grp_info->external &&
+           group < MLX5_MAX_TABLES_EXTERNAL)
                group *= MLX5_FLOW_TABLE_FACTOR;
        if (is_tunnel_offload_active(dev)) {
-               standard_translation = !grp_info.external ||
-                                       grp_info.std_tbl_fix;
+               standard_translation = !grp_info->external ||
+                                       grp_info->std_tbl_fix;
        } else {
                standard_translation = true;
        }
        DRV_LOG(DEBUG,
-               "port %u group=%#x transfer=%d external=%d fdb_def_rule=%d translate=%s",
-               dev->data->port_id, group, grp_info.transfer,
-               grp_info.external, grp_info.fdb_def_rule,
+               "port %u group=%u transfer=%d external=%d fdb_def_rule=%d translate=%s",
+               dev->data->port_id, group, grp_info->transfer,
+               grp_info->external, grp_info->fdb_def_rule,
                standard_translation ? "STANDARD" : "TUNNEL");
        if (standard_translation)
                ret = flow_group_to_table(dev->data->port_id, group, table,
@@ -7124,8 +6895,7 @@ mlx5_flow_discover_mreg_c(struct rte_eth_dev *dev)
                                      flow_idx);
                if (!flow)
                        continue;
-               if (dev->data->dev_started || !flow_drv_apply(dev, flow, NULL))
-                       config->flow_mreg_c[n++] = idx;
+               config->flow_mreg_c[n++] = idx;
                flow_list_destroy(dev, NULL, flow_idx);
        }
        for (; n < MLX5_MREG_C_NUM; ++n)
@@ -7273,6 +7043,25 @@ flow_drv_action_update(struct rte_eth_dev *dev,
        return fops->action_update(dev, action, action_conf, error);
 }
 
+/* Wrapper for driver action_destroy op callback */
+static int
+flow_drv_action_query(struct rte_eth_dev *dev,
+                     const struct rte_flow_shared_action *action,
+                     void *data,
+                     const struct mlx5_flow_driver_ops *fops,
+                     struct rte_flow_error *error)
+{
+       static const char err_msg[] = "shared action query unsupported";
+
+       if (!fops->action_query) {
+               DRV_LOG(ERR, "port %u %s.", dev->data->port_id, err_msg);
+               rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
+                                  NULL, err_msg);
+               return -rte_errno;
+       }
+       return fops->action_query(dev, action, data, error);
+}
+
 /**
  * Create shared action for reuse in multiple flow rules.
  *
@@ -7375,11 +7164,11 @@ mlx5_shared_action_query(struct rte_eth_dev *dev,
                         void *data,
                         struct rte_flow_error *error)
 {
-       (void)dev;
-       (void)action;
-       (void)data;
-       return rte_flow_error_set(error, ENOTSUP, RTE_FLOW_ERROR_TYPE_ACTION,
-                                 NULL, "action type query not supported");
+       struct rte_flow_attr attr = { .transfer = 0 };
+       const struct mlx5_flow_driver_ops *fops =
+                       flow_get_drv_ops(flow_get_drv_type(dev, &attr));
+
+       return flow_drv_action_query(dev, action, data, fops, error);
 }
 
 /**
@@ -7408,172 +7197,6 @@ mlx5_shared_action_flush(struct rte_eth_dev *dev)
        return ret;
 }
 
-static void
-mlx5_flow_tunnel_free(struct rte_eth_dev *dev,
-                     struct mlx5_flow_tunnel *tunnel)
-{
-       struct mlx5_priv *priv = dev->data->dev_private;
-
-       DRV_LOG(DEBUG, "port %u release pmd tunnel id=0x%x",
-               dev->data->port_id, tunnel->tunnel_id);
-       RTE_VERIFY(!__atomic_load_n(&tunnel->refctn, __ATOMIC_RELAXED));
-       mlx5_ipool_free(priv->sh->ipool[MLX5_IPOOL_TUNNEL_ID],
-                       tunnel->tunnel_id);
-       mlx5_hlist_destroy(tunnel->groups);
-       mlx5_free(tunnel);
-}
-
-static struct mlx5_flow_tunnel *
-mlx5_find_tunnel_id(struct rte_eth_dev *dev, uint32_t id)
-{
-       struct mlx5_flow_tunnel_hub *thub = mlx5_tunnel_hub(dev);
-       struct mlx5_flow_tunnel *tun;
-
-       LIST_FOREACH(tun, &thub->tunnels, chain) {
-               if (tun->tunnel_id == id)
-                       break;
-       }
-
-       return tun;
-}
-
-static struct mlx5_flow_tunnel *
-mlx5_flow_tunnel_allocate(struct rte_eth_dev *dev,
-                         const struct rte_flow_tunnel *app_tunnel)
-{
-       struct mlx5_priv *priv = dev->data->dev_private;
-       struct mlx5_flow_tunnel *tunnel;
-       uint32_t id;
-
-       mlx5_ipool_malloc(priv->sh->ipool[MLX5_IPOOL_RSS_EXPANTION_FLOW_ID],
-                         &id);
-       if (id >= MLX5_MAX_TUNNELS) {
-               mlx5_ipool_free(priv->sh->ipool
