On Wed, 2018-04-11 at 10:15 -0700, tj@xxxxxxxxxx wrote:
> On Wed, Apr 11, 2018 at 05:06:41PM +0000, Bart Van Assche wrote:
> > A simple and effective solution is to dissociate a request queue from the
> > block cgroup controller before blk_cleanup_queue() returns. This is why commit
> > a063057d7c73 ("block: Fix a race between request queue removal and the block
> > cgroup controller") moved the blkcg_exit_queue() call from __blk_release_queue()
> > into blk_cleanup_queue().
>
> which is broken. We can try to switch the lifetime model to revoking
> all live objects but that likely is a lot more involving than blindly
> moving blkg shootdown from release to cleanup. Implementing sever
> semantics is usually a lot more involved / fragile because it requires
> explicit participation from all users (exactly the same way revoking
> ->queue_lock is difficult).
>
> I'm not necessarily against switching to sever model, but what the
> patch did isn't that. It just moved some code without actually
> understanding or auditing what the implications are.
Hello Tejun,
Please explain what you wrote further. It's not clear to me why you think
that anything is broken nor what a "sever model" is.
I think we really need the blkcg_exit_queue() call in blk_cleanup_queue()
to avoid race conditions between request queue cleanup and the block cgroup
controller. Additionally, since it is guaranteed that no new requests will
be submitted to a queue after it has been marked dead I don't see why it
would make sense to keep the association between a request queue and the
blkcg controller until the last reference on a queue is dropped.
Bart.
