net/mlx5: fix flow counter container resize
[dpdk.git] / drivers / net / mlx5 / mlx5_flow_dv.c
index c6d132c..fd002f3 100644 (file)
@@ -24,6 +24,7 @@
 #include <rte_flow.h>
 #include <rte_flow_driver.h>
 #include <rte_malloc.h>
+#include <rte_cycles.h>
 #include <rte_ip.h>
 #include <rte_gre.h>
 #include <rte_vxlan.h>
@@ -1638,6 +1639,91 @@ flow_dv_validate_item_port_id(struct rte_eth_dev *dev,
        return 0;
 }
 
+/*
+ * GTP flags are contained in 1 byte of the format:
+ * -------------------------------------------
+ * | bit   | 0 - 2   | 3  | 4   | 5 | 6 | 7  |
+ * |-----------------------------------------|
+ * | value | Version | PT | Res | E | S | PN |
+ * -------------------------------------------
+ *
+ * Matching is supported only for GTP flags E, S, PN.
+ */
+#define MLX5_GTP_FLAGS_MASK    0x07
+
+/**
+ * Validate VLAN item.
+ *
+ * @param[in] item
+ *   Item specification.
+ * @param[in] item_flags
+ *   Bit-fields that holds the items detected until now.
+ * @param[in] dev
+ *   Ethernet device flow is being created on.
+ * @param[out] error
+ *   Pointer to error structure.
+ *
+ * @return
+ *   0 on success, a negative errno value otherwise and rte_errno is set.
+ */
+static int
+flow_dv_validate_item_vlan(const struct rte_flow_item *item,
+                          uint64_t item_flags,
+                          struct rte_eth_dev *dev,
+                          struct rte_flow_error *error)
+{
+       const struct rte_flow_item_vlan *mask = item->mask;
+       const struct rte_flow_item_vlan nic_mask = {
+               .tci = RTE_BE16(UINT16_MAX),
+               .inner_type = RTE_BE16(UINT16_MAX),
+       };
+       const int tunnel = !!(item_flags & MLX5_FLOW_LAYER_TUNNEL);
+       int ret;
+       const uint64_t l34m = tunnel ? (MLX5_FLOW_LAYER_INNER_L3 |
+                                       MLX5_FLOW_LAYER_INNER_L4) :
+                                      (MLX5_FLOW_LAYER_OUTER_L3 |
+                                       MLX5_FLOW_LAYER_OUTER_L4);
+       const uint64_t vlanm = tunnel ? MLX5_FLOW_LAYER_INNER_VLAN :
+                                       MLX5_FLOW_LAYER_OUTER_VLAN;
+
+       if (item_flags & vlanm)
+               return rte_flow_error_set(error, EINVAL,
+                                         RTE_FLOW_ERROR_TYPE_ITEM, item,
+                                         "multiple VLAN layers not supported");
+       else if ((item_flags & l34m) != 0)
+               return rte_flow_error_set(error, EINVAL,
+                                         RTE_FLOW_ERROR_TYPE_ITEM, item,
+                                         "VLAN cannot follow L3/L4 layer");
+       if (!mask)
+               mask = &rte_flow_item_vlan_mask;
+       ret = mlx5_flow_item_acceptable(item, (const uint8_t *)mask,
+                                       (const uint8_t *)&nic_mask,
+                                       sizeof(struct rte_flow_item_vlan),
+                                       error);
+       if (ret)
+               return ret;
+       if (!tunnel && mask->tci != RTE_BE16(0x0fff)) {
+               struct mlx5_priv *priv = dev->data->dev_private;
+
+               if (priv->vmwa_context) {
+                       /*
+                        * Non-NULL context means we have a virtual machine
+                        * and SR-IOV enabled, we have to create VLAN interface
+                        * to make hypervisor to setup E-Switch vport
+                        * context correctly. We avoid creating the multiple
+                        * VLAN interfaces, so we cannot support VLAN tag mask.
+                        */
+                       return rte_flow_error_set(error, EINVAL,
+                                                 RTE_FLOW_ERROR_TYPE_ITEM,
+                                                 item,
+                                                 "VLAN tag mask is not"
+                                                 " supported in virtual"
+                                                 " environment");
+               }
+       }
+       return 0;
+}
+
 /**
  * Validate GTP item.
  *
@@ -1660,8 +1746,10 @@ flow_dv_validate_item_gtp(struct rte_eth_dev *dev,
                          struct rte_flow_error *error)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
+       const struct rte_flow_item_gtp *spec = item->spec;
        const struct rte_flow_item_gtp *mask = item->mask;
        const struct rte_flow_item_gtp nic_mask = {
+               .v_pt_rsv_flags = MLX5_GTP_FLAGS_MASK,
                .msg_type = 0xff,
                .teid = RTE_BE32(0xffffffff),
        };
@@ -1681,6 +1769,11 @@ flow_dv_validate_item_gtp(struct rte_eth_dev *dev,
                                          "no outer UDP layer found");
        if (!mask)
                mask = &rte_flow_item_gtp_mask;
+       if (spec && spec->v_pt_rsv_flags & ~MLX5_GTP_FLAGS_MASK)
+               return rte_flow_error_set(error, ENOTSUP,
+                                         RTE_FLOW_ERROR_TYPE_ITEM, item,
+                                         "Match is supported for GTP"
+                                         " flags only");
        return mlx5_flow_item_acceptable
                (item, (const uint8_t *)mask,
                 (const uint8_t *)&nic_mask,
@@ -1964,7 +2057,7 @@ flow_dv_validate_action_set_vlan_vid(uint64_t item_flags,
        const struct rte_flow_action *action = actions;
        const struct rte_flow_action_of_set_vlan_vid *conf = action->conf;
 
-       if (conf->vlan_vid > RTE_BE16(0xFFE))
+       if (rte_be_to_cpu_16(conf->vlan_vid) > 0xFFE)
                return rte_flow_error_set(error, EINVAL,
                                          RTE_FLOW_ERROR_TYPE_ACTION, action,
                                          "VLAN VID value is too big");
@@ -3639,21 +3732,18 @@ flow_dv_validate_action_port_id(struct rte_eth_dev *dev,
  * @return
  *   Max number of modify header actions device can support.
  */
-static unsigned int
-flow_dv_modify_hdr_action_max(struct rte_eth_dev *dev, uint64_t flags)
+static inline unsigned int
+flow_dv_modify_hdr_action_max(struct rte_eth_dev *dev __rte_unused,
+                             uint64_t flags)
 {
        /*
-        * There's no way to directly query the max cap. Although it has to be
-        * acquried by iterative trial, it is a safe assumption that more
-        * actions are supported by FW if extensive metadata register is
-        * supported. (Only in the root table)
+        * There's no way to directly query the max capacity from FW.
+        * The maximal value on root table should be assumed to be supported.
         */
        if (!(flags & MLX5DV_DR_ACTION_FLAGS_ROOT_LEVEL))
                return MLX5_MAX_MODIFY_NUM;
        else
-               return mlx5_flow_ext_mreg_supported(dev) ?
-                                       MLX5_ROOT_TBL_MODIFY_NUM :
-                                       MLX5_ROOT_TBL_MODIFY_NUM_NO_MREG;
+               return MLX5_ROOT_TBL_MODIFY_NUM;
 }
 
 /**
@@ -3718,6 +3808,50 @@ mlx5_flow_validate_action_meter(struct rte_eth_dev *dev,
        return 0;
 }
 
