On Fri, 2022-02-18 at 10:08 -0600, Eric W. Biederman wrote:
> 
> Maybe I am reading the lifetimes wrong but is there
> any chance the code can just do something like the diff below?
> 
> AKA have a special version of kern_umount that does the call_rcu?
> 
> Looking at rcu_reclaim_tiny I think this use of mnt_rcu is valid.
> AKA reusing the rcu_head in the rcu callback.
> 
> 
> diff --git a/fs/namespace.c b/fs/namespace.c
> index 40b994a29e90..7d7aaef1592e 100644
> --- a/fs/namespace.c
> +++ b/fs/namespace.c
> @@ -4395,6 +4395,22 @@ void kern_unmount(struct vfsmount *mnt)
>  }
>  EXPORT_SYMBOL(kern_unmount);
>  
> +static void rcu_mntput(struct rcu_head *head)
> +{
> +       struct mount *mnt = container_of(head, struct mount,
> mnt_rcu);
> +       mntput(&mnt->mnt);
> +}
> +
> +void kern_rcu_unmount(struct vfsmount *mnt)
> +{
> +       /* release long term mount so mount point can be released */
> +       if (!IS_ERR_OR_NULL(mnt)) {
> +               struct mount *m = real_mount(mnt);
> +               m->mnt_ns = NULL;
> +               call_rcu(&m->mnt_rcu, rcu_mntput);
> +       }
> +}

OK, two comments here:

1) As Paul pointed out, we need to use call_rcu_work here,
   because rcu_mntput needs to run from a work item (or well,
   any process context) because of the rwlock.
2) No user of kern_unmount can use the vfsmount structure
   after kern_unmount returns. That means they could all use
   the RCU version, and no special version is needed.

Let me go test a patch that does that...

-- 
All Rights Reversed.