-                               [MLX5_IPOOL_RSS_EXPANTION_FLOW_ID], id);
-               DRV_LOG(ERR, "Tunnel ID %d exceed max limit.", id);
-               return NULL;
-       } else if (!id) {
-               return NULL;
-       }
-       /**
-        * mlx5 flow tunnel is an auxlilary data structure
-        * It's not part of IO. No need to allocate it from
-        * huge pages pools dedicated for IO
-        */
-       tunnel = mlx5_malloc(MLX5_MEM_SYS | MLX5_MEM_ZERO, sizeof(*tunnel),
-                            0, SOCKET_ID_ANY);
-       if (!tunnel) {
-               mlx5_ipool_free(priv->sh->ipool
-                               [MLX5_IPOOL_RSS_EXPANTION_FLOW_ID], id);
-               return NULL;
-       }
-       tunnel->groups = mlx5_hlist_create("tunnel groups", 1024, 0, 0,
-                                          mlx5_flow_tunnel_grp2tbl_create_cb,
-                                          NULL,
-                                          mlx5_flow_tunnel_grp2tbl_remove_cb);
-       if (!tunnel->groups) {
-               mlx5_ipool_free(priv->sh->ipool
-                               [MLX5_IPOOL_RSS_EXPANTION_FLOW_ID], id);
-               mlx5_free(tunnel);
-               return NULL;
-       }
-       tunnel->groups->ctx = priv->sh;
-       /* initiate new PMD tunnel */
-       memcpy(&tunnel->app_tunnel, app_tunnel, sizeof(*app_tunnel));
-       tunnel->tunnel_id = id;
-       tunnel->action.type = (typeof(tunnel->action.type))
-                             MLX5_RTE_FLOW_ACTION_TYPE_TUNNEL_SET;
-       tunnel->action.conf = tunnel;
-       tunnel->item.type = (typeof(tunnel->item.type))
-                           MLX5_RTE_FLOW_ITEM_TYPE_TUNNEL;
-       tunnel->item.spec = tunnel;
-       tunnel->item.last = NULL;
-       tunnel->item.mask = NULL;
-
-       DRV_LOG(DEBUG, "port %u new pmd tunnel id=0x%x",
-               dev->data->port_id, tunnel->tunnel_id);
-
-       return tunnel;
-}
-
-static int
-mlx5_get_flow_tunnel(struct rte_eth_dev *dev,
-                    const struct rte_flow_tunnel *app_tunnel,
-                    struct mlx5_flow_tunnel **tunnel)
-{
-       int ret;
-       struct mlx5_flow_tunnel_hub *thub = mlx5_tunnel_hub(dev);
-       struct mlx5_flow_tunnel *tun;
-
-       rte_spinlock_lock(&thub->sl);
-       LIST_FOREACH(tun, &thub->tunnels, chain) {
-               if (!memcmp(app_tunnel, &tun->app_tunnel,
-                           sizeof(*app_tunnel))) {
-                       *tunnel = tun;
-                       ret = 0;
-                       break;
-               }
-       }
-       if (!tun) {
-               tun = mlx5_flow_tunnel_allocate(dev, app_tunnel);
-               if (tun) {
-                       LIST_INSERT_HEAD(&thub->tunnels, tun, chain);
-                       *tunnel = tun;
-               } else {
-                       ret = -ENOMEM;
-               }
-       }
-       rte_spinlock_unlock(&thub->sl);
-       if (tun)
-               __atomic_add_fetch(&tun->refctn, 1, __ATOMIC_RELAXED);
-
-       return ret;
-}
-
-void mlx5_release_tunnel_hub(struct mlx5_dev_ctx_shared *sh, uint16_t port_id)
-{
-       struct mlx5_flow_tunnel_hub *thub = sh->tunnel_hub;
-
-       if (!thub)
-               return;
-       if (!LIST_EMPTY(&thub->tunnels))
-               DRV_LOG(WARNING, "port %u tunnels present\n", port_id);
-       mlx5_hlist_destroy(thub->groups);
-       mlx5_free(thub);
-}
-
-int mlx5_alloc_tunnel_hub(struct mlx5_dev_ctx_shared *sh)
-{
-       int err;
-       struct mlx5_flow_tunnel_hub *thub;
-
-       thub = mlx5_malloc(MLX5_MEM_SYS | MLX5_MEM_ZERO, sizeof(*thub),
-                          0, SOCKET_ID_ANY);
-       if (!thub)
-               return -ENOMEM;
-       LIST_INIT(&thub->tunnels);
-       rte_spinlock_init(&thub->sl);
-       thub->groups = mlx5_hlist_create("flow groups", MLX5_MAX_TABLES, 0,
-                                        0, mlx5_flow_tunnel_grp2tbl_create_cb,
-                                        NULL,
-                                        mlx5_flow_tunnel_grp2tbl_remove_cb);
-       if (!thub->groups) {
-               err = -rte_errno;
-               goto err;
-       }
-       thub->groups->ctx = sh;
-       sh->tunnel_hub = thub;
-
-       return 0;
-
-err:
-       if (thub->groups)
-               mlx5_hlist_destroy(thub->groups);
-       if (thub)
-               mlx5_free(thub);
-       return err;
-}
-
 #ifndef HAVE_MLX5DV_DR
 #define MLX5_DOMAIN_SYNC_FLOW ((1 << 0) | (1 << 1))
 #else
@@ -7594,3 +7217,773 @@ int rte_pmd_mlx5_sync_flow(uint16_t port_id, uint32_t domains)
                ret = -ret;
        return ret;
 }
+
+/**
+ * tunnel offload functionalilty is defined for DV environment only
+ */
+#ifdef HAVE_IBV_FLOW_DV_SUPPORT
+__extension__
+union tunnel_offload_mark {
+       uint32_t val;
+       struct {
+               uint32_t app_reserve:8;
+               uint32_t table_id:15;
+               uint32_t transfer:1;
+               uint32_t _unused_:8;
+       };
+};
+
+static bool
+mlx5_access_tunnel_offload_db
+       (struct rte_eth_dev *dev,
+        bool (*match)(struct rte_eth_dev *,
+                      struct mlx5_flow_tunnel *, const void *),
+        void (*hit)(struct rte_eth_dev *, struct mlx5_flow_tunnel *, void *),
+        void (*miss)(struct rte_eth_dev *, void *),
+        void *ctx, bool lock_op);
+
+static int
+flow_tunnel_add_default_miss(struct rte_eth_dev *dev,
+                            struct rte_flow *flow,
+                            const struct rte_flow_attr *attr,
+                            const struct rte_flow_action *app_actions,
+                            uint32_t flow_idx,
+                            struct tunnel_default_miss_ctx *ctx,
+                            struct rte_flow_error *error)
+{
+       struct mlx5_priv *priv = dev->data->dev_private;
+       struct mlx5_flow *dev_flow;
+       struct rte_flow_attr miss_attr = *attr;
+       const struct mlx5_flow_tunnel *tunnel = app_actions[0].conf;
+       const struct rte_flow_item miss_items[2] = {
+               {
+                       .type = RTE_FLOW_ITEM_TYPE_ETH,
+                       .spec = NULL,
+                       .last = NULL,
+                       .mask = NULL
+               },
+               {
+                       .type = RTE_FLOW_ITEM_TYPE_END,
+                       .spec = NULL,
+                       .last = NULL,
+                       .mask = NULL
+               }
+       };
+       union tunnel_offload_mark mark_id;
+       struct rte_flow_action_mark miss_mark;
+       struct rte_flow_action miss_actions[3] = {
+               [0] = { .type = RTE_FLOW_ACTION_TYPE_MARK, .conf = &miss_mark },
+               [2] = { .type = RTE_FLOW_ACTION_TYPE_END,  .conf = NULL }
+       };
+       const struct rte_flow_action_jump *jump_data;
+       uint32_t i, flow_table = 0; /* prevent compilation warning */
+       struct flow_grp_info grp_info = {
+               .external = 1,
+               .transfer = attr->transfer,
+               .fdb_def_rule = !!priv->fdb_def_rule,
+               .std_tbl_fix = 0,
+       };
+       int ret;
+
+       if (!attr->transfer) {
+               uint32_t q_size;
+
+               miss_actions[1].type = RTE_FLOW_ACTION_TYPE_RSS;
+               q_size = priv->reta_idx_n * sizeof(ctx->queue[0]);
+               ctx->queue = mlx5_malloc(MLX5_MEM_SYS | MLX5_MEM_ZERO, q_size,
+                                        0, SOCKET_ID_ANY);
+               if (!ctx->queue)
+                       return rte_flow_error_set
+                               (error, ENOMEM,
+                               RTE_FLOW_ERROR_TYPE_ACTION_CONF,
+                               NULL, "invalid default miss RSS");
+               ctx->action_rss.func = RTE_ETH_HASH_FUNCTION_DEFAULT,
+               ctx->action_rss.level = 0,
+               ctx->action_rss.types = priv->rss_conf.rss_hf,
+               ctx->action_rss.key_len = priv->rss_conf.rss_key_len,
+               ctx->action_rss.queue_num = priv->reta_idx_n,
+               ctx->action_rss.key = priv->rss_conf.rss_key,
+               ctx->action_rss.queue = ctx->queue;
+               if (!priv->reta_idx_n || !priv->rxqs_n)
+                       return rte_flow_error_set
+                               (error, EINVAL,
+                               RTE_FLOW_ERROR_TYPE_ACTION_CONF,
+                               NULL, "invalid port configuration");
+               if (!(dev->data->dev_conf.rxmode.mq_mode & ETH_MQ_RX_RSS_FLAG))
+                       ctx->action_rss.types = 0;
+               for (i = 0; i != priv->reta_idx_n; ++i)
+                       ctx->queue[i] = (*priv->reta_idx)[i];
+       } else {
+               miss_actions[1].type = RTE_FLOW_ACTION_TYPE_JUMP;
+               ctx->miss_jump.group = MLX5_TNL_MISS_FDB_JUMP_GRP;
+       }
+       miss_actions[1].conf = (typeof(miss_actions[1].conf))ctx->raw;
+       for (; app_actions->type != RTE_FLOW_ACTION_TYPE_JUMP; app_actions++);
+       jump_data = app_actions->conf;
+       miss_attr.priority = MLX5_TNL_MISS_RULE_PRIORITY;
+       miss_attr.group = jump_data->group;
+       ret = mlx5_flow_group_to_table(dev, tunnel, jump_data->group,
+                                      &flow_table, &grp_info, error);
+       if (ret)
+               return rte_flow_error_set(error, EINVAL,
