ceph/qa/tasks/cephfs/test_failover.py

   1 import time
   2 import signal
   3 import logging
   4 import operator
   5 from random import randint
   6
   7 from tasks.cephfs.cephfs_test_case import CephFSTestCase
   8 from teuthology.exceptions import CommandFailedError
   9 from tasks.cephfs.fuse_mount import FuseMount
  10
  11 log = logging.getLogger(__name__)
  12
  13 class TestClusterAffinity(CephFSTestCase):
  14     CLIENTS_REQUIRED = 0
  15     MDSS_REQUIRED = 4
  16
  17     def _verify_join_fs(self, target, status=None):
  18         if status is None:
  19             status = self.fs.wait_for_daemons(timeout=30)
  20             log.debug("%s", status)
  21         target = sorted(target, key=operator.itemgetter('name'))
  22         log.info("target = %s", target)
  23         current = list(status.get_all())
  24         current = sorted(current, key=operator.itemgetter('name'))
  25         log.info("current = %s", current)
  26         self.assertEqual(len(current), len(target))
  27         for i in range(len(current)):
  28             for attr in target[i]:
  29                 self.assertIn(attr, current[i])
  30                 self.assertEqual(target[i][attr], current[i][attr])
  31
  32     def _change_target_state(self, state, name, changes):
  33         for entity in state:
  34             if entity['name'] == name:
  35                 for k, v in changes.items():
  36                     entity[k] = v
  37                 return
  38         self.fail("no entity")
  39
  40     def _verify_init(self):
  41         status = self.fs.status()
  42         log.info("status = {0}".format(status))
  43         target = [{'join_fscid': -1, 'name': info['name']} for info in status.get_all()]
  44         self._verify_join_fs(target, status=status)
  45         return (status, target)
  46
  47     def _reach_target(self, target):
  48         def takeover():
  49             try:
  50                 self._verify_join_fs(target)
  51                 return True
  52             except AssertionError as e:
  53                 log.debug("%s", e)
  54                 return False
  55         self.wait_until_true(takeover, 30)
  56
  57     def test_join_fs_runtime(self):
  58         """
  59         That setting mds_join_fs at runtime affects the cluster layout.
  60         """
  61         status, target = self._verify_init()
  62         standbys = list(status.get_standbys())
  63         self.config_set('mds.'+standbys[0]['name'], 'mds_join_fs', 'cephfs')
  64         self._change_target_state(target, standbys[0]['name'], {'join_fscid': self.fs.id, 'state': 'up:active'})
  65         self._reach_target(target)
  66
  67     def test_join_fs_unset(self):
  68         """
  69         That unsetting mds_join_fs will cause failover if another high-affinity standby exists.
  70         """
  71         status, target = self._verify_init()
  72         standbys = list(status.get_standbys())
  73         names = (standbys[0]['name'], standbys[1]['name'])
  74         self.config_set('mds.'+names[0], 'mds_join_fs', 'cephfs')
  75         self.config_set('mds.'+names[1], 'mds_join_fs', 'cephfs')
  76         self._change_target_state(target, names[0], {'join_fscid': self.fs.id})
  77         self._change_target_state(target, names[1], {'join_fscid': self.fs.id})
  78         self._reach_target(target)
  79         status = self.fs.status()
  80         active = self.fs.get_active_names(status=status)[0]
  81         self.assertIn(active, names)
  82         self.config_rm('mds.'+active, 'mds_join_fs')
  83         self._change_target_state(target, active, {'join_fscid': -1})
  84         new_active = (set(names) - set((active,))).pop()
  85         self._change_target_state(target, new_active, {'state': 'up:active'})
  86         self._reach_target(target)
  87
  88     def test_join_fs_drop(self):
  89         """
