ceph/qa/tasks/cephfs/test_misc.py

   1
   2 from unittest import SkipTest
   3 from tasks.cephfs.fuse_mount import FuseMount
   4 from tasks.cephfs.cephfs_test_case import CephFSTestCase
   5 from teuthology.orchestra.run import CommandFailedError, ConnectionLostError
   6 import errno
   7 import time
   8 import json
   9 import logging
  10 import time
  11
  12 log = logging.getLogger(__name__)
  13
  14 class TestMisc(CephFSTestCase):
  15     CLIENTS_REQUIRED = 2
  16
  17     def test_getattr_caps(self):
  18         """
  19         Check if MDS recognizes the 'mask' parameter of open request.
  20         The parameter allows client to request caps when opening file
  21         """
  22
  23         if not isinstance(self.mount_a, FuseMount):
  24             raise SkipTest("Require FUSE client")
  25
  26         # Enable debug. Client will requests CEPH_CAP_XATTR_SHARED
  27         # on lookup/open
  28         self.mount_b.umount_wait()
  29         self.set_conf('client', 'client debug getattr caps', 'true')
  30         self.mount_b.mount()
  31         self.mount_b.wait_until_mounted()
  32
  33         # create a file and hold it open. MDS will issue CEPH_CAP_EXCL_*
  34         # to mount_a
  35         p = self.mount_a.open_background("testfile")
  36         self.mount_b.wait_for_visible("testfile")
  37
  38         # this triggers a lookup request and an open request. The debug
  39         # code will check if lookup/open reply contains xattrs
  40         self.mount_b.run_shell(["cat", "testfile"])
  41
  42         self.mount_a.kill_background(p)
  43
  44     def test_root_rctime(self):
  45         """
  46         Check that the root inode has a non-default rctime on startup.
  47         """
  48
  49         t = time.time()
  50         rctime = self.mount_a.getfattr(".", "ceph.dir.rctime")
  51         log.info("rctime = {}".format(rctime))
  52         self.assertGreaterEqual(rctime, t-10)
  53
  54     def test_fs_new(self):
  55         self.mount_a.umount_wait()
  56         self.mount_b.umount_wait()
  57
  58         data_pool_name = self.fs.get_data_pool_name()
  59
  60         self.fs.mds_stop()
  61         self.fs.mds_fail()
  62
  63         self.fs.mon_manager.raw_cluster_cmd('fs', 'rm', self.fs.name,
  64                                             '--yes-i-really-mean-it')
  65
  66         self.fs.mon_manager.raw_cluster_cmd('osd', 'pool', 'delete',
  67                                             self.fs.metadata_pool_name,
  68                                             self.fs.metadata_pool_name,
  69                                             '--yes-i-really-really-mean-it')
  70         self.fs.mon_manager.raw_cluster_cmd('osd', 'pool', 'create',
  71                                             self.fs.metadata_pool_name,
  72                                             self.fs.get_pgs_per_fs_pool().__str__())
  73
  74         dummyfile = '/etc/fstab'
  75
  76         self.fs.put_metadata_object_raw("key", dummyfile)
  77
  78         def get_pool_df(fs, name):
  79             try:
  80                 return fs.get_pool_df(name)['objects'] > 0
  81             except RuntimeError as e:
  82                 return False
  83
  84         self.wait_until_true(lambda: get_pool_df(self.fs, self.fs.metadata_pool_name), timeout=30)
  85
  86         try:
  87             self.fs.mon_manager.raw_cluster_cmd('fs', 'new', self.fs.name,
  88                                                 self.fs.metadata_pool_name,
  89                                                 data_pool_name)
  90         except CommandFailedError as e:
  91             self.assertEqual(e.exitstatus, errno.EINVAL)
  92         else:
  93             raise AssertionError("Expected EINVAL")
  94
  95         self.fs.mon_manager.raw_cluster_cmd('fs', 'new', self.fs.name,
  96                                             self.fs.metadata_pool_name,
  97                                             data_pool_name, "--force")
  98
  99         self.fs.mon_manager.raw_cluster_cmd('fs', 'rm', self.fs.name,
 100                                             '--yes-i-really-mean-it')
 101
 102
 103         self.fs.mon_manager.raw_cluster_cmd('osd', 'pool', 'delete',
 104                                             self.fs.metadata_pool_name,
 105                                             self.fs.metadata_pool_name,
 106                                             '--yes-i-really-really-mean-it')
 107         self.fs.mon_manager.raw_cluster_cmd('osd', 'pool', 'create',
 108                                             self.fs.metadata_pool_name,
 109                                             self.fs.get_pgs_per_fs_pool().__str__())
 110         self.fs.mon_manager.raw_cluster_cmd('fs', 'new', self.fs.name,
 111                                             self.fs.metadata_pool_name,
 112                                             data_pool_name)
 113
 114     def test_cap_revoke_nonresponder(self):
 115         """
 116         Check that a client is evicted if it has not responded to cap revoke
 117         request for configured number of seconds.
 118         """
 119         session_timeout = self.fs.get_var("session_timeout")
 120         eviction_timeout = session_timeout / 2.0
 121
 122         self.fs.mds_asok(['config', 'set', 'mds_cap_revoke_eviction_timeout',
 123                           str(eviction_timeout)])
 124
 125         cap_holder = self.mount_a.open_background()
 126
 127         # Wait for the file to be visible from another client, indicating
 128         # that mount_a has completed its network ops
 129         self.mount_b.wait_for_visible()
 130
 131         # Simulate client death
 132         self.mount_a.kill()
 133
 134         try:
 135             # The waiter should get stuck waiting for the capability
 136             # held on the MDS by the now-dead client A
 137             cap_waiter = self.mount_b.write_background()
 138
 139             a = time.time()
 140             time.sleep(eviction_timeout)
 141             cap_waiter.wait()
 142             b = time.time()
 143             cap_waited = b - a
 144             log.info("cap_waiter waited {0}s".format(cap_waited))
 145
 146             # check if the cap is transferred before session timeout kicked in.
 147             # this is a good enough check to ensure that the client got evicted
 148             # by the cap auto evicter rather than transitioning to stale state
 149             # and then getting evicted.
 150             self.assertLess(cap_waited, session_timeout,
 151                             "Capability handover took {0}, expected less than {1}".format(
 152                                 cap_waited, session_timeout
 153                             ))
 154
 155             self.assertTrue(self.mount_a.is_blacklisted())
 156             cap_holder.stdin.close()
 157             try:
 158                 cap_holder.wait()
 159             except (CommandFailedError, ConnectionLostError):
 160                 # We killed it (and possibly its node), so it raises an error
 161                 pass
 162         finally:
 163             self.mount_a.kill_cleanup()
 164
 165         self.mount_a.mount()
 166         self.mount_a.wait_until_mounted()
 167
 168     def test_filtered_df(self):
 169         pool_name = self.fs.get_data_pool_name()
 170         raw_df = self.fs.get_pool_df(pool_name)
 171         raw_avail = float(raw_df["max_avail"])
 172         out = self.fs.mon_manager.raw_cluster_cmd('osd', 'pool', 'get',
 173                                                   pool_name, 'size',
 174                                                   '-f', 'json-pretty')
 175         j = json.loads(out)
 176         pool_size = int(j['size'])
 177
 178         proc = self.mount_a.run_shell(['df', '.'])
 179         output = proc.stdout.getvalue()
 180         fs_avail = output.split('\n')[1].split()[3]
 181         fs_avail = float(fs_avail) * 1024
 182
 183         ratio = raw_avail / fs_avail
 184         assert 0.9 < ratio < 1.1
 185
 186     def test_dump_inode(self):
 187         info = self.fs.mds_asok(['dump', 'inode', '1'])
 188         assert(info['path'] == "/")
 189
 190     def test_dump_inode_hexademical(self):
 191         self.mount_a.run_shell(["mkdir", "-p", "foo"])
 192         ino = self.mount_a.path_to_ino("foo")
 193         assert type(ino) is int
 194         info = self.fs.mds_asok(['dump', 'inode', hex(ino)])
 195         assert info['path'] == "/foo"
 196
 197
 198 class TestCacheDrop(CephFSTestCase):
 199     CLIENTS_REQUIRED = 1
 200
 201     def _run_drop_cache_cmd(self, timeout=None):
 202         result = None
 203         mds_id = self.fs.get_lone_mds_id()
 204         if timeout is not None:
 205             result = self.fs.mon_manager.raw_cluster_cmd("tell", "mds.{0}".format(mds_id),
 206                                                     "cache", "drop", str(timeout))
 207         else:
 208             result = self.fs.mon_manager.raw_cluster_cmd("tell", "mds.{0}".format(mds_id),
 209                                                     "cache", "drop")
 210         return json.loads(result)
 211
 212     def _setup(self, max_caps=20, threshold=400):
 213         # create some files
 214         self.mount_a.create_n_files("dc-dir/dc-file", 1000, sync=True)
 215
 216         # Reduce this so the MDS doesn't rkcall the maximum for simple tests
 217         self.fs.rank_asok(['config', 'set', 'mds_recall_max_caps', str(max_caps)])
 218         self.fs.rank_asok(['config', 'set', 'mds_recall_max_decay_threshold', str(threshold)])
 219
 220     def test_drop_cache_command(self):
 221         """
