lkml.org 
[lkml]   [2002]   [Feb]   [21]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH] O(1) scheduler set_cpus_allowed for non-current tasks
Ingo,

thanks for the patch, I am glad to have an "official" solution! Though
I'll wait for your 2.4.X patch to give it a try on Itanium SMP
systems. But this lazy migration scheme looks nice and safe.

Best regards,
Erich


On Thu, 21 Feb 2002, Ingo Molnar wrote:

>
> On Thu, 21 Feb 2002, Erich Focht wrote:
>
> > well I whished this kind of feedback came at my first attempt to have
> > a discussion on this subject. Would have saved me lots of reboots :-)
>
> i did not find out about the most fundamental problem until today, when i
> tried to merge your patch. Much of Linux's development is trial & err, i
> do that all the time as well.
>
> > The reaction those days was:
> >
> > Ingo> your patch does not solve the problem, the situation is more
> > Ingo> complex. What happens if the target task is not 'current' and is
> > Ingo> running on some other CPU? If we send the migration interrupt then
> > Ingo> nothing guarantees that the task will reschedule anytime soon, so
> > Ingo> the target CPU will keep spinning indefinitely. There are other
> > Ingo> problems too, like crossing calls to set_cpus_allowed(), etc. Right
> > Ingo> now set_cpus_allowed() can only be used for
> > Ingo> the current task, and must be used by kernel code that knows what it
> > Ingo> does.
> >
> > and later:
> >
> > Ingo> well, there is a way, by fixing the current mechanizm. But since
> > Ingo> nothing uses it currently it wont get much testing. I only pointed
> > Ingo> out that the patch does not solve some of the races.
> >
> > So I kept Ingo's design idea of sending IPIs. And I made it survive
> > crossing calls and avoid spinning around for long time, specially in
> > interrupt.
>
> the fundamental problem is with wait_task_inactive() called from IRQ
> contexts - it can spin indefinitely. The above comments list some of the
> other problems, and those can indeed be solved by extending the IPI
> delivery mechanism, which you did.
>
> migration threads are a variation of the current scheduler as well -the
> schedule()/wakeup() hotpath did not have to be touched.
>
> Ingo

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2005-03-22 13:24    [W:0.039 / U:6.948 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site