Re: [PATCH 1/2] cpuset: fix cpuset_cpus_allowed_fallback() don't update tsk->rt.nr_cpus_allowed

[Yong Zhang <yong.zhang0@gmail.com>]

On Fri, May 13, 2011 at 3:33 PM, KOSAKI Motohiro
<kosaki.motohiro@jp.fujitsu.com> wrote:
> Hello,
>
>>>>> (3) schedule() makes fallback and cpuset_cpus_allowed_fallback
>>>>>     change task->cpus_allowed
>>>>
>>>> I'm failing to see how this is happening, surely that kthread isn't
>>>> actually running that early?
>>>
>>> If my understand correctly, current call graph is below.
>>>
>>> kernel_init()
>>>        smp_init();
>>>                cpu_up()
>>>                        ... cpu hotplug notification
>>>                                kthread_create()
>>>        sched_init_smp();
>>>
>>>
>>> So, cpu hotplug event is happen before sched_init_smp(). The old rule is,
>>> all kthreads of using cpu-up notification have to use kthread_bind(). It
>>> protected from sched load balance.
>>>
>>> but, now cpuset_cpus_allowed_fallback() forcedly change kthread's
>>> cpumask.
>>> Why is this works? the point are two.
>>>
>>> - rcu_cpu_kthread_should_stop() call set_cpus_allowed_ptr() again
>>> periodically.
>>>  then, it can reset cpumask if cpuset_cpus_allowed_fallback() change it.
>>>  my debug print obseve following cpumask change occur at boot time.
>>>     1) kthread_bind: bind cpu1
>>>     2) cpuset_cpus_allowed_fallback: bind possible cpu
>>>     3) rcu_cpu_kthread_should_stop: rebind cpu1
>>> - while tsk->rt.nr_cpus_allowed == 1, sched load balancer never be crash.
>>
>> Seems rcu_spawn_one_cpu_kthread() call wake_up_process() directly,
>> which is under hotplug event CPU_UP_PREPARE. Maybe it should be
>> under CPU_ONLINE.
>
> Hmm..
>
> I haven't catch your point. cpu_up() call both CPU_UP_PREPARE and CPU_ONLINE
> notification. Thus, CPU_ONLINE still be called before sched_init_smp().

The difference is that CPU_ONLINE will set cpu_online_mask and cpu_active_mask,
so wake_up_process()->ttwu()->select_task_rq() will not go back to
cpuset_cpus_allowed_fallback().

But the comments for rcu_cpu_kthread_should_stop() shot me, since
rcu_spawn_one_cpu_kthread() must be called under CPU_UP_PREPARE
for some reason.

Maybe we should find a more suitable way for your issue :)

Thanks,
Yong

>
> Am I missing something?
>
>
>



-- 
Only stand for myself