Delete allocations even if _confirm_resize raises

When we are confirming a resize, the guest is on the dest
host and the instance host/node values in the database
are pointing at the dest host, so the _confirm_resize method
on the source is really best effort. If something fails, we
should not leak allocations in placement for the source compute
node resource provider since the instance is not actually
consuming the source node provider resources.

This change refactors the error handling around the _confirm_resize
call so the big nesting for _error_out_instance_on_exception is
moved to confirm_resize and then a try/finally is added around
_confirm_resize so we can be sure to try and cleanup the allocations
even if _confirm_resize fails in some obscure way. If _confirm_resize
does fail, the error gets re-raised along with logging a traceback
and hint about how to correct the instance state in the DB by hard
rebooting the server on the dest host.

Change-Id: I29c5f491ec20a71283190a1599e7732541de736f
Closes-Bug: #1821594

Author: mriedem

nova/compute/manager.py

@@ -3973,7 +3973,29 @@ def do_confirm_resize(context, instance, migration_id):
                          instance=instance)
                 return
 
-            self._confirm_resize(context, instance, migration=migration)
+            with self._error_out_instance_on_exception(context, instance):
+                try:
+                    self._confirm_resize(
+                        context, instance, migration=migration)
+                except Exception:
+                    # Something failed when cleaning up the source host so
+                    # log a traceback and leave a hint about hard rebooting
+                    # the server to correct its state in the DB.
+                    with excutils.save_and_reraise_exception(logger=LOG):
+                        LOG.exception(
+                            'Confirm resize failed on source host %s. '
+                            'Resource allocations in the placement service '
+                            'will be removed regardless because the instance '
+                            'is now on the destination host %s. You can try '
+                            'hard rebooting the instance to correct its '
+                            'state.', self.host, migration.dest_compute,
+                            instance=instance)
+                finally:
+                    # Whether an error occurred or not, at this point the
+                    # instance is on the dest host so to avoid leaking
+                    # allocations in placement, delete them here.
+                    self._delete_allocation_after_move(
+                        context, instance, migration)
 
         do_confirm_resize(context, instance, migration.id)
 
@@ -3985,59 +4007,57 @@ def _confirm_resize(self, context, instance, migration=None):
             self.host, action=fields.NotificationAction.RESIZE_CONFIRM,
             phase=fields.NotificationPhase.START)
 
-        with self._error_out_instance_on_exception(context, instance):
-            # NOTE(danms): delete stashed migration information
-            old_instance_type = instance.old_flavor
-            instance.old_flavor = None
-            instance.new_flavor = None
-            instance.system_metadata.pop('old_vm_state', None)
-            instance.save()
-
-            # NOTE(tr3buchet): tear down networks on source host
-            self.network_api.setup_networks_on_host(context, instance,
-                               migration.source_compute, teardown=True)
+        # NOTE(danms): delete stashed migration information
+        old_instance_type = instance.old_flavor
+        instance.old_flavor = None
+        instance.new_flavor = None
+        instance.system_metadata.pop('old_vm_state', None)
+        instance.save()
 
-            network_info = self.network_api.get_instance_nw_info(context,
-                                                                 instance)
-            # TODO(mriedem): Get BDMs here and pass them to the driver.
-            self.driver.confirm_migration(context, migration, instance,
-                                          network_info)
+        # NOTE(tr3buchet): tear down networks on source host
+        self.network_api.setup_networks_on_host(context, instance,
+                           migration.source_compute, teardown=True)
 
-            migration.status = 'confirmed'
-            migration.save()
+        network_info = self.network_api.get_instance_nw_info(context,
+                                                             instance)
+        # TODO(mriedem): Get BDMs here and pass them to the driver.
+        self.driver.confirm_migration(context, migration, instance,
+                                      network_info)
 
-            self.rt.drop_move_claim(context, instance, migration.source_node,
-                                    old_instance_type, prefix='old_')
-            self._delete_allocation_after_move(context, instance, migration)
-            instance.drop_migration_context()
+        migration.status = 'confirmed'
+        migration.save()
 
