net/mlx5: optimize flow RSS struct
[dpdk.git] / drivers / net / mlx5 / mlx5_flow.c
index aad9689..0565f6f 100644 (file)
@@ -658,13 +658,12 @@ mlx5_flow_item_acceptable(const struct rte_flow_item *item,
  *   The hash fields that should be used.
  */
 uint64_t
-mlx5_flow_hashfields_adjust(struct mlx5_flow *dev_flow,
+mlx5_flow_hashfields_adjust(struct mlx5_flow_rss_desc *rss_desc,
                            int tunnel __rte_unused, uint64_t layer_types,
                            uint64_t hash_fields)
 {
-       struct rte_flow *flow = dev_flow->flow;
 #ifdef HAVE_IBV_DEVICE_TUNNEL_SUPPORT
-       int rss_request_inner = flow->rss.level >= 2;
+       int rss_request_inner = rss_desc->level >= 2;
 
        /* Check RSS hash level for tunnel. */
        if (tunnel && rss_request_inner)
@@ -673,7 +672,7 @@ mlx5_flow_hashfields_adjust(struct mlx5_flow *dev_flow,
                return 0;
 #endif
        /* Check if requested layer matches RSS hash fields. */
-       if (!(flow->rss.types & layer_types))
+       if (!(rss_desc->types & layer_types))
                return 0;
        return hash_fields;
 }
@@ -712,21 +711,27 @@ flow_rxq_tunnel_ptype_update(struct mlx5_rxq_ctrl *rxq_ctrl)
  *
  * @param[in] dev
  *   Pointer to the Ethernet device structure.
- * @param[in] dev_flow
- *   Pointer to device flow structure.
+ * @param[in] dev_handle
+ *   Pointer to device flow handle structure.
  */
 static void
-flow_drv_rxq_flags_set(struct rte_eth_dev *dev, struct mlx5_flow *dev_flow)
+flow_drv_rxq_flags_set(struct rte_eth_dev *dev,
+                      struct mlx5_flow_handle *dev_handle)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
-       struct rte_flow *flow = dev_flow->flow;
-       const int mark = !!(dev_flow->actions &
-                           (MLX5_FLOW_ACTION_FLAG | MLX5_FLOW_ACTION_MARK));
-       const int tunnel = !!(dev_flow->layers & MLX5_FLOW_LAYER_TUNNEL);
+       const int mark = dev_handle->mark;
+       const int tunnel = !!(dev_handle->layers & MLX5_FLOW_LAYER_TUNNEL);
+       struct mlx5_hrxq *hrxq;
        unsigned int i;
 
-       for (i = 0; i != flow->rss.queue_num; ++i) {
-               int idx = (*flow->rss.queue)[i];
+       if (dev_handle->fate_action != MLX5_FLOW_FATE_QUEUE)
+               return;
+       hrxq = mlx5_ipool_get(priv->sh->ipool[MLX5_IPOOL_HRXQ],
+                             dev_handle->rix_hrxq);
+       if (!hrxq)
+               return;
+       for (i = 0; i != hrxq->ind_table->queues_n; ++i) {
+               int idx = hrxq->ind_table->queues[i];
                struct mlx5_rxq_ctrl *rxq_ctrl =
                        container_of((*priv->rxqs)[idx],
                                     struct mlx5_rxq_ctrl, rxq);
@@ -751,7 +756,7 @@ flow_drv_rxq_flags_set(struct rte_eth_dev *dev, struct mlx5_flow *dev_flow)
                        /* Increase the counter matching the flow. */
                        for (j = 0; j != MLX5_FLOW_TUNNEL; ++j) {
                                if ((tunnels_info[j].tunnel &
-                                    dev_flow->layers) ==
+                                    dev_handle->layers) ==
                                    tunnels_info[j].tunnel) {
                                        rxq_ctrl->flow_tunnels_n[j]++;
                                        break;
@@ -773,10 +778,13 @@ flow_drv_rxq_flags_set(struct rte_eth_dev *dev, struct mlx5_flow *dev_flow)
 static void
 flow_rxq_flags_set(struct rte_eth_dev *dev, struct rte_flow *flow)
 {
-       struct mlx5_flow *dev_flow;
+       struct mlx5_priv *priv = dev->data->dev_private;
+       uint32_t handle_idx;
+       struct mlx5_flow_handle *dev_handle;
 
-       LIST_FOREACH(dev_flow, &flow->dev_flows, next)
-               flow_drv_rxq_flags_set(dev, dev_flow);
+       SILIST_FOREACH(priv->sh->ipool[MLX5_IPOOL_MLX5_FLOW], flow->dev_handles,
+                      handle_idx, dev_handle, next)
+               flow_drv_rxq_flags_set(dev, dev_handle);
 }
 
