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         fs2 = self.mds_cluster.newfs(name="cephfs2")
 107         status, target = self._verify_init()
 108         active = self.fs.get_active_names(status=status)[0]
 109         standbys = [info['name'] for info in status.get_standbys()]
 110         victim = standbys.pop()
 111         # Set a bogus fs on the others
 112         for mds in standbys:
 113             self.config_set('mds.'+mds, 'mds_join_fs', 'cephfs2')
 114             self._change_target_state(target, mds, {'join_fscid': fs2.id})
 115         self.fs.rank_fail()
 116         self._change_target_state(target, victim, {'state': 'up:active'})
 117         self._reach_target(target)
 118         status = self.fs.status()
 119         active = self.fs.get_active_names(status=status)[0]
 120         self.assertEqual(active, victim)
 121
 122     def test_join_fs_last_resort(self):
 123         """
 124         That a standby with mds_join_fs set to another fs is still used if necessary.
 125         """
 126         status, target = self._verify_init()
 127         standbys = [info['name'] for info in status.get_standbys()]
 128         for mds in standbys:
 129             self.config_set('mds.'+mds, 'mds_join_fs', 'cephfs2')
 130         fs2 = self.mds_cluster.newfs(name="cephfs2")
 131         for mds in standbys:
 132             self._change_target_state(target, mds, {'join_fscid': fs2.id})
 133         self.fs.rank_fail()
 134         status = self.fs.status()
 135         ranks = list(self.fs.get_ranks(status=status))
 136         self.assertEqual(len(ranks), 1)
 137         self.assertIn(ranks[0]['name'], standbys)
 138         # Note that we would expect the former active to reclaim its spot, but
 139         # we're not testing that here.
 140
 141     def test_join_fs_steady(self):
 142         """
 143         That a sole MDS with mds_join_fs set will come back as active eventually even after failover.
 144         """
 145         status, target = self._verify_init()
 146         active = self.fs.get_active_names(status=status)[0]
 147         self.config_set('mds.'+active, 'mds_join_fs', 'cephfs')
 148         self._change_target_state(target, active, {'join_fscid': self.fs.id})
 149         self._reach_target(target)
 150         self.fs.rank_fail()
 151         self._reach_target(target)
 152
 153     def test_join_fs_standby_replay(self):
 154         """
 155         That a standby-replay daemon with weak affinity is replaced by a stronger one.
 156         """
 157         status, target = self._verify_init()
 158         standbys = [info['name'] for info in status.get_standbys()]
 159         self.config_set('mds.'+standbys[0], 'mds_join_fs', 'cephfs')
 160         self._change_target_state(target, standbys[0], {'join_fscid': self.fs.id, 'state': 'up:active'})
 161         self._reach_target(target)
 162         self.fs.set_allow_standby_replay(True)
 163         status = self.fs.status()
 164         standbys = [info['name'] for info in status.get_standbys()]
 165         self.config_set('mds.'+standbys[0], 'mds_join_fs', 'cephfs')
 166         self._change_target_state(target, standbys[0], {'join_fscid': self.fs.id, 'state': 'up:standby-replay'})
 167         self._reach_target(target)
 168
 169 class TestClusterResize(CephFSTestCase):
 170     CLIENTS_REQUIRED = 0
 171     MDSS_REQUIRED = 3
 172
 173     def test_grow(self):
 174         """
 175         That the MDS cluster grows after increasing max_mds.
 176         """
 177
 178         # Need all my standbys up as well as the active daemons
 179         # self.wait_for_daemon_start() necessary?
