Handle @safe_connect returns None side effect in _ensure_resource_provider

Change I0c4ca6a81f213277fe7219cb905a805712f81e36 added more error
handling to the _ensure_resource_provider flow but didn't account
for @safe_connect returning None when calling _create_resource_provider
in the case that nova-compute is started before placement is running.
If that happens, we fail with a TypeError during the nova-compute
startup because we put None in the resource provider cache and then
later blindly try to use it because the compute node resource provider
uuid is in the cache, but mapped to None.

This adds the None check back in _ensure_resource_provider and if
None is returned from _create_resource_provider we raise the same
exception that _create_resource_provider would raise if it couldn't
create the provider.

Change-Id: If9e1581db9c1ae14340b787d03c815d243d5a50c
Closes-Bug: #1767139
(cherry picked from commit 80a0019893)
This commit is contained in:
Matt Riedemann 2018-05-03 11:21:47 -04:00 committed by Oliver Walsh
parent f9ab466c6d
commit f95a10b26e
2 changed files with 40 additions and 0 deletions

View File

@ -662,6 +662,14 @@ class SchedulerReportClient(object):
created_rp = self._create_resource_provider(
context, uuid, name or uuid,
parent_provider_uuid=parent_provider_uuid)
# If @safe_connect can't establish a connection to the placement
# service, like if placement isn't running or nova-compute is
# mis-configured for authentication, we'll get None back and need
# to treat it like we couldn't create the provider (because we
# couldn't).
if created_rp is None:
raise exception.ResourceProviderCreationFailed(
name=name or uuid)
# Don't add the created_rp to rps_to_refresh. Since we just
# created it, it has no aggregates or traits.

View File

@ -1358,6 +1358,38 @@ class TestProviderOperations(SchedulerReportClientTestCase):
ValueError,
self.client._provider_tree.has_traits, uuids.compute_node, [])
@mock.patch('nova.scheduler.client.report.SchedulerReportClient.'
'_create_resource_provider', return_value=None)
@mock.patch('nova.scheduler.client.report.SchedulerReportClient.'
'_refresh_associations')
@mock.patch('nova.scheduler.client.report.SchedulerReportClient.'
'_get_providers_in_tree')
def test_ensure_resource_provider_create_no_placement(self, get_rpt_mock,
refresh_mock, create_rp_mock):
# No resource provider exists in the client's cache, and
# @safe_connect on _create_resource_provider returns None because
# Placement isn't running yet. Ensure we don't populate the resource
# provider cache.
get_rpt_mock.return_value = []
self.assertRaises(
exception.ResourceProviderCreationFailed,
self.client._ensure_resource_provider, self.context,
uuids.compute_node)
get_rpt_mock.assert_called_once_with(self.context, uuids.compute_node)
create_rp_mock.assert_called_once_with(
self.context, uuids.compute_node, uuids.compute_node,
parent_provider_uuid=None)
self.assertFalse(self.client._provider_tree.exists(uuids.compute_node))
refresh_mock.assert_not_called()
self.assertRaises(
ValueError,
self.client._provider_tree.in_aggregates, uuids.compute_node, [])
self.assertRaises(
ValueError,
self.client._provider_tree.has_traits, uuids.compute_node, [])
@mock.patch('nova.scheduler.client.report.SchedulerReportClient.'
'_create_resource_provider')
@mock.patch('nova.scheduler.client.report.SchedulerReportClient.'