  90         That unsetting mds_join_fs will not cause failover if no high-affinity standby exists.
  91         """
  92         status, target = self._verify_init()
  93         standbys = list(status.get_standbys())
  94         active = standbys[0]['name']
  95         self.config_set('mds.'+active, 'mds_join_fs', 'cephfs')
  96         self._change_target_state(target, active, {'join_fscid': self.fs.id, 'state': 'up:active'})
  97         self._reach_target(target)
  98         self.config_rm('mds.'+active, 'mds_join_fs')
  99         self._change_target_state(target, active, {'join_fscid': -1})
 100         self._reach_target(target)
 101
 102     def test_join_fs_vanilla(self):
 103         """
 104         That a vanilla standby is preferred over others with mds_join_fs set to another fs.
 105         """
 106         # After Octopus is EOL, we can remove this setting:
 107         self.fs.set_allow_multifs()
 108         fs2 = self.mds_cluster.newfs(name="cephfs2")
 109         status, target = self._verify_init()
 110         active = self.fs.get_active_names(status=status)[0]
 111         standbys = [info['name'] for info in status.get_standbys()]
 112         victim = standbys.pop()
 113         # Set a bogus fs on the others
 114         for mds in standbys:
 115             self.config_set('mds.'+mds, 'mds_join_fs', 'cephfs2')
 116             self._change_target_state(target, mds, {'join_fscid': fs2.id})
 117         self.fs.rank_fail()
 118         self._change_target_state(target, victim, {'state': 'up:active'})
 119         self._reach_target(target)
 120         status = self.fs.status()
 121         active = self.fs.get_active_names(status=status)[0]
 122         self.assertEqual(active, victim)
 123
 124     def test_join_fs_last_resort(self):
 125         """
 126         That a standby with mds_join_fs set to another fs is still used if necessary.
 127         """
 128         status, target = self._verify_init()
 129         standbys = [info['name'] for info in status.get_standbys()]
 130         for mds in standbys:
 131             self.config_set('mds.'+mds, 'mds_join_fs', 'cephfs2')
 132         # After Octopus is EOL, we can remove this setting:
 133         self.fs.set_allow_multifs()
 134         fs2 = self.mds_cluster.newfs(name="cephfs2")
 135         for mds in standbys:
 136             self._change_target_state(target, mds, {'join_fscid': fs2.id})
 137         self.fs.rank_fail()
 138         status = self.fs.status()
 139         ranks = list(self.fs.get_ranks(status=status))
 140         self.assertEqual(len(ranks), 1)
 141         self.assertIn(ranks[0]['name'], standbys)
 142         # Note that we would expect the former active to reclaim its spot, but
 143         # we're not testing that here.
 144
 145     def test_join_fs_steady(self):
 146         """
 147         That a sole MDS with mds_join_fs set will come back as active eventually even after failover.
 148         """
 149         status, target = self._verify_init()
 150         active = self.fs.get_active_names(status=status)[0]
 151         self.config_set('mds.'+active, 'mds_join_fs', 'cephfs')
 152         self._change_target_state(target, active, {'join_fscid': self.fs.id})
 153         self._reach_target(target)
 154         self.fs.rank_fail()
 155         self._reach_target(target)
 156
 157     def test_join_fs_standby_replay(self):
 158         """
 159         That a standby-replay daemon with weak affinity is replaced by a stronger one.
 160         """
 161         status, target = self._verify_init()
 162         standbys = [info['name'] for info in status.get_standbys()]
 163         self.config_set('mds.'+standbys[0], 'mds_join_fs', 'cephfs')
 164         self._change_target_state(target, standbys[0], {'join_fscid': self.fs.id, 'state': 'up:active'})
 165         self._reach_target(target)
 166         self.fs.set_allow_standby_replay(True)
 167         status = self.fs.status()
 168         standbys = [info['name'] for info in status.get_standbys()]
 169         self.config_set('mds.'+standbys[0], 'mds_join_fs', 'cephfs')
 170         self._change_target_state(target, standbys[0], {'join_fscid': self.fs.id, 'state': 'up:standby-replay'})
 171         self._reach_target(target)
 172
 173 class TestClusterResize(CephFSTestCase):
 174     CLIENTS_REQUIRED = 0
 175     MDSS_REQUIRED = 3
 176
 177     def test_grow(self):
 178         """
 179         That the MDS cluster grows after increasing max_mds.
 180         """
 181
 182         # Need all my standbys up as well as the active daemons
 183         # self.wait_for_daemon_start() necessary?
