]>
git.proxmox.com Git - ceph.git/blob - ceph/qa/tasks/cephfs/test_full.py
4 from textwrap
import dedent
7 from typing
import Optional
9 # make it work for python2
11 from teuthology
.orchestra
.run
import CommandFailedError
12 from tasks
.cephfs
.fuse_mount
import FuseMount
13 from tasks
.cephfs
.cephfs_test_case
import CephFSTestCase
16 log
= logging
.getLogger(__name__
)
19 class FullnessTestCase(CephFSTestCase
):
22 # Subclasses define whether they're filling whole cluster or just data pool
25 # Subclasses define how many bytes should be written to achieve fullness
26 pool_capacity
= None # type: Optional[int]
29 # Subclasses define what fullness means to them
31 raise NotImplementedError()
34 CephFSTestCase
.setUp(self
)
36 mds_status
= self
.fs
.rank_asok(["status"])
38 # Capture the initial OSD map epoch for later use
39 self
.initial_osd_epoch
= mds_status
['osdmap_epoch_barrier']
41 def test_barrier(self
):
43 That when an OSD epoch barrier is set on an MDS, subsequently
44 issued capabilities cause clients to update their OSD map to that
48 # script that sync up client with MDS OSD map barrier. The barrier should
49 # be updated by cap flush ack message.
52 fd = os.open("{path}", os.O_CREAT | os.O_RDWR, 0O600)
58 # Sync up client with initial MDS OSD map barrier.
59 path
= os
.path
.join(self
.mount_a
.mountpoint
, "foo")
60 self
.mount_a
.run_python(pyscript
.format(path
=path
))
62 # Grab mounts' initial OSD epochs: later we will check that
63 # it hasn't advanced beyond this point.
64 mount_a_initial_epoch
, mount_a_initial_barrier
= self
.mount_a
.get_osd_epoch()
66 # Freshly mounted at start of test, should be up to date with OSD map
67 self
.assertGreaterEqual(mount_a_initial_epoch
, self
.initial_osd_epoch
)
69 # Set and unset a flag to cause OSD epoch to increment
70 self
.fs
.mon_manager
.raw_cluster_cmd("osd", "set", "pause")
71 self
.fs
.mon_manager
.raw_cluster_cmd("osd", "unset", "pause")
73 out
= self
.fs
.mon_manager
.raw_cluster_cmd("osd", "dump", "--format=json").strip()
74 new_epoch
= json
.loads(out
)['epoch']
75 self
.assertNotEqual(self
.initial_osd_epoch
, new_epoch
)
77 # Do a metadata operation on clients, witness that they end up with
78 # the old OSD map from startup time (nothing has prompted client
80 path
= os
.path
.join(self
.mount_a
.mountpoint
, "foo")
81 self
.mount_a
.run_python(pyscript
.format(path
=path
))
82 mount_a_epoch
, mount_a_barrier
= self
.mount_a
.get_osd_epoch()
83 self
.assertEqual(mount_a_epoch
, mount_a_initial_epoch
)
84 self
.assertEqual(mount_a_barrier
, mount_a_initial_barrier
)
86 # Set a barrier on the MDS
87 self
.fs
.rank_asok(["osdmap", "barrier", new_epoch
.__str
__()])
89 # Sync up client with new MDS OSD map barrier
90 path
= os
.path
.join(self
.mount_a
.mountpoint
, "baz")
91 self
.mount_a
.run_python(pyscript
.format(path
=path
))
92 mount_a_epoch
, mount_a_barrier
= self
.mount_a
.get_osd_epoch()
93 self
.assertEqual(mount_a_barrier
, new_epoch
)
95 # Some time passes here because the metadata part of the operation
96 # completes immediately, while the resulting OSD map update happens
97 # asynchronously (it's an Objecter::_maybe_request_map) as a result
98 # of seeing the new epoch barrier.
100 lambda: self
.mount_a
.get_osd_epoch()[0] >= new_epoch
,
103 def _data_pool_name(self
):
104 data_pool_names
= self
.fs
.get_data_pool_names()
105 if len(data_pool_names
) > 1:
106 raise RuntimeError("This test can't handle multiple data pools")
108 return data_pool_names
[0]
110 def _test_full(self
, easy_case
):
112 - That a client trying to write data to a file is prevented
113 from doing so with an -EFULL result
114 - That they are also prevented from creating new files by the MDS.