 /**
@@ -785,22 +793,28 @@ flow_rxq_flags_set(struct rte_eth_dev *dev, struct rte_flow *flow)
  *
  * @param dev
  *   Pointer to Ethernet device.
- * @param[in] dev_flow
- *   Pointer to the device flow.
+ * @param[in] dev_handle
+ *   Pointer to the device flow handle structure.
  */
 static void
-flow_drv_rxq_flags_trim(struct rte_eth_dev *dev, struct mlx5_flow *dev_flow)
+flow_drv_rxq_flags_trim(struct rte_eth_dev *dev,
+                       struct mlx5_flow_handle *dev_handle)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
-       struct rte_flow *flow = dev_flow->flow;
-       const int mark = !!(dev_flow->actions &
-                           (MLX5_FLOW_ACTION_FLAG | MLX5_FLOW_ACTION_MARK));
-       const int tunnel = !!(dev_flow->layers & MLX5_FLOW_LAYER_TUNNEL);
+       const int mark = dev_handle->mark;
+       const int tunnel = !!(dev_handle->layers & MLX5_FLOW_LAYER_TUNNEL);
+       struct mlx5_hrxq *hrxq;
        unsigned int i;
 
+       if (dev_handle->fate_action != MLX5_FLOW_FATE_QUEUE)
+               return;
+       hrxq = mlx5_ipool_get(priv->sh->ipool[MLX5_IPOOL_HRXQ],
+                             dev_handle->rix_hrxq);
+       if (!hrxq)
+               return;
        MLX5_ASSERT(dev->data->dev_started);
-       for (i = 0; i != flow->rss.queue_num; ++i) {
-               int idx = (*flow->rss.queue)[i];
+       for (i = 0; i != hrxq->ind_table->queues_n; ++i) {
+               int idx = hrxq->ind_table->queues[i];
                struct mlx5_rxq_ctrl *rxq_ctrl =
                        container_of((*priv->rxqs)[idx],
                                     struct mlx5_rxq_ctrl, rxq);
@@ -820,7 +834,7 @@ flow_drv_rxq_flags_trim(struct rte_eth_dev *dev, struct mlx5_flow *dev_flow)
                        /* Decrease the counter matching the flow. */
                        for (j = 0; j != MLX5_FLOW_TUNNEL; ++j) {
                                if ((tunnels_info[j].tunnel &
-                                    dev_flow->layers) ==
+                                    dev_handle->layers) ==
                                    tunnels_info[j].tunnel) {
                                        rxq_ctrl->flow_tunnels_n[j]--;
                                        break;
@@ -843,10 +857,13 @@ flow_drv_rxq_flags_trim(struct rte_eth_dev *dev, struct mlx5_flow *dev_flow)
 static void
 flow_rxq_flags_trim(struct rte_eth_dev *dev, struct rte_flow *flow)
 {
-       struct mlx5_flow *dev_flow;
+       struct mlx5_priv *priv = dev->data->dev_private;
+       uint32_t handle_idx;
+       struct mlx5_flow_handle *dev_handle;
 
-       LIST_FOREACH(dev_flow, &flow->dev_flows, next)
-               flow_drv_rxq_flags_trim(dev, dev_flow);
+       SILIST_FOREACH(priv->sh->ipool[MLX5_IPOOL_MLX5_FLOW], flow->dev_handles,
+                      handle_idx, dev_handle, next)
+               flow_drv_rxq_flags_trim(dev, dev_handle);
 }
 
 /**
@@ -2309,11 +2326,14 @@ static void
 flow_mreg_split_qrss_release(struct rte_eth_dev *dev,
                             struct rte_flow *flow)
 {
-       struct mlx5_flow *dev_flow;
+       struct mlx5_priv *priv = dev->data->dev_private;
+       uint32_t handle_idx;
+       struct mlx5_flow_handle *dev_handle;
 
-       LIST_FOREACH(dev_flow, &flow->dev_flows, next)
-               if (dev_flow->qrss_id)
-                       flow_qrss_free_id(dev, dev_flow->qrss_id);
+       SILIST_FOREACH(priv->sh->ipool[MLX5_IPOOL_MLX5_FLOW], flow->dev_handles,
+                      handle_idx, dev_handle, next)
+               if (dev_handle->split_flow_id)
+                       flow_qrss_free_id(dev, dev_handle->split_flow_id);
 }
 
 static int
@@ -2329,7 +2349,8 @@ flow_null_validate(struct rte_eth_dev *dev __rte_unused,
 }
 
 static struct mlx5_flow *
-flow_null_prepare(const struct rte_flow_attr *attr __rte_unused,
+flow_null_prepare(struct rte_eth_dev *dev __rte_unused,
+                 const struct rte_flow_attr *attr __rte_unused,
                  const struct rte_flow_item items[] __rte_unused,
                  const struct rte_flow_action actions[] __rte_unused,
                  struct rte_flow_error *error)
@@ -2469,6 +2490,8 @@ flow_drv_validate(struct rte_eth_dev *dev,
  *   setting backward reference to the flow should be done out of this function.
  *   layers field is not filled either.
  *
+ * @param[in] dev
+ *   Pointer to the dev structure.
  * @param[in] attr
  *   Pointer to the flow attributes.
  * @param[in] items
@@ -2482,7 +2505,8 @@ flow_drv_validate(struct rte_eth_dev *dev,
  *   Pointer to device flow on success, otherwise NULL and rte_errno is set.
  */
 static inline struct mlx5_flow *
-flow_drv_prepare(const struct rte_flow *flow,
+flow_drv_prepare(struct rte_eth_dev *dev,
+                const struct rte_flow *flow,
                 const struct rte_flow_attr *attr,
                 const struct rte_flow_item items[],
                 const struct rte_flow_action actions[],
@@ -2493,7 +2517,7 @@ flow_drv_prepare(const struct rte_flow *flow,
 
