net/mlx5: fix validation of push VLAN without full mask
[dpdk.git] / drivers / net / mlx5 / mlx5_flow_dv.c
index 1c04eeb..ae00ca4 100644 (file)
@@ -73,6 +73,10 @@ union flow_dv_attr {
        uint32_t attr;
 };
 
+static int
+flow_dv_tbl_resource_release(struct rte_eth_dev *dev,
+                            struct mlx5_flow_tbl_resource *tbl);
+
 /**
  * Initialize flow attributes structure according to flow items' types.
  *
@@ -1829,7 +1833,7 @@ flow_dev_get_vlan_info_from_items(const struct rte_flow_item *items,
 static int
 flow_dv_validate_action_push_vlan(struct rte_eth_dev *dev,
                                  uint64_t action_flags,
-                                 uint64_t item_flags __rte_unused,
+                                 const struct rte_flow_item_vlan *vlan_m,
                                  const struct rte_flow_action *action,
                                  const struct rte_flow_attr *attr,
                                  struct rte_flow_error *error)
@@ -1863,6 +1867,32 @@ flow_dv_validate_action_push_vlan(struct rte_eth_dev *dev,
                                          RTE_FLOW_ERROR_TYPE_UNSPECIFIED, NULL,
                                          "push vlan action for VF representor "
                                          "not supported on NIC table");
+       if (vlan_m &&
+           (vlan_m->tci & MLX5DV_FLOW_VLAN_PCP_MASK_BE) &&
+           (vlan_m->tci & MLX5DV_FLOW_VLAN_PCP_MASK_BE) !=
+               MLX5DV_FLOW_VLAN_PCP_MASK_BE &&
+           !(action_flags & MLX5_FLOW_ACTION_OF_SET_VLAN_PCP) &&
+           !(mlx5_flow_find_action
+               (action + 1, RTE_FLOW_ACTION_TYPE_OF_SET_VLAN_PCP)))
+               return rte_flow_error_set(error, EINVAL,
+                                         RTE_FLOW_ERROR_TYPE_ACTION, action,
+                                         "not full match mask on VLAN PCP and "
+                                         "there is no of_set_vlan_pcp action, "
+                                         "push VLAN action cannot figure out "
+                                         "PCP value");
+       if (vlan_m &&
+           (vlan_m->tci & MLX5DV_FLOW_VLAN_VID_MASK_BE) &&
+           (vlan_m->tci & MLX5DV_FLOW_VLAN_VID_MASK_BE) !=
+               MLX5DV_FLOW_VLAN_VID_MASK_BE &&
+           !(action_flags & MLX5_FLOW_ACTION_OF_SET_VLAN_VID) &&
+           !(mlx5_flow_find_action
+               (action + 1, RTE_FLOW_ACTION_TYPE_OF_SET_VLAN_VID)))
+               return rte_flow_error_set(error, EINVAL,
+                                         RTE_FLOW_ERROR_TYPE_ACTION, action,
+                                         "not full match mask on VLAN VID and "
+                                         "there is no of_set_vlan_vid action, "
+                                         "push VLAN action cannot figure out "
+                                         "VID value");
        (void)attr;
        return 0;
 }
@@ -2530,6 +2560,8 @@ flow_dv_jump_tbl_resource_register
                DRV_LOG(DEBUG, "new jump table resource %p: refcnt %d++",
                        (void *)&tbl_data->jump, cnt);
        } else {
+               /* old jump should not make the table ref++. */
+               flow_dv_tbl_resource_release(dev, &tbl_data->tbl);
                MLX5_ASSERT(tbl_data->jump.action);
                DRV_LOG(DEBUG, "existed jump table resource %p: refcnt %d++",
                        (void *)&tbl_data->jump, cnt);
@@ -3820,132 +3852,41 @@ flow_dv_modify_hdr_resource_register
        return 0;
 }
 
