-import json
+import time
+import signal
import logging
-from unittest import case, SkipTest
+import operator
+from random import randint, choice
-from cephfs_test_case import CephFSTestCase
+from tasks.cephfs.cephfs_test_case import CephFSTestCase
from teuthology.exceptions import CommandFailedError
-from teuthology import misc as teuthology
from tasks.cephfs.fuse_mount import FuseMount
log = logging.getLogger(__name__)
+class TestClusterAffinity(CephFSTestCase):
+ CLIENTS_REQUIRED = 0
+ MDSS_REQUIRED = 4
+
+ def _verify_join_fs(self, target, status=None):
+ if status is None:
+ status = self.fs.wait_for_daemons(timeout=30)
+ log.debug("%s", status)
+ target = sorted(target, key=operator.itemgetter('name'))
+ log.info("target = %s", target)
+ current = list(status.get_all())
+ current = sorted(current, key=operator.itemgetter('name'))
+ log.info("current = %s", current)
+ self.assertEqual(len(current), len(target))
+ for i in range(len(current)):
+ for attr in target[i]:
+ self.assertIn(attr, current[i])
+ self.assertEqual(target[i][attr], current[i][attr])
+
+ def _change_target_state(self, state, name, changes):
+ for entity in state:
+ if entity['name'] == name:
+ for k, v in changes.items():
+ entity[k] = v
+ return
+ self.fail("no entity")
+
+ def _verify_init(self):
+ status = self.fs.status()
+ log.info("status = {0}".format(status))
+ target = [{'join_fscid': -1, 'name': info['name']} for info in status.get_all()]
+ self._verify_join_fs(target, status=status)
+ return (status, target)
+
+ def _reach_target(self, target):
+ def takeover():
+ try:
+ self._verify_join_fs(target)
+ return True
+ except AssertionError as e:
+ log.debug("%s", e)
+ return False
+ self.wait_until_true(takeover, 30)
+
+ def test_join_fs_runtime(self):
+ """
+ That setting mds_join_fs at runtime affects the cluster layout.
+ """
+ status, target = self._verify_init()
+ standbys = list(status.get_standbys())
+ self.config_set('mds.'+standbys[0]['name'], 'mds_join_fs', 'cephfs')
+ self._change_target_state(target, standbys[0]['name'], {'join_fscid': self.fs.id, 'state': 'up:active'})
+ self._reach_target(target)
+
+ def test_join_fs_unset(self):
+ """
+ That unsetting mds_join_fs will cause failover if another high-affinity standby exists.
+ """
+ status, target = self._verify_init()
+ standbys = list(status.get_standbys())
+ names = (standbys[0]['name'], standbys[1]['name'])
+ self.config_set('mds.'+names[0], 'mds_join_fs', 'cephfs')
+ self.config_set('mds.'+names[1], 'mds_join_fs', 'cephfs')
+ self._change_target_state(target, names[0], {'join_fscid': self.fs.id})
+ self._change_target_state(target, names[1], {'join_fscid': self.fs.id})
+ self._reach_target(target)
+ status = self.fs.status()
+ active = self.fs.get_active_names(status=status)[0]
+ self.assertIn(active, names)
+ self.config_rm('mds.'+active, 'mds_join_fs')
+ self._change_target_state(target, active, {'join_fscid': -1})
+ new_active = (set(names) - set((active,))).pop()
+ self._change_target_state(target, new_active, {'state': 'up:active'})
+ self._reach_target(target)
+
+ def test_join_fs_drop(self):
+ """
+ That unsetting mds_join_fs will not cause failover if no high-affinity standby exists.
+ """
+ status, target = self._verify_init()
+ standbys = list(status.get_standbys())
+ active = standbys[0]['name']
+ self.config_set('mds.'+active, 'mds_join_fs', 'cephfs')
+ self._change_target_state(target, active, {'join_fscid': self.fs.id, 'state': 'up:active'})
+ self._reach_target(target)
+ self.config_rm('mds.'+active, 'mds_join_fs')
+ self._change_target_state(target, active, {'join_fscid': -1})
+ self._reach_target(target)
+
+ def test_join_fs_vanilla(self):
+ """
+ That a vanilla standby is preferred over others with mds_join_fs set to another fs.
+ """
+ fs2 = self.mds_cluster.newfs(name="cephfs2")
+ status, target = self._verify_init()
+ active = self.fs.get_active_names(status=status)[0]
+ standbys = [info['name'] for info in status.get_standbys()]
+ victim = standbys.pop()
+ # Set a bogus fs on the others
+ for mds in standbys:
+ self.config_set('mds.'+mds, 'mds_join_fs', 'cephfs2')
+ self._change_target_state(target, mds, {'join_fscid': fs2.id})
+ self.fs.rank_fail()
+ self._change_target_state(target, victim, {'state': 'up:active'})
+ self._reach_target(target)
+ status = self.fs.status()
+ active = self.fs.get_active_names(status=status)[0]
+ self.assertEqual(active, victim)
+
+ def test_join_fs_last_resort(self):
+ """
+ That a standby with mds_join_fs set to another fs is still used if necessary.
