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

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

2008/7/13 Linus Torvalds <torvalds@linux-foundation.org>:
>
> On Sun, 13 Jul 2008, Linus Torvalds wrote:
>>
>> The thing is, we should fix the top level code to never even _consider_ an
>> invalid CPU as a target, and that in turn should mean that all the other
>> code should be able to just totally ignore CPU hotplug events.
>
> IOW, I think we should totally remove the whole "update_sched_domains()"
> thing too. Any logic that needs it is broken. We shouldn't detach the
> scheduler domains in DOWN_PREPARE (much less UP_PREPARE), we should just
> leave them damn well alone.
>
> As the comment says, "The domains and groups cannot be updated in place
> without racing with the balancing code". The thing is, we shouldn't even
> try. The correct way to handle all this is to make the balancing code use
> the domains regardless, but protect against CPU's going down with
> _another_ data structure that is much easier to update.
>
> Namely something like 'cpu_active_map'.
>
> Then we just get rid of all the crap in update_sched_domains() entirely,
> and then we can make the cpusets code do the *sane* thing, which is to
> rebuild the scheduler domains only when the CPU up/down has completed.
>
> So instead of this illogical and crazy mess:
>
>        +       switch (phase) {
>        +       case CPU_UP_CANCELED:
>        +       case CPU_UP_CANCELED_FROZEN:
>        +       case CPU_DOWN_FAILED:
>        +       case CPU_DOWN_FAILED_FROZEN:
>        +       case CPU_ONLINE:
>        +       case CPU_ONLINE_FROZEN:
>        +       case CPU_DEAD:
>        +       case CPU_DEAD_FROZEN:
>        +               common_cpu_mem_hotplug_unplug(1);
>
> it should just say
>
>        +       switch (phase) {
>        +       case CPU_ONLINE:
>        +       case CPU_ONLINE_FROZEN:
>        +       case CPU_DEAD:
>        +       case CPU_DEAD_FROZEN:
>        +               common_cpu_mem_hotplug_unplug(1);
>
> because it only makes sense to rebuild the scheduler domains when the
> thing SUCCEEDS.
>
> See? By having a sane design, the code is not just more robust and easy to
> follow, you can also simplify it and make it more logical.

Yes, I agree. And I did _not_ say that the current design is sane. My
impression about changes acceptable during a late release cycle was
utterly CRAPPY (indeed, it's always better to immediately fix a
problem the right way, not just add another patch and pray it doesn't
break somewhere else).


>
>                        Linus
>

-- 
Best regards,
Dmitry Adamushko