 184
 185         self.fs.grow(2)
 186         self.fs.grow(3)
 187
 188
 189     def test_shrink(self):
 190         """
 191         That the MDS cluster shrinks automatically after decreasing max_mds.
 192         """
 193
 194         self.fs.grow(3)
 195         self.fs.shrink(1)
 196
 197     def test_up_less_than_max(self):
 198         """
 199         That a health warning is generated when max_mds is greater than active count.
 200         """
 201
 202         status = self.fs.status()
 203         mdss = [info['gid'] for info in status.get_all()]
 204         self.fs.set_max_mds(len(mdss)+1)
 205         self.wait_for_health("MDS_UP_LESS_THAN_MAX", 30)
 206         self.fs.shrink(2)
 207         self.wait_for_health_clear(30)
 208
 209     def test_down_health(self):
 210         """
 211         That marking a FS down does not generate a health warning
 212         """
 213
 214         self.fs.set_down()
 215         try:
 216             self.wait_for_health("", 30)
 217             raise RuntimeError("got health warning?")
 218         except RuntimeError as e:
 219             if "Timed out after" in str(e):
 220                 pass
 221             else:
 222                 raise
 223
 224     def test_down_twice(self):
 225         """
 226         That marking a FS down twice does not wipe old_max_mds.
 227         """
 228
 229         self.fs.grow(2)
 230         self.fs.set_down()
 231         self.fs.wait_for_daemons()
 232         self.fs.set_down(False)
 233         self.assertEqual(self.fs.get_var("max_mds"), 2)
 234         self.fs.wait_for_daemons(timeout=60)
 235
 236     def test_down_grow(self):
 237         """
 238         That setting max_mds undoes down.
 239         """
 240
 241         self.fs.set_down()
 242         self.fs.wait_for_daemons()
 243         self.fs.grow(2)
 244         self.fs.wait_for_daemons()
 245
 246     def test_down(self):
 247         """
 248         That down setting toggles and sets max_mds appropriately.
 249         """
 250
 251         self.fs.set_down()
 252         self.fs.wait_for_daemons()
 253         self.assertEqual(self.fs.get_var("max_mds"), 0)
 254         self.fs.set_down(False)
 255         self.assertEqual(self.fs.get_var("max_mds"), 1)
 256         self.fs.wait_for_daemons()
 257         self.assertEqual(self.fs.get_var("max_mds"), 1)
 258
 259     def test_hole(self):
 260         """
 261         Test that a hole cannot be created in the FS ranks.
 262         """
 263
 264         fscid = self.fs.id
 265
 266         self.fs.grow(2)
 267
 268         # Now add a delay which should slow down how quickly rank 1 stops
 269         self.config_set('mds', 'ms_inject_delay_max', '5.0')
 270         self.config_set('mds', 'ms_inject_delay_probability', '1.0')
 271         self.fs.set_max_mds(1)
 272         log.info("status = {0}".format(self.fs.status()))
 273
 274         # Don't wait for rank 1 to stop
 275         self.fs.set_max_mds(3)
 276         log.info("status = {0}".format(self.fs.status()))
 277
 278         # Now check that the mons didn't try to promote a standby to rank 2
 279         self.fs.set_max_mds(2)
 280         status = self.fs.status()
 281         try:
 282             status = self.fs.wait_for_daemons(timeout=90)
 283             ranks = set([info['rank'] for info in status.get_ranks(fscid)])
 284             self.assertEqual(ranks, set([0, 1]))
 285         finally:
 286             log.info("status = {0}".format(status))
 287
 288     def test_thrash(self):
 289         """
 290         Test that thrashing max_mds does not fail.
 291         """
 292
 293         max_mds = 2
 294         for i in range(0, 100):
 295             self.fs.set_max_mds(max_mds)
 296             max_mds = (max_mds+1)%3+1
 297
 298         self.fs.wait_for_daemons(timeout=90)
 299
 300 class TestFailover(CephFSTestCase):
 301     CLIENTS_REQUIRED = 1
 302     MDSS_REQUIRED = 2
 303
 304     def test_simple(self):
 305         """
 306         That when the active MDS is killed, a standby MDS is promoted into
 307         its rank after the grace period.