-#define MLX5_CNT_CONTAINER_RESIZE 64
-
 /**
- * Get or create a flow counter.
+ * Get DV flow counter by index.
  *
  * @param[in] dev
  *   Pointer to the Ethernet device structure.
- * @param[in] shared
- *   Indicate if this counter is shared with other flows.
- * @param[in] id
- *   Counter identifier.
+ * @param[in] idx
+ *   mlx5 flow counter index in the container.
+ * @param[out] ppool
+ *   mlx5 flow counter pool in the container,
  *
  * @return
- *   pointer to flow counter on success, NULL otherwise and rte_errno is set.
+ *   Pointer to the counter, NULL otherwise.
  */
 static struct mlx5_flow_counter *
-flow_dv_counter_alloc_fallback(struct rte_eth_dev *dev, uint32_t shared,
-                              uint32_t id)
-{
-       struct mlx5_priv *priv = dev->data->dev_private;
-       struct mlx5_flow_counter *cnt = NULL;
-       struct mlx5_devx_obj *dcs = NULL;
-
-       if (!priv->config.devx) {
-               rte_errno = ENOTSUP;
-               return NULL;
-       }
-       if (shared) {
-               TAILQ_FOREACH(cnt, &priv->sh->cmng.flow_counters, next) {
-                       if (cnt->shared && cnt->id == id) {
-                               cnt->ref_cnt++;
-                               return cnt;
-                       }
-               }
-       }
-       dcs = mlx5_devx_cmd_flow_counter_alloc(priv->sh->ctx, 0);
-       if (!dcs)
-               return NULL;
-       cnt = rte_calloc(__func__, 1, sizeof(*cnt), 0);
-       if (!cnt) {
-               claim_zero(mlx5_devx_cmd_destroy(cnt->dcs));
-               rte_errno = ENOMEM;
-               return NULL;
-       }
-       struct mlx5_flow_counter tmpl = {
-               .shared = shared,
-               .ref_cnt = 1,
-               .id = id,
-               .dcs = dcs,
-       };
-       tmpl.action = mlx5_glue->dv_create_flow_action_counter(dcs->obj, 0);
-       if (!tmpl.action) {
-               claim_zero(mlx5_devx_cmd_destroy(cnt->dcs));
-               rte_errno = errno;
-               rte_free(cnt);
-               return NULL;
-       }
-       *cnt = tmpl;
-       TAILQ_INSERT_HEAD(&priv->sh->cmng.flow_counters, cnt, next);
-       return cnt;
-}
-
-/**
- * Release a flow counter.
- *
- * @param[in] dev
- *   Pointer to the Ethernet device structure.
- * @param[in] counter
- *   Pointer to the counter handler.
- */
-static void
-flow_dv_counter_release_fallback(struct rte_eth_dev *dev,
-                                struct mlx5_flow_counter *counter)
+flow_dv_counter_get_by_idx(struct rte_eth_dev *dev,
+                          uint32_t idx,
+                          struct mlx5_flow_counter_pool **ppool)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
+       struct mlx5_pools_container *cont;
+       struct mlx5_flow_counter_pool *pool;
+       uint32_t batch = 0;
 
-       if (!counter)
-               return;
-       if (--counter->ref_cnt == 0) {
-               TAILQ_REMOVE(&priv->sh->cmng.flow_counters, counter, next);
-               claim_zero(mlx5_devx_cmd_destroy(counter->dcs));
-               rte_free(counter);
+       idx--;
+       if (idx >= MLX5_CNT_BATCH_OFFSET) {
+               idx -= MLX5_CNT_BATCH_OFFSET;
+               batch = 1;
        }
-}
-
-/**
- * Query a devx flow counter.
- *
- * @param[in] dev
- *   Pointer to the Ethernet device structure.
- * @param[in] cnt
- *   Pointer to the flow counter.
- * @param[out] pkts
- *   The statistics value of packets.
- * @param[out] bytes
- *   The statistics value of bytes.
- *
- * @return
- *   0 on success, otherwise a negative errno value and rte_errno is set.
- */
-static inline int
-_flow_dv_query_count_fallback(struct rte_eth_dev *dev __rte_unused,
-                    struct mlx5_flow_counter *cnt, uint64_t *pkts,
-                    uint64_t *bytes)
-{
-       return mlx5_devx_cmd_flow_counter_query(cnt->dcs, 0, 0, pkts, bytes,
-                                               0, NULL, NULL, 0);
-}
-
-/**
- * Get a pool by a counter.
- *
- * @param[in] cnt
- *   Pointer to the counter.
- *
- * @return
- *   The counter pool.
- */
-static struct mlx5_flow_counter_pool *
-flow_dv_counter_pool_get(struct mlx5_flow_counter *cnt)
-{
-       if (!cnt->batch) {
-               cnt -= cnt->dcs->id % MLX5_COUNTERS_PER_POOL;
-               return (struct mlx5_flow_counter_pool *)cnt - 1;
-       }
-       return cnt->pool;
+       cont = MLX5_CNT_CONTAINER(priv->sh, batch, 0);
+       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];
 }
 
 /**
@@ -3962,15 +3903,26 @@ flow_dv_counter_pool_get(struct mlx5_flow_counter *cnt)
 static struct mlx5_flow_counter_pool *
 flow_dv_find_pool_by_id(struct mlx5_pools_container *cont, int id)
 {
-       struct mlx5_flow_counter_pool *pool;
+       uint32_t i;
+       uint32_t n_valid = rte_atomic16_read(&cont->n_valid);
 
-       TAILQ_FOREACH(pool, &cont->pool_list, next) {
+       for (i = 0; i < n_valid; i++) {
+               struct mlx5_flow_counter_pool *pool = cont->pools[i];
                int base = (pool->min_dcs->id / MLX5_COUNTERS_PER_POOL) *
-                               MLX5_COUNTERS_PER_POOL;
+                          MLX5_COUNTERS_PER_POOL;
 
-               if (id >= base && id < base + MLX5_COUNTERS_PER_POOL)
+               if (id >= base && id < base + MLX5_COUNTERS_PER_POOL) {
+                       /*
+                        * Move the pool to the head, as counter allocate
+                        * always gets the first pool in the container.
+                        */
+                       if (pool != TAILQ_FIRST(&cont->pool_list)) {
+                               TAILQ_REMOVE(&cont->pool_list, pool, next);
+                               TAILQ_INSERT_HEAD(&cont->pool_list, pool, next);
+                       }
                        return pool;
-       };
+               }
+       }
        return NULL;
 }
 