 180
 181         self.fs.grow(2)
 182         self.fs.grow(3)
 183
 184
 185     def test_shrink(self):
 186         """
 187         That the MDS cluster shrinks automatically after decreasing max_mds.
 188         """
 189
 190         self.fs.grow(3)
 191         self.fs.shrink(1)
 192
 193     def test_up_less_than_max(self):
 194         """
 195         That a health warning is generated when max_mds is greater than active count.
 196         """
 197
 198         status = self.fs.status()
 199         mdss = [info['gid'] for info in status.get_all()]
 200         self.fs.set_max_mds(len(mdss)+1)
 201         self.wait_for_health("MDS_UP_LESS_THAN_MAX", 30)
 202         self.fs.shrink(2)
 203         self.wait_for_health_clear(30)
 204
 205     def test_down_health(self):
 206         """
 207         That marking a FS down does not generate a health warning
 208         """
 209
 210         self.fs.set_down()
 211         try:
 212             self.wait_for_health("", 30)
 213             raise RuntimeError("got health warning?")
 214         except RuntimeError as e:
 215             if "Timed out after" in str(e):
 216                 pass
 217             else:
 218                 raise
 219
 220     def test_down_twice(self):
 221         """
 222         That marking a FS down twice does not wipe old_max_mds.
 223         """
 224
 225         self.fs.grow(2)
 226         self.fs.set_down()
 227         self.fs.wait_for_daemons()
 228         self.fs.set_down(False)
 229         self.assertEqual(self.fs.get_var("max_mds"), 2)
 230         self.fs.wait_for_daemons(timeout=60)
 231
 232     def test_down_grow(self):
 233         """
 234         That setting max_mds undoes down.
 235         """
 236
 237         self.fs.set_down()
 238         self.fs.wait_for_daemons()
 239         self.fs.grow(2)
 240         self.fs.wait_for_daemons()
 241
 242     def test_down(self):
 243         """
 244         That down setting toggles and sets max_mds appropriately.
 245         """
 246
 247         self.fs.set_down()
 248         self.fs.wait_for_daemons()
 249         self.assertEqual(self.fs.get_var("max_mds"), 0)
 250         self.fs.set_down(False)
 251         self.assertEqual(self.fs.get_var("max_mds"), 1)
 252         self.fs.wait_for_daemons()
 253         self.assertEqual(self.fs.get_var("max_mds"), 1)
 254
 255     def test_hole(self):
 256         """
 257         Test that a hole cannot be created in the FS ranks.
 258         """
 259
 260         fscid = self.fs.id
 261
 262         self.fs.grow(2)
 263
 264         # Now add a delay which should slow down how quickly rank 1 stops
 265         self.config_set('mds', 'ms_inject_delay_max', '5.0')
 266         self.config_set('mds', 'ms_inject_delay_probability', '1.0')
 267         self.fs.set_max_mds(1)
 268         log.info("status = {0}".format(self.fs.status()))
 269
 270         # Don't wait for rank 1 to stop
 271         self.fs.set_max_mds(3)
 272         log.info("status = {0}".format(self.fs.status()))
 273
 274         # Now check that the mons didn't try to promote a standby to rank 2
 275         self.fs.set_max_mds(2)
 276         status = self.fs.status()
 277         try:
 278             status = self.fs.wait_for_daemons(timeout=90)
 279             ranks = set([info['rank'] for info in status.get_ranks(fscid)])
 280             self.assertEqual(ranks, set([0, 1]))
 281         finally:
 282             log.info("status = {0}".format(status))
 283
 284     def test_thrash(self):
 285         """
 286         Test that thrashing max_mds does not fail.
 287         """
 288
 289         max_mds = 2
 290         for i in range(0, 100):
 291             self.fs.set_max_mds(max_mds)
 292             max_mds = (max_mds+1)%3+1
 293
 294         self.fs.wait_for_daemons(timeout=90)
 295
 296 class TestFailover(CephFSTestCase):
 297     CLIENTS_REQUIRED = 1
 298     MDSS_REQUIRED = 2
 299
 300     def test_simple(self):
 301         """
 302         That when the active MDS is killed, a standby MDS is promoted into
 303         its rank after the grace period.
 304
 305         This is just a simple unit test, the harder cases are covered
 306         in thrashing tests.
