On Wed, 6 Jun 2012, Thomas Gleixner wrote:

> On Wed, 6 Jun 2012, Chen Gong wrote:
> > 于 2012/6/6 18:23, Thomas Gleixner 写道:
> > I think you mean
> > 
> > -               del_timer_sync(t);
> >                 smp_call_function_single(cpu, mce_disable_cpu, &action, 1);
> > +		del_timer_sync(t);
> >                 break;
> 
> No I meant it the way I wrote, but as you don't execute hotplug it's
> irrelevant.
> 
> So the obvious candidate is the mce-injection code, which was
> obviously never tested with DEBUG_PREEMPT enabled.
> 
> raise_local() can be called with preemption enabled from
> raise_mce(). Fix for that below.
> 
> Though I can't see how that would do anything with the timer.

I think I found it. Do you have CONFIG_NO_HZ enabled? Then mod_timer()
will try to move the timer to a different cpu, when the cpu which is
running that code is idle. Bloody obvious :(

I'll send out a combo patch with all changes so far later.

Thanks,
	
	tglx

Index: linux-2.6/arch/x86/kernel/cpu/mcheck/mce.c
===================================================================
--- linux-2.6.orig/arch/x86/kernel/cpu/mcheck/mce.c
+++ linux-2.6/arch/x86/kernel/cpu/mcheck/mce.c
@@ -1307,7 +1307,7 @@ void mce_timer_kick(unsigned long interv
 
 	if (timer_pending(t)) {
 		if (time_before(when, t->expires))
-			mod_timer(t, when);
+			mod_timer_pinned(t, when);
 	} else {
 		t->expires = round_jiffies(when);
 		add_timer_on(t, smp_processor_id());