On Mon, Apr 16, 2018 at 19:16:05 -0400, Collin Walling wrote:
[Overview]
This patch series implements an interface to "query-cpu-model-comparison"
(available QEMU ~2.8.0) via virsh cpu-compare.
[Using This Feature]
Run virsh cpu-compare (ensure you are running the virsh in your build dir) and
pass it an xml file describing a cpu definition. You can copy the cpu xml from
virsh capabilities (if you want to compare the host cpu to itself), or a cpu
defined in any guest xml. Additionally, you can create a cpu xml as such (e.g.
for s390x):
<cpu>
<arch>s390x</arch>
<model fallback='forbid'>model_name</model>
<feature policy='require|disable' name='feature_name'/>
</cpu>
NOTE: the presence of <arch> is optional and it will treat the cpu defined in
the xml as a host cpu. This will disregard all feature policies (i.e.
all features listed will behave with policy='require', even if disable
is specified).
NOTE: as s390x only supports feature policies 'require' and 'disable', I
am
uncertain how to handle the other policies when parsing CPUDef to JSON.
[Example Output]
On an s390x system running a z13.2, this is the expected output (where each file
describes a CPU model corresponding to the name of the file):
$ virsh cpu-compare zEC12.xml
Host CPU is a superset of CPU described in zEC12.xml
$ virsh cpu-compare z13.2.xml
CPU described in z13.2.xml is identical to host CPU
$ virsh cpu-compare z14.xml
CPU described in z14.xml is incompatible with host CPU
$ virsh cpu-compare z14.xml --error
error: Failed to compare host CPU with z14.xml
error: the CPU is incompatible with host CPU
$ virsh cpu-compare z12345.xml
error: Failed to compare host CPU with z12345cpu.xml
error: internal error: unable to execute QEMU command
'query-cpu-model-comparison': The CPU definition 'z12345-base' is
unknown.
[Patch Rundown]
The first patch copies the host CPU definition from qemuCaps to virCaps so
we can easily access the host CPU model and features when we handle the CPU
comparison in qemu_driver. Note that we take care to not clobber anything
already stored in the host CPU definition until we have successfully
constructed a new host CPU definition.
I think this is a wrong approach. You'd be basically giving random
answers depending on which QEMU binary is probed first. The reason for
storing the CPU model in qemuCaps is that it is tight to a particular
QEMU binary rather than the host itself. The model reported by one
binary may not be usable with other binaries and this applies to any
comparisons or other operations done with this CPU model.
In other words, we need to introduce a new set of CPU related APIs which
will take more arguments so that the caller may specify what binary,
virt type, and machine type they want to use. In other words, the APIs
should support parameters similar to virConnectGetDomainCapabilities().
I'm currently starting to work on these new APIs.
Jirka