Re: [PATCH 1/4] irq: correct CPUMASKS_OFFSTACK typo -v2

[Andrew Morton <akpm@linux-foundation.org>]

On Wed, 15 Apr 2009 12:01:05 +0200
Ingo Molnar <mingo@elte.hu> wrote:

> 
> * Andrew Morton <akpm@linux-foundation.org> wrote:
> 
> > On Tue, 14 Apr 2009 13:41:55 -0700
> > Yinghai Lu <yinghai@kernel.org> wrote:
> > 
> > > irq
> > 
> > Speaking of which, could someone please take a look at
> > http://www.gossamer-threads.com/lists/linux/kernel/1060580?do=post_view_threaded#1060580
> > ?
> [...]
> 
> > But try_one_irq() is running tifm_7xx1_isr() with local interrupts 
> > enabled, which upsets lockdep.
> 
> It doesnt just upset lockdep, it could also cause real lockups. 

Only if the interrupt were to magically come back to life, then
trigger.  I think.  Possibly in the case of shared interrupts we'd lock
up because of an interrupt from another device.

> Lockdep is just the canary, the lockup is the methane explosion.
> 
> > But I suspect that the code as it stands is non-buggy. Unless the 
> > interrupt can magically come back to life. In which case any 
> > change we make is purely a make-lockdep-shut-up thing.
> 
> Hm, i'd suggest we go for the methane leak instead of squashing the 
> canary. Which in this case would be try_one_irq() ignoring 
> IRQF_DISABLED or so? Affecting (much) more ISRs than just 
> tifm_7xx1_isr()?

tifm_7xx1_isr() is requested with bare IRQF_SHARED, so that function is
supposed to be called with local interrupts enabled.

And indeed, try_one_irq() is calling it with local interrupts enabled. 
That gets lockdep upset.  I assume there's magic in lockdep somewhere
which recognises the case where a function is called in hard irq
context with local interrupts enabled and, knowing that the controller
won't generate another interrupt, treats this as
local-interrupt-disabled.  Or something.

I don't know how to fix this really.  We _could_ fudge it by disabling
local interrupts in try_one_irq().  But the ISR could legitimately do
(say) spin_unlock_irq() and muck everything up.