+                                         RTE_FLOW_ERROR_TYPE_ACTION_CONF,
+                                         NULL, "invalid tunnel id");
+       mark_id.app_reserve = 0;
+       mark_id.table_id = tunnel_flow_tbl_to_id(flow_table);
+       mark_id.transfer = !!attr->transfer;
+       mark_id._unused_ = 0;
+       miss_mark.id = mark_id.val;
+       dev_flow = flow_drv_prepare(dev, flow, &miss_attr,
+                                   miss_items, miss_actions, flow_idx, error);
+       if (!dev_flow)
+               return -rte_errno;
+       dev_flow->flow = flow;
+       dev_flow->external = true;
+       dev_flow->tunnel = tunnel;
+       /* Subflow object was created, we must include one in the list. */
+       SILIST_INSERT(&flow->dev_handles, dev_flow->handle_idx,
+                     dev_flow->handle, next);
+       DRV_LOG(DEBUG,
+               "port %u tunnel type=%d id=%u miss rule priority=%u group=%u",
+               dev->data->port_id, tunnel->app_tunnel.type,
+               tunnel->tunnel_id, miss_attr.priority, miss_attr.group);
+       ret = flow_drv_translate(dev, dev_flow, &miss_attr, miss_items,
+                                 miss_actions, error);
+       if (!ret)
+               ret = flow_mreg_update_copy_table(dev, flow, miss_actions,
+                                                 error);
+
+       return ret;
+}
+
+static const struct mlx5_flow_tbl_data_entry  *
+tunnel_mark_decode(struct rte_eth_dev *dev, uint32_t mark)
+{
+       struct mlx5_priv *priv = dev->data->dev_private;
+       struct mlx5_dev_ctx_shared *sh = priv->sh;
+       struct mlx5_hlist_entry *he;
+       union tunnel_offload_mark mbits = { .val = mark };
+       union mlx5_flow_tbl_key table_key = {
+               {
+                       .table_id = tunnel_id_to_flow_tbl(mbits.table_id),
+                       .dummy = 0,
+                       .domain = !!mbits.transfer,
+                       .direction = 0,
+               }
+       };
+       he = mlx5_hlist_lookup(sh->flow_tbls, table_key.v64, NULL);
+       return he ?
+              container_of(he, struct mlx5_flow_tbl_data_entry, entry) : NULL;
+}
+
+static void
+mlx5_flow_tunnel_grp2tbl_remove_cb(struct mlx5_hlist *list,
+                                  struct mlx5_hlist_entry *entry)
+{
+       struct mlx5_dev_ctx_shared *sh = list->ctx;
+       struct tunnel_tbl_entry *tte = container_of(entry, typeof(*tte), hash);
+
+       mlx5_ipool_free(sh->ipool[MLX5_IPOOL_TNL_TBL_ID],
+                       tunnel_flow_tbl_to_id(tte->flow_table));
+       mlx5_free(tte);
+}
+
+static int
+mlx5_flow_tunnel_grp2tbl_match_cb(struct mlx5_hlist *list __rte_unused,
+                                 struct mlx5_hlist_entry *entry,
+                                 uint64_t key, void *cb_ctx __rte_unused)
+{
+       union tunnel_tbl_key tbl = {
+               .val = key,
+       };
+       struct tunnel_tbl_entry *tte = container_of(entry, typeof(*tte), hash);
+
+       return tbl.tunnel_id != tte->tunnel_id || tbl.group != tte->group;
+}
+
+static struct mlx5_hlist_entry *
+mlx5_flow_tunnel_grp2tbl_create_cb(struct mlx5_hlist *list, uint64_t key,
+                                  void *ctx __rte_unused)
+{
+       struct mlx5_dev_ctx_shared *sh = list->ctx;
+       struct tunnel_tbl_entry *tte;
+       union tunnel_tbl_key tbl = {
+               .val = key,
+       };
+
+       tte = mlx5_malloc(MLX5_MEM_SYS | MLX5_MEM_ZERO,
+                         sizeof(*tte), 0,
+                         SOCKET_ID_ANY);
+       if (!tte)
+               goto err;
+       mlx5_ipool_malloc(sh->ipool[MLX5_IPOOL_TNL_TBL_ID],
+                         &tte->flow_table);
+       if (tte->flow_table >= MLX5_MAX_TABLES) {
+               DRV_LOG(ERR, "Tunnel TBL ID %d exceed max limit.",
+                       tte->flow_table);
+               mlx5_ipool_free(sh->ipool[MLX5_IPOOL_TNL_TBL_ID],
+                               tte->flow_table);
+               goto err;
+       } else if (!tte->flow_table) {
+               goto err;
+       }
+       tte->flow_table = tunnel_id_to_flow_tbl(tte->flow_table);
+       tte->tunnel_id = tbl.tunnel_id;
+       tte->group = tbl.group;
+       return &tte->hash;
+err:
+       if (tte)
+               mlx5_free(tte);
+       return NULL;
+}
+
+static uint32_t
+tunnel_flow_group_to_flow_table(struct rte_eth_dev *dev,
+                               const struct mlx5_flow_tunnel *tunnel,
+                               uint32_t group, uint32_t *table,
+                               struct rte_flow_error *error)
+{
+       struct mlx5_hlist_entry *he;
+       struct tunnel_tbl_entry *tte;
+       union tunnel_tbl_key key = {