@@ -4060,12 +4012,14 @@ flow_dv_container_resize(struct rte_eth_dev *dev, uint32_t batch)
                        MLX5_CNT_CONTAINER(priv->sh, batch, 0);
        struct mlx5_pools_container *new_cont =
                        MLX5_CNT_CONTAINER_UNUSED(priv->sh, batch, 0);
-       struct mlx5_counter_stats_mem_mng *mem_mng;
+       struct mlx5_counter_stats_mem_mng *mem_mng = NULL;
        uint32_t resize = cont->n + MLX5_CNT_CONTAINER_RESIZE;
        uint32_t mem_size = sizeof(struct mlx5_flow_counter_pool *) * resize;
        int i;
 
-       if (cont != MLX5_CNT_CONTAINER(priv->sh, batch, 1)) {
+       /* 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;
@@ -4078,16 +4032,29 @@ flow_dv_container_resize(struct rte_eth_dev *dev, uint32_t batch)
        if (cont->n)
                memcpy(new_cont->pools, cont->pools, cont->n *
                       sizeof(struct mlx5_flow_counter_pool *));
-       mem_mng = flow_dv_create_counter_stat_mem_mng(dev,
-               MLX5_CNT_CONTAINER_RESIZE + MLX5_MAX_PENDING_QUERIES);
-       if (!mem_mng) {
-               rte_free(new_cont->pools);
-               return NULL;
+       /*
+        * Fallback mode query the counter directly, no background query
+        * resources are needed.
+        */
+       if (!priv->counter_fallback) {
+               mem_mng = flow_dv_create_counter_stat_mem_mng(dev,
+                       MLX5_CNT_CONTAINER_RESIZE + MLX5_MAX_PENDING_QUERIES);
+               if (!mem_mng) {
+                       rte_free(new_cont->pools);
+                       return NULL;
+               }
+               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);
        }
-       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);
        new_cont->n = resize;
        rte_atomic16_set(&new_cont->n_valid, rte_atomic16_read(&cont->n_valid));
        TAILQ_INIT(&new_cont->pool_list);
@@ -4105,7 +4072,7 @@ flow_dv_container_resize(struct rte_eth_dev *dev, uint32_t batch)
  * @param[in] dev
  *   Pointer to the Ethernet device structure.
  * @param[in] cnt
- *   Pointer to the flow counter.
+ *   Index to the flow counter.
  * @param[out] pkts
  *   The statistics value of packets.
  * @param[out] bytes
@@ -4115,17 +4082,23 @@ flow_dv_container_resize(struct rte_eth_dev *dev, uint32_t batch)
  *   0 on success, otherwise a negative errno value and rte_errno is set.
  */
 static inline int
-_flow_dv_query_count(struct rte_eth_dev *dev,
-                    struct mlx5_flow_counter *cnt, uint64_t *pkts,
+_flow_dv_query_count(struct rte_eth_dev *dev, uint32_t counter, uint64_t *pkts,
                     uint64_t *bytes)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
-       struct mlx5_flow_counter_pool *pool =
-                       flow_dv_counter_pool_get(cnt);
-       int offset = cnt - &pool->counters_raw[0];
+       struct mlx5_flow_counter_pool *pool = NULL;
+       struct mlx5_flow_counter *cnt;
+       struct mlx5_flow_counter_ext *cnt_ext = NULL;
+       int offset;
 
-       if (priv->counter_fallback)
-               return _flow_dv_query_count_fallback(dev, cnt, pkts, bytes);
+       cnt = flow_dv_counter_get_by_idx(dev, counter, &pool);
+       MLX5_ASSERT(pool);
+       if (counter < MLX5_CNT_BATCH_OFFSET) {
+               cnt_ext = MLX5_CNT_TO_CNT_EXT(pool, cnt);
+               if (priv->counter_fallback)
+                       return mlx5_devx_cmd_flow_counter_query(cnt_ext->dcs, 0,
+                                       0, pkts, bytes, 0, NULL, NULL, 0);
+       }
 
