Re: [PATCH v8 01/10] kernfs: Remove reference counting for kernfs_open_node.

[Tejun Heo]

On Tue, Apr 26, 2022 at 11:43:38AM +1000, Imran Khan wrote:
> Hello Tejun,
> 
> On 23/4/22 2:03 am, Tejun Heo wrote:
> > On Sun, Apr 10, 2022 at 12:37:10PM +1000, Imran Khan wrote:
> >> @@ -768,15 +765,15 @@ void kernfs_drain_open_files(struct kernfs_node *kn)
> >>  	if (!(kn->flags & (KERNFS_HAS_MMAP | KERNFS_HAS_RELEASE)))
> >>  		return;
> >>  
> >> -	spin_lock_irq(&kernfs_open_node_lock);
> >>  	on = kn->attr.open;
> >> -	if (on)
> >> -		atomic_inc(&on->refcnt);
> >> -	spin_unlock_irq(&kernfs_open_node_lock);
> >>  	if (!on)
> >>  		return;
> >>
> >>  	mutex_lock(&kernfs_open_file_mutex);
> >> +	if (!kn->attr.open) {
> >> +		mutex_unlock(&kernfs_open_file_mutex);
> >> +		return;
> >> +	}
> > 
> > What if @on got freed and new one got allocated between the lockless check
> > and the locked check? Is there a reason to keep the lockless check at all?
> 
> The only reason for lockless check is to opportunistically check and
> return if ->attr.open is already NULL, without waiting to acquire the
> mutex. This is because no one will be adding to ->attr.open at this
> point of time.
> But we can live with just the locked check as well.
> Please let me know if you think of lockless check as an overkill in this
> case.

The code is just wrong. You can end up:

        on = kn->attr.open;
        if (!on)
                return;

        // we get preempted here and someone else puts @on and then
        // recreates it

        mutex_lock();
        if (!kn->attr.open) {
                mutex_unlock();
                return;
        }

        // we're here but @on is a pointer to an already freed memory

Thanks.

-- 
tejun