XenAPI should implement list_instance_uuids()

[Chris Behrens]

compute manager init_host() does a check to see if instances have been moved to other hosts and nukes them locally, if so. It makes use of driver.list_instance_uuids() but this doesn't exist in XenAPI. It attempts to fallback to another method that is somewhat kludgy and won't work every time. So... we should try to implement this method. We've been storing nova-uuid in other-data (or whatever) with VM records, so this shouldn't be too difficult to do.

[OpenStack Infra (hudson-openstack)]

Fix proposed to branch: master
Review: https://review.openstack.org/24704

[OpenStack Infra (hudson-openstack)]

Reviewed: https://review.openstack.org/24704
Committed: http://github.com/openstack/nova/commit/f9a89b7065ff4a15bf7ce14a983e8934cd4710f4
Submitter: Jenkins
Branch: master

commit f9a89b7065ff4a15bf7ce14a983e8934cd4710f4
Author: Chris Behrens <email address hidden>
Date: Mon Mar 18 20:26:02 2013 +0000

    Fix XenAPI performance issue

    This patch implements 'list_instance_uuids' in the xenapi virt driver so
    that compute manager's '_get_instances_on_driver' can operate more
    efficiently.

    Fixes bug 1097980

    The cleanup_running_deleted_instances periodic task uses the above call
    while context has been modified to be read_deleted='yes'. Without
    list_instance_uuids being implemented in xenapi, there's a fallback to
    querying all instances on the host. Because of read_deleted='yes', this
    queries all instances that have ever lived on the host. In a very busy
    environment where instances are repeatedly built and destroyed, one can
    end up with thousands of deleted instances. Now that we are storing
    instance_type data in system_metadata and system_metadata is joined with
    every instance_get, this results in 10x the number of rows being
    returned with sqlalchemy... the fallback doesn't perform well enough.

    Change-Id: I4bbfd69c9769807cec813af757665f03d9643460