+               .tunnel_id = tunnel ? tunnel->tunnel_id : 0,
+               .group = group
+       };
+       struct mlx5_flow_tunnel_hub *thub = mlx5_tunnel_hub(dev);
+       struct mlx5_hlist *group_hash;
+
+       group_hash = tunnel ? tunnel->groups : thub->groups;
+       he = mlx5_hlist_register(group_hash, key.val, NULL);
+       if (!he)
+               return rte_flow_error_set(error, EINVAL,
+                                         RTE_FLOW_ERROR_TYPE_ATTR_GROUP,
+                                         NULL,
+                                         "tunnel group index not supported");
+       tte = container_of(he, typeof(*tte), hash);
+       *table = tte->flow_table;
+       DRV_LOG(DEBUG, "port %u tunnel %u group=%#x table=%#x",
+               dev->data->port_id, key.tunnel_id, group, *table);
+       return 0;
+}
+
+static void
+mlx5_flow_tunnel_free(struct rte_eth_dev *dev,
+                     struct mlx5_flow_tunnel *tunnel)
+{
+       struct mlx5_priv *priv = dev->data->dev_private;
+       struct mlx5_indexed_pool *ipool;
+
+       DRV_LOG(DEBUG, "port %u release pmd tunnel id=0x%x",
+               dev->data->port_id, tunnel->tunnel_id);
+       LIST_REMOVE(tunnel, chain);
+       mlx5_hlist_destroy(tunnel->groups);
+       ipool = priv->sh->ipool[MLX5_IPOOL_TUNNEL_ID];
+       mlx5_ipool_free(ipool, tunnel->tunnel_id);
+}
+
+static bool
+mlx5_access_tunnel_offload_db
+       (struct rte_eth_dev *dev,
+        bool (*match)(struct rte_eth_dev *,
+                      struct mlx5_flow_tunnel *, const void *),
+        void (*hit)(struct rte_eth_dev *, struct mlx5_flow_tunnel *, void *),
+        void (*miss)(struct rte_eth_dev *, void *),
+        void *ctx, bool lock_op)
+{
+       bool verdict = false;
+       struct mlx5_flow_tunnel_hub *thub = mlx5_tunnel_hub(dev);
+       struct mlx5_flow_tunnel *tunnel;
+
+       rte_spinlock_lock(&thub->sl);
+       LIST_FOREACH(tunnel, &thub->tunnels, chain) {
+               verdict = match(dev, tunnel, (const void *)ctx);
+               if (verdict)
+                       break;
+       }
+       if (!lock_op)
+               rte_spinlock_unlock(&thub->sl);
+       if (verdict && hit)
+               hit(dev, tunnel, ctx);
+       if (!verdict && miss)
+               miss(dev, ctx);
+       if (lock_op)
+               rte_spinlock_unlock(&thub->sl);
+
+       return verdict;
+}
+
+struct tunnel_db_find_tunnel_id_ctx {
+       uint32_t tunnel_id;
+       struct mlx5_flow_tunnel *tunnel;
+};
+
+static bool
+find_tunnel_id_match(struct rte_eth_dev *dev,
+                    struct mlx5_flow_tunnel *tunnel, const void *x)
+{
+       const struct tunnel_db_find_tunnel_id_ctx *ctx = x;
+
+       RTE_SET_USED(dev);
+       return tunnel->tunnel_id == ctx->tunnel_id;
+}
+
+static void
+find_tunnel_id_hit(struct rte_eth_dev *dev,
+                  struct mlx5_flow_tunnel *tunnel, void *x)
+{
+       struct tunnel_db_find_tunnel_id_ctx *ctx = x;
+       RTE_SET_USED(dev);
+       ctx->tunnel = tunnel;
+}
+
+static struct mlx5_flow_tunnel *
+mlx5_find_tunnel_id(struct rte_eth_dev *dev, uint32_t id)
+{
+       struct tunnel_db_find_tunnel_id_ctx ctx = {
+               .tunnel_id = id,
+       };
+
+       mlx5_access_tunnel_offload_db(dev, find_tunnel_id_match,
+                                     find_tunnel_id_hit, NULL, &ctx, true);
+
+       return ctx.tunnel;
+}
+
+static struct mlx5_flow_tunnel *
+mlx5_flow_tunnel_allocate(struct rte_eth_dev *dev,
+                         const struct rte_flow_tunnel *app_tunnel)
+{
+       struct mlx5_priv *priv = dev->data->dev_private;
+       struct mlx5_indexed_pool *ipool;
+       struct mlx5_flow_tunnel *tunnel;
+       uint32_t id;
+
+       ipool = priv->sh->ipool[MLX5_IPOOL_TUNNEL_ID];
+       tunnel = mlx5_ipool_zmalloc(ipool, &id);
+       if (!tunnel)
+               return NULL;
+       if (id >= MLX5_MAX_TUNNELS) {
+               mlx5_ipool_free(ipool, id);
+               DRV_LOG(ERR, "Tunnel ID %d exceed max limit.", id);
+               return NULL;
+       }
+       tunnel->groups = mlx5_hlist_create("tunnel groups", 1024, 0, 0,
+                                          mlx5_flow_tunnel_grp2tbl_create_cb,
+                                          mlx5_flow_tunnel_grp2tbl_match_cb,
+                                          mlx5_flow_tunnel_grp2tbl_remove_cb);
+       if (!tunnel->groups) {
+               mlx5_ipool_free(ipool, id);
+               return NULL;
+       }
+       tunnel->groups->ctx = priv->sh;
+       /* initiate new PMD tunnel */
+       memcpy(&tunnel->app_tunnel, app_tunnel, sizeof(*app_tunnel));
+       tunnel->tunnel_id = id;
+       tunnel->action.type = (typeof(tunnel->action.type))
