RDMA/mlx5: Remove the redundant MLX5_IB_STAGE_UAR stage
authorYishai Hadas <yishaih@nvidia.com>
Tue, 13 May 2025 11:02:40 +0000 (14:02 +0300)
committerLeon Romanovsky <leon@kernel.org>
Sun, 18 May 2025 07:52:03 +0000 (03:52 -0400)
commit972db388d40ded1a5ef8ce09d92ef1e2b9e40f07
tree3bc3057535b12b98b8d6af26fa6deec358a870c2
parent21508c8c972ca0ff06b07af37adb4021ab527de2
RDMA/mlx5: Remove the redundant MLX5_IB_STAGE_UAR stage

The MLX5_IB_STAGE_UAR stage in the RDMA driver is redundant and should
be removed.

Responsibility for initializing the device's UAR pointer
(mdev->priv.uar) lies with mlx5_core, which already sets it during the
mlx5_load() process.

At present, the RDMA UAR stage overwrites this pointer, which was
correctly initialized by mlx5_core, creating the risk of inconsistency.

Ownership and management of the UAR pointer should remain exclusively
within mlx5_core.

In the current upstream code, we luckily receive the same pointer, since
mlx5_get_uars_page() still finds available BF registers for that UAR,
allowing it to be shared.

However, future changes in mlx5_core may expose this flaw.
For instance, if mlx5_alloc_bfreg() is invoked twice before the RDMA UAR
stage runs, the RDMA driver may overwrite the UAR allocated by
mlx5_core.

This could lead to real bugs. For example, if mlx5_ib is unloaded
(rmmod), it might free the UAR, leaving mlx5_core with a dangling
reference to an invalid UAR.

Signed-off-by: Yishai Hadas <yishaih@nvidia.com>
Reviewed-by: Fan Li <fanl@nvidia.com>
Link: https://patch.msgid.link/feaa84ec6f20468b4935c439923e9266122a93d0.1747134130.git.leon@kernel.org
Signed-off-by: Leon Romanovsky <leon@kernel.org>
drivers/infiniband/hw/mlx5/main.c
drivers/infiniband/hw/mlx5/mlx5_ib.h