Reproduce bug/2097359

In Victoria InstanceNUMACell ovo got the new pcpuset field and a
connected in place data migration. If the cpu_policy is dedicated the
content of the cpuset is moved to to the new pcpuset field during the
load of the data from the DB and the ovo is persisted back to the DB.
However during this data migration the version of the ovo is not bumped
to the latest, 1.6, version. Therefore the DB contains an inconsistent
object as it has the new pcpuset field from 1.6 but the
nova_object.version is still set to 1.4. It turned out that this can
cause that an old compute node having ovo 1.4 code will not request a
back levelling of the ovo even if it is already data migrated to 1.6
causing data loss from the compute perspective. Also if the compute
saves the object back to DB then the data loss becomes persistent.

Related-Bug: #2097359
Change-Id: I76ee9d59abc39e29c54be7217491e911b88a0621
(cherry picked from commit ae2f9bd)
(cherry picked from commit 82b5bd6)
(cherry picked from commit a3d3b89)
(cherry picked from commit db9a810)

Author: gibizer

nova/tests/unit/objects/test_instance_numa.py

@@ -13,6 +13,7 @@
 import copy
 from unittest import mock
 
+from oslo_serialization import jsonutils
 from oslo_utils.fixture import uuidsentinel as uuids
 from oslo_versionedobjects import base as ovo_base
 import testtools
@@ -411,6 +412,46 @@ def test_obj_from_db_obj_no_pinning(self):
             self.assertEqual(topo_cell.cpuset, obj_cell.cpuset)
             self.assertEqual(set(), obj_cell.pcpuset)
 
+    def test__migrate_legacy_dedicated_instance_cpuset(self):
+        # Create a topology with a cell on latest version. Would be nice
+        # to create one the old 1.4 cell version directly but that is only
+        # possible indirectly as done below.
+        topo = objects.InstanceNUMATopology(
+            instance_uuid=fake_instance_uuid,
+            cells=[
+                objects.InstanceNUMACell(id=0, cpuset=set(), pcpuset={0, 1}),
+            ])
+        topo.cells[0].cpu_policy = objects.fields.CPUAllocationPolicy.DEDICATED
+
+        # Use the builtin backlevelling logic to pull it back to old cell
+        # version
+        topo_with_cell_1_4 = topo.obj_to_primitive(
+            target_version='1.3', version_manifest={'InstanceNUMACell': '1.4'})
+
+        # Just check that the backlevelling works, and we have a cell with
+        # version and data on 1.4 level
+        cell_1_4_primitive = topo_with_cell_1_4['nova_object.data']['cells'][0]
+        self.assertEqual('1.4', cell_1_4_primitive['nova_object.version'])
+        self.assertEqual(
+            (0, 1), cell_1_4_primitive['nova_object.data']['cpuset'])
+        self.assertNotIn('pcpuset', cell_1_4_primitive['nova_object.data'])
+
+        # Now simulate that such old data is loaded from the DB and migrated
+        # from 1.4 to 1.6 by the data migration
+        topo_loaded = objects.InstanceNUMATopology.obj_from_db_obj(
+            self.context, fake_instance_uuid,
+            jsonutils.dumps(topo_with_cell_1_4))
+
+        # In place data migration happens during load, cpuset data is moved to
+        # pcpuset
+        self.assertEqual(set(), topo_loaded.cells[0].cpuset)
+        self.assertEqual({0, 1}, topo_loaded.cells[0].pcpuset)
+        # but the object version isn't bumped. So when the
+        # data is saved back to the DB it still has the old version 1.4, but
+        # also it has the new pcpuset field from version 1.6. This is bug
+        # https://bugs.launchpad.net/nova/+bug/2097359.
+        self.assertEqual('1.4', topo_loaded.cells[0].VERSION)
+
 
 class TestInstanceNUMATopology(
     test_objects._LocalTest, _TestInstanceNUMATopology,