3 Exercise the MDS's behaviour when clients and the MDCache reach or
4 exceed the limits of how many caps/inodes they should hold.
8 from textwrap
import dedent
9 from teuthology
.orchestra
.run
import CommandFailedError
10 from tasks
.cephfs
.cephfs_test_case
import CephFSTestCase
, needs_trimming
11 from tasks
.cephfs
.fuse_mount
import FuseMount
15 log
= logging
.getLogger(__name__
)
18 # Arbitrary timeouts for operations involving restarting
19 # an MDS or waiting for it to come up
20 MDS_RESTART_GRACE
= 60
22 # Hardcoded values from Server::recall_client_state
23 CAP_RECALL_RATIO
= 0.8
27 class TestClientLimits(CephFSTestCase
):
28 REQUIRE_KCLIENT_REMOTE
= True
31 def _test_client_pin(self
, use_subdir
, open_files
):
33 When a client pins an inode in its cache, for example because the file is held open,
34 it should reject requests from the MDS to trim these caps. The MDS should complain
35 to the user that it is unable to enforce its cache size limits because of this
38 :param use_subdir: whether to put test files in a subdir or use root
41 # Set MDS cache memory limit to a low value that will make the MDS to
42 # ask the client to trim the caps.
43 cache_memory_limit
= "1K"
45 self
.set_conf('mds', 'mds_cache_memory_limit', cache_memory_limit
)
46 self
.set_conf('mds', 'mds_recall_max_caps', open_files
/2)
47 self
.set_conf('mds', 'mds_recall_warning_threshold', open_files
)
48 self
.fs
.mds_fail_restart()
49 self
.fs
.wait_for_daemons()
51 mds_min_caps_per_client
= int(self
.fs
.get_config("mds_min_caps_per_client"))
52 mds_max_caps_per_client
= int(self
.fs
.get_config("mds_max_caps_per_client"))
53 mds_recall_warning_decay_rate
= float(self
.fs
.get_config("mds_recall_warning_decay_rate"))
54 self
.assertTrue(open_files
>= mds_min_caps_per_client
)
56 mount_a_client_id
= self
.mount_a
.get_global_id()
57 path
= "subdir/mount_a" if use_subdir
else "mount_a"
58 open_proc
= self
.mount_a
.open_n_background(path
, open_files
)
60 # Client should now hold:
61 # `open_files` caps for the open files
64 self
.wait_until_equal(lambda: self
.get_session(mount_a_client_id
)['num_caps'],
65 open_files
+ (2 if use_subdir
else 1),
67 reject_fn
=lambda x
: x
> open_files
+ 2)
69 # MDS should not be happy about that, as the client is failing to comply
70 # with the SESSION_RECALL messages it is being sent
71 self
.wait_for_health("MDS_CLIENT_RECALL", mds_recall_warning_decay_rate
*2)
73 # We can also test that the MDS health warning for oversized
74 # cache is functioning as intended.
75 self
.wait_for_health("MDS_CACHE_OVERSIZED", mds_recall_warning_decay_rate
*2)
77 # When the client closes the files, it should retain only as many caps as allowed
78 # under the SESSION_RECALL policy
79 log
.info("Terminating process holding files open")
80 open_proc
.stdin
.close()
83 except CommandFailedError
:
84 # We killed it, so it raises an error
87 # The remaining caps should comply with the numbers sent from MDS in SESSION_RECALL message,
88 # which depend on the caps outstanding, cache size and overall ratio
90 num_caps
= self
.get_session(mount_a_client_id
)['num_caps']
91 if num_caps
<= mds_min_caps_per_client
:
93 elif num_caps
<= mds_max_caps_per_client
:
98 self
.wait_until_true(expected_caps
, timeout
=60)
101 def test_client_pin_root(self
):
102 self
._test
_client
_pin
(False, 400)
105 def test_client_pin(self
):
106 self
._test
_client
_pin
(True, 800)
109 def test_client_pin_mincaps(self
):
110 self
._test
_client
_pin
(True, 200)
112 def test_client_release_bug(self
):
114 When a client has a bug (which we will simulate) preventing it from releasing caps,
115 the MDS should notice that releases are not being sent promptly, and generate a health
116 metric to that effect.