+/**
+ * Validate the age action.
+ *
+ * @param[in] action_flags
+ *   Holds the actions detected until now.
+ * @param[in] action
+ *   Pointer to the age action.
+ * @param[in] dev
+ *   Pointer to the Ethernet device structure.
+ * @param[out] error
+ *   Pointer to error structure.
+ *
+ * @return
+ *   0 on success, a negative errno value otherwise and rte_errno is set.
+ */
+static int
+flow_dv_validate_action_age(uint64_t action_flags,
+                           const struct rte_flow_action *action,
+                           struct rte_eth_dev *dev,
+                           struct rte_flow_error *error)
+{
+       struct mlx5_priv *priv = dev->data->dev_private;
+       const struct rte_flow_action_age *age = action->conf;
+
+       if (!priv->config.devx || priv->counter_fallback)
+               return rte_flow_error_set(error, ENOTSUP,
+                                         RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
+                                         NULL,
+                                         "age action not supported");
+       if (!(action->conf))
+               return rte_flow_error_set(error, EINVAL,
+                                         RTE_FLOW_ERROR_TYPE_ACTION, action,
+                                         "configuration cannot be null");
+       if (age->timeout >= UINT16_MAX / 2 / 10)
+               return rte_flow_error_set(error, ENOTSUP,
+                                         RTE_FLOW_ERROR_TYPE_ACTION, action,
+                                         "Max age time: 3275 seconds");
+       if (action_flags & MLX5_FLOW_ACTION_AGE)
+               return rte_flow_error_set(error, EINVAL,
+                                         RTE_FLOW_ERROR_TYPE_ACTION, NULL,
+                                         "Duplicate age ctions set");
+       return 0;
+}
+
 /**
  * Validate the modify-header IPv4 DSCP actions.
  *
@@ -3895,20 +4029,22 @@ flow_dv_counter_get_by_idx(struct rte_eth_dev *dev,
        struct mlx5_priv *priv = dev->data->dev_private;
        struct mlx5_pools_container *cont;
        struct mlx5_flow_counter_pool *pool;
-       uint32_t batch = 0;
+       uint32_t batch = 0, age = 0;
 
        idx--;
+       age = MLX_CNT_IS_AGE(idx);
+       idx = age ? idx - MLX5_CNT_AGE_OFFSET : idx;
        if (idx >= MLX5_CNT_BATCH_OFFSET) {
                idx -= MLX5_CNT_BATCH_OFFSET;
                batch = 1;
        }
-       cont = MLX5_CNT_CONTAINER(priv->sh, batch, 0);
+       cont = MLX5_CNT_CONTAINER(priv->sh, batch, age);
        MLX5_ASSERT(idx / MLX5_COUNTERS_PER_POOL < cont->n);
        pool = cont->pools[idx / MLX5_COUNTERS_PER_POOL];
        MLX5_ASSERT(pool);
        if (ppool)
                *ppool = pool;
-       return &pool->counters_raw[idx % MLX5_COUNTERS_PER_POOL];
+       return MLX5_POOL_GET_CNT(pool, idx % MLX5_COUNTERS_PER_POOL);
 }
 
 /**
@@ -3964,8 +4100,8 @@ flow_dv_find_pool_by_id(struct mlx5_pools_container *cont, int id)
 static struct mlx5_counter_stats_mem_mng *
 flow_dv_create_counter_stat_mem_mng(struct rte_eth_dev *dev, int raws_n)
 {
-       struct mlx5_ibv_shared *sh = ((struct mlx5_priv *)
-                                       (dev->data->dev_private))->sh;
+       struct mlx5_priv *priv = dev->data->dev_private;
+       struct mlx5_ibv_shared *sh = priv->sh;
        struct mlx5_devx_mkey_attr mkey_attr;
        struct mlx5_counter_stats_mem_mng *mem_mng;
        volatile struct flow_counter_stats *raw_data;
@@ -3997,7 +4133,10 @@ flow_dv_create_counter_stat_mem_mng(struct rte_eth_dev *dev, int raws_n)
        mkey_attr.pg_access = 0;
        mkey_attr.klm_array = NULL;
        mkey_attr.klm_num = 0;
-       mkey_attr.relaxed_ordering = 1;
+       if (priv->config.hca_attr.relaxed_ordering_write &&
+               priv->config.hca_attr.relaxed_ordering_read  &&
+               !haswell_broadwell_cpu)
+               mkey_attr.relaxed_ordering = 1;
        mem_mng->dm = mlx5_devx_cmd_mkey_create(sh->ctx, &mkey_attr);
        if (!mem_mng->dm) {
                mlx5_glue->devx_umem_dereg(mem_mng->umem);
@@ -4022,70 +4161,59 @@ flow_dv_create_counter_stat_mem_mng(struct rte_eth_dev *dev, int raws_n)
  *   Pointer to the Ethernet device structure.
  * @param[in] batch
  *   Whether the pool is for counter that was allocated by batch command.
+ * @param[in] age
+ *   Whether the pool is for Aging counter.
  *
  * @return
- *   The new container pointer on success, otherwise NULL and rte_errno is set.
+ *   0 on success, otherwise negative errno value and rte_errno is set.
  */
-static struct mlx5_pools_container *
-flow_dv_container_resize(struct rte_eth_dev *dev, uint32_t batch)
+static int
+flow_dv_container_resize(struct rte_eth_dev *dev,
+                               uint32_t batch, uint32_t age)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
-       struct mlx5_pools_container *cont =
-                       MLX5_CNT_CONTAINER(priv->sh, batch, 0);
-       struct mlx5_pools_container *new_cont =
-                       MLX5_CNT_CONTAINER_UNUSED(priv->sh, batch, 0);
+       struct mlx5_pools_container *cont = MLX5_CNT_CONTAINER(priv->sh, batch,
+                                                              age);
        struct mlx5_counter_stats_mem_mng *mem_mng = NULL;
+       void *old_pools = cont->pools;
        uint32_t resize = cont->n + MLX5_CNT_CONTAINER_RESIZE;
        uint32_t mem_size = sizeof(struct mlx5_flow_counter_pool *) * resize;
-       int i;
+       void *pools = rte_calloc(__func__, 1, mem_size, 0);
 
-       /* Fallback mode has no background thread. Skip the check. */
-       if (!priv->counter_fallback &&
-           cont != MLX5_CNT_CONTAINER(priv->sh, batch, 1)) {
-               /* The last resize still hasn't detected by the host thread. */
-               rte_errno = EAGAIN;
-               return NULL;
-       }
-       new_cont->pools = rte_calloc(__func__, 1, mem_size, 0);
-       if (!new_cont->pools) {
+       if (!pools) {
                rte_errno = ENOMEM;
-               return NULL;
+               return -ENOMEM;
        }
-       if (cont->n)
-               memcpy(new_cont->pools, cont->pools, cont->n *
-                      sizeof(struct mlx5_flow_counter_pool *));
+       if (old_pools)
+               memcpy(pools, old_pools, cont->n *
+                                      sizeof(struct mlx5_flow_counter_pool *));
        /*
         * Fallback mode query the counter directly, no background query
         * resources are needed.
         */
        if (!priv->counter_fallback) {
+               int i;
+
                mem_mng = flow_dv_create_counter_stat_mem_mng(dev,
-                       MLX5_CNT_CONTAINER_RESIZE + MLX5_MAX_PENDING_QUERIES);
+                         MLX5_CNT_CONTAINER_RESIZE + MLX5_MAX_PENDING_QUERIES);
                if (!mem_mng) {
-                       rte_free(new_cont->pools);
-                       return NULL;
+                       rte_free(pools);
+                       return -ENOMEM;
                }
                for (i = 0; i < MLX5_MAX_PENDING_QUERIES; ++i)
                        LIST_INSERT_HEAD(&priv->sh->cmng.free_stat_raws,
                                         mem_mng->raws +
                                         MLX5_CNT_CONTAINER_RESIZE +
                                         i, next);
-       } else {
-               /*
-                * Release the old container pools directly as no background
-                * thread helps that.
-                */
-               rte_free(cont->pools);
        }
-       new_cont->n = resize;
-       rte_atomic16_set(&new_cont->n_valid, rte_atomic16_read(&cont->n_valid));
-       TAILQ_INIT(&new_cont->pool_list);
-       TAILQ_CONCAT(&new_cont->pool_list, &cont->pool_list, next);
-       new_cont->init_mem_mng = mem_mng;
-       rte_cio_wmb();
-        /* Flip the master container. */
-       priv->sh->cmng.mhi[batch] ^= (uint8_t)1;
-       return new_cont;
+       rte_spinlock_lock(&cont->resize_sl);
+       cont->n = resize;
+       cont->mem_mng = mem_mng;
+       cont->pools = pools;
+       rte_spinlock_unlock(&cont->resize_sl);
+       if (old_pools)
+               rte_free(old_pools);
+       return 0;
 }
 
