Re: workqueue deadlock

[Srivatsa Vaddagiri <vatsa@in.ibm.com>]

On Wed, Dec 06, 2006 at 05:26:14PM -0700, Bjorn Helgaas wrote:
> loadkeys is holding the cpu_hotplug lock (acquired in flush_workqueue())
> and waiting in flush_cpu_workqueue() until the cpu_workqueue drains.
> 
> But events/4 is responsible for draining it, and it is blocked waiting
> to acquire the cpu_hotplug lock.
> 
> In current upstream, the cpu_hotplug lock has been replaced with
> workqueue_mutex, but it looks to me like the same deadlock is still
> possible.

Yes I think so too.

> Is there some rule that workqueue functions shouldn't try to
> flush a workqueue?  

In general, workqueue functions wanting to flush workqueue seems wierd
to me. But in this case, I think the requirement is to block until all
queued work is complete, which is what flush_workqueue is supposed to
do. Hence I dont see any way to avoid it ..

> Or should flush_workqueue() be smarter by
> releasing the workqueue_mutex once in a while?

IMHO, rehauling lock_cpu_hotplug() to support scenarios like this is a
better approach. 

	- Make it rw-sem
	- Make it per-cpu mutex, which could be either:

		http://lkml.org/lkml/2006/11/30/110 - Ingo's suggestion
		http://lkml.org/lkml/2006/10/26/65 - Gautham's work based on RCU

In Ingo's suggestion, I really dont know if the task_struct
modifications is a good thing (to support recursive requirements).
Gautham's patches avoid modifications to task_struct, is fast but can
starve writers (who want to bring down/up a CPU).

-- 
Regards,
vatsa