Call Neutron immediately upon _post_live_migration() start

Previously, we first called to Cinder to cleanp source volume
connections, then to Neutron to activate destination port bindings.
This means that, if Cinder or the storage backend were slow, the live
migrated instance would be without network connectivity during that
entire process.

This patch immediately activates the port bindings on the destination
(and thus gives the instance network connectivity). We just need to
get the old network_info first, in order to use it in notifications
and to stash it for the later call to
driver.post_live_migration_at_source().

This is a smaller and safer change than the parallelization attempt in
the subsequent patch, so it's done in its own patch because it might
be backportable, and would help with network downtime during live
migration.

To avoid any potential data leaks, we want to be certain to cleanup
volumes on the source. To that end we wrap the code that is being
moved before the source volume cleanup code in a try/finally block
in order to prevent any uncaught exception from blocking the cleanup.

Change-Id: I700943723a32e732e3e3be825f3fd44a9f923a0b

Author: notartom

nova/compute/manager.py

@@ -9337,41 +9337,50 @@ def _post_live_migration(self, ctxt, instance, dest,
         LOG.info('_post_live_migration() is started..',
                  instance=instance)
 
-        # Cleanup source host post live-migration
-        block_device_info = self._get_instance_block_device_info(
-                            ctxt, instance, bdms=source_bdms)
-        self.driver.post_live_migration(ctxt, instance, block_device_info,
-                                        migrate_data)
-
-        # Disconnect volumes from this (the source) host.
-        self._post_live_migration_remove_source_vol_connections(
-            ctxt, instance, source_bdms)
-
-        # NOTE(artom) At this point in time we have not bound the ports to the
-        # destination host yet (this happens in migrate_instance_start()
-        # below). Therefore, the "old" source network info that's still in the
-        # instance info cache is safe to use here, since it'll be used below
-        # during driver.post_live_migration_at_source() to unplug the VIFs on
-        # the source.
-        network_info = instance.get_network_info()
+        # NOTE(artom) The ordering and exception handling are important here.
+        # We want to immediately activate the port bindings on the destination
+        # host to minimize network downtime. This happens in
+        # migrate_instance_start(). It's also crucial that we call
+        # _post_live_migration_remove_source_vol_connections() to clean up
+        # source volume connections and prevent potential data leaks. We
+        # therefore activate the port bindings in a try block, and, regardless
+        # of any expcetions during that process, clean up volume connections in
+        # a finally block.
+        try:
+            # NOTE(artom) At this point in time we have not bound the ports to
+            # the destination host yet (this happens in
+            # migrate_instance_start() below). Therefore, the "old" source
+            # network info that's still in the instance info cache is safe to
+            # use here, since it'll be used below during
+            # driver.post_live_migration_at_source() to unplug the VIFs on the
+            # source.
+            network_info = instance.get_network_info()
 
-        self._notify_about_instance_usage(ctxt, instance,
-                                          "live_migration._post.start",
-                                          network_info=network_info)
-        compute_utils.notify_about_instance_action(
-            ctxt, instance, self.host,
-            action=fields.NotificationAction.LIVE_MIGRATION_POST,
-            phase=fields.NotificationPhase.START)
+            self._notify_about_instance_usage(ctxt, instance,
+                                              "live_migration._post.start",
+                                              network_info=network_info)
+            compute_utils.notify_about_instance_action(
+                ctxt, instance, self.host,
+                action=fields.NotificationAction.LIVE_MIGRATION_POST,
+                phase=fields.NotificationPhase.START)
 
-        migration = objects.Migration(
-            source_compute=self.host, dest_compute=dest,
-        )
-        # For neutron, migrate_instance_start will activate the destination
-        # host port bindings, if there are any created by conductor before live
-        # migration started.
-        self.network_api.migrate_instance_start(ctxt,
-                                                instance,
-                                                migration)
+            migration = objects.Migration(
+                source_compute=self.host, dest_compute=dest,
+            )
+            # For neutron, migrate_instance_start will activate the destination
+            # host port bindings, if there are any created by conductor before
+            # live migration started.
+            self.network_api.migrate_instance_start(ctxt, instance, migration)
+        finally:
+            # Cleanup source host post live-migration
+            block_device_info = self._get_instance_block_device_info(
+                                ctxt, instance, bdms=source_bdms)
+            self.driver.post_live_migration(ctxt, instance, block_device_info,
+                                            migrate_data)
+
+            # Disconnect volumes from this (the source) host.
+            self._post_live_migration_remove_source_vol_connections(
+                ctxt, instance, source_bdms)
 
         destroy_vifs = False
         try: