arm64: Migration pre-check error: CPU doesn't have compatibility.

[dann frazier]

Pike/openstack-base running on identical servers (HiSilicon D05):

ubuntu@ike-hisi-maas:~$ openstack server migrate --live strong-emu dannf
Migration pre-check error: CPU doesn't have compatibility.

XML error: Missing CPU model name

Refer to http://libvirt.org/html/libvirt-libvirt-host.html#virCPUCompareResult (HTTP 400) (Request-ID: req-c5ec9320-d111-40b7-af0e-d8414df3925c)

[James Page (james-page)]

My initial hunch was something was wonky with the cpu configuration in nova; however:

        if config('cpu-mode'):
            ctxt['cpu_mode'] = config('cpu-mode')
        elif ctxt['arch'] in ('ppc64el', 'ppc64le', 'aarch64'):
            ctxt['cpu_mode'] = 'host-passthrough'

reveals that for aarch64, we actually do a direct host-passthrough so for identical machines the CPU features should be a 100% match; I'm wondering whether something is not exposed correctly in libvirt on aarch64 which nova is looking for for the pre-flight checks.

[James Page (james-page)]

"Missing CPU model name" is bubbling up from libvirt's pre-migration checks.

[James Page (james-page)]

Dann - does this work in releases prior to pike? Reading the nova code, not a huge amount has changed in this area, so I'm wondering whether this is a bug somewhere in the virt stack @ pike.

[dann frazier (dannf)]

James - I never tried it with earlier releases. However, prior to pike, there was the additional reason that QEMU didn't yet support save/restore of these guests at all. See LP: #1710019.

[Jason Hobbs (jason-hobbs)]

Live migration test: https://github.com/openstack/tempest/blob/master/tempest/api/compute/admin/test_live_migration.py

[Andrew McLeod (admcleod)]

The issue is that the XML does not contain the CPU model type, following is what is actually being passed.

https://github.com/openstack/nova/blob/master/nova/virt/libvirt/driver.py#L6337

This is the actual XML provided on an arm64 deploy (can be confirmed using "virsh cpu-compare file.xml" -

<cpu>
  <arch>aarch64</arch>
  <topology sockets="1" cores="48" threads="1"/>
</cpu>

more logs: https://pastebin.canonical.com/209211/

If you insert <model>anything</model> then cpu-compare works, e.g.

https://pastebin.canonical.com/209216/

However, https://github.com/openstack/nova/blob/f94c728bdddc9d1d3e3ae5f8f34b0111b9168bc9/releasenotes/notes/aarch64-set-proper-cpu-mode-8455bad7d69dc6fd.yaml indicates that no host-model will be passed.

[Christian Ehrhardt  (paelzer)]

While I agree for cpu-compare needing a cpu to compare.
I happen to migrate guests like http://paste.ubuntu.com/26530819/ all the time which have none as well.

So there must be way to pass it even with no cpu model is defined.

[Marcin Juszkiewicz (hrw)]

aarch64 does not have cpu_model = "host-model" funtionality which would be required for it.

There is no sane way of knowing how cpu1 compare to cpu2 yet.

[Sylvain Bauza (sylvain-bauza)]

Since Aarch64 architectures doesn't provide CPU models, but since we compare CPU models when migrating, that's why we have that issue.

I think tacking that bug requires more than a single bugfix and is probably a feature, but I leave that bug open for records, and we'll see how to address that in the future.

[Sylvain Bauza (sylvain-bauza)]

Anyway, I feel we should amend https://docs.openstack.org/nova/latest/user/feature-classification.html and https://docs.openstack.org/nova/latest/user/support-matrix.html to tell what ARM64 can support.