+                             MLX5_RTE_FLOW_ACTION_TYPE_TUNNEL_SET;
+       tunnel->action.conf = tunnel;
+       tunnel->item.type = (typeof(tunnel->item.type))
+                           MLX5_RTE_FLOW_ITEM_TYPE_TUNNEL;
+       tunnel->item.spec = tunnel;
+       tunnel->item.last = NULL;
+       tunnel->item.mask = NULL;
+
+       DRV_LOG(DEBUG, "port %u new pmd tunnel id=0x%x",
+               dev->data->port_id, tunnel->tunnel_id);
+
+       return tunnel;
+}
+
+struct tunnel_db_get_tunnel_ctx {
+       const struct rte_flow_tunnel *app_tunnel;
+       struct mlx5_flow_tunnel *tunnel;
+};
+
+static bool get_tunnel_match(struct rte_eth_dev *dev,
+                            struct mlx5_flow_tunnel *tunnel, const void *x)
+{
+       const struct tunnel_db_get_tunnel_ctx *ctx = x;
+
+       RTE_SET_USED(dev);
+       return !memcmp(ctx->app_tunnel, &tunnel->app_tunnel,
+                      sizeof(*ctx->app_tunnel));
+}
+
+static void get_tunnel_hit(struct rte_eth_dev *dev,
+                          struct mlx5_flow_tunnel *tunnel, void *x)
+{
+       /* called under tunnel spinlock protection */
+       struct tunnel_db_get_tunnel_ctx *ctx = x;
+
+       RTE_SET_USED(dev);
+       tunnel->refctn++;
+       ctx->tunnel = tunnel;
+}
+
+static void get_tunnel_miss(struct rte_eth_dev *dev, void *x)
+{
+       /* called under tunnel spinlock protection */
+       struct mlx5_flow_tunnel_hub *thub = mlx5_tunnel_hub(dev);
+       struct tunnel_db_get_tunnel_ctx *ctx = x;
+
+       rte_spinlock_unlock(&thub->sl);
+       ctx->tunnel = mlx5_flow_tunnel_allocate(dev, ctx->app_tunnel);
+       ctx->tunnel->refctn = 1;
+       rte_spinlock_lock(&thub->sl);
+       if (ctx->tunnel)
+               LIST_INSERT_HEAD(&thub->tunnels, ctx->tunnel, chain);
+}
+
+
+static int
+mlx5_get_flow_tunnel(struct rte_eth_dev *dev,
+                    const struct rte_flow_tunnel *app_tunnel,
+                    struct mlx5_flow_tunnel **tunnel)
+{
+       struct tunnel_db_get_tunnel_ctx ctx = {
+               .app_tunnel = app_tunnel,
+       };
+
+       mlx5_access_tunnel_offload_db(dev, get_tunnel_match, get_tunnel_hit,
+                                     get_tunnel_miss, &ctx, true);
+       *tunnel = ctx.tunnel;
+       return ctx.tunnel ? 0 : -ENOMEM;
+}
+
+void mlx5_release_tunnel_hub(struct mlx5_dev_ctx_shared *sh, uint16_t port_id)
+{
+       struct mlx5_flow_tunnel_hub *thub = sh->tunnel_hub;
+
+       if (!thub)
+               return;
+       if (!LIST_EMPTY(&thub->tunnels))
+               DRV_LOG(WARNING, "port %u tunnels present\n", port_id);
+       mlx5_hlist_destroy(thub->groups);
+       mlx5_free(thub);
+}
+
+int mlx5_alloc_tunnel_hub(struct mlx5_dev_ctx_shared *sh)
+{
+       int err;
+       struct mlx5_flow_tunnel_hub *thub;
+
+       thub = mlx5_malloc(MLX5_MEM_SYS | MLX5_MEM_ZERO, sizeof(*thub),
+                          0, SOCKET_ID_ANY);
+       if (!thub)
+               return -ENOMEM;
+       LIST_INIT(&thub->tunnels);
+       rte_spinlock_init(&thub->sl);
+       thub->groups = mlx5_hlist_create("flow groups", MLX5_MAX_TABLES, 0,
+                                        0, mlx5_flow_tunnel_grp2tbl_create_cb,
+                                        mlx5_flow_tunnel_grp2tbl_match_cb,
+                                        mlx5_flow_tunnel_grp2tbl_remove_cb);
+       if (!thub->groups) {
+               err = -rte_errno;
+               goto err;
+       }
+       thub->groups->ctx = sh;
+       sh->tunnel_hub = thub;
+
+       return 0;
+
+err:
+       if (thub->groups)
+               mlx5_hlist_destroy(thub->groups);
+       if (thub)
+               mlx5_free(thub);
+       return err;
+}
+
+static inline bool
+mlx5_flow_tunnel_validate(struct rte_eth_dev *dev,
+                         struct rte_flow_tunnel *tunnel,
+                         const char *err_msg)
+{
+       err_msg = NULL;
+       if (!is_tunnel_offload_active(dev)) {
+               err_msg = "tunnel offload was not activated";
+               goto out;
+       } else if (!tunnel) {
+               err_msg = "no application tunnel";
+               goto out;
+       }
+
+       switch (tunnel->type) {
+       default:
+               err_msg = "unsupported tunnel type";
+               goto out;
+       case RTE_FLOW_ITEM_TYPE_VXLAN:
+               break;
+       }
+
+out:
+       return !err_msg;
+}
+
+static int
+mlx5_flow_tunnel_decap_set(struct rte_eth_dev *dev,
+                   struct rte_flow_tunnel *app_tunnel,
+                   struct rte_flow_action **actions,
+                   uint32_t *num_of_actions,
+                   struct rte_flow_error *error)
+{
+       int ret;