 307         """
 308
 309         (original_active, ) = self.fs.get_active_names()
 310         original_standbys = self.mds_cluster.get_standby_daemons()
 311
 312         # Kill the rank 0 daemon's physical process
 313         self.fs.mds_stop(original_active)
 314
 315         # Wait until the monitor promotes his replacement
 316         def promoted():
 317             ranks = list(self.fs.get_ranks())
 318             return len(ranks) > 0 and ranks[0]['name'] in original_standbys
 319
 320         log.info("Waiting for promotion of one of the original standbys {0}".format(
 321             original_standbys))
 322         self.wait_until_true(promoted, timeout=self.fs.beacon_timeout)
 323
 324         # Start the original rank 0 daemon up again, see that he becomes a standby
 325         self.fs.mds_restart(original_active)
 326         self.wait_until_true(
 327             lambda: original_active in self.mds_cluster.get_standby_daemons(),
 328             timeout=60  # Approximately long enough for MDS to start and mon to notice
 329         )
 330
 331     def test_client_abort(self):
 332         """
 333         That a client will respect fuse_require_active_mds and error out
 334         when the cluster appears to be unavailable.
 335         """
 336
 337         if not isinstance(self.mount_a, FuseMount):
 338             self.skipTest("Requires FUSE client to inject client metadata")
 339
 340         require_active = self.fs.get_config("fuse_require_active_mds", service_type="mon").lower() == "true"
 341         if not require_active:
 342             self.skipTest("fuse_require_active_mds is not set")
 343
 344         # Check it's not laggy to begin with
 345         (original_active, ) = self.fs.get_active_names()
 346         self.assertNotIn("laggy_since", self.fs.status().get_mds(original_active))
 347
 348         self.mounts[0].umount_wait()
 349
 350         # Control: that we can mount and unmount usually, while the cluster is healthy
 351         self.mounts[0].mount_wait()
 352         self.mounts[0].umount_wait()
 353
 354         # Stop the daemon processes
 355         self.fs.mds_stop()
 356
 357         # Wait for everyone to go laggy
 358         def laggy():
 359             mdsmap = self.fs.get_mds_map()
 360             for info in mdsmap['info'].values():
 361                 if "laggy_since" not in info:
 362                     return False
 363
 364             return True
 365
 366         self.wait_until_true(laggy, self.fs.beacon_timeout)
 367         with self.assertRaises(CommandFailedError):
 368             self.mounts[0].mount_wait()
 369
 370     def test_standby_count_wanted(self):
 371         """
 372         That cluster health warnings are generated by insufficient standbys available.
 373         """
 374
 375         # Need all my standbys up as well as the active daemons
 376         self.wait_for_daemon_start()
 377
 378         standbys = self.mds_cluster.get_standby_daemons()
 379         self.assertGreaterEqual(len(standbys), 1)
 380         self.fs.mon_manager.raw_cluster_cmd('fs', 'set', self.fs.name, 'standby_count_wanted', str(len(standbys)))
 381
 382         # Kill a standby and check for warning
 383         victim = standbys.pop()
 384         self.fs.mds_stop(victim)
 385         self.wait_for_health("MDS_INSUFFICIENT_STANDBY", self.fs.beacon_timeout)
 386
 387         # restart the standby, see that he becomes a standby, check health clears
 388         self.fs.mds_restart(victim)
 389         self.wait_until_true(
 390             lambda: victim in self.mds_cluster.get_standby_daemons(),
 391             timeout=60  # Approximately long enough for MDS to start and mon to notice
 392         )
 393         self.wait_for_health_clear(timeout=30)
 394
 395         # Set it one greater than standbys ever seen
 396         standbys = self.mds_cluster.get_standby_daemons()
 397         self.assertGreaterEqual(len(standbys), 1)
 398         self.fs.mon_manager.raw_cluster_cmd('fs', 'set', self.fs.name, 'standby_count_wanted', str(len(standbys)+1))
 399         self.wait_for_health("MDS_INSUFFICIENT_STANDBY", self.fs.beacon_timeout)
 400
 401         # Set it to 0
 402         self.fs.mon_manager.raw_cluster_cmd('fs', 'set', self.fs.name, 'standby_count_wanted', '0')
 403         self.wait_for_health_clear(timeout=30)
 404
 405     def test_discontinuous_mdsmap(self):
 406         """
