[libvirt PATCH v2] qemu: substitute missing model name for host-passthrough

[Jiri Denemark]

On Wed, Sep 23, 2020 at 18:29:54 -0400, Collin Walling wrote:
> On 9/23/20 5:01 PM, Jiri Denemark wrote:
> > On Wed, Sep 23, 2020 at 16:17:04 -0400, Collin Walling wrote:
> >> On 9/23/20 2:52 PM, Collin Walling wrote:
> >> Actually, it might help to extend this functionality for baseline as
> >> well. If anything to at least catch the case when a CPU definition in
> >> the XML file is missing a <model> tag. Right now, virsh will either
> >> report "an unknown error occurred" when the XML file contains a _single_
> >> <cpu> element without a <model> tag, or it will report the same "Invalid
> >> parameter type for 'modela.name', expected: string" mentioned above when
> >> there are multiple definitions in the file, and at least one of them is
> >> missing a <model> tag.
> > 
> > Hmm, I would expect libvirt to complain about missing CPU model. If
> > that's not the case, we need to fix it. But we should not fix it the
> > same way. This change in the Compare API is hidden inside libvirt, we
> > just tell QEMU we're comparing "host" when the cpu->mode is
> > host-passthrough and get the result, which is basically yes/no.
> > 
> > But with baseline we get the result and make a guest CPU definition out
> > of it. By setting cpu->model to "host" we would could end up with
> > <model>host</model> in the result or even random result depending on the
> > host performing the baseline API. This API should just reject any CPU
> > definition without <model> as invalid argument.
> > 
> > Jirka
> > 
> 
> On s390x, QMP will internally convert the model "host" to a proper CPU
> model.
> 
> How about when baselining and if there is only a single CPU definition
> and the model is "host" (either provided verbatim by the file, or
> converted when the mode is host-passthrough), then perform a CPU model
> expansion on the single CPU definition? No baselining would technically
> be performed in this case.

Baseline is meant to be called on CPU definitions created by libvirt in
either domain capabilities (this is the best way) or capabilities.
Neither of them will contain mode='host-passthrough'. And since no
<model> means the host CPU could not be detected, baseline should just
return an error. I don't see why s390x should be any special. This is
what happens if you try this in x86:

    virsh hypervisor-cpu-baseline cpu-host.xml
    error: invalid argument: no CPU model specified at index 0

All we need to do is to make sure we report the right error message.

Overloading baseline to perform model expansion is not needed as what
you suggest is already provided by domain capabilities.

> I think this would support the behavior reported by the virsh man page:
> 
> """
> When FILE contains only a single CPU definition, the command will print
>      the same CPU with restrictions  imposed  by  the  capabilities  of
> the       hypervisor.

Host-passthrough CPU is by definition already restricted to what the
hypervisor is able to do so there's no need to ask baseline for the
result.

Jirka