openstack
/
ironic
Code Issues Proposed changes

Fixes a race condition in the hash ring code 

The current hash ring code suffers from several problems:

1. The cache is reset on any get_topic_for call, which means that the
   cache is not used between API calls. Previously it was done to work
   around the situation when a new conductor is not visible until the API
   process restarts. Currently we refresh the hash rings periodically.

   This patch removes resetting the cache. To avoid waiting 2 minutes to
   be able to use a new driver:
   1) the hash ring cache is always rebuilt when a ring is not found,
   2) the hash_ring_reset_interval option was changed to 15 seconds.

2. The reset of the cache races with the hot path in the get_ring call.
   It is possible that the reset happens after the class-level cache
   variable is checked but before it is used, yielding None.

   This patch stores the value of the class-level variable to a local
   variable before checking, thus ensuring None is never returned.

Finally, some logging was added to the modified code to make this kind
of problems more debugable in the future.

Conflicts:
	ironic/common/hash_ring.py
	ironic/conductor/rpcapi.py
	ironic/tests/unit/common/test_hash_ring.py

Change-Id: I6e18c6ec23a053b59c76fcadd52b13d84d81b4fb
Story: #2003966
Task: #26896
Partial-Bug: #1792872
(cherry picked from commit 5471157e4f)
This commit is contained in:
Dmitry Tantsur 2018-10-04 12:51:52 +02:00
parent cb02c676a1
commit d8e09aec55
5 changed files with 78 additions and 29 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

33
ironic/common/hash_ring.py
View File

@ -16,6 +16,7 @@
import threading
import time
from oslo_log import log
from tooz import hashring
from ironic.common import exception
@ -24,6 +25,9 @@ from ironic.conf import CONF
from ironic.db import api as dbapi
LOG = log.getLogger(__name__)
class HashRingManager(object):
_hash_rings = None
_lock = threading.Lock()
@ -36,15 +40,20 @@ class HashRingManager(object):
def ring(self):
interval = CONF.hash_ring_reset_interval
limit = time.time() - interval
# Hot path, no lock
if self.__class__._hash_rings is not None and self.updated_at >= limit:
return self.__class__._hash_rings
# Hot path, no lock. Using a local variable to avoid races with code
# changing the class variable.
hash_rings = self.__class__._hash_rings
if hash_rings is not None and self.updated_at >= limit:
return hash_rings
with self._lock:
if self.__class__._hash_rings is None or self.updated_at < limit:
LOG.debug('Rebuilding cached hash rings')
rings = self._load_hash_rings()
self.__class__._hash_rings = rings
self.updated_at = time.time()
LOG.debug('Finished rebuilding hash rings, available drivers '
'are %s', ', '.join(rings))
return self.__class__._hash_rings
def _load_hash_rings(self):
@ -60,9 +69,27 @@ class HashRingManager(object):
@classmethod
def reset(cls):
with cls._lock:
LOG.debug('Resetting cached hash rings')
cls._hash_rings = None
def __getitem__(self, driver_name):
try:
return self._get_ring(driver_name)
except (exception.DriverNotFound, exception.TemporaryFailure):
# NOTE(dtantsur): we assume that this case is more often caused by
# conductors coming and leaving, so we try to rebuild the rings.
LOG.debug('No conductor found for driver %(driver)s, trying '
'to rebuild the hash rings',
{'driver': driver_name})
self.__class__.reset()
return self._get_ring(driver_name)
def _get_ring(self, driver_name):
# There are no conductors, temporary failure - 503 Service Unavailable
if not self.ring:
raise exception.TemporaryFailure()
try:
return self.ring[driver_name]
except KeyError:

14
ironic/conductor/rpcapi.py
View File

@ -126,12 +126,6 @@ class ConductorAPI(object):
:raises: NoValidHost
"""
self.ring_manager.reset()
# There are no conductors, temporary failure - 503 Service Unavailable
if not self.ring_manager.ring:
raise exception.TemporaryFailure()
try:
ring = self.ring_manager[node.driver]
dest = ring.get_nodes(node.uuid.encode('utf-8'),
@ -154,9 +148,13 @@ class ConductorAPI(object):
:raises: DriverNotFound
"""
self.ring_manager.reset()
try:
ring = self.ring_manager[driver_name]
except exception.TemporaryFailure:
# NOTE(dtantsur): even if no conductors are registered, it makes
# sense to report 404 on any driver request.
raise exception.DriverNotFound(_("No conductors registered."))
ring = self.ring_manager[driver_name]
host = random.choice(list(ring.nodes))
return self.topic + "." + host

6
ironic/conf/default.py
View File

@ -191,8 +191,10 @@ hash_opts = [
'and potentially allow the Ironic cluster to recover '
'more quickly if a conductor instance is terminated.')),
cfg.IntOpt('hash_ring_reset_interval',
default=180,
help=_('Interval (in seconds) between hash ring resets.')),
default=15,
help=_('Time (in seconds) after which the hash ring is '
'considered outdated and is refreshed on the next '
'access.')),
]
image_opts = [

41
ironic/tests/unit/common/test_hash_ring.py
View File

@ -59,25 +59,34 @@ class HashRingManagerTestCase(db_base.DbTestCase):
self.ring_manager.__getitem__,
'driver3')
def test_hash_ring_manager_no_refresh(self):
# If a new conductor is registered after the ring manager is
# initialized, it won't be seen. Long term this is probably
# undesirable, but today is the intended behavior.
self.assertRaises(exception.DriverNotFound,
def test_hash_ring_manager_automatic_retry(self):
self.assertRaises(exception.TemporaryFailure,
self.ring_manager.__getitem__,
'driver1')
self.register_conductors()
self.assertRaises(exception.DriverNotFound,
self.ring_manager.__getitem__,
'driver1')
self.assertIsNotNone(self.ring_manager['driver1'])
def test_hash_ring_manager_refresh(self):
def test_hash_ring_manager_reset_interval(self):
CONF.set_override('hash_ring_reset_interval', 30)
# Initialize the ring manager to make _hash_rings not None, then
# hash ring will refresh only when time interval exceeded.
self.assertRaises(exception.DriverNotFound,
self.ring_manager.__getitem__,
'driver1')
self.register_conductors()
# Just to simplify calculations
CONF.set_override('hash_partition_exponent', 0)
self.dbapi.register_conductor({
'hostname': 'host1',
'drivers': ['driver1', 'driver2'],
})
ring = self.ring_manager['driver1']
self.assertEqual(1, len(ring))
self.dbapi.register_conductor({
'hostname': 'host2',
'drivers': ['driver1'],
})
ring = self.ring_manager['driver1']
# The new conductor is not known yet. Automatic retry does not kick in,
# since there is an active conductor for the requested driver.
self.assertEqual(1, len(ring))
self.ring_manager.updated_at = time.time() - 31
self.ring_manager.__getitem__('driver1')
ring = self.ring_manager['driver1']
self.assertEqual(2, len(ring))

13
releasenotes/notes/hash-ring-race-da0d584de1f46788.yaml Normal file
View File

@ -0,0 +1,13 @@
---
fixes:
- |
Fixes a race condition in the hash ring implementation that could cause
an internal server error on any request. See `story 2003966
<https://storyboard.openstack.org/#!/story/2003966>`_ for details.
upgrade:
- |
The ``hash_ring_reset_interval`` configuration option was changed from 180
to 15 seconds. Previously, this option was essentially ignored on the API
side, becase the hash ring was reset on each API access. The lower value
minimizes the probability of a request routed to a wrong conductor when the
ring needs rebalancing.