+ """
+ status, target = self._verify_init()
+ standbys = [info['name'] for info in status.get_standbys()]
+ for mds in standbys:
+ self.config_set('mds.'+mds, 'mds_join_fs', 'cephfs2')
+ fs2 = self.mds_cluster.newfs(name="cephfs2")
+ for mds in standbys:
+ self._change_target_state(target, mds, {'join_fscid': fs2.id})
+ self.fs.rank_fail()
+ status = self.fs.status()
+ ranks = list(self.fs.get_ranks(status=status))
+ self.assertEqual(len(ranks), 1)
+ self.assertIn(ranks[0]['name'], standbys)
+ # Note that we would expect the former active to reclaim its spot, but
+ # we're not testing that here.
+
+ def test_join_fs_steady(self):
+ """
+ That a sole MDS with mds_join_fs set will come back as active eventually even after failover.
+ """
+ status, target = self._verify_init()
+ active = self.fs.get_active_names(status=status)[0]
+ self.config_set('mds.'+active, 'mds_join_fs', 'cephfs')
+ self._change_target_state(target, active, {'join_fscid': self.fs.id})
+ self._reach_target(target)
+ self.fs.rank_fail()
+ self._reach_target(target)
+
+ def test_join_fs_standby_replay(self):
+ """
+ That a standby-replay daemon with weak affinity is replaced by a stronger one.
+ """
+ status, target = self._verify_init()
+ standbys = [info['name'] for info in status.get_standbys()]
+ self.config_set('mds.'+standbys[0], 'mds_join_fs', 'cephfs')
+ self._change_target_state(target, standbys[0], {'join_fscid': self.fs.id, 'state': 'up:active'})
+ self._reach_target(target)
+ self.fs.set_allow_standby_replay(True)
+ status = self.fs.status()
+ standbys = [info['name'] for info in status.get_standbys()]
+ self.config_set('mds.'+standbys[0], 'mds_join_fs', 'cephfs')
+ self._change_target_state(target, standbys[0], {'join_fscid': self.fs.id, 'state': 'up:standby-replay'})
+ self._reach_target(target)
+
+class TestClusterResize(CephFSTestCase):
+ CLIENTS_REQUIRED = 0
+ MDSS_REQUIRED = 3
+
+ def test_grow(self):
+ """
+ That the MDS cluster grows after increasing max_mds.
+ """
+
+ # Need all my standbys up as well as the active daemons
+ # self.wait_for_daemon_start() necessary?
+
+ self.fs.grow(2)
+ self.fs.grow(3)
+
+
+ def test_shrink(self):
+ """
+ That the MDS cluster shrinks automatically after decreasing max_mds.
+ """
+
+ self.fs.grow(3)
+ self.fs.shrink(1)
+
+ def test_up_less_than_max(self):
+ """
+ That a health warning is generated when max_mds is greater than active count.
+ """
+
+ status = self.fs.status()
+ mdss = [info['gid'] for info in status.get_all()]
+ self.fs.set_max_mds(len(mdss)+1)
+ self.wait_for_health("MDS_UP_LESS_THAN_MAX", 30)
+ self.fs.shrink(2)
+ self.wait_for_health_clear(30)
+
+ def test_down_health(self):
+ """
+ That marking a FS down does not generate a health warning
+ """
+
+ self.fs.set_down()
+ try:
+ self.wait_for_health("", 30)
+ raise RuntimeError("got health warning?")
+ except RuntimeError as e:
+ if "Timed out after" in str(e):
+ pass
+ else:
+ raise
+
+ def test_down_twice(self):
+ """
+ That marking a FS down twice does not wipe old_max_mds.
+ """
+
+ self.fs.grow(2)
+ self.fs.set_down()
+ self.fs.wait_for_daemons()
+ self.fs.set_down(False)
+ self.assertEqual(self.fs.get_var("max_mds"), 2)
+ self.fs.wait_for_daemons(timeout=60)
+
+ def test_down_grow(self):
+ """
+ That setting max_mds undoes down.
+ """
+
+ self.fs.set_down()
+ self.fs.wait_for_daemons()
+ self.fs.grow(2)
+ self.fs.wait_for_daemons()
+
+ def test_down(self):
+ """
+ That down setting toggles and sets max_mds appropriately.
+ """
+
+ self.fs.set_down()
+ self.fs.wait_for_daemons()
+ self.assertEqual(self.fs.get_var("max_mds"), 0)
+ self.fs.set_down(False)
+ self.assertEqual(self.fs.get_var("max_mds"), 1)
+ self.fs.wait_for_daemons()
+ self.assertEqual(self.fs.get_var("max_mds"), 1)
+
+ def test_hole(self):
+ """
+ Test that a hole cannot be created in the FS ranks.
+ """
+
+ fscid = self.fs.id
+
+ self.fs.grow(2)
+
+ # Now add a delay which should slow down how quickly rank 1 stops
+ self.config_set('mds', 'ms_inject_delay_max', '5.0')
+ self.config_set('mds', 'ms_inject_delay_probability', '1.0')
+ self.fs.set_max_mds(1)
+ log.info("status = {0}".format(self.fs.status()))
+
+ # Don't wait for rank 1 to stop
+ self.fs.set_max_mds(3)
+ log.info("status = {0}".format(self.fs.status()))
+
+ # Now check that the mons didn't try to promote a standby to rank 2
+ self.fs.set_max_mds(2)
+ status = self.fs.status()
+ try:
+ status = self.fs.wait_for_daemons(timeout=90)
+ ranks = set([info['rank'] for info in status.get_ranks(fscid)])
+ self.assertEqual(ranks, set([0, 1]))
+ finally:
+ log.info("status = {0}".format(status))
+
+ def test_thrash(self):
+ """
+ Test that thrashing max_mds does not fail.