        rte_spinlock_lock(&pool->sl);
        /*
@@ -4133,10 +4106,11 @@ _flow_dv_query_count(struct rte_eth_dev *dev,
         * current allocated in parallel to the host reading.
         * In this case the new counter values must be reported as 0.
         */
-       if (unlikely(!cnt->batch && cnt->dcs->id < pool->raw->min_dcs_id)) {
+       if (unlikely(cnt_ext && cnt_ext->dcs->id < pool->raw->min_dcs_id)) {
                *pkts = 0;
                *bytes = 0;
        } else {
+               offset = cnt - &pool->counters_raw[0];
                *pkts = rte_be_to_cpu_64(pool->raw->data[offset].hits);
                *bytes = rte_be_to_cpu_64(pool->raw->data[offset].bytes);
        }
@@ -4175,15 +4149,18 @@ flow_dv_pool_create(struct rte_eth_dev *dev, struct mlx5_devx_obj *dcs,
                if (!cont)
                        return NULL;
        }
-       size = sizeof(*pool) + MLX5_COUNTERS_PER_POOL *
-                       sizeof(struct mlx5_flow_counter);
+       size = sizeof(*pool);
+       if (!batch)
+               size += MLX5_COUNTERS_PER_POOL *
+                       sizeof(struct mlx5_flow_counter_ext);
        pool = rte_calloc(__func__, 1, size, 0);
        if (!pool) {
                rte_errno = ENOMEM;
                return NULL;
        }
        pool->min_dcs = dcs;
-       pool->raw = cont->init_mem_mng->raws + n_valid %
+       if (!priv->counter_fallback)
+               pool->raw = cont->init_mem_mng->raws + n_valid %
                                                     MLX5_CNT_CONTAINER_RESIZE;
        pool->raw_hw = NULL;
        rte_spinlock_init(&pool->sl);
@@ -4195,9 +4172,16 @@ flow_dv_pool_create(struct rte_eth_dev *dev, struct mlx5_devx_obj *dcs,
         * without the last query finished and stats updated to the memory.
         */
        rte_atomic64_set(&pool->start_query_gen, 0x2);
-       rte_atomic64_set(&pool->end_query_gen, 0x2);
+       /*
+        * There's no background query thread for fallback mode, set the
+        * end_query_gen to the maximum value since no need to wait for
+        * statistics update.
+        */
+       rte_atomic64_set(&pool->end_query_gen, priv->counter_fallback ?
+                        INT64_MAX : 0x2);
        TAILQ_INIT(&pool->counters);
        TAILQ_INSERT_HEAD(&cont->pool_list, pool, next);
+       pool->index = n_valid;
        cont->pools[n_valid] = pool;
        /* Pool initialization must be updated before host thread access. */
        rte_cio_wmb();
@@ -4249,9 +4233,10 @@ flow_dv_counter_pool_prepare(struct rte_eth_dev *dev,
                        rte_atomic64_set(&pool->a64_dcs,
                                         (int64_t)(uintptr_t)dcs);
                }
-               cnt = &pool->counters_raw[dcs->id % MLX5_COUNTERS_PER_POOL];
+               i = dcs->id % MLX5_COUNTERS_PER_POOL;
+               cnt = &pool->counters_raw[i];
                TAILQ_INSERT_HEAD(&pool->counters, cnt, next);
-               cnt->dcs = dcs;
+               MLX5_GET_POOL_CNT_EXT(pool, i)->dcs = dcs;
                *cnt_free = cnt;
                return cont;
        }
@@ -4270,7 +4255,6 @@ flow_dv_counter_pool_prepare(struct rte_eth_dev *dev,
        pool = TAILQ_FIRST(&cont->pool_list);
        for (i = 0; i < MLX5_COUNTERS_PER_POOL; ++i) {
                cnt = &pool->counters_raw[i];
-               cnt->pool = pool;
                TAILQ_INSERT_HEAD(&pool->counters, cnt, next);
        }
        *cnt_free = &pool->counters_raw[0];
@@ -4284,23 +4268,30 @@ flow_dv_counter_pool_prepare(struct rte_eth_dev *dev,
  *   Pointer to the relevant counter pool container.
  * @param[in] id
  *   The shared counter ID to search.
+ * @param[out] ppool
+ *   mlx5 flow counter pool in the container,
  *
  * @return
- *   NULL if not existed, otherwise pointer to the shared counter.
+ *   NULL if not existed, otherwise pointer to the shared extend counter.
  */
-static struct mlx5_flow_counter *
-flow_dv_counter_shared_search(struct mlx5_pools_container *cont,
-                             uint32_t id)
+static struct mlx5_flow_counter_ext *
+flow_dv_counter_shared_search(struct mlx5_pools_container *cont, uint32_t id,
+                             struct mlx5_flow_counter_pool **ppool)
 {
-       static struct mlx5_flow_counter *cnt;
+       static struct mlx5_flow_counter_ext *cnt;
        struct mlx5_flow_counter_pool *pool;
-       int i;
+       uint32_t i;
+       uint32_t n_valid = rte_atomic16_read(&cont->n_valid);
 
-       TAILQ_FOREACH(pool, &cont->pool_list, next) {
+       for (i = 0; i < n_valid; i++) {
+               pool = cont->pools[i];
                for (i = 0; i < MLX5_COUNTERS_PER_POOL; ++i) {
-                       cnt = &pool->counters_raw[i];
-                       if (cnt->ref_cnt && cnt->shared && cnt->id == id)
+                       cnt = MLX5_GET_POOL_CNT_EXT(pool, i);
+                       if (cnt->ref_cnt && cnt->shared && cnt->id == id) {
+                               if (ppool)
+                                       *ppool = cont->pools[i];
                                return cnt;
+                       }
                }
        }
        return NULL;
@@ -4319,15 +4310,16 @@ flow_dv_counter_shared_search(struct mlx5_pools_container *cont,
  *   Counter flow group.
  *
  * @return
- *   pointer to flow counter on success, NULL otherwise and rte_errno is set.
+ *   Index to flow counter on success, 0 otherwise and rte_errno is set.
  */
-static struct mlx5_flow_counter *
+static uint32_t
 flow_dv_counter_alloc(struct rte_eth_dev *dev, uint32_t shared, uint32_t id,
                      uint16_t group)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
        struct mlx5_flow_counter_pool *pool = NULL;
        struct mlx5_flow_counter *cnt_free = NULL;
+       struct mlx5_flow_counter_ext *cnt_ext = NULL;
        /*
         * Currently group 0 flow counter cannot be assigned to a flow if it is
         * not the first one in the batch counter allocation, so it is better
@@ -4336,25 +4328,27 @@ flow_dv_counter_alloc(struct rte_eth_dev *dev, uint32_t shared, uint32_t id,
         * A counter can be shared between different groups so need to take
         * shared counters from the single container.
         */
-       uint32_t batch = (group && !shared) ? 1 : 0;
+       uint32_t batch = (group && !shared && !priv->counter_fallback) ? 1 : 0;
        struct mlx5_pools_container *cont = MLX5_CNT_CONTAINER(priv->sh, batch,
                                                               0);
+       uint32_t cnt_idx;
 
