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:
> Now that blkg additions / removals are always done under both q and
> blkcg locks, the only place RCU locking is used is blkg_lookup() for
> lockless lookup.  This patch drops unncessary RCU locking replacing it
> with plain blkcg / q locking as necessary.
> 
> * blkg_lookup_create() and blkiocg_pre_destroy() already perform
>   proper locking and don't need RCU.  Dropped.

But blkg_lookup_create() is called under rcu() to protect blkcg pointer.
And blkg_lookup() is also happening under same rcu read lock. So I think
you can't drop rcu from blkg_lookup_create().

>  {
>  	struct blkio_group *blkg, *new_blkg;
>  
> -	WARN_ON_ONCE(!rcu_read_lock_held());

Don't we want to be called with rcu lock held needed for blkg_lookup()?

>  	lockdep_assert_held(q->queue_lock);
>  
>  	/*

[..]
> @@ -581,11 +580,9 @@ struct blkio_group *blkg_lookup_create(struct blkio_cgroup *blkcg,
>  	 * allocation is fixed.
>  	 */
>  	spin_unlock_irq(q->queue_lock);
> -	rcu_read_unlock();
>  
>  	new_blkg = blkg_alloc(blkcg, q);
>  
> -	rcu_read_lock();
>  	spin_lock_irq(q->queue_lock);

blkg_alloc() might sleep here with rcu lock held?

Thanks
Vivek