Re: [PATCH 1/4] locking/atomic/x86: Silence intentional wrapping addition

[Kees Cook <keescook@chromium.org>]

On Thu, Apr 25, 2024 at 01:05:00AM +0200, Peter Zijlstra wrote:
> On Thu, Apr 25, 2024 at 12:54:36AM +0200, Peter Zijlstra wrote:
> > On Wed, Apr 24, 2024 at 03:45:07PM -0700, Kees Cook wrote:
> > > On Thu, Apr 25, 2024 at 12:41:41AM +0200, Peter Zijlstra wrote:
> > > > On Wed, Apr 24, 2024 at 12:17:34PM -0700, Kees Cook wrote:
> > > > 
> > > > > @@ -82,7 +83,7 @@ static __always_inline bool arch_atomic_add_negative(int i, atomic_t *v)
> > > > >  
> > > > >  static __always_inline int arch_atomic_add_return(int i, atomic_t *v)
> > > > >  {
> > > > > -	return i + xadd(&v->counter, i);
> > > > > +	return wrapping_add(int, i, xadd(&v->counter, i));
> > > > >  }
> > > > >  #define arch_atomic_add_return arch_atomic_add_return
> > > > 
> > > > this is going to get old *real* quick :-/
> > > > 
> > > > This must be the ugliest possible way to annotate all this, and then
> > > > litter the kernel with all this... urgh.
> > > 
> > > I'm expecting to have explicit wrapping type annotations soon[1], but for
> > > the atomics, it's kind of a wash on how intrusive the annotations get. I
> > > had originally wanted to mark the function (as I did in other cases)
> > > rather than using the helper, but Mark preferred it this way. I'm happy
> > > to do whatever! :)
> > > 
> > > -Kees
> > > 
> > > [1] https://github.com/llvm/llvm-project/pull/86618
> > 
> > This is arse-about-face. Signed stuff wraps per -fno-strict-overflow.
> > We've been writing code for years under that assumption.
> > 
> > You want to mark the non-wrapping case.
> 
> That is, anything that actively warns about signed overflow when build
> with -fno-strict-overflow is a bug. If you want this warning you have to
> explicitly mark things.

This is confusing UB with "overflow detection". We're doing the latter.

> Signed overflow is not UB, is not a bug.
> 
> Now, it might be unexpected in some places, but fundamentally we run on
> 2s complement and expect 2s complement. If you want more, mark it so.

Regular C never provided us with enough choice in types to be able to
select the overflow resolution strategy. :( So we're stuck mixing
expectations into our types. (One early defense you were involved in
touched on this too: refcount_t uses a saturating overflow strategy, as
that works best for how it gets used.)

Regardless, yes, someone intent on wrapping gets their expected 2s
complement results, but in the cases were a few values started collecting
in some dark corner of protocol handling, having a calculation wrap around
is at best a behavioral bug and at worst a total system compromise.
Wrapping is the uncommon case here, so we mark those.

-- 
Kees Cook