Re: 2.6.26-rc4: RIP __call_for_each_cic+0x20/0x50

[Jens Axboe <jens.axboe@oracle.com>]

On Thu, May 29 2008, Paul E. McKenney wrote:
> On Thu, May 29, 2008 at 12:13:54PM +0200, Jens Axboe wrote:
> > On Thu, May 29 2008, Paul E. McKenney wrote:
> > > On Thu, May 29, 2008 at 08:42:02AM +0200, Jens Axboe wrote:
> > > > On Thu, May 29 2008, Jens Axboe wrote:
> > > > > > But one additional question...
> > > > > > 
> > > > > > 	static void cfq_cic_free_rcu(struct rcu_head *head)
> > > > > > 	{
> > > > > > 		struct cfq_io_context *cic;
> > > > > > 
> > > > > > 		cic = container_of(head, struct cfq_io_context, rcu_head);
> > > > > > 
> > > > > > 		kmem_cache_free(cfq_ioc_pool, cic);
> > > > > > 		elv_ioc_count_dec(ioc_count);
> > > > > > 
> > > > > > 		if (ioc_gone && !elv_ioc_count_read(ioc_count))
> > > > > > 			complete(ioc_gone);
> > > > > > 	}
> > > > > > 
> > > > > > Suppose that a pair of tasks both execute the elv_ioc_count_dec()
> > > > > > at the same time, so that all counters are now zero.  Both then
> > > > > > find that there is still an ioc_gone, and that the count is
> > > > > > now zero.  One of the tasks invokes complete(ioc_gone).  This
> > > > > > awakens the corresponding cfq_exit(), which now returns, getting
> > > > > > rid of its stack frame -- and corrupting the all_gone auto variable
> > > > > > that ioc_gone references.
> > > > > > 
> > > > > > Now the second task gets a big surprise when it tries to invoke
> > > > > > complete(ioc_gone).
> > > > > > 
> > > > > > Or is there something else that I am missing here?
> > > > > 
> > > > > No, I think that's a problem spot as well. To my knowledge, nobody has
> > > > > ever hit that. The anticipatory scheduler has the same code.
> > > > > 
> > > > > What we want to avoid here is making cfq_cic_free_rcu() a lot more
> > > > > expensive, which is why the elv_ioc_count_read() is behind that
> > > > > ioc_gone check. I'll need to think a bit on how to handle that
> > > > > better :-)
> > > > 
> > > > So how about this? Add a spinlock for checking and clearing ioc_gone
> > > > back to NULL. It doesn't matter if we make the ioc_gone != NULL
> > > > case a little more expensive, as it will only happen on cfq-iosched
> > > > module unload. And it seems the clearest way of making this safe.
> > > > The last hunk should really not be necessary, as ioc_gone wont be
> > > > set back to NULL before wait_for_completion() is entered.
> > > 
> > > Looks better!  I do have one scenario that seems troublesome, but
> > > it should be easy to fix, see below.  (Assuming it really is a
> > > problem, that is...)
> > > 
> > > 						Thanx, Paul
> > > 
> > > > An identical patch is needed in AS as well.
> > > > 
> > > > diff --git a/block/cfq-iosched.c b/block/cfq-iosched.c
> > > > index d01b411..32aa367 100644
> > > > --- a/block/cfq-iosched.c
> > > > +++ b/block/cfq-iosched.c
> > > > @@ -48,6 +48,7 @@ static struct kmem_cache *cfq_ioc_pool;
> > > > 
> > > >  static DEFINE_PER_CPU(unsigned long, ioc_count);
> > > >  static struct completion *ioc_gone;
> > > > +static DEFINE_SPINLOCK(ioc_gone_lock);
> > > > 
> > > >  #define CFQ_PRIO_LISTS		IOPRIO_BE_NR
> > > >  #define cfq_class_idle(cfqq)	((cfqq)->ioprio_class == IOPRIO_CLASS_IDLE)
> > > > @@ -1177,8 +1178,19 @@ static void cfq_cic_free_rcu(struct rcu_head *head)
> > > >  	kmem_cache_free(cfq_ioc_pool, cic);
> > > >  	elv_ioc_count_dec(ioc_count);
> > > > 
> > > > -	if (ioc_gone && !elv_ioc_count_read(ioc_count))
> > > > -		complete(ioc_gone);
> > > > +	if (ioc_gone) {
> > > > +		/*
> > > > +		 * CFQ scheduler is exiting, grab exit lock and check
> > > > +		 * the pending io context count. If it hits zero,
> > > > +		 * complete ioc_gone and set it back to NULL
> > > > +		 */
> > > 
> > > Suppose that at this point some other CPU does the last complete().
> > > They have set ioc_gone to NULL, so everything is fine.  But suppose
> > > that in the meantime, some other CPU sets up a cfq and then starts
> > > tearing it down.  Then ioc_gone would be non-NULL, and we would cause
> > > this new teardown to end prematurely.
> > > 
> > > If this is a real problem, one way to get around it is to have a
> > > generation number.  We capture this before doing the elv_ioc_count_dec()
> > > (alas, with a memory barrier between the capture and the elv_ioc_count_dec()),
> > > and then check it under the lock.  If it has changed, we know someone else
> > > has already done the awakening for us.  Increment the generation number
> > > in the same place that ioc_gone is set to NULL.
> > > 
> > > Seem reasonable?
> > 
> > This isn't a problem, since cfq_exit() cannot be called before
> > all block queues in the system have been detached from CFQ.
> 
> And once all block queues have been detached, no future block queues
> can ever be attached again?  Or perhaps a better way of putting it,
> once CFQ has been shut down, can it be restarted without rebooting
> the system?  If it can be restarted without reboot, then I do not
> yet see how the scenario above is avoided.

Certainly, you can modprobe cfq-iosched and switch any device to cfq
again. But I really don't see the problem in this case. Once CFQ
has exited, all tasks are detached from CFQ. They have to be, or
we could not unload the module.

> > cfq_exit() calls elv_unregister() before setting ioc_gone, so
> > when elv_unregister() has returned, CFQ is in its own little world.
> > Do we need an smp_wmb() between elv_unregister() and the ioc_gone
> > assignment to ensure this ordering as well? IIRC, the spin_lock
> > and spin_unlock in elv_unregister() isn't enough to guarentee this.
> > We are really down to splitting hairs now, but better safe than
> > sorry :-)
> 
> I believe that the spinlock takes care of that ordering issue.  I am
> instead worried about a "Rip Van Winkle" effect where a given task
> is delayed at a crucial point.  By the time it wakes back up, CFQ
> has been not only restarted, but is now in the process of being torn
> down again.  (Assuming that it can in fact be restarted without
> a reboot.)

Why would we have done the ioc dec and complete() before that task
had finished?

-- 
Jens Axboe