+ """
+
+ max_mds = 2
+ for i in range(0, 100):
+ self.fs.set_max_mds(max_mds)
+ max_mds = (max_mds+1)%3+1
+
+ self.fs.wait_for_daemons(timeout=90)
class TestFailover(CephFSTestCase):
CLIENTS_REQUIRED = 1
MDSS_REQUIRED = 2
+ def test_repeated_boot(self):
+ """
+ That multiple boot messages do not result in the MDS getting evicted.
+ """
+
+ interval = 10
+ self.config_set("mon", "paxos_propose_interval", interval)
+
+ mds = choice(list(self.fs.status().get_all()))
+
+ with self.assert_cluster_log(f"daemon mds.{mds['name']} restarted", present=False):
+ # Avoid a beacon to the monitors with down:dne by restarting:
+ self.fs.mds_fail(mds_id=mds['name'])
+ # `ceph mds fail` won't return until the FSMap is committed, double-check:
+ self.assertIsNone(self.fs.status().get_mds_gid(mds['gid']))
+ time.sleep(2) # for mds to restart and accept asok commands
+ status1 = self.fs.mds_asok(['status'], mds_id=mds['name'])
+ time.sleep(interval*1.5)
+ status2 = self.fs.mds_asok(['status'], mds_id=mds['name'])
+ self.assertEqual(status1['id'], status2['id'])
+
def test_simple(self):
"""
That when the active MDS is killed, a standby MDS is promoted into
in thrashing tests.
"""
- # Need all my standbys up as well as the active daemons
- self.wait_for_daemon_start()
-
(original_active, ) = self.fs.get_active_names()
original_standbys = self.mds_cluster.get_standby_daemons()
# Kill the rank 0 daemon's physical process
self.fs.mds_stop(original_active)
- grace = int(self.fs.get_config("mds_beacon_grace", service_type="mon"))
-
# Wait until the monitor promotes his replacement
def promoted():
- active = self.fs.get_active_names()
- return active and active[0] in original_standbys
+ ranks = list(self.fs.get_ranks())
+ return len(ranks) > 0 and ranks[0]['name'] in original_standbys
log.info("Waiting for promotion of one of the original standbys {0}".format(
original_standbys))
- self.wait_until_true(
- promoted,
- timeout=grace*2)
+ self.wait_until_true(promoted, timeout=self.fs.beacon_timeout)
# Start the original rank 0 daemon up again, see that he becomes a standby
self.fs.mds_restart(original_active)
"""
if not isinstance(self.mount_a, FuseMount):
- raise SkipTest("Requires FUSE client to inject client metadata")
+ self.skipTest("Requires FUSE client to inject client metadata")
require_active = self.fs.get_config("fuse_require_active_mds", service_type="mon").lower() == "true"
if not require_active:
- raise case.SkipTest("fuse_require_active_mds is not set")
-
- grace = int(self.fs.get_config("mds_beacon_grace", service_type="mon"))
+ self.skipTest("fuse_require_active_mds is not set")
# Check it's not laggy to begin with
(original_active, ) = self.fs.get_active_names()
- self.assertNotIn("laggy_since", self.fs.mon_manager.get_mds_status(original_active))
+ self.assertNotIn("laggy_since", self.fs.status().get_mds(original_active))
self.mounts[0].umount_wait()
# Control: that we can mount and unmount usually, while the cluster is healthy
- self.mounts[0].mount()
- self.mounts[0].wait_until_mounted()
+ self.mounts[0].mount_wait()
self.mounts[0].umount_wait()
# Stop the daemon processes
return True
- self.wait_until_true(laggy, grace * 2)
+ self.wait_until_true(laggy, self.fs.beacon_timeout)
with self.assertRaises(CommandFailedError):
- self.mounts[0].mount()
+ self.mounts[0].mount_wait()
def test_standby_count_wanted(self):
"""
# Need all my standbys up as well as the active daemons
self.wait_for_daemon_start()
- grace = int(self.fs.get_config("mds_beacon_grace", service_type="mon"))
-
standbys = self.mds_cluster.get_standby_daemons()
self.assertGreaterEqual(len(standbys), 1)
self.fs.mon_manager.raw_cluster_cmd('fs', 'set', self.fs.name, 'standby_count_wanted', str(len(standbys)))
# Kill a standby and check for warning
victim = standbys.pop()
self.fs.mds_stop(victim)
- log.info("waiting for insufficient standby daemon warning")
- self.wait_for_health("MDS_INSUFFICIENT_STANDBY", grace*2)
+ self.wait_for_health("MDS_INSUFFICIENT_STANDBY", self.fs.beacon_timeout)
# restart the standby, see that he becomes a standby, check health clears
self.fs.mds_restart(victim)
standbys = self.mds_cluster.get_standby_daemons()
self.assertGreaterEqual(len(standbys), 1)
self.fs.mon_manager.raw_cluster_cmd('fs', 'set', self.fs.name, 'standby_count_wanted', str(len(standbys)+1))
- log.info("waiting for insufficient standby daemon warning")
- self.wait_for_health("MDS_INSUFFICIENT_STANDBY", grace*2)
+ self.wait_for_health("MDS_INSUFFICIENT_STANDBY", self.fs.beacon_timeout)
# Set it to 0
self.fs.mon_manager.raw_cluster_cmd('fs', 'set', self.fs.name, 'standby_count_wanted', '0')
self.wait_for_health_clear(timeout=30)
+ def test_discontinuous_mdsmap(self):
+ """
+ That discontinuous mdsmap does not affect failover.
