Re: ring_buffer_attach && cond_synchronize_rcu (Was: percpu-rwsem: Optimize readers and reduce global impact)

[Peter Zijlstra <peterz@infradead.org>]

On Sat, May 30, 2015 at 10:04:25PM +0200, Oleg Nesterov wrote:
> And. I tried to find other users of get_state/cond_sync. Found
> ring_buffer_attach() and it looks obviously buggy?

Urgh, indeed.

> IOW. Suppose that ring_buffer_attach() preempts right_after
> get_state_synchronize_rcu() and gp completes before spin_lock().
> 
> In this case cond_synchronize_rcu() does nothing and we reuse
> ->rb_entry without waiting for gp in between?

Yes.

> Don't we need the patch below? (it also moves the ->rcu_pending check
> under "if (rb)", to make it more readable imo).

> --- x/kernel/events/core.c
> +++ x/kernel/events/core.c
> @@ -4310,20 +4310,20 @@ static void ring_buffer_attach(struct pe
>  		WARN_ON_ONCE(event->rcu_pending);
>  
>  		spin_lock_irqsave(&old_rb->event_lock, flags);
>  		list_del_rcu(&event->rb_entry);
>  		spin_unlock_irqrestore(&old_rb->event_lock, flags);
>  
> +		event->rcu_batches = get_state_synchronize_rcu();
> +		event->rcu_pending = 1;
>  	}
>  
>  	if (rb) {
> +		if (event->rcu_pending) {
> +			cond_synchronize_rcu(event->rcu_batches);
> +			event->rcu_pending = 0;
> +		}
> +
>  		spin_lock_irqsave(&rb->event_lock, flags);
>  		list_add_rcu(&event->rb_entry, &rb->event_list);
>  		spin_unlock_irqrestore(&rb->event_lock, flags);

Agreed.