[PATCH] fuse: fix 'inval_wq' parameter setting during module initialisation

[PATCH] fuse: fix 'inval_wq' parameter setting during module initialisation

[Luis Henriques]

If the fuse module is loaded with a 'inval_wq' parameter the inval_wq_set()
function will schedule work into dentry_tree_work.  However this worker is
not initialised at this point because fuse_dentry_tree_init() will not have
been executed yet.

This can be fixed by delaying the work scheduling until later if the module
is still being initialised.

While there, also simplify the sched/cancel by using mod_delayed_work()
instead of schedule_delayed_work().  This will result in having the period
being changed immediately instead of waiting for the next reschedule.

Fixes: ab84ad597386 ("fuse: new work queue to periodically invalidate expired dentries")
Cc: stable@vger.kernel.org
Signed-off-by: Luis Henriques <luis@igalia.com>
---
Hi!

My initial approach to fix this was to use DECLARE_DELAYED_WORK() to
initialise the wq statically.  But this would then move the problem to
fuse_dentry_tree_work(), as the worker could still be scheduled before
the module finishes initialisation.

Cheers,
-- 
Luis

 fs/fuse/dir.c | 22 ++++++++++++++++------
 1 file changed, 16 insertions(+), 6 deletions(-)

diff --git a/fs/fuse/dir.c b/fs/fuse/dir.c
index e49b4e874b15..23c8259725c9 100644
--- a/fs/fuse/dir.c
+++ b/fs/fuse/dir.c
@@ -43,7 +43,6 @@ unsigned __read_mostly inval_wq;
 static int inval_wq_set(const char *val, const struct kernel_param *kp)
 {
 	unsigned int num;
-	unsigned int old = inval_wq;
 	int ret;
 
 	if (!val)
@@ -62,11 +61,14 @@ static int inval_wq_set(const char *val, const struct kernel_param *kp)
 
 	*((unsigned int *)kp->arg) = num;
 
-	if (num && !old)
-		schedule_delayed_work(&dentry_tree_work,
-				      secs_to_jiffies(num));
-	else if (!num && old)
-		cancel_delayed_work_sync(&dentry_tree_work);
+	/* Only start/stop workqueue if we're past module initialisation */
+	if (!module_is_coming(kp->mod)) {
+		if (num)
+			mod_delayed_work(system_percpu_wq, &dentry_tree_work,
+					 secs_to_jiffies(num));
+		else
+			cancel_delayed_work_sync(&dentry_tree_work);
+	}
 
 	return 0;
 }
@@ -224,6 +226,14 @@ void fuse_dentry_tree_init(void)
 		dentry_hash[i].tree = RB_ROOT;
 	}
 	INIT_DELAYED_WORK(&dentry_tree_work, fuse_dentry_tree_work);
+
+	/*
+	 * Start workqueue immediately if the module was loaded with the
+	 * 'inval_wq' parameter set
+	 */
+	if (inval_wq)
+		schedule_delayed_work(&dentry_tree_work,
+				      secs_to_jiffies(inval_wq));
 }
 
 void fuse_dentry_tree_cleanup(void)

Re: [PATCH] fuse: fix 'inval_wq' parameter setting during module initialisation

[Luis Henriques]

On Mon, Oct 19, 2026 at 01:53:29PM +0100, Luis Henriques wrote:
> If the fuse module is loaded with a 'inval_wq' parameter the inval_wq_set()
> function will schedule work into dentry_tree_work.  However this worker is
> not initialised at this point because fuse_dentry_tree_init() will not have
> been executed yet.
> 
> This can be fixed by delaying the work scheduling until later if the module
> is still being initialised.
> 
> While there, also simplify the sched/cancel by using mod_delayed_work()
> instead of schedule_delayed_work().  This will result in having the period
> being changed immediately instead of waiting for the next reschedule.
> 
> Fixes: ab84ad597386 ("fuse: new work queue to periodically invalidate expired dentries")
> Cc: stable@vger.kernel.org
> Signed-off-by: Luis Henriques <luis@igalia.com>
> ---
> Hi!
> 
> My initial approach to fix this was to use DECLARE_DELAYED_WORK() to
> initialise the wq statically.  But this would then move the problem to
> fuse_dentry_tree_work(), as the worker could still be scheduled before
> the module finishes initialisation.

Ping.

Any comments, or suggestion on a better way to fix this problem?