+ See http://tracker.ceph.com/issues/24856.
+ """
+ self.fs.set_max_mds(2)
+ status = self.fs.wait_for_daemons()
+
+ self.mount_a.umount_wait()
+
+ monc_timeout = float(self.fs.get_config("mon_client_ping_timeout", service_type="mds"))
+
+ mds_0 = self.fs.get_rank(rank=0, status=status)
+ self.fs.rank_freeze(True, rank=0) # prevent failover
+ self.fs.rank_signal(signal.SIGSTOP, rank=0, status=status)
+ self.wait_until_true(
+ lambda: "laggy_since" in self.fs.get_rank(),
+ timeout=self.fs.beacon_timeout
+ )
+
+ self.fs.rank_fail(rank=1)
+ self.fs.wait_for_state('up:resolve', rank=1, timeout=30)
+
+ # Make sure of mds_0's monitor connection gets reset
+ time.sleep(monc_timeout * 2)
+
+ # Continue rank 0, it will get discontinuous mdsmap
+ self.fs.rank_signal(signal.SIGCONT, rank=0)
+ self.wait_until_true(
+ lambda: "laggy_since" not in self.fs.get_rank(rank=0),
+ timeout=self.fs.beacon_timeout
+ )
+
+ # mds.b will be stuck at 'reconnect' state if snapserver gets confused
+ # by discontinuous mdsmap
+ self.fs.wait_for_state('up:active', rank=1, timeout=30)
+ self.assertEqual(mds_0['gid'], self.fs.get_rank(rank=0)['gid'])
+ self.fs.rank_freeze(False, rank=0)
+ def test_connect_bootstrapping(self):
+ self.config_set("mds", "mds_sleep_rank_change", 10000000.0)
+ self.config_set("mds", "mds_connect_bootstrapping", True)
+ self.fs.set_max_mds(2)
+ self.fs.wait_for_daemons()
+ self.fs.rank_fail(rank=0)
+ # rank 0 will get stuck in up:resolve, see https://tracker.ceph.com/issues/53194
+ self.fs.wait_for_daemons()
class TestStandbyReplay(CephFSTestCase):
+ CLIENTS_REQUIRED = 0
MDSS_REQUIRED = 4
- REQUIRE_FILESYSTEM = False
- def set_standby_for(self, leader, follower, replay):
- self.set_conf("mds.{0}".format(follower), "mds_standby_for_name", leader)
+ def _confirm_no_replay(self):
+ status = self.fs.status()
+ _ = len(list(status.get_standbys()))
+ self.assertEqual(0, len(list(self.fs.get_replays(status=status))))
+ return status
+
+ def _confirm_single_replay(self, full=True, status=None, retries=3):
+ status = self.fs.wait_for_daemons(status=status)
+ ranks = sorted(self.fs.get_mds_map(status=status)['in'])
+ replays = list(self.fs.get_replays(status=status))
+ checked_replays = set()
+ for rank in ranks:
+ has_replay = False
+ for replay in replays:
+ if replay['rank'] == rank:
+ self.assertFalse(has_replay)
+ has_replay = True
+ checked_replays.add(replay['gid'])
+ if full and not has_replay:
+ if retries <= 0:
+ raise RuntimeError("rank "+str(rank)+" has no standby-replay follower")
+ else:
+ retries = retries-1
+ time.sleep(2)
+ self.assertEqual(checked_replays, set(info['gid'] for info in replays))
+ return status
+
+ def _check_replay_takeover(self, status, rank=0):
+ replay = self.fs.get_replay(rank=rank, status=status)
+ new_status = self.fs.wait_for_daemons()
+ new_active = self.fs.get_rank(rank=rank, status=new_status)
if replay:
- self.set_conf("mds.{0}".format(follower), "mds_standby_replay", "true")
-
- def get_info_by_name(self, mds_name):
- status = self.mds_cluster.status()
- info = status.get_mds(mds_name)
- if info is None:
- log.warn(str(status))
- raise RuntimeError("MDS '{0}' not found".format(mds_name))
+ self.assertEqual(replay['gid'], new_active['gid'])
else:
- return info
-
- def test_standby_replay_unused(self):
- # Pick out exactly 3 daemons to be run during test
- use_daemons = sorted(self.mds_cluster.mds_ids[0:3])
- mds_a, mds_b, mds_c = use_daemons
- log.info("Using MDS daemons: {0}".format(use_daemons))
-
- # B and C should both follow A, but only one will
- # really get into standby replay state.