+       struct mlx5_flow_tunnel *tunnel;
+       const char *err_msg = NULL;
+       bool verdict = mlx5_flow_tunnel_validate(dev, app_tunnel, err_msg);
+
+       if (!verdict)
+               return rte_flow_error_set(error, EINVAL,
+                                         RTE_FLOW_ERROR_TYPE_ACTION_CONF, NULL,
+                                         err_msg);
+       ret = mlx5_get_flow_tunnel(dev, app_tunnel, &tunnel);
+       if (ret < 0) {
+               return rte_flow_error_set(error, ret,
+                                         RTE_FLOW_ERROR_TYPE_ACTION_CONF, NULL,
+                                         "failed to initialize pmd tunnel");
+       }
+       *actions = &tunnel->action;
+       *num_of_actions = 1;
+       return 0;
+}
+
+static int
+mlx5_flow_tunnel_match(struct rte_eth_dev *dev,
+                      struct rte_flow_tunnel *app_tunnel,
+                      struct rte_flow_item **items,
+                      uint32_t *num_of_items,
+                      struct rte_flow_error *error)
+{
+       int ret;
+       struct mlx5_flow_tunnel *tunnel;
+       const char *err_msg = NULL;
+       bool verdict = mlx5_flow_tunnel_validate(dev, app_tunnel, err_msg);
+
+       if (!verdict)
+               return rte_flow_error_set(error, EINVAL,
+                                         RTE_FLOW_ERROR_TYPE_HANDLE, NULL,
+                                         err_msg);
+       ret = mlx5_get_flow_tunnel(dev, app_tunnel, &tunnel);
+       if (ret < 0) {
+               return rte_flow_error_set(error, ret,
+                                         RTE_FLOW_ERROR_TYPE_HANDLE, NULL,
+                                         "failed to initialize pmd tunnel");
+       }
+       *items = &tunnel->item;
+       *num_of_items = 1;
+       return 0;
+}
+
+struct tunnel_db_element_release_ctx {
+       struct rte_flow_item *items;
+       struct rte_flow_action *actions;
+       uint32_t num_elements;
+       struct rte_flow_error *error;
+       int ret;
+};
+
+static bool
+tunnel_element_release_match(struct rte_eth_dev *dev,
+                            struct mlx5_flow_tunnel *tunnel, const void *x)
+{
+       const struct tunnel_db_element_release_ctx *ctx = x;
+
+       RTE_SET_USED(dev);
+       if (ctx->num_elements != 1)
+               return false;
+       else if (ctx->items)
+               return ctx->items == &tunnel->item;
+       else if (ctx->actions)
+               return ctx->actions == &tunnel->action;
+
+       return false;
+}
+
+static void
+tunnel_element_release_hit(struct rte_eth_dev *dev,
+                          struct mlx5_flow_tunnel *tunnel, void *x)
+{
+       struct tunnel_db_element_release_ctx *ctx = x;
+       ctx->ret = 0;
+       if (!__atomic_sub_fetch(&tunnel->refctn, 1, __ATOMIC_RELAXED))
+               mlx5_flow_tunnel_free(dev, tunnel);
+}
+
+static void
+tunnel_element_release_miss(struct rte_eth_dev *dev, void *x)
+{
+       struct tunnel_db_element_release_ctx *ctx = x;
+       RTE_SET_USED(dev);
+       ctx->ret = rte_flow_error_set(ctx->error, EINVAL,
+                                     RTE_FLOW_ERROR_TYPE_HANDLE, NULL,
+                                     "invalid argument");
+}
+
+static int
+mlx5_flow_tunnel_item_release(struct rte_eth_dev *dev,
+                      struct rte_flow_item *pmd_items,
+                      uint32_t num_items, struct rte_flow_error *err)
+{
+       struct tunnel_db_element_release_ctx ctx = {
+               .items = pmd_items,
+               .actions = NULL,
+               .num_elements = num_items,
+               .error = err,
+       };
+
+       mlx5_access_tunnel_offload_db(dev, tunnel_element_release_match,
+                                     tunnel_element_release_hit,
+                                     tunnel_element_release_miss, &ctx, false);
+
+       return ctx.ret;
+}
+
+static int
+mlx5_flow_tunnel_action_release(struct rte_eth_dev *dev,
+                        struct rte_flow_action *pmd_actions,
+                        uint32_t num_actions, struct rte_flow_error *err)
+{
+       struct tunnel_db_element_release_ctx ctx = {
+               .items = NULL,
+               .actions = pmd_actions,
+               .num_elements = num_actions,
+               .error = err,
+       };
+
+       mlx5_access_tunnel_offload_db(dev, tunnel_element_release_match,
+                                     tunnel_element_release_hit,
+                                     tunnel_element_release_miss, &ctx, false);
+
+       return ctx.ret;
+}
+
+static int
+mlx5_flow_tunnel_get_restore_info(struct rte_eth_dev *dev,
+                                 struct rte_mbuf *m,
