Re: 2.6.30: hibernation/swsusp lockup due to acpi-cpufreq

["Pallipadi, Venkatesh" <venkatesh.pallipadi@intel.com>]

On Tue, Jun 16, 2009 at 12:57:50PM -0700, Johannes Stezenbach wrote:
> On Tue, Jun 16, 2009 at 11:55:40AM -0700, Andrew Morton wrote:
> > On Tue, 16 Jun 2009 16:22:17 +0200
> > Johannes Stezenbach <js@sig21.net> wrote:
> > 
> > > Fix swsusp failure on !SMP
> > > 
> > > Commit 01599fca6758d2cd133e78f87426fc851c9ea725 introduced
> > > a regression which caused a backtrace on suspend and
> > > a hang on resume on a Thinkpad T42p (Pentium M CPU).
> > > 
> > > Signed-off-by: Johannes Stezenbach <js@sig21.net>
> > > 
> > > 
> > > --- linux-2.6.30/kernel/up.c.orig	2009-06-16 15:56:28.000000000 +0200
> > > +++ linux-2.6.30/kernel/up.c	2009-06-16 15:57:27.000000000 +0200
> > > @@ -10,11 +10,13 @@
> > >  int smp_call_function_single(int cpu, void (*func) (void *info), void *info,
> > >  				int wait)
> > >  {
> > > +	unsigned long flags;
> > > +
> > >  	WARN_ON(cpu != 0);
> > >  
> > > -	local_irq_disable();
> > > +	local_irq_save(flags);
> > >  	(func)(info);
> > > -	local_irq_enable();
> > > +	local_irq_restore(flags);
> > >  
> > >  	return 0;
> > >  }
> > 
> > ok, what's going on here?  The patch implies that someone (presumably
> > acpi-cpufreq) is calling smp_call_function_single() with local
> > interrupts disabled.  That's a bug on SMP kernels.  And it'll generate
> > a trace if it happens:
> > 
> > 	/* Can deadlock when called with interrupts disabled */
> > 	WARN_ON_ONCE(irqs_disabled() && !oops_in_progress);
> > 
> > but nobody has reported such a trace AFAIK?
> 
> This problem apparently only exists on !SMP kernels...
> 
> > Also, prior to 01599fca6758d2cd133e78f87426fc851c9ea725, acpi-cpufreq
> > was using work_on_cpu().  If it was calling work_on_cpu() with local
> > interrupts disabled then that would have been a bug too, which could
> > generate might_sleep() or scheduling-while-atomic warnings.
> 
> On !SMP, work_on_cpu() is just a function call:
> http://lxr.linux.no/linux+v2.6.30/include/linux/workqueue.h#L261
> 
> > Because it is a bug to call the SMP version of
> > smp_call_function_single() with local interrupts disabled, I don't
> > think we should need to apply the above patch.
> 
> and on SMP, smp_call_function_single() also uses
> local_irq_save/restore() iff  cpu == this_cpu:
> http://lxr.linux.no/linux+v2.6.30/kernel/smp.c#L272
> 
> > But I don't know what we _should_ do because I don't know what the bug
> > is.  Are you able to get us a copy of that stack trace?
> 
> Unfortunately my laptop doesn't have a serial port, and the
> stack trace is large and scrolls off the screen, I can only
> see the last part of it and I would need to find someone with
> a camera to take a picture...


Can you try the patch below (your changes + a warnon). That should give
the stack trace with successful suspend-resume.

acpi-cpufreq will not directly disable interrupt and call these routines.
So, it will be interesting to see how we are ending up in this state.

Thanks,
Venki


diff --git a/kernel/up.c b/kernel/up.c
index 1ff27a2..a4318ff 100644
--- a/kernel/up.c
+++ b/kernel/up.c
@@ -10,11 +10,15 @@
 int smp_call_function_single(int cpu, void (*func) (void *info), void *info,
 				int wait)
 {
+	unsigned long flags;
+
 	WARN_ON(cpu != 0);
 
-	local_irq_disable();
+	WARN_ON(irqs_disabled());
+
+	local_irq_save(flags);
 	(func)(info);
-	local_irq_enable();
+	local_irq_restore(flags);
 
 	return 0;
 }