Re: docker crashes rcuos in __blkg_release_rcu

[Vivek Goyal <vgoyal@redhat.com>]

On Mon, Jun 09, 2014 at 01:47:08PM -0400, Vivek Goyal wrote:
> On Sun, Jun 08, 2014 at 06:22:00PM -0400, Joe Lawrence wrote:
> 
> [..]
> > Summary thus far:
> > 
> > R12: ffff88103c17a130 = struct rcu_head *rcu_head 
> > R13: ffff88103c17a080 = struct blkcg_gq *blkg
> >      ffff88103fc7df90 = struct request_queue *blkg->q (contains 0x6b
> >                                                        poison-pattern)
> > 
> > commit 2a4fd070 "blkcg: move bulk of blkcg_gq release operations to the
> > RCU callback" shuffled around some code in this space, introducing the
> > the calls to spin_[un]lock_irq(blkg->q->queue_lock).
> > 
> 
> Hi Joe,
> 
> Thanks for reporting and debugging this issue. So in summary it looks
> like that we have freed request queue associated with the blkg and
> when blkg is freed later and tries to access spin lock embedded in
> request queue, it crashes.
> 
> So the question is why request queue is being freed early. Are there any
> reference counting issues.

I am wondering if we need to take a reference on the queue
(blk_get_queue()) in blkg_alloc(), to make sure request queue is
still around when blkg is being freed.

I will try to reproduce the issue locally.

Thanks
Vivek