-       if (priv->counter_fallback)
-               return flow_dv_counter_alloc_fallback(dev, shared, id);
        if (!priv->config.devx) {
                rte_errno = ENOTSUP;
-               return NULL;
+               return 0;
        }
        if (shared) {
-               cnt_free = flow_dv_counter_shared_search(cont, id);
-               if (cnt_free) {
-                       if (cnt_free->ref_cnt + 1 == 0) {
+               cnt_ext = flow_dv_counter_shared_search(cont, id, &pool);
+               if (cnt_ext) {
+                       if (cnt_ext->ref_cnt + 1 == 0) {
                                rte_errno = E2BIG;
-                               return NULL;
+                               return 0;
                        }
-                       cnt_free->ref_cnt++;
-                       return cnt_free;
+                       cnt_ext->ref_cnt++;
+                       cnt_idx = pool->index * MLX5_COUNTERS_PER_POOL +
+                                 (cnt_ext->dcs->id % MLX5_COUNTERS_PER_POOL)
+                                 + 1;
+                       return cnt_idx;
                }
        }
        /* Pools which has a free counters are in the start. */
@@ -4377,10 +4371,11 @@ flow_dv_counter_alloc(struct rte_eth_dev *dev, uint32_t shared, uint32_t id,
        if (!cnt_free) {
                cont = flow_dv_counter_pool_prepare(dev, &cnt_free, batch);
                if (!cont)
-                       return NULL;
+                       return 0;
                pool = TAILQ_FIRST(&cont->pool_list);
        }
-       cnt_free->batch = batch;
+       if (!batch)
+               cnt_ext = MLX5_CNT_TO_CNT_EXT(pool, cnt_free);
        /* Create a DV counter action only in the first time usage. */
        if (!cnt_free->action) {
                uint16_t offset;
@@ -4391,23 +4386,28 @@ flow_dv_counter_alloc(struct rte_eth_dev *dev, uint32_t shared, uint32_t id,
                        dcs = pool->min_dcs;
                } else {
                        offset = 0;
-                       dcs = cnt_free->dcs;
+                       dcs = cnt_ext->dcs;
                }
                cnt_free->action = mlx5_glue->dv_create_flow_action_counter
                                        (dcs->obj, offset);
                if (!cnt_free->action) {
                        rte_errno = errno;
-                       return NULL;
+                       return 0;
                }
        }
+       cnt_idx = MLX5_MAKE_CNT_IDX(pool->index,
+                                   (cnt_free - pool->counters_raw));
+       cnt_idx += batch * MLX5_CNT_BATCH_OFFSET;
        /* Update the counter reset values. */
-       if (_flow_dv_query_count(dev, cnt_free, &cnt_free->hits,
+       if (_flow_dv_query_count(dev, cnt_idx, &cnt_free->hits,
                                 &cnt_free->bytes))
-               return NULL;
-       cnt_free->shared = shared;
-       cnt_free->ref_cnt = 1;
-       cnt_free->id = id;
-       if (!priv->sh->cmng.query_thread_on)
+               return 0;
+       if (cnt_ext) {
+               cnt_ext->shared = shared;
+               cnt_ext->ref_cnt = 1;
+               cnt_ext->id = id;
+       }
+       if (!priv->counter_fallback && !priv->sh->cmng.query_thread_on)
                /* Start the asynchronous batch query by the host thread. */
                mlx5_set_query_alarm(priv->sh);
        TAILQ_REMOVE(&pool->counters, cnt_free, next);
@@ -4416,7 +4416,7 @@ flow_dv_counter_alloc(struct rte_eth_dev *dev, uint32_t shared, uint32_t id,
                TAILQ_REMOVE(&cont->pool_list, pool, next);
                TAILQ_INSERT_TAIL(&cont->pool_list, pool, next);
        }
-       return cnt_free;
+       return cnt_idx;
 }
 
 /**
@@ -4425,34 +4425,33 @@ flow_dv_counter_alloc(struct rte_eth_dev *dev, uint32_t shared, uint32_t id,
  * @param[in] dev
  *   Pointer to the Ethernet device structure.
  * @param[in] counter
- *   Pointer to the counter handler.
+ *   Index to the counter handler.
  */
 static void
-flow_dv_counter_release(struct rte_eth_dev *dev,
-                       struct mlx5_flow_counter *counter)
+flow_dv_counter_release(struct rte_eth_dev *dev, uint32_t counter)
 {
-       struct mlx5_priv *priv = dev->data->dev_private;
+       struct mlx5_flow_counter_pool *pool = NULL;
+       struct mlx5_flow_counter *cnt;
+       struct mlx5_flow_counter_ext *cnt_ext = NULL;
 