Cheers,
-- 
Luis

>  fs/fuse/dir.c | 22 ++++++++++++++++------
>  1 file changed, 16 insertions(+), 6 deletions(-)
> 
> diff --git a/fs/fuse/dir.c b/fs/fuse/dir.c
> index e49b4e874b15..23c8259725c9 100644
> --- a/fs/fuse/dir.c
> +++ b/fs/fuse/dir.c
> @@ -43,7 +43,6 @@ unsigned __read_mostly inval_wq;
>  static int inval_wq_set(const char *val, const struct kernel_param *kp)
>  {
>  	unsigned int num;
> -	unsigned int old = inval_wq;
>  	int ret;
>  
>  	if (!val)
> @@ -62,11 +61,14 @@ static int inval_wq_set(const char *val, const struct kernel_param *kp)
>  
>  	*((unsigned int *)kp->arg) = num;
>  
> -	if (num && !old)
> -		schedule_delayed_work(&dentry_tree_work,
> -				      secs_to_jiffies(num));
> -	else if (!num && old)
> -		cancel_delayed_work_sync(&dentry_tree_work);
> +	/* Only start/stop workqueue if we're past module initialisation */
> +	if (!module_is_coming(kp->mod)) {
> +		if (num)
> +			mod_delayed_work(system_percpu_wq, &dentry_tree_work,
> +					 secs_to_jiffies(num));
> +		else
> +			cancel_delayed_work_sync(&dentry_tree_work);
> +	}
>  
>  	return 0;
>  }
> @@ -224,6 +226,14 @@ void fuse_dentry_tree_init(void)
>  		dentry_hash[i].tree = RB_ROOT;
>  	}
>  	INIT_DELAYED_WORK(&dentry_tree_work, fuse_dentry_tree_work);
> +
> +	/*
> +	 * Start workqueue immediately if the module was loaded with the
> +	 * 'inval_wq' parameter set
> +	 */
> +	if (inval_wq)
> +		schedule_delayed_work(&dentry_tree_work,
> +				      secs_to_jiffies(inval_wq));
>  }
>  
>  void fuse_dentry_tree_cleanup(void)

Re: [PATCH] fuse: fix 'inval_wq' parameter setting during module initialisation

[Miklos Szeredi]

On Mon, 19 Oct 2026 at 14:52, Luis Henriques <luis@igalia.com> wrote:

> @@ -62,11 +61,14 @@ static int inval_wq_set(const char *val, const struct kernel_param *kp)
>
>         *((unsigned int *)kp->arg) = num;
>
> -       if (num && !old)
> -               schedule_delayed_work(&dentry_tree_work,
> -                                     secs_to_jiffies(num));
> -       else if (!num && old)
> -               cancel_delayed_work_sync(&dentry_tree_work);
> +       /* Only start/stop workqueue if we're past module initialisation */
> +       if (!module_is_coming(kp->mod)) {

See sashiko review:

- kp->mod might be NULL which will Oops module_is_coming()

- it's still racy against fuse_dentry_tree_init

My question is: does the inval_wq_set() even need to work before the
module is initialized?  Can't it just return an error and it wouldn't
need to deal with the race.

Thanks,
Miklos

Re: [PATCH] fuse: fix 'inval_wq' parameter setting during module initialisation

[Luis Henriques]

On Fri, Nov 06, 2026 at 12:33:50PM +0100, Miklos Szeredi wrote:
> On Mon, 19 Oct 2026 at 14:52, Luis Henriques <luis@igalia.com> wrote:
> 
> > @@ -62,11 +61,14 @@ static int inval_wq_set(const char *val, const struct kernel_param *kp)
> >
> >         *((unsigned int *)kp->arg) = num;
> >
> > -       if (num && !old)
> > -               schedule_delayed_work(&dentry_tree_work,
> > -                                     secs_to_jiffies(num));
> > -       else if (!num && old)
> > -               cancel_delayed_work_sync(&dentry_tree_work);
> > +       /* Only start/stop workqueue if we're past module initialisation */
> > +       if (!module_is_coming(kp->mod)) {
> 
> See sashiko review:

*sigh*
I keep forgetting to check sashiko website.  Sorry about that :-/

> - kp->mod might be NULL which will Oops module_is_coming()

Hmm... I guess this should be fixable by adding a '!kp->mod' to the
condition:

	if (!kp->mod || !module_is_coming(kp->mod)) ...

> - it's still racy against fuse_dentry_tree_init

This one is a bit trickier indeed.  I could try to play a bit with the
modules notifiers (register_module_notifier).  But I haven't used that
before, so not sure it can be used, for example, kicking off the worker
later when the module has been initialised.

> 
> My question is: does the inval_wq_set() even need to work before the
> module is initialized?  Can't it just return an error and it wouldn't
> need to deal with the race.

That's probably the simplest option, but that would prevent us from, for
example, setting inval_wq as a kernel parameter, right?  Or doing:

  # modprobe fuse inval_wq=10

Do you think that's acceptable?

Cheers,
--
Luis