Re: [RFC PATCH v2 12/12] i386/sev: update query-sev QAPI format to handle SEV-SNP

[Michael Roth]

On Wed, Sep 01, 2021 at 04:14:10PM +0200, Markus Armbruster wrote:
> Michael Roth <michael.roth@amd.com> writes:
> 
> > Most of the current 'query-sev' command is relevant to both legacy
> > SEV/SEV-ES guests and SEV-SNP guests, with 2 exceptions:
> >
> >   - 'policy' is a 64-bit field for SEV-SNP, not 32-bit, and
> >     the meaning of the bit positions has changed
> >   - 'handle' is not relevant to SEV-SNP
> >
> > To address this, this patch adds a new 'sev-type' field that can be
> > used as a discriminator to select between SEV and SEV-SNP-specific
> > fields/formats without breaking compatibility for existing management
> > tools (so long as management tools that add support for launching
> > SEV-SNP guest update their handling of query-sev appropriately).
> 
> Technically a compatibility break: query-sev can now return an object
> that whose member @policy has different meaning, and also lacks @handle.
> 
> Matrix:
> 
>                             Old mgmt app    New mgmt app
>     Old QEMU, SEV/SEV-ES       good            good(1)
>     New QEMU, SEV/SEV-ES       good(2)         good
>     New QEMU, SEV-SNP           bad(3)         good
> 
> Notes:
> 
> (1) As long as the management application can cope with absent member
> @sev-type.
> 
> (2) As long as the management application ignores unknown member
> @sev-type.
> 
> (3) Management application may choke on missing member @handle, or
> worse, misinterpret member @policy.  Can only happen when something
> other than the management application created the SEV-SNP guest (or the
> user somehow made the management application create one even though it
> doesn't know how, say with CLI option passthrough, but that's always
> fragile, and I wouldn't worry about it here).
> 
> I think (1) and (2) are reasonable.  (3) is an issue for management
> applications that support attaching to existing guests.  Thoughts?

Hmm... yah I hadn't considering 'old mgmt' trying to interact with a SNP
guest started through some other means.

Don't really see an alternative other than introducing a new
'query-sev-snp', but that would still leave 'old mgmt' broken, since
it might still call do weird stuff like try to interpret the SNP policy
as an SEV/SEV-ES and end up with some very unexpected results. So if I
did go this route, I would need to have QMP begin returning an error if
query-sev is run against an SNP guest. But currently for non-SEV guests
it already does:

  error_setg(errp, "SEV feature is not available")

so 'old mgmt' should be able to handle the error just fine.

Would that approach be reasonable?