 /**
@@ -4132,7 +4260,7 @@ _flow_dv_query_count(struct rte_eth_dev *dev, uint32_t counter, uint64_t *pkts,
                *pkts = 0;
                *bytes = 0;
        } else {
-               offset = cnt - &pool->counters_raw[0];
+               offset = MLX5_CNT_ARRAY_IDX(pool, cnt);
                *pkts = rte_be_to_cpu_64(pool->raw->data[offset].hits);
                *bytes = rte_be_to_cpu_64(pool->raw->data[offset].bytes);
        }
@@ -4149,32 +4277,30 @@ _flow_dv_query_count(struct rte_eth_dev *dev, uint32_t counter, uint64_t *pkts,
  *   The devX counter handle.
  * @param[in] batch
  *   Whether the pool is for counter that was allocated by batch command.
+ * @param[in] age
+ *   Whether the pool is for counter that was allocated for aging.
  * @param[in/out] cont_cur
  *   Pointer to the container pointer, it will be update in pool resize.
  *
  * @return
  *   The pool container pointer on success, NULL otherwise and rte_errno is set.
  */
-static struct mlx5_pools_container *
+static struct mlx5_flow_counter_pool *
 flow_dv_pool_create(struct rte_eth_dev *dev, struct mlx5_devx_obj *dcs,
-                   uint32_t batch)
+                   uint32_t batch, uint32_t age)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
        struct mlx5_flow_counter_pool *pool;
        struct mlx5_pools_container *cont = MLX5_CNT_CONTAINER(priv->sh, batch,
-                                                              0);
+                                                              age);
        int16_t n_valid = rte_atomic16_read(&cont->n_valid);
-       uint32_t size;
+       uint32_t size = sizeof(*pool);
 
-       if (cont->n == n_valid) {
-               cont = flow_dv_container_resize(dev, batch);
-               if (!cont)
-                       return NULL;
-       }
-       size = sizeof(*pool);
-       if (!batch)
-               size += MLX5_COUNTERS_PER_POOL *
-                       sizeof(struct mlx5_flow_counter_ext);
+       if (cont->n == n_valid && flow_dv_container_resize(dev, batch, age))
+               return NULL;
+       size += MLX5_COUNTERS_PER_POOL * CNT_SIZE;
+       size += (batch ? 0 : MLX5_COUNTERS_PER_POOL * CNTEXT_SIZE);
+       size += (!age ? 0 : MLX5_COUNTERS_PER_POOL * AGE_SIZE);
        pool = rte_calloc(__func__, 1, size, 0);
        if (!pool) {
                rte_errno = ENOMEM;
@@ -4182,9 +4308,12 @@ flow_dv_pool_create(struct rte_eth_dev *dev, struct mlx5_devx_obj *dcs,
        }
        pool->min_dcs = dcs;
        if (!priv->counter_fallback)
-               pool->raw = cont->init_mem_mng->raws + n_valid %
-                                                    MLX5_CNT_CONTAINER_RESIZE;
+               pool->raw = cont->mem_mng->raws + n_valid %
+                                                     MLX5_CNT_CONTAINER_RESIZE;
        pool->raw_hw = NULL;
+       pool->type = 0;
+       pool->type |= (batch ? 0 :  CNT_POOL_TYPE_EXT);
+       pool->type |= (!age ? 0 :  CNT_POOL_TYPE_AGE);
        rte_spinlock_init(&pool->sl);
        /*
         * The generation of the new allocated counters in this pool is 0, 2 in
@@ -4208,9 +4337,42 @@ flow_dv_pool_create(struct rte_eth_dev *dev, struct mlx5_devx_obj *dcs,
        /* Pool initialization must be updated before host thread access. */
        rte_cio_wmb();
        rte_atomic16_add(&cont->n_valid, 1);
-       return cont;
+       return pool;
 }
 
+/**
+ * Update the minimum dcs-id for aged or no-aged counter pool.
+ *
+ * @param[in] dev
+ *   Pointer to the Ethernet device structure.
+ * @param[in] pool
+ *   Current counter pool.
+ * @param[in] batch
+ *   Whether the pool is for counter that was allocated by batch command.
+ * @param[in] age
+ *   Whether the counter is for aging.
+ */
+static void
+flow_dv_counter_update_min_dcs(struct rte_eth_dev *dev,
+                       struct mlx5_flow_counter_pool *pool,
+                       uint32_t batch, uint32_t age)
+{
+       struct mlx5_priv *priv = dev->data->dev_private;
+       struct mlx5_flow_counter_pool *other;
+       struct mlx5_pools_container *cont;
+
+       cont = MLX5_CNT_CONTAINER(priv->sh, batch, (age ^ 0x1));
+       other = flow_dv_find_pool_by_id(cont, pool->min_dcs->id);
+       if (!other)
+               return;
+       if (pool->min_dcs->id < other->min_dcs->id) {
+               rte_atomic64_set(&other->a64_dcs,
+                       rte_atomic64_read(&pool->a64_dcs));
+       } else {
+               rte_atomic64_set(&pool->a64_dcs,
+                       rte_atomic64_read(&other->a64_dcs));
+       }
+}
 /**
  * Prepare a new counter and/or a new counter pool.
  *
@@ -4220,15 +4382,17 @@ flow_dv_pool_create(struct rte_eth_dev *dev, struct mlx5_devx_obj *dcs,
  *   Where to put the pointer of a new counter.
  * @param[in] batch
  *   Whether the pool is for counter that was allocated by batch command.
+ * @param[in] age
+ *   Whether the pool is for counter that was allocated for aging.
  *
  * @return
- *   The counter container pointer and @p cnt_free is set on success,
+ *   The counter pool pointer and @p cnt_free is set on success,
  *   NULL otherwise and rte_errno is set.
  */
-static struct mlx5_pools_container *
+static struct mlx5_flow_counter_pool *
 flow_dv_counter_pool_prepare(struct rte_eth_dev *dev,
                             struct mlx5_flow_counter **cnt_free,
-                            uint32_t batch)
+                            uint32_t batch, uint32_t age)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
        struct mlx5_pools_container *cont;
@@ -4237,7 +4401,7 @@ flow_dv_counter_pool_prepare(struct rte_eth_dev *dev,
        struct mlx5_flow_counter *cnt;
        uint32_t i;
 
-       cont = MLX5_CNT_CONTAINER(priv->sh, batch, 0);
+       cont = MLX5_CNT_CONTAINER(priv->sh, batch, age);
        if (!batch) {
                /* bulk_bitmap must be 0 for single counter allocation. */
                dcs = mlx5_devx_cmd_flow_counter_alloc(priv->sh->ctx, 0);
@@ -4245,22 +4409,23 @@ flow_dv_counter_pool_prepare(struct rte_eth_dev *dev,
                        return NULL;
                pool = flow_dv_find_pool_by_id(cont, dcs->id);
                if (!pool) {
-                       cont = flow_dv_pool_create(dev, dcs, batch);
-                       if (!cont) {
+                       pool = flow_dv_pool_create(dev, dcs, batch, age);
+                       if (!pool) {
                                mlx5_devx_cmd_destroy(dcs);
                                return NULL;
                        }
-                       pool = TAILQ_FIRST(&cont->pool_list);
                } else if (dcs->id < pool->min_dcs->id) {
                        rte_atomic64_set(&pool->a64_dcs,
                                         (int64_t)(uintptr_t)dcs);
                }
+               flow_dv_counter_update_min_dcs(dev,
+                                               pool, batch, age);
                i = dcs->id % MLX5_COUNTERS_PER_POOL;
-               cnt = &pool->counters_raw[i];
+               cnt = MLX5_POOL_GET_CNT(pool, i);
                TAILQ_INSERT_HEAD(&pool->counters, cnt, next);
                MLX5_GET_POOL_CNT_EXT(pool, i)->dcs = dcs;
                *cnt_free = cnt;
-               return cont;
+               return pool;
        }
        /* bulk_bitmap is in 128 counters units. */
        if (priv->config.hca_attr.flow_counter_bulk_alloc_bitmap & 0x4)
@@ -4269,18 +4434,17 @@ flow_dv_counter_pool_prepare(struct rte_eth_dev *dev,
                rte_errno = ENODATA;
                return NULL;
        }
-       cont = flow_dv_pool_create(dev, dcs, batch);
-       if (!cont) {
+       pool = flow_dv_pool_create(dev, dcs, batch, age);
+       if (!pool) {
                mlx5_devx_cmd_destroy(dcs);
                return NULL;
        }
-       pool = TAILQ_FIRST(&cont->pool_list);
        for (i = 0; i < MLX5_COUNTERS_PER_POOL; ++i) {
-               cnt = &pool->counters_raw[i];
+               cnt = MLX5_POOL_GET_CNT(pool, i);
                TAILQ_INSERT_HEAD(&pool->counters, cnt, next);
        }
-       *cnt_free = &pool->counters_raw[0];
-       return cont;
+       *cnt_free = MLX5_POOL_GET_CNT(pool, 0);
+       return pool;
 }
 
