Re: [PATCH] RDMA/mlx5: Wait for in-flight page faults on implicit MR null_mkey dereg

[Michael Gur]

On 9/20/2026 12:06 PM, lirongqing wrote:
> External email: Use caution opening links or attachments
>
>
> From: Li RongQing <lirongqing@xxxxxxxxx>
>
> An implicit ODP MR (imr) publishes two mkeys into dev->odp_mkeys: the
> main imr->mmkey and imr->null_mmkey (MLX5_MKEY_NULL). Both are stored via
> mlx5r_store_odp_mkey(), which initialises their usecount to 1, and
> find_odp_mkey() takes a reference on whichever mkey it finds for the
> duration of a page fault.
>
> __mlx5_ib_dereg_mr() erases and waits on the *main* mmkey usecount only.
> For the null_mmkey, mlx5_ib_free_odp_mr() merely xa_erase()s it from
> odp_mkeys and calls mlx5_core_destroy_mkey() -- it never waits for an
> in-flight memory-scheme page fault (MLX5_MKEY_NULL) that holds a
> reference on null_mmkey. After xa_erase(), find_odp_mkey() stops
> returning the null_mmkey, but a fault already past the lookup still
> holds a reference and dereferences the imr (via container_of and
> pagefault_mr) after __mlx5_ib_dereg_mr() proceeds to kfree(mr).
>
> Mirror the main-mmkey handling: after xa_erase() of null_mmkey, call
> mlx5r_deref_wait_odp_mkey() to drop the reference taken at store time
> and wait for any in-flight fault to finish before destroying the mkey
> and freeing the imr.

Adding the wait after the implicit MR mkey was cleaned up and its umem 
was released doesn't solve the issue.

I suggest to make the implicit MR dereg wait for its null-mkey reference 
to drop, before any cleanup, or change how we handle the null-mkey: when 
a page fault hits it, take a reference on the implicit MR mkey instead.

> Fixes: 6f2487bfafce ("RDMA/mlx5: Add implicit MR handling to ODP memory scheme")
> Signed-off-by: Li RongQing <lirongqing@xxxxxxxxx>
> ---
>   drivers/infiniband/hw/mlx5/odp.c | 2 ++
>   1 file changed, 2 insertions(+)
>
> diff --git a/drivers/infiniband/hw/mlx5/odp.c b/drivers/infiniband/hw/mlx5/odp.c
> index b861861..f27ca77 100644
> --- a/drivers/infiniband/hw/mlx5/odp.c
> +++ b/drivers/infiniband/hw/mlx5/odp.c
> @@ -683,6 +683,8 @@ void mlx5_ib_free_odp_mr(struct mlx5_ib_mr *mr)
>                  xa_erase(&mr_to_mdev(mr)->odp_mkeys,
>                           mlx5_base_mkey(mr->null_mmkey.key));
>
> +               mlx5r_deref_wait_odp_mkey(&mr->null_mmkey);
> +
>                  mlx5_core_destroy_mkey(mr_to_mdev(mr)->mdev,
>                                         mr->null_mmkey.key);
>          }
> --
> 2.9.4