Re: [PATCH 1/2] irq_flags_t: intro and core annotations

[Peter Zijlstra <peterz@infradead.org>]

On Sun, 2007-10-28 at 00:14 +0400, Alexey Dobriyan wrote:
> On Sat, Oct 27, 2007 at 09:20:43PM +0200, Roman Zippel wrote:
> > On Sun, 21 Oct 2007, Alexey Dobriyan wrote:
> > 
> > > So far remedies were:
> > > a) grep(1) -- obviously fragile. I tried at some point grepping for
> > >    spin_lock_irqsave(), found quite a few, but it became booooring quickly.
> > > b) BUILD_BUG_ON(sizeof(flags) != sizeof(unsigned long)) -- was tried,
> > >    brutally broke some arches, survived one commit before revert :^)
> > >    Doesn't work on i386 where sizeof(unsigned int) == sizeof(unsigned long).
> > > 
> > > So it would be nice to have something more robust.
> > 
> > If it's just about the type checking, something like below should pretty 
> > much do the same.
> 
> It won't catch, the following if both variables are unsigned long:
> 
> 	spin_lock_irqsave(&lock, flags);
> 		[stuff]
> 	spin_unlock_irqrestore(&lock, foo->flags);
> 
> It won't catch "static unsigned long flags;". With sparse, we can
> eventually mark type as "on-stack only".

> > +static __always_inline void __irq_flags_check(unsigned long *flags)
> > +{
+	BUILD_BUG_ON(!__builtin_stack_addr(flags));
> > +}
> > +

obviously gcc doesn't (yet) support that __builtin function, but you
could make it work for sparse and define a dummy for gcc.