RE: [PATCH] crypto/caam: Avoid GCC constprop bug warning

[David Laight <David.Laight@ACULAB.COM>]

From: Kees Cook
> Sent: 02 December 2022 18:58
> 
> On Fri, Dec 02, 2022 at 10:01:50AM +0000, David Laight wrote:
> > From: Anders Roxell
> > > Sent: 02 December 2022 00:58
> > >
> > > On 2022-10-28 14:05, Kees Cook wrote:
> > > > GCC 12 appears to perform constant propagation incompletely(?) and can
> > > > no longer notice that "len" is always 0 when "data" is NULL. Expand the
> > > > check to avoid warnings about memcpy() having a NULL argument:
> > > >
> > > >    ...
> > > >                     from drivers/crypto/caam/key_gen.c:8:
> > > >    drivers/crypto/caam/desc_constr.h: In function 'append_data.constprop':
> > > >    include/linux/fortify-string.h:48:33: warning: argument 2 null where non-null expected [-
> > > Wnonnull]
> > > >       48 | #define __underlying_memcpy     __builtin_memcpy
> > > >          |                                 ^
> > > >    include/linux/fortify-string.h:438:9: note: in expansion of macro '__underlying_memcpy'
> > > >      438 |         __underlying_##op(p, q, __fortify_size);                        \
> > > >          |         ^~~~~~~~~~~~~
> > ...
> >
> > Is this really a bug in the fortify-string wrappers?
> > IIRC the call is memcpy(NULL, ptr, 0) (or maybe memcpy(ptr, NULL, 0).
> > In either case call can be removed at compile time.
> >
> > I'd bet that the constant propagation of 'len' fails because
> > of all the intermediate variables that get used in order to
> > avoid multiple evaluation.
> >
> > The some 'tricks' that are used in min() (see minmax.h) to
> > generate a constant output for constant input could be
> > use to detect a compile-time zero length.
> >
> > Something like:
> > #define memcpy(dst, src, len) \
> > 	(__is_constzero(len) ? (dst) : memcpy_check(dst, src, len))
> >
> > With:
> > #define __is_constzero(x) sizeof(*(1 ? (void *)(x) : (int *)0) != 1)
> > Which could go into const.h and used in the definition of __is_constexpr().
> 
> While it could be possible to strip the nonnull attribute, I think it's
> not an unreasonable check to have. This is literally the only case in
> the entire kernel that is tripped, for example.

It is probably the only place that calls memcpy() with compile-time
NULL and zero length.
IIRC the memcpy() call comes from a #define expansion where some
expansions don't need anything copied.
A simple 'builtin_constant' check and then one for zero in the
#define itself would probably suffice - and avoid the call
being compiled in at all.

	David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)