115 - That they may delete another file to get the system healthy again
117 :param easy_case: if true, delete a successfully written file to
118 free up space. else, delete the file that experienced
122 osd_mon_report_interval
= int(self
.fs
.get_config("osd_mon_report_interval", service_type
='osd'))
124 log
.info("Writing {0}MB should fill this cluster".format(self
.fill_mb
))
126 # Fill up the cluster. This dd may or may not fail, as it depends on
127 # how soon the cluster recognises its own fullness
128 self
.mount_a
.write_n_mb("large_file_a", self
.fill_mb
/ 2)
130 self
.mount_a
.write_n_mb("large_file_b", self
.fill_mb
/ 2)
131 except CommandFailedError
:
132 log
.info("Writing file B failed (full status happened already)")
133 assert self
.is_full()
135 log
.info("Writing file B succeeded (full status will happen soon)")
136 self
.wait_until_true(lambda: self
.is_full(),
137 timeout
=osd_mon_report_interval
* 5)
139 # Attempting to write more data should give me ENOSPC
140 with self
.assertRaises(CommandFailedError
) as ar
:
141 self
.mount_a
.write_n_mb("large_file_b", 50, seek
=self
.fill_mb
/ 2)
142 self
.assertEqual(ar
.exception
.exitstatus
, 1) # dd returns 1 on "No space"
144 # Wait for the MDS to see the latest OSD map so that it will reliably
145 # be applying the policy of rejecting non-deletion metadata operations
146 # while in the full state.
147 osd_epoch
= json
.loads(self
.fs
.mon_manager
.raw_cluster_cmd("osd", "dump", "--format=json-pretty"))['epoch']
148 self
.wait_until_true(
149 lambda: self
.fs
.rank_asok(['status'])['osdmap_epoch'] >= osd_epoch
,
152 if not self
.data_only
:
153 with self
.assertRaises(CommandFailedError
):
154 self
.mount_a
.write_n_mb("small_file_1", 0)
156 # Clear out some space
158 self
.mount_a
.run_shell(['rm', '-f', 'large_file_a'])
159 self
.mount_a
.run_shell(['rm', '-f', 'large_file_b'])
161 # In the hard case it is the file that filled the system.
162 # Before the new #7317 (ENOSPC, epoch barrier) changes, this
163 # would fail because the last objects written would be
164 # stuck in the client cache as objecter operations.
165 self
.mount_a
.run_shell(['rm', '-f', 'large_file_b'])
166 self
.mount_a
.run_shell(['rm', '-f', 'large_file_a'])
168 # Here we are waiting for two things to happen:
169 # * The MDS to purge the stray folder and execute object deletions
170 # * The OSDs to inform the mon that they are no longer full
171 self
.wait_until_true(lambda: not self
.is_full(),
172 timeout
=osd_mon_report_interval
* 5)
174 # Wait for the MDS to see the latest OSD map so that it will reliably
175 # be applying the free space policy
176 osd_epoch
= json
.loads(self
.fs
.mon_manager
.raw_cluster_cmd("osd", "dump", "--format=json-pretty"))['epoch']
177 self
.wait_until_true(
178 lambda: self
.fs
.rank_asok(['status'])['osdmap_epoch'] >= osd_epoch
,
181 # Now I should be able to write again
182 self
.mount_a
.write_n_mb("large_file", 50, seek
=0)
184 # Ensure that the MDS keeps its OSD epoch barrier across a restart
186 def test_full_different_file(self
):
187 self
._test
_full
(True)
189 def test_full_same_file(self
):
190 self
._test
_full
(False)
192 def _remote_write_test(self
, template
):
194 Run some remote python in a way that's useful for
195 testing free space behaviour (see test_* methods using this)
197 file_path
= os
.path
.join(self
.mount_a
.mountpoint
, "full_test_file")
199 # Enough to trip the full flag
200 osd_mon_report_interval
= int(self
.fs
.get_config("osd_mon_report_interval", service_type
='osd'))
201 mon_tick_interval
= int(self
.fs
.get_config("mon_tick_interval", service_type
="mon"))
203 # Sufficient data to cause RADOS cluster to go 'full'
204 log
.info("pool capacity {0}, {1}MB should be enough to fill it".format(self
.pool_capacity
, self
.fill_mb
))
206 # Long enough for RADOS cluster to notice it is full and set flag on mons
207 # (report_interval for mon to learn PG stats, tick interval for it to update OSD map,
208 # factor of 1.5 for I/O + network latency in committing OSD map and distributing it
210 full_wait
= (osd_mon_report_interval
+ mon_tick_interval
) * 1.5
212 # Configs for this test should bring this setting down in order to
213 # run reasonably quickly
214 if osd_mon_report_interval
> 10:
215 log
.warn("This test may run rather slowly unless you decrease"
216 "osd_mon_report_interval (5 is a good setting)!")