 407         That discontinuous mdsmap does not affect failover.
 408         See http://tracker.ceph.com/issues/24856.
 409         """
 410         self.fs.set_max_mds(2)
 411         status = self.fs.wait_for_daemons()
 412
 413         self.mount_a.umount_wait()
 414
 415         monc_timeout = float(self.fs.get_config("mon_client_ping_timeout", service_type="mds"))
 416
 417         mds_0 = self.fs.get_rank(rank=0, status=status)
 418         self.fs.rank_freeze(True, rank=0) # prevent failover
 419         self.fs.rank_signal(signal.SIGSTOP, rank=0, status=status)
 420         self.wait_until_true(
 421             lambda: "laggy_since" in self.fs.get_rank(),
 422             timeout=self.fs.beacon_timeout
 423         )
 424
 425         self.fs.rank_fail(rank=1)
 426         self.fs.wait_for_state('up:resolve', rank=1, timeout=30)
 427
 428         # Make sure of mds_0's monitor connection gets reset
 429         time.sleep(monc_timeout * 2)
 430
 431         # Continue rank 0, it will get discontinuous mdsmap
 432         self.fs.rank_signal(signal.SIGCONT, rank=0)
 433         self.wait_until_true(
 434             lambda: "laggy_since" not in self.fs.get_rank(rank=0),
 435             timeout=self.fs.beacon_timeout
 436         )
 437
 438         # mds.b will be stuck at 'reconnect' state if snapserver gets confused
 439         # by discontinuous mdsmap
 440         self.fs.wait_for_state('up:active', rank=1, timeout=30)
 441         self.assertEqual(mds_0['gid'], self.fs.get_rank(rank=0)['gid'])
 442         self.fs.rank_freeze(False, rank=0)
 443
 444     def test_connect_bootstrapping(self):
 445         self.config_set("mds", "mds_sleep_rank_change", 10000000.0)
 446         self.config_set("mds", "mds_connect_bootstrapping", True)
 447         self.fs.set_max_mds(2)
 448         self.fs.wait_for_daemons()
 449         self.fs.rank_fail(rank=0)
 450         # rank 0 will get stuck in up:resolve, see https://tracker.ceph.com/issues/53194
 451         self.fs.wait_for_daemons()
 452
 453
 454 class TestStandbyReplay(CephFSTestCase):
 455     CLIENTS_REQUIRED = 0
 456     MDSS_REQUIRED = 4
 457
 458     def _confirm_no_replay(self):
 459         status = self.fs.status()
 460         _ = len(list(status.get_standbys()))
 461         self.assertEqual(0, len(list(self.fs.get_replays(status=status))))
 462         return status
 463
 464     def _confirm_single_replay(self, full=True, status=None, retries=3):
 465         status = self.fs.wait_for_daemons(status=status)
 466         ranks = sorted(self.fs.get_mds_map(status=status)['in'])
 467         replays = list(self.fs.get_replays(status=status))
 468         checked_replays = set()
 469         for rank in ranks:
 470             has_replay = False
 471             for replay in replays:
 472                 if replay['rank'] == rank:
 473                     self.assertFalse(has_replay)
 474                     has_replay = True
 475                     checked_replays.add(replay['gid'])
 476             if full and not has_replay:
 477                 if retries <= 0:
 478                     raise RuntimeError("rank "+str(rank)+" has no standby-replay follower")
 479                 else:
 480                     retries = retries-1
 481                     time.sleep(2)
 482         self.assertEqual(checked_replays, set(info['gid'] for info in replays))
 483         return status
 484
 485     def _check_replay_takeover(self, status, rank=0):
 486         replay = self.fs.get_replay(rank=rank, status=status)
 487         new_status = self.fs.wait_for_daemons()
 488         new_active = self.fs.get_rank(rank=rank, status=new_status)
 489         if replay:
 490             self.assertEqual(replay['gid'], new_active['gid'])
 491         else:
 492             # double check takeover came from a standby (or some new daemon via restart)
 493             found = False
 494             for info in status.get_standbys():
 495                 if info['gid'] == new_active['gid']:
 496                     found = True
 497                     break
 498             if not found:
 499                 for info in status.get_all():
 500                     self.assertNotEqual(info['gid'], new_active['gid'])
 501         return new_status
 502
 503     def test_standby_replay_singleton(self):
 504         """