-            # NOTE(mriedem): The old_vm_state could be STOPPED but the user
-            # might have manually powered up the instance to confirm the
-            # resize/migrate, so we need to check the current power state
-            # on the instance and set the vm_state appropriately. We default
-            # to ACTIVE because if the power state is not SHUTDOWN, we
-            # assume _sync_instance_power_state will clean it up.
-            p_state = instance.power_state
-            vm_state = None
-            if p_state == power_state.SHUTDOWN:
-                vm_state = vm_states.STOPPED
-                LOG.debug("Resized/migrated instance is powered off. "
-                          "Setting vm_state to '%s'.", vm_state,
-                          instance=instance)
-            else:
-                vm_state = vm_states.ACTIVE
+        self.rt.drop_move_claim(context, instance, migration.source_node,
+                                old_instance_type, prefix='old_')
+        instance.drop_migration_context()
+
+        # NOTE(mriedem): The old_vm_state could be STOPPED but the user
+        # might have manually powered up the instance to confirm the
+        # resize/migrate, so we need to check the current power state
+        # on the instance and set the vm_state appropriately. We default
+        # to ACTIVE because if the power state is not SHUTDOWN, we
+        # assume _sync_instance_power_state will clean it up.
+        p_state = instance.power_state
+        vm_state = None
+        if p_state == power_state.SHUTDOWN:
+            vm_state = vm_states.STOPPED
+            LOG.debug("Resized/migrated instance is powered off. "
+                      "Setting vm_state to '%s'.", vm_state,
+                      instance=instance)
+        else:
+            vm_state = vm_states.ACTIVE
 
-            instance.vm_state = vm_state
-            instance.task_state = None
-            instance.save(expected_task_state=[None, task_states.DELETING,
-                                               task_states.SOFT_DELETING])
+        instance.vm_state = vm_state
+        instance.task_state = None
+        instance.save(expected_task_state=[None, task_states.DELETING,
+                                           task_states.SOFT_DELETING])
 
-            self._notify_about_instance_usage(
-                context, instance, "resize.confirm.end",
-                network_info=network_info)
-            compute_utils.notify_about_instance_action(context, instance,
-                   self.host, action=fields.NotificationAction.RESIZE_CONFIRM,
-                   phase=fields.NotificationPhase.END)
+        self._notify_about_instance_usage(
+            context, instance, "resize.confirm.end",
+            network_info=network_info)
+        compute_utils.notify_about_instance_action(context, instance,
+               self.host, action=fields.NotificationAction.RESIZE_CONFIRM,
+               phase=fields.NotificationPhase.END)
 
     def _delete_allocation_after_move(self, context, instance, migration):
         """Deletes resource allocations held by the migration record against

nova/tests/unit/compute/test_compute_mgr.py

@@ -7183,6 +7183,8 @@ def do_finish_revert_resize(mock_attachment_complete,
         do_finish_revert_resize()
 
     def test_confirm_resize_deletes_allocations(self):
+        @mock.patch('nova.objects.Instance.get_by_uuid')
+        @mock.patch('nova.objects.Migration.get_by_id')
         @mock.patch.object(self.migration, 'save')
         @mock.patch.object(self.compute, '_notify_about_instance_usage')
         @mock.patch.object(self.compute, 'network_api')
@@ -7192,13 +7194,16 @@ def test_confirm_resize_deletes_allocations(self):
         @mock.patch.object(self.instance, 'save')
         def do_confirm_resize(mock_save, mock_drop, mock_delete,
                               mock_confirm, mock_nwapi, mock_notify,
-                              mock_mig_save):
+                              mock_mig_save, mock_mig_get, mock_inst_get):
             self._mock_rt()
             self.instance.migration_context = objects.MigrationContext()
             self.migration.source_compute = self.instance['host']
             self.migration.source_node = self.instance['node']
-            self.compute._confirm_resize(self.context, self.instance,
-                                         self.migration)
+            self.migration.status = 'confirming'
+            mock_mig_get.return_value = self.migration
+            mock_inst_get.return_value = self.instance
+            self.compute.confirm_resize(self.context, self.instance,
+                                        self.migration)
             mock_delete.assert_called_once_with(self.context, self.instance,
                                                 self.migration)
             mock_save.assert_called_with(expected_task_state=
@@ -7229,9 +7234,10 @@ def test_confirm_resize_driver_confirm_migration_fails(
         with test.nested(
             mock.patch.object(self.compute, 'network_api'),
             mock.patch.object(self.compute.driver, 'confirm_migration',
-                              side_effect=error)
+                              side_effect=error),
+            mock.patch.object(self.compute, '_delete_allocation_after_move')
         ) as (
-            network_api, confirm_migration
+            network_api, confirm_migration, delete_allocation
         ):
             self.assertRaises(exception.HypervisorUnavailable,
                               self.compute.confirm_resize,
@@ -7245,6 +7251,10 @@ def test_confirm_resize_driver_confirm_migration_fails(
         self.assertEqual(2, instance_save.call_count)
         # The migration.status should have been saved.
         self.migration.save.assert_called_once_with()
+        # Allocations should always be cleaned up even if cleaning up the
+        # source host fails.
+        delete_allocation.assert_called_once_with(
+            self.context, self.instance, self.migration)
         # Assert other mocks we care less about.
         notify_usage.assert_called_once()
         notify_action.assert_called_once()