- self.set_standby_for(mds_a, mds_b, True)
- self.set_standby_for(mds_a, mds_c, True)
-
- # Create FS and start A
- fs_a = self.mds_cluster.newfs("alpha")
- self.mds_cluster.mds_restart(mds_a)
- fs_a.wait_for_daemons()
- self.assertEqual(fs_a.get_active_names(), [mds_a])
-
- # Start B, he should go into standby replay
- self.mds_cluster.mds_restart(mds_b)
- self.wait_for_daemon_start([mds_b])
- info_b = self.get_info_by_name(mds_b)
- self.assertEqual(info_b['state'], "up:standby-replay")
- self.assertEqual(info_b['standby_for_name'], mds_a)
- self.assertEqual(info_b['rank'], 0)
-
- # Start C, he should go into standby (*not* replay)
- self.mds_cluster.mds_restart(mds_c)
- self.wait_for_daemon_start([mds_c])
- info_c = self.get_info_by_name(mds_c)
- self.assertEqual(info_c['state'], "up:standby")
- self.assertEqual(info_c['standby_for_name'], mds_a)
- self.assertEqual(info_c['rank'], -1)
-
- # Kill B, C should go into standby replay
- self.mds_cluster.mds_stop(mds_b)
- self.mds_cluster.mds_fail(mds_b)
- self.wait_until_equal(
- lambda: self.get_info_by_name(mds_c)['state'],
- "up:standby-replay",
- 60)
- info_c = self.get_info_by_name(mds_c)
- self.assertEqual(info_c['state'], "up:standby-replay")
- self.assertEqual(info_c['standby_for_name'], mds_a)
- self.assertEqual(info_c['rank'], 0)
-
- def test_standby_failure(self):
+ # double check takeover came from a standby (or some new daemon via restart)
+ found = False
+ for info in status.get_standbys():
+ if info['gid'] == new_active['gid']:
+ found = True
+ break
+ if not found:
+ for info in status.get_all():
+ self.assertNotEqual(info['gid'], new_active['gid'])
+ return new_status
+
+ def test_standby_replay_singleton(self):
+ """
+ That only one MDS becomes standby-replay.
+ """
+
+ self._confirm_no_replay()
+ self.fs.set_allow_standby_replay(True)
+ time.sleep(30)
+ self._confirm_single_replay()
+
+ def test_standby_replay_damaged(self):
+ """
+ That a standby-replay daemon can cause the rank to go damaged correctly.
+ """
+
+ self._confirm_no_replay()
+ self.config_set("mds", "mds_standby_replay_damaged", True)
+ self.fs.set_allow_standby_replay(True)
+ self.wait_until_true(
+ lambda: len(self.fs.get_damaged()) > 0,
+ timeout=30
+ )
+ status = self.fs.status()
+ self.assertListEqual([], list(self.fs.get_ranks(status=status)))
+ self.assertListEqual([0], self.fs.get_damaged(status=status))
+
+ def test_standby_replay_disable(self):
+ """
+ That turning off allow_standby_replay fails all standby-replay daemons.
+ """
+
+ self._confirm_no_replay()
+ self.fs.set_allow_standby_replay(True)
+ time.sleep(30)
+ self._confirm_single_replay()
+ self.fs.set_allow_standby_replay(False)
+ self._confirm_no_replay()
+
+ def test_standby_replay_singleton_fail(self):
+ """
+ That failures don't violate singleton constraint.
+ """
+
+ self._confirm_no_replay()
+ self.fs.set_allow_standby_replay(True)
+ status = self._confirm_single_replay()
+
+ for i in range(10):
+ time.sleep(randint(1, 5))
+ self.fs.rank_restart(status=status)
+ status = self._check_replay_takeover(status)
+ status = self._confirm_single_replay(status=status)
+
+ for i in range(10):
+ time.sleep(randint(1, 5))
+ self.fs.rank_fail()
+ status = self._check_replay_takeover(status)
+ status = self._confirm_single_replay(status=status)
+
+ def test_standby_replay_singleton_fail_multimds(self):
+ """
+ That failures don't violate singleton constraint with multiple actives.
+ """
+
+ status = self._confirm_no_replay()
+ new_max_mds = randint(2, len(list(status.get_standbys())))
+ self.fs.set_max_mds(new_max_mds)
+ self.fs.wait_for_daemons() # wait for actives to come online!
+ self.fs.set_allow_standby_replay(True)
+ status = self._confirm_single_replay(full=False)
+
+ for i in range(10):
+ time.sleep(randint(1, 5))
+ victim = randint(0, new_max_mds-1)
+ self.fs.rank_restart(rank=victim, status=status)
+ status = self._check_replay_takeover(status, rank=victim)
+ status = self._confirm_single_replay(status=status, full=False)
+
+ for i in range(10):
+ time.sleep(randint(1, 5))
+ victim = randint(0, new_max_mds-1)
+ self.fs.rank_fail(rank=victim)
+ status = self._check_replay_takeover(status, rank=victim)
+ status = self._confirm_single_replay(status=status, full=False)
+
+ def test_standby_replay_failure(self):
"""
That the failure of a standby-replay daemon happens cleanly
and doesn't interrupt anything else.
