libvirt users - Dec 2018 - Usable and non-usable CPU models in nested virtualization

Hi, some custom CPU models are reported from
virConnectGetDomainCapabilities as usable='yes' on a physical machine
while as usable='no' inside a VM running on the same machine. 
That's
not completely surprising.

But what surprises me is that those models are still reported from
virConnectCompareCPU as supported (VIR_CPU_COMPARE_SUPERSET) in the
nested environment and VMs can be started happily with them.

For instance, virConnectGetDomainCapabilities reports

  <model usable='no'>Skylake-Client</model>

but when I try to use that model anyway, the VM starts fine with it:

  <cpu mode='custom' match='exact' check='full'>
    <model fallback='forbid'>Skylake-Client</model>
    <topology sockets='16' cores='1' threads='1'/>
    <feature policy='require' name='hypervisor'/>
    <feature policy='disable' name='invpcid'/>
    <numa>
      <cell id='0' cpus='0' memory='524288'
unit='KiB'/>
    </numa>
  </cpu>

That's actually good news, but unexpected.  Do I miss something?

Thanks,
Milan

On Fri, Dec 07, 2018 at 11:52:38 +0100, Milan Zamazal
wrote:
> Hi, some custom CPU models are reported from
> virConnectGetDomainCapabilities as usable='yes' on a physical
machine
> while as usable='no' inside a VM running on the same machine. 
That's
> not completely surprising.
> 
> But what surprises me is that those models are still reported from
> virConnectCompareCPU as supported (VIR_CPU_COMPARE_SUPERSET) in the
virConnectCompareCPU uses CPUID data for comparison, which is not the
same as a list of features QEMU/KVM can provide on the host. You should
use virConnectCompareHypervisorCPU to check whether a given CPU can be
used on the host.
> nested environment and VMs can be started happily with them.
> 
> For instance, virConnectGetDomainCapabilities reports
> 
>   <model usable='no'>Skylake-Client</model>
> 
> but when I try to use that model anyway, the VM starts fine with it:
> 
>   <cpu mode='custom' match='exact'
check='full'>
>     <model fallback='forbid'>Skylake-Client</model>
>     <topology sockets='16' cores='1'
threads='1'/>
>     <feature policy='require' name='hypervisor'/>
>     <feature policy='disable' name='invpcid'/>
This is not the same as Skylake-Client, it's Skylake-Client without
invpcid. The usable='no' attribute says the Skylake-Client CPU model is
not usable unless you disable some features. You did that and it works.
If you asked for just Skylake-Client without any <feature> element, the
domain should fail to start.

Actually QEMU even reports what features need to be disabled to run each
CPU model, but I don't think that's really useful. You don't want to
disable all of them mechanically anyway since that can result in strange
CPU models which would confuse guests. That's why we only report the
usable=yes/no attribute.

Jirka

Jiri Denemark <jdenemar@redhat.com> writes:
> On Fri, Dec 07, 2018 at 11:52:38 +0100, Milan Zamazal wrote:
>> Hi, some custom CPU models are reported from
>> virConnectGetDomainCapabilities as usable='yes' on a physical
machine
>
>> while as usable='no' inside a VM running on the same machine. 
That's
>> not completely surprising.
>> 
>> But what surprises me is that those models are still reported from
>> virConnectCompareCPU as supported (VIR_CPU_COMPARE_SUPERSET) in the
>
> virConnectCompareCPU uses CPUID data for comparison, which is not the
> same as a list of features QEMU/KVM can provide on the host. You should
> use virConnectCompareHypervisorCPU to check whether a given CPU can be
> used on the host.
>
>> nested environment and VMs can be started happily with them.
>> 
>> For instance, virConnectGetDomainCapabilities reports
>> 
>>   <model usable='no'>Skylake-Client</model>
>> 
>> but when I try to use that model anyway, the VM starts fine with it:
>> 
>>   <cpu mode='custom' match='exact'
check='full'>
>>     <model fallback='forbid'>Skylake-Client</model>
>>     <topology sockets='16' cores='1'
threads='1'/>
>>     <feature policy='require' name='hypervisor'/>
>>     <feature policy='disable' name='invpcid'/>
>
> This is not the same as Skylake-Client, it's Skylake-Client without
> invpcid. The usable='no' attribute says the Skylake-Client CPU
model is
> not usable unless you disable some features. You did that and it works.
> If you asked for just Skylake-Client without any <feature> element,
the
> domain should fail to start.
Thank you for explanation.  However the behavior I observe is still not
clear to me.  The <cpu> snippet above is from a running domain,
successfully started from this definition:

  <cpu match="exact">
    <model>Skylake-Client</model>
    <topology cores="1" sockets="16"
threads="1" />
    <numa>
        <cell cpus="0" id="0" memory="524288"
/>
    </numa>
  </cpu>

When this definition is fed to compare CPU, I get:

  # virsh hypervisor-cpu-compare cpu.xml
  CPU described in cpu.xml is incompatible with the CPU provided by hypervisor
on the host

  # virsh cpu-compare cpu.xml
  Host CPU is a superset of CPU described in cpu.xml

It's not clear to me:

- Why is the domain successfully started despite hypervisor-cpu-compare
  rejects it?

- Why is `invpcid' disabled when `invpcid' is present in /proc/cpuinfo?

- What's the basic difference between virConnectCompareCPU and
  virConnectCompareHypervisorCPU?  Does "specific hypervisor and its
  abilities" (as stated in the documentation) mean that the hypervisor
  may extend CPU capabilities (by emulation), restrict CPU capabilities,
  or both (depending or particular feature etc.)?
> Actually QEMU even reports what features need to be disabled to run each
> CPU model, but I don't think that's really useful. You don't
want to
> disable all of them mechanically anyway since that can result in strange
> CPU models which would confuse guests. That's why we only report the
> usable=yes/no attribute.
>
> Jirka