Re: [PATCH v3 0/2] x86: enforce and cleanup RIP-relative accesses in early boot code

[Ard Biesheuvel]

On Sat, 3 Feb 2024 at 01:22, Kevin Loughlin <kevinloughlin@xxxxxxxxxx> wrote:
>
> On Wed, Jan 31, 2024 at 10:30 AM Borislav Petkov <bp@xxxxxxxxx> wrote:
> >
> > On Wed, Jan 31, 2024 at 10:16:55AM -0800, Jacob Xu wrote:
> > > We're interested in fixing SEV-SNP guest boots which are currently
> > > broken when using a guest kernel compiled with clang. It seems like
> > > every other user of SEV/SNP linux kernel code uses GCC to compile the
> > > kernel so they've avoided this issue.
> >
> > Lemme give that a try here.
> >
> > > E.g. Google COS uses clang to compile the kernel and we've made do
> > > with an internal fix for a while.
> >
> > Which means that, theoretically, you could forward-port this internal
> > fix until the issue is fixed for real, I'd say.
>
> True. I just think it would be better to have an upstream fix for
> clang builds of SEV-SNP guests; I believe the first such SEV-SNP code
> was merged in 5.19 if I'm not mistaken.

The problem is not only Clang. The problem is that we try to keep the
stable trees working with newer compilers in general, and we are
relying heavily on behavior on the part of the compiler that could
change in the future. Those references that GCC happens to emit as
RIP-relative today even without the workarounds could easily turn into
absolute references on tomorrow's version, given that both are
permitted by the code model under -fno-pic.

I've compared notes with Kevin internally, and we'll get a minimal,
simplified version of these changes into my v4 SEV PIC series so that
we can easily cherry-pick the fixes, either into linux-stable or into
our downstream fork.