[PATCH] RDMA/mlx5: Fix refcount inconsistency in mlx5_netdev_event

[PATCH] RDMA/mlx5: Fix refcount inconsistency in mlx5_netdev_event

[Shuhao Fu]

Fix refcount inconsistency related to `mlx5_ib_get_native_port_mdev`.

Function `mlx5_ib_get_native_port_mdev` could increase the counter of 
`mpi->mdev_refcnt` if mpi is not master. To ensure refcount consistency,
each call to `mlx5_ib_get_native_port_mdev` should have a corresponding
call to `mlx5_ib_put_native_port_mdev`. In `mlx5_netdev_event`, two
branches fail to do so, leading to a possible bug when unbinding.

Fixes: 379013776222 ("RDMA/mlx5: Handle link status event only for LAG device")
Fixes: 35b0aa67b298 ("RDMA/mlx5: Refactor netdev affinity code")
Signed-off-by: Shuhao Fu <sfual@cse.ust.hk>
---
 drivers/infiniband/hw/mlx5/main.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/drivers/infiniband/hw/mlx5/main.c b/drivers/infiniband/hw/mlx5/main.c
index fc1e86f6c..0c4aa7c50 100644
--- a/drivers/infiniband/hw/mlx5/main.c
+++ b/drivers/infiniband/hw/mlx5/main.c
@@ -247,7 +247,7 @@ static int mlx5_netdev_event(struct notifier_block *this,
 
 		if (!netif_is_lag_master(ndev) && !netif_is_lag_port(ndev) &&
 		    !mlx5_core_mp_enabled(mdev))
-			return NOTIFY_DONE;
+			goto done;
 
 		if (mlx5_lag_is_roce(mdev) || mlx5_lag_is_sriov(mdev)) {
 			struct net_device *lag_ndev;
@@ -268,7 +268,7 @@ static int mlx5_netdev_event(struct notifier_block *this,
 		if (ibdev->is_rep)
 			roce = mlx5_get_rep_roce(ibdev, ndev, upper, &port_num);
 		if (!roce)
-			return NOTIFY_DONE;
+			goto done;
 
 		ib_ndev = ib_device_get_netdev(&ibdev->ib_dev, port_num);
 
-- 
2.39.5 (Apple Git-154)

Re: [PATCH] RDMA/mlx5: Fix refcount inconsistency in mlx5_netdev_event

[Leon Romanovsky]

On Tue, Oct 21, 2025 at 04:40:58PM +0000, Shuhao Fu wrote:
> Fix refcount inconsistency related to `mlx5_ib_get_native_port_mdev`.
> 
> Function `mlx5_ib_get_native_port_mdev` could increase the counter of 
> `mpi->mdev_refcnt` if mpi is not master. To ensure refcount consistency,
> each call to `mlx5_ib_get_native_port_mdev` should have a corresponding
> call to `mlx5_ib_put_native_port_mdev`. In `mlx5_netdev_event`, two
> branches fail to do so, leading to a possible bug when unbinding.
> 
> Fixes: 379013776222 ("RDMA/mlx5: Handle link status event only for LAG device")
> Fixes: 35b0aa67b298 ("RDMA/mlx5: Refactor netdev affinity code")
> Signed-off-by: Shuhao Fu <sfual@cse.ust.hk>
> ---
>  drivers/infiniband/hw/mlx5/main.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/drivers/infiniband/hw/mlx5/main.c b/drivers/infiniband/hw/mlx5/main.c
> index fc1e86f6c..0c4aa7c50 100644
> --- a/drivers/infiniband/hw/mlx5/main.c
> +++ b/drivers/infiniband/hw/mlx5/main.c
> @@ -247,7 +247,7 @@ static int mlx5_netdev_event(struct notifier_block *this,
>  
>  		if (!netif_is_lag_master(ndev) && !netif_is_lag_port(ndev) &&
>  		    !mlx5_core_mp_enabled(mdev))
> -			return NOTIFY_DONE;
> +			goto done;

mdev_refcnt is increased only for MPV slave, but here we are checking if
MPV is enabled. It is not in this leaf and "return NOTIFY_DONE" is the
right thing to do.

>  
>  		if (mlx5_lag_is_roce(mdev) || mlx5_lag_is_sriov(mdev)) {
>  			struct net_device *lag_ndev;
> @@ -268,7 +268,7 @@ static int mlx5_netdev_event(struct notifier_block *this,
>  		if (ibdev->is_rep)
>  			roce = mlx5_get_rep_roce(ibdev, ndev, upper, &port_num);
>  		if (!roce)
> -			return NOTIFY_DONE;
> +			goto done;

It is not MPV flow either. There is nothing to change here.

Thanks

>  
>  		ib_ndev = ib_device_get_netdev(&ibdev->ib_dev, port_num);
>  
> -- 
> 2.39.5 (Apple Git-154)
>