Re: [PATCH v2] x86/traps: Enable UBSAN traps on x86

[Kees Cook <kees@kernel.org>]

On Tue, Jun 18, 2024 at 12:13:27AM +0200, Thomas Gleixner wrote:
> On Wed, Jun 12 2024 at 11:42, Kees Cook wrote:
> > On Tue, Jun 11, 2024 at 01:26:09PM -0700, Gatlin Newhouse wrote:
> >> It seems that is_valid_bugaddr() needs to be implemented on all architectures
> >> and the function get_ud_type() replaces it here. So how should the patch handle
> >> is_valid_bugaddr()? Should the function remain as-is in traps.c despite no
> >> longer being used?
> >
> > Yeah, this is why I'd suggested to Gatlin in early designs to reuse
> > is_valid_bugaddr()'s int value. It's a required function, so it seemed
> > sensible to just repurpose it from yes/no to no/type1/type2/type3/etc.
> 
> It's not sensible, it's just tasteless.
> 
> If is_valid_bugaddr() is globaly required in it's boolean form then it
> should just stay that way and not be abused just because it can be
> abused.
> 
> What's wrong with doing:
> 
> __always_inline u16 get_ud_type(unsigned long addr)
> {
>         ....
> }
> 
> int is_valid_bugaddr(unsigned long addr)
> {
> 	return get_ud_type() != BUG_UD_NONE;
> }
> 
> Hmm?
> 
> In fact is_valid_bugaddr() should be globally fixed up to return bool to
> match what the function name suggests.
> 
> The UD type information is x86 specific and has zero business in a
> generic architecture agnostic function return value.
> 
> It's a sad state of affairs that I have to explain this to people who
> care about code correctness. Readability and consistency are substantial
> parts of correctness, really.

Well, it's trade-offs. If get_ud_type() is in is_valid_bugaddr(), we
have to call it _again_ outside of is_valid_bugaddr(). That's suboptimal
as well. I was trying to find a reasonable way to avoid refactoring all
architectures and to avoid code code.

Looking at it all again, I actually think arch/x86/kernel/traps.c
shouldn't call is_valid_bugaddr() at all. That usage can continue to
stay in lib/bug.c, which is only ever used by x86 during very early
boot, according to the comments in early_fixup_exception(). So just a
direct replacement of is_valid_bugaddr() with the proposed get_ud_type()
should be fine in arch/x86/kernel/traps.c.

What do you think?

-- 
Kees Cook