Re: 2.6.12-rc2-mm3

[Oleg Nesterov <oleg@tv-sign.ru>]

Benjamin Herrenschmidt wrote:
>
> I'll have a look at the timer patch next week, they might have some
> subtle race caused by a lack of memory barrier. I've had to debug some
> of those in early timer code, and those are really nasty, they usually
> only trigger under some subtle conditions, like ... heavy networking.

Before this timer patch del_timer(pending_timer) worked as
a memory barrier for the caller, now it does not.

For example, sk_stop_timer() does:

	if (del_timer(timer))
		// no more wmb() here
		atomic_dec(&sk->sk_refcnt);

I think that this particular case is ok, but may be there is
some user of timers which lacks the memory barrier?

Juergen Kreileder wrote:
>
> It only happens when running Azareus with IBM's Java (our's isn't ready yet).
> So far I was able to reproduce the problem on all -mm versions within
> one hour.  Otherwise the kernels seem to work fine -- no lockup unless
> I run Azareus.

By any chance, could you please try this patch?
It restores "deleting timer acts as a barrier" behaviour.

--- 2.6.12-rc2+timer_patches/kernel/timer.c~	Sun Apr 24 11:59:31 2005
+++ 2.6.12-rc2+timer_patches/kernel/timer.c	Sun Apr 24 13:35:01 2005
@@ -341,6 +341,9 @@ int del_timer(struct timer_list *timer)
 		spin_unlock_irqrestore(&base->lock, flags);
 	}
 
+	if (ret)
+		smp_wmb();
+
 	return ret;
 }
 
@@ -387,6 +390,10 @@ unlock:
 		spin_unlock_irqrestore(&base->lock, flags);
 	} while (ret < 0);
 
+	if (ret)
+		smp_wmb();
+	smp_rmb();
+
 	return ret;
 }
 
@@ -457,6 +464,7 @@ repeat:
 
 			set_running_timer(base, timer);
 			detach_timer(timer, 1);
+			smp_wmb();
 			spin_unlock_irq(&base->t_base.lock);
 			{
 				u32 preempt_count = preempt_count();
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/