Re: Default CPU models on s390x and ppc64

[David Hildenbrand <david@redhat.com>]

On 17.10.19 17:31, David Hildenbrand wrote:
> On 17.10.19 17:18, David Hildenbrand wrote:
>> On 17.10.19 17:16, Jiri Denemark wrote:
>>> Hi David and David,
>>>
>>> I'm working on libvirt's support [1] for query-machines'
>>> default-cpu-type, which is supposed to return the type of the default
>>> CPU model that QEMU uses for each machine type. Rather than hard coding
>>> the default in libvirt (which we currently do on x86), we ask QEMU for
>>> the default CPU model and use it unless a user asks for a specific CPU
>>> model explicitly.
>>>
>>> We use query-cpu-definitions for translating the default CPU type to the
>>> actual CPU model we need to pass to -cpu by looking up the CPU model
>>> with matching typename.
>>>
>>> However, all this seems to work only with TCG on both s390x and ppc64.
>>> The issues I met with KVM on each architecture are described below.
>>>
>>> On ppc64 the default CPU type reported by query-machines is
>>> power*-powerpc64-cpu, while IIUC QEMU would effectively use -cpu host by
>>> default. In fact -cpu power8 is mostly just a fancy alias to -cpu host
>>> on a Power8 machine. But QEMU even rewrites typename of the current host
>>> CPU model to host-powerpc64-cpu. Which means on a Power8 host the power8
>>> CPU model would have typename=host-powerpc64-cpu while the default CPU
>>> type would still use power8*-powerpc64-cpu. Thus we may just fail to
>>> find any CPU model corresponding to the default CPU model.
>>>
>>> And to make it even worse, the default CPU model changes with machine
>>> type. E.g., pseries-3.1 uses power8_v2.0-powerpc64-cpu while pseries-4.2
>>> uses power9_v2.0-powerpc64-cpu. However, starting QEMU with pseries-4.2
>>> machine type and the reported default -cpu power9 fails on any
>>> non-Power9 host. That said QEMU just lies about the default CPU model.
>>>
>>> So for all combinations of (pseries-3.1, pseries-4.2) machine types and
>>> (Power8, Power9) hosts, one will get the following mixed results:
>>>
>>> - pseries-3.1 on Power8: no default CPU model will be detected, QEMU
>>>      starts fine with it's own default
>>> - pseries-4.2 on Power9: same as above
>>> - pseries-3.1 on Power9: -cpu power8 (not sure if this works, though)
>>> - pseries-4.2 on Power8: -cpu power9, QEMU doesn't start
>>>
>>>
>>> This situation on s390x is not so complicated, but not really better.
>>> The default CPU is said to be "qemu" for all machine types, which works
>>> fine for TCG domains, but it doesn't work on KVM because QEMU complains
>>> that some features requested in the CPU model are not available. In
>>> other words the "qemu" CPU model is not runnable on KVM. This a bit
>>> similar to what happens on x86_64, but QEMU just ignores missing
>>> features and starts happily there.
>>
>> The default model under KVM is "host", under TCG it's "qemu". We should
>> not use "qemu" under KVM, although it might work on some setups ...
>>
>> Where/how is this default model detected?
>>
> 
> ... target/s390x/kvm.c:kvm_arch_init()
> 
> MachineClass *mc = MACHINE_GET_CLASS(ms);
> 
> mc->default_cpu_type = S390_CPU_TYPE_NAME("host");
> 
> 
> I think the call order should make sure that "host" is set after "qemu"
> is set. I'll go ahead and verify that.kvm_arch_init(
> 

configure_accelerator(current_machine, argv[0]) -> kvm_arch_init()

is called after

current_machine = MACHINE(object_new(object_class_get_name(
		  OBJECT_CLASS(machine_class))));

and therefore after the .class_init function of the machine was called.

I don't see how the default cpu model could not be "host" if qemu was 
started with "--enable-kvm"

-- 

Thanks,

David / dhildenb