Re: [tip:x86/seves] BUILD SUCCESS WITH WARNING e6eb15c9ba3165698488ae5c34920eea20eaa38e

[Marco Elver]

On Wed, 16 Sep 2020 at 00:34, Nick Desaulniers <ndesaulniers@xxxxxxxxxx> wrote:
> On Tue, Sep 15, 2020 at 2:02 PM Josh Poimboeuf <jpoimboe@xxxxxxxxxx> wrote:
> >
> > panic() is noreturn, so the compiler is enforcing the fact that it
> > doesn't return, by trapping if it does return.
> >
> > I seem to remember that's caused by CONFIG_UBSAN_TRAP.
>
> Indeed, if I remove CONFIG_UBSAN_TRAP from the 0day report's
> randconfig, these unreachable instruction warnings all go away.
>
> So what's the right way to fix this?
>
> CONFIG_UBSAN_TRAP enables -fsanitize-undefined-trap-on-error  (not
> sure why that's wrapped in cc-option; it shouldn't be selectable via
> Kconfig if unsupported by the toolchain).
>
> Should clang not be emitting `ud2` trapping instructions for this flag
> for no-return functions?

I think this would defeat the purpose of this UBSAN feature. Certain
UBSAN checks are done fully statically, like is done by
fsanitize=unreachable, and could actually be enabled in production
kernels; trapping the kernel in these cases would be a reasonable way
to avoid further damage to the system.

(You could in theory force it to not emit a trap by using
fno-sanitize-trap=unreachable, but I think it's a bad idea.)

> or
>
> Should objtool be made aware of the config option and then not check
> traps after no-returns?

I'd vote for this. And it seems Ilie implemented this already.

> I suspect the latter, but I'm not sure how feasible it is to
> implement.  Josh, Marco, do you have thoughts on the above?

Thanks,
-- Marco