119 # The debug hook to inject the failure only exists in the fuse client
120 if not isinstance(self
.mount_a
, FuseMount
):
121 self
.skipTest("Require FUSE client to inject client release failure")
123 self
.set_conf('client.{0}'.format(self
.mount_a
.client_id
), 'client inject release failure', 'true')
124 self
.mount_a
.teardown()
125 self
.mount_a
.mount_wait()
126 mount_a_client_id
= self
.mount_a
.get_global_id()
128 # Client A creates a file. He will hold the write caps on the file, and later (simulated bug) fail
129 # to comply with the MDSs request to release that cap
130 self
.mount_a
.run_shell(["touch", "file1"])
132 # Client B tries to stat the file that client A created
133 rproc
= self
.mount_b
.write_background("file1")
135 # After session_timeout, we should see a health warning (extra lag from
137 session_timeout
= self
.fs
.get_var("session_timeout")
138 self
.wait_for_health("MDS_CLIENT_LATE_RELEASE", session_timeout
+ 10)
140 # Client B should still be stuck
141 self
.assertFalse(rproc
.finished
)
145 self
.mount_a
.kill_cleanup()
147 # Client B should complete
148 self
.fs
.mds_asok(['session', 'evict', "%s" % mount_a_client_id
])
151 def test_client_oldest_tid(self
):
153 When a client does not advance its oldest tid, the MDS should notice that
154 and generate health warnings.
157 # num of requests client issues
160 # The debug hook to inject the failure only exists in the fuse client
161 if not isinstance(self
.mount_a
, FuseMount
):
162 self
.skipTest("Require FUSE client to inject client release failure")
164 self
.set_conf('client', 'client inject fixed oldest tid', 'true')
165 self
.mount_a
.teardown()
166 self
.mount_a
.mount_wait()
168 self
.fs
.mds_asok(['config', 'set', 'mds_max_completed_requests', '{0}'.format(max_requests
)])
170 # Create lots of files
171 self
.mount_a
.create_n_files("testdir/file1", max_requests
+ 100)
173 # Create a few files synchronously. This makes sure previous requests are completed
174 self
.mount_a
.create_n_files("testdir/file2", 5, True)
176 # Wait for the health warnings. Assume mds can handle 10 request per second at least
177 self
.wait_for_health("MDS_CLIENT_OLDEST_TID", max_requests
// 10)
179 def _test_client_cache_size(self
, mount_subdir
):
181 check if client invalidate kernel dcache according to its cache size config
184 # The debug hook to inject the failure only exists in the fuse client
185 if not isinstance(self
.mount_a
, FuseMount
):
186 self
.skipTest("Require FUSE client to inject client release failure")
189 # fuse assigns a fix inode number (1) to root inode. But in mounting into
190 # subdir case, the actual inode number of root is not 1. This mismatch
191 # confuses fuse_lowlevel_notify_inval_entry() when invalidating dentries
193 self
.mount_a
.run_shell(["mkdir", "subdir"])
194 self
.mount_a
.umount_wait()
195 self
.set_conf('client', 'client mountpoint', '/subdir')
196 self
.mount_a
.mount_wait()
197 root_ino
= self
.mount_a
.path_to_ino(".")
198 self
.assertEqual(root_ino
, 1);
200 dir_path
= os
.path
.join(self
.mount_a
.mountpoint
, "testdir")
202 mkdir_script
= dedent("""
205 for n in range(0, {num_dirs}):
206 os.mkdir("{path}/dir{{0}}".format(n))
210 self
.mount_a
.run_python(mkdir_script
.format(path
=dir_path
, num_dirs
=num_dirs
))
211 self
.mount_a
.run_shell(["sync"])
213 dentry_count
, dentry_pinned_count
= self
.mount_a
.get_dentry_count()
214 self
.assertGreaterEqual(dentry_count
, num_dirs
)
215 self
.assertGreaterEqual(dentry_pinned_count
, num_dirs
)
217 cache_size
= num_dirs
// 10
218 self
.mount_a
.set_cache_size(cache_size
)
221 dentry_count
, dentry_pinned_count
= self
.mount_a
.get_dentry_count()
222 log
.info("waiting, dentry_count, dentry_pinned_count: {0}, {1}".format(
223 dentry_count
, dentry_pinned_count
225 if dentry_count
> cache_size
or dentry_pinned_count
> cache_size
:
230 self
.wait_until_true(trimmed
, 30)
233 def test_client_cache_size(self
):
234 self
._test
_client
_cache
_size
(False)
235 self
._test
_client
_cache
_size
(True)
237 def test_client_max_caps(self
):
239 That the MDS will not let a client sit above mds_max_caps_per_client caps.
242 mds_min_caps_per_client
= int(self
.fs
.get_config("mds_min_caps_per_client"))
243 mds_max_caps_per_client
= 2*mds_min_caps_per_client
244 self
.set_conf('mds', 'mds_max_caps_per_client', mds_max_caps_per_client
)
245 self
.fs
.mds_fail_restart()
246 self
.fs
.wait_for_daemons()
248 self
.mount_a
.create_n_files("foo/", 3*mds_max_caps_per_client
, sync
=True)
250 mount_a_client_id
= self
.mount_a
.get_global_id()
252 num_caps
= self
.get_session(mount_a_client_id
)['num_caps']
253 if num_caps
<= mds_max_caps_per_client
:
258 self
.wait_until_true(expected_caps
, timeout
=60)