Re: [PATCH 2/9] blkcg: drop unnecessary RCU locking

[Vivek Goyal <vgoyal@redhat.com>]

On Thu, Feb 16, 2012 at 02:37:51PM -0800, Tejun Heo wrote:

[..]
> * queue_lock coverage extended to cover @blkg usage in
>   blkio_policy_parse_and_set() and RCU dropped.  This means all config
>   update callbacks are now called under queue_lock.
> 

[..]
> @@ -1041,11 +1034,8 @@ static int blkio_policy_parse_and_set(char *buf, enum blkio_policy_id plid,
>  	if (!disk || part)
>  		goto out;
>  
> -	rcu_read_lock();
> -
>  	spin_lock_irq(disk->queue->queue_lock);
>  	blkg = blkg_lookup_create(blkcg, disk->queue, plid, false);
> -	spin_unlock_irq(disk->queue->queue_lock);
>  

So now in some cases we call blkg_lookup_create() with both queue and rcu
read lock held (cfq_lookup_create_cfqg()) and in this case hold only queue
lock. blkg_lookup_create() calls blkg_lookup() which expects a rcu_read_lock()
to be held and we will be travesing that list without rcu_read_lock()
held. Isn't that a problem? We might be examining a blkg belonging to
a different queue and it might be being freed parallely.

Or blkg destruction in this cgroup is serialized by cgroup_mutex() or
by something else in this policy parse and set path?

Thanks
Vivek