        MLX5_ASSERT(type > MLX5_FLOW_TYPE_MIN && type < MLX5_FLOW_TYPE_MAX);
        fops = flow_get_drv_ops(type);
-       return fops->prepare(attr, items, actions, error);
+       return fops->prepare(dev, attr, items, actions, error);
 }
 
 /**
@@ -2696,18 +2720,22 @@ flow_get_prefix_layer_flags(struct mlx5_flow *dev_flow)
 {
        uint64_t layers = 0;
 
-       /* If no decap actions, use the layers directly. */
-       if (!(dev_flow->actions & MLX5_FLOW_ACTION_DECAP))
-               return dev_flow->layers;
+       /*
+        * Layers bits could be localization, but usually the compiler will
+        * help to do the optimization work for source code.
+        * If no decap actions, use the layers directly.
+        */
+       if (!(dev_flow->act_flags & MLX5_FLOW_ACTION_DECAP))
+               return dev_flow->handle->layers;
        /* Convert L3 layers with decap action. */
-       if (dev_flow->layers & MLX5_FLOW_LAYER_INNER_L3_IPV4)
+       if (dev_flow->handle->layers & MLX5_FLOW_LAYER_INNER_L3_IPV4)
                layers |= MLX5_FLOW_LAYER_OUTER_L3_IPV4;
-       else if (dev_flow->layers & MLX5_FLOW_LAYER_INNER_L3_IPV6)
+       else if (dev_flow->handle->layers & MLX5_FLOW_LAYER_INNER_L3_IPV6)
                layers |= MLX5_FLOW_LAYER_OUTER_L3_IPV6;
        /* Convert L4 layers with decap action.  */
-       if (dev_flow->layers & MLX5_FLOW_LAYER_INNER_L4_TCP)
+       if (dev_flow->handle->layers & MLX5_FLOW_LAYER_INNER_L4_TCP)
                layers |= MLX5_FLOW_LAYER_OUTER_L4_TCP;
-       else if (dev_flow->layers & MLX5_FLOW_LAYER_INNER_L4_UDP)
+       else if (dev_flow->handle->layers & MLX5_FLOW_LAYER_INNER_L4_UDP)
                layers |= MLX5_FLOW_LAYER_OUTER_L4_UDP;
        return layers;
 }
@@ -2943,6 +2971,7 @@ flow_mreg_add_copy_action(struct rte_eth_dev *dev, uint32_t mark_id,
                [3] = { .type = RTE_FLOW_ACTION_TYPE_END, },
        };
        struct mlx5_flow_mreg_copy_resource *mcp_res;
+       uint32_t idx = 0;
        int ret;
 
        /* Fill the register fileds in the flow. */
@@ -3011,11 +3040,12 @@ flow_mreg_add_copy_action(struct rte_eth_dev *dev, uint32_t mark_id,
                };
        }
        /* Build a new entry. */
-       mcp_res = rte_zmalloc(__func__, sizeof(*mcp_res), 0);
+       mcp_res = mlx5_ipool_zmalloc(priv->sh->ipool[MLX5_IPOOL_MCP], &idx);
        if (!mcp_res) {
                rte_errno = ENOMEM;
                return NULL;
        }
+       mcp_res->idx = idx;
        /*
         * The copy Flows are not included in any list. There
         * ones are referenced from other Flows and can not
@@ -3037,7 +3067,7 @@ flow_mreg_add_copy_action(struct rte_eth_dev *dev, uint32_t mark_id,
 error:
        if (mcp_res->flow)
                flow_list_destroy(dev, NULL, mcp_res->flow);
-       rte_free(mcp_res);
+       mlx5_ipool_free(priv->sh->ipool[MLX5_IPOOL_MCP], mcp_res->idx);
        return NULL;
 }
 
@@ -3053,9 +3083,13 @@ static void
 flow_mreg_del_copy_action(struct rte_eth_dev *dev,
                          struct rte_flow *flow)
 {
-       struct mlx5_flow_mreg_copy_resource *mcp_res = flow->mreg_copy;
+       struct mlx5_flow_mreg_copy_resource *mcp_res;
        struct mlx5_priv *priv = dev->data->dev_private;
 
+       if (!flow->rix_mreg_copy)
+               return;
+       mcp_res = mlx5_ipool_get(priv->sh->ipool[MLX5_IPOOL_MCP],
+                                flow->rix_mreg_copy);
        if (!mcp_res || !priv->mreg_cp_tbl)
                return;
        if (flow->copy_applied) {
@@ -3074,8 +3108,8 @@ flow_mreg_del_copy_action(struct rte_eth_dev *dev,
        MLX5_ASSERT(mcp_res->flow);
        flow_list_destroy(dev, NULL, mcp_res->flow);
        mlx5_hlist_remove(priv->mreg_cp_tbl, &mcp_res->hlist_ent);
-       rte_free(mcp_res);
-       flow->mreg_copy = NULL;
+       mlx5_ipool_free(priv->sh->ipool[MLX5_IPOOL_MCP], mcp_res->idx);
+       flow->rix_mreg_copy = 0;
 }
 
