[PATCH 20.11] net/mlx5: fix flow shared age action reference counting

[PATCH 20.11] net/mlx5: fix flow shared age action reference counting

[Dmitry Kozlyuk]

[ upstream commit b09c65fa4f8bb55880b6b36c849e4ed1bb815227 ]

When an shared AGE action is used in a flow rule with a pattern that
causes RSS expansion, each device flow generated by the expansion
incremented the reference counter of the action. When such a flow was
destroyed, its action reference counter had been decremented only once.
The action remained marked as being used and could not be destroyed.
The error was only visible with --log-level=pmd.net.mlx5:debug
("...references 1" is not an error):

    mlx5_pci: Shared age action 65536 was released with references 4.

Increment action counter only once for the original flow rule.

Fixes: 81073e1f8ce1 ("net/mlx5: support shared age action")

Signed-off-by: Dmitry Kozlyuk <dkozlyuk@nvidia.com>
---
 drivers/net/mlx5/mlx5_flow_dv.c | 12 ++++++++----
 1 file changed, 8 insertions(+), 4 deletions(-)

diff --git a/drivers/net/mlx5/mlx5_flow_dv.c b/drivers/net/mlx5/mlx5_flow_dv.c
index 1793683421..66732cbdd4 100644
--- a/drivers/net/mlx5/mlx5_flow_dv.c
+++ b/drivers/net/mlx5/mlx5_flow_dv.c
@@ -9812,6 +9812,7 @@ flow_dv_translate(struct rte_eth_dev *dev,
 		const struct rte_flow_action_meter *mtr;
 		struct mlx5_flow_tbl_resource *tbl;
 		struct mlx5_aso_age_action *age_act;
+		uint32_t owner_idx;
 		uint32_t port_id = 0;
 		struct mlx5_flow_dv_port_id_action_resource port_id_resource;
 		int action_type = actions->type;
@@ -9951,10 +9952,13 @@ flow_dv_translate(struct rte_eth_dev *dev,
 				MLX5_FLOW_FATE_QUEUE;
 			break;
 		case MLX5_RTE_FLOW_ACTION_TYPE_AGE:
-			flow->age = (uint32_t)(uintptr_t)(action->conf);
-			age_act = flow_aso_age_get_by_idx(dev, flow->age);
-			__atomic_fetch_add(&age_act->refcnt, 1,
-					   __ATOMIC_RELAXED);
+			owner_idx = (uint32_t)(uintptr_t)action->conf;
+			age_act = flow_aso_age_get_by_idx(dev, owner_idx);
+			if (flow->age == 0) {
+				flow->age = owner_idx;
+				__atomic_fetch_add(&age_act->refcnt, 1,
+						   __ATOMIC_RELAXED);
+			}
 			dev_flow->dv.actions[actions_n++] = age_act->dr_action;
 			action_flags |= MLX5_FLOW_ACTION_AGE;
 			break;
-- 
2.25.1

Re: [PATCH 20.11] net/mlx5: fix flow shared age action reference counting

[Xueming(Steven) Li]

Thanks, applied.

On Fri, 2021-12-03 at 22:11 +0200, Dmitry Kozlyuk wrote:
> [ upstream commit b09c65fa4f8bb55880b6b36c849e4ed1bb815227 ]
> 
> When an shared AGE action is used in a flow rule with a pattern that
> causes RSS expansion, each device flow generated by the expansion
> incremented the reference counter of the action. When such a flow was
> destroyed, its action reference counter had been decremented only once.
> The action remained marked as being used and could not be destroyed.
> The error was only visible with --log-level=pmd.net.mlx5:debug
> ("...references 1" is not an error):
> 
>     mlx5_pci: Shared age action 65536 was released with references 4.
> 
> Increment action counter only once for the original flow rule.
> 
> Fixes: 81073e1f8ce1 ("net/mlx5: support shared age action")
> 
> Signed-off-by: Dmitry Kozlyuk <dkozlyuk@nvidia.com>
> ---
>  drivers/net/mlx5/mlx5_flow_dv.c | 12 ++++++++----
>  1 file changed, 8 insertions(+), 4 deletions(-)
> 
> diff --git a/drivers/net/mlx5/mlx5_flow_dv.c b/drivers/net/mlx5/mlx5_flow_dv.c
> index 1793683421..66732cbdd4 100644
> --- a/drivers/net/mlx5/mlx5_flow_dv.c
> +++ b/drivers/net/mlx5/mlx5_flow_dv.c
> @@ -9812,6 +9812,7 @@ flow_dv_translate(struct rte_eth_dev *dev,
>  		const struct rte_flow_action_meter *mtr;
>  		struct mlx5_flow_tbl_resource *tbl;
>  		struct mlx5_aso_age_action *age_act;
> +		uint32_t owner_idx;
>  		uint32_t port_id = 0;
>  		struct mlx5_flow_dv_port_id_action_resource port_id_resource;
>  		int action_type = actions->type;
> @@ -9951,10 +9952,13 @@ flow_dv_translate(struct rte_eth_dev *dev,
>  				MLX5_FLOW_FATE_QUEUE;
>  			break;
>  		case MLX5_RTE_FLOW_ACTION_TYPE_AGE:
> -			flow->age = (uint32_t)(uintptr_t)(action->conf);
> -			age_act = flow_aso_age_get_by_idx(dev, flow->age);
> -			__atomic_fetch_add(&age_act->refcnt, 1,
> -					   __ATOMIC_RELAXED);
> +			owner_idx = (uint32_t)(uintptr_t)action->conf;
> +			age_act = flow_aso_age_get_by_idx(dev, owner_idx);
> +			if (flow->age == 0) {
> +				flow->age = owner_idx;
> +				__atomic_fetch_add(&age_act->refcnt, 1,
> +						   __ATOMIC_RELAXED);
> +			}
>  			dev_flow->dv.actions[actions_n++] = age_act->dr_action;
>  			action_flags |= MLX5_FLOW_ACTION_AGE;
>  			break;