Retry after hitting libvirt error code VIR_ERR_OPERATION_INVALID in live migration.

Fix proposed to branch: master
Review: https://review.openstack.org/612272

This bug looks valid, marking as High as it involves data loss.

i dont think i agree that we should retry as if the vm has been killed its unlikely that we woudl be able to recover.

as noted on the review, if you are doing a post copy migration and either vm is killed in the post copy phase then you will lose the contentce of the guest ram.

if we are doing a pre copy migration and the dest vm is killed before we pause the source vm we could revert the migiration, mark it as faild and allow the vm to continue to execute on the source node.
the alternitve woudl be to catach the invalid operation and restart the live migratrion unless libvirt internally will detect the dest vm exited and recreate it? i doubt it will do that so if we get an invalid operation when we get the job stats i dont think a retry will ever succeed.

In our case, we didn't permit post copy live migration, a pre copy migration was executing, and during migration process, vm, aka qemu process on the *source node* was killed due to host OOM. The domain status is SHUTOFF, then in get_job_info(), self._domain.jobStats() got a libvirt error 'VIR_ERR_OPERATION_INVALID'. In previous code, nova thinks the domain is shutdown or gone away, so it happily return JobInfo(type=libvirt.VIR_DOMAIN_JOB_COMPLETED), but it will eventually trigger post_live_migration() to delete source vm files. That's why I report this bug.

IMHO, if qemu-kvm process was killed by source host OOM, we would get error code VIR_ERR_OPERATION_INVALID reported by libvirt because domain state is SHUTOFF and we try to execute `self._domain.jobStats()`. In this case, migration job should be considered failed. If migration succeeded, libvirt would also kill qemu-kvm process, and domain state is SHUTOFF. Then we could get error code VIR_ERR_OPERATION_INVALID, but in such case, we should consider VIR_ERR_OPERATION_INVALID as nothing, just return JobInfo with type=VIR_DOMAIN_JOB_COMPLETED. The difference between these two cases is that under the latter circumstances, we would eventually get VIR_ERR_NO_DOMAIN if we try to get job info for couple of more times.

i think this is actully another example of https://bugs.launchpad.net/nova/+bug/1788014
in that we had a general problem where we tread all event as success.

i agree that its a bug that we treat it as sucess and then end up deleting the vm but not with the retry. when we get VIR_ERR_OPERATION_INVALID i think we should fail the migration immediately and rollback without retrying.

the fix for https://bugs.launchpad.net/nova/+bug/1788014 has not been backported to queens yet
but if we look at the change https://review.opendev.org/#/c/700774/1/nova/virt/libvirt/host.py
the primary thing we are doing is looking at the detail of the event to determin if the
libvirt.VIR_DOMAIN_EVENT_SUSPENDED_MIGRATED event was a success and signals the completion of the migration or if it was an error.

by the way i assume you are refering to https://github.com/open stack/nova/blob/stable/queens/nova/virt/libvirt/migration.py#L240-L258 when you say "nova thinks the domain is shutdown or gone away, so it happily return JobInfo(type=libvirt.VIR_DOMAIN_JOB_COMPLETED),"

on queens we treat all the VIR_DOMAIN_EVENT_* as sucess and our huristic fo determining if a migration suceeded wont handel the OOM case so we proceed to post live migrate when we should have failed the migration and rolled back. when we recive an invalid operation error form libvirt and called find_job_type we really shoudl end up taking the exception path and return libvirt.VIR_DOMAIN_JOB_FAILED

im not sure if backporting https://review.opendev.org/#/c/700774/1 to queens would also solve this issue but i think that is the direction we shoudl go to adress this.

I was refering to [1], when I said nova returns JobInfo(type=libvirt.VIR_DOMAIN_JOB_COMPLETED). The
key point I want to say is that in our case, we meet VIR_ERR_OPERATION_INVALID error when trying to
get job stats by self._domain.jobStats(). In [1], nova already returns JobInfo(type=libvirt.VIR_DOMAIN_JOB_COMPLETED), so post live migration is executing later.

As for the patch you mentioned, I'll check it later. Thanks.

[1]. https://github.com/openstack/nova/blob/stable/queens/nova/virt/libvirt/guest.py#L719-L725,