 222         Basic test for checking drop cache command.
 223         Confirm it halts without a timeout.
 224         Note that the cache size post trimming is not checked here.
 225         """
 226         mds_min_caps_per_client = int(self.fs.get_config("mds_min_caps_per_client"))
 227         self._setup()
 228         result = self._run_drop_cache_cmd()
 229         self.assertEqual(result['client_recall']['return_code'], 0)
 230         self.assertEqual(result['flush_journal']['return_code'], 0)
 231         # It should take at least 1 second
 232         self.assertGreater(result['duration'], 1)
 233         self.assertGreaterEqual(result['trim_cache']['trimmed'], 1000-2*mds_min_caps_per_client)
 234
 235     def test_drop_cache_command_timeout(self):
 236         """
 237         Basic test for checking drop cache command.
 238         Confirm recall halts early via a timeout.
 239         Note that the cache size post trimming is not checked here.
 240         """
 241         self._setup()
 242         result = self._run_drop_cache_cmd(timeout=10)
 243         self.assertEqual(result['client_recall']['return_code'], -errno.ETIMEDOUT)
 244         self.assertEqual(result['flush_journal']['return_code'], 0)
 245         self.assertGreater(result['duration'], 10)
 246         self.assertGreaterEqual(result['trim_cache']['trimmed'], 100) # we did something, right?
 247
 248     def test_drop_cache_command_dead_timeout(self):
 249         """
 250         Check drop cache command with non-responding client using tell
 251         interface. Note that the cache size post trimming is not checked
 252         here.
 253         """
 254         self._setup()
 255         self.mount_a.kill()
 256         # Note: recall is subject to the timeout. The journal flush will
 257         # be delayed due to the client being dead.
 258         result = self._run_drop_cache_cmd(timeout=5)
 259         self.assertEqual(result['client_recall']['return_code'], -errno.ETIMEDOUT)
 260         self.assertEqual(result['flush_journal']['return_code'], 0)
 261         self.assertGreater(result['duration'], 5)
 262         self.assertLess(result['duration'], 120)
 263         # Note: result['trim_cache']['trimmed'] may be >0 because dropping the
 264         # cache now causes the Locker to drive eviction of stale clients (a
 265         # stale session will be autoclosed at mdsmap['session_timeout']). The
 266         # particular operation causing this is journal flush which causes the
 267         # MDS to wait wait for cap revoke.
 268         #self.assertEqual(0, result['trim_cache']['trimmed'])
 269         self.mount_a.kill_cleanup()
 270         self.mount_a.mount()
 271         self.mount_a.wait_until_mounted()
 272
 273     def test_drop_cache_command_dead(self):
 274         """
 275         Check drop cache command with non-responding client using tell
 276         interface. Note that the cache size post trimming is not checked
 277         here.
 278         """
 279         self._setup()
 280         self.mount_a.kill()
 281         result = self._run_drop_cache_cmd()
 282         self.assertEqual(result['client_recall']['return_code'], 0)
 283         self.assertEqual(result['flush_journal']['return_code'], 0)
 284         self.assertGreater(result['duration'], 5)
 285         self.assertLess(result['duration'], 120)
 286         # Note: result['trim_cache']['trimmed'] may be >0 because dropping the
 287         # cache now causes the Locker to drive eviction of stale clients (a
 288         # stale session will be autoclosed at mdsmap['session_timeout']). The
 289         # particular operation causing this is journal flush which causes the
 290         # MDS to wait wait for cap revoke.
 291         self.mount_a.kill_cleanup()
 292         self.mount_a.mount()
 293         self.mount_a.wait_until_mounted()