 /**
@@ -4330,13 +4494,15 @@ flow_dv_counter_shared_search(struct mlx5_pools_container *cont, uint32_t id,
  *   Counter identifier.
  * @param[in] group
  *   Counter flow group.
+ * @param[in] age
+ *   Whether the counter was allocated for aging.
  *
  * @return
  *   Index to flow counter on success, 0 otherwise and rte_errno is set.
  */
 static uint32_t
 flow_dv_counter_alloc(struct rte_eth_dev *dev, uint32_t shared, uint32_t id,
-                     uint16_t group)
+                     uint16_t group, uint32_t age)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
        struct mlx5_flow_counter_pool *pool = NULL;
@@ -4352,7 +4518,7 @@ flow_dv_counter_alloc(struct rte_eth_dev *dev, uint32_t shared, uint32_t id,
         */
        uint32_t batch = (group && !shared && !priv->counter_fallback) ? 1 : 0;
        struct mlx5_pools_container *cont = MLX5_CNT_CONTAINER(priv->sh, batch,
-                                                              0);
+                                                              age);
        uint32_t cnt_idx;
 
        if (!priv->config.devx) {
@@ -4391,10 +4557,9 @@ flow_dv_counter_alloc(struct rte_eth_dev *dev, uint32_t shared, uint32_t id,
                cnt_free = NULL;
        }
        if (!cnt_free) {
-               cont = flow_dv_counter_pool_prepare(dev, &cnt_free, batch);
-               if (!cont)
+               pool = flow_dv_counter_pool_prepare(dev, &cnt_free, batch, age);
+               if (!pool)
                        return 0;
-               pool = TAILQ_FIRST(&cont->pool_list);
        }
        if (!batch)
                cnt_ext = MLX5_CNT_TO_CNT_EXT(pool, cnt_free);
@@ -4404,7 +4569,7 @@ flow_dv_counter_alloc(struct rte_eth_dev *dev, uint32_t shared, uint32_t id,
                struct mlx5_devx_obj *dcs;
 
                if (batch) {
-                       offset = cnt_free - &pool->counters_raw[0];
+                       offset = MLX5_CNT_ARRAY_IDX(pool, cnt_free);
                        dcs = pool->min_dcs;
                } else {
                        offset = 0;
@@ -4418,8 +4583,9 @@ flow_dv_counter_alloc(struct rte_eth_dev *dev, uint32_t shared, uint32_t id,
                }
        }
        cnt_idx = MLX5_MAKE_CNT_IDX(pool->index,
-                                   (cnt_free - pool->counters_raw));
+                               MLX5_CNT_ARRAY_IDX(pool, cnt_free));
        cnt_idx += batch * MLX5_CNT_BATCH_OFFSET;
+       cnt_idx += age * MLX5_CNT_AGE_OFFSET;
        /* Update the counter reset values. */
        if (_flow_dv_query_count(dev, cnt_idx, &cnt_free->hits,
                                 &cnt_free->bytes))
@@ -4441,6 +4607,64 @@ flow_dv_counter_alloc(struct rte_eth_dev *dev, uint32_t shared, uint32_t id,
        return cnt_idx;
 }
 
+/**
+ * Get age param from counter index.
+ *
+ * @param[in] dev
+ *   Pointer to the Ethernet device structure.
+ * @param[in] counter
+ *   Index to the counter handler.
+ *
+ * @return
+ *   The aging parameter specified for the counter index.
+ */
+static struct mlx5_age_param*
+flow_dv_counter_idx_get_age(struct rte_eth_dev *dev,
+                               uint32_t counter)
+{
+       struct mlx5_flow_counter *cnt;
+       struct mlx5_flow_counter_pool *pool = NULL;
+
+       flow_dv_counter_get_by_idx(dev, counter, &pool);
+       counter = (counter - 1) % MLX5_COUNTERS_PER_POOL;
+       cnt = MLX5_POOL_GET_CNT(pool, counter);
+       return MLX5_CNT_TO_AGE(cnt);
+}
+
+/**
+ * Remove a flow counter from aged counter list.
+ *
+ * @param[in] dev
+ *   Pointer to the Ethernet device structure.
+ * @param[in] counter
+ *   Index to the counter handler.
+ * @param[in] cnt
+ *   Pointer to the counter handler.
+ */
+static void
+flow_dv_counter_remove_from_age(struct rte_eth_dev *dev,
+                               uint32_t counter, struct mlx5_flow_counter *cnt)
+{
+       struct mlx5_age_info *age_info;
+       struct mlx5_age_param *age_param;
+       struct mlx5_priv *priv = dev->data->dev_private;
+
+       age_info = GET_PORT_AGE_INFO(priv);
+       age_param = flow_dv_counter_idx_get_age(dev, counter);
+       if (rte_atomic16_cmpset((volatile uint16_t *)
+                       &age_param->state,
+                       AGE_CANDIDATE, AGE_FREE)
+                       != AGE_CANDIDATE) {
+               /**
+                * We need the lock even it is age timeout,
+                * since counter may still in process.
+                */
+               rte_spinlock_lock(&age_info->aged_sl);
+               TAILQ_REMOVE(&age_info->aged_counters, cnt, next);
+               rte_spinlock_unlock(&age_info->aged_sl);
+       }
+       rte_atomic16_set(&age_param->state, AGE_FREE);
+}
 /**
  * Release a flow counter.
  *
@@ -4465,6 +4689,8 @@ flow_dv_counter_release(struct rte_eth_dev *dev, uint32_t counter)
                if (cnt_ext && --cnt_ext->ref_cnt)
                        return;
        }
+       if (IS_AGE_POOL(pool))
+               flow_dv_counter_remove_from_age(dev, counter, cnt);
        /* Put the counter in the end - the last updated one. */
        TAILQ_INSERT_TAIL(&pool->counters, cnt, next);
        /*
@@ -4490,7 +4716,9 @@ flow_dv_counter_release(struct rte_eth_dev *dev, uint32_t counter)
  *   Pointer to error structure.
  *
  * @return
- *   0 on success, a negative errno value otherwise and rte_errno is set.
+ *   - 0 on success and non root table.
+ *   - 1 on success and root table.
+ *   - a negative errno value otherwise and rte_errno is set.
  */
 static int
 flow_dv_validate_attributes(struct rte_eth_dev *dev,
@@ -4500,6 +4728,7 @@ flow_dv_validate_attributes(struct rte_eth_dev *dev,
 {
        struct mlx5_priv *priv = dev->data->dev_private;
        uint32_t priority_max = priv->config.flow_prio - 1;
+       int ret = 0;
 
 #ifndef HAVE_MLX5DV_DR
        if (attributes->group)
@@ -4508,14 +4737,15 @@ flow_dv_validate_attributes(struct rte_eth_dev *dev,
                                          NULL,
                                          "groups are not supported");
 #else
-       uint32_t table;
-       int ret;
+       uint32_t table = 0;
 
        ret = mlx5_flow_group_to_table(attributes, external,
                                       attributes->group, !!priv->fdb_def_rule,
                                       &table, error);
        if (ret)
                return ret;
+       if (!table)
+               ret = MLX5DV_DR_ACTION_FLAGS_ROOT_LEVEL;
 #endif
        if (attributes->priority != MLX5_FLOW_PRIO_RSVD &&
            attributes->priority >= priority_max)
@@ -4545,7 +4775,7 @@ flow_dv_validate_attributes(struct rte_eth_dev *dev,
                                          RTE_FLOW_ERROR_TYPE_ATTR, NULL,
                                          "must specify exactly one of "
                                          "ingress or egress");
-       return 0;
+       return ret;
 }
 
