Messages in this thread | | | Date | Wed, 26 Oct 2011 12:25:48 -0700 | From | Tejun Heo <> | Subject | Re: [PATCH 4/6] block: add blk_queue_dead() |
| |
Hello,
On Wed, Oct 26, 2011 at 01:20:49PM -0400, Vivek Goyal wrote: > On Tue, Oct 25, 2011 at 06:02:05PM -0700, Tejun Heo wrote: > > @@ -603,7 +603,7 @@ EXPORT_SYMBOL(blk_init_allocated_queue_node); > > > > int blk_get_queue(struct request_queue *q) > > { > > - if (likely(!test_bit(QUEUE_FLAG_DEAD, &q->queue_flags))) { > > + if (likely(!blk_queue_dead(q))) { > > kobject_get(&q->kobj); > > return 0; > > I thought DEAD flag is now synchronized with queue lock. So the protocol > is that caller should be holding queue lock here first?
The requirement is that issue and processing of requests don't happen once DEAD is set and to guarantee that it's necessary to set DEAD and check DEAD in rq alloc/issue paths.
blk_get_queue() is inherently opportunistic as holding a reference doesn't prevent it from being killed. It doesn't make any sense to require its holder to grab spinlock - either the caller doesn't need a refcnt (as it's holding spinlock) or the result of the check becomes stale as soon as it drops the spinlock. Whether testing DEAD is helpful (it isn't necessary per-se) is another question tho.
Thanks.
-- tejun
| |