        if (!counter)
                return;
-       if (priv->counter_fallback) {
-               flow_dv_counter_release_fallback(dev, counter);
-               return;
-       }
-       if (--counter->ref_cnt == 0) {
-               struct mlx5_flow_counter_pool *pool =
-                               flow_dv_counter_pool_get(counter);
-
-               /* Put the counter in the end - the last updated one. */
-               TAILQ_INSERT_TAIL(&pool->counters, counter, next);
-               /*
-                * Counters released between query trigger and handler need
-                * to wait the next round of query. Since the packets arrive
-                * in the gap period will not be taken into account to the
-                * old counter.
-                */
-               counter->query_gen = rte_atomic64_read(&pool->start_query_gen);
+       cnt = flow_dv_counter_get_by_idx(dev, counter, &pool);
+       MLX5_ASSERT(pool);
+       if (counter < MLX5_CNT_BATCH_OFFSET) {
+               cnt_ext = MLX5_CNT_TO_CNT_EXT(pool, cnt);
+               if (cnt_ext && --cnt_ext->ref_cnt)
+                       return;
        }
+       /* Put the counter in the end - the last updated one. */
+       TAILQ_INSERT_TAIL(&pool->counters, cnt, next);
+       /*
+        * Counters released between query trigger and handler need
+        * to wait the next round of query. Since the packets arrive
+        * in the gap period will not be taken into account to the
+        * old counter.
+        */
+       cnt->query_gen = rte_atomic64_read(&pool->start_query_gen);
 }
 
 /**
@@ -4596,6 +4595,7 @@ flow_dv_validate(struct rte_eth_dev *dev, const struct rte_flow_attr *attr,
        struct mlx5_priv *priv = dev->data->dev_private;
        struct mlx5_dev_config *dev_conf = &priv->config;
        uint16_t queue_index = 0xFFFF;
+       const struct rte_flow_item_vlan *vlan_m = NULL;
 
        if (items == NULL)
                return -1;
@@ -4653,6 +4653,9 @@ flow_dv_validate(struct rte_eth_dev *dev, const struct rte_flow_attr *attr,
                        } else {
                                ether_type = 0;
                        }
+                       /* Store outer VLAN mask for of_push_vlan action. */
+                       if (!tunnel)
+                               vlan_m = items->mask;
                        break;
                case RTE_FLOW_ITEM_TYPE_IPV4:
                        mlx5_flow_tunnel_ip_check(items, next_protocol,
@@ -4968,7 +4971,7 @@ flow_dv_validate(struct rte_eth_dev *dev, const struct rte_flow_attr *attr,
                case RTE_FLOW_ACTION_TYPE_OF_PUSH_VLAN:
                        ret = flow_dv_validate_action_push_vlan(dev,
                                                                action_flags,
-                                                               item_flags,
+                                                               vlan_m,
                                                                actions, attr,
                                                                error);
                        if (ret < 0)
@@ -7516,10 +7519,11 @@ __flow_dv_translate(struct rte_eth_dev *dev,
                                                        count->shared,
                                                        count->id,
                                                        dev_flow->dv.group);
-                       if (flow->counter == NULL)
+                       if (!flow->counter)
                                goto cnt_err;
                        dev_flow->dv.actions[actions_n++] =
-                               flow->counter->action;
+                                 (flow_dv_counter_get_by_idx(dev,
+                                 flow->counter, NULL))->action;
                        action_flags |= MLX5_FLOW_ACTION_COUNT;
                        break;
 cnt_err:
@@ -7543,7 +7547,9 @@ cnt_err:
                        action_flags |= MLX5_FLOW_ACTION_OF_POP_VLAN;
                        break;
                case RTE_FLOW_ACTION_TYPE_OF_PUSH_VLAN:
-                       flow_dev_get_vlan_info_from_items(items, &vlan);
+                       if (!(action_flags &
+                             MLX5_FLOW_ACTION_OF_SET_VLAN_VID))
+                               flow_dev_get_vlan_info_from_items(items, &vlan);
                        vlan.eth_proto = rte_be_to_cpu_16
                             ((((const struct rte_flow_action_of_push_vlan *)
                                                   actions->conf)->ethertype));
@@ -8053,7 +8059,7 @@ __flow_dv_apply(struct rte_eth_dev *dev, struct rte_flow *flow,
        int err;
        int idx;
 
-       for (idx = priv->flow_idx - 1; idx >= 0; idx--) {
+       for (idx = priv->flow_idx - 1; idx >= priv->flow_nested_idx; idx--) {
                dev_flow = &((struct mlx5_flow *)priv->inter_flows)[idx];
                dv = &dev_flow->dv;
                dh = dev_flow->handle;
@@ -8392,7 +8398,7 @@ __flow_dv_destroy(struct rte_eth_dev *dev, struct rte_flow *flow)
        __flow_dv_remove(dev, flow);
        if (flow->counter) {
                flow_dv_counter_release(dev, flow->counter);
-               flow->counter = NULL;
+               flow->counter = 0;
        }
        if (flow->meter) {
                mlx5_flow_meter_detach(flow->meter);
@@ -8449,6 +8455,10 @@ flow_dv_query_count(struct rte_eth_dev *dev, struct rte_flow *flow,
                                          "counters are not supported");
        if (flow->counter) {
                uint64_t pkts, bytes;
+               struct mlx5_flow_counter *cnt;
+
+               cnt = flow_dv_counter_get_by_idx(dev, flow->counter,
+                                                NULL);
                int err = _flow_dv_query_count(dev, flow->counter, &pkts,
                                               &bytes);
 