 /**
@@ -4561,6 +4791,8 @@ flow_dv_validate_attributes(struct rte_eth_dev *dev,
  *   Pointer to the list of actions.
  * @param[in] external
  *   This flow rule is created by request external to PMD.
+ * @param[in] hairpin
+ *   Number of hairpin TX actions, 0 means classic flow.
  * @param[out] error
  *   Pointer to the error structure.
  *
@@ -4571,7 +4803,7 @@ static int
 flow_dv_validate(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, struct rte_flow_error *error)
+                bool external, int hairpin, struct rte_flow_error *error)
 {
        int ret;
        uint64_t action_flags = 0;
@@ -4618,12 +4850,15 @@ flow_dv_validate(struct rte_eth_dev *dev, const struct rte_flow_attr *attr,
        struct mlx5_dev_config *dev_conf = &priv->config;
        uint16_t queue_index = 0xFFFF;
        const struct rte_flow_item_vlan *vlan_m = NULL;
+       int16_t rw_act_num = 0;
+       uint64_t is_root;
 
        if (items == NULL)
                return -1;
        ret = flow_dv_validate_attributes(dev, attr, external, error);
        if (ret < 0)
                return ret;
+       is_root = (uint64_t)ret;
        for (; items->type != RTE_FLOW_ITEM_TYPE_END; items++) {
                int tunnel = !!(item_flags & MLX5_FLOW_LAYER_TUNNEL);
                int type = items->type;
@@ -4658,8 +4893,8 @@ flow_dv_validate(struct rte_eth_dev *dev, const struct rte_flow_attr *attr,
                        }
                        break;
                case RTE_FLOW_ITEM_TYPE_VLAN:
-                       ret = mlx5_flow_validate_item_vlan(items, item_flags,
-                                                          dev, error);
+                       ret = flow_dv_validate_item_vlan(items, item_flags,
+                                                        dev, error);
                        if (ret < 0)
                                return ret;
                        last_item = tunnel ? MLX5_FLOW_LAYER_INNER_VLAN :
@@ -4900,6 +5135,7 @@ flow_dv_validate(struct rte_eth_dev *dev, const struct rte_flow_attr *attr,
                                action_flags |= MLX5_FLOW_ACTION_FLAG;
                                ++actions_n;
                        }
+                       rw_act_num += MLX5_ACT_NUM_SET_MARK;
                        break;
                case RTE_FLOW_ACTION_TYPE_MARK:
                        ret = flow_dv_validate_action_mark(dev, actions,
@@ -4918,6 +5154,7 @@ flow_dv_validate(struct rte_eth_dev *dev, const struct rte_flow_attr *attr,
                                action_flags |= MLX5_FLOW_ACTION_MARK;
                                ++actions_n;
                        }
+                       rw_act_num += MLX5_ACT_NUM_SET_MARK;
                        break;
                case RTE_FLOW_ACTION_TYPE_SET_META:
                        ret = flow_dv_validate_action_set_meta(dev, actions,
@@ -4929,6 +5166,7 @@ flow_dv_validate(struct rte_eth_dev *dev, const struct rte_flow_attr *attr,
                        if (!(action_flags & MLX5_FLOW_MODIFY_HDR_ACTIONS))
                                ++actions_n;
                        action_flags |= MLX5_FLOW_ACTION_SET_META;
+                       rw_act_num += MLX5_ACT_NUM_SET_META;
                        break;
                case RTE_FLOW_ACTION_TYPE_SET_TAG:
                        ret = flow_dv_validate_action_set_tag(dev, actions,
@@ -4940,6 +5178,7 @@ flow_dv_validate(struct rte_eth_dev *dev, const struct rte_flow_attr *attr,
                        if (!(action_flags & MLX5_FLOW_MODIFY_HDR_ACTIONS))
                                ++actions_n;
                        action_flags |= MLX5_FLOW_ACTION_SET_TAG;
+                       rw_act_num += MLX5_ACT_NUM_SET_TAG;
                        break;
                case RTE_FLOW_ACTION_TYPE_DROP:
                        ret = mlx5_flow_validate_action_drop(action_flags,
@@ -5017,6 +5256,7 @@ flow_dv_validate(struct rte_eth_dev *dev, const struct rte_flow_attr *attr,
                                return ret;
                        /* Count VID with push_vlan command. */
                        action_flags |= MLX5_FLOW_ACTION_OF_SET_VLAN_VID;
+                       rw_act_num += MLX5_ACT_NUM_MDF_VID;
                        break;
                case RTE_FLOW_ACTION_TYPE_VXLAN_ENCAP:
                case RTE_FLOW_ACTION_TYPE_NVGRE_ENCAP:
@@ -5078,8 +5318,15 @@ flow_dv_validate(struct rte_eth_dev *dev, const struct rte_flow_attr *attr,
                                        RTE_FLOW_ACTION_TYPE_SET_MAC_SRC ?
                                                MLX5_FLOW_ACTION_SET_MAC_SRC :
                                                MLX5_FLOW_ACTION_SET_MAC_DST;
+                       /*
+                        * Even if the source and destination MAC addresses have
+                        * overlap in the header with 4B alignment, the convert
+                        * function will handle them separately and 4 SW actions
+                        * will be created. And 2 actions will be added each
+                        * time no matter how many bytes of address will be set.
+                        */
+                       rw_act_num += MLX5_ACT_NUM_MDF_MAC;
                        break;
-
                case RTE_FLOW_ACTION_TYPE_SET_IPV4_SRC:
                case RTE_FLOW_ACTION_TYPE_SET_IPV4_DST:
                        ret = flow_dv_validate_action_modify_ipv4(action_flags,
@@ -5095,6 +5342,7 @@ flow_dv_validate(struct rte_eth_dev *dev, const struct rte_flow_attr *attr,
                                        RTE_FLOW_ACTION_TYPE_SET_IPV4_SRC ?
                                                MLX5_FLOW_ACTION_SET_IPV4_SRC :
                                                MLX5_FLOW_ACTION_SET_IPV4_DST;
+                       rw_act_num += MLX5_ACT_NUM_MDF_IPV4;
                        break;
                case RTE_FLOW_ACTION_TYPE_SET_IPV6_SRC:
                case RTE_FLOW_ACTION_TYPE_SET_IPV6_DST:
@@ -5117,6 +5365,7 @@ flow_dv_validate(struct rte_eth_dev *dev, const struct rte_flow_attr *attr,
                                        RTE_FLOW_ACTION_TYPE_SET_IPV6_SRC ?
                                                MLX5_FLOW_ACTION_SET_IPV6_SRC :
                                                MLX5_FLOW_ACTION_SET_IPV6_DST;
+                       rw_act_num += MLX5_ACT_NUM_MDF_IPV6;
                        break;
                case RTE_FLOW_ACTION_TYPE_SET_TP_SRC:
                case RTE_FLOW_ACTION_TYPE_SET_TP_DST:
@@ -5133,6 +5382,7 @@ flow_dv_validate(struct rte_eth_dev *dev, const struct rte_flow_attr *attr,
                                        RTE_FLOW_ACTION_TYPE_SET_TP_SRC ?
                                                MLX5_FLOW_ACTION_SET_TP_SRC :
                                                MLX5_FLOW_ACTION_SET_TP_DST;
+                       rw_act_num += MLX5_ACT_NUM_MDF_PORT;
                        break;
                case RTE_FLOW_ACTION_TYPE_DEC_TTL:
                case RTE_FLOW_ACTION_TYPE_SET_TTL:
@@ -5149,6 +5399,7 @@ flow_dv_validate(struct rte_eth_dev *dev, const struct rte_flow_attr *attr,
                                        RTE_FLOW_ACTION_TYPE_SET_TTL ?
                                                MLX5_FLOW_ACTION_SET_TTL :
                                                MLX5_FLOW_ACTION_DEC_TTL;
+                       rw_act_num += MLX5_ACT_NUM_MDF_TTL;
                        break;
                case RTE_FLOW_ACTION_TYPE_JUMP:
                        ret = flow_dv_validate_action_jump(actions,
@@ -5176,6 +5427,7 @@ flow_dv_validate(struct rte_eth_dev *dev, const struct rte_flow_attr *attr,
                                        RTE_FLOW_ACTION_TYPE_INC_TCP_SEQ ?
                                                MLX5_FLOW_ACTION_INC_TCP_SEQ :
                                                MLX5_FLOW_ACTION_DEC_TCP_SEQ;
+                       rw_act_num += MLX5_ACT_NUM_MDF_TCPSEQ;
                        break;
                case RTE_FLOW_ACTION_TYPE_INC_TCP_ACK:
                case RTE_FLOW_ACTION_TYPE_DEC_TCP_ACK:
@@ -5193,10 +5445,13 @@ flow_dv_validate(struct rte_eth_dev *dev, const struct rte_flow_attr *attr,
                                        RTE_FLOW_ACTION_TYPE_INC_TCP_ACK ?
                                                MLX5_FLOW_ACTION_INC_TCP_ACK :
                                                MLX5_FLOW_ACTION_DEC_TCP_ACK;
+                       rw_act_num += MLX5_ACT_NUM_MDF_TCPACK;
                        break;
-               case MLX5_RTE_FLOW_ACTION_TYPE_TAG:
                case MLX5_RTE_FLOW_ACTION_TYPE_MARK:
+                       break;
+               case MLX5_RTE_FLOW_ACTION_TYPE_TAG:
                case MLX5_RTE_FLOW_ACTION_TYPE_COPY_MREG:
+                       rw_act_num += MLX5_ACT_NUM_SET_TAG;
                        break;
                case RTE_FLOW_ACTION_TYPE_METER:
                        ret = mlx5_flow_validate_action_meter(dev,
@@ -5207,6 +5462,17 @@ flow_dv_validate(struct rte_eth_dev *dev, const struct rte_flow_attr *attr,
                                return ret;
                        action_flags |= MLX5_FLOW_ACTION_METER;
                        ++actions_n;
+                       /* Meter action will add one more TAG action. */
+                       rw_act_num += MLX5_ACT_NUM_SET_TAG;
+                       break;
+               case RTE_FLOW_ACTION_TYPE_AGE:
+                       ret = flow_dv_validate_action_age(action_flags,
+                                                         actions, dev,
+                                                         error);
+                       if (ret < 0)
+                               return ret;
+                       action_flags |= MLX5_FLOW_ACTION_AGE;
+                       ++actions_n;
                        break;
                case RTE_FLOW_ACTION_TYPE_SET_IPV4_DSCP:
                        ret = flow_dv_validate_action_modify_ipv4_dscp
@@ -5220,6 +5486,7 @@ flow_dv_validate(struct rte_eth_dev *dev, const struct rte_flow_attr *attr,
                        if (!(action_flags & MLX5_FLOW_MODIFY_HDR_ACTIONS))
                                ++actions_n;
                        action_flags |= MLX5_FLOW_ACTION_SET_IPV4_DSCP;
+                       rw_act_num += MLX5_ACT_NUM_SET_DSCP;
                        break;
                case RTE_FLOW_ACTION_TYPE_SET_IPV6_DSCP:
                        ret = flow_dv_validate_action_modify_ipv6_dscp
@@ -5233,6 +5500,7 @@ flow_dv_validate(struct rte_eth_dev *dev, const struct rte_flow_attr *attr,
                        if (!(action_flags & MLX5_FLOW_MODIFY_HDR_ACTIONS))
                                ++actions_n;
                        action_flags |= MLX5_FLOW_ACTION_SET_IPV6_DSCP;
+                       rw_act_num += MLX5_ACT_NUM_SET_DSCP;
                        break;
                default:
                        return rte_flow_error_set(error, ENOTSUP,
@@ -5305,6 +5573,21 @@ flow_dv_validate(struct rte_eth_dev *dev, const struct rte_flow_attr *attr,
                                                  NULL, "encap is not supported"
                                                  " for ingress traffic");
        }
+       /* Hairpin flow will add one more TAG action. */
+       if (hairpin > 0)
+               rw_act_num += MLX5_ACT_NUM_SET_TAG;
+       /* extra metadata enabled: one more TAG action will be add. */
+       if (dev_conf->dv_flow_en &&
+           dev_conf->dv_xmeta_en != MLX5_XMETA_MODE_LEGACY &&
+           mlx5_flow_ext_mreg_supported(dev))
+               rw_act_num += MLX5_ACT_NUM_SET_TAG;
+       if ((uint32_t)rw_act_num >
+                       flow_dv_modify_hdr_action_max(dev, is_root)) {
+               return rte_flow_error_set(error, ENOTSUP,
+                                         RTE_FLOW_ERROR_TYPE_ACTION,
+                                         NULL, "too many header modify"
+                                         " actions to support");
+       }
        return 0;
 }
 
