Re: [Qemu-devel] [PATCH 4/6] i386: Infrastructure for versioned CPU models

[Eduardo Habkost <ehabkost@redhat.com>]

On Tue, Jun 25, 2019 at 05:26:03PM +0100, Daniel P. Berrangé wrote:
> On Tue, Jun 25, 2019 at 02:00:06AM -0300, Eduardo Habkost wrote:
> > Base code for versioned CPU models.  This will register a "-4.1"
> > version of all existing CPU models, and make the unversioned CPU
> > models be an alias for the -4.1 versions on the pc-*-4.1 machine
> > types.
> > 
> > On older machine types, the unversioned CPU models will keep the
> > old behavior.  This way, management software can use old machine
> > types while resolving aliases if compatibility with older QEMU
> > versions is required.
> > 
> > Using "-machine none", the unversioned CPU models will be aliases
> > to the latest CPU model version.
> > 
> > Includes a test case to ensure that:
> > old machine types won't report any alias to versioned CPU models;
> > "pc-*-4.1" will return aliases to -4.1 CPU models;
> > and "-machine none" will report aliases to some versioned CPU model.
> 
> I'm wondering about the of tieing CPU versions to the release version
> number and whether its a good idea or not ?
> 
> Could there be a reason for us to introduce 2 or more variants
> of a CPU in the same release & would that be a problem if we needed
> todo it ?

I don't see a problem, we can use 3-digit versions that won't be
enabled by any machine type by default.

> 
> Consider if we did not have a Broadwell CPU model defined yet
> and we were adding it at the same time as Spectre came out. We
> might have needed to add "Broadwell-NN" and "Broadwell-MM" one
> with "spec-ctrl" and one without, in order to ensure runability
> on hosts with & without the microcode upgrade.  "Broadwell" alias
> would resolve to either the NN or MM variant according to what
> the current host supported.
> 
> One way to cope with that would have been to add a 3rd digit
> after the version number. eg a Broadwell-4.1.1 and Broadwell-4.1.2

That's exactly what I did for Cascadelake-Server, see patch 6/6.

> 
> An alternative could consider using a plain counter for the CPU
> versions eg Broadwell-1, Broadwell-2, etc.... ?

This is possible too.  It would require a more complex mapping
between machine types and CPU model versions, though.  Maybe this
is worth the extra complexity because it would make the external
interfaces simpler.

> 
> 
> If we want to backport the newly added CPU model variants to
> exist branches, plain counters don't have the unsightly mismatch.
> eg we'd backport Broadwell-3 to QEMU 3.1, not Broadwell-4.1 to
> QEMU 3.1.  This isn't a functional problem, just something looking
> a bit odd.

I think I'm liking this approach.  If we're untying CPU models
from machine types, let's do it all the way.

-- 
Eduardo