@@ -8458,11 +8468,11 @@ flow_dv_query_count(struct rte_eth_dev *dev, struct rte_flow *flow,
                                        NULL, "cannot read counters");
                qc->hits_set = 1;
                qc->bytes_set = 1;
-               qc->hits = pkts - flow->counter->hits;
-               qc->bytes = bytes - flow->counter->bytes;
+               qc->hits = pkts - cnt->hits;
+               qc->bytes = bytes - cnt->bytes;
                if (qc->reset) {
-                       flow->counter->hits = pkts;
-                       flow->counter->bytes = bytes;
+                       cnt->hits = pkts;
+                       cnt->bytes = bytes;
                }
                return 0;
        }
@@ -8544,6 +8554,9 @@ flow_dv_destroy_mtr_tbl(struct rte_eth_dev *dev,
        if (mtd->egress.tbl)
                claim_zero(flow_dv_tbl_resource_release(dev,
                                                        mtd->egress.tbl));
+       if (mtd->egress.sfx_tbl)
+               claim_zero(flow_dv_tbl_resource_release(dev,
+                                                       mtd->egress.sfx_tbl));
        if (mtd->ingress.color_matcher)
                claim_zero(mlx5_glue->dv_destroy_flow_matcher
                          (mtd->ingress.color_matcher));
@@ -8553,6 +8566,9 @@ flow_dv_destroy_mtr_tbl(struct rte_eth_dev *dev,
        if (mtd->ingress.tbl)
                claim_zero(flow_dv_tbl_resource_release(dev,
                                                        mtd->ingress.tbl));
+       if (mtd->ingress.sfx_tbl)
+               claim_zero(flow_dv_tbl_resource_release(dev,
+                                                       mtd->ingress.sfx_tbl));
        if (mtd->transfer.color_matcher)
                claim_zero(mlx5_glue->dv_destroy_flow_matcher
                          (mtd->transfer.color_matcher));
@@ -8562,6 +8578,9 @@ flow_dv_destroy_mtr_tbl(struct rte_eth_dev *dev,
        if (mtd->transfer.tbl)
                claim_zero(flow_dv_tbl_resource_release(dev,
                                                        mtd->transfer.tbl));
+       if (mtd->transfer.sfx_tbl)
+               claim_zero(flow_dv_tbl_resource_release(dev,
+                                                       mtd->transfer.sfx_tbl));
        if (mtd->drop_actn)
                claim_zero(mlx5_glue->destroy_flow_action(mtd->drop_actn));
        rte_free(mtd);
@@ -8609,31 +8628,16 @@ flow_dv_prepare_mtr_tables(struct rte_eth_dev *dev,
                .match_mask = (void *)&mask,
        };
        void *actions[METER_ACTIONS];
-       struct mlx5_flow_tbl_resource **sfx_tbl;
        struct mlx5_meter_domain_info *dtb;
        struct rte_flow_error error;
        int i = 0;
 
-       if (transfer) {
-               sfx_tbl = &sh->fdb_mtr_sfx_tbl;
+       if (transfer)
                dtb = &mtb->transfer;
-       } else if (egress) {
-               sfx_tbl = &sh->tx_mtr_sfx_tbl;
+       else if (egress)
                dtb = &mtb->egress;
-       } else {
-               sfx_tbl = &sh->rx_mtr_sfx_tbl;
+       else
                dtb = &mtb->ingress;
-       }
-       /* If the suffix table in missing, create it. */
-       if (!(*sfx_tbl)) {
-               *sfx_tbl = flow_dv_tbl_resource_get(dev,
-                                               MLX5_FLOW_TABLE_LEVEL_SUFFIX,
-                                               egress, transfer, &error);
-               if (!(*sfx_tbl)) {
-                       DRV_LOG(ERR, "Failed to create meter suffix table.");
-                       return -1;
-               }
-       }
        /* Create the meter table with METER level. */
        dtb->tbl = flow_dv_tbl_resource_get(dev, MLX5_FLOW_TABLE_LEVEL_METER,
                                            egress, transfer, &error);
@@ -8641,6 +8645,14 @@ flow_dv_prepare_mtr_tables(struct rte_eth_dev *dev,
                DRV_LOG(ERR, "Failed to create meter policer table.");
                return -1;
        }
+       /* Create the meter suffix table with SUFFIX level. */
+       dtb->sfx_tbl = flow_dv_tbl_resource_get(dev,
+                                           MLX5_FLOW_TABLE_LEVEL_SUFFIX,
+                                           egress, transfer, &error);
+       if (!dtb->sfx_tbl) {
+               DRV_LOG(ERR, "Failed to create meter suffix table.");
+               return -1;
+       }
        /* Create matchers, Any and Color. */
        dv_attr.priority = 3;
        dv_attr.match_criteria_enable = 0;
@@ -8712,9 +8724,12 @@ flow_dv_create_mtr_tbl(struct rte_eth_dev *dev,
        }
        /* Create meter count actions */
        for (i = 0; i <= RTE_MTR_DROPPED; i++) {
+               struct mlx5_flow_counter *cnt;
                if (!fm->policer_stats.cnt[i])
                        continue;
-               mtb->count_actns[i] = fm->policer_stats.cnt[i]->action;
+               cnt = flow_dv_counter_get_by_idx(dev,
+                     fm->policer_stats.cnt[i], NULL);
+               mtb->count_actns[i] = cnt->action;
        }
        /* Create drop action. */
        mtb->drop_actn = mlx5_glue->dr_create_flow_action_drop();
@@ -8811,8 +8826,6 @@ flow_dv_destroy_policer_rules(struct rte_eth_dev *dev __rte_unused,
  *   Pointer to flow meter structure.
  * @param[in] mtb
  *   Pointer to DV meter table set.
- * @param[in] sfx_tb
- *   Pointer to suffix table.
  * @param[in] mtr_reg_c
  *   Color match REG_C.
  *
@@ -8822,7 +8835,6 @@ flow_dv_destroy_policer_rules(struct rte_eth_dev *dev __rte_unused,
 static int
 flow_dv_create_policer_forward_rule(struct mlx5_flow_meter *fm,
                                    struct mlx5_meter_domain_info *dtb,
-                                   struct mlx5_flow_tbl_resource *sfx_tb,
                                    uint8_t mtr_reg_c)
 {
        struct mlx5_flow_dv_match_params matcher = {
@@ -8836,12 +8848,10 @@ flow_dv_create_policer_forward_rule(struct mlx5_flow_meter *fm,
        int i;
 
