Originally reported in Red Hat Bugzilla: https://bugzilla.redhat.com/show_bug.cgi?id=1648931
Create a cinder volume, attach it to a nova instance, and migrate the volume to a different storage host:
$ cinder create 1 --volume-type foo --name myvol
$ nova volume-attach myinstance myvol
$ cinder migrate myvol c-vol2
Everything seems to work correctly, but if we look at myinstance we see that it's now connected to a new volume, and the original volume is still present on the original storage host.
This is because nova didn't call cinder's migrate_volume_completion. migrate_volume_completion would have deleted the original volume, and changed the volume id of the new volume to be the same as the original. The result would be that myinstance would appear to be connected to the same volume as before.
Note that there are 2 ways (that I'm aware of) to intiate a cinder volume migration: retype and migrate. AFAICT retype is *not* affected. In fact, I updated the relevant tempest test to try to trip it up and it didn't fail. However, an exlicit migrate *is* affected. They are different top-level entry points in cinder, and set different state, which is what triggers the Nova bug.
This appears to be a regression which was introduced by https://review.openstack.org/#/c/456971/ :
# Yes this is a tightly-coupled state check of what's going on inside
# cinder, but we need this while we still support old (v1/v2) and
# new style attachments (v3.44). Once we drop support for old style
# attachments we could think about cleaning up the cinder-initiated
# swap volume API flows.
is_cinder_migration = (
True if old_volume['status'] in ('retyping',
'migrating') else False)
There's a bug here because AFAICT cinder never sets status to 'migrating' during any operation: it sets migration_status to 'migrating' during both retype and migrate. During retype it sets status to 'retyping', but not during an explicit migrate.
Fix proposed to branch: master
Review: https:/