Re: [RFC][PATCH 16/17] sched: Move the second half of ttwu() to the remote cpu

[Oleg Nesterov <oleg@redhat.com>]

On 12/24, Peter Zijlstra wrote:
>
> +static void
> +ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
> +{
> +#ifdef CONFIG_SMP
> +	if (task_cpu(p) != cpu_of(rq))
> +		set_task_cpu(p, cpu_of(rq));
> +#endif

This looks a bit suspicious.

If this is called by sched_ttwu_pending() we are holding rq->lock,
not task_rq_lock(). It seems, we can race with, say, migration
thread running on task_cpu().



OK, p->state = TASK_WAKING protects us against, say, set_cpus_allowed_ptr()
which does task_rq_lock(p) and thus checks task_is_waking().

But, at the same time,

> +#ifdef CONFIG_SMP
> +static void ttwu_queue_remote(struct task_struct *p, int cpu)
> +{
> +	struct task_struct *next = NULL;
> +	struct rq *rq = cpu_rq(cpu);
> +
> +	for (;;) {
> +		struct task_struct *old = next;
> +
> +		p->wake_entry = next;
> +		next = cmpxchg(&rq->wake_list, old, p);
> +		if (next == old)
> +			break;
> +	}
> +
> +	if (!next)
> +		smp_send_reschedule(cpu);

what if that cpu does set_cpus_allowed_ptr(p) ?

It spins with irq disabled. Once the caller, try_to_wake_up(),
drops ->pi_lock it will wait for !task_is_waking() forever, no?

Oleg.