 308
 309         This is just a simple unit test, the harder cases are covered
 310         in thrashing tests.
 311         """
 312
 313         # Need all my standbys up as well as the active daemons
 314         self.wait_for_daemon_start()
 315
 316         (original_active, ) = self.fs.get_active_names()
 317         original_standbys = self.mds_cluster.get_standby_daemons()
 318
 319         # Kill the rank 0 daemon's physical process
 320         self.fs.mds_stop(original_active)
 321
 322         grace = float(self.fs.get_config("mds_beacon_grace", service_type="mon"))
 323
 324         # Wait until the monitor promotes his replacement
 325         def promoted():
 326             active = self.fs.get_active_names()
 327             return active and active[0] in original_standbys
 328
 329         log.info("Waiting for promotion of one of the original standbys {0}".format(
 330             original_standbys))
 331         self.wait_until_true(
 332             promoted,
 333             timeout=grace*2)
 334
 335         # Start the original rank 0 daemon up again, see that he becomes a standby
 336         self.fs.mds_restart(original_active)
 337         self.wait_until_true(
 338             lambda: original_active in self.mds_cluster.get_standby_daemons(),
 339             timeout=60  # Approximately long enough for MDS to start and mon to notice
 340         )
 341
 342     def test_client_abort(self):
 343         """
 344         That a client will respect fuse_require_active_mds and error out
 345         when the cluster appears to be unavailable.
 346         """
 347
 348         if not isinstance(self.mount_a, FuseMount):
 349             self.skipTest("Requires FUSE client to inject client metadata")
 350
 351         require_active = self.fs.get_config("fuse_require_active_mds", service_type="mon").lower() == "true"
 352         if not require_active:
 353             self.skipTest("fuse_require_active_mds is not set")
 354
 355         grace = float(self.fs.get_config("mds_beacon_grace", service_type="mon"))
 356
 357         # Check it's not laggy to begin with
 358         (original_active, ) = self.fs.get_active_names()
 359         self.assertNotIn("laggy_since", self.fs.status().get_mds(original_active))
 360
 361         self.mounts[0].umount_wait()
 362
 363         # Control: that we can mount and unmount usually, while the cluster is healthy
 364         self.mounts[0].mount_wait()
 365         self.mounts[0].umount_wait()
 366
 367         # Stop the daemon processes
 368         self.fs.mds_stop()
 369
 370         # Wait for everyone to go laggy
 371         def laggy():
 372             mdsmap = self.fs.get_mds_map()
 373             for info in mdsmap['info'].values():
 374                 if "laggy_since" not in info:
 375                     return False
 376
 377             return True
 378
 379         self.wait_until_true(laggy, grace * 2)
 380         with self.assertRaises(CommandFailedError):
 381             self.mounts[0].mount_wait()
 382
 383     def test_standby_count_wanted(self):
 384         """
