Re: [PATCH] sched/fair: Fix dereference NULL sched domain during select_idle_sibling

[Peter Zijlstra <peterz@infradead.org>]

On Sat, Oct 08, 2016 at 06:24:38PM +0800, Wanpeng Li wrote:

> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> index 543b2f2..03a6620 100644
> --- a/kernel/sched/fair.c
> +++ b/kernel/sched/fair.c
> @@ -5472,19 +5472,29 @@ static inline int select_idle_smt(struct task_struct *p, struct sched_domain *sd
>   */
>  static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int target)
>  {
> -	struct sched_domain *this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));

So select_idle_cpu() <- select_idle_sibling() is called from two places,
both which already hold rcu_read_lock() afaict.

This would've insta-triggered a rcu-lockdep splat otherwise I think.

That is, selsect_task_rq_fair() has rcu_read_lock() taken when calling
this, and task_numa_compare() does too.

> +	struct sched_domain *this_sd;
>  	u64 avg_idle = this_rq()->avg_idle;
> -	u64 avg_cost = this_sd->avg_scan_cost;
> +	u64 avg_cost;
>  	u64 time, cost;
>  	s64 delta;
>  	int cpu, wrap;
>  
> +	rcu_read_lock();
> +	this_sd = rcu_dereference(*this_cpu_ptr(&sd_llc));
> +	if (!this_sd) {
> +		cpu = -1;
> +		goto unlock;
> +	}

Yes, this is the part that was missing. We need to test this_sd after
the lookup.

Thanks for looking at this!