Re: Scheduler Bug (set_cpus_allowed)

[Mike Kravetz]

On Tue, Jun 11, 2002 at 12:35:38AM +0200, Ingo Molnar wrote:
> 
> agreed. I removed the _sync code mainly because there was no
> idle_resched() to migrate a task actively, and the migration bits i tried
> were incomplete. But with your above conditional it should cover all the
> practical cases we care about, in an elegant way.
> 
> i ported your sync wakeup resurrection patch to 2.5.21 (attached). I did
> some modifications:
> 
> - wake_up() needs to check (rq->curr != p) as well, not only !p->array.
> 
> - make __wake_up_sync dependent on CONFIG_SMP
> 
> - export __wake_up_sync().
> 
> (the attached patch includes both the ->frozen change plus the sync wakeup
> resurrection, it's against vanilla 2.5.21.)
> 
> appears to work for me just fine (compiles, boots and works under SMP & UP
> alike), and does the trick for bw_pipe and lat_pipe. Comments?
> 
> 	Ingo

Great!  Thanks!

You might also consider adding the optimization/fast path to
set_cpus_allowed().  Once again, I don't expect this routine
(or this code path) to be used much, but I just hate to see
us scheudle a migration task to set the cpu field when it is
safe to do it within the routine.

-- 
Mike