 385         That cluster health warnings are generated by insufficient standbys available.
 386         """
 387
 388         # Need all my standbys up as well as the active daemons
 389         self.wait_for_daemon_start()
 390
 391         grace = float(self.fs.get_config("mds_beacon_grace", service_type="mon"))
 392
 393         standbys = self.mds_cluster.get_standby_daemons()
 394         self.assertGreaterEqual(len(standbys), 1)
 395         self.fs.mon_manager.raw_cluster_cmd('fs', 'set', self.fs.name, 'standby_count_wanted', str(len(standbys)))
 396
 397         # Kill a standby and check for warning
 398         victim = standbys.pop()
 399         self.fs.mds_stop(victim)
 400         log.info("waiting for insufficient standby daemon warning")
 401         self.wait_for_health("MDS_INSUFFICIENT_STANDBY", grace*2)
 402
 403         # restart the standby, see that he becomes a standby, check health clears
 404         self.fs.mds_restart(victim)
 405         self.wait_until_true(
 406             lambda: victim in self.mds_cluster.get_standby_daemons(),
 407             timeout=60  # Approximately long enough for MDS to start and mon to notice
 408         )
 409         self.wait_for_health_clear(timeout=30)
 410
 411         # Set it one greater than standbys ever seen
 412         standbys = self.mds_cluster.get_standby_daemons()
 413         self.assertGreaterEqual(len(standbys), 1)
 414         self.fs.mon_manager.raw_cluster_cmd('fs', 'set', self.fs.name, 'standby_count_wanted', str(len(standbys)+1))
 415         log.info("waiting for insufficient standby daemon warning")
 416         self.wait_for_health("MDS_INSUFFICIENT_STANDBY", grace*2)
 417
 418         # Set it to 0
 419         self.fs.mon_manager.raw_cluster_cmd('fs', 'set', self.fs.name, 'standby_count_wanted', '0')
 420         self.wait_for_health_clear(timeout=30)
 421
 422     def test_discontinuous_mdsmap(self):
 423         """
 424         That discontinuous mdsmap does not affect failover.
 425         See http://tracker.ceph.com/issues/24856.
 426         """
 427         self.fs.set_max_mds(2)
 428         status = self.fs.wait_for_daemons()
 429
 430         self.mount_a.umount_wait()
 431
 432         grace = float(self.fs.get_config("mds_beacon_grace", service_type="mon"))
 433         monc_timeout = float(self.fs.get_config("mon_client_ping_timeout", service_type="mds"))
 434
 435         mds_0 = self.fs.get_rank(rank=0, status=status)
 436         self.fs.rank_freeze(True, rank=0) # prevent failover
 437         self.fs.rank_signal(signal.SIGSTOP, rank=0, status=status)
 438         self.wait_until_true(
 439             lambda: "laggy_since" in self.fs.get_rank(),
 440             timeout=grace * 2
 441         )
 442
 443         self.fs.rank_fail(rank=1)
 444         self.fs.wait_for_state('up:resolve', rank=1, timeout=30)
 445
 446         # Make sure of mds_0's monitor connection gets reset
 447         time.sleep(monc_timeout * 2)
 448
 449         # Continue rank 0, it will get discontinuous mdsmap
 450         self.fs.rank_signal(signal.SIGCONT, rank=0)
 451         self.wait_until_true(
 452             lambda: "laggy_since" not in self.fs.get_rank(rank=0),
 453             timeout=grace * 2
 454         )
 455
 456         # mds.b will be stuck at 'reconnect' state if snapserver gets confused
 457         # by discontinuous mdsmap
 458         self.fs.wait_for_state('up:active', rank=1, timeout=30)
 459         self.assertEqual(mds_0['gid'], self.fs.get_rank(rank=0)['gid'])
 460         self.fs.rank_freeze(False, rank=0)
 461
 462 class TestStandbyReplay(CephFSTestCase):
 463     CLIENTS_REQUIRED = 0
 464     MDSS_REQUIRED = 4
 465
 466     def _confirm_no_replay(self):
 467         status = self.fs.status()
 468         _ = len(list(status.get_standbys()))
 469         self.assertEqual(0, len(list(self.fs.get_replays(status=status))))
 470         return status
 471
 472     def _confirm_single_replay(self, full=True, status=None, retries=3):
 473         status = self.fs.wait_for_daemons(status=status)
 474         ranks = sorted(self.fs.get_mds_map(status=status)['in'])
 475         replays = list(self.fs.get_replays(status=status))
 476         checked_replays = set()
 477         for rank in ranks:
 478             has_replay = False
 479             for replay in replays:
 480                 if replay['rank'] == rank:
 481                     self.assertFalse(has_replay)
 482                     has_replay = True
 483                     checked_replays.add(replay['gid'])
 484             if full and not has_replay:
 485                 if retries <= 0:
 486                     raise RuntimeError("rank "+str(rank)+" has no standby-replay follower")
 487                 else:
 488                     retries = retries-1
 489                     time.sleep(2)
 490         self.assertEqual(checked_replays, set(info['gid'] for info in replays))
 491         return status
 492
 493     def _check_replay_takeover(self, status, rank=0):
 494         replay = self.fs.get_replay(rank=rank, status=status)
 495         new_status = self.fs.wait_for_daemons()
 496         new_active = self.fs.get_rank(rank=rank, status=new_status)
 497         if replay:
 498             self.assertEqual(replay['gid'], new_active['gid'])
 499         else:
 500             # double check takeover came from a standby (or some new daemon via restart)
 501             found = False
 502             for info in status.get_standbys():
 503                 if info['gid'] == new_active['gid']:
 504                     found = True
 505                     break
 506             if not found:
 507                 for info in status.get_all():
 508                     self.assertNotEqual(info['gid'], new_active['gid'])
 509         return new_status
 510
 511     def test_standby_replay_singleton(self):
 512         """