        /* Create jump action. */
-       if (!sfx_tb)
-               return -1;
        if (!dtb->jump_actn)
                dtb->jump_actn =
                        mlx5_glue->dr_create_flow_action_dest_flow_tbl
-                                                       (sfx_tb->obj);
+                                                       (dtb->sfx_tbl->obj);
        if (!dtb->jump_actn) {
                DRV_LOG(ERR, "Failed to create policer jump action.");
                goto error;
@@ -8896,7 +8906,6 @@ flow_dv_create_policer_rules(struct rte_eth_dev *dev,
 
        if (attr->egress) {
                ret = flow_dv_create_policer_forward_rule(fm, &mtb->egress,
-                                               priv->sh->tx_mtr_sfx_tbl,
                                                priv->mtr_color_reg);
                if (ret) {
                        DRV_LOG(ERR, "Failed to create egress policer.");
@@ -8905,7 +8914,6 @@ flow_dv_create_policer_rules(struct rte_eth_dev *dev,
        }
        if (attr->ingress) {
                ret = flow_dv_create_policer_forward_rule(fm, &mtb->ingress,
-                                               priv->sh->rx_mtr_sfx_tbl,
                                                priv->mtr_color_reg);
                if (ret) {
                        DRV_LOG(ERR, "Failed to create ingress policer.");
@@ -8914,7 +8922,6 @@ flow_dv_create_policer_rules(struct rte_eth_dev *dev,
        }
        if (attr->transfer) {
                ret = flow_dv_create_policer_forward_rule(fm, &mtb->transfer,
-                                               priv->sh->fdb_mtr_sfx_tbl,
                                                priv->mtr_color_reg);
                if (ret) {
                        DRV_LOG(ERR, "Failed to create transfer policer.");
@@ -8933,7 +8940,7 @@ error:
  * @param[in] dev
  *   Pointer to the Ethernet device structure.
  * @param[in] cnt
- *   Pointer to the flow counter.
+ *   Index to the flow counter.
  * @param[in] clear
  *   Set to clear the counter statistics.
  * @param[out] pkts
@@ -8945,19 +8952,21 @@ error:
  *   0 on success, otherwise return -1.
  */
 static int
-flow_dv_counter_query(struct rte_eth_dev *dev,
-                     struct mlx5_flow_counter *cnt, bool clear,
+flow_dv_counter_query(struct rte_eth_dev *dev, uint32_t counter, bool clear,
                      uint64_t *pkts, uint64_t *bytes)
 {
        struct mlx5_priv *priv = dev->data->dev_private;
+       struct mlx5_flow_counter *cnt;
        uint64_t inn_pkts, inn_bytes;
        int ret;
 
        if (!priv->config.devx)
                return -1;
-       ret = _flow_dv_query_count(dev, cnt, &inn_pkts, &inn_bytes);
+
+       ret = _flow_dv_query_count(dev, counter, &inn_pkts, &inn_bytes);
        if (ret)
                return -1;
+       cnt = flow_dv_counter_get_by_idx(dev, counter, NULL);
        *pkts = inn_pkts - cnt->hits;
        *bytes = inn_bytes - cnt->bytes;
        if (clear) {
@@ -9027,10 +9036,10 @@ flow_dv_destroy(struct rte_eth_dev *dev, struct rte_flow *flow)
 /*
  * Mutex-protected thunk to lock-free flow_dv_counter_alloc().
  */
-static struct mlx5_flow_counter *
+static uint32_t
 flow_dv_counter_allocate(struct rte_eth_dev *dev)
 {
-       struct mlx5_flow_counter *cnt;
+       uint32_t cnt;
 
        flow_dv_shared_lock(dev);
        cnt = flow_dv_counter_alloc(dev, 0, 0, 1);
@@ -9042,7 +9051,7 @@ flow_dv_counter_allocate(struct rte_eth_dev *dev)
  * Mutex-protected thunk to lock-free flow_dv_counter_release().
  */
 static void
-flow_dv_counter_free(struct rte_eth_dev *dev, struct mlx5_flow_counter *cnt)
+flow_dv_counter_free(struct rte_eth_dev *dev, uint32_t cnt)
 {
        flow_dv_shared_lock(dev);
        flow_dv_counter_release(dev, cnt);