[1/2] net/mlx5: fix queue used to deallocate counter
Checks
Commit Message
When ports are configured to share flow engine resources
(configured through rte_flow_configure()),
counter objects used during flow rule creation are pulled
directly from the shared counter pool.
This is done by calling mlx5_hws_cnt_pool_get() with NULL pointer
passed as queue pointer.
When a flow rule was destroyed on the host port,
counter object was not returned to the pool, because
mlx5_hws_cnt_pool_put() was always called with a pointer
to the currently used queue.
This forced placing the counter in the local queue cache.
As a result, during the application runtime, some counter objects
became inaccessible, since this queue is never fully flushed.
This patch fixes that behavior, by changing the flow rule destroying
logic. If port is configured to use shared resources,
mlx5_hws_cnt_pool_put() is called with NULL pointer
passed as queue pointer.
Fixes: 13ea6bdcc7ee ("net/mlx5: support counters in cross port shared mode")
Cc: viacheslavo@nvidia.com
Signed-off-by: Dariusz Sosnowski <dsosnowski@nvidia.com>
Acked-by: Ori Kam <orika@nvidia.com>
---
drivers/net/mlx5/mlx5_flow_hw.c | 14 +++++++++-----
drivers/net/mlx5/mlx5_hws_cnt.h | 16 ++++++++++++++++
2 files changed, 25 insertions(+), 5 deletions(-)
@@ -2273,6 +2273,7 @@ flow_hw_actions_construct(struct rte_eth_dev *dev,
struct mlx5_hrxq *hrxq;
uint32_t ct_idx;
cnt_id_t cnt_id;
+ uint32_t *cnt_queue;
uint32_t mtr_id;
action = &actions[act_data->action_src];
@@ -2429,10 +2430,9 @@ flow_hw_actions_construct(struct rte_eth_dev *dev,
break;
/* Fall-through. */
case RTE_FLOW_ACTION_TYPE_COUNT:
- ret = mlx5_hws_cnt_pool_get(priv->hws_cpool,
- (priv->shared_refcnt ||
- priv->hws_cpool->cfg.host_cpool) ?
- NULL : &queue, &cnt_id, age_idx);
+ /* If the port is engaged in resource sharing, do not use queue cache. */
+ cnt_queue = mlx5_hws_cnt_is_pool_shared(priv) ? NULL : &queue;
+ ret = mlx5_hws_cnt_pool_get(priv->hws_cpool, cnt_queue, &cnt_id, age_idx);
if (ret != 0)
return ret;
ret = mlx5_hws_cnt_pool_get_action_offset
@@ -3014,6 +3014,8 @@ flow_hw_age_count_release(struct mlx5_priv *priv, uint32_t queue,
struct rte_flow_hw *flow,
struct rte_flow_error *error)
{
+ uint32_t *cnt_queue;
+
if (mlx5_hws_cnt_is_shared(priv->hws_cpool, flow->cnt_id)) {
if (flow->age_idx && !mlx5_hws_age_is_indirect(flow->age_idx)) {
/* Remove this AGE parameter from indirect counter. */
@@ -3024,8 +3026,10 @@ flow_hw_age_count_release(struct mlx5_priv *priv, uint32_t queue,
}
return;
}
+ /* If the port is engaged in resource sharing, do not use queue cache. */
+ cnt_queue = mlx5_hws_cnt_is_pool_shared(priv) ? NULL : &queue;
/* Put the counter first to reduce the race risk in BG thread. */
- mlx5_hws_cnt_pool_put(priv->hws_cpool, &queue, &flow->cnt_id);
+ mlx5_hws_cnt_pool_put(priv->hws_cpool, cnt_queue, &flow->cnt_id);
flow->cnt_id = 0;
if (flow->age_idx) {
if (mlx5_hws_age_is_indirect(flow->age_idx)) {
@@ -553,6 +553,22 @@ mlx5_hws_cnt_pool_get(struct mlx5_hws_cnt_pool *cpool, uint32_t *queue,
return 0;
}
+/**
+ * Check if counter pool allocated for HWS is shared between ports.
+ *
+ * @param[in] priv
+ * Pointer to the port private data structure.
+ *
+ * @return
+ * True if counter pools is shared between ports. False otherwise.
+ */
+static __rte_always_inline bool
+mlx5_hws_cnt_is_pool_shared(struct mlx5_priv *priv)
+{
+ return priv && priv->hws_cpool &&
+ (priv->shared_refcnt || priv->hws_cpool->cfg.host_cpool != NULL);
+}
+
static __rte_always_inline unsigned int
mlx5_hws_cnt_pool_get_size(struct mlx5_hws_cnt_pool *cpool)
{