Re: [PATCH] fix-flush_workqueue-vs-cpu_dead-race-update

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

On Sun, Jan 07, 2007 at 11:59:57AM -0800, Andrew Morton wrote:
> > How would this provide a stable access to cpu_online_map in functions
> > that need to block while accessing it (as flush_workqueue requires)?
> 
> If a thread simply blocks, that will not permit a cpu plug/unplug to proceed.
> 
> The thread had to explicitly call try_to_freeze().  CPU plug/unplug will
> not occur (and cpu_online_map will not change) until every process in the
> machine has called try_to_freeze()).

Maybe my misunderstanding of the code, but:

Looking at the code, it seems to me that try_to_freeze() will be called very 
likely from the signal-delivery path (get_signal_to_deliver()). Isnt
that correct? If so, consider a thread as below:


 Thread1 :
	for_each_online_cpu() { online_map = 0x1111 at this point }
		do_some_thing();
			kmalloc(); <- blocks


Can't Thread1 be frozen in the above blocking state w/o it voluntarily 
calling try_to_freeze?  If so, online_map can change when it returns
from kmalloc() ..

> So the problem which you're referring to will only occur if a workqueue
> callback function calls try_to_freeze(), which would be mad.
> 
> 
> Plus flush_workqueue() is on the way out.  We're slowly edging towards a
> working cancel_work() which will only block if the work which you're trying
> to cancel is presently running.  With that, pretty much all the
> flush_workqueue() calls go away, and all these accidental rarely-occurring
> deadlocks go away too.

Fundamentally, I think it is important to give the ability to block
concurrent hotplug operations from happening ..

-- 
Regards,
vatsa