"""
- # Pick out exactly 2 daemons to be run during test
- use_daemons = sorted(self.mds_cluster.mds_ids[0:2])
- mds_a, mds_b = use_daemons
- log.info("Using MDS daemons: {0}".format(use_daemons))
- # Configure two pairs of MDSs that are standby for each other
- self.set_standby_for(mds_a, mds_b, True)
- self.set_standby_for(mds_b, mds_a, False)
+ status = self._confirm_no_replay()
+ self.fs.set_max_mds(1)
+ self.fs.set_allow_standby_replay(True)
+ status = self._confirm_single_replay()
- # Create FS alpha and get mds_a to come up as active
- fs_a = self.mds_cluster.newfs("alpha")
- self.mds_cluster.mds_restart(mds_a)
- fs_a.wait_for_daemons()
- self.assertEqual(fs_a.get_active_names(), [mds_a])
-
- # Start the standbys
- self.mds_cluster.mds_restart(mds_b)
- self.wait_for_daemon_start([mds_b])
-
- # See the standby come up as the correct rank
- info_b = self.get_info_by_name(mds_b)
- self.assertEqual(info_b['state'], "up:standby-replay")
- self.assertEqual(info_b['standby_for_name'], mds_a)
- self.assertEqual(info_b['rank'], 0)
-
- # Kill the standby
- self.mds_cluster.mds_stop(mds_b)
- self.mds_cluster.mds_fail(mds_b)
-
- # See that the standby is gone and the active remains
- self.assertEqual(fs_a.get_active_names(), [mds_a])
- mds_map = fs_a.get_mds_map()
- self.assertEqual(len(mds_map['info']), 1)
- self.assertEqual(mds_map['failed'], [])
- self.assertEqual(mds_map['damaged'], [])
- self.assertEqual(mds_map['stopped'], [])
+ for i in range(10):
+ time.sleep(randint(1, 5))
+ victim = self.fs.get_replay(status=status)
+ self.fs.mds_restart(mds_id=victim['name'])
+ status = self._confirm_single_replay(status=status)
+
+ def test_standby_replay_prepare_beacon(self):
+ """
+ That a MDSMonitor::prepare_beacon handles standby-replay daemons
+ correctly without removing the standby. (Note, usually a standby-replay
+ beacon will just be replied to by MDSMonitor::preprocess_beacon.)
+ """
+
+ status = self._confirm_no_replay()
+ self.fs.set_max_mds(1)
+ self.fs.set_allow_standby_replay(True)
+ status = self._confirm_single_replay()
+ replays = list(status.get_replays(self.fs.id))
+ self.assertEqual(len(replays), 1)
+ self.config_set('mds.'+replays[0]['name'], 'mds_inject_health_dummy', True)
+ time.sleep(10) # for something not to happen...
+ status = self._confirm_single_replay()
+ replays2 = list(status.get_replays(self.fs.id))
+ self.assertEqual(replays[0]['gid'], replays2[0]['gid'])
def test_rank_stopped(self):
"""
That when a rank is STOPPED, standby replays for
that rank get torn down
"""
- # Pick out exactly 2 daemons to be run during test
- use_daemons = sorted(self.mds_cluster.mds_ids[0:4])
- mds_a, mds_b, mds_a_s, mds_b_s = use_daemons
- log.info("Using MDS daemons: {0}".format(use_daemons))
-
- # a and b both get a standby
- self.set_standby_for(mds_a, mds_a_s, True)
- self.set_standby_for(mds_b, mds_b_s, True)
- # Create FS alpha and get mds_a to come up as active
- fs_a = self.mds_cluster.newfs("alpha")
- fs_a.set_max_mds(2)
+ status = self._confirm_no_replay()
+ standby_count = len(list(status.get_standbys()))
+ self.fs.set_max_mds(2)
+ self.fs.set_allow_standby_replay(True)
+ status = self._confirm_single_replay()
- self.mds_cluster.mds_restart(mds_a)
- self.wait_until_equal(lambda: fs_a.get_active_names(), [mds_a], 30)
- self.mds_cluster.mds_restart(mds_b)
- fs_a.wait_for_daemons()
- self.assertEqual(sorted(fs_a.get_active_names()), [mds_a, mds_b])
-
- # Start the standbys
- self.mds_cluster.mds_restart(mds_b_s)
- self.wait_for_daemon_start([mds_b_s])
- self.mds_cluster.mds_restart(mds_a_s)
- self.wait_for_daemon_start([mds_a_s])
- info_b_s = self.get_info_by_name(mds_b_s)
- self.assertEqual(info_b_s['state'], "up:standby-replay")
- info_a_s = self.get_info_by_name(mds_a_s)
- self.assertEqual(info_a_s['state'], "up:standby-replay")
-
- # Shrink the cluster
- fs_a.set_max_mds(1)
- fs_a.mon_manager.raw_cluster_cmd("mds", "stop", "{0}:1".format(fs_a.name))
- self.wait_until_equal(
- lambda: fs_a.get_active_names(), [mds_a],
- 60
- )
+ self.fs.set_max_mds(1) # stop rank 1
- # Both 'b' and 'b_s' should go back to being standbys
- self.wait_until_equal(
- lambda: self.mds_cluster.get_standby_daemons(), {mds_b, mds_b_s},
- 60
- )
+ status = self._confirm_single_replay()
+ self.assertTrue(standby_count, len(list(status.get_standbys())))
class TestMultiFilesystems(CephFSTestCase):
"--yes-i-really-mean-it")
def _setup_two(self):
- fs_a = self.mds_cluster.newfs("alpha")
- fs_b = self.mds_cluster.newfs("bravo")
+ fs_a = self.mds_cluster.newfs(name="alpha")
+ fs_b = self.mds_cluster.newfs(name="bravo")
self.mds_cluster.mds_restart()
fs_a, fs_b = self._setup_two()
# Mount a client on fs_a
- self.mount_a.mount(mount_fs_name=fs_a.name)
+ self.mount_a.mount_wait(cephfs_name=fs_a.name)
self.mount_a.write_n_mb("pad.bin", 1)
self.mount_a.write_n_mb("test.bin", 2)
