Re: current linux-2.6.git: cpusets completely broken

["Dmitry Adamushko" <dmitry.adamushko@gmail.com>]

2008/7/13 Dmitry Adamushko <dmitry.adamushko@gmail.com>:
> 2008/7/13 Linus Torvalds <torvalds@linux-foundation.org>:
>>
>>> (A little pet horse for the occasion: Testing can show the presence of
>>> errors, but not their absence. But that's a different story.)
>>
>> Absolutely. Which is actually why I prefer my patch. I think it fixes - in
>> general - the issue of CPU migration migrating tasks back to the CPU that
>> we're taking down.
>>
>> The other patches seem to work around just the problem that _triggers_ the
>> bug. They don't actually make it impossible to migrate to a CPU that is
>> getting shut down - they just try to avoid the particular sequence that
>> made it happen for you.
>
> Well, they try to make sched-domains consistent for all possible
> cases, not just any particular case. So no, they don't allow a
> possibility to leave tasks on a dead CPU (unless there is another
> bug).
>
> With your patch (and a cpusets :: hotplug handler from the current
> -git) sched-domains are still broken and they are used in a number of
> places. So why keep them at all?
>
> I'm really surprised that Vegard says this "cpu_active_map" patch
> alone fixes the problem.
>
> With your modifications of common_cpu_mem_hotplug_unplug() - yes

No, not even with your modifications of common_mem_hotplug_unplug().

-       if (phase == CPU_DYING || phase == CPU_DYING_FROZEN)
+       if (phase == CPU_DYING || phase == CPU_DYING_FROZEN ||
+           phase == CPU_UP_PREPARE || phase == CPU_UP_PREPARE_FROZEN)
               return NOTIFY_DONE;

You should have added "phase == CPU_DOWN_PREPARE || phase ==
CPU_DOWN_PREPARE_FROZEN" additionally. So I'm really surprised by
Vegard's assertion :-)


 but
> then it will work even without "cpu_active_map".
>
> (ok, unless I'm really blind at this late hour so please direct me to
> the right way :-)
>
>
>>
>>                        Linus
>>
>
>
> --
> Best regards,
> Dmitry Adamushko
>



-- 
Best regards,
Dmitry Adamushko