 513         That only one MDS becomes standby-replay.
 514         """
 515
 516         self._confirm_no_replay()
 517         self.fs.set_allow_standby_replay(True)
 518         time.sleep(30)
 519         self._confirm_single_replay()
 520
 521     def test_standby_replay_damaged(self):
 522         """
 523         That a standby-replay daemon can cause the rank to go damaged correctly.
 524         """
 525
 526         self._confirm_no_replay()
 527         self.config_set("mds", "mds_standby_replay_damaged", True)
 528         self.fs.set_allow_standby_replay(True)
 529         self.wait_until_true(
 530             lambda: len(self.fs.get_damaged()) > 0,
 531             timeout=30
 532         )
 533         status = self.fs.status()
 534         self.assertListEqual([], list(self.fs.get_ranks(status=status)))
 535         self.assertListEqual([0], self.fs.get_damaged(status=status))
 536
 537     def test_standby_replay_disable(self):
 538         """
 539         That turning off allow_standby_replay fails all standby-replay daemons.
 540         """
 541
 542         self._confirm_no_replay()
 543         self.fs.set_allow_standby_replay(True)
 544         time.sleep(30)
 545         self._confirm_single_replay()
 546         self.fs.set_allow_standby_replay(False)
 547         self._confirm_no_replay()
 548
 549     def test_standby_replay_singleton_fail(self):
 550         """
 551         That failures don't violate singleton constraint.
 552         """
 553
 554         self._confirm_no_replay()
 555         self.fs.set_allow_standby_replay(True)
 556         status = self._confirm_single_replay()
 557
 558         for i in range(10):
 559             time.sleep(randint(1, 5))
 560             self.fs.rank_restart(status=status)
 561             status = self._check_replay_takeover(status)
 562             status = self._confirm_single_replay(status=status)
 563
 564         for i in range(10):
 565             time.sleep(randint(1, 5))
 566             self.fs.rank_fail()
 567             status = self._check_replay_takeover(status)
 568             status = self._confirm_single_replay(status=status)
 569
 570     def test_standby_replay_singleton_fail_multimds(self):
 571         """
 572         That failures don't violate singleton constraint with multiple actives.
 573         """
 574
 575         status = self._confirm_no_replay()
 576         new_max_mds = randint(2, len(list(status.get_standbys())))
 577         self.fs.set_max_mds(new_max_mds)
 578         self.fs.wait_for_daemons() # wait for actives to come online!
 579         self.fs.set_allow_standby_replay(True)
 580         status = self._confirm_single_replay(full=False)
 581
 582         for i in range(10):
 583             time.sleep(randint(1, 5))
 584             victim = randint(0, new_max_mds-1)
 585             self.fs.rank_restart(rank=victim, status=status)
 586             status = self._check_replay_takeover(status, rank=victim)
 587             status = self._confirm_single_replay(status=status, full=False)
 588
 589         for i in range(10):
 590             time.sleep(randint(1, 5))
 591             victim = randint(0, new_max_mds-1)
 592             self.fs.rank_fail(rank=victim)
 593             status = self._check_replay_takeover(status, rank=victim)
 594             status = self._confirm_single_replay(status=status, full=False)
 595
 596     def test_standby_replay_failure(self):
 597         """
 598         That the failure of a standby-replay daemon happens cleanly
 599         and doesn't interrupt anything else.