 505         That only one MDS becomes standby-replay.
 506         """
 507
 508         self._confirm_no_replay()
 509         self.fs.set_allow_standby_replay(True)
 510         time.sleep(30)
 511         self._confirm_single_replay()
 512
 513     def test_standby_replay_damaged(self):
 514         """
 515         That a standby-replay daemon can cause the rank to go damaged correctly.
 516         """
 517
 518         self._confirm_no_replay()
 519         self.config_set("mds", "mds_standby_replay_damaged", True)
 520         self.fs.set_allow_standby_replay(True)
 521         self.wait_until_true(
 522             lambda: len(self.fs.get_damaged()) > 0,
 523             timeout=30
 524         )
 525         status = self.fs.status()
 526         self.assertListEqual([], list(self.fs.get_ranks(status=status)))
 527         self.assertListEqual([0], self.fs.get_damaged(status=status))
 528
 529     def test_standby_replay_disable(self):
 530         """
 531         That turning off allow_standby_replay fails all standby-replay daemons.
 532         """
 533
 534         self._confirm_no_replay()
 535         self.fs.set_allow_standby_replay(True)
 536         time.sleep(30)
 537         self._confirm_single_replay()
 538         self.fs.set_allow_standby_replay(False)
 539         self._confirm_no_replay()
 540
 541     def test_standby_replay_singleton_fail(self):
 542         """
 543         That failures don't violate singleton constraint.
 544         """
 545
 546         self._confirm_no_replay()
 547         self.fs.set_allow_standby_replay(True)
 548         status = self._confirm_single_replay()
 549
 550         for i in range(10):
 551             time.sleep(randint(1, 5))
 552             self.fs.rank_restart(status=status)
 553             status = self._check_replay_takeover(status)
 554             status = self._confirm_single_replay(status=status)
 555
 556         for i in range(10):
 557             time.sleep(randint(1, 5))
 558             self.fs.rank_fail()
 559             status = self._check_replay_takeover(status)
 560             status = self._confirm_single_replay(status=status)
 561
 562     def test_standby_replay_singleton_fail_multimds(self):
 563         """
 564         That failures don't violate singleton constraint with multiple actives.
 565         """
 566
 567         status = self._confirm_no_replay()
 568         new_max_mds = randint(2, len(list(status.get_standbys())))
 569         self.fs.set_max_mds(new_max_mds)
 570         self.fs.wait_for_daemons() # wait for actives to come online!
 571         self.fs.set_allow_standby_replay(True)
 572         status = self._confirm_single_replay(full=False)
 573
 574         for i in range(10):
 575             time.sleep(randint(1, 5))
 576             victim = randint(0, new_max_mds-1)
 577             self.fs.rank_restart(rank=victim, status=status)
 578             status = self._check_replay_takeover(status, rank=victim)
 579             status = self._confirm_single_replay(status=status, full=False)
 580
 581         for i in range(10):
 582             time.sleep(randint(1, 5))
 583             victim = randint(0, new_max_mds-1)
 584             self.fs.rank_fail(rank=victim)
 585             status = self._check_replay_takeover(status, rank=victim)
 586             status = self._confirm_single_replay(status=status, full=False)
 587
 588     def test_standby_replay_failure(self):
 589         """
 590         That the failure of a standby-replay daemon happens cleanly
 591         and doesn't interrupt anything else.