@@ -5403,6 +5686,34 @@ flow_dv_check_valid_spec(void *match_mask, void *match_value)
 }
 #endif
 
+/**
+ * Add match of ip_version.
+ *
+ * @param[in] group
+ *   Flow group.
+ * @param[in] headers_v
+ *   Values header pointer.
+ * @param[in] headers_m
+ *   Masks header pointer.
+ * @param[in] ip_version
+ *   The IP version to set.
+ */
+static inline void
+flow_dv_set_match_ip_version(uint32_t group,
+                            void *headers_v,
+                            void *headers_m,
+                            uint8_t ip_version)
+{
+       if (group == 0)
+               MLX5_SET(fte_match_set_lyr_2_4, headers_m, ip_version, 0xf);
+       else
+               MLX5_SET(fte_match_set_lyr_2_4, headers_m, ip_version,
+                        ip_version);
+       MLX5_SET(fte_match_set_lyr_2_4, headers_v, ip_version, ip_version);
+       MLX5_SET(fte_match_set_lyr_2_4, headers_v, ethertype, 0);
+       MLX5_SET(fte_match_set_lyr_2_4, headers_m, ethertype, 0);
+}
+
 /**
  * Add Ethernet item to matcher and to the value.
  *
@@ -5417,7 +5728,8 @@ flow_dv_check_valid_spec(void *match_mask, void *match_value)
  */
 static void
 flow_dv_translate_item_eth(void *matcher, void *key,
-                          const struct rte_flow_item *item, int inner)
+                          const struct rte_flow_item *item, int inner,
+                          uint32_t group)
 {
        const struct rte_flow_item_eth *eth_m = item->mask;
        const struct rte_flow_item_eth *eth_v = item->spec;
@@ -5472,11 +5784,22 @@ flow_dv_translate_item_eth(void *matcher, void *key,
         * HW supports match on one Ethertype, the Ethertype following the last
         * VLAN tag of the packet (see PRM).
         * Set match on ethertype only if ETH header is not followed by VLAN.
+        * HW is optimized for IPv4/IPv6. In such cases, avoid setting
+        * ethertype, and use ip_version field instead.
         */
-       MLX5_SET(fte_match_set_lyr_2_4, headers_m, ethertype,
-                rte_be_to_cpu_16(eth_m->type));
-       l24_v = MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v, ethertype);
-       *(uint16_t *)(l24_v) = eth_m->type & eth_v->type;
+       if (eth_v->type == RTE_BE16(RTE_ETHER_TYPE_IPV4) &&
+           eth_m->type == 0xFFFF) {
+               flow_dv_set_match_ip_version(group, headers_v, headers_m, 4);
+       } else if (eth_v->type == RTE_BE16(RTE_ETHER_TYPE_IPV6) &&
+                  eth_m->type == 0xFFFF) {
+               flow_dv_set_match_ip_version(group, headers_v, headers_m, 6);
+       } else {
+               MLX5_SET(fte_match_set_lyr_2_4, headers_m, ethertype,
+                        rte_be_to_cpu_16(eth_m->type));
+               l24_v = MLX5_ADDR_OF(fte_match_set_lyr_2_4, headers_v,
+                                    ethertype);
+               *(uint16_t *)(l24_v) = eth_m->type & eth_v->type;
+       }
 }
 
 /**
@@ -5497,7 +5820,7 @@ static void
 flow_dv_translate_item_vlan(struct mlx5_flow *dev_flow,
                            void *matcher, void *key,
                            const struct rte_flow_item *item,
-                           int inner)
+                           int inner, uint32_t group)
 {
        const struct rte_flow_item_vlan *vlan_m = item->mask;
        const struct rte_flow_item_vlan *vlan_v = item->spec;
@@ -5506,10 +5829,6 @@ flow_dv_translate_item_vlan(struct mlx5_flow *dev_flow,
        uint16_t tci_m;
        uint16_t tci_v;
 