 /**
@@ -3093,10 +3127,15 @@ static int
 flow_mreg_start_copy_action(struct rte_eth_dev *dev,
                            struct rte_flow *flow)
 {
-       struct mlx5_flow_mreg_copy_resource *mcp_res = flow->mreg_copy;
+       struct mlx5_flow_mreg_copy_resource *mcp_res;
+       struct mlx5_priv *priv = dev->data->dev_private;
        int ret;
 
-       if (!mcp_res || flow->copy_applied)
+       if (!flow->rix_mreg_copy || flow->copy_applied)
+               return 0;
+       mcp_res = mlx5_ipool_get(priv->sh->ipool[MLX5_IPOOL_MCP],
+                                flow->rix_mreg_copy);
+       if (!mcp_res)
                return 0;
        if (!mcp_res->appcnt) {
                ret = flow_drv_apply(dev, mcp_res->flow, NULL);
@@ -3120,9 +3159,14 @@ static void
 flow_mreg_stop_copy_action(struct rte_eth_dev *dev,
                           struct rte_flow *flow)
 {
-       struct mlx5_flow_mreg_copy_resource *mcp_res = flow->mreg_copy;
+       struct mlx5_flow_mreg_copy_resource *mcp_res;
+       struct mlx5_priv *priv = dev->data->dev_private;
 
-       if (!mcp_res || !flow->copy_applied)
+       if (!flow->rix_mreg_copy || !flow->copy_applied)
+               return;
+       mcp_res = mlx5_ipool_get(priv->sh->ipool[MLX5_IPOOL_MCP],
+                                flow->rix_mreg_copy);
+       if (!mcp_res)
                return;
        MLX5_ASSERT(mcp_res->appcnt);
        --mcp_res->appcnt;
@@ -3153,7 +3197,7 @@ flow_mreg_del_default_copy_action(struct rte_eth_dev *dev)
        MLX5_ASSERT(mcp_res->flow);
        flow_list_destroy(dev, NULL, mcp_res->flow);
        mlx5_hlist_remove(priv->mreg_cp_tbl, &mcp_res->hlist_ent);
-       rte_free(mcp_res);
+       mlx5_ipool_free(priv->sh->ipool[MLX5_IPOOL_MCP], mcp_res->idx);
 }
 
 /**
@@ -3245,7 +3289,7 @@ flow_mreg_update_copy_table(struct rte_eth_dev *dev,
                                (dev, MLX5_FLOW_MARK_DEFAULT, error);
                        if (!mcp_res)
                                return -rte_errno;
-                       flow->mreg_copy = mcp_res;
+                       flow->rix_mreg_copy = mcp_res->idx;
                        if (dev->data->dev_started) {
                                mcp_res->appcnt++;
                                flow->copy_applied = 1;
@@ -3258,7 +3302,7 @@ flow_mreg_update_copy_table(struct rte_eth_dev *dev,
                                flow_mreg_add_copy_action(dev, mark->id, error);
                        if (!mcp_res)
                                return -rte_errno;
-                       flow->mreg_copy = mcp_res;
+                       flow->rix_mreg_copy = mcp_res->idx;
                        if (dev->data->dev_started) {
                                mcp_res->appcnt++;
                                flow->copy_applied = 1;
@@ -3408,7 +3452,7 @@ flow_hairpin_split(struct rte_eth_dev *dev,
  * The last stage of splitting chain, just creates the subflow
  * without any modification.
  *
- * @param dev
+ * @param[in] dev
  *   Pointer to Ethernet device.
  * @param[in] flow
  *   Parent flow structure pointer.
@@ -3441,19 +3485,20 @@ flow_create_split_inner(struct rte_eth_dev *dev,
 {
        struct mlx5_flow *dev_flow;
 
-       dev_flow = flow_drv_prepare(flow, attr, items, actions, error);
+       dev_flow = flow_drv_prepare(dev, flow, attr, items, actions, error);
        if (!dev_flow)
                return -rte_errno;
        dev_flow->flow = flow;
        dev_flow->external = external;
        /* Subflow object was created, we must include one in the list. */
-       LIST_INSERT_HEAD(&flow->dev_flows, dev_flow, next);
+       SILIST_INSERT(&flow->dev_handles, dev_flow->handle_idx,
+                     dev_flow->handle, next);
        /*
         * If dev_flow is as one of the suffix flow, some actions in suffix
         * flow may need some user defined item layer flags.
         */
        if (prefix_layers)
-               dev_flow->layers = prefix_layers;
+               dev_flow->handle->layers = prefix_layers;
        if (sub_flow)
                *sub_flow = dev_flow;
        return flow_drv_translate(dev, dev_flow, attr, items, actions, error);
@@ -3968,8 +4013,7 @@ flow_create_split_metadata(struct rte_eth_dev *dev,
                         * reallocation becomes possible (for example, for
                         * other flows in other threads).
                         */
-                       dev_flow->qrss_id = qrss_id;
-                       qrss_id = 0;
+                       dev_flow->handle->split_flow_id = qrss_id;
                        ret = mlx5_flow_get_reg_id(dev, MLX5_COPY_MARK, 0,
                                                   error);
                        if (ret < 0)
@@ -3984,6 +4028,8 @@ flow_create_split_metadata(struct rte_eth_dev *dev,
                                              external, error);
                if (ret < 0)
                        goto exit;
+               /* qrss ID should be freed if failed. */
+               qrss_id = 0;
                MLX5_ASSERT(dev_flow);
        }
 
