Re: [RFC PATCH v2] x86/sev: enforce RIP-relative accesses in early SEV/SME code

[Ard Biesheuvel]

On Sun, 21 Jan 2024 at 16:38, Borislav Petkov <bp@xxxxxxxxx> wrote:
>
> On Sun, Jan 21, 2024 at 03:12:56PM +0100, Ard Biesheuvel wrote:
> > The SEV boot code is especially tricky here as very few
> > people can even test it,
>
> No worries about that - us, the Google cloud folks, AWS and a bunch of
> others are people I could think of who could help out. :-)
>

Yeah. I have been trying to find people internally at Google that can
help me set up some CI that I can throw kernel builds at and they will
be test booted in a SEV guest, but so far progress has been slow.

> > 1)
> > WARNING: modpost: vmlinux: section mismatch in reference:
> > startup_64_pi+0x33 (section: .pi.text) -> sme_enable (section:
> > .init.text)
>
> sme_enable() is in the 1:1 mapping TU
> arch/x86/mm/mem_encrypt_identity.c, see
>
>   1cd9c22fee3a ("x86/mm/encrypt: Move page table helpers into separate translation unit")
>
> so might as well move it to .pi.text
>

Ack.

> The rest below look like they'd need more serious untangling.
>
> Btw, I just had another idea: we could remove -mcmodel=kernel from the
> build flags of the whole kernel once -fPIC is enabled so that gcc can be
> forced to do rIP-relative addressing.
>
> I'm being told the reason it doesn't allow mcmodel=kernel with -fPIC is
> only a matter of removing that check and that it *should* otherwise work
> but someone needs to try that. And then there are older gccs which we
> cannot fix.
>

-fPIE -mcmodel=small should work fine afaik. The only problem i
encountered is that it changes the default per-CPU base register to FS
but that can be overridden on the command line.

The problem with building the entire kernel -fPIE is that it increases
code size: RIP-relative LEA instructions are 1 byte longer than
absolute 32-bit MOVs.