Detach Volume Ordering Has Bad User Experience

[Vish Ishaya]

Currently when detaching a volume we do:

1. driver.detach() -> detaches the volume in the hypervisor
2. cinder.detach() -> marks the volume detached in cinder
3. bdm.delete() -> stops the nova side from reporting an attached volume.

This leads to bad UX for two reasons:

a. If the cinder detach fails, the bdm still exists so nova reports a volume attached even when the hypervisor
    has detached it.
b. There is a window where cinder reports the volume available but nova still thinks the attachment exists

I propose we reverse the order of 2. and 3.

This reverses a. so that a detach fail will show the volume as attached in cinder but nova will not show it. The nova side more accurately reflects what nova knows about.
This reverses b. so that there is a window where nova has removed the attachment but cinder still reports it in use. This is a fairly minor difference, but leads to a nicer experience when you are detaching and re-attaching the same volume because it can be reattached once it becomes available.

This also may help with https://bugs.launchpad.net/nova/+bug/1172695

[Openstack Gerrit (openstack-gerrit)]

Fix proposed to branch: master
Review: https://review.openstack.org/88416

[OpenStack Infra (hudson-openstack)]

Reviewed: https://review.openstack.org/88416
Committed: https://git.openstack.org/cgit/openstack/nova/commit/?id=94a3b83f9f1fd52a78b9d49b32ddfae40182f852
Submitter: Jenkins
Branch: master

commit 94a3b83f9f1fd52a78b9d49b32ddfae40182f852
Author: Vishvananda Ishaya <email address hidden>
Date: Thu Apr 17 12:45:25 2014 -0700

    Reverse order of cinder.detach() and bdm.delete()

    The user experience for detaching a volume is better if the bdm
    is destroyed as soon as the driver has detached the volume instead
    of waiting until after we have updated cinder.

    Change-Id: I6f3de25f7737db270af4ba41d7451ae95f3950f5
    Closes-Bug: #1309154