a_created_ino = self.mount_a.path_to_ino("test.bin")
self.mount_a.create_files()
# Mount a client on fs_b
- self.mount_b.mount(mount_fs_name=fs_b.name)
+ self.mount_b.mount_wait(cephfs_name=fs_b.name)
self.mount_b.write_n_mb("test.bin", 1)
b_created_ino = self.mount_b.path_to_ino("test.bin")
self.mount_b.create_files()
# Shrink fs_b back to 1, see a daemon go back to standby
fs_b.set_max_mds(1)
- fs_b.deactivate(1)
self.wait_until_equal(lambda: len(fs_b.get_active_names()), 1, 30,
reject_fn=lambda v: v > 2 or v < 1)
fs_a.set_max_mds(3)
self.wait_until_equal(lambda: len(fs_a.get_active_names()), 3, 60,
reject_fn=lambda v: v > 3 or v < 2)
-
- def test_standby_for_name(self):
- # Pick out exactly 4 daemons to be run during test
- use_daemons = sorted(self.mds_cluster.mds_ids[0:4])
- mds_a, mds_b, mds_c, mds_d = use_daemons
- log.info("Using MDS daemons: {0}".format(use_daemons))
-
- def set_standby_for(leader, follower, replay):
- self.set_conf("mds.{0}".format(follower), "mds_standby_for_name", leader)
- if replay:
- self.set_conf("mds.{0}".format(follower), "mds_standby_replay", "true")
-
- # Configure two pairs of MDSs that are standby for each other
- set_standby_for(mds_a, mds_b, True)
- set_standby_for(mds_b, mds_a, False)
- set_standby_for(mds_c, mds_d, True)
- set_standby_for(mds_d, mds_c, False)
-
- # Create FS alpha and get mds_a to come up as active
- fs_a = self.mds_cluster.newfs("alpha")
- self.mds_cluster.mds_restart(mds_a)
- fs_a.wait_for_daemons()
- self.assertEqual(fs_a.get_active_names(), [mds_a])
-
- # Create FS bravo and get mds_c to come up as active
- fs_b = self.mds_cluster.newfs("bravo")
- self.mds_cluster.mds_restart(mds_c)
- fs_b.wait_for_daemons()
- self.assertEqual(fs_b.get_active_names(), [mds_c])
-
- # Start the standbys
- self.mds_cluster.mds_restart(mds_b)
- self.mds_cluster.mds_restart(mds_d)
- self.wait_for_daemon_start([mds_b, mds_d])
-
- def get_info_by_name(fs, mds_name):
- mds_map = fs.get_mds_map()
- for gid_str, info in mds_map['info'].items():
- if info['name'] == mds_name:
- return info
-
- log.warn(json.dumps(mds_map, indent=2))
- raise RuntimeError("MDS '{0}' not found in filesystem MDSMap".format(mds_name))
-
- # See both standbys come up as standby replay for the correct ranks
- # mds_b should be in filesystem alpha following mds_a
- info_b = get_info_by_name(fs_a, mds_b)
- self.assertEqual(info_b['state'], "up:standby-replay")
- self.assertEqual(info_b['standby_for_name'], mds_a)
- self.assertEqual(info_b['rank'], 0)
- # mds_d should be in filesystem alpha following mds_c
- info_d = get_info_by_name(fs_b, mds_d)
- self.assertEqual(info_d['state'], "up:standby-replay")
- self.assertEqual(info_d['standby_for_name'], mds_c)
- self.assertEqual(info_d['rank'], 0)
-
- # Kill both active daemons
- self.mds_cluster.mds_stop(mds_a)
- self.mds_cluster.mds_fail(mds_a)
- self.mds_cluster.mds_stop(mds_c)
- self.mds_cluster.mds_fail(mds_c)
-
- # Wait for standbys to take over
- fs_a.wait_for_daemons()
- self.assertEqual(fs_a.get_active_names(), [mds_b])
- fs_b.wait_for_daemons()
- self.assertEqual(fs_b.get_active_names(), [mds_d])
-
- # Start the original active daemons up again
- self.mds_cluster.mds_restart(mds_a)
- self.mds_cluster.mds_restart(mds_c)
- self.wait_for_daemon_start([mds_a, mds_c])
-
- self.assertEqual(set(self.mds_cluster.get_standby_daemons()),
- {mds_a, mds_c})
-
- def test_standby_for_rank(self):
- use_daemons = sorted(self.mds_cluster.mds_ids[0:4])
- mds_a, mds_b, mds_c, mds_d = use_daemons
- log.info("Using MDS daemons: {0}".format(use_daemons))
-
- def set_standby_for(leader_rank, leader_fs, follower_id):
- self.set_conf("mds.{0}".format(follower_id),
- "mds_standby_for_rank", leader_rank)
-
- fscid = leader_fs.get_namespace_id()
- self.set_conf("mds.{0}".format(follower_id),
- "mds_standby_for_fscid", fscid)
-
- fs_a = self.mds_cluster.newfs("alpha")
- fs_b = self.mds_cluster.newfs("bravo")
- set_standby_for(0, fs_a, mds_a)
- set_standby_for(0, fs_a, mds_b)
- set_standby_for(0, fs_b, mds_c)
- set_standby_for(0, fs_b, mds_d)
-
- self.mds_cluster.mds_restart(mds_a)
- fs_a.wait_for_daemons()
- self.assertEqual(fs_a.get_active_names(), [mds_a])
-
- self.mds_cluster.mds_restart(mds_c)
- fs_b.wait_for_daemons()
- self.assertEqual(fs_b.get_active_names(), [mds_c])
-
- self.mds_cluster.mds_restart(mds_b)
- self.mds_cluster.mds_restart(mds_d)
- self.wait_for_daemon_start([mds_b, mds_d])
-
- self.mds_cluster.mds_stop(mds_a)
- self.mds_cluster.mds_fail(mds_a)
- self.mds_cluster.mds_stop(mds_c)
- self.mds_cluster.mds_fail(mds_c)
-
- fs_a.wait_for_daemons()
- self.assertEqual(fs_a.get_active_names(), [mds_b])
- fs_b.wait_for_daemons()
- self.assertEqual(fs_b.get_active_names(), [mds_d])
-
- def test_standby_for_fscid(self):
- """
- That I can set a standby FSCID with no rank, and the result is
- that daemons join any rank for that filesystem.