 600         """
 601
 602         status = self._confirm_no_replay()
 603         self.fs.set_max_mds(1)
 604         self.fs.set_allow_standby_replay(True)
 605         status = self._confirm_single_replay()
 606
 607         for i in range(10):
 608             time.sleep(randint(1, 5))
 609             victim = self.fs.get_replay(status=status)
 610             self.fs.mds_restart(mds_id=victim['name'])
 611             status = self._confirm_single_replay(status=status)
 612
 613     def test_rank_stopped(self):
 614         """
 615         That when a rank is STOPPED, standby replays for
 616         that rank get torn down
 617         """
 618
 619         status = self._confirm_no_replay()
 620         standby_count = len(list(status.get_standbys()))
 621         self.fs.set_max_mds(2)
 622         self.fs.set_allow_standby_replay(True)
 623         status = self._confirm_single_replay()
 624
 625         self.fs.set_max_mds(1) # stop rank 1
 626
 627         status = self._confirm_single_replay()
 628         self.assertTrue(standby_count, len(list(status.get_standbys())))
 629
 630
 631 class TestMultiFilesystems(CephFSTestCase):
 632     CLIENTS_REQUIRED = 2
 633     MDSS_REQUIRED = 4
 634
 635     # We'll create our own filesystems and start our own daemons
 636     REQUIRE_FILESYSTEM = False
 637
 638     def setUp(self):
 639         super(TestMultiFilesystems, self).setUp()
 640         self.mds_cluster.mon_manager.raw_cluster_cmd("fs", "flag", "set",
 641             "enable_multiple", "true",
 642             "--yes-i-really-mean-it")
 643
 644     def _setup_two(self):
 645         fs_a = self.mds_cluster.newfs(name="alpha")
 646         fs_b = self.mds_cluster.newfs(name="bravo")
 647
 648         self.mds_cluster.mds_restart()
 649
 650         # Wait for both filesystems to go healthy
 651         fs_a.wait_for_daemons()
 652         fs_b.wait_for_daemons()
 653
 654         # Reconfigure client auth caps
 655         for mount in self.mounts:
 656             self.mds_cluster.mon_manager.raw_cluster_cmd_result(
 657                 'auth', 'caps', "client.{0}".format(mount.client_id),
 658                 'mds', 'allow',
 659                 'mon', 'allow r',
 660                 'osd', 'allow rw pool={0}, allow rw pool={1}'.format(
 661                     fs_a.get_data_pool_name(), fs_b.get_data_pool_name()))
 662
 663         return fs_a, fs_b
 664
 665     def test_clients(self):
 666         fs_a, fs_b = self._setup_two()
 667
 668         # Mount a client on fs_a
 669         self.mount_a.mount_wait(cephfs_name=fs_a.name)
 670         self.mount_a.write_n_mb("pad.bin", 1)
 671         self.mount_a.write_n_mb("test.bin", 2)
 672         a_created_ino = self.mount_a.path_to_ino("test.bin")
 673         self.mount_a.create_files()
 674
 675         # Mount a client on fs_b
 676         self.mount_b.mount_wait(cephfs_name=fs_b.name)
 677         self.mount_b.write_n_mb("test.bin", 1)
 678         b_created_ino = self.mount_b.path_to_ino("test.bin")
 679         self.mount_b.create_files()
 680
 681         # Check that a non-default filesystem mount survives an MDS
 682         # failover (i.e. that map subscription is continuous, not
 683         # just the first time), reproduces #16022
 684         old_fs_b_mds = fs_b.get_active_names()[0]
 685         self.mds_cluster.mds_stop(old_fs_b_mds)
 686         self.mds_cluster.mds_fail(old_fs_b_mds)
 687         fs_b.wait_for_daemons()
 688         background = self.mount_b.write_background()
 689         # Raise exception if the write doesn't finish (i.e. if client
 690         # has not kept up with MDS failure)
 691         try:
 692             self.wait_until_true(lambda: background.finished, timeout=30)
 693         except RuntimeError:
 694             # The mount is stuck, we'll have to force it to fail cleanly
 695             background.stdin.close()
