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

[Peter Zijlstra <a.p.zijlstra@chello.nl>]

On Mon, 2011-05-02 at 19:55 +0900, KOSAKI Motohiro wrote:
> The rule is, we have to update tsk->rt.nr_cpus_allowed too if we change
> tsk->cpus_allowed. Otherwise RT scheduler may confuse.
> 
> This patch fixes it.
> 
> btw, system_state checking is very important. current boot sequence is (1) smp_init
> (ie secondary cpus up and created cpu bound kthreads). (2) sched_init_smp().
> Then following bad scenario can be happen,
> 
> (1) cpuup call notifier(CPU_UP_PREPARE)
> (2) A cpu notifier consumer create FIFO kthread
> (3) It call kthread_bind()
>    ... but, now secondary cpu haven't ONLINE

isn't

> (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?

> (4) find_lowest_rq() touch local_cpu_mask if task->rt.nr_cpus_allowed != 1,
>     but it haven't been initialized.
> 
> RCU folks plan to introduce such FIFO kthread and our testing hitted the
> above issue. Then this patch also protect it.

I'm fairly sure it doesn't, normal cpu-hotplug doesn't poke at
system_state.

> 
> Signed-off-by: KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com>
> Cc: Oleg Nesterov <oleg@redhat.com>
> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
> Cc: Ingo Molnar <mingo@elte.hu>
> ---
>  include/linux/cpuset.h |    1 +
>  kernel/cpuset.c        |    1 +
>  kernel/sched.c         |    4 ++++
>  3 files changed, 6 insertions(+), 0 deletions(-)
> 
> diff --git a/include/linux/cpuset.h b/include/linux/cpuset.h
> index f20eb8f..42dcbdc 100644
> --- a/include/linux/cpuset.h
> +++ b/include/linux/cpuset.h
> @@ -147,6 +147,7 @@ static inline void cpuset_cpus_allowed(struct task_struct *p,
>  static inline int cpuset_cpus_allowed_fallback(struct task_struct *p)
>  {
>  	cpumask_copy(&p->cpus_allowed, cpu_possible_mask);
> +	p->rt.nr_cpus_allowed = cpumask_weight(&p->cpus_allowed);
>  	return cpumask_any(cpu_active_mask);
>  }
>  
> diff --git a/kernel/cpuset.c b/kernel/cpuset.c
> index 1ceeb04..6e5bbe8 100644
> --- a/kernel/cpuset.c
> +++ b/kernel/cpuset.c
> @@ -2220,6 +2220,7 @@ int cpuset_cpus_allowed_fallback(struct task_struct *tsk)
>  		cpumask_copy(&tsk->cpus_allowed, cpu_possible_mask);
>  		cpu = cpumask_any(cpu_active_mask);
>  	}
> +	tsk->rt.nr_cpus_allowed = cpumask_weight(&tsk->cpus_allowed);
>  
>  	return cpu;
>  }

I don't really see the point of doing this separately from your second
patch, please fold them.

> diff --git a/kernel/sched.c b/kernel/sched.c
> index fd4625f..bfcd219 100644
> --- a/kernel/sched.c
> +++ b/kernel/sched.c
> @@ -2352,6 +2352,10 @@ static int select_fallback_rq(int cpu, struct task_struct *p)
>  	if (dest_cpu < nr_cpu_ids)
>  		return dest_cpu;
>  
> +	/* Don't worry. It's temporary mismatch. */
> +	if (system_state < SYSTEM_RUNNING)
> +		return cpu;
> +
>  	/* No more Mr. Nice Guy. */
>  	dest_cpu = cpuset_cpus_allowed_fallback(p);
>  	/*

Like explained, I don't believe this actually fixes your problem (its
also disgusting).