@@ -4080,7 +4126,7 @@ flow_create_split_meter(struct rte_eth_dev *dev,
                        ret = -rte_errno;
                        goto exit;
                }
-               dev_flow->mtr_flow_id = mtr_tag_id;
+               dev_flow->handle->split_flow_id = mtr_tag_id;
                /* Setting the sfx group atrr. */
                sfx_attr.group = sfx_attr.transfer ?
                                (MLX5_FLOW_TABLE_LEVEL_SUFFIX - 1) :
@@ -4202,9 +4248,10 @@ flow_list_create(struct rte_eth_dev *dev, struct mlx5_flows *list,
                uint8_t buffer[2048];
        } items_tx;
        struct rte_flow_expand_rss *buf = &expand_buffer.buf;
+       struct mlx5_flow_rss_desc *rss_desc = &((struct mlx5_flow_rss_desc *)
+                                             priv->rss_desc)[!!priv->flow_idx];
        const struct rte_flow_action *p_actions_rx = actions;
        uint32_t i;
-       uint32_t flow_size;
        int hairpin_flow = 0;
        uint32_t hairpin_id = 0;
        struct rte_flow_attr attr_tx = { .priority = 0 };
@@ -4224,14 +4271,7 @@ flow_list_create(struct rte_eth_dev *dev, struct mlx5_flows *list,
                                   &hairpin_id);
                p_actions_rx = actions_rx.actions;
        }
-       flow_size = sizeof(struct rte_flow);
-       rss = flow_get_rss_action(p_actions_rx);
-       if (rss)
-               flow_size += RTE_ALIGN_CEIL(rss->queue_num * sizeof(uint16_t),
-                                           sizeof(void *));
-       else
-               flow_size += RTE_ALIGN_CEIL(sizeof(uint16_t), sizeof(void *));
-       flow = rte_calloc(__func__, 1, flow_size, 0);
+       flow = rte_calloc(__func__, 1, sizeof(struct rte_flow), 0);
        if (!flow) {
                rte_errno = ENOMEM;
                goto error_before_flow;
@@ -4241,17 +4281,18 @@ flow_list_create(struct rte_eth_dev *dev, struct mlx5_flows *list,
                flow->hairpin_flow_id = hairpin_id;
        MLX5_ASSERT(flow->drv_type > MLX5_FLOW_TYPE_MIN &&
                    flow->drv_type < MLX5_FLOW_TYPE_MAX);
-       flow->rss.queue = (void *)(flow + 1);
+       memset(rss_desc, 0, sizeof(*rss_desc));
+       rss = flow_get_rss_action(p_actions_rx);
        if (rss) {
                /*
                 * The following information is required by
                 * mlx5_flow_hashfields_adjust() in advance.
                 */
-               flow->rss.level = rss->level;
+               rss_desc->level = rss->level;
                /* RSS type 0 indicates default RSS type (ETH_RSS_IP). */
-               flow->rss.types = !rss->types ? ETH_RSS_IP : rss->types;
+               rss_desc->types = !rss->types ? ETH_RSS_IP : rss->types;
        }
-       LIST_INIT(&flow->dev_flows);
+       flow->dev_handles = 0;
        if (rss && rss->types) {
                unsigned int graph_root;
 