-       if (!vlan_v)
-               return;
-       if (!vlan_m)
-               vlan_m = &rte_flow_item_vlan_mask;
        if (inner) {
                headers_m = MLX5_ADDR_OF(fte_match_param, matcher,
                                         inner_headers);
@@ -5522,23 +5841,45 @@ flow_dv_translate_item_vlan(struct mlx5_flow *dev_flow,
                 * This is workaround, masks are not supported,
                 * and pre-validated.
                 */
-               dev_flow->handle->vf_vlan.tag =
-                       rte_be_to_cpu_16(vlan_v->tci) & 0x0fff;
+               if (vlan_v)
+                       dev_flow->handle->vf_vlan.tag =
+                                       rte_be_to_cpu_16(vlan_v->tci) & 0x0fff;
        }
-       tci_m = rte_be_to_cpu_16(vlan_m->tci);
-       tci_v = rte_be_to_cpu_16(vlan_m->tci & vlan_v->tci);
+       /*
+        * When VLAN item exists in flow, mark packet as tagged,
+        * even if TCI is not specified.
+        */
        MLX5_SET(fte_match_set_lyr_2_4, headers_m, cvlan_tag, 1);
        MLX5_SET(fte_match_set_lyr_2_4, headers_v, cvlan_tag, 1);
+       if (!vlan_v)
+               return;
+       if (!vlan_m)
+               vlan_m = &rte_flow_item_vlan_mask;
+       tci_m = rte_be_to_cpu_16(vlan_m->tci);
+       tci_v = rte_be_to_cpu_16(vlan_m->tci & vlan_v->tci);
        MLX5_SET(fte_match_set_lyr_2_4, headers_m, first_vid, tci_m);
        MLX5_SET(fte_match_set_lyr_2_4, headers_v, first_vid, tci_v);
        MLX5_SET(fte_match_set_lyr_2_4, headers_m, first_cfi, tci_m >> 12);
        MLX5_SET(fte_match_set_lyr_2_4, headers_v, first_cfi, tci_v >> 12);
        MLX5_SET(fte_match_set_lyr_2_4, headers_m, first_prio, tci_m >> 13);
        MLX5_SET(fte_match_set_lyr_2_4, headers_v, first_prio, tci_v >> 13);
-       MLX5_SET(fte_match_set_lyr_2_4, headers_m, ethertype,
-                rte_be_to_cpu_16(vlan_m->inner_type));
-       MLX5_SET(fte_match_set_lyr_2_4, headers_v, ethertype,
-                rte_be_to_cpu_16(vlan_m->inner_type & vlan_v->inner_type));
+       /*
+        * HW is optimized for IPv4/IPv6. In such cases, avoid setting
+        * ethertype, and use ip_version field instead.
+        */
+       if (vlan_v->inner_type == RTE_BE16(RTE_ETHER_TYPE_IPV4) &&
+           vlan_m->inner_type == 0xFFFF) {
+               flow_dv_set_match_ip_version(group, headers_v, headers_m, 4);
+       } else if (vlan_v->inner_type == RTE_BE16(RTE_ETHER_TYPE_IPV6) &&
+                  vlan_m->inner_type == 0xFFFF) {
+               flow_dv_set_match_ip_version(group, headers_v, headers_m, 6);
+       } else {
+               MLX5_SET(fte_match_set_lyr_2_4, headers_m, ethertype,
+                        rte_be_to_cpu_16(vlan_m->inner_type));
+               MLX5_SET(fte_match_set_lyr_2_4, headers_v, ethertype,
+                        rte_be_to_cpu_16(vlan_m->inner_type &
+                                         vlan_v->inner_type));
+       }
 }
 
 /**
@@ -5589,11 +5930,7 @@ flow_dv_translate_item_ipv4(void *matcher, void *key,
                                         outer_headers);
                headers_v = MLX5_ADDR_OF(fte_match_param, key, outer_headers);
        }
-       if (group == 0)
-               MLX5_SET(fte_match_set_lyr_2_4, headers_m, ip_version, 0xf);
-       else
-               MLX5_SET(fte_match_set_lyr_2_4, headers_m, ip_version, 0x4);
-       MLX5_SET(fte_match_set_lyr_2_4, headers_v, ip_version, 4);
+       flow_dv_set_match_ip_version(group, headers_v, headers_m, 4);
        /*
         * On outer header (which must contains L2), or inner header with L2,
         * set cvlan_tag mask bit to mark this packet as untagged.
@@ -5691,11 +6028,7 @@ flow_dv_translate_item_ipv6(void *matcher, void *key,
                                         outer_headers);
                headers_v = MLX5_ADDR_OF(fte_match_param, key, outer_headers);
        }
-       if (group == 0)
-               MLX5_SET(fte_match_set_lyr_2_4, headers_m, ip_version, 0xf);
-       else
-               MLX5_SET(fte_match_set_lyr_2_4, headers_m, ip_version, 0x6);
-       MLX5_SET(fte_match_set_lyr_2_4, headers_v, ip_version, 6);
+       flow_dv_set_match_ip_version(group, headers_v, headers_m, 6);
        /*
         * On outer header (which must contains L2), or inner header with L2,
         * set cvlan_tag mask bit to mark this packet as untagged.
@@ -6794,6 +7127,10 @@ flow_dv_translate_item_gtp(void *matcher, void *key,
                return;
        if (!gtp_m)
                gtp_m = &rte_flow_item_gtp_mask;
+       MLX5_SET(fte_match_set_misc3, misc3_m, gtpu_msg_flags,
+                gtp_m->v_pt_rsv_flags);
+       MLX5_SET(fte_match_set_misc3, misc3_v, gtpu_msg_flags,
+                gtp_v->v_pt_rsv_flags & gtp_m->v_pt_rsv_flags);
        MLX5_SET(fte_match_set_misc3, misc3_m, gtpu_msg_type, gtp_m->msg_type);
        MLX5_SET(fte_match_set_misc3, misc3_v, gtpu_msg_type,
                 gtp_v->msg_type & gtp_m->msg_type);
@@ -7229,6 +7566,53 @@ flow_dv_translate_action_port_id(struct rte_eth_dev *dev,
        return 0;
 }
 
