On 22/09/2016 16:35, Borislav Petkov wrote:
>> > @@ -230,6 +230,10 @@ int __init efi_setup_page_tables(unsigned long 
>> > pa_memmap, unsigned num_pages)
>> >    efi_scratch.efi_pgt = (pgd_t *)__sme_pa(efi_pgd);
>> >    pgd = efi_pgd;
>> >  
>> > +  flags = _PAGE_NX | _PAGE_RW;
>> > +  if (sev_active)
>> > +          flags |= _PAGE_ENC;
> So this is confusing me. There's this patch which says EFI data is
> accessed in the clear:
> https://lkml.kernel.org/r/20160822223738.29880.6909.st...@tlendack-t1.amdoffice.net
> but now here it is encrypted when SEV is enabled.
> Do you mean, it is encrypted here because we're in the guest kernel?

I suspect this patch is untested, and also wrong. :)

The main difference between the SME and SEV encryption, from the point
of view of the kernel, is that real-mode always writes unencrypted in
SME and always writes encrypted in SEV.  But UEFI can run in 64-bit mode
and learn about the C bit, so EFI boot data should be unprotected in SEV

Because the firmware volume is written to high memory in encrypted form,
and because the PEI phase runs in 32-bit mode, the firmware code will be
encrypted; on the other hand, data that is placed in low memory for the
kernel can be unencrypted, thus limiting differences between SME and SEV.

        Important: I don't know what you guys are doing for SEV and
        Windows guests, but if you are doing something I would really
        appreciate doing things in the open.  If Linux and Windows end
        up doing different things with EFI boot data, ACPI tables, etc.
        it will be a huge pain.  On the other hand, if we can enjoy
        being first, that's great.

In fact, I have suggested in the QEMU list that SEV guests should always
use UEFI; because BIOS runs in real-mode or 32-bit non-paging protected
mode, BIOS must always write encrypted data, which becomes painful in
the kernel.

And regarding the above "important" point, all I know is that Microsoft
for sure will be happy to restrict SEV to UEFI guests. :)

There are still some differences, mostly around the real mode trampoline
executed by the kernel, but they should be much smaller.

devel mailing list

Reply via email to