Fix pre_live_migration rollback

During the pre live migration process, Nova performs most of the
tasks related to the creation and operation of the VM in the destination
host. That is done without interrupting any of the hardware in the source
host. If the pre_live_migration fails, those same operations should be
rolled back.

Currently nova is sharing the _rollback_live_migration for both
live and pre_live migration rollbacks, and that is causing the source
host to try to re-attach network interfaces on the source host where
they weren't actually de-attached.

This patch fixes that by adding a conditional to allow nova to do
different paths for migration and pre_live_migration rollbacks.

Closes-bug: #1944619
Change-Id: I784190ac356695dd508e0ad8ec31d8eaa3ebee56
(cherry picked from commit 63ffba7)

Author: sombrafam

nova/compute/manager.py

@@ -8413,7 +8413,8 @@ def _cleanup_pre_live_migration(self, context, dest, instance,
         migrate_data.migration = migration
         self._rollback_live_migration(context, instance, dest,
                                       migrate_data=migrate_data,
-                                      source_bdms=source_bdms)
+                                      source_bdms=source_bdms,
+                                      pre_live_migration=True)
 
     def _do_pre_live_migration_from_source(self, context, dest, instance,
                                            block_migration, migration,
@@ -9167,7 +9168,8 @@ def _rollback_volume_bdms(self, context, bdms, original_bdms, instance):
     def _rollback_live_migration(self, context, instance,
                                  dest, migrate_data=None,
                                  migration_status='failed',
-                                 source_bdms=None):
+                                 source_bdms=None,
+                                 pre_live_migration=False):
         """Recovers Instance/volume state from migrating -> running.
 
         :param context: security context
@@ -9217,8 +9219,14 @@ def _rollback_live_migration(self, context, instance,
         #                  for nova-network)
         # NOTE(mriedem): This is a no-op for neutron.
         self.network_api.setup_networks_on_host(context, instance, self.host)
-        self.driver.rollback_live_migration_at_source(context, instance,
-                                                      migrate_data)
+
+        # NOTE(erlon): We should make sure that rollback_live_migration_at_src
+        # is not called in the pre_live_migration rollback as that will trigger
+        # the src host to re-attach interfaces which were not detached
+        # previously.
+        if not pre_live_migration:
+            self.driver.rollback_live_migration_at_source(context, instance,
+                                                          migrate_data)
 
         # NOTE(lyarwood): Fetch the current list of BDMs, disconnect any
         # connected volumes from the dest and delete any volume attachments

nova/tests/functional/regressions/test_bug_1944619.py

@@ -72,11 +72,5 @@ def test_rollback_pre_live_migration(self):
                 self._live_migrate(self.server,
                                    migration_expected_state='failed',
                                    server_expected_state='MIGRATING')
-        # FIXME(erlon): In the current behavior,
-        # rollback_live_migration_at_source is called if an error happens
-        # during the  pre_live_migration phase on the destination and therefore
-        # triggers the observed bug. rollback_live_migration_at_source should
-        # *not* be called for when errors happen during pre_live_migration
-        # phase.
-        mlpr.assert_called_once()
+        mlpr.assert_not_called()
         mlpp.assert_called_once()

nova/tests/unit/compute/test_compute_mgr.py

@@ -9539,7 +9539,8 @@ def test_live_migration_wait_vif_plugged_vif_plug_error(
         self.assertEqual('error', self.migration.status)
         mock_rollback_live_mig.assert_called_once_with(
             self.context, self.instance, 'dest-host',
-            migrate_data=migrate_data, source_bdms=source_bdms)
+            migrate_data=migrate_data, source_bdms=source_bdms,
+            pre_live_migration=True)
 
     @mock.patch('nova.compute.rpcapi.ComputeAPI.pre_live_migration')
     @mock.patch('nova.compute.manager.ComputeManager._rollback_live_migration')
@@ -9574,7 +9575,8 @@ def test_live_migration_wait_vif_plugged_timeout_error(
         self.assertEqual('error', self.migration.status)
         mock_rollback_live_mig.assert_called_once_with(
             self.context, self.instance, 'dest-host',
-            migrate_data=migrate_data, source_bdms=source_bdms)
+            migrate_data=migrate_data, source_bdms=source_bdms,
+            pre_live_migration=True)
 
     @mock.patch('nova.compute.rpcapi.ComputeAPI.pre_live_migration')
     @mock.patch('nova.compute.manager.ComputeManager._rollback_live_migration')

releasenotes/notes/bug-1944619-fix-live-migration-rollback.yaml

@@ -0,0 +1,10 @@
+---
+fixes:
+  - |
+    Instances with hardware offloaded ovs ports no longer lose connectivity
+    after failed live migrations. The driver.rollback_live_migration_at_source
+    function is no longer called during during pre_live_migration rollback
+    which previously resulted in connectivity loss following a failed live
+    migration. See `Bug 1944619`_ for more details.
+
+    .. _Bug 1944619: https://bugs.launchpad.net/nova/+bug/1944619