 696             self.mount_b.umount_wait(force=True)
 697             raise
 698
 699         self.mount_a.umount_wait()
 700         self.mount_b.umount_wait()
 701
 702         # See that the client's files went into the correct pool
 703         self.assertTrue(fs_a.data_objects_present(a_created_ino, 1024 * 1024))
 704         self.assertTrue(fs_b.data_objects_present(b_created_ino, 1024 * 1024))
 705
 706     def test_standby(self):
 707         fs_a, fs_b = self._setup_two()
 708
 709         # Assert that the remaining two MDS daemons are now standbys
 710         a_daemons = fs_a.get_active_names()
 711         b_daemons = fs_b.get_active_names()
 712         self.assertEqual(len(a_daemons), 1)
 713         self.assertEqual(len(b_daemons), 1)
 714         original_a = a_daemons[0]
 715         original_b = b_daemons[0]
 716         expect_standby_daemons = set(self.mds_cluster.mds_ids) - (set(a_daemons) | set(b_daemons))
 717
 718         # Need all my standbys up as well as the active daemons
 719         self.wait_for_daemon_start()
 720         self.assertEqual(expect_standby_daemons, self.mds_cluster.get_standby_daemons())
 721
 722         # Kill fs_a's active MDS, see a standby take over
 723         self.mds_cluster.mds_stop(original_a)
 724         self.mds_cluster.mon_manager.raw_cluster_cmd("mds", "fail", original_a)
 725         self.wait_until_equal(lambda: len(fs_a.get_active_names()), 1, 30,
 726                               reject_fn=lambda v: v > 1)
 727         # Assert that it's a *different* daemon that has now appeared in the map for fs_a
 728         self.assertNotEqual(fs_a.get_active_names()[0], original_a)
 729
 730         # Kill fs_b's active MDS, see a standby take over
 731         self.mds_cluster.mds_stop(original_b)
 732         self.mds_cluster.mon_manager.raw_cluster_cmd("mds", "fail", original_b)
 733         self.wait_until_equal(lambda: len(fs_b.get_active_names()), 1, 30,
 734                               reject_fn=lambda v: v > 1)
 735         # Assert that it's a *different* daemon that has now appeared in the map for fs_a
 736         self.assertNotEqual(fs_b.get_active_names()[0], original_b)
 737
 738         # Both of the original active daemons should be gone, and all standbys used up
 739         self.assertEqual(self.mds_cluster.get_standby_daemons(), set())
 740
 741         # Restart the ones I killed, see them reappear as standbys
 742         self.mds_cluster.mds_restart(original_a)
 743         self.mds_cluster.mds_restart(original_b)
 744         self.wait_until_true(
 745             lambda: {original_a, original_b} == self.mds_cluster.get_standby_daemons(),
 746             timeout=30
 747         )
 748
 749     def test_grow_shrink(self):
 750         # Usual setup...
 751         fs_a, fs_b = self._setup_two()
 752
 753         # Increase max_mds on fs_b, see a standby take up the role
 754         fs_b.set_max_mds(2)
 755         self.wait_until_equal(lambda: len(fs_b.get_active_names()), 2, 30,
 756                               reject_fn=lambda v: v > 2 or v < 1)
 757
 758         # Increase max_mds on fs_a, see a standby take up the role
 759         fs_a.set_max_mds(2)
 760         self.wait_until_equal(lambda: len(fs_a.get_active_names()), 2, 30,
 761                               reject_fn=lambda v: v > 2 or v < 1)
 762
 763         # Shrink fs_b back to 1, see a daemon go back to standby
 764         fs_b.set_max_mds(1)
 765         self.wait_until_equal(lambda: len(fs_b.get_active_names()), 1, 30,
 766                               reject_fn=lambda v: v > 2 or v < 1)
 767
 768         # Grow fs_a up to 3, see the former fs_b daemon join it.
 769         fs_a.set_max_mds(3)
 770         self.wait_until_equal(lambda: len(fs_a.get_active_names()), 3, 60,
 771                               reject_fn=lambda v: v > 3 or v < 2)