+/**
+ * Create a counter with aging configuration.
+ *
+ * @param[in] dev
+ *   Pointer to rte_eth_dev structure.
+ * @param[out] count
+ *   Pointer to the counter action configuration.
+ * @param[in] age
+ *   Pointer to the aging action configuration.
+ *
+ * @return
+ *   Index to flow counter on success, 0 otherwise.
+ */
+static uint32_t
+flow_dv_translate_create_counter(struct rte_eth_dev *dev,
+                               struct mlx5_flow *dev_flow,
+                               const struct rte_flow_action_count *count,
+                               const struct rte_flow_action_age *age)
+{
+       uint32_t counter;
+       struct mlx5_age_param *age_param;
+
+       counter = flow_dv_counter_alloc(dev,
+                               count ? count->shared : 0,
+                               count ? count->id : 0,
+                               dev_flow->dv.group, !!age);
+       if (!counter || age == NULL)
+               return counter;
+       age_param  = flow_dv_counter_idx_get_age(dev, counter);
+       /*
+        * The counter age accuracy may have a bit delay. Have 3/4
+        * second bias on the timeount in order to let it age in time.
+        */
+       age_param->context = age->context ? age->context :
+               (void *)(uintptr_t)(dev_flow->flow_idx);
+       /*
+        * The counter age accuracy may have a bit delay. Have 3/4
+        * second bias on the timeount in order to let it age in time.
+        */
+       age_param->timeout = age->timeout * 10 - MLX5_AGING_TIME_DELAY;
+       /* Set expire time in unit of 0.1 sec. */
+       age_param->port_id = dev->data->port_id;
+       age_param->expire = age_param->timeout +
+                       rte_rdtsc() / (rte_get_tsc_hz() / 10);
+       rte_atomic16_set(&age_param->state, AGE_CANDIDATE);
+       return counter;
+}
 /**
  * Add Tx queue matcher
  *
@@ -7398,6 +7782,8 @@ __flow_dv_translate(struct rte_eth_dev *dev,
                            (MLX5_MAX_MODIFY_NUM + 1)];
        } mhdr_dummy;
        struct mlx5_flow_dv_modify_hdr_resource *mhdr_res = &mhdr_dummy.res;
+       const struct rte_flow_action_count *count = NULL;
+       const struct rte_flow_action_age *age = NULL;
        union flow_dv_attr flow_attr = { .attr = 0 };
        uint32_t tag_be;
        union mlx5_flow_tbl_key tbl_key;
@@ -7426,7 +7812,6 @@ __flow_dv_translate(struct rte_eth_dev *dev,
                const struct rte_flow_action_queue *queue;
                const struct rte_flow_action_rss *rss;
                const struct rte_flow_action *action = actions;
-               const struct rte_flow_action_count *count = action->conf;
                const uint8_t *rss_key;
                const struct rte_flow_action_jump *jump_data;
                const struct rte_flow_action_meter *mtr;
@@ -7555,36 +7940,21 @@ __flow_dv_translate(struct rte_eth_dev *dev,
                        action_flags |= MLX5_FLOW_ACTION_RSS;
                        dev_flow->handle->fate_action = MLX5_FLOW_FATE_QUEUE;
                        break;
+               case RTE_FLOW_ACTION_TYPE_AGE:
                case RTE_FLOW_ACTION_TYPE_COUNT:
                        if (!dev_conf->devx) {
-                               rte_errno = ENOTSUP;
-                               goto cnt_err;
-                       }
-                       flow->counter = flow_dv_counter_alloc(dev,
-                                                       count->shared,
-                                                       count->id,
-                                                       dev_flow->dv.group);
-                       if (!flow->counter)
-                               goto cnt_err;
-                       dev_flow->dv.actions[actions_n++] =
-                                 (flow_dv_counter_get_by_idx(dev,
-                                 flow->counter, NULL))->action;
-                       action_flags |= MLX5_FLOW_ACTION_COUNT;
-                       break;
-cnt_err:
-                       if (rte_errno == ENOTSUP)
                                return rte_flow_error_set
                                              (error, ENOTSUP,
                                               RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
                                               NULL,
                                               "count action not supported");
+                       }
+                       /* Save information first, will apply later. */
+                       if (actions->type == RTE_FLOW_ACTION_TYPE_COUNT)
+                               count = action->conf;
                        else
-                               return rte_flow_error_set
-                                               (error, rte_errno,
-                                                RTE_FLOW_ERROR_TYPE_ACTION,
-                                                action,
-                                                "cannot create counter"
-                                                 " object.");
+                               age = action->conf;
+                       action_flags |= MLX5_FLOW_ACTION_COUNT;
                        break;
                case RTE_FLOW_ACTION_TYPE_OF_POP_VLAN:
                        dev_flow->dv.actions[actions_n++] =
@@ -7818,11 +8188,12 @@ cnt_err:
                                                NULL,
                                                "meter not found "
                                                "or invalid parameters");
-                               flow->meter = fm->meter_id;
+                               flow->meter = fm->idx;
                        }
                        /* Set the meter action. */
                        if (!fm) {
-                               fm = mlx5_flow_meter_find(priv, flow->meter);
+                               fm = mlx5_ipool_get(priv->sh->ipool
+                                               [MLX5_IPOOL_MTR], flow->meter);
                                if (!fm)
                                        return rte_flow_error_set(error,
                                                rte_errno,
@@ -7857,6 +8228,22 @@ cnt_err:
                                dev_flow->dv.actions[modify_action_position] =
                                        handle->dvh.modify_hdr->verbs_action;
                        }
+                       if (action_flags & MLX5_FLOW_ACTION_COUNT) {
+                               flow->counter =
+                                       flow_dv_translate_create_counter(dev,
+                                               dev_flow, count, age);
+
+                               if (!flow->counter)
+                                       return rte_flow_error_set
+                                               (error, rte_errno,
+                                               RTE_FLOW_ERROR_TYPE_ACTION,
+                                               NULL,
+                                               "cannot create counter"
+                                               " object.");
+                               dev_flow->dv.actions[actions_n++] =
+                                         (flow_dv_counter_get_by_idx(dev,
+                                         flow->counter, NULL))->action;
+                       }
                        break;
                default:
                        break;
@@ -7879,7 +8266,8 @@ cnt_err:
                        break;
                case RTE_FLOW_ITEM_TYPE_ETH:
                        flow_dv_translate_item_eth(match_mask, match_value,
-                                                  items, tunnel);
+                                                  items, tunnel,
+                                                  dev_flow->dv.group);
                        matcher.priority = MLX5_PRIORITY_MAP_L2;
                        last_item = tunnel ? MLX5_FLOW_LAYER_INNER_L2 :
                                             MLX5_FLOW_LAYER_OUTER_L2;
@@ -7887,7 +8275,8 @@ cnt_err:
                case RTE_FLOW_ITEM_TYPE_VLAN:
                        flow_dv_translate_item_vlan(dev_flow,
                                                    match_mask, match_value,
-                                                   items, tunnel);
+                                                   items, tunnel,
+                                                   dev_flow->dv.group);
                        matcher.priority = MLX5_PRIORITY_MAP_L2;
                        last_item = tunnel ? (MLX5_FLOW_LAYER_INNER_L2 |
                                              MLX5_FLOW_LAYER_INNER_VLAN) :
@@ -8542,7 +8931,8 @@ __flow_dv_destroy(struct rte_eth_dev *dev, struct rte_flow *flow)
        if (flow->meter) {
                struct mlx5_flow_meter *fm;
 
-               fm  = mlx5_flow_meter_find(priv, flow->meter);
+               fm = mlx5_ipool_get(priv->sh->ipool[MLX5_IPOOL_MTR],
+                                   flow->meter);
                if (fm)
                        mlx5_flow_meter_detach(fm);
                flow->meter = 0;
@@ -9117,6 +9507,60 @@ flow_dv_counter_query(struct rte_eth_dev *dev, uint32_t counter, bool clear,
        return 0;
 }
 
+/**
+ * Get aged-out flows.
+ *
+ * @param[in] dev
+ *   Pointer to the Ethernet device structure.
+ * @param[in] context
+ *   The address of an array of pointers to the aged-out flows contexts.
+ * @param[in] nb_contexts
+ *   The length of context array pointers.
+ * @param[out] error
+ *   Perform verbose error reporting if not NULL. Initialized in case of
+ *   error only.
+ *
+ * @return
+ *   how many contexts get in success, otherwise negative errno value.
+ *   if nb_contexts is 0, return the amount of all aged contexts.
+ *   if nb_contexts is not 0 , return the amount of aged flows reported
+ *   in the context array.
+ * @note: only stub for now
+ */
+static int
+flow_get_aged_flows(struct rte_eth_dev *dev,
+                   void **context,
+                   uint32_t nb_contexts,
+                   struct rte_flow_error *error)
+{
+       struct mlx5_priv *priv = dev->data->dev_private;
+       struct mlx5_age_info *age_info;
+       struct mlx5_age_param *age_param;
+       struct mlx5_flow_counter *counter;
+       int nb_flows = 0;
+
+       if (nb_contexts && !context)
+               return rte_flow_error_set(error, EINVAL,
+                                         RTE_FLOW_ERROR_TYPE_UNSPECIFIED,
+                                         NULL,
+                                         "Should assign at least one flow or"
+                                         " context to get if nb_contexts != 0");
+       age_info = GET_PORT_AGE_INFO(priv);
+       rte_spinlock_lock(&age_info->aged_sl);
+       TAILQ_FOREACH(counter, &age_info->aged_counters, next) {
+               nb_flows++;
+               if (nb_contexts) {
+                       age_param = MLX5_CNT_TO_AGE(counter);
+                       context[nb_flows - 1] = age_param->context;
+                       if (!(--nb_contexts))
+                               break;
+               }
+       }
+       rte_spinlock_unlock(&age_info->aged_sl);
+       MLX5_AGE_SET(age_info, MLX5_AGE_TRIGGER);
+       return nb_flows;
+}
+
 /*
  * Mutex-protected thunk to lock-free  __flow_dv_translate().
  */
@@ -9183,7 +9627,7 @@ flow_dv_counter_allocate(struct rte_eth_dev *dev)
        uint32_t cnt;
 
        flow_dv_shared_lock(dev);
-       cnt = flow_dv_counter_alloc(dev, 0, 0, 1);
+       cnt = flow_dv_counter_alloc(dev, 0, 0, 1, 0);
        flow_dv_shared_unlock(dev);
        return cnt;
 }
@@ -9214,6 +9658,7 @@ const struct mlx5_flow_driver_ops mlx5_flow_dv_drv_ops = {
        .counter_alloc = flow_dv_counter_allocate,
        .counter_free = flow_dv_counter_free,
        .counter_query = flow_dv_counter_query,
+       .get_aged_flows = flow_get_aged_flows,
 };
 
 #endif /* HAVE_IBV_FLOW_DV_SUPPORT */