Re: [PATCH 07/11] blkcg: make request_queue bypassing on allocation

[Vivek Goyal <vgoyal-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>]

On Fri, Apr 13, 2012 at 02:38:52PM -0700, Tejun Heo wrote:
> On Fri, Apr 13, 2012 at 05:33:44PM -0400, Vivek Goyal wrote:
> > On Fri, Apr 13, 2012 at 02:05:48PM -0700, Tejun Heo wrote:
> > > On Fri, Apr 13, 2012 at 04:55:01PM -0400, Vivek Goyal wrote:
> > > > But neither seems to be the case here. So to make sure that blkg_lookup()
> > > > under rcu will see the updated value of queue flag (bypass), are we
> > > > relying on the fact that caller should see the DEAD flag and not go
> > > > ahead with blkg_lookup()?  If yes, atleast it is not obivious.
> > > 
> > > We're relying on the fact that it doesn't matter anymore because all
> > > blkgs will be shoot down in queue cleanup path which goes through rcu
> > > free, which is different from deactivating individual policies.  It
> > > indeed is subtle.  Umm... this is starting to get ridiculous.  Why the
> > > hell was megaraid messing with so many queues anyways?
> > 
> > Well, blkcg_deactivate_policy() frees the policy data in a non-rcu
> > manner. So group is around but policy data is gone. So technically if some
> > IO submitter does not see the queue bypass flag, he might still try to
> > access blkg->pd[pol->plid] after being freed.
> 
> No, we always go through blkg_destroy_all() and each blkg along with
> any attached policy_data will go through RCU grace period before
> getting destroyed.  It is stupid subtle but nevertheless correct.

Ok, I see that we are calling blkg_destroy_all() before we call
blk_throtl_exit() or elevator_exit(). So yes, this should be fine.

Thanks
Vivek