Re: [PATCH] kernfs: recheck of->released after acquiring the active reference

[Tejun Heo]

Hello,

On Thu, Aug 20, 2026 at 02:18:10AM +0000, Fan Wu wrote:
> @@ -73,12 +73,22 @@ static struct kernfs_open_node *of_on(struct kernfs_open_file *of)
>  /* Get active reference to kernfs node for an open file */
>  static struct kernfs_open_file *kernfs_get_active_of(struct kernfs_open_file *of)
>  {
> -	/* Skip if file was already released */
> -	if (unlikely(of->released))
> +	if (!kernfs_get_active(of->kn))
>  		return NULL;
>  
> -	if (!kernfs_get_active(of->kn))
> +	/*
> +	 * @of->released is set under kernfs_open_file_mutex.  While the
> +	 * active reference is held, @kn can't be drained anymore and
> +	 * kernfs_fop_release() can't run, so re-reading @of->released
> +	 * here settles whether @of was released for good.
> +	 */
> +	mutex_lock(kernfs_open_file_mutex_ptr(of->kn));
> +	if (unlikely(of->released)) {
> +		mutex_unlock(kernfs_open_file_mutex_ptr(of->kn));
> +		kernfs_put_active(of->kn);
>  		return NULL;
> +	}
> +	mutex_unlock(kernfs_open_file_mutex_ptr(of->kn));

"lock -> test something -> unlock" pattern is usually unnecessary. What's
the mutex achieving doing? Doesn't kernfs_get_active() already have strong
enough memory barrier? If not, it'd be better to solve it by iterlocking
kernfs_get_active() and released clearing.

Thanks.

-- 
tejun