Re: [PATCH] flush_cpu_workqueue: don't flush an empty ->worklist

[Oleg Nesterov <oleg@tv-sign.ru>]

On 01/09, Srivatsa Vaddagiri wrote:
>
> On Tue, Jan 09, 2007 at 06:07:55PM +0300, Oleg Nesterov wrote:
> > but at some point we should thaw processes, including cwq->thread which
> > should die.
> 
> I am presuming we will thaw processes after all CPU_DEAD handlers have
> run.
> 
> > So we are doing things like take_over_work() and this is the
> > source of races, because the dead CPU is not on cpu_online_map.
> >
> > flush_workqueue() doesn't use any locks now. If we use freezer to implement
> > cpu-hotplug nothing will change, we still have races.
> 
> We have races -if- CPU_DEAD handling can run concurrently with a ongoing
> flush_workqueue. From my recent understanding of process freezer, this
> is not possible. In other words, flush_workqueue() can be its old
> implementation as below w/o any races:
> 
> 	some_thread:
> 
> 	for_each_online_cpu(i)
> 		flush_cpu_workqueue(i);
> 
> As long as this loop is running, cpu_down/up will not proceed. This means, 
> cpu_online_map is stable even if flush_cpu_workqueue blocks ..
> 
> Once this loop is complete and all threads have called try_to_freeze,
> cpu_down will proceed to change the bit map and run CPU_DEAD handlers
> of everyone. I am presuimg we will thaw processes only after all
> CPU_DEAD/ONLINE handlers have run.

We can't do this. We should thaw cwq->thread (which was bound to the
dead CPU) to complete CPU_DEAD event. So we still need some changes.

Oleg.