[OVN] Handle OVN agents when "Chassis" register is deleted

If an "ovn-controller" ends not gracefully, the node "Chassis" and
"Chassis_Private" registers will remain in the OVN SB database.
Because there is no a mandatory procedure to delete the "Chassis"
and "Chassis_Private" registers, the administrator can manually
delete, from the OVN SB database, any register in any order.

If the "Chassis" register is deleted and the Neutron server restarted,
the updated "Chassis_Private" register will be read from the database.
That won't contain the "Chassis" information as this register has been
deleted. In this case, the ``NeutronAgent`` returns ``DeletedChassis``,
an empty chassis register with no information.

NOTE: the sequence of actions ("Chassis" register deletion, Neutron
server restart) must be follow to reproduce this issue. If the
"Chassis" register is deleted, the Neutron server OVN agent local cache
won't update the stored information and will keep the previous value.
It is when the Neutron server is restarted when the OVN agent local
cache is retrieved again; at this time the "Chassis_Private" register
won't have any related "Chassis" register.

Closes-Bug: #1951149

Change-Id: I17aa53cea6aba8ea83187c99102a6f25fd33cfff
(cherry picked from commit f1a5511)

Author: ralonsoh

neutron/plugins/ml2/drivers/ovn/agent/neutron_agent.py

@@ -23,6 +23,12 @@
 from neutron.common import utils
 
 
+class DeletedChassis(object):
+    external_ids = {}
+    hostname = '("Chassis" register deleted)'
+    name = '("Chassis" register deleted)'
+
+
 class NeutronAgent(abc.ABC):
     types = {}
 
@@ -45,9 +51,12 @@ def update(self, chassis_private, updated_at=None, clear_down=False):
     def chassis_from_private(chassis_private):
         try:
             return chassis_private.chassis[0]
-        except (AttributeError, IndexError):
+        except AttributeError:
             # No Chassis_Private support, just use Chassis
             return chassis_private
+        except IndexError:
+            # Chassis register has been deleted but not Chassis_Private.
+            return DeletedChassis
 
     @property
     def chassis(self):

neutron/tests/functional/plugins/ml2/drivers/ovn/mech_driver/test_mech_driver.py

@@ -997,9 +997,9 @@ class AgentWaitEvent(event.WaitEvent):
 
     ONETIME = False
 
-    def __init__(self, driver, chassis_names):
+    def __init__(self, driver, chassis_names, events=None):
         table = driver.agent_chassis_table
-        events = (self.ROW_CREATE,)
+        events = events or (self.ROW_CREATE,)
         self.chassis_names = chassis_names
         super().__init__(events, table, None)
         self.event_name = "AgentWaitEvent"
@@ -1059,11 +1059,41 @@ def test_agent_list(self):
             self.context, filters={'host': self.host})]
         self.assertCountEqual(list(self.agent_types.values()), agent_ids)
 
+        # "ovn-controller" ends without deleting "Chassis" and
+        # "Chassis_Private" registers. If "Chassis" register is deleted,
+        # then Chassis_Private.chassis = []; both metadata and controller
+        # agents will still be present in the agent list.
+        agent_event = AgentWaitEvent(self.mech_driver, [self.chassis],
+                                     events=(event.RowEvent.ROW_UPDATE,))
+        self.sb_api.idl.notify_handler.watch_event(agent_event)
+        self.sb_api.chassis_del(self.chassis).execute(check_error=True)
+        self.assertTrue(agent_event.wait())
+        agent_ids = [a['id'] for a in self.plugin.get_agents(
+            self.context, filters={'host': self.host})]
+        self.assertCountEqual(list(self.agent_types.values()), agent_ids)
+
     def test_agent_delete(self):
-        for agent_id in self.agent_types.values():
-            self.plugin.delete_agent(self.context, agent_id)
-            self.assertRaises(agent_exc.AgentNotFound, self.plugin.get_agent,
-                              self.context, agent_id)
+        # Non OVN agent deletion.
+        agent_id = self.agent_types[self.TEST_AGENT]
+        self.plugin.delete_agent(self.context, agent_id)
+        self.assertRaises(agent_exc.AgentNotFound, self.plugin.get_agent,
+                          self.context, agent_id)
+
+        # OVN controller agent deletion, that triggers the "Chassis" register
+        # deletion. The "Chassis" register deletion triggers the host OVN
+        # agents deletion, both controller and metadata if present.
+        controller_id = self.agent_types[ovn_const.OVN_CONTROLLER_AGENT]
+        metadata_id = self.agent_types[ovn_const.OVN_METADATA_AGENT]
+        self.plugin.delete_agent(self.context, controller_id)
+        self.assertRaises(agent_exc.AgentNotFound, self.plugin.get_agent,
+                          self.context, controller_id)
+        self.assertEqual(
+            metadata_id,
+            self.plugin.get_agent(self.context, metadata_id)['id'])
+
+        self.plugin.delete_agent(self.context, metadata_id)
+        self.assertRaises(agent_exc.AgentNotFound, self.plugin.get_agent,
+                          self.context, metadata_id)
 
 
 class ConnectionInactivityProbeSetEvent(event.WaitEvent):