Re: [BUG] 2.6.25-rc3-mm1 kernel panic while bootup on powerpc ()

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

On Thu, 06 Mar 2008 11:03:31 +1100
Benjamin Herrenschmidt <benh@kernel.crashing.org> wrote:

> 
> > Yes, we are - it's the semaphore rewrite which is doing this in
> > start_kernel().  It's being discussed.
> > 
> > Enabling interrupts too early on powerpc was discovered to be fatal on
> > powerpc years ago.  It looks like that remains the case.
> 
> Regarding these issues. I could make it non fatal and just WARN_ON,
> provided that I have a way to differentiate legal vs. illegal calls
> to local_irq_enable().

And local_irq_restore() and various other things.

> We already have that function mostly out of
> line in C code due to our lazy irq disabling scheme, so the overhead of
> testing some global kernel state would be minimum here.
> 
> However, I don't see anything around init/main.c:start_kernel() that I
> can use. What do you reckon here we should do ? Add some kind of global
> we set before calling local_irq_enable() ? Or make early_boot_irqs_on()
> do that generically 
> 
> It's currently defined as an empty inline without CONFIG_TRACE_IRQFLAGS
> but we could make it set a flag instead.
> 
> I'm pretty sure other archs have similar problems, especially in the
> embedded world where you are booted with random junk firmwares that may
> leave devices, interrupt controllers etc... in random state, and
> enabling incoming IRQs before the arch code properly initializes the
> main interrupt controller can be fatal. I know at least of an ARM board
> I worked on a while ago that had a similar issues.
> 
> On ppc32, unfortunately, our local_irq_enable/restore are nice inlines
> that whack the appropriate MSR bits directly, thus adding a test for a
> global flag would add some bloat/overhead that I'd like to avoid, at
> least until we decide to also do lazy disabling on those, if ever...

I'd have thought that the way to do this would be to add it to lockdep -
lockdep already has all the infrastructure and code sites to do this.

Set some special flag saying its-ok-to-enable-interrupts-now and test that
in lockdep.

akpm:/usr/src/25> grep LOCKDEP arch/powerpc/Kconfig 
akpm:/usr/src/25> 

losers ;)

Still, doing it for

akpm:/usr/src/25> grep -l LOCKDEP arch/*/Kconfig 
arch/arm/Kconfig
arch/avr32/Kconfig
arch/mips/Kconfig
arch/s390/Kconfig
arch/sh/Kconfig
arch/sparc64/Kconfig
arch/um/Kconfig
arch/x86/Kconfig

should give pretty good coverage.