 592         """
 593
 594         status = self._confirm_no_replay()
 595         self.fs.set_max_mds(1)
 596         self.fs.set_allow_standby_replay(True)
 597         status = self._confirm_single_replay()
 598
 599         for i in range(10):
 600             time.sleep(randint(1, 5))
 601             victim = self.fs.get_replay(status=status)
 602             self.fs.mds_restart(mds_id=victim['name'])
 603             status = self._confirm_single_replay(status=status)
 604
 605     def test_rank_stopped(self):
 606         """
 607         That when a rank is STOPPED, standby replays for
 608         that rank get torn down
 609         """
 610
 611         status = self._confirm_no_replay()
 612         standby_count = len(list(status.get_standbys()))
 613         self.fs.set_max_mds(2)
 614         self.fs.set_allow_standby_replay(True)
 615         status = self._confirm_single_replay()
 616
 617         self.fs.set_max_mds(1) # stop rank 1
 618
 619         status = self._confirm_single_replay()
 620         self.assertTrue(standby_count, len(list(status.get_standbys())))
 621
 622
 623 class TestMultiFilesystems(CephFSTestCase):
 624     CLIENTS_REQUIRED = 2
 625     MDSS_REQUIRED = 4
 626
 627     # We'll create our own filesystems and start our own daemons
 628     REQUIRE_FILESYSTEM = False
 629
 630     def setUp(self):
 631         super(TestMultiFilesystems, self).setUp()
 632         self.mds_cluster.mon_manager.raw_cluster_cmd("fs", "flag", "set",
 633             "enable_multiple", "true",
 634             "--yes-i-really-mean-it")
 635
 636     def _setup_two(self):
 637         fs_a = self.mds_cluster.newfs(name="alpha")
 638         fs_b = self.mds_cluster.newfs(name="bravo")
 639
 640         self.mds_cluster.mds_restart()
 641
 642         # Wait for both filesystems to go healthy
 643         fs_a.wait_for_daemons()
 644         fs_b.wait_for_daemons()
 645
 646         # Reconfigure client auth caps
 647         for mount in self.mounts:
 648             self.mds_cluster.mon_manager.raw_cluster_cmd_result(
 649                 'auth', 'caps', "client.{0}".format(mount.client_id),
 650                 'mds', 'allow',
 651                 'mon', 'allow r',
 652                 'osd', 'allow rw pool={0}, allow rw pool={1}'.format(
 653                     fs_a.get_data_pool_name(), fs_b.get_data_pool_name()))
 654
 655         return fs_a, fs_b
 656
 657     def test_clients(self):
 658         fs_a, fs_b = self._setup_two()
 659
 660         # Mount a client on fs_a
 661         self.mount_a.mount_wait(cephfs_name=fs_a.name)
 662         self.mount_a.write_n_mb("pad.bin", 1)
 663         self.mount_a.write_n_mb("test.bin", 2)
 664         a_created_ino = self.mount_a.path_to_ino("test.bin")
 665         self.mount_a.create_files()
 666
 667         # Mount a client on fs_b
 668         self.mount_b.mount_wait(cephfs_name=fs_b.name)
 669         self.mount_b.write_n_mb("test.bin", 1)
 670         b_created_ino = self.mount_b.path_to_ino("test.bin")
 671         self.mount_b.create_files()
 672
 673         # Check that a non-default filesystem mount survives an MDS
 674         # failover (i.e. that map subscription is continuous, not
 675         # just the first time), reproduces #16022
 676         old_fs_b_mds = fs_b.get_active_names()[0]
 677         self.mds_cluster.mds_stop(old_fs_b_mds)
 678         self.mds_cluster.mds_fail(old_fs_b_mds)
 679         fs_b.wait_for_daemons()
 680         background = self.mount_b.write_background()
 681         # Raise exception if the write doesn't finish (i.e. if client
 682         # has not kept up with MDS failure)
 683         try:
 684             self.wait_until_true(lambda: background.finished, timeout=30)
 685         except RuntimeError:
 686             # The mount is stuck, we'll have to force it to fail cleanly
 687             background.stdin.close()
 688             self.mount_b.umount_wait(force=True)
 689             raise
 690
 691         self.mount_a.umount_wait()
 692         self.mount_b.umount_wait()
 693
 694         # See that the client's files went into the correct pool
 695         self.assertTrue(fs_a.data_objects_present(a_created_ino, 1024 * 1024))
 696         self.assertTrue(fs_b.data_objects_present(b_created_ino, 1024 * 1024))
 697
 698     def test_standby(self):
 699         fs_a, fs_b = self._setup_two()
 700
 701         # Assert that the remaining two MDS daemons are now standbys
 702         a_daemons = fs_a.get_active_names()
 703         b_daemons = fs_b.get_active_names()
 704         self.assertEqual(len(a_daemons), 1)
 705         self.assertEqual(len(b_daemons), 1)
 706         original_a = a_daemons[0]
 707         original_b = b_daemons[0]
 708         expect_standby_daemons = set(self.mds_cluster.mds_ids) - (set(a_daemons) | set(b_daemons))
 709
 710         # Need all my standbys up as well as the active daemons
 711         self.wait_for_daemon_start()
 712         self.assertEqual(expect_standby_daemons, self.mds_cluster.get_standby_daemons())
 713
 714         # Kill fs_a's active MDS, see a standby take over
 715         self.mds_cluster.mds_stop(original_a)
 716         self.mds_cluster.mon_manager.raw_cluster_cmd("mds", "fail", original_a)
 717         self.wait_until_equal(lambda: len(fs_a.get_active_names()), 1, 30,
 718                               reject_fn=lambda v: v > 1)
 719         # Assert that it's a *different* daemon that has now appeared in the map for fs_a
 720         self.assertNotEqual(fs_a.get_active_names()[0], original_a)
 721
 722         # Kill fs_b's active MDS, see a standby take over
 723         self.mds_cluster.mds_stop(original_b)
 724         self.mds_cluster.mon_manager.raw_cluster_cmd("mds", "fail", original_b)
 725         self.wait_until_equal(lambda: len(fs_b.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_b.get_active_names()[0], original_b)
 729
 730         # Both of the original active daemons should be gone, and all standbys used up
 731         self.assertEqual(self.mds_cluster.get_standby_daemons(), set())
 732
 733         # Restart the ones I killed, see them reappear as standbys
 734         self.mds_cluster.mds_restart(original_a)
 735         self.mds_cluster.mds_restart(original_b)
 736         self.wait_until_true(
 737             lambda: {original_a, original_b} == self.mds_cluster.get_standby_daemons(),
 738             timeout=30
 739         )
 740
 741     def test_grow_shrink(self):
 742         # Usual setup...
 743         fs_a, fs_b = self._setup_two()
 744
 745         # Increase max_mds on fs_b, see a standby take up the role
 746         fs_b.set_max_mds(2)
 747         self.wait_until_equal(lambda: len(fs_b.get_active_names()), 2, 30,
 748                               reject_fn=lambda v: v > 2 or v < 1)
 749
 750         # Increase max_mds on fs_a, see a standby take up the role
 751         fs_a.set_max_mds(2)
 752         self.wait_until_equal(lambda: len(fs_a.get_active_names()), 2, 30,
 753                               reject_fn=lambda v: v > 2 or v < 1)
 754
 755         # Shrink fs_b back to 1, see a daemon go back to standby
 756         fs_b.set_max_mds(1)
 757         self.wait_until_equal(lambda: len(fs_b.get_active_names()), 1, 30,
 758                               reject_fn=lambda v: v > 2 or v < 1)
 759
 760         # Grow fs_a up to 3, see the former fs_b daemon join it.
 761         fs_a.set_max_mds(3)
 762         self.wait_until_equal(lambda: len(fs_a.get_active_names()), 3, 60,
 763                               reject_fn=lambda v: v > 3 or v < 2)