Re: [PATCH v8 29/40] x86/compressed/64: add support for SEV-SNP CPUID table in #VC handlers

[Borislav Petkov]

On Tue, Jan 18, 2022 at 07:18:06PM -0600, Michael Roth wrote:
> If 'fake_count'/'reported_count' is greater than the actual number of
> entries in the table, 'actual_count', then all table entries up to
> 'fake_count' will also need to pass validation. Generally the table
> will be zero'd out initially, so those additional/bogus entries will
> be interpreted as a CPUID leaves where all fields are 0. Unfortunately,
> that's still considered a valid leaf, even if it's a duplicate of the
> *actual* 0x0 leaf present earlier in the table. The current code will
> handle this fine, since it scans the table in order, and uses the
> valid 0x0 leaf earlier in the table.

I guess it would be prudent to have some warnings when enumerating those
leafs and when the count index "goes off into the weeds", so to speak,
and starts reading 0-CPUID entries. I.e., "dear guest owner, your HV is
giving you a big lie: a weird/bogus CPUID leaf count..."

:-)

And lemme make sure I understand it: the ->count itself is not
measured/encrypted because you want to be flexible here and supply
different blobs with different CPUID leafs?

> This is isn't really a special case though, it falls under the general
> category of a hypervisor inserting garbage entries that happen to pass
> validation, but don't reflect values that a guest would normally see.
> This will be detectable as part of guest owner attestation, since the
> guest code is careful to guarantee that the values seen after boot,
> once the attestation stage is reached, will be identical to the values
> seen during boot, so if this sort of manipulation of CPUID values
> occurred, the guest owner will notice this during attestation, and can
> abort the boot at that point. The Documentation patch addresses this
> in more detail.

Yap, it is important this is properly explained there so that people can
pay attention to during attestation.

> If 'fake_count' is less than 'actual_count', then the PSP skips
> validation for anything >= 'fake_count', and leaves them in the table.
> That should also be fine though, since guest code should never exceed
> 'fake_count'/'reported_count', as that's a blatant violation of the
> spec, and it doesn't make any sense for a guest to do this. This will
> effectively 'hide' entries, but those resulting missing CPUID leaves
> will be noticeable to the guest owner once attestation phase is
> reached.

Noticeable because the guest owner did supply a CPUID table with X
entries but the HV is reporting Y?

If so, you can make this part of the attestation process: guest owners
should always check the CPUID entries count to be of a certain value.

> This does all highlight the need for some very thorough guidelines
> on how a guest owner should implement their attestation checks for
> cpuid, however. I think a section in the reference implementation
> notes/document that covers this would be a good starting point. I'll
> also check with the PSP team on tightening up some of these CPUID
> page checks to rule out some of these possibilities in the future.

Now you're starting to grow the right amount of paranoia - I'm glad I
was able to sensitize you properly!

:-)))

> Nevermind, that doesn't work since snp_cpuid_info_get_ptr() is also called
> by snp_cpuid_info_get_ptr() *prior* to initializing the table, so it ends
> seeing cpuid->count==0 and fails right away. So your initial suggestion
> of checking cpuid->count==0 at the call-sites to determine if the table
> is enabled is probably the best option.
> 
> Sorry for the noise/confusion.

No worries - the end result is important!

Thx.

-- 
Regards/Gruss,
    Boris.

https://people.kernel.org/tglx/notes-about-netiquette