Re: [Qemu-devel] [PATCH 3/3] S390: Enable -cpu help and QMP query-cpu-definitions

[Alexander Graf <agraf@suse.de>]

On 12/12/2012 05:28 PM, Andreas Färber wrote:
> Am 12.12.2012 16:05, schrieb Viktor Mihajlovski:
>> On 12/12/2012 02:51 PM, Andreas Färber wrote:
>>>> +    (*cpu_fprintf)(f, "s390 %16s\n", "[host]");
>>> Note that the square-bracket notation was specific to x86 when it
>>> distinguished between built-in and config-based models.
>>>
>> OK, since libvirt capable of dealing with s390 cpu models will
>> never see this, we can change it any way that is wanted.
>> So, host without brackets?
> Yes, but see below...

Yes, and with #ifdef CONFIG_KVM :). That way the TCG behavior stays 
untouched.

>
>>> "host" only makes sense for KVM, not for TCG. So we would need one other
>>> placeholder model for libvirt.
>> see my reply to Alex' comment: the placeholder name must be chosen
>> carefully, i.e. future-proof
> On further thoughts, didn't we discuss that the issue libvirt wants to
> address is that migration from z10 to z9 must fail? That's not solved
> with -cpu host, we would need two other models then. IMO ideally -cpu
> host should have the same semantics as on x86, that is passing the host
> features through mostly 1:1. IIUC there is currently no way to not do so?

That is another problem that we need to solve, but one thing at a time.

> What about future-wise having -cpu host not be a subclass and instead
> behaving like Alex' -cpu best, given the above semantics? What I am just
> worried about with this patch is cementing the use of -cpu host into
> libvirt when that is not a mid-term solution.

I think on s390 we can get away with the same method as ppc for -cpu 
host, which basically is -cpu best without fuzziness.


Alex