Re: [PATCH v2 1/3] compiler_types: Introduce the Clang __preserve_most function attribute

[Marco Elver]

On Mon, 7 Aug 2023 at 14:37, Florian Weimer <fweimer@xxxxxxxxxx> wrote:
>
> * Marco Elver:
>
> > Good idea. I had already created
> > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110899, and we need
> > better spec to proceed for GCC anyway.
>
> Thanks for the reference.
>
> >> Doesn't this change impact the kernel module ABI?
> >>
> >> I would really expect a check here
> >>
> >> > +#if __has_attribute(__preserve_most__)
> >> > +# define __preserve_most notrace __attribute__((__preserve_most__))
> >> > +#else
> >> > +# define __preserve_most
> >> > +#endif
> >>
> >> that this is not a compilation for a module.  Otherwise modules built
> >> with a compiler with __preserve_most__ attribute support are
> >> incompatible with kernels built with a compiler without that attribute.
> >
> > That's true, but is it a real problem? Isn't it known that trying to
> > make kernel modules built for a kernel with a different config (incl.
> > compiler) is not guaranteed to work? See IBT, CFI schemes, kernel
> > sanitizers, etc?
> >
> > If we were to start trying to introduce some kind of minimal kernel to
> > module ABI so that modules and kernels built with different toolchains
> > keep working together, we'd need a mechanism to guarantee this minimal
> > ABI or prohibit incompatible modules and kernels somehow. Is there a
> > precedence for this somewhere?
>
> I think the GCC vs Clang thing is expected to work today, isn't it?

I, personally, wouldn't bet on it. It very much depends on the kernel
config used.

> Using the Clang-based BPF tools with a GCC-compiled kernel requires a
> matching ABI.

BPF is a different story altogether, and falls more into the category
of user space to kernel ABI, which the kernel has strong guarantees
on.

> The other things you listed result in fairly obvious breakage, sometimes
> even module loading failures.  Unconditional crashes are possible as
> well.  With __preserve_most__, the issues are much more subtle and may
> only appear for some kernel/module compielr combinations and
> optimization settings.  The impact of incorrectly clobbered registers
> tends to be like that.

One way around this would be to make the availability of the attribute
a Kconfig variable. Then externally compiled kernel modules should do
the right thing, since they ought to use the right .config when being
built.

I can make that change for v3.