Re: [PATCH 05/10] efi/libstub: distinguish between native/mixed not 32/64 bit

[Arvind Sankar]

On Sat, Dec 14, 2019 at 07:54:25PM +0000, Ard Biesheuvel wrote:
> On Sat, 14 Dec 2019 at 20:49, Arvind Sankar <nivedita@xxxxxxxxxxxx> wrote:
> >
> > On Sat, Dec 14, 2019 at 02:46:27PM -0500, Arvind Sankar wrote:
> > > On Sat, Dec 14, 2019 at 06:57:30PM +0100, Ard Biesheuvel wrote:
> > > > +
> > > > +#define efi_table_attr(table, attr, instance) ({                   \
> > > > +   __typeof__(((table##_t *)0)->attr) __ret;                       \
> > > > +   if (efi_is_native()) {                                          \
> > > > +           __ret = ((table##_t *)instance)->attr;                  \
> > > > +   } else {                                                        \
> > > > +           __typeof__(((table##_32_t *)0)->attr) at;               \
> > > > +           at = (((table##_32_t *)(unsigned long)instance)->attr); \
> > > > +           __ret = (__typeof__(__ret))(unsigned long)at;           \
> > > > +   }                                                               \
> > > > +   __ret;                                                          \
> > > > +})
> > >
> > > The casting of `at' is appropriate if the attr is a pointer type which
> > > needs to be zero-extended to 64-bit, but for other fields it is
> > > unnecessary at best and possibly dangerous.  There are probably no
> > > instances currently where it is called for a non-pointer field, but is
> > > it possible to detect if the type is pointer and avoid the cast if not?
> >
> > To clarify, I mean the casting via `unsigned long' -- casting to type of
> > __ret should be ok. We could also use uintptr_t for cleanliness when the
> > cast is required?
> 
> Could you give an example of how it could break?

eg, if the field is actually a structure. Nobody seems to do this
currently, but say for
	efi_table_attr(efi_boot_services, hdr, instance)
you shouldn't cast hdr to unsigned long.

There's also the case of nested 32/64-bit structures that breaks, but
that might be too hard to try to handle:
	efi_table_attr(efi_pci_io_protocol, io, instance)
where io is a structure of two pointers which would need to be
individually casted. It's properly accessible as io.read/io.write
though.

In general though, everything is either a pointer or a u64, so if it's
too messy to detect pointer type, the cast is still probably safe in
practice.