[libvirt] [RFC] cpu_map: Remove pconfig from Icelake-Server CPU model

[Paolo Bonzini]

On 30/09/19 18:16, Jiri Denemark wrote:
> On Mon, Sep 30, 2019 at 17:16:27 +0200, Paolo Bonzini wrote:
>> On 30/09/19 16:31, Hu, Robert wrote:
>>>> This might be a problem if there are plans to eventually make KVM support
>>>> pconfig, though.  Paolo, Robert, are there plans to support pconfig in KVM in the
>>>> future?
>>> [Robert Hoo] 
>>> Thanks Eduardo for efforts in resolving this issue, introduced from my Icelake CPU
>>> model patch.
>>> I've no idea about PCONFIG's detail and plan. Let me sync with Huang, Kai and answer
>>> you soon.
>>
>> It's really, really unlikely.  It's possible that some future processor
>> overloads PCONFIG in such a way that it will become virtualizable, but
>> not IceLake.
> 
> I guess, the likelihood of this happening would be similar to
> reintroducing other features, such as osxsave or ospke, right?

No, haveing osxsave and ospke was a mistake in the first place (they are
not CPU features at all; they are more like a special way to let
unprivileged programs read some bits of CR4).  For pconfig, it's just
very unlikely.

>> Would it make sense for libvirt to treat absent CPU flags as "default
>> off" during migration, so that it can leave out the flag in the command
>> line if it's off?  If it's on, libvirt would pass pconfig=on as usual.
>> This is a variant of [2], but more generally applicable:
>>
>>> [2] However starting a domain with Icelake-Server so that it can be
>>> migrated or saved/restored on QEMU in 3.1.1 and 4.0.0 would be
>>> impossible. This can be solved by a different hack, which would drop
>>> pconfig=off from QEMU command line.
> 
> The domain XML does not contain a complete list of all CPU features.
> Features which are implicitly included in a CPU model are not listed in
> the XML. Count in the differences in libvirt's vs QEMU's definitions of
> a particular CPU model and you can see feat=off cannot be mechanically
> dropped from the command line as the CPU model itself could turn it on
> by default and thus feat=off is not redundant.

I think I wasn't very clear, I meant "unsupported by QEMU" when I said
"absent".  Libvirt on the destination knows that from
query-cpu-model-expansion, so it can leave off pconfig if it is not
supported by the destination QEMU.

Paolo