218 self
.mount_a
.run_python(template
.format(
219 fill_mb
=self
.fill_mb
,
222 is_fuse
=isinstance(self
.mount_a
, FuseMount
)
225 def test_full_fclose(self
):
226 # A remote script which opens a file handle, fills up the filesystem, and then
227 # checks that ENOSPC errors on buffered writes appear correctly as errors in fsync
228 remote_script
= dedent("""
234 # Write some buffered data through before going full, all should be well
235 print("writing some data through which we expect to succeed")
237 f = os.open("{file_path}", os.O_WRONLY | os.O_CREAT)
238 bytes += os.write(f, b'a' * 512 * 1024)
240 print("fsync'ed data successfully, will now attempt to fill fs")
242 # Okay, now we're going to fill up the filesystem, and then keep
243 # writing until we see an error from fsync. As long as we're doing
244 # buffered IO, the error should always only appear from fsync and not
248 for n in range(0, int({fill_mb} * 0.9)):
249 bytes += os.write(f, b'x' * 1024 * 1024)
250 print("wrote {{0}} bytes via buffered write, may repeat".format(bytes))
251 print("done writing {{0}} bytes".format(bytes))
253 # OK, now we should sneak in under the full condition
254 # due to the time it takes the OSDs to report to the
255 # mons, and get a successful fsync on our full-making data
257 print("successfully fsync'ed prior to getting full state reported")
259 # buffered write, add more dirty data to the buffer
260 print("starting buffered write")
262 for n in range(0, int({fill_mb} * 0.2)):
263 bytes += os.write(f, b'x' * 1024 * 1024)
264 print("sleeping a bit as we've exceeded 90% of our expected full ratio")
265 time.sleep({full_wait})
269 print("wrote, now waiting 30s and then doing a close we expect to fail")
271 # Wait long enough for a background flush that should fail
275 # ...and check that the failed background flush is reflected in fclose
279 print("close() returned an error as expected")
281 raise RuntimeError("close() failed to raise error")
283 # The kernel cephfs client does not raise errors on fclose
286 os.unlink("{file_path}")
288 self
._remote
_write
_test
(remote_script
)
290 def test_full_fsync(self
):
292 That when the full flag is encountered during asynchronous
293 flushes, such that an fwrite() succeeds but an fsync/fclose()
294 should return the ENOSPC error.
297 # A remote script which opens a file handle, fills up the filesystem, and then
298 # checks that ENOSPC errors on buffered writes appear correctly as errors in fsync
299 remote_script
= dedent("""
305 # Write some buffered data through before going full, all should be well
306 print("writing some data through which we expect to succeed")
308 f = os.open("{file_path}", os.O_WRONLY | os.O_CREAT)
309 bytes += os.write(f, b'a' * 4096)
311 print("fsync'ed data successfully, will now attempt to fill fs")
313 # Okay, now we're going to fill up the filesystem, and then keep
314 # writing until we see an error from fsync. As long as we're doing
315 # buffered IO, the error should always only appear from fsync and not
319 for n in range(0, int({fill_mb} * 1.1)):
321 bytes += os.write(f, b'x' * 1024 * 1024)
322 print("wrote bytes via buffered write, moving on to fsync")
324 print("Unexpected error %s from write() instead of fsync()" % e)
329 print("fsync'ed successfully")
331 print("Reached fullness after %.2f MB" % (bytes / (1024.0 * 1024.0)))
335 print("Not full yet after %.2f MB" % (bytes / (1024.0 * 1024.0)))
337 if n > {fill_mb} * 0.9:
338 # Be cautious in the last region where we expect to hit
339 # the full condition, so that we don't overshoot too dramatically
340 print("sleeping a bit as we've exceeded 90% of our expected full ratio")
341 time.sleep({full_wait})
344 raise RuntimeError("Failed to reach fullness after writing %d bytes" % bytes)
346 # close() should not raise an error because we already caught it in
347 # fsync. There shouldn't have been any more writeback errors
348 # since then because all IOs got cancelled on the full flag.
349 print("calling close")
351 print("close() did not raise error")
353 os.unlink("{file_path}")
356 self
._remote
_write
_test
(remote_script
)
359 class TestQuotaFull(FullnessTestCase
):
361 Test per-pool fullness, which indicates quota limits exceeded
363 pool_capacity
= 1024 * 1024 * 32 # arbitrary low-ish limit
364 fill_mb
= pool_capacity
/ (1024 * 1024) # type: ignore
366 # We are only testing quota handling on the data pool, not the metadata
371 super(TestQuotaFull
, self
).setUp()
373 pool_name
= self
.fs
.get_data_pool_name()
374 self
.fs
.mon_manager
.raw_cluster_cmd("osd", "pool", "set-quota", pool_name
,
375 "max_bytes", "{0}".format(self
.pool_capacity
))
378 return self
.fs
.is_full()
381 class TestClusterFull(FullnessTestCase
):
383 Test data pool fullness, which indicates that an OSD has become too full
386 REQUIRE_MEMSTORE
= True
389 super(TestClusterFull
, self
).setUp()
391 if self
.pool_capacity
is None:
392 max_avail
= self
.fs
.get_pool_df(self
._data
_pool
_name
())['max_avail']
393 full_ratio
= float(self
.fs
.get_config("mon_osd_full_ratio", service_type
="mon"))
394 TestClusterFull
.pool_capacity
= int(max_avail
* full_ratio
)
395 TestClusterFull
.fill_mb
= (self
.pool_capacity
/ (1024 * 1024))
398 return self
.fs
.is_full()
400 # Hide the parent class so that unittest.loader doesn't try to run it.
401 del globals()['FullnessTestCase']