Re: [PATCH 4/5] mm/swapfile: refcount block and queue before using blkcg_schedule_throttle()

[Christoph Hellwig <hch@infradead.org>]

On Wed, Apr 15, 2020 at 01:19:15PM +0000, Luis Chamberlain wrote:
> >  	if (current->throttle_queue)
> >  		return;
> > +	if (unlikely(current->flags & PF_KTHREAD))
> > +		return;
> >  
> >  	spin_lock(&swap_avail_lock);
> >  	plist_for_each_entry_safe(si, next, &swap_avail_heads[node],
> >  				  avail_lists[node]) {
> > -		if (si->bdev) {
> > -			blkcg_schedule_throttle(bdev_get_queue(si->bdev),
> > -						true);
> > -			break;
> > +		if (!si->bdev)
> > +			continue;
> > +		if (blk_get_queue(dev_get_queue(si->bdev))) {
> > +			current->throttle_queue = dev_get_queue(si->bdev);
> > +			current->use_memdelay = true;
> > +			set_notify_resume(current);
> >  		}
> > +		break;
> >  	}
> >  	spin_unlock(&swap_avail_lock);
> >  }
> 
> Sorry, its not clear to me  who calls the respective blk_put_queue()
> here?

If you look at blkcg_schedule_throttle, it only puts the queue that
was in current->throttle_queue.  But mem_cgroup_throttle_swaprate
exits early when current->throttle_queue is non-zero (first two lines
quote above).  So when called from mem_cgroup_throttle_swaprate,
blkcg_schedule_throttle should never actually put a queue.  Open
coding the few relevant lines from blkcg_schedule_throttle in
mem_cgroup_throttle_swaprate makes that obvious.