Re: [PATCH, RFC] reimplement flush_workqueue()

[Srivatsa Vaddagiri <vatsa@in.ibm.com>]

On Mon, Dec 18, 2006 at 01:34:16AM +0300, Oleg Nesterov wrote:
>  void fastcall flush_workqueue(struct workqueue_struct *wq)
>  {
> -	might_sleep();
> -
> +	mutex_lock(&workqueue_mutex);
>  	if (is_single_threaded(wq)) {
>  		/* Always use first cpu's area. */
> -		flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, singlethread_cpu),
> -					-1);
> +		flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, singlethread_cpu));
>  	} else {
>  		int cpu;
> 
> -		mutex_lock(&workqueue_mutex);
>  		for_each_online_cpu(cpu)


Can compiler optimizations lead to cpu_online_map being cached in a register 
while running this loop? AFAICS cpu_online_map is not declared to be
volatile. If it can be cached, then we have the danger of invoking 
flush_cpu_workqueue() on a dead cpu (because flush_cpu_workqueue drops
workqueue_mutex, cpu hp events can change cpu_online_map while we are in
flush_cpu_workqueue).

> -			flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, cpu), cpu);
> -		mutex_unlock(&workqueue_mutex);
> +			flush_cpu_workqueue(per_cpu_ptr(wq->cpu_wq, cpu));


-- 
Regards,
vatsa