- """
- use_daemons = sorted(self.mds_cluster.mds_ids[0:4])
- mds_a, mds_b, mds_c, mds_d = use_daemons
-
- log.info("Using MDS daemons: {0}".format(use_daemons))
-
- def set_standby_for(leader_fs, follower_id):
- fscid = leader_fs.get_namespace_id()
- self.set_conf("mds.{0}".format(follower_id),
- "mds_standby_for_fscid", fscid)
-
- # Create two filesystems which should have two ranks each
- fs_a = self.mds_cluster.newfs("alpha")
-
- fs_b = self.mds_cluster.newfs("bravo")
-
- fs_a.set_max_mds(2)
- fs_b.set_max_mds(2)
-
- # Set all the daemons to have a FSCID assignment but no other
- # standby preferences.
- set_standby_for(fs_a, mds_a)
- set_standby_for(fs_a, mds_b)
- set_standby_for(fs_b, mds_c)
- set_standby_for(fs_b, mds_d)
-
- # Now when we start all daemons at once, they should fall into
- # ranks in the right filesystem
- self.mds_cluster.mds_restart(mds_a)
- self.mds_cluster.mds_restart(mds_b)
- self.mds_cluster.mds_restart(mds_c)
- self.mds_cluster.mds_restart(mds_d)
- self.wait_for_daemon_start([mds_a, mds_b, mds_c, mds_d])
- fs_a.wait_for_daemons()
- fs_b.wait_for_daemons()
- self.assertEqual(set(fs_a.get_active_names()), {mds_a, mds_b})
- self.assertEqual(set(fs_b.get_active_names()), {mds_c, mds_d})
-
- def test_standby_for_invalid_fscid(self):
- """
- That an invalid standby_fscid does not cause a mon crash
- """
- use_daemons = sorted(self.mds_cluster.mds_ids[0:3])
- mds_a, mds_b, mds_c = use_daemons
- log.info("Using MDS daemons: {0}".format(use_daemons))
-
- def set_standby_for_rank(leader_rank, follower_id):
- self.set_conf("mds.{0}".format(follower_id),
- "mds_standby_for_rank", leader_rank)
-
- # Create one fs
- fs_a = self.mds_cluster.newfs("cephfs")
-
- # Get configured mons in the cluster, so we can see if any
- # crashed later.
- configured_mons = fs_a.mon_manager.get_mon_quorum()
-
- # Set all the daemons to have a rank assignment but no other
- # standby preferences.
- set_standby_for_rank(0, mds_a)
- set_standby_for_rank(0, mds_b)
-
- # Set third daemon to have invalid fscid assignment and no other
- # standby preferences
- invalid_fscid = 123
- self.set_conf("mds.{0}".format(mds_c), "mds_standby_for_fscid", invalid_fscid)
-
- #Restart all the daemons to make the standby preference applied
- self.mds_cluster.mds_restart(mds_a)
- self.mds_cluster.mds_restart(mds_b)
- self.mds_cluster.mds_restart(mds_c)
- self.wait_for_daemon_start([mds_a, mds_b, mds_c])
-
- #Stop active mds daemon service of fs
- if (fs_a.get_active_names(), [mds_a]):
- self.mds_cluster.mds_stop(mds_a)
- self.mds_cluster.mds_fail(mds_a)
- fs_a.wait_for_daemons()
- else:
- self.mds_cluster.mds_stop(mds_b)
- self.mds_cluster.mds_fail(mds_b)
- fs_a.wait_for_daemons()
-
- #Get active mons from cluster
- active_mons = fs_a.mon_manager.get_mon_quorum()
-
- #Check for active quorum mon status and configured mon status
- self.assertEqual(active_mons, configured_mons,
- "Not all mons are in quorum Invalid standby invalid fscid test failed!")