Re: [Qemu-devel] [Patch v1 01/29] qmp: details about CPU definitions in query-cpu-definitions

[David Hildenbrand <dahi@linux.vnet.ibm.com>]

> >  # @name: the name of the CPU definition
> >  #
> > +# @migration-safe: #optional whether a CPU definition can be safely used for
> > +#                  migration in combination with a QEMU compatibility machine
> > +#                  when migrating between different QMU versions and hosts.
> > +#                  If not provided, information is not available.  
> 
> I would be more explicit about migration between different hosts.
> I suggest "between different QEMU versions and between hosts with
> different sets of (hardware or software) capabilities".

Sounds good to me.

> 
> Maybe we should make the "if not provided" case clearer. Maybe
> "if not provided, information is not available and caller should
> not assume the CPU model is migration-safe". We know that
> existing libvirt x86 code assumes all CPU models (except "host")
> are migration-safe, but it's better to advise people to not try
> to make any assumptions in new code.

Also sounds good to me.

> 
> Later, we need to document somewhere that the "migratable"
> property in "host" does not mean "migration-safe" (at least in
> x86), because migration of "host" is safe only if the host
> (software and hardware) capabilities are exactly the same.
> 

Right, this will be a special case, also once we have that for s390x.

> For reference: in x86, all CPU models except "host" are
> migration-safe.

Yes, that was my conclusion and that's also why I turned all s390x models
(except host) into migration-safe models. Makes a lot of things easier to
handle.

> 
> > +#
> > +# @static: #optional whether a CPU definition is static and will not change
> > +#          between QEMU versions / QEMU machines. A static model is always
> > +#          migration-safe. If not provided, information is not available.  
> 
> I assume static models don't change depending on the
> machine-type, either. If that's case, we should document that.

That's what I meant with "QEMU machines", should that be "QEMU machine types"
instead?

> 
> I believe in this case we don't need to make it optional: just
> make the field always present and set it to "false" by default.

That is true for x86, do you know about the other architectures (arm, ppc)?
I'd like to avoid returning false information here for other architectures.

Thanks Eduardo!

David