Re: Debugging Thinkpad T430s occasional suspend failure.

[Thomas Gleixner <tglx@linutronix.de>]

CC'ing linux-arch, so the arch folks can scream murder if they have
any objections.

On Tue, 19 Feb 2013, Frederic Weisbecker wrote:
> 2013/2/18 Thomas Gleixner <tglx@linutronix.de>:
> > On Mon, 18 Feb 2013, Frederic Weisbecker wrote:
> >> diff --git a/kernel/softirq.c b/kernel/softirq.c
> >> index ed567ba..69fbefd 100644
> >> --- a/kernel/softirq.c
> >> +++ b/kernel/softirq.c
> >> @@ -277,6 +277,17 @@ restart:
> >>       tsk_restore_flags(current, old_flags, PF_MEMALLOC);
> >>  }
> >>
> >> +#ifndef __ARCH_IRQ_EXIT_IRQS_DISABLED
> >
> > We really should get rid of that and make irqs disabled on irq exit
> > mandatory.
> 
> Until now I even thought that having irqs disabled in irq_exit() was
> mandatory. I need to have a second look at tick_nohz_irq_exit() as I
> modified it with this requirement in mind...

Duh.
 
> Also with this patch, if a nesting interrupt happens right after
> tick_nohz_irq_exit(), this new irq will call tick_nohz_stop_idle() but
> not tick_nohz_start_idle() on the end. Leaving that unbalanced. I
> could work that around but I don't want to spaghettize further around
> that ifdef.
> 
> So this patch actually does depend on __ARCH_IRQ_EXIT_IRQS_DISABLED. I
> need to convert all archs to force this first. Then think about the
> softirq.c cleanup once its done.
> 
> I wonder what's the right path to this, may be loop over archs until
> they all have __ARCH_IRQ_EXIT_IRQS_DISABLED, then remove that? Sounds
> like non-sense but the safest way at the same time.

Yeah.
 
> >>  static inline void invoke_softirq(void)
> >>  {
> >> -     if (!force_irqthreads) {
> >> +     /* Can we run softirq's at all? We migth be nesting interrupts */
> >> +     if (in_nesting_interrupt())
> >> +             return;
> >
> > We might also be in a softirq disabled region where it makes no sense
> > to continue. So this should be
> >
> >    in_nesting_interrupt() || in_softirq()
> 
> in_nesting_interrupt() takes care of that as well.
 
Gah, misread the macro maze :(

Thanks,

	tglx