@@ -4266,6 +4307,15 @@ flow_list_create(struct rte_eth_dev *dev, struct mlx5_flows *list,
                buf->entries = 1;
                buf->entry[0].pattern = (void *)(uintptr_t)items;
        }
+       /*
+        * 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 (priv->flow_idx) {
+               MLX5_ASSERT(!priv->flow_nested_idx);
+               priv->flow_nested_idx = priv->flow_idx;
+       }
        for (i = 0; i < buf->entries; ++i) {
                /*
                 * The splitter may create multiple dev_flows,
@@ -4284,13 +4334,14 @@ flow_list_create(struct rte_eth_dev *dev, struct mlx5_flows *list,
                attr_tx.group = MLX5_HAIRPIN_TX_TABLE;
                attr_tx.ingress = 0;
                attr_tx.egress = 1;
-               dev_flow = flow_drv_prepare(flow, &attr_tx, items_tx.items,
+               dev_flow = flow_drv_prepare(dev, flow, &attr_tx, items_tx.items,
                                            actions_hairpin_tx.actions, error);
                if (!dev_flow)
                        goto error;
                dev_flow->flow = flow;
                dev_flow->external = 0;
-               LIST_INSERT_HEAD(&flow->dev_flows, dev_flow, next);
+               SILIST_INSERT(&flow->dev_handles, dev_flow->handle_idx,
+                             dev_flow->handle, next);
                ret = flow_drv_translate(dev, dev_flow, &attr_tx,
                                         items_tx.items,
                                         actions_hairpin_tx.actions, error);
@@ -4313,7 +4364,11 @@ flow_list_create(struct rte_eth_dev *dev, struct mlx5_flows *list,
                if (ret)
                        goto error;
        }
-       if (dev->data->dev_started) {
+       /*
+        * If the flow is external (from application) OR device is started, then
+        * the flow will be applied immediately.
+        */
+       if (external || dev->data->dev_started) {
                ret = flow_drv_apply(dev, flow, error);
                if (ret < 0)
                        goto error;
@@ -4321,23 +4376,27 @@ flow_list_create(struct rte_eth_dev *dev, struct mlx5_flows *list,
        if (list)
                TAILQ_INSERT_TAIL(list, flow, next);
        flow_rxq_flags_set(dev, flow);
+       /* Nested flow creation index recovery. */
+       priv->flow_idx = priv->flow_nested_idx;
+       if (priv->flow_nested_idx)
+               priv->flow_nested_idx = 0;
        return flow;
-error_before_flow:
-       if (hairpin_id)
-               mlx5_flow_id_release(priv->sh->flow_id_pool,
-                                    hairpin_id);
-       return NULL;
 error:
        MLX5_ASSERT(flow);
-       flow_mreg_del_copy_action(dev, flow);
        ret = rte_errno; /* Save rte_errno before cleanup. */
-       if (flow->hairpin_flow_id)
-               mlx5_flow_id_release(priv->sh->flow_id_pool,
-                                    flow->hairpin_flow_id);
-       MLX5_ASSERT(flow);
+       flow_mreg_del_copy_action(dev, flow);
        flow_drv_destroy(dev, flow);
        rte_free(flow);
        rte_errno = ret; /* Restore rte_errno. */
+error_before_flow:
+       ret = rte_errno;
+       if (hairpin_id)
+               mlx5_flow_id_release(priv->sh->flow_id_pool,
+                                    hairpin_id);
+       rte_errno = ret;
+       priv->flow_idx = priv->flow_nested_idx;
+       if (priv->flow_nested_idx)
+               priv->flow_nested_idx = 0;
        return NULL;
 }
 
@@ -4405,6 +4464,17 @@ mlx5_flow_create(struct rte_eth_dev *dev,
 {
        struct mlx5_priv *priv = dev->data->dev_private;
 
+       /*
+        * If the device is not started yet, it is not allowed to created a
+        * flow from application. PMD default flows and traffic control flows
+        * are not affected.
+        */
+       if (unlikely(!dev->data->dev_started)) {
+               rte_errno = ENODEV;
+               DRV_LOG(DEBUG, "port %u is not started when "
+                       "inserting a flow", dev->data->port_id);
+               return NULL;
+       }
        return flow_list_create(dev, &priv->flows,
                                attr, items, actions, true, error);
 }
@@ -4425,6 +4495,7 @@ flow_list_destroy(struct rte_eth_dev *dev, struct mlx5_flows *list,
                  struct rte_flow *flow)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
+       struct mlx5_fdir_flow *priv_fdir_flow = NULL;
 
        /*
         * Update RX queue flags only if port is started, otherwise it is
@@ -4439,7 +4510,17 @@ flow_list_destroy(struct rte_eth_dev *dev, struct mlx5_flows *list,
        if (list)
                TAILQ_REMOVE(list, flow, next);
        flow_mreg_del_copy_action(dev, flow);
-       rte_free(flow->fdir);
+       if (flow->fdir) {
+               LIST_FOREACH(priv_fdir_flow, &priv->fdir_flows, next) {
+                       if (priv_fdir_flow->flow == flow)
+                               break;
+               }
+               if (priv_fdir_flow) {
+                       LIST_REMOVE(priv_fdir_flow, next);
+                       rte_free(priv_fdir_flow->fdir);
+                       rte_free(priv_fdir_flow);
+               }
+       }
        rte_free(flow);
 }
 
@@ -4538,8 +4619,6 @@ error:
  *
  * @param dev
  *   Pointer to Ethernet device.
- * @param list
- *   Pointer to a TAILQ flow list.
  */
 void
 mlx5_flow_stop_default(struct rte_eth_dev *dev)
@@ -4564,6 +4643,50 @@ mlx5_flow_start_default(struct rte_eth_dev *dev)
        return flow_mreg_add_default_copy_action(dev, &error);
 }
 
+/**
+ * Allocate intermediate resources for flow creation.
+ *
+ * @param dev
+ *   Pointer to Ethernet device.
+ */
+void
+mlx5_flow_alloc_intermediate(struct rte_eth_dev *dev)
+{
+       struct mlx5_priv *priv = dev->data->dev_private;
+
+       if (!priv->inter_flows) {
+               priv->inter_flows = rte_calloc(__func__, 1,
+                                   MLX5_NUM_MAX_DEV_FLOWS *
+                                   sizeof(struct mlx5_flow) +
+                                   (sizeof(struct mlx5_flow_rss_desc) +
+                                   sizeof(uint16_t) * UINT16_MAX) * 2, 0);
+               if (!priv->inter_flows) {
+                       DRV_LOG(ERR, "can't allocate intermediate memory.");
+                       return;
+               }
+       }
+       priv->rss_desc = &((struct mlx5_flow *)priv->inter_flows)
+                        [MLX5_NUM_MAX_DEV_FLOWS];
+       /* Reset the index. */
+       priv->flow_idx = 0;
+       priv->flow_nested_idx = 0;
+}
+
+/**
+ * Free intermediate resources for flows.
+ *
+ * @param dev
+ *   Pointer to Ethernet device.
+ */
+void
+mlx5_flow_free_intermediate(struct rte_eth_dev *dev)
+{
+       struct mlx5_priv *priv = dev->data->dev_private;
+
+       rte_free(priv->inter_flows);
+       priv->inter_flows = NULL;
+}
+
 /**
  * Verify the flow list is empty
  *
@@ -5091,12 +5214,14 @@ flow_fdir_filter_lookup(struct rte_eth_dev *dev, struct mlx5_fdir *fdir_flow)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
        struct rte_flow *flow = NULL;
+       struct mlx5_fdir_flow *priv_fdir_flow = NULL;
 
        MLX5_ASSERT(fdir_flow);
-       TAILQ_FOREACH(flow, &priv->flows, next) {
-               if (flow->fdir && !flow_fdir_cmp(flow->fdir, fdir_flow)) {
+       LIST_FOREACH(priv_fdir_flow, &priv->fdir_flows, next) {
+               if (!flow_fdir_cmp(priv_fdir_flow->fdir, fdir_flow)) {
                        DRV_LOG(DEBUG, "port %u found FDIR flow %p",
                                dev->data->port_id, (void *)flow);
+                       flow = priv_fdir_flow->flow;
                        break;
                }
        }
@@ -5121,6 +5246,7 @@ flow_fdir_filter_add(struct rte_eth_dev *dev,
        struct mlx5_priv *priv = dev->data->dev_private;
        struct mlx5_fdir *fdir_flow;
        struct rte_flow *flow;
+       struct mlx5_fdir_flow *priv_fdir_flow = NULL;
        int ret;
 
        fdir_flow = rte_zmalloc(__func__, sizeof(*fdir_flow), 0);
@@ -5136,17 +5262,26 @@ flow_fdir_filter_add(struct rte_eth_dev *dev,
                rte_errno = EEXIST;
                goto error;
        }
+       priv_fdir_flow = rte_zmalloc(__func__, sizeof(struct mlx5_fdir_flow),
+                                    0);
+       if (!priv_fdir_flow) {
+               rte_errno = ENOMEM;
+               goto error;
+       }
        flow = flow_list_create(dev, &priv->flows, &fdir_flow->attr,
                                fdir_flow->items, fdir_flow->actions, true,
                                NULL);
        if (!flow)
                goto error;
-       MLX5_ASSERT(!flow->fdir);
-       flow->fdir = fdir_flow;
+       flow->fdir = 1;
+       priv_fdir_flow->fdir = fdir_flow;
+       priv_fdir_flow->flow = flow;
+       LIST_INSERT_HEAD(&priv->fdir_flows, priv_fdir_flow, next);
        DRV_LOG(DEBUG, "port %u created FDIR flow %p",
                dev->data->port_id, (void *)flow);
        return 0;
 error:
+       rte_free(priv_fdir_flow);
        rte_free(fdir_flow);
        return -rte_errno;
 }
@@ -5171,17 +5306,26 @@ flow_fdir_filter_delete(struct rte_eth_dev *dev,
        struct mlx5_fdir fdir_flow = {
                .attr.group = 0,
        };
+       struct mlx5_fdir_flow *priv_fdir_flow = NULL;
        int ret;
 
        ret = flow_fdir_filter_convert(dev, fdir_filter, &fdir_flow);
        if (ret)
                return -rte_errno;
-       flow = flow_fdir_filter_lookup(dev, &fdir_flow);
-       if (!flow) {
-               rte_errno = ENOENT;
-               return -rte_errno;
+       LIST_FOREACH(priv_fdir_flow, &priv->fdir_flows, next) {
+               /* Find the fdir in priv list */
+               if (!flow_fdir_cmp(priv_fdir_flow->fdir, &fdir_flow))
+                       break;
        }
+       if (!priv_fdir_flow)
+               return 0;
+       LIST_REMOVE(priv_fdir_flow, next);
+       flow = priv_fdir_flow->flow;
+       /* Fdir resource will be releasd after flow destroy. */
+       flow->fdir = 0;
        flow_list_destroy(dev, &priv->flows, flow);
+       rte_free(priv_fdir_flow->fdir);
+       rte_free(priv_fdir_flow);
        DRV_LOG(DEBUG, "port %u deleted FDIR flow %p",
                dev->data->port_id, (void *)flow);
        return 0;
@@ -5220,8 +5364,16 @@ static void
 flow_fdir_filter_flush(struct rte_eth_dev *dev)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
-
-       mlx5_flow_list_flush(dev, &priv->flows, false);
+       struct mlx5_fdir_flow *priv_fdir_flow = NULL;
+
+       while (!LIST_EMPTY(&priv->fdir_flows)) {
+               priv_fdir_flow = LIST_FIRST(&priv->fdir_flows);
+               LIST_REMOVE(priv_fdir_flow, next);
+               priv_fdir_flow->flow->fdir = 0;
+               flow_list_destroy(dev, &priv->flows, priv_fdir_flow->flow);
+               rte_free(priv_fdir_flow->fdir);
+               rte_free(priv_fdir_flow);
+       }
 }
 
