Re: [patch] sched: fix set_task_cpu() and provide an unlocked runqueue variant

[Peter Zijlstra <peterz@infradead.org>]

On Sun, 2009-11-22 at 13:09 +0100, Mike Galbraith wrote:
> sched: fix set_task_cpu() and provide an unlocked runqueue variant.
> 
> set_task_cpu() falsifies migration stats by unconditionally generating migration
> stats whether a task's cpu actually changed or not.  As used in copy_process(),
> the runqueue is unlocked, so we need to provide an unlocked variant which does
> the locking to provide a write barrier.
> 
> Signed-off-by: Mike Galbraith <efault@gmx.de>
> Cc: Ingo Molnar <mingo@elte.hu>
> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
> LKML-Reference: <new-submission>
> 
> ---

> +void set_task_cpu_unlocked(struct task_struct *p, unsigned int new_cpu)
> +{
> +	unsigned long flags;
> +	struct rq *rq, *new_rq = cpu_rq(new_cpu);
> +
> +	smp_wmb();
> +	rq = task_rq_lock(p, &flags);
> +	update_rq_clock(rq);
> +	if (rq != new_rq)
> +		update_rq_clock(new_rq);
> +	set_task_cpu(p, new_cpu);
> +	task_rq_unlock(rq, &flags);
> +}

I've got to ask, what's that barrier for?