+                                 struct rte_flow_restore_info *info,
+                                 struct rte_flow_error *err)
+{
+       uint64_t ol_flags = m->ol_flags;
+       const struct mlx5_flow_tbl_data_entry *tble;
+       const uint64_t mask = PKT_RX_FDIR | PKT_RX_FDIR_ID;
+
+       if (!is_tunnel_offload_active(dev)) {
+               info->flags = 0;
+               return 0;
+       }
+
+       if ((ol_flags & mask) != mask)
+               goto err;
+       tble = tunnel_mark_decode(dev, m->hash.fdir.hi);
+       if (!tble) {
+               DRV_LOG(DEBUG, "port %u invalid miss tunnel mark %#x",
+                       dev->data->port_id, m->hash.fdir.hi);
+               goto err;
+       }
+       MLX5_ASSERT(tble->tunnel);
+       memcpy(&info->tunnel, &tble->tunnel->app_tunnel, sizeof(info->tunnel));
+       info->group_id = tble->group_id;
+       info->flags = RTE_FLOW_RESTORE_INFO_TUNNEL |
+                     RTE_FLOW_RESTORE_INFO_GROUP_ID |
+                     RTE_FLOW_RESTORE_INFO_ENCAPSULATED;
+
+       return 0;
+
+err:
+       return rte_flow_error_set(err, EINVAL,
+                                 RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
+                                 "failed to get restore info");
+}
+
+#else /* HAVE_IBV_FLOW_DV_SUPPORT */
+static int
+mlx5_flow_tunnel_decap_set(__rte_unused struct rte_eth_dev *dev,
+                          __rte_unused struct rte_flow_tunnel *app_tunnel,
+                          __rte_unused struct rte_flow_action **actions,
+                          __rte_unused uint32_t *num_of_actions,
+                          __rte_unused struct rte_flow_error *error)
+{
+       return -ENOTSUP;
+}
+
+static int
+mlx5_flow_tunnel_match(__rte_unused struct rte_eth_dev *dev,
+                      __rte_unused struct rte_flow_tunnel *app_tunnel,
+                      __rte_unused struct rte_flow_item **items,
+                      __rte_unused uint32_t *num_of_items,
+                      __rte_unused struct rte_flow_error *error)
+{
+       return -ENOTSUP;
+}
+
+static int
+mlx5_flow_tunnel_item_release(__rte_unused struct rte_eth_dev *dev,
+                             __rte_unused struct rte_flow_item *pmd_items,
+                             __rte_unused uint32_t num_items,
+                             __rte_unused struct rte_flow_error *err)
+{
+       return -ENOTSUP;
+}
+
+static int
+mlx5_flow_tunnel_action_release(__rte_unused struct rte_eth_dev *dev,
+                               __rte_unused struct rte_flow_action *pmd_action,
+                               __rte_unused uint32_t num_actions,
+                               __rte_unused struct rte_flow_error *err)
+{
+       return -ENOTSUP;
+}
+
+static int
+mlx5_flow_tunnel_get_restore_info(__rte_unused struct rte_eth_dev *dev,
+                                 __rte_unused struct rte_mbuf *m,
+                                 __rte_unused struct rte_flow_restore_info *i,
+                                 __rte_unused struct rte_flow_error *err)
+{
+       return -ENOTSUP;
+}
+
+static int
+flow_tunnel_add_default_miss(__rte_unused struct rte_eth_dev *dev,
+                            __rte_unused struct rte_flow *flow,
+                            __rte_unused const struct rte_flow_attr *attr,
+                            __rte_unused const struct rte_flow_action *actions,
+                            __rte_unused uint32_t flow_idx,
+                            __rte_unused struct tunnel_default_miss_ctx *ctx,
+                            __rte_unused struct rte_flow_error *error)
+{
+       return -ENOTSUP;
+}
+
+static struct mlx5_flow_tunnel *
+mlx5_find_tunnel_id(__rte_unused struct rte_eth_dev *dev,
+                   __rte_unused uint32_t id)
+{
+       return NULL;
+}
+
+static void
+mlx5_flow_tunnel_free(__rte_unused struct rte_eth_dev *dev,
+                     __rte_unused struct mlx5_flow_tunnel *tunnel)
+{
+}
+
+static uint32_t
+tunnel_flow_group_to_flow_table(__rte_unused struct rte_eth_dev *dev,
+                               __rte_unused const struct mlx5_flow_tunnel *t,
+                               __rte_unused uint32_t group,
+                               __rte_unused uint32_t *table,
+                               struct rte_flow_error *error)
+{
+       return rte_flow_error_set(error, ENOTSUP,
+                                 RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
+                                 "tunnel offload requires DV support");
+}
+
+void
+mlx5_release_tunnel_hub(__rte_unused struct mlx5_dev_ctx_shared *sh,
+                       __rte_unused  uint16_t port_id)
+{
+}
+#endif /* HAVE_IBV_FLOW_DV_SUPPORT */
+