 /**
@@ -5435,9 +5587,9 @@ mlx5_flow_destroy_policer_rules(struct rte_eth_dev *dev,
  *   Pointer to Ethernet device structure.
  *
  * @return
- *   Pointer to allocated counter  on success, NULL otherwise.
+ *   Index to allocated counter  on success, 0 otherwise.
  */
-struct mlx5_flow_counter *
+uint32_t
 mlx5_counter_alloc(struct rte_eth_dev *dev)
 {
        const struct mlx5_flow_driver_ops *fops;
@@ -5450,7 +5602,7 @@ mlx5_counter_alloc(struct rte_eth_dev *dev)
        DRV_LOG(ERR,
                "port %u counter allocate is not supported.",
                 dev->data->port_id);
-       return NULL;
+       return 0;
 }
 
 /**
@@ -5459,10 +5611,10 @@ mlx5_counter_alloc(struct rte_eth_dev *dev)
  * @param[in] dev
  *   Pointer to Ethernet device structure.
  * @param[in] cnt
- *   Pointer to counter to be free.
+ *   Index to counter to be free.
  */
 void
-mlx5_counter_free(struct rte_eth_dev *dev, struct mlx5_flow_counter *cnt)
+mlx5_counter_free(struct rte_eth_dev *dev, uint32_t cnt)
 {
        const struct mlx5_flow_driver_ops *fops;
        struct rte_flow_attr attr = { .transfer = 0 };
@@ -5483,7 +5635,7 @@ mlx5_counter_free(struct rte_eth_dev *dev, struct mlx5_flow_counter *cnt)
  * @param[in] dev
  *   Pointer to Ethernet device structure.
  * @param[in] cnt
- *   Pointer to counter to query.
+ *   Index to counter to query.
  * @param[in] clear
  *   Set to clear counter statistics.
  * @param[out] pkts
@@ -5495,7 +5647,7 @@ mlx5_counter_free(struct rte_eth_dev *dev, struct mlx5_flow_counter *cnt)
  *   0 on success, a negative errno value otherwise.
  */
 int
-mlx5_counter_query(struct rte_eth_dev *dev, struct mlx5_flow_counter *cnt,
+mlx5_counter_query(struct rte_eth_dev *dev, uint32_t cnt,
                   bool clear, uint64_t *pkts, uint64_t *bytes)
 {
        const struct mlx5_flow_driver_ops *fops;
@@ -5596,6 +5748,13 @@ next_container:
        dcs = (struct mlx5_devx_obj *)(uintptr_t)rte_atomic64_read
                                                              (&pool->a64_dcs);
        offset = batch ? 0 : dcs->id % MLX5_COUNTERS_PER_POOL;
+       /*
+        * Identify the counters released between query trigger and query
+        * handle more effiecntly. The counter released in this gap period
+        * should wait for a new round of query as the new arrived packets
+        * will not be taken into account.
+        */
+       rte_atomic64_add(&pool->start_query_gen, 1);
        ret = mlx5_devx_cmd_flow_counter_query(dcs, 0, MLX5_COUNTERS_PER_POOL -
                                               offset, NULL, NULL,
                                               pool->raw_hw->mem_mng->dm->id,
@@ -5604,6 +5763,7 @@ next_container:
                                               sh->devx_comp,
                                               (uint64_t)(uintptr_t)pool);
        if (ret) {
+               rte_atomic64_sub(&pool->start_query_gen, 1);
                DRV_LOG(ERR, "Failed to trigger asynchronous query for dcs ID"
                        " %d", pool->min_dcs->id);
                pool->raw_hw = NULL;
@@ -5643,13 +5803,17 @@ mlx5_flow_async_pool_query_handle(struct mlx5_ibv_shared *sh,
        struct mlx5_counter_stats_raw *raw_to_free;
 
        if (unlikely(status)) {
+               rte_atomic64_sub(&pool->start_query_gen, 1);
                raw_to_free = pool->raw_hw;
        } else {
                raw_to_free = pool->raw;
                rte_spinlock_lock(&pool->sl);
                pool->raw = pool->raw_hw;
                rte_spinlock_unlock(&pool->sl);
-               rte_atomic64_add(&pool->query_gen, 1);
+               MLX5_ASSERT(rte_atomic64_read(&pool->end_query_gen) + 1 ==
+                           rte_atomic64_read(&pool->start_query_gen));
+               rte_atomic64_set(&pool->end_query_gen,
+                                rte_atomic64_read(&pool->start_query_gen));
                /* Be sure the new raw counters data is updated in memory. */
                rte_cio_wmb();
        }