Re: workqueue code needing preemption disabled

[Tejun Heo <tj@kernel.org>]

On Mon, Mar 18, 2013 at 01:08:07PM -0400, Steven Rostedt wrote:
> On Mon, 2013-03-18 at 09:43 -0700, Tejun Heo wrote:
> 
> > Making gcwq locks disable preemption would be much safer / easier, but
> > if that's not desirable, anything touching gcwq->idle_list would be a
> > good place to start - worker_enter_idle() and worker_leave_idle().
> > Hmmm... ignoring CPU hotplug, I think those two might just do it.
> > Give it a try?  How reproducible is the problem?
> 
> Not very :-( I triggered it twice on a 40 CPU box. It can go
> approximately 1 month before it triggers. And the box we are testing on
> is currently a loaner, and we have it on extension right now. Which
> means we wont have it much longer.
> 
> But perhaps that's the place to fix things.

I've been thinking about it and AFAICS the only way that BUG_ON()
could trigger from preemption is if preemption happens while the
idle_list head is becoming or stopping being empty.
ie. pool->worklist is half updated so list_empty() isn't true but the
first next entry is already pointing back to itself.  If there's a
crashdump, it shouldn't be too difficult to verify and wrapping the
above two functions should resolve it.

Thanks.

-- 
tejun