Re: [PATCH] edac, poll timeout cannot be zero

[Tejun Heo <tj@kernel.org>]

Hello,

On Wed, Feb 12, 2014 at 04:07:48PM +0100, Borislav Petkov wrote:
> While testing, however, I keep seeing the splat below and that's:
> 
>         if (WARN_ON(!list_empty(&work->entry))) {
>                 spin_unlock(&pwq->pool->lock);
>                 return;
>         }
> 
> and there seems to be some interference with edac_mc_workq_setup()
> which does mod_delayed_work() and then the workqueue callback
> edac_mc_workq_function() which does queue_delayed_work().
> 
> What I'm seeing in the splat is that when the timer fires to run the
> delayed work, __queue_work() complains that the work list is not empty
> even though we've done mod_delayed_work() which is supposed to cancel
> any pending work.
> 
> Tejun, any ideas what's happening? Do we need synchronization here or do
> you have a _sync version of mod_delayed_work() which makes sure any work
> is cancelled?

No, you don't need to.  All workqueue operations should be able to
synchronize with each other.

> Or does this mean that once the work is getting queued from the timer
> callback delayed_work_timer_fn, it cannot be cancelled anymore? Or
> something else I'm missing...?

Looking at edac_mc_workq_setup().... it contains INIT_DELAYED_WORK().
Does this race with other workqueue operations on the work item?  If
so, it of course breaks.  It's like doing spin_lock_init() while other
spinlock operations are in progress.

Thanks.

-- 
tejun