Volume detach leaves volume attached to instance on start/rebuild/reboot

[Roman Podoliaka]

When starting/restarting/rebuilding instances, it may happen that a volume detach request comes right in the middle of attaching a volume in the driver. In this case the hypervisor (e.g. libvirt) will throw DiskNotFound exception in driver.detach_volume() call, but still the volume gets attached eventually, when the instance starts.

This leaves the instance in the state, when the volume is `de-facto` attached to it (i.e. shown in the `virsh dumpxmp $instance` output for libvirt), but both Nova and Cinder think the volume is actually *not* in-use.

Steps to reproduce:

1. Create an instance and attach a volume to it.
2. Stop the instance.
3. Start the instance and send a couple of volume-detach requests in a row, like:

   nova start demo && nova volume-detach demo $volume_id || nova volume-detach demo $volume_id || nova volume-detach demo $volume_id
 || nova volume-detach demo $volume_id

4. Check the cinder list, nova show $inst, virsh dumpxml $inst output.

Expected result:

Both cinder list and nova show report volume is not in-use anymore. There is no volume related elements in virsh dumpxml output.

Actual result:

Both cinder list and nova show report volume is not in-use anymore. But virsh dumpxml shows that the volume is still attached to the instance.

[OpenStack Infra (hudson-openstack)]

Change abandoned by Roman Podoliaka (<email address hidden>) on branch: master
Review: https://review.openstack.org